Thu, 05 Apr 2012 13:57:23 -0400
7127697: G1: remove dead code after recent concurrent mark changes
Summary: Removed lots of dead code after some recent conc mark changes.
Reviewed-by: brutisso, johnc
1 /*
2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
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 // Different defaults for different number of GC threads
40 // They were chosen by running GCOld and SPECjbb on debris with different
41 // numbers of GC threads and choosing them based on the results
43 // all the same
44 static double rs_length_diff_defaults[] = {
45 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
46 };
48 static double cost_per_card_ms_defaults[] = {
49 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
50 };
52 // all the same
53 static double young_cards_per_entry_ratio_defaults[] = {
54 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
55 };
57 static double cost_per_entry_ms_defaults[] = {
58 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005
59 };
61 static double cost_per_byte_ms_defaults[] = {
62 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
63 };
65 // these should be pretty consistent
66 static double constant_other_time_ms_defaults[] = {
67 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
68 };
71 static double young_other_cost_per_region_ms_defaults[] = {
72 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
73 };
75 static double non_young_other_cost_per_region_ms_defaults[] = {
76 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
77 };
79 // Help class for avoiding interleaved logging
80 class LineBuffer: public StackObj {
82 private:
83 static const int BUFFER_LEN = 1024;
84 static const int INDENT_CHARS = 3;
85 char _buffer[BUFFER_LEN];
86 int _indent_level;
87 int _cur;
89 void vappend(const char* format, va_list ap) {
90 int res = vsnprintf(&_buffer[_cur], BUFFER_LEN - _cur, format, ap);
91 if (res != -1) {
92 _cur += res;
93 } else {
94 DEBUG_ONLY(warning("buffer too small in LineBuffer");)
95 _buffer[BUFFER_LEN -1] = 0;
96 _cur = BUFFER_LEN; // vsnprintf above should not add to _buffer if we are called again
97 }
98 }
100 public:
101 explicit LineBuffer(int indent_level): _indent_level(indent_level), _cur(0) {
102 for (; (_cur < BUFFER_LEN && _cur < (_indent_level * INDENT_CHARS)); _cur++) {
103 _buffer[_cur] = ' ';
104 }
105 }
107 #ifndef PRODUCT
108 ~LineBuffer() {
109 assert(_cur == _indent_level * INDENT_CHARS, "pending data in buffer - append_and_print_cr() not called?");
110 }
111 #endif
113 void append(const char* format, ...) {
114 va_list ap;
115 va_start(ap, format);
116 vappend(format, ap);
117 va_end(ap);
118 }
120 void append_and_print_cr(const char* format, ...) {
121 va_list ap;
122 va_start(ap, format);
123 vappend(format, ap);
124 va_end(ap);
125 gclog_or_tty->print_cr("%s", _buffer);
126 _cur = _indent_level * INDENT_CHARS;
127 }
128 };
130 G1CollectorPolicy::G1CollectorPolicy() :
131 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads()
132 ? ParallelGCThreads : 1),
134 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
135 _all_pause_times_ms(new NumberSeq()),
136 _stop_world_start(0.0),
137 _all_stop_world_times_ms(new NumberSeq()),
138 _all_yield_times_ms(new NumberSeq()),
140 _summary(new Summary()),
142 _cur_clear_ct_time_ms(0.0),
143 _root_region_scan_wait_time_ms(0.0),
145 _cur_ref_proc_time_ms(0.0),
146 _cur_ref_enq_time_ms(0.0),
148 #ifndef PRODUCT
149 _min_clear_cc_time_ms(-1.0),
150 _max_clear_cc_time_ms(-1.0),
151 _cur_clear_cc_time_ms(0.0),
152 _cum_clear_cc_time_ms(0.0),
153 _num_cc_clears(0L),
154 #endif
156 _aux_num(10),
157 _all_aux_times_ms(new NumberSeq[_aux_num]),
158 _cur_aux_start_times_ms(new double[_aux_num]),
159 _cur_aux_times_ms(new double[_aux_num]),
160 _cur_aux_times_set(new bool[_aux_num]),
162 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
163 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
165 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
166 _prev_collection_pause_end_ms(0.0),
167 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
168 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
169 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
170 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
171 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
172 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
173 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
174 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
175 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
176 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
177 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
178 _non_young_other_cost_per_region_ms_seq(
179 new TruncatedSeq(TruncatedSeqLength)),
181 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
182 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
184 _pause_time_target_ms((double) MaxGCPauseMillis),
186 _gcs_are_young(true),
187 _young_pause_num(0),
188 _mixed_pause_num(0),
190 _during_marking(false),
191 _in_marking_window(false),
192 _in_marking_window_im(false),
194 _known_garbage_ratio(0.0),
195 _known_garbage_bytes(0),
197 _young_gc_eff_seq(new TruncatedSeq(TruncatedSeqLength)),
199 _recent_prev_end_times_for_all_gcs_sec(
200 new TruncatedSeq(NumPrevPausesForHeuristics)),
202 _recent_avg_pause_time_ratio(0.0),
204 _all_full_gc_times_ms(new NumberSeq()),
206 _initiate_conc_mark_if_possible(false),
207 _during_initial_mark_pause(false),
208 _last_young_gc(false),
209 _last_gc_was_young(false),
211 _eden_bytes_before_gc(0),
212 _survivor_bytes_before_gc(0),
213 _capacity_before_gc(0),
215 _eden_cset_region_length(0),
216 _survivor_cset_region_length(0),
217 _old_cset_region_length(0),
219 _collection_set(NULL),
220 _collection_set_bytes_used_before(0),
222 // Incremental CSet attributes
223 _inc_cset_build_state(Inactive),
224 _inc_cset_head(NULL),
225 _inc_cset_tail(NULL),
226 _inc_cset_bytes_used_before(0),
227 _inc_cset_max_finger(NULL),
228 _inc_cset_recorded_rs_lengths(0),
229 _inc_cset_recorded_rs_lengths_diffs(0),
230 _inc_cset_predicted_elapsed_time_ms(0.0),
231 _inc_cset_predicted_elapsed_time_ms_diffs(0.0),
233 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
234 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
235 #endif // _MSC_VER
237 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
238 G1YoungSurvRateNumRegionsSummary)),
239 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
240 G1YoungSurvRateNumRegionsSummary)),
241 // add here any more surv rate groups
242 _recorded_survivor_regions(0),
243 _recorded_survivor_head(NULL),
244 _recorded_survivor_tail(NULL),
245 _survivors_age_table(true),
247 _gc_overhead_perc(0.0) {
249 // Set up the region size and associated fields. Given that the
250 // policy is created before the heap, we have to set this up here,
251 // so it's done as soon as possible.
252 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
253 HeapRegionRemSet::setup_remset_size();
255 G1ErgoVerbose::initialize();
256 if (PrintAdaptiveSizePolicy) {
257 // Currently, we only use a single switch for all the heuristics.
258 G1ErgoVerbose::set_enabled(true);
259 // Given that we don't currently have a verboseness level
260 // parameter, we'll hardcode this to high. This can be easily
261 // changed in the future.
262 G1ErgoVerbose::set_level(ErgoHigh);
263 } else {
264 G1ErgoVerbose::set_enabled(false);
265 }
267 // Verify PLAB sizes
268 const size_t region_size = HeapRegion::GrainWords;
269 if (YoungPLABSize > region_size || OldPLABSize > region_size) {
270 char buffer[128];
271 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT,
272 OldPLABSize > region_size ? "Old" : "Young", region_size);
273 vm_exit_during_initialization(buffer);
274 }
276 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
277 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
279 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads];
280 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads];
281 _par_last_satb_filtering_times_ms = new double[_parallel_gc_threads];
283 _par_last_update_rs_times_ms = new double[_parallel_gc_threads];
284 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads];
286 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads];
288 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads];
290 _par_last_termination_times_ms = new double[_parallel_gc_threads];
291 _par_last_termination_attempts = new double[_parallel_gc_threads];
292 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads];
293 _par_last_gc_worker_times_ms = new double[_parallel_gc_threads];
294 _par_last_gc_worker_other_times_ms = new double[_parallel_gc_threads];
296 int index;
297 if (ParallelGCThreads == 0)
298 index = 0;
299 else if (ParallelGCThreads > 8)
300 index = 7;
301 else
302 index = ParallelGCThreads - 1;
304 _pending_card_diff_seq->add(0.0);
305 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
306 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
307 _young_cards_per_entry_ratio_seq->add(
308 young_cards_per_entry_ratio_defaults[index]);
309 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
310 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
311 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
312 _young_other_cost_per_region_ms_seq->add(
313 young_other_cost_per_region_ms_defaults[index]);
314 _non_young_other_cost_per_region_ms_seq->add(
315 non_young_other_cost_per_region_ms_defaults[index]);
317 // Below, we might need to calculate the pause time target based on
318 // the pause interval. When we do so we are going to give G1 maximum
319 // flexibility and allow it to do pauses when it needs to. So, we'll
320 // arrange that the pause interval to be pause time target + 1 to
321 // ensure that a) the pause time target is maximized with respect to
322 // the pause interval and b) we maintain the invariant that pause
323 // time target < pause interval. If the user does not want this
324 // maximum flexibility, they will have to set the pause interval
325 // explicitly.
327 // First make sure that, if either parameter is set, its value is
328 // reasonable.
329 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
330 if (MaxGCPauseMillis < 1) {
331 vm_exit_during_initialization("MaxGCPauseMillis should be "
332 "greater than 0");
333 }
334 }
335 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
336 if (GCPauseIntervalMillis < 1) {
337 vm_exit_during_initialization("GCPauseIntervalMillis should be "
338 "greater than 0");
339 }
340 }
342 // Then, if the pause time target parameter was not set, set it to
343 // the default value.
344 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
345 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
346 // The default pause time target in G1 is 200ms
347 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
348 } else {
349 // We do not allow the pause interval to be set without the
350 // pause time target
351 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
352 "without setting MaxGCPauseMillis");
353 }
354 }
356 // Then, if the interval parameter was not set, set it according to
357 // the pause time target (this will also deal with the case when the
358 // pause time target is the default value).
359 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
360 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
361 }
363 // Finally, make sure that the two parameters are consistent.
364 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
365 char buffer[256];
366 jio_snprintf(buffer, 256,
367 "MaxGCPauseMillis (%u) should be less than "
368 "GCPauseIntervalMillis (%u)",
369 MaxGCPauseMillis, GCPauseIntervalMillis);
370 vm_exit_during_initialization(buffer);
371 }
373 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
374 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
375 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
376 _sigma = (double) G1ConfidencePercent / 100.0;
378 // start conservatively (around 50ms is about right)
379 _concurrent_mark_remark_times_ms->add(0.05);
380 _concurrent_mark_cleanup_times_ms->add(0.20);
381 _tenuring_threshold = MaxTenuringThreshold;
382 // _max_survivor_regions will be calculated by
383 // update_young_list_target_length() during initialization.
384 _max_survivor_regions = 0;
386 assert(GCTimeRatio > 0,
387 "we should have set it to a default value set_g1_gc_flags() "
388 "if a user set it to 0");
389 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
391 uintx reserve_perc = G1ReservePercent;
392 // Put an artificial ceiling on this so that it's not set to a silly value.
393 if (reserve_perc > 50) {
394 reserve_perc = 50;
395 warning("G1ReservePercent is set to a value that is too large, "
396 "it's been updated to %u", reserve_perc);
397 }
398 _reserve_factor = (double) reserve_perc / 100.0;
399 // This will be set when the heap is expanded
400 // for the first time during initialization.
401 _reserve_regions = 0;
403 initialize_all();
404 _collectionSetChooser = new CollectionSetChooser();
405 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags
406 }
408 void G1CollectorPolicy::initialize_flags() {
409 set_min_alignment(HeapRegion::GrainBytes);
410 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
411 if (SurvivorRatio < 1) {
412 vm_exit_during_initialization("Invalid survivor ratio specified");
413 }
414 CollectorPolicy::initialize_flags();
415 }
417 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true) {
418 assert(G1DefaultMinNewGenPercent <= G1DefaultMaxNewGenPercent, "Min larger than max");
419 assert(G1DefaultMinNewGenPercent > 0 && G1DefaultMinNewGenPercent < 100, "Min out of bounds");
420 assert(G1DefaultMaxNewGenPercent > 0 && G1DefaultMaxNewGenPercent < 100, "Max out of bounds");
422 if (FLAG_IS_CMDLINE(NewRatio)) {
423 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
424 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
425 } else {
426 _sizer_kind = SizerNewRatio;
427 _adaptive_size = false;
428 return;
429 }
430 }
432 if (FLAG_IS_CMDLINE(NewSize)) {
433 _min_desired_young_length = MAX2((size_t) 1, NewSize / HeapRegion::GrainBytes);
434 if (FLAG_IS_CMDLINE(MaxNewSize)) {
435 _max_desired_young_length = MAX2((size_t) 1, MaxNewSize / HeapRegion::GrainBytes);
436 _sizer_kind = SizerMaxAndNewSize;
437 _adaptive_size = _min_desired_young_length == _max_desired_young_length;
438 } else {
439 _sizer_kind = SizerNewSizeOnly;
440 }
441 } else if (FLAG_IS_CMDLINE(MaxNewSize)) {
442 _max_desired_young_length = MAX2((size_t) 1, MaxNewSize / HeapRegion::GrainBytes);
443 _sizer_kind = SizerMaxNewSizeOnly;
444 }
445 }
447 size_t G1YoungGenSizer::calculate_default_min_length(size_t new_number_of_heap_regions) {
448 size_t default_value = (new_number_of_heap_regions * G1DefaultMinNewGenPercent) / 100;
449 return MAX2((size_t)1, default_value);
450 }
452 size_t G1YoungGenSizer::calculate_default_max_length(size_t new_number_of_heap_regions) {
453 size_t default_value = (new_number_of_heap_regions * G1DefaultMaxNewGenPercent) / 100;
454 return MAX2((size_t)1, default_value);
455 }
457 void G1YoungGenSizer::heap_size_changed(size_t new_number_of_heap_regions) {
458 assert(new_number_of_heap_regions > 0, "Heap must be initialized");
460 switch (_sizer_kind) {
461 case SizerDefaults:
462 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
463 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
464 break;
465 case SizerNewSizeOnly:
466 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
467 _max_desired_young_length = MAX2(_min_desired_young_length, _max_desired_young_length);
468 break;
469 case SizerMaxNewSizeOnly:
470 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
471 _min_desired_young_length = MIN2(_min_desired_young_length, _max_desired_young_length);
472 break;
473 case SizerMaxAndNewSize:
474 // Do nothing. Values set on the command line, don't update them at runtime.
475 break;
476 case SizerNewRatio:
477 _min_desired_young_length = new_number_of_heap_regions / (NewRatio + 1);
478 _max_desired_young_length = _min_desired_young_length;
479 break;
480 default:
481 ShouldNotReachHere();
482 }
484 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values");
485 }
487 void G1CollectorPolicy::init() {
488 // Set aside an initial future to_space.
489 _g1 = G1CollectedHeap::heap();
491 assert(Heap_lock->owned_by_self(), "Locking discipline.");
493 initialize_gc_policy_counters();
495 if (adaptive_young_list_length()) {
496 _young_list_fixed_length = 0;
497 } else {
498 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length();
499 }
500 _free_regions_at_end_of_collection = _g1->free_regions();
501 update_young_list_target_length();
502 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes;
504 // We may immediately start allocating regions and placing them on the
505 // collection set list. Initialize the per-collection set info
506 start_incremental_cset_building();
507 }
509 // Create the jstat counters for the policy.
510 void G1CollectorPolicy::initialize_gc_policy_counters() {
511 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
512 }
514 bool G1CollectorPolicy::predict_will_fit(size_t young_length,
515 double base_time_ms,
516 size_t base_free_regions,
517 double target_pause_time_ms) {
518 if (young_length >= base_free_regions) {
519 // end condition 1: not enough space for the young regions
520 return false;
521 }
523 double accum_surv_rate = accum_yg_surv_rate_pred((int)(young_length - 1));
524 size_t bytes_to_copy =
525 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
526 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
527 double young_other_time_ms = predict_young_other_time_ms(young_length);
528 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
529 if (pause_time_ms > target_pause_time_ms) {
530 // end condition 2: prediction is over the target pause time
531 return false;
532 }
534 size_t free_bytes =
535 (base_free_regions - young_length) * HeapRegion::GrainBytes;
536 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
537 // end condition 3: out-of-space (conservatively!)
538 return false;
539 }
541 // success!
542 return true;
543 }
545 void G1CollectorPolicy::record_new_heap_size(size_t new_number_of_regions) {
546 // re-calculate the necessary reserve
547 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
548 // We use ceiling so that if reserve_regions_d is > 0.0 (but
549 // smaller than 1.0) we'll get 1.
550 _reserve_regions = (size_t) ceil(reserve_regions_d);
552 _young_gen_sizer->heap_size_changed(new_number_of_regions);
553 }
555 size_t G1CollectorPolicy::calculate_young_list_desired_min_length(
556 size_t base_min_length) {
557 size_t desired_min_length = 0;
558 if (adaptive_young_list_length()) {
559 if (_alloc_rate_ms_seq->num() > 3) {
560 double now_sec = os::elapsedTime();
561 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
562 double alloc_rate_ms = predict_alloc_rate_ms();
563 desired_min_length = (size_t) ceil(alloc_rate_ms * when_ms);
564 } else {
565 // otherwise we don't have enough info to make the prediction
566 }
567 }
568 desired_min_length += base_min_length;
569 // make sure we don't go below any user-defined minimum bound
570 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length);
571 }
573 size_t G1CollectorPolicy::calculate_young_list_desired_max_length() {
574 // Here, we might want to also take into account any additional
575 // constraints (i.e., user-defined minimum bound). Currently, we
576 // effectively don't set this bound.
577 return _young_gen_sizer->max_desired_young_length();
578 }
580 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
581 if (rs_lengths == (size_t) -1) {
582 // if it's set to the default value (-1), we should predict it;
583 // otherwise, use the given value.
584 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
585 }
587 // Calculate the absolute and desired min bounds.
589 // This is how many young regions we already have (currently: the survivors).
590 size_t base_min_length = recorded_survivor_regions();
591 // This is the absolute minimum young length, which ensures that we
592 // can allocate one eden region in the worst-case.
593 size_t absolute_min_length = base_min_length + 1;
594 size_t desired_min_length =
595 calculate_young_list_desired_min_length(base_min_length);
596 if (desired_min_length < absolute_min_length) {
597 desired_min_length = absolute_min_length;
598 }
600 // Calculate the absolute and desired max bounds.
602 // We will try our best not to "eat" into the reserve.
603 size_t absolute_max_length = 0;
604 if (_free_regions_at_end_of_collection > _reserve_regions) {
605 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
606 }
607 size_t desired_max_length = calculate_young_list_desired_max_length();
608 if (desired_max_length > absolute_max_length) {
609 desired_max_length = absolute_max_length;
610 }
612 size_t young_list_target_length = 0;
613 if (adaptive_young_list_length()) {
614 if (gcs_are_young()) {
615 young_list_target_length =
616 calculate_young_list_target_length(rs_lengths,
617 base_min_length,
618 desired_min_length,
619 desired_max_length);
620 _rs_lengths_prediction = rs_lengths;
621 } else {
622 // Don't calculate anything and let the code below bound it to
623 // the desired_min_length, i.e., do the next GC as soon as
624 // possible to maximize how many old regions we can add to it.
625 }
626 } else {
627 // The user asked for a fixed young gen so we'll fix the young gen
628 // whether the next GC is young or mixed.
629 young_list_target_length = _young_list_fixed_length;
630 }
632 // Make sure we don't go over the desired max length, nor under the
633 // desired min length. In case they clash, desired_min_length wins
634 // which is why that test is second.
635 if (young_list_target_length > desired_max_length) {
636 young_list_target_length = desired_max_length;
637 }
638 if (young_list_target_length < desired_min_length) {
639 young_list_target_length = desired_min_length;
640 }
642 assert(young_list_target_length > recorded_survivor_regions(),
643 "we should be able to allocate at least one eden region");
644 assert(young_list_target_length >= absolute_min_length, "post-condition");
645 _young_list_target_length = young_list_target_length;
647 update_max_gc_locker_expansion();
648 }
650 size_t
651 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
652 size_t base_min_length,
653 size_t desired_min_length,
654 size_t desired_max_length) {
655 assert(adaptive_young_list_length(), "pre-condition");
656 assert(gcs_are_young(), "only call this for young GCs");
658 // In case some edge-condition makes the desired max length too small...
659 if (desired_max_length <= desired_min_length) {
660 return desired_min_length;
661 }
663 // We'll adjust min_young_length and max_young_length not to include
664 // the already allocated young regions (i.e., so they reflect the
665 // min and max eden regions we'll allocate). The base_min_length
666 // will be reflected in the predictions by the
667 // survivor_regions_evac_time prediction.
668 assert(desired_min_length > base_min_length, "invariant");
669 size_t min_young_length = desired_min_length - base_min_length;
670 assert(desired_max_length > base_min_length, "invariant");
671 size_t max_young_length = desired_max_length - base_min_length;
673 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
674 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
675 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
676 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
677 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
678 double base_time_ms =
679 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
680 survivor_regions_evac_time;
681 size_t available_free_regions = _free_regions_at_end_of_collection;
682 size_t base_free_regions = 0;
683 if (available_free_regions > _reserve_regions) {
684 base_free_regions = available_free_regions - _reserve_regions;
685 }
687 // Here, we will make sure that the shortest young length that
688 // makes sense fits within the target pause time.
690 if (predict_will_fit(min_young_length, base_time_ms,
691 base_free_regions, target_pause_time_ms)) {
692 // The shortest young length will fit into the target pause time;
693 // we'll now check whether the absolute maximum number of young
694 // regions will fit in the target pause time. If not, we'll do
695 // a binary search between min_young_length and max_young_length.
696 if (predict_will_fit(max_young_length, base_time_ms,
697 base_free_regions, target_pause_time_ms)) {
698 // The maximum young length will fit into the target pause time.
699 // We are done so set min young length to the maximum length (as
700 // the result is assumed to be returned in min_young_length).
701 min_young_length = max_young_length;
702 } else {
703 // The maximum possible number of young regions will not fit within
704 // the target pause time so we'll search for the optimal
705 // length. The loop invariants are:
706 //
707 // min_young_length < max_young_length
708 // min_young_length is known to fit into the target pause time
709 // max_young_length is known not to fit into the target pause time
710 //
711 // Going into the loop we know the above hold as we've just
712 // checked them. Every time around the loop we check whether
713 // the middle value between min_young_length and
714 // max_young_length fits into the target pause time. If it
715 // does, it becomes the new min. If it doesn't, it becomes
716 // the new max. This way we maintain the loop invariants.
718 assert(min_young_length < max_young_length, "invariant");
719 size_t diff = (max_young_length - min_young_length) / 2;
720 while (diff > 0) {
721 size_t young_length = min_young_length + diff;
722 if (predict_will_fit(young_length, base_time_ms,
723 base_free_regions, target_pause_time_ms)) {
724 min_young_length = young_length;
725 } else {
726 max_young_length = young_length;
727 }
728 assert(min_young_length < max_young_length, "invariant");
729 diff = (max_young_length - min_young_length) / 2;
730 }
731 // The results is min_young_length which, according to the
732 // loop invariants, should fit within the target pause time.
734 // These are the post-conditions of the binary search above:
735 assert(min_young_length < max_young_length,
736 "otherwise we should have discovered that max_young_length "
737 "fits into the pause target and not done the binary search");
738 assert(predict_will_fit(min_young_length, base_time_ms,
739 base_free_regions, target_pause_time_ms),
740 "min_young_length, the result of the binary search, should "
741 "fit into the pause target");
742 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
743 base_free_regions, target_pause_time_ms),
744 "min_young_length, the result of the binary search, should be "
745 "optimal, so no larger length should fit into the pause target");
746 }
747 } else {
748 // Even the minimum length doesn't fit into the pause time
749 // target, return it as the result nevertheless.
750 }
751 return base_min_length + min_young_length;
752 }
754 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
755 double survivor_regions_evac_time = 0.0;
756 for (HeapRegion * r = _recorded_survivor_head;
757 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
758 r = r->get_next_young_region()) {
759 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
760 }
761 return survivor_regions_evac_time;
762 }
764 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
765 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
767 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
768 if (rs_lengths > _rs_lengths_prediction) {
769 // add 10% to avoid having to recalculate often
770 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
771 update_young_list_target_length(rs_lengths_prediction);
772 }
773 }
777 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
778 bool is_tlab,
779 bool* gc_overhead_limit_was_exceeded) {
780 guarantee(false, "Not using this policy feature yet.");
781 return NULL;
782 }
784 // This method controls how a collector handles one or more
785 // of its generations being fully allocated.
786 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
787 bool is_tlab) {
788 guarantee(false, "Not using this policy feature yet.");
789 return NULL;
790 }
793 #ifndef PRODUCT
794 bool G1CollectorPolicy::verify_young_ages() {
795 HeapRegion* head = _g1->young_list()->first_region();
796 return
797 verify_young_ages(head, _short_lived_surv_rate_group);
798 // also call verify_young_ages on any additional surv rate groups
799 }
801 bool
802 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
803 SurvRateGroup *surv_rate_group) {
804 guarantee( surv_rate_group != NULL, "pre-condition" );
806 const char* name = surv_rate_group->name();
807 bool ret = true;
808 int prev_age = -1;
810 for (HeapRegion* curr = head;
811 curr != NULL;
812 curr = curr->get_next_young_region()) {
813 SurvRateGroup* group = curr->surv_rate_group();
814 if (group == NULL && !curr->is_survivor()) {
815 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
816 ret = false;
817 }
819 if (surv_rate_group == group) {
820 int age = curr->age_in_surv_rate_group();
822 if (age < 0) {
823 gclog_or_tty->print_cr("## %s: encountered negative age", name);
824 ret = false;
825 }
827 if (age <= prev_age) {
828 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
829 "(%d, %d)", name, age, prev_age);
830 ret = false;
831 }
832 prev_age = age;
833 }
834 }
836 return ret;
837 }
838 #endif // PRODUCT
840 void G1CollectorPolicy::record_full_collection_start() {
841 _cur_collection_start_sec = os::elapsedTime();
842 // Release the future to-space so that it is available for compaction into.
843 _g1->set_full_collection();
844 }
846 void G1CollectorPolicy::record_full_collection_end() {
847 // Consider this like a collection pause for the purposes of allocation
848 // since last pause.
849 double end_sec = os::elapsedTime();
850 double full_gc_time_sec = end_sec - _cur_collection_start_sec;
851 double full_gc_time_ms = full_gc_time_sec * 1000.0;
853 _all_full_gc_times_ms->add(full_gc_time_ms);
855 update_recent_gc_times(end_sec, full_gc_time_ms);
857 _g1->clear_full_collection();
859 // "Nuke" the heuristics that control the young/mixed GC
860 // transitions and make sure we start with young GCs after the Full GC.
861 set_gcs_are_young(true);
862 _last_young_gc = false;
863 clear_initiate_conc_mark_if_possible();
864 clear_during_initial_mark_pause();
865 _known_garbage_bytes = 0;
866 _known_garbage_ratio = 0.0;
867 _in_marking_window = false;
868 _in_marking_window_im = false;
870 _short_lived_surv_rate_group->start_adding_regions();
871 // also call this on any additional surv rate groups
873 record_survivor_regions(0, NULL, NULL);
875 _free_regions_at_end_of_collection = _g1->free_regions();
876 // Reset survivors SurvRateGroup.
877 _survivor_surv_rate_group->reset();
878 update_young_list_target_length();
879 _collectionSetChooser->clearMarkedHeapRegions();
880 }
882 void G1CollectorPolicy::record_stop_world_start() {
883 _stop_world_start = os::elapsedTime();
884 }
886 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
887 size_t start_used) {
888 if (PrintGCDetails) {
889 gclog_or_tty->stamp(PrintGCTimeStamps);
890 gclog_or_tty->print("[GC pause");
891 gclog_or_tty->print(" (%s)", gcs_are_young() ? "young" : "mixed");
892 }
894 // We only need to do this here as the policy will only be applied
895 // to the GC we're about to start. so, no point is calculating this
896 // every time we calculate / recalculate the target young length.
897 update_survivors_policy();
899 assert(_g1->used() == _g1->recalculate_used(),
900 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
901 _g1->used(), _g1->recalculate_used()));
903 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
904 _all_stop_world_times_ms->add(s_w_t_ms);
905 _stop_world_start = 0.0;
907 _cur_collection_start_sec = start_time_sec;
908 _cur_collection_pause_used_at_start_bytes = start_used;
909 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
910 _pending_cards = _g1->pending_card_num();
911 _max_pending_cards = _g1->max_pending_card_num();
913 _bytes_in_collection_set_before_gc = 0;
914 _bytes_copied_during_gc = 0;
916 YoungList* young_list = _g1->young_list();
917 _eden_bytes_before_gc = young_list->eden_used_bytes();
918 _survivor_bytes_before_gc = young_list->survivor_used_bytes();
919 _capacity_before_gc = _g1->capacity();
921 #ifdef DEBUG
922 // initialise these to something well known so that we can spot
923 // if they are not set properly
925 for (int i = 0; i < _parallel_gc_threads; ++i) {
926 _par_last_gc_worker_start_times_ms[i] = -1234.0;
927 _par_last_ext_root_scan_times_ms[i] = -1234.0;
928 _par_last_satb_filtering_times_ms[i] = -1234.0;
929 _par_last_update_rs_times_ms[i] = -1234.0;
930 _par_last_update_rs_processed_buffers[i] = -1234.0;
931 _par_last_scan_rs_times_ms[i] = -1234.0;
932 _par_last_obj_copy_times_ms[i] = -1234.0;
933 _par_last_termination_times_ms[i] = -1234.0;
934 _par_last_termination_attempts[i] = -1234.0;
935 _par_last_gc_worker_end_times_ms[i] = -1234.0;
936 _par_last_gc_worker_times_ms[i] = -1234.0;
937 _par_last_gc_worker_other_times_ms[i] = -1234.0;
938 }
939 #endif
941 for (int i = 0; i < _aux_num; ++i) {
942 _cur_aux_times_ms[i] = 0.0;
943 _cur_aux_times_set[i] = false;
944 }
946 // This is initialized to zero here and is set during the evacuation
947 // pause if we actually waited for the root region scanning to finish.
948 _root_region_scan_wait_time_ms = 0.0;
950 _last_gc_was_young = false;
952 // do that for any other surv rate groups
953 _short_lived_surv_rate_group->stop_adding_regions();
954 _survivors_age_table.clear();
956 assert( verify_young_ages(), "region age verification" );
957 }
959 void G1CollectorPolicy::record_concurrent_mark_init_end(double
960 mark_init_elapsed_time_ms) {
961 _during_marking = true;
962 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
963 clear_during_initial_mark_pause();
964 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
965 }
967 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
968 _mark_remark_start_sec = os::elapsedTime();
969 _during_marking = false;
970 }
972 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
973 double end_time_sec = os::elapsedTime();
974 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
975 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
976 _cur_mark_stop_world_time_ms += elapsed_time_ms;
977 _prev_collection_pause_end_ms += elapsed_time_ms;
979 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
980 }
982 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
983 _mark_cleanup_start_sec = os::elapsedTime();
984 }
986 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
987 _last_young_gc = true;
988 _in_marking_window = false;
989 }
991 void G1CollectorPolicy::record_concurrent_pause() {
992 if (_stop_world_start > 0.0) {
993 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
994 _all_yield_times_ms->add(yield_ms);
995 }
996 }
998 void G1CollectorPolicy::record_concurrent_pause_end() {
999 }
1001 template<class T>
1002 T sum_of(T* sum_arr, int start, int n, int N) {
1003 T sum = (T)0;
1004 for (int i = 0; i < n; i++) {
1005 int j = (start + i) % N;
1006 sum += sum_arr[j];
1007 }
1008 return sum;
1009 }
1011 void G1CollectorPolicy::print_par_stats(int level,
1012 const char* str,
1013 double* data) {
1014 double min = data[0], max = data[0];
1015 double total = 0.0;
1016 LineBuffer buf(level);
1017 buf.append("[%s (ms):", str);
1018 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1019 double val = data[i];
1020 if (val < min)
1021 min = val;
1022 if (val > max)
1023 max = val;
1024 total += val;
1025 buf.append(" %3.1lf", val);
1026 }
1027 buf.append_and_print_cr("");
1028 double avg = total / (double) no_of_gc_threads();
1029 buf.append_and_print_cr(" Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf, Diff: %5.1lf]",
1030 avg, min, max, max - min);
1031 }
1033 void G1CollectorPolicy::print_par_sizes(int level,
1034 const char* str,
1035 double* data) {
1036 double min = data[0], max = data[0];
1037 double total = 0.0;
1038 LineBuffer buf(level);
1039 buf.append("[%s :", str);
1040 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1041 double val = data[i];
1042 if (val < min)
1043 min = val;
1044 if (val > max)
1045 max = val;
1046 total += val;
1047 buf.append(" %d", (int) val);
1048 }
1049 buf.append_and_print_cr("");
1050 double avg = total / (double) no_of_gc_threads();
1051 buf.append_and_print_cr(" Sum: %d, Avg: %d, Min: %d, Max: %d, Diff: %d]",
1052 (int)total, (int)avg, (int)min, (int)max, (int)max - (int)min);
1053 }
1055 void G1CollectorPolicy::print_stats(int level,
1056 const char* str,
1057 double value) {
1058 LineBuffer(level).append_and_print_cr("[%s: %5.1lf ms]", str, value);
1059 }
1061 void G1CollectorPolicy::print_stats(int level,
1062 const char* str,
1063 int value) {
1064 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value);
1065 }
1067 double G1CollectorPolicy::avg_value(double* data) {
1068 if (G1CollectedHeap::use_parallel_gc_threads()) {
1069 double ret = 0.0;
1070 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1071 ret += data[i];
1072 }
1073 return ret / (double) no_of_gc_threads();
1074 } else {
1075 return data[0];
1076 }
1077 }
1079 double G1CollectorPolicy::max_value(double* data) {
1080 if (G1CollectedHeap::use_parallel_gc_threads()) {
1081 double ret = data[0];
1082 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1083 if (data[i] > ret) {
1084 ret = data[i];
1085 }
1086 }
1087 return ret;
1088 } else {
1089 return data[0];
1090 }
1091 }
1093 double G1CollectorPolicy::sum_of_values(double* data) {
1094 if (G1CollectedHeap::use_parallel_gc_threads()) {
1095 double sum = 0.0;
1096 for (uint i = 0; i < no_of_gc_threads(); i++) {
1097 sum += data[i];
1098 }
1099 return sum;
1100 } else {
1101 return data[0];
1102 }
1103 }
1105 double G1CollectorPolicy::max_sum(double* data1, double* data2) {
1106 double ret = data1[0] + data2[0];
1108 if (G1CollectedHeap::use_parallel_gc_threads()) {
1109 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1110 double data = data1[i] + data2[i];
1111 if (data > ret) {
1112 ret = data;
1113 }
1114 }
1115 }
1116 return ret;
1117 }
1119 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) {
1120 if (_g1->concurrent_mark()->cmThread()->during_cycle()) {
1121 return false;
1122 }
1124 size_t marking_initiating_used_threshold =
1125 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
1126 size_t cur_used_bytes = _g1->non_young_capacity_bytes();
1127 size_t alloc_byte_size = alloc_word_size * HeapWordSize;
1129 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) {
1130 if (gcs_are_young()) {
1131 ergo_verbose5(ErgoConcCycles,
1132 "request concurrent cycle initiation",
1133 ergo_format_reason("occupancy higher than threshold")
1134 ergo_format_byte("occupancy")
1135 ergo_format_byte("allocation request")
1136 ergo_format_byte_perc("threshold")
1137 ergo_format_str("source"),
1138 cur_used_bytes,
1139 alloc_byte_size,
1140 marking_initiating_used_threshold,
1141 (double) InitiatingHeapOccupancyPercent,
1142 source);
1143 return true;
1144 } else {
1145 ergo_verbose5(ErgoConcCycles,
1146 "do not request concurrent cycle initiation",
1147 ergo_format_reason("still doing mixed collections")
1148 ergo_format_byte("occupancy")
1149 ergo_format_byte("allocation request")
1150 ergo_format_byte_perc("threshold")
1151 ergo_format_str("source"),
1152 cur_used_bytes,
1153 alloc_byte_size,
1154 marking_initiating_used_threshold,
1155 (double) InitiatingHeapOccupancyPercent,
1156 source);
1157 }
1158 }
1160 return false;
1161 }
1163 // Anything below that is considered to be zero
1164 #define MIN_TIMER_GRANULARITY 0.0000001
1166 void G1CollectorPolicy::record_collection_pause_end(int no_of_gc_threads) {
1167 double end_time_sec = os::elapsedTime();
1168 double elapsed_ms = _last_pause_time_ms;
1169 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1170 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(),
1171 "otherwise, the subtraction below does not make sense");
1172 size_t rs_size =
1173 _cur_collection_pause_used_regions_at_start - cset_region_length();
1174 size_t cur_used_bytes = _g1->used();
1175 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1176 bool last_pause_included_initial_mark = false;
1177 bool update_stats = !_g1->evacuation_failed();
1178 set_no_of_gc_threads(no_of_gc_threads);
1180 #ifndef PRODUCT
1181 if (G1YoungSurvRateVerbose) {
1182 gclog_or_tty->print_cr("");
1183 _short_lived_surv_rate_group->print();
1184 // do that for any other surv rate groups too
1185 }
1186 #endif // PRODUCT
1188 last_pause_included_initial_mark = during_initial_mark_pause();
1189 if (last_pause_included_initial_mark) {
1190 record_concurrent_mark_init_end(0.0);
1191 } else if (!_last_young_gc && need_to_start_conc_mark("end of GC")) {
1192 // Note: this might have already been set, if during the last
1193 // pause we decided to start a cycle but at the beginning of
1194 // this pause we decided to postpone it. That's OK.
1195 set_initiate_conc_mark_if_possible();
1196 }
1198 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1199 end_time_sec, false);
1201 // This assert is exempted when we're doing parallel collection pauses,
1202 // because the fragmentation caused by the parallel GC allocation buffers
1203 // can lead to more memory being used during collection than was used
1204 // before. Best leave this out until the fragmentation problem is fixed.
1205 // Pauses in which evacuation failed can also lead to negative
1206 // collections, since no space is reclaimed from a region containing an
1207 // object whose evacuation failed.
1208 // Further, we're now always doing parallel collection. But I'm still
1209 // leaving this here as a placeholder for a more precise assertion later.
1210 // (DLD, 10/05.)
1211 assert((true || parallel) // Always using GC LABs now.
1212 || _g1->evacuation_failed()
1213 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1214 "Negative collection");
1216 size_t freed_bytes =
1217 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1218 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1220 double survival_fraction =
1221 (double)surviving_bytes/
1222 (double)_collection_set_bytes_used_before;
1224 // These values are used to update the summary information that is
1225 // displayed when TraceGen0Time is enabled, and are output as part
1226 // of the PrintGCDetails output, in the non-parallel case.
1228 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1229 double satb_filtering_time = avg_value(_par_last_satb_filtering_times_ms);
1230 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1231 double update_rs_processed_buffers =
1232 sum_of_values(_par_last_update_rs_processed_buffers);
1233 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1234 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1235 double termination_time = avg_value(_par_last_termination_times_ms);
1237 double known_time = ext_root_scan_time +
1238 satb_filtering_time +
1239 update_rs_time +
1240 scan_rs_time +
1241 obj_copy_time;
1243 double other_time_ms = elapsed_ms;
1245 // Subtract the root region scanning wait time. It's initialized to
1246 // zero at the start of the pause.
1247 other_time_ms -= _root_region_scan_wait_time_ms;
1249 if (parallel) {
1250 other_time_ms -= _cur_collection_par_time_ms;
1251 } else {
1252 other_time_ms -= known_time;
1253 }
1255 // Now subtract the time taken to fix up roots in generated code
1256 other_time_ms -= _cur_collection_code_root_fixup_time_ms;
1258 // Subtract the time taken to clean the card table from the
1259 // current value of "other time"
1260 other_time_ms -= _cur_clear_ct_time_ms;
1262 // TraceGen0Time and TraceGen1Time summary info updating.
1263 _all_pause_times_ms->add(elapsed_ms);
1265 if (update_stats) {
1266 _summary->record_total_time_ms(elapsed_ms);
1267 _summary->record_other_time_ms(other_time_ms);
1269 MainBodySummary* body_summary = _summary->main_body_summary();
1270 assert(body_summary != NULL, "should not be null!");
1272 body_summary->record_root_region_scan_wait_time_ms(
1273 _root_region_scan_wait_time_ms);
1274 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1275 body_summary->record_satb_filtering_time_ms(satb_filtering_time);
1276 body_summary->record_update_rs_time_ms(update_rs_time);
1277 body_summary->record_scan_rs_time_ms(scan_rs_time);
1278 body_summary->record_obj_copy_time_ms(obj_copy_time);
1280 if (parallel) {
1281 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1282 body_summary->record_termination_time_ms(termination_time);
1284 double parallel_known_time = known_time + termination_time;
1285 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time;
1286 body_summary->record_parallel_other_time_ms(parallel_other_time);
1287 }
1289 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1291 // We exempt parallel collection from this check because Alloc Buffer
1292 // fragmentation can produce negative collections. Same with evac
1293 // failure.
1294 // Further, we're now always doing parallel collection. But I'm still
1295 // leaving this here as a placeholder for a more precise assertion later.
1296 // (DLD, 10/05.
1297 assert((true || parallel)
1298 || _g1->evacuation_failed()
1299 || surviving_bytes <= _collection_set_bytes_used_before,
1300 "Or else negative collection!");
1302 // this is where we update the allocation rate of the application
1303 double app_time_ms =
1304 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1305 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1306 // This usually happens due to the timer not having the required
1307 // granularity. Some Linuxes are the usual culprits.
1308 // We'll just set it to something (arbitrarily) small.
1309 app_time_ms = 1.0;
1310 }
1311 // We maintain the invariant that all objects allocated by mutator
1312 // threads will be allocated out of eden regions. So, we can use
1313 // the eden region number allocated since the previous GC to
1314 // calculate the application's allocate rate. The only exception
1315 // to that is humongous objects that are allocated separately. But
1316 // given that humongous object allocations do not really affect
1317 // either the pause's duration nor when the next pause will take
1318 // place we can safely ignore them here.
1319 size_t regions_allocated = eden_cset_region_length();
1320 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1321 _alloc_rate_ms_seq->add(alloc_rate_ms);
1323 double interval_ms =
1324 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1325 update_recent_gc_times(end_time_sec, elapsed_ms);
1326 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1327 if (recent_avg_pause_time_ratio() < 0.0 ||
1328 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1329 #ifndef PRODUCT
1330 // Dump info to allow post-facto debugging
1331 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1332 gclog_or_tty->print_cr("-------------------------------------------");
1333 gclog_or_tty->print_cr("Recent GC Times (ms):");
1334 _recent_gc_times_ms->dump();
1335 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1336 _recent_prev_end_times_for_all_gcs_sec->dump();
1337 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1338 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1339 // In debug mode, terminate the JVM if the user wants to debug at this point.
1340 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1341 #endif // !PRODUCT
1342 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1343 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1344 if (_recent_avg_pause_time_ratio < 0.0) {
1345 _recent_avg_pause_time_ratio = 0.0;
1346 } else {
1347 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1348 _recent_avg_pause_time_ratio = 1.0;
1349 }
1350 }
1351 }
1353 for (int i = 0; i < _aux_num; ++i) {
1354 if (_cur_aux_times_set[i]) {
1355 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1356 }
1357 }
1359 // PrintGCDetails output
1360 if (PrintGCDetails) {
1361 bool print_marking_info =
1362 _g1->mark_in_progress() && !last_pause_included_initial_mark;
1364 gclog_or_tty->print_cr("%s, %1.8lf secs]",
1365 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1366 elapsed_ms / 1000.0);
1368 if (_root_region_scan_wait_time_ms > 0.0) {
1369 print_stats(1, "Root Region Scan Waiting", _root_region_scan_wait_time_ms);
1370 }
1371 if (parallel) {
1372 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1373 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms);
1374 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1375 if (print_marking_info) {
1376 print_par_stats(2, "SATB Filtering", _par_last_satb_filtering_times_ms);
1377 }
1378 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1379 print_par_sizes(3, "Processed Buffers", _par_last_update_rs_processed_buffers);
1380 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1381 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1382 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1383 print_par_sizes(3, "Termination Attempts", _par_last_termination_attempts);
1385 for (int i = 0; i < _parallel_gc_threads; i++) {
1386 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] -
1387 _par_last_gc_worker_start_times_ms[i];
1389 double worker_known_time = _par_last_ext_root_scan_times_ms[i] +
1390 _par_last_satb_filtering_times_ms[i] +
1391 _par_last_update_rs_times_ms[i] +
1392 _par_last_scan_rs_times_ms[i] +
1393 _par_last_obj_copy_times_ms[i] +
1394 _par_last_termination_times_ms[i];
1396 _par_last_gc_worker_other_times_ms[i] = _par_last_gc_worker_times_ms[i] -
1397 worker_known_time;
1398 }
1400 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms);
1401 print_par_stats(2, "GC Worker Total", _par_last_gc_worker_times_ms);
1402 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms);
1403 } else {
1404 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1405 if (print_marking_info) {
1406 print_stats(1, "SATB Filtering", satb_filtering_time);
1407 }
1408 print_stats(1, "Update RS", update_rs_time);
1409 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers);
1410 print_stats(1, "Scan RS", scan_rs_time);
1411 print_stats(1, "Object Copying", obj_copy_time);
1412 }
1413 print_stats(1, "Code Root Fixup", _cur_collection_code_root_fixup_time_ms);
1414 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1415 #ifndef PRODUCT
1416 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1417 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1418 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1419 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1420 if (_num_cc_clears > 0) {
1421 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1422 }
1423 #endif
1424 print_stats(1, "Other", other_time_ms);
1425 print_stats(2, "Choose CSet",
1426 (_recorded_young_cset_choice_time_ms +
1427 _recorded_non_young_cset_choice_time_ms));
1428 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms);
1429 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms);
1430 print_stats(2, "Free CSet",
1431 (_recorded_young_free_cset_time_ms +
1432 _recorded_non_young_free_cset_time_ms));
1434 for (int i = 0; i < _aux_num; ++i) {
1435 if (_cur_aux_times_set[i]) {
1436 char buffer[96];
1437 sprintf(buffer, "Aux%d", i);
1438 print_stats(1, buffer, _cur_aux_times_ms[i]);
1439 }
1440 }
1441 }
1443 // Update the efficiency-since-mark vars.
1444 double proc_ms = elapsed_ms * (double) _parallel_gc_threads;
1445 if (elapsed_ms < MIN_TIMER_GRANULARITY) {
1446 // This usually happens due to the timer not having the required
1447 // granularity. Some Linuxes are the usual culprits.
1448 // We'll just set it to something (arbitrarily) small.
1449 proc_ms = 1.0;
1450 }
1451 double cur_efficiency = (double) freed_bytes / proc_ms;
1453 bool new_in_marking_window = _in_marking_window;
1454 bool new_in_marking_window_im = false;
1455 if (during_initial_mark_pause()) {
1456 new_in_marking_window = true;
1457 new_in_marking_window_im = true;
1458 }
1460 if (_last_young_gc) {
1461 // This is supposed to to be the "last young GC" before we start
1462 // doing mixed GCs. Here we decide whether to start mixed GCs or not.
1464 if (!last_pause_included_initial_mark) {
1465 if (next_gc_should_be_mixed("start mixed GCs",
1466 "do not start mixed GCs")) {
1467 set_gcs_are_young(false);
1468 }
1469 } else {
1470 ergo_verbose0(ErgoMixedGCs,
1471 "do not start mixed GCs",
1472 ergo_format_reason("concurrent cycle is about to start"));
1473 }
1474 _last_young_gc = false;
1475 }
1477 if (!_last_gc_was_young) {
1478 // This is a mixed GC. Here we decide whether to continue doing
1479 // mixed GCs or not.
1481 if (!next_gc_should_be_mixed("continue mixed GCs",
1482 "do not continue mixed GCs")) {
1483 set_gcs_are_young(true);
1484 }
1485 }
1487 if (_last_gc_was_young && !_during_marking) {
1488 _young_gc_eff_seq->add(cur_efficiency);
1489 }
1491 _short_lived_surv_rate_group->start_adding_regions();
1492 // do that for any other surv rate groupsx
1494 if (update_stats) {
1495 double pause_time_ms = elapsed_ms;
1497 size_t diff = 0;
1498 if (_max_pending_cards >= _pending_cards)
1499 diff = _max_pending_cards - _pending_cards;
1500 _pending_card_diff_seq->add((double) diff);
1502 double cost_per_card_ms = 0.0;
1503 if (_pending_cards > 0) {
1504 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1505 _cost_per_card_ms_seq->add(cost_per_card_ms);
1506 }
1508 size_t cards_scanned = _g1->cards_scanned();
1510 double cost_per_entry_ms = 0.0;
1511 if (cards_scanned > 10) {
1512 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1513 if (_last_gc_was_young) {
1514 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1515 } else {
1516 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1517 }
1518 }
1520 if (_max_rs_lengths > 0) {
1521 double cards_per_entry_ratio =
1522 (double) cards_scanned / (double) _max_rs_lengths;
1523 if (_last_gc_was_young) {
1524 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1525 } else {
1526 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1527 }
1528 }
1530 // This is defensive. For a while _max_rs_lengths could get
1531 // smaller than _recorded_rs_lengths which was causing
1532 // rs_length_diff to get very large and mess up the RSet length
1533 // predictions. The reason was unsafe concurrent updates to the
1534 // _inc_cset_recorded_rs_lengths field which the code below guards
1535 // against (see CR 7118202). This bug has now been fixed (see CR
1536 // 7119027). However, I'm still worried that
1537 // _inc_cset_recorded_rs_lengths might still end up somewhat
1538 // inaccurate. The concurrent refinement thread calculates an
1539 // RSet's length concurrently with other CR threads updating it
1540 // which might cause it to calculate the length incorrectly (if,
1541 // say, it's in mid-coarsening). So I'll leave in the defensive
1542 // conditional below just in case.
1543 size_t rs_length_diff = 0;
1544 if (_max_rs_lengths > _recorded_rs_lengths) {
1545 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1546 }
1547 _rs_length_diff_seq->add((double) rs_length_diff);
1549 size_t copied_bytes = surviving_bytes;
1550 double cost_per_byte_ms = 0.0;
1551 if (copied_bytes > 0) {
1552 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1553 if (_in_marking_window) {
1554 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1555 } else {
1556 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1557 }
1558 }
1560 double all_other_time_ms = pause_time_ms -
1561 (update_rs_time + scan_rs_time + obj_copy_time + termination_time);
1563 double young_other_time_ms = 0.0;
1564 if (young_cset_region_length() > 0) {
1565 young_other_time_ms =
1566 _recorded_young_cset_choice_time_ms +
1567 _recorded_young_free_cset_time_ms;
1568 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1569 (double) young_cset_region_length());
1570 }
1571 double non_young_other_time_ms = 0.0;
1572 if (old_cset_region_length() > 0) {
1573 non_young_other_time_ms =
1574 _recorded_non_young_cset_choice_time_ms +
1575 _recorded_non_young_free_cset_time_ms;
1577 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1578 (double) old_cset_region_length());
1579 }
1581 double constant_other_time_ms = all_other_time_ms -
1582 (young_other_time_ms + non_young_other_time_ms);
1583 _constant_other_time_ms_seq->add(constant_other_time_ms);
1585 double survival_ratio = 0.0;
1586 if (_bytes_in_collection_set_before_gc > 0) {
1587 survival_ratio = (double) _bytes_copied_during_gc /
1588 (double) _bytes_in_collection_set_before_gc;
1589 }
1591 _pending_cards_seq->add((double) _pending_cards);
1592 _rs_lengths_seq->add((double) _max_rs_lengths);
1593 }
1595 _in_marking_window = new_in_marking_window;
1596 _in_marking_window_im = new_in_marking_window_im;
1597 _free_regions_at_end_of_collection = _g1->free_regions();
1598 update_young_list_target_length();
1600 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1601 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1602 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1604 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end.");
1605 }
1607 #define EXT_SIZE_FORMAT "%d%s"
1608 #define EXT_SIZE_PARAMS(bytes) \
1609 byte_size_in_proper_unit((bytes)), \
1610 proper_unit_for_byte_size((bytes))
1612 void G1CollectorPolicy::print_heap_transition() {
1613 if (PrintGCDetails) {
1614 YoungList* young_list = _g1->young_list();
1615 size_t eden_bytes = young_list->eden_used_bytes();
1616 size_t survivor_bytes = young_list->survivor_used_bytes();
1617 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes;
1618 size_t used = _g1->used();
1619 size_t capacity = _g1->capacity();
1620 size_t eden_capacity =
1621 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes;
1623 gclog_or_tty->print_cr(
1624 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1625 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1626 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1627 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1628 EXT_SIZE_PARAMS(_eden_bytes_before_gc),
1629 EXT_SIZE_PARAMS(_prev_eden_capacity),
1630 EXT_SIZE_PARAMS(eden_bytes),
1631 EXT_SIZE_PARAMS(eden_capacity),
1632 EXT_SIZE_PARAMS(_survivor_bytes_before_gc),
1633 EXT_SIZE_PARAMS(survivor_bytes),
1634 EXT_SIZE_PARAMS(used_before_gc),
1635 EXT_SIZE_PARAMS(_capacity_before_gc),
1636 EXT_SIZE_PARAMS(used),
1637 EXT_SIZE_PARAMS(capacity));
1639 _prev_eden_capacity = eden_capacity;
1640 } else if (PrintGC) {
1641 _g1->print_size_transition(gclog_or_tty,
1642 _cur_collection_pause_used_at_start_bytes,
1643 _g1->used(), _g1->capacity());
1644 }
1645 }
1647 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1648 double update_rs_processed_buffers,
1649 double goal_ms) {
1650 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1651 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1653 if (G1UseAdaptiveConcRefinement) {
1654 const int k_gy = 3, k_gr = 6;
1655 const double inc_k = 1.1, dec_k = 0.9;
1657 int g = cg1r->green_zone();
1658 if (update_rs_time > goal_ms) {
1659 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1660 } else {
1661 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1662 g = (int)MAX2(g * inc_k, g + 1.0);
1663 }
1664 }
1665 // Change the refinement threads params
1666 cg1r->set_green_zone(g);
1667 cg1r->set_yellow_zone(g * k_gy);
1668 cg1r->set_red_zone(g * k_gr);
1669 cg1r->reinitialize_threads();
1671 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1672 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1673 cg1r->yellow_zone());
1674 // Change the barrier params
1675 dcqs.set_process_completed_threshold(processing_threshold);
1676 dcqs.set_max_completed_queue(cg1r->red_zone());
1677 }
1679 int curr_queue_size = dcqs.completed_buffers_num();
1680 if (curr_queue_size >= cg1r->yellow_zone()) {
1681 dcqs.set_completed_queue_padding(curr_queue_size);
1682 } else {
1683 dcqs.set_completed_queue_padding(0);
1684 }
1685 dcqs.notify_if_necessary();
1686 }
1688 double
1689 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1690 size_t rs_length = predict_rs_length_diff();
1691 size_t card_num;
1692 if (gcs_are_young()) {
1693 card_num = predict_young_card_num(rs_length);
1694 } else {
1695 card_num = predict_non_young_card_num(rs_length);
1696 }
1697 return predict_base_elapsed_time_ms(pending_cards, card_num);
1698 }
1700 double
1701 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1702 size_t scanned_cards) {
1703 return
1704 predict_rs_update_time_ms(pending_cards) +
1705 predict_rs_scan_time_ms(scanned_cards) +
1706 predict_constant_other_time_ms();
1707 }
1709 double
1710 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1711 bool young) {
1712 size_t rs_length = hr->rem_set()->occupied();
1713 size_t card_num;
1714 if (gcs_are_young()) {
1715 card_num = predict_young_card_num(rs_length);
1716 } else {
1717 card_num = predict_non_young_card_num(rs_length);
1718 }
1719 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1721 double region_elapsed_time_ms =
1722 predict_rs_scan_time_ms(card_num) +
1723 predict_object_copy_time_ms(bytes_to_copy);
1725 if (young)
1726 region_elapsed_time_ms += predict_young_other_time_ms(1);
1727 else
1728 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1730 return region_elapsed_time_ms;
1731 }
1733 size_t
1734 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1735 size_t bytes_to_copy;
1736 if (hr->is_marked())
1737 bytes_to_copy = hr->max_live_bytes();
1738 else {
1739 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant");
1740 int age = hr->age_in_surv_rate_group();
1741 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1742 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1743 }
1744 return bytes_to_copy;
1745 }
1747 void
1748 G1CollectorPolicy::init_cset_region_lengths(size_t eden_cset_region_length,
1749 size_t survivor_cset_region_length) {
1750 _eden_cset_region_length = eden_cset_region_length;
1751 _survivor_cset_region_length = survivor_cset_region_length;
1752 _old_cset_region_length = 0;
1753 }
1755 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1756 _recorded_rs_lengths = rs_lengths;
1757 }
1759 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1760 double elapsed_ms) {
1761 _recent_gc_times_ms->add(elapsed_ms);
1762 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1763 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1764 }
1766 size_t G1CollectorPolicy::expansion_amount() {
1767 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1768 double threshold = _gc_overhead_perc;
1769 if (recent_gc_overhead > threshold) {
1770 // We will double the existing space, or take
1771 // G1ExpandByPercentOfAvailable % of the available expansion
1772 // space, whichever is smaller, bounded below by a minimum
1773 // expansion (unless that's all that's left.)
1774 const size_t min_expand_bytes = 1*M;
1775 size_t reserved_bytes = _g1->max_capacity();
1776 size_t committed_bytes = _g1->capacity();
1777 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1778 size_t expand_bytes;
1779 size_t expand_bytes_via_pct =
1780 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1781 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1782 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1783 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1785 ergo_verbose5(ErgoHeapSizing,
1786 "attempt heap expansion",
1787 ergo_format_reason("recent GC overhead higher than "
1788 "threshold after GC")
1789 ergo_format_perc("recent GC overhead")
1790 ergo_format_perc("threshold")
1791 ergo_format_byte("uncommitted")
1792 ergo_format_byte_perc("calculated expansion amount"),
1793 recent_gc_overhead, threshold,
1794 uncommitted_bytes,
1795 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
1797 return expand_bytes;
1798 } else {
1799 return 0;
1800 }
1801 }
1803 class CountCSClosure: public HeapRegionClosure {
1804 G1CollectorPolicy* _g1_policy;
1805 public:
1806 CountCSClosure(G1CollectorPolicy* g1_policy) :
1807 _g1_policy(g1_policy) {}
1808 bool doHeapRegion(HeapRegion* r) {
1809 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
1810 return false;
1811 }
1812 };
1814 void G1CollectorPolicy::count_CS_bytes_used() {
1815 CountCSClosure cs_closure(this);
1816 _g1->collection_set_iterate(&cs_closure);
1817 }
1819 void G1CollectorPolicy::print_summary(int level,
1820 const char* str,
1821 NumberSeq* seq) const {
1822 double sum = seq->sum();
1823 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
1824 str, sum / 1000.0, seq->avg());
1825 }
1827 void G1CollectorPolicy::print_summary_sd(int level,
1828 const char* str,
1829 NumberSeq* seq) const {
1830 print_summary(level, str, seq);
1831 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
1832 seq->num(), seq->sd(), seq->maximum());
1833 }
1835 void G1CollectorPolicy::check_other_times(int level,
1836 NumberSeq* other_times_ms,
1837 NumberSeq* calc_other_times_ms) const {
1838 bool should_print = false;
1839 LineBuffer buf(level + 2);
1841 double max_sum = MAX2(fabs(other_times_ms->sum()),
1842 fabs(calc_other_times_ms->sum()));
1843 double min_sum = MIN2(fabs(other_times_ms->sum()),
1844 fabs(calc_other_times_ms->sum()));
1845 double sum_ratio = max_sum / min_sum;
1846 if (sum_ratio > 1.1) {
1847 should_print = true;
1848 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
1849 }
1851 double max_avg = MAX2(fabs(other_times_ms->avg()),
1852 fabs(calc_other_times_ms->avg()));
1853 double min_avg = MIN2(fabs(other_times_ms->avg()),
1854 fabs(calc_other_times_ms->avg()));
1855 double avg_ratio = max_avg / min_avg;
1856 if (avg_ratio > 1.1) {
1857 should_print = true;
1858 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
1859 }
1861 if (other_times_ms->sum() < -0.01) {
1862 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
1863 }
1865 if (other_times_ms->avg() < -0.01) {
1866 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
1867 }
1869 if (calc_other_times_ms->sum() < -0.01) {
1870 should_print = true;
1871 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
1872 }
1874 if (calc_other_times_ms->avg() < -0.01) {
1875 should_print = true;
1876 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
1877 }
1879 if (should_print)
1880 print_summary(level, "Other(Calc)", calc_other_times_ms);
1881 }
1883 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
1884 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1885 MainBodySummary* body_summary = summary->main_body_summary();
1886 if (summary->get_total_seq()->num() > 0) {
1887 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
1888 if (body_summary != NULL) {
1889 print_summary(1, "Root Region Scan Wait", body_summary->get_root_region_scan_wait_seq());
1890 if (parallel) {
1891 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
1892 print_summary(2, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
1893 print_summary(2, "SATB Filtering", body_summary->get_satb_filtering_seq());
1894 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
1895 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
1896 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
1897 print_summary(2, "Termination", body_summary->get_termination_seq());
1898 print_summary(2, "Parallel Other", body_summary->get_parallel_other_seq());
1899 {
1900 NumberSeq* other_parts[] = {
1901 body_summary->get_ext_root_scan_seq(),
1902 body_summary->get_satb_filtering_seq(),
1903 body_summary->get_update_rs_seq(),
1904 body_summary->get_scan_rs_seq(),
1905 body_summary->get_obj_copy_seq(),
1906 body_summary->get_termination_seq()
1907 };
1908 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
1909 6, other_parts);
1910 check_other_times(2, body_summary->get_parallel_other_seq(),
1911 &calc_other_times_ms);
1912 }
1913 } else {
1914 print_summary(1, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
1915 print_summary(1, "SATB Filtering", body_summary->get_satb_filtering_seq());
1916 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
1917 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
1918 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
1919 }
1920 }
1921 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
1922 print_summary(1, "Other", summary->get_other_seq());
1923 {
1924 if (body_summary != NULL) {
1925 NumberSeq calc_other_times_ms;
1926 if (parallel) {
1927 // parallel
1928 NumberSeq* other_parts[] = {
1929 body_summary->get_root_region_scan_wait_seq(),
1930 body_summary->get_parallel_seq(),
1931 body_summary->get_clear_ct_seq()
1932 };
1933 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
1934 3, other_parts);
1935 } else {
1936 // serial
1937 NumberSeq* other_parts[] = {
1938 body_summary->get_root_region_scan_wait_seq(),
1939 body_summary->get_update_rs_seq(),
1940 body_summary->get_ext_root_scan_seq(),
1941 body_summary->get_satb_filtering_seq(),
1942 body_summary->get_scan_rs_seq(),
1943 body_summary->get_obj_copy_seq()
1944 };
1945 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
1946 6, other_parts);
1947 }
1948 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
1949 }
1950 }
1951 } else {
1952 LineBuffer(1).append_and_print_cr("none");
1953 }
1954 LineBuffer(0).append_and_print_cr("");
1955 }
1957 void G1CollectorPolicy::print_tracing_info() const {
1958 if (TraceGen0Time) {
1959 gclog_or_tty->print_cr("ALL PAUSES");
1960 print_summary_sd(0, "Total", _all_pause_times_ms);
1961 gclog_or_tty->print_cr("");
1962 gclog_or_tty->print_cr("");
1963 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
1964 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
1965 gclog_or_tty->print_cr("");
1967 gclog_or_tty->print_cr("EVACUATION PAUSES");
1968 print_summary(_summary);
1970 gclog_or_tty->print_cr("MISC");
1971 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
1972 print_summary_sd(0, "Yields", _all_yield_times_ms);
1973 for (int i = 0; i < _aux_num; ++i) {
1974 if (_all_aux_times_ms[i].num() > 0) {
1975 char buffer[96];
1976 sprintf(buffer, "Aux%d", i);
1977 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
1978 }
1979 }
1980 }
1981 if (TraceGen1Time) {
1982 if (_all_full_gc_times_ms->num() > 0) {
1983 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
1984 _all_full_gc_times_ms->num(),
1985 _all_full_gc_times_ms->sum() / 1000.0);
1986 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
1987 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
1988 _all_full_gc_times_ms->sd(),
1989 _all_full_gc_times_ms->maximum());
1990 }
1991 }
1992 }
1994 void G1CollectorPolicy::print_yg_surv_rate_info() const {
1995 #ifndef PRODUCT
1996 _short_lived_surv_rate_group->print_surv_rate_summary();
1997 // add this call for any other surv rate groups
1998 #endif // PRODUCT
1999 }
2001 #ifndef PRODUCT
2002 // for debugging, bit of a hack...
2003 static char*
2004 region_num_to_mbs(int length) {
2005 static char buffer[64];
2006 double bytes = (double) (length * HeapRegion::GrainBytes);
2007 double mbs = bytes / (double) (1024 * 1024);
2008 sprintf(buffer, "%7.2lfMB", mbs);
2009 return buffer;
2010 }
2011 #endif // PRODUCT
2013 size_t G1CollectorPolicy::max_regions(int purpose) {
2014 switch (purpose) {
2015 case GCAllocForSurvived:
2016 return _max_survivor_regions;
2017 case GCAllocForTenured:
2018 return REGIONS_UNLIMITED;
2019 default:
2020 ShouldNotReachHere();
2021 return REGIONS_UNLIMITED;
2022 };
2023 }
2025 void G1CollectorPolicy::update_max_gc_locker_expansion() {
2026 size_t expansion_region_num = 0;
2027 if (GCLockerEdenExpansionPercent > 0) {
2028 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
2029 double expansion_region_num_d = perc * (double) _young_list_target_length;
2030 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
2031 // less than 1.0) we'll get 1.
2032 expansion_region_num = (size_t) ceil(expansion_region_num_d);
2033 } else {
2034 assert(expansion_region_num == 0, "sanity");
2035 }
2036 _young_list_max_length = _young_list_target_length + expansion_region_num;
2037 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
2038 }
2040 // Calculates survivor space parameters.
2041 void G1CollectorPolicy::update_survivors_policy() {
2042 double max_survivor_regions_d =
2043 (double) _young_list_target_length / (double) SurvivorRatio;
2044 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
2045 // smaller than 1.0) we'll get 1.
2046 _max_survivor_regions = (size_t) ceil(max_survivor_regions_d);
2048 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2049 HeapRegion::GrainWords * _max_survivor_regions);
2050 }
2052 #ifndef PRODUCT
2053 class HRSortIndexIsOKClosure: public HeapRegionClosure {
2054 CollectionSetChooser* _chooser;
2055 public:
2056 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) :
2057 _chooser(chooser) {}
2059 bool doHeapRegion(HeapRegion* r) {
2060 if (!r->continuesHumongous()) {
2061 assert(_chooser->regionProperlyOrdered(r), "Ought to be.");
2062 }
2063 return false;
2064 }
2065 };
2067 bool G1CollectorPolicy::assertMarkedBytesDataOK() {
2068 HRSortIndexIsOKClosure cl(_collectionSetChooser);
2069 _g1->heap_region_iterate(&cl);
2070 return true;
2071 }
2072 #endif
2074 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
2075 GCCause::Cause gc_cause) {
2076 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2077 if (!during_cycle) {
2078 ergo_verbose1(ErgoConcCycles,
2079 "request concurrent cycle initiation",
2080 ergo_format_reason("requested by GC cause")
2081 ergo_format_str("GC cause"),
2082 GCCause::to_string(gc_cause));
2083 set_initiate_conc_mark_if_possible();
2084 return true;
2085 } else {
2086 ergo_verbose1(ErgoConcCycles,
2087 "do not request concurrent cycle initiation",
2088 ergo_format_reason("concurrent cycle already in progress")
2089 ergo_format_str("GC cause"),
2090 GCCause::to_string(gc_cause));
2091 return false;
2092 }
2093 }
2095 void
2096 G1CollectorPolicy::decide_on_conc_mark_initiation() {
2097 // We are about to decide on whether this pause will be an
2098 // initial-mark pause.
2100 // First, during_initial_mark_pause() should not be already set. We
2101 // will set it here if we have to. However, it should be cleared by
2102 // the end of the pause (it's only set for the duration of an
2103 // initial-mark pause).
2104 assert(!during_initial_mark_pause(), "pre-condition");
2106 if (initiate_conc_mark_if_possible()) {
2107 // We had noticed on a previous pause that the heap occupancy has
2108 // gone over the initiating threshold and we should start a
2109 // concurrent marking cycle. So we might initiate one.
2111 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2112 if (!during_cycle) {
2113 // The concurrent marking thread is not "during a cycle", i.e.,
2114 // it has completed the last one. So we can go ahead and
2115 // initiate a new cycle.
2117 set_during_initial_mark_pause();
2118 // We do not allow mixed GCs during marking.
2119 if (!gcs_are_young()) {
2120 set_gcs_are_young(true);
2121 ergo_verbose0(ErgoMixedGCs,
2122 "end mixed GCs",
2123 ergo_format_reason("concurrent cycle is about to start"));
2124 }
2126 // And we can now clear initiate_conc_mark_if_possible() as
2127 // we've already acted on it.
2128 clear_initiate_conc_mark_if_possible();
2130 ergo_verbose0(ErgoConcCycles,
2131 "initiate concurrent cycle",
2132 ergo_format_reason("concurrent cycle initiation requested"));
2133 } else {
2134 // The concurrent marking thread is still finishing up the
2135 // previous cycle. If we start one right now the two cycles
2136 // overlap. In particular, the concurrent marking thread might
2137 // be in the process of clearing the next marking bitmap (which
2138 // we will use for the next cycle if we start one). Starting a
2139 // cycle now will be bad given that parts of the marking
2140 // information might get cleared by the marking thread. And we
2141 // cannot wait for the marking thread to finish the cycle as it
2142 // periodically yields while clearing the next marking bitmap
2143 // and, if it's in a yield point, it's waiting for us to
2144 // finish. So, at this point we will not start a cycle and we'll
2145 // let the concurrent marking thread complete the last one.
2146 ergo_verbose0(ErgoConcCycles,
2147 "do not initiate concurrent cycle",
2148 ergo_format_reason("concurrent cycle already in progress"));
2149 }
2150 }
2151 }
2153 class KnownGarbageClosure: public HeapRegionClosure {
2154 G1CollectedHeap* _g1h;
2155 CollectionSetChooser* _hrSorted;
2157 public:
2158 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2159 _g1h(G1CollectedHeap::heap()), _hrSorted(hrSorted) { }
2161 bool doHeapRegion(HeapRegion* r) {
2162 // We only include humongous regions in collection
2163 // sets when concurrent mark shows that their contained object is
2164 // unreachable.
2166 // Do we have any marking information for this region?
2167 if (r->is_marked()) {
2168 // We will skip any region that's currently used as an old GC
2169 // alloc region (we should not consider those for collection
2170 // before we fill them up).
2171 if (_hrSorted->shouldAdd(r) && !_g1h->is_old_gc_alloc_region(r)) {
2172 _hrSorted->addMarkedHeapRegion(r);
2173 }
2174 }
2175 return false;
2176 }
2177 };
2179 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2180 G1CollectedHeap* _g1h;
2181 CollectionSetChooser* _hrSorted;
2182 jint _marked_regions_added;
2183 size_t _reclaimable_bytes_added;
2184 jint _chunk_size;
2185 jint _cur_chunk_idx;
2186 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2187 int _worker;
2188 int _invokes;
2190 void get_new_chunk() {
2191 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size);
2192 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2193 }
2194 void add_region(HeapRegion* r) {
2195 if (_cur_chunk_idx == _cur_chunk_end) {
2196 get_new_chunk();
2197 }
2198 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2199 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r);
2200 _marked_regions_added++;
2201 _reclaimable_bytes_added += r->reclaimable_bytes();
2202 _cur_chunk_idx++;
2203 }
2205 public:
2206 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2207 jint chunk_size,
2208 int worker) :
2209 _g1h(G1CollectedHeap::heap()),
2210 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker),
2211 _marked_regions_added(0), _reclaimable_bytes_added(0),
2212 _cur_chunk_idx(0), _cur_chunk_end(0), _invokes(0) { }
2214 bool doHeapRegion(HeapRegion* r) {
2215 // We only include humongous regions in collection
2216 // sets when concurrent mark shows that their contained object is
2217 // unreachable.
2218 _invokes++;
2220 // Do we have any marking information for this region?
2221 if (r->is_marked()) {
2222 // We will skip any region that's currently used as an old GC
2223 // alloc region (we should not consider those for collection
2224 // before we fill them up).
2225 if (_hrSorted->shouldAdd(r) && !_g1h->is_old_gc_alloc_region(r)) {
2226 add_region(r);
2227 }
2228 }
2229 return false;
2230 }
2231 jint marked_regions_added() { return _marked_regions_added; }
2232 size_t reclaimable_bytes_added() { return _reclaimable_bytes_added; }
2233 int invokes() { return _invokes; }
2234 };
2236 class ParKnownGarbageTask: public AbstractGangTask {
2237 CollectionSetChooser* _hrSorted;
2238 jint _chunk_size;
2239 G1CollectedHeap* _g1;
2240 public:
2241 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) :
2242 AbstractGangTask("ParKnownGarbageTask"),
2243 _hrSorted(hrSorted), _chunk_size(chunk_size),
2244 _g1(G1CollectedHeap::heap()) { }
2246 void work(uint worker_id) {
2247 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted,
2248 _chunk_size,
2249 worker_id);
2250 // Back to zero for the claim value.
2251 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id,
2252 _g1->workers()->active_workers(),
2253 HeapRegion::InitialClaimValue);
2254 jint regions_added = parKnownGarbageCl.marked_regions_added();
2255 size_t reclaimable_bytes_added =
2256 parKnownGarbageCl.reclaimable_bytes_added();
2257 _hrSorted->updateTotals(regions_added, reclaimable_bytes_added);
2258 if (G1PrintParCleanupStats) {
2259 gclog_or_tty->print_cr(" Thread %d called %d times, added %d regions to list.",
2260 worker_id, parKnownGarbageCl.invokes(), regions_added);
2261 }
2262 }
2263 };
2265 void
2266 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) {
2267 double start_sec;
2268 if (G1PrintParCleanupStats) {
2269 start_sec = os::elapsedTime();
2270 }
2272 _collectionSetChooser->clearMarkedHeapRegions();
2273 double clear_marked_end_sec;
2274 if (G1PrintParCleanupStats) {
2275 clear_marked_end_sec = os::elapsedTime();
2276 gclog_or_tty->print_cr(" clear marked regions: %8.3f ms.",
2277 (clear_marked_end_sec - start_sec) * 1000.0);
2278 }
2280 if (G1CollectedHeap::use_parallel_gc_threads()) {
2281 const size_t OverpartitionFactor = 4;
2282 size_t WorkUnit;
2283 // The use of MinChunkSize = 8 in the original code
2284 // causes some assertion failures when the total number of
2285 // region is less than 8. The code here tries to fix that.
2286 // Should the original code also be fixed?
2287 if (no_of_gc_threads > 0) {
2288 const size_t MinWorkUnit =
2289 MAX2(_g1->n_regions() / no_of_gc_threads, (size_t) 1U);
2290 WorkUnit =
2291 MAX2(_g1->n_regions() / (no_of_gc_threads * OverpartitionFactor),
2292 MinWorkUnit);
2293 } else {
2294 assert(no_of_gc_threads > 0,
2295 "The active gc workers should be greater than 0");
2296 // In a product build do something reasonable to avoid a crash.
2297 const size_t MinWorkUnit =
2298 MAX2(_g1->n_regions() / ParallelGCThreads, (size_t) 1U);
2299 WorkUnit =
2300 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor),
2301 MinWorkUnit);
2302 }
2303 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(),
2304 WorkUnit);
2305 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2306 (int) WorkUnit);
2307 _g1->workers()->run_task(&parKnownGarbageTask);
2309 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2310 "sanity check");
2311 } else {
2312 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2313 _g1->heap_region_iterate(&knownGarbagecl);
2314 }
2315 double known_garbage_end_sec;
2316 if (G1PrintParCleanupStats) {
2317 known_garbage_end_sec = os::elapsedTime();
2318 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.",
2319 (known_garbage_end_sec - clear_marked_end_sec) * 1000.0);
2320 }
2322 _collectionSetChooser->sortMarkedHeapRegions();
2323 double end_sec = os::elapsedTime();
2324 if (G1PrintParCleanupStats) {
2325 gclog_or_tty->print_cr(" sorting: %8.3f ms.",
2326 (end_sec - known_garbage_end_sec) * 1000.0);
2327 }
2329 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
2330 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
2331 _cur_mark_stop_world_time_ms += elapsed_time_ms;
2332 _prev_collection_pause_end_ms += elapsed_time_ms;
2333 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
2334 }
2336 // Add the heap region at the head of the non-incremental collection set
2337 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
2338 assert(_inc_cset_build_state == Active, "Precondition");
2339 assert(!hr->is_young(), "non-incremental add of young region");
2341 assert(!hr->in_collection_set(), "should not already be in the CSet");
2342 hr->set_in_collection_set(true);
2343 hr->set_next_in_collection_set(_collection_set);
2344 _collection_set = hr;
2345 _collection_set_bytes_used_before += hr->used();
2346 _g1->register_region_with_in_cset_fast_test(hr);
2347 size_t rs_length = hr->rem_set()->occupied();
2348 _recorded_rs_lengths += rs_length;
2349 _old_cset_region_length += 1;
2350 }
2352 // Initialize the per-collection-set information
2353 void G1CollectorPolicy::start_incremental_cset_building() {
2354 assert(_inc_cset_build_state == Inactive, "Precondition");
2356 _inc_cset_head = NULL;
2357 _inc_cset_tail = NULL;
2358 _inc_cset_bytes_used_before = 0;
2360 _inc_cset_max_finger = 0;
2361 _inc_cset_recorded_rs_lengths = 0;
2362 _inc_cset_recorded_rs_lengths_diffs = 0;
2363 _inc_cset_predicted_elapsed_time_ms = 0.0;
2364 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
2365 _inc_cset_build_state = Active;
2366 }
2368 void G1CollectorPolicy::finalize_incremental_cset_building() {
2369 assert(_inc_cset_build_state == Active, "Precondition");
2370 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2372 // The two "main" fields, _inc_cset_recorded_rs_lengths and
2373 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread
2374 // that adds a new region to the CSet. Further updates by the
2375 // concurrent refinement thread that samples the young RSet lengths
2376 // are accumulated in the *_diffs fields. Here we add the diffs to
2377 // the "main" fields.
2379 if (_inc_cset_recorded_rs_lengths_diffs >= 0) {
2380 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs;
2381 } else {
2382 // This is defensive. The diff should in theory be always positive
2383 // as RSets can only grow between GCs. However, given that we
2384 // sample their size concurrently with other threads updating them
2385 // it's possible that we might get the wrong size back, which
2386 // could make the calculations somewhat inaccurate.
2387 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs);
2388 if (_inc_cset_recorded_rs_lengths >= diffs) {
2389 _inc_cset_recorded_rs_lengths -= diffs;
2390 } else {
2391 _inc_cset_recorded_rs_lengths = 0;
2392 }
2393 }
2394 _inc_cset_predicted_elapsed_time_ms +=
2395 _inc_cset_predicted_elapsed_time_ms_diffs;
2397 _inc_cset_recorded_rs_lengths_diffs = 0;
2398 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
2399 }
2401 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
2402 // This routine is used when:
2403 // * adding survivor regions to the incremental cset at the end of an
2404 // evacuation pause,
2405 // * adding the current allocation region to the incremental cset
2406 // when it is retired, and
2407 // * updating existing policy information for a region in the
2408 // incremental cset via young list RSet sampling.
2409 // Therefore this routine may be called at a safepoint by the
2410 // VM thread, or in-between safepoints by mutator threads (when
2411 // retiring the current allocation region) or a concurrent
2412 // refine thread (RSet sampling).
2414 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2415 size_t used_bytes = hr->used();
2416 _inc_cset_recorded_rs_lengths += rs_length;
2417 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
2418 _inc_cset_bytes_used_before += used_bytes;
2420 // Cache the values we have added to the aggregated informtion
2421 // in the heap region in case we have to remove this region from
2422 // the incremental collection set, or it is updated by the
2423 // rset sampling code
2424 hr->set_recorded_rs_length(rs_length);
2425 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
2426 }
2428 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr,
2429 size_t new_rs_length) {
2430 // Update the CSet information that is dependent on the new RS length
2431 assert(hr->is_young(), "Precondition");
2432 assert(!SafepointSynchronize::is_at_safepoint(),
2433 "should not be at a safepoint");
2435 // We could have updated _inc_cset_recorded_rs_lengths and
2436 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do
2437 // that atomically, as this code is executed by a concurrent
2438 // refinement thread, potentially concurrently with a mutator thread
2439 // allocating a new region and also updating the same fields. To
2440 // avoid the atomic operations we accumulate these updates on two
2441 // separate fields (*_diffs) and we'll just add them to the "main"
2442 // fields at the start of a GC.
2444 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length();
2445 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length;
2446 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff;
2448 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
2449 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2450 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms;
2451 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff;
2453 hr->set_recorded_rs_length(new_rs_length);
2454 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms);
2455 }
2457 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
2458 assert(hr->is_young(), "invariant");
2459 assert(hr->young_index_in_cset() > -1, "should have already been set");
2460 assert(_inc_cset_build_state == Active, "Precondition");
2462 // We need to clear and set the cached recorded/cached collection set
2463 // information in the heap region here (before the region gets added
2464 // to the collection set). An individual heap region's cached values
2465 // are calculated, aggregated with the policy collection set info,
2466 // and cached in the heap region here (initially) and (subsequently)
2467 // by the Young List sampling code.
2469 size_t rs_length = hr->rem_set()->occupied();
2470 add_to_incremental_cset_info(hr, rs_length);
2472 HeapWord* hr_end = hr->end();
2473 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
2475 assert(!hr->in_collection_set(), "invariant");
2476 hr->set_in_collection_set(true);
2477 assert( hr->next_in_collection_set() == NULL, "invariant");
2479 _g1->register_region_with_in_cset_fast_test(hr);
2480 }
2482 // Add the region at the RHS of the incremental cset
2483 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
2484 // We should only ever be appending survivors at the end of a pause
2485 assert( hr->is_survivor(), "Logic");
2487 // Do the 'common' stuff
2488 add_region_to_incremental_cset_common(hr);
2490 // Now add the region at the right hand side
2491 if (_inc_cset_tail == NULL) {
2492 assert(_inc_cset_head == NULL, "invariant");
2493 _inc_cset_head = hr;
2494 } else {
2495 _inc_cset_tail->set_next_in_collection_set(hr);
2496 }
2497 _inc_cset_tail = hr;
2498 }
2500 // Add the region to the LHS of the incremental cset
2501 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
2502 // Survivors should be added to the RHS at the end of a pause
2503 assert(!hr->is_survivor(), "Logic");
2505 // Do the 'common' stuff
2506 add_region_to_incremental_cset_common(hr);
2508 // Add the region at the left hand side
2509 hr->set_next_in_collection_set(_inc_cset_head);
2510 if (_inc_cset_head == NULL) {
2511 assert(_inc_cset_tail == NULL, "Invariant");
2512 _inc_cset_tail = hr;
2513 }
2514 _inc_cset_head = hr;
2515 }
2517 #ifndef PRODUCT
2518 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
2519 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
2521 st->print_cr("\nCollection_set:");
2522 HeapRegion* csr = list_head;
2523 while (csr != NULL) {
2524 HeapRegion* next = csr->next_in_collection_set();
2525 assert(csr->in_collection_set(), "bad CS");
2526 st->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
2527 "age: %4d, y: %d, surv: %d",
2528 csr->bottom(), csr->end(),
2529 csr->top(),
2530 csr->prev_top_at_mark_start(),
2531 csr->next_top_at_mark_start(),
2532 csr->top_at_conc_mark_count(),
2533 csr->age_in_surv_rate_group_cond(),
2534 csr->is_young(),
2535 csr->is_survivor());
2536 csr = next;
2537 }
2538 }
2539 #endif // !PRODUCT
2541 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str,
2542 const char* false_action_str) {
2543 CollectionSetChooser* cset_chooser = _collectionSetChooser;
2544 if (cset_chooser->isEmpty()) {
2545 ergo_verbose0(ErgoMixedGCs,
2546 false_action_str,
2547 ergo_format_reason("candidate old regions not available"));
2548 return false;
2549 }
2550 size_t reclaimable_bytes = cset_chooser->remainingReclaimableBytes();
2551 size_t capacity_bytes = _g1->capacity();
2552 double perc = (double) reclaimable_bytes * 100.0 / (double) capacity_bytes;
2553 double threshold = (double) G1HeapWastePercent;
2554 if (perc < threshold) {
2555 ergo_verbose4(ErgoMixedGCs,
2556 false_action_str,
2557 ergo_format_reason("reclaimable percentage lower than threshold")
2558 ergo_format_region("candidate old regions")
2559 ergo_format_byte_perc("reclaimable")
2560 ergo_format_perc("threshold"),
2561 cset_chooser->remainingRegions(),
2562 reclaimable_bytes, perc, threshold);
2563 return false;
2564 }
2566 ergo_verbose4(ErgoMixedGCs,
2567 true_action_str,
2568 ergo_format_reason("candidate old regions available")
2569 ergo_format_region("candidate old regions")
2570 ergo_format_byte_perc("reclaimable")
2571 ergo_format_perc("threshold"),
2572 cset_chooser->remainingRegions(),
2573 reclaimable_bytes, perc, threshold);
2574 return true;
2575 }
2577 void G1CollectorPolicy::finalize_cset(double target_pause_time_ms) {
2578 // Set this here - in case we're not doing young collections.
2579 double non_young_start_time_sec = os::elapsedTime();
2581 YoungList* young_list = _g1->young_list();
2582 finalize_incremental_cset_building();
2584 guarantee(target_pause_time_ms > 0.0,
2585 err_msg("target_pause_time_ms = %1.6lf should be positive",
2586 target_pause_time_ms));
2587 guarantee(_collection_set == NULL, "Precondition");
2589 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2590 double predicted_pause_time_ms = base_time_ms;
2591 double time_remaining_ms = target_pause_time_ms - base_time_ms;
2593 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2594 "start choosing CSet",
2595 ergo_format_ms("predicted base time")
2596 ergo_format_ms("remaining time")
2597 ergo_format_ms("target pause time"),
2598 base_time_ms, time_remaining_ms, target_pause_time_ms);
2600 HeapRegion* hr;
2601 double young_start_time_sec = os::elapsedTime();
2603 _collection_set_bytes_used_before = 0;
2604 _last_gc_was_young = gcs_are_young() ? true : false;
2606 if (_last_gc_was_young) {
2607 ++_young_pause_num;
2608 } else {
2609 ++_mixed_pause_num;
2610 }
2612 // The young list is laid with the survivor regions from the previous
2613 // pause are appended to the RHS of the young list, i.e.
2614 // [Newly Young Regions ++ Survivors from last pause].
2616 size_t survivor_region_length = young_list->survivor_length();
2617 size_t eden_region_length = young_list->length() - survivor_region_length;
2618 init_cset_region_lengths(eden_region_length, survivor_region_length);
2619 hr = young_list->first_survivor_region();
2620 while (hr != NULL) {
2621 assert(hr->is_survivor(), "badly formed young list");
2622 hr->set_young();
2623 hr = hr->get_next_young_region();
2624 }
2626 // Clear the fields that point to the survivor list - they are all young now.
2627 young_list->clear_survivors();
2629 _collection_set = _inc_cset_head;
2630 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2631 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms;
2632 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
2634 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2635 "add young regions to CSet",
2636 ergo_format_region("eden")
2637 ergo_format_region("survivors")
2638 ergo_format_ms("predicted young region time"),
2639 eden_region_length, survivor_region_length,
2640 _inc_cset_predicted_elapsed_time_ms);
2642 // The number of recorded young regions is the incremental
2643 // collection set's current size
2644 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2646 double young_end_time_sec = os::elapsedTime();
2647 _recorded_young_cset_choice_time_ms =
2648 (young_end_time_sec - young_start_time_sec) * 1000.0;
2650 // We are doing young collections so reset this.
2651 non_young_start_time_sec = young_end_time_sec;
2653 if (!gcs_are_young()) {
2654 CollectionSetChooser* cset_chooser = _collectionSetChooser;
2655 assert(cset_chooser->verify(), "CSet Chooser verification - pre");
2656 const size_t min_old_cset_length = cset_chooser->calcMinOldCSetLength();
2657 const size_t max_old_cset_length = cset_chooser->calcMaxOldCSetLength();
2659 size_t expensive_region_num = 0;
2660 bool check_time_remaining = adaptive_young_list_length();
2661 HeapRegion* hr = cset_chooser->peek();
2662 while (hr != NULL) {
2663 if (old_cset_region_length() >= max_old_cset_length) {
2664 // Added maximum number of old regions to the CSet.
2665 ergo_verbose2(ErgoCSetConstruction,
2666 "finish adding old regions to CSet",
2667 ergo_format_reason("old CSet region num reached max")
2668 ergo_format_region("old")
2669 ergo_format_region("max"),
2670 old_cset_region_length(), max_old_cset_length);
2671 break;
2672 }
2674 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
2675 if (check_time_remaining) {
2676 if (predicted_time_ms > time_remaining_ms) {
2677 // Too expensive for the current CSet.
2679 if (old_cset_region_length() >= min_old_cset_length) {
2680 // We have added the minimum number of old regions to the CSet,
2681 // we are done with this CSet.
2682 ergo_verbose4(ErgoCSetConstruction,
2683 "finish adding old regions to CSet",
2684 ergo_format_reason("predicted time is too high")
2685 ergo_format_ms("predicted time")
2686 ergo_format_ms("remaining time")
2687 ergo_format_region("old")
2688 ergo_format_region("min"),
2689 predicted_time_ms, time_remaining_ms,
2690 old_cset_region_length(), min_old_cset_length);
2691 break;
2692 }
2694 // We'll add it anyway given that we haven't reached the
2695 // minimum number of old regions.
2696 expensive_region_num += 1;
2697 }
2698 } else {
2699 if (old_cset_region_length() >= min_old_cset_length) {
2700 // In the non-auto-tuning case, we'll finish adding regions
2701 // to the CSet if we reach the minimum.
2702 ergo_verbose2(ErgoCSetConstruction,
2703 "finish adding old regions to CSet",
2704 ergo_format_reason("old CSet region num reached min")
2705 ergo_format_region("old")
2706 ergo_format_region("min"),
2707 old_cset_region_length(), min_old_cset_length);
2708 break;
2709 }
2710 }
2712 // We will add this region to the CSet.
2713 time_remaining_ms -= predicted_time_ms;
2714 predicted_pause_time_ms += predicted_time_ms;
2715 cset_chooser->remove_and_move_to_next(hr);
2716 _g1->old_set_remove(hr);
2717 add_old_region_to_cset(hr);
2719 hr = cset_chooser->peek();
2720 }
2721 if (hr == NULL) {
2722 ergo_verbose0(ErgoCSetConstruction,
2723 "finish adding old regions to CSet",
2724 ergo_format_reason("candidate old regions not available"));
2725 }
2727 if (expensive_region_num > 0) {
2728 // We print the information once here at the end, predicated on
2729 // whether we added any apparently expensive regions or not, to
2730 // avoid generating output per region.
2731 ergo_verbose4(ErgoCSetConstruction,
2732 "added expensive regions to CSet",
2733 ergo_format_reason("old CSet region num not reached min")
2734 ergo_format_region("old")
2735 ergo_format_region("expensive")
2736 ergo_format_region("min")
2737 ergo_format_ms("remaining time"),
2738 old_cset_region_length(),
2739 expensive_region_num,
2740 min_old_cset_length,
2741 time_remaining_ms);
2742 }
2744 assert(cset_chooser->verify(), "CSet Chooser verification - post");
2745 }
2747 stop_incremental_cset_building();
2749 count_CS_bytes_used();
2751 ergo_verbose5(ErgoCSetConstruction,
2752 "finish choosing CSet",
2753 ergo_format_region("eden")
2754 ergo_format_region("survivors")
2755 ergo_format_region("old")
2756 ergo_format_ms("predicted pause time")
2757 ergo_format_ms("target pause time"),
2758 eden_region_length, survivor_region_length,
2759 old_cset_region_length(),
2760 predicted_pause_time_ms, target_pause_time_ms);
2762 double non_young_end_time_sec = os::elapsedTime();
2763 _recorded_non_young_cset_choice_time_ms =
2764 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
2765 }