Mon, 16 Jan 2012 22:10:05 +0100
6976060: G1: humongous object allocations should initiate marking cycles when necessary
Reviewed-by: tonyp, 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 _mark_closure_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 _should_revert_to_young_gcs(false),
209 _last_young_gc(false),
210 _last_gc_was_young(false),
212 _eden_bytes_before_gc(0),
213 _survivor_bytes_before_gc(0),
214 _capacity_before_gc(0),
216 _eden_cset_region_length(0),
217 _survivor_cset_region_length(0),
218 _old_cset_region_length(0),
220 _collection_set(NULL),
221 _collection_set_bytes_used_before(0),
223 // Incremental CSet attributes
224 _inc_cset_build_state(Inactive),
225 _inc_cset_head(NULL),
226 _inc_cset_tail(NULL),
227 _inc_cset_bytes_used_before(0),
228 _inc_cset_max_finger(NULL),
229 _inc_cset_recorded_rs_lengths(0),
230 _inc_cset_recorded_rs_lengths_diffs(0),
231 _inc_cset_predicted_elapsed_time_ms(0.0),
232 _inc_cset_predicted_elapsed_time_ms_diffs(0.0),
234 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
235 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
236 #endif // _MSC_VER
238 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
239 G1YoungSurvRateNumRegionsSummary)),
240 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
241 G1YoungSurvRateNumRegionsSummary)),
242 // add here any more surv rate groups
243 _recorded_survivor_regions(0),
244 _recorded_survivor_head(NULL),
245 _recorded_survivor_tail(NULL),
246 _survivors_age_table(true),
248 _gc_overhead_perc(0.0) {
250 // Set up the region size and associated fields. Given that the
251 // policy is created before the heap, we have to set this up here,
252 // so it's done as soon as possible.
253 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
254 HeapRegionRemSet::setup_remset_size();
256 G1ErgoVerbose::initialize();
257 if (PrintAdaptiveSizePolicy) {
258 // Currently, we only use a single switch for all the heuristics.
259 G1ErgoVerbose::set_enabled(true);
260 // Given that we don't currently have a verboseness level
261 // parameter, we'll hardcode this to high. This can be easily
262 // changed in the future.
263 G1ErgoVerbose::set_level(ErgoHigh);
264 } else {
265 G1ErgoVerbose::set_enabled(false);
266 }
268 // Verify PLAB sizes
269 const size_t region_size = HeapRegion::GrainWords;
270 if (YoungPLABSize > region_size || OldPLABSize > region_size) {
271 char buffer[128];
272 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT,
273 OldPLABSize > region_size ? "Old" : "Young", region_size);
274 vm_exit_during_initialization(buffer);
275 }
277 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
278 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
280 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads];
281 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads];
282 _par_last_satb_filtering_times_ms = new double[_parallel_gc_threads];
284 _par_last_update_rs_times_ms = new double[_parallel_gc_threads];
285 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads];
287 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads];
289 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads];
291 _par_last_termination_times_ms = new double[_parallel_gc_threads];
292 _par_last_termination_attempts = new double[_parallel_gc_threads];
293 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads];
294 _par_last_gc_worker_times_ms = new double[_parallel_gc_threads];
295 _par_last_gc_worker_other_times_ms = new double[_parallel_gc_threads];
297 // start conservatively
298 _expensive_region_limit_ms = 0.5 * (double) MaxGCPauseMillis;
300 int index;
301 if (ParallelGCThreads == 0)
302 index = 0;
303 else if (ParallelGCThreads > 8)
304 index = 7;
305 else
306 index = ParallelGCThreads - 1;
308 _pending_card_diff_seq->add(0.0);
309 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
310 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
311 _young_cards_per_entry_ratio_seq->add(
312 young_cards_per_entry_ratio_defaults[index]);
313 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
314 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
315 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
316 _young_other_cost_per_region_ms_seq->add(
317 young_other_cost_per_region_ms_defaults[index]);
318 _non_young_other_cost_per_region_ms_seq->add(
319 non_young_other_cost_per_region_ms_defaults[index]);
321 // Below, we might need to calculate the pause time target based on
322 // the pause interval. When we do so we are going to give G1 maximum
323 // flexibility and allow it to do pauses when it needs to. So, we'll
324 // arrange that the pause interval to be pause time target + 1 to
325 // ensure that a) the pause time target is maximized with respect to
326 // the pause interval and b) we maintain the invariant that pause
327 // time target < pause interval. If the user does not want this
328 // maximum flexibility, they will have to set the pause interval
329 // explicitly.
331 // First make sure that, if either parameter is set, its value is
332 // reasonable.
333 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
334 if (MaxGCPauseMillis < 1) {
335 vm_exit_during_initialization("MaxGCPauseMillis should be "
336 "greater than 0");
337 }
338 }
339 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
340 if (GCPauseIntervalMillis < 1) {
341 vm_exit_during_initialization("GCPauseIntervalMillis should be "
342 "greater than 0");
343 }
344 }
346 // Then, if the pause time target parameter was not set, set it to
347 // the default value.
348 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
349 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
350 // The default pause time target in G1 is 200ms
351 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
352 } else {
353 // We do not allow the pause interval to be set without the
354 // pause time target
355 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
356 "without setting MaxGCPauseMillis");
357 }
358 }
360 // Then, if the interval parameter was not set, set it according to
361 // the pause time target (this will also deal with the case when the
362 // pause time target is the default value).
363 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
364 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
365 }
367 // Finally, make sure that the two parameters are consistent.
368 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
369 char buffer[256];
370 jio_snprintf(buffer, 256,
371 "MaxGCPauseMillis (%u) should be less than "
372 "GCPauseIntervalMillis (%u)",
373 MaxGCPauseMillis, GCPauseIntervalMillis);
374 vm_exit_during_initialization(buffer);
375 }
377 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
378 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
379 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
380 _sigma = (double) G1ConfidencePercent / 100.0;
382 // start conservatively (around 50ms is about right)
383 _concurrent_mark_remark_times_ms->add(0.05);
384 _concurrent_mark_cleanup_times_ms->add(0.20);
385 _tenuring_threshold = MaxTenuringThreshold;
386 // _max_survivor_regions will be calculated by
387 // update_young_list_target_length() during initialization.
388 _max_survivor_regions = 0;
390 assert(GCTimeRatio > 0,
391 "we should have set it to a default value set_g1_gc_flags() "
392 "if a user set it to 0");
393 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
395 uintx reserve_perc = G1ReservePercent;
396 // Put an artificial ceiling on this so that it's not set to a silly value.
397 if (reserve_perc > 50) {
398 reserve_perc = 50;
399 warning("G1ReservePercent is set to a value that is too large, "
400 "it's been updated to %u", reserve_perc);
401 }
402 _reserve_factor = (double) reserve_perc / 100.0;
403 // This will be set when the heap is expanded
404 // for the first time during initialization.
405 _reserve_regions = 0;
407 initialize_all();
408 _collectionSetChooser = new CollectionSetChooser();
409 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags
410 }
412 void G1CollectorPolicy::initialize_flags() {
413 set_min_alignment(HeapRegion::GrainBytes);
414 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
415 if (SurvivorRatio < 1) {
416 vm_exit_during_initialization("Invalid survivor ratio specified");
417 }
418 CollectorPolicy::initialize_flags();
419 }
421 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true) {
422 assert(G1DefaultMinNewGenPercent <= G1DefaultMaxNewGenPercent, "Min larger than max");
423 assert(G1DefaultMinNewGenPercent > 0 && G1DefaultMinNewGenPercent < 100, "Min out of bounds");
424 assert(G1DefaultMaxNewGenPercent > 0 && G1DefaultMaxNewGenPercent < 100, "Max out of bounds");
426 if (FLAG_IS_CMDLINE(NewRatio)) {
427 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
428 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
429 } else {
430 _sizer_kind = SizerNewRatio;
431 _adaptive_size = false;
432 return;
433 }
434 }
436 if (FLAG_IS_CMDLINE(NewSize)) {
437 _min_desired_young_length = MAX2((size_t) 1, NewSize / HeapRegion::GrainBytes);
438 if (FLAG_IS_CMDLINE(MaxNewSize)) {
439 _max_desired_young_length = MAX2((size_t) 1, MaxNewSize / HeapRegion::GrainBytes);
440 _sizer_kind = SizerMaxAndNewSize;
441 _adaptive_size = _min_desired_young_length == _max_desired_young_length;
442 } else {
443 _sizer_kind = SizerNewSizeOnly;
444 }
445 } else if (FLAG_IS_CMDLINE(MaxNewSize)) {
446 _max_desired_young_length = MAX2((size_t) 1, MaxNewSize / HeapRegion::GrainBytes);
447 _sizer_kind = SizerMaxNewSizeOnly;
448 }
449 }
451 size_t G1YoungGenSizer::calculate_default_min_length(size_t new_number_of_heap_regions) {
452 size_t default_value = (new_number_of_heap_regions * G1DefaultMinNewGenPercent) / 100;
453 return MAX2((size_t)1, default_value);
454 }
456 size_t G1YoungGenSizer::calculate_default_max_length(size_t new_number_of_heap_regions) {
457 size_t default_value = (new_number_of_heap_regions * G1DefaultMaxNewGenPercent) / 100;
458 return MAX2((size_t)1, default_value);
459 }
461 void G1YoungGenSizer::heap_size_changed(size_t new_number_of_heap_regions) {
462 assert(new_number_of_heap_regions > 0, "Heap must be initialized");
464 switch (_sizer_kind) {
465 case SizerDefaults:
466 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
467 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
468 break;
469 case SizerNewSizeOnly:
470 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
471 _max_desired_young_length = MAX2(_min_desired_young_length, _max_desired_young_length);
472 break;
473 case SizerMaxNewSizeOnly:
474 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
475 _min_desired_young_length = MIN2(_min_desired_young_length, _max_desired_young_length);
476 break;
477 case SizerMaxAndNewSize:
478 // Do nothing. Values set on the command line, don't update them at runtime.
479 break;
480 case SizerNewRatio:
481 _min_desired_young_length = new_number_of_heap_regions / (NewRatio + 1);
482 _max_desired_young_length = _min_desired_young_length;
483 break;
484 default:
485 ShouldNotReachHere();
486 }
488 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values");
489 }
491 void G1CollectorPolicy::init() {
492 // Set aside an initial future to_space.
493 _g1 = G1CollectedHeap::heap();
495 assert(Heap_lock->owned_by_self(), "Locking discipline.");
497 initialize_gc_policy_counters();
499 if (adaptive_young_list_length()) {
500 _young_list_fixed_length = 0;
501 } else {
502 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length();
503 }
504 _free_regions_at_end_of_collection = _g1->free_regions();
505 update_young_list_target_length();
506 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes;
508 // We may immediately start allocating regions and placing them on the
509 // collection set list. Initialize the per-collection set info
510 start_incremental_cset_building();
511 }
513 // Create the jstat counters for the policy.
514 void G1CollectorPolicy::initialize_gc_policy_counters() {
515 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
516 }
518 bool G1CollectorPolicy::predict_will_fit(size_t young_length,
519 double base_time_ms,
520 size_t base_free_regions,
521 double target_pause_time_ms) {
522 if (young_length >= base_free_regions) {
523 // end condition 1: not enough space for the young regions
524 return false;
525 }
527 double accum_surv_rate = accum_yg_surv_rate_pred((int)(young_length - 1));
528 size_t bytes_to_copy =
529 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
530 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
531 double young_other_time_ms = predict_young_other_time_ms(young_length);
532 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
533 if (pause_time_ms > target_pause_time_ms) {
534 // end condition 2: prediction is over the target pause time
535 return false;
536 }
538 size_t free_bytes =
539 (base_free_regions - young_length) * HeapRegion::GrainBytes;
540 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
541 // end condition 3: out-of-space (conservatively!)
542 return false;
543 }
545 // success!
546 return true;
547 }
549 void G1CollectorPolicy::record_new_heap_size(size_t new_number_of_regions) {
550 // re-calculate the necessary reserve
551 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
552 // We use ceiling so that if reserve_regions_d is > 0.0 (but
553 // smaller than 1.0) we'll get 1.
554 _reserve_regions = (size_t) ceil(reserve_regions_d);
556 _young_gen_sizer->heap_size_changed(new_number_of_regions);
557 }
559 size_t G1CollectorPolicy::calculate_young_list_desired_min_length(
560 size_t base_min_length) {
561 size_t desired_min_length = 0;
562 if (adaptive_young_list_length()) {
563 if (_alloc_rate_ms_seq->num() > 3) {
564 double now_sec = os::elapsedTime();
565 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
566 double alloc_rate_ms = predict_alloc_rate_ms();
567 desired_min_length = (size_t) ceil(alloc_rate_ms * when_ms);
568 } else {
569 // otherwise we don't have enough info to make the prediction
570 }
571 }
572 desired_min_length += base_min_length;
573 // make sure we don't go below any user-defined minimum bound
574 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length);
575 }
577 size_t G1CollectorPolicy::calculate_young_list_desired_max_length() {
578 // Here, we might want to also take into account any additional
579 // constraints (i.e., user-defined minimum bound). Currently, we
580 // effectively don't set this bound.
581 return _young_gen_sizer->max_desired_young_length();
582 }
584 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
585 if (rs_lengths == (size_t) -1) {
586 // if it's set to the default value (-1), we should predict it;
587 // otherwise, use the given value.
588 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
589 }
591 // Calculate the absolute and desired min bounds.
593 // This is how many young regions we already have (currently: the survivors).
594 size_t base_min_length = recorded_survivor_regions();
595 // This is the absolute minimum young length, which ensures that we
596 // can allocate one eden region in the worst-case.
597 size_t absolute_min_length = base_min_length + 1;
598 size_t desired_min_length =
599 calculate_young_list_desired_min_length(base_min_length);
600 if (desired_min_length < absolute_min_length) {
601 desired_min_length = absolute_min_length;
602 }
604 // Calculate the absolute and desired max bounds.
606 // We will try our best not to "eat" into the reserve.
607 size_t absolute_max_length = 0;
608 if (_free_regions_at_end_of_collection > _reserve_regions) {
609 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
610 }
611 size_t desired_max_length = calculate_young_list_desired_max_length();
612 if (desired_max_length > absolute_max_length) {
613 desired_max_length = absolute_max_length;
614 }
616 size_t young_list_target_length = 0;
617 if (adaptive_young_list_length()) {
618 if (gcs_are_young()) {
619 young_list_target_length =
620 calculate_young_list_target_length(rs_lengths,
621 base_min_length,
622 desired_min_length,
623 desired_max_length);
624 _rs_lengths_prediction = rs_lengths;
625 } else {
626 // Don't calculate anything and let the code below bound it to
627 // the desired_min_length, i.e., do the next GC as soon as
628 // possible to maximize how many old regions we can add to it.
629 }
630 } else {
631 if (gcs_are_young()) {
632 young_list_target_length = _young_list_fixed_length;
633 } else {
634 // A bit arbitrary: during mixed GCs we allocate half
635 // the young regions to try to add old regions to the CSet.
636 young_list_target_length = _young_list_fixed_length / 2;
637 // We choose to accept that we might go under the desired min
638 // length given that we intentionally ask for a smaller young gen.
639 desired_min_length = absolute_min_length;
640 }
641 }
643 // Make sure we don't go over the desired max length, nor under the
644 // desired min length. In case they clash, desired_min_length wins
645 // which is why that test is second.
646 if (young_list_target_length > desired_max_length) {
647 young_list_target_length = desired_max_length;
648 }
649 if (young_list_target_length < desired_min_length) {
650 young_list_target_length = desired_min_length;
651 }
653 assert(young_list_target_length > recorded_survivor_regions(),
654 "we should be able to allocate at least one eden region");
655 assert(young_list_target_length >= absolute_min_length, "post-condition");
656 _young_list_target_length = young_list_target_length;
658 update_max_gc_locker_expansion();
659 }
661 size_t
662 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
663 size_t base_min_length,
664 size_t desired_min_length,
665 size_t desired_max_length) {
666 assert(adaptive_young_list_length(), "pre-condition");
667 assert(gcs_are_young(), "only call this for young GCs");
669 // In case some edge-condition makes the desired max length too small...
670 if (desired_max_length <= desired_min_length) {
671 return desired_min_length;
672 }
674 // We'll adjust min_young_length and max_young_length not to include
675 // the already allocated young regions (i.e., so they reflect the
676 // min and max eden regions we'll allocate). The base_min_length
677 // will be reflected in the predictions by the
678 // survivor_regions_evac_time prediction.
679 assert(desired_min_length > base_min_length, "invariant");
680 size_t min_young_length = desired_min_length - base_min_length;
681 assert(desired_max_length > base_min_length, "invariant");
682 size_t max_young_length = desired_max_length - base_min_length;
684 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
685 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
686 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
687 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
688 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
689 double base_time_ms =
690 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
691 survivor_regions_evac_time;
692 size_t available_free_regions = _free_regions_at_end_of_collection;
693 size_t base_free_regions = 0;
694 if (available_free_regions > _reserve_regions) {
695 base_free_regions = available_free_regions - _reserve_regions;
696 }
698 // Here, we will make sure that the shortest young length that
699 // makes sense fits within the target pause time.
701 if (predict_will_fit(min_young_length, base_time_ms,
702 base_free_regions, target_pause_time_ms)) {
703 // The shortest young length will fit into the target pause time;
704 // we'll now check whether the absolute maximum number of young
705 // regions will fit in the target pause time. If not, we'll do
706 // a binary search between min_young_length and max_young_length.
707 if (predict_will_fit(max_young_length, base_time_ms,
708 base_free_regions, target_pause_time_ms)) {
709 // The maximum young length will fit into the target pause time.
710 // We are done so set min young length to the maximum length (as
711 // the result is assumed to be returned in min_young_length).
712 min_young_length = max_young_length;
713 } else {
714 // The maximum possible number of young regions will not fit within
715 // the target pause time so we'll search for the optimal
716 // length. The loop invariants are:
717 //
718 // min_young_length < max_young_length
719 // min_young_length is known to fit into the target pause time
720 // max_young_length is known not to fit into the target pause time
721 //
722 // Going into the loop we know the above hold as we've just
723 // checked them. Every time around the loop we check whether
724 // the middle value between min_young_length and
725 // max_young_length fits into the target pause time. If it
726 // does, it becomes the new min. If it doesn't, it becomes
727 // the new max. This way we maintain the loop invariants.
729 assert(min_young_length < max_young_length, "invariant");
730 size_t diff = (max_young_length - min_young_length) / 2;
731 while (diff > 0) {
732 size_t young_length = min_young_length + diff;
733 if (predict_will_fit(young_length, base_time_ms,
734 base_free_regions, target_pause_time_ms)) {
735 min_young_length = young_length;
736 } else {
737 max_young_length = young_length;
738 }
739 assert(min_young_length < max_young_length, "invariant");
740 diff = (max_young_length - min_young_length) / 2;
741 }
742 // The results is min_young_length which, according to the
743 // loop invariants, should fit within the target pause time.
745 // These are the post-conditions of the binary search above:
746 assert(min_young_length < max_young_length,
747 "otherwise we should have discovered that max_young_length "
748 "fits into the pause target and not done the binary search");
749 assert(predict_will_fit(min_young_length, base_time_ms,
750 base_free_regions, target_pause_time_ms),
751 "min_young_length, the result of the binary search, should "
752 "fit into the pause target");
753 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
754 base_free_regions, target_pause_time_ms),
755 "min_young_length, the result of the binary search, should be "
756 "optimal, so no larger length should fit into the pause target");
757 }
758 } else {
759 // Even the minimum length doesn't fit into the pause time
760 // target, return it as the result nevertheless.
761 }
762 return base_min_length + min_young_length;
763 }
765 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
766 double survivor_regions_evac_time = 0.0;
767 for (HeapRegion * r = _recorded_survivor_head;
768 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
769 r = r->get_next_young_region()) {
770 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
771 }
772 return survivor_regions_evac_time;
773 }
775 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
776 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
778 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
779 if (rs_lengths > _rs_lengths_prediction) {
780 // add 10% to avoid having to recalculate often
781 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
782 update_young_list_target_length(rs_lengths_prediction);
783 }
784 }
788 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
789 bool is_tlab,
790 bool* gc_overhead_limit_was_exceeded) {
791 guarantee(false, "Not using this policy feature yet.");
792 return NULL;
793 }
795 // This method controls how a collector handles one or more
796 // of its generations being fully allocated.
797 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
798 bool is_tlab) {
799 guarantee(false, "Not using this policy feature yet.");
800 return NULL;
801 }
804 #ifndef PRODUCT
805 bool G1CollectorPolicy::verify_young_ages() {
806 HeapRegion* head = _g1->young_list()->first_region();
807 return
808 verify_young_ages(head, _short_lived_surv_rate_group);
809 // also call verify_young_ages on any additional surv rate groups
810 }
812 bool
813 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
814 SurvRateGroup *surv_rate_group) {
815 guarantee( surv_rate_group != NULL, "pre-condition" );
817 const char* name = surv_rate_group->name();
818 bool ret = true;
819 int prev_age = -1;
821 for (HeapRegion* curr = head;
822 curr != NULL;
823 curr = curr->get_next_young_region()) {
824 SurvRateGroup* group = curr->surv_rate_group();
825 if (group == NULL && !curr->is_survivor()) {
826 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
827 ret = false;
828 }
830 if (surv_rate_group == group) {
831 int age = curr->age_in_surv_rate_group();
833 if (age < 0) {
834 gclog_or_tty->print_cr("## %s: encountered negative age", name);
835 ret = false;
836 }
838 if (age <= prev_age) {
839 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
840 "(%d, %d)", name, age, prev_age);
841 ret = false;
842 }
843 prev_age = age;
844 }
845 }
847 return ret;
848 }
849 #endif // PRODUCT
851 void G1CollectorPolicy::record_full_collection_start() {
852 _cur_collection_start_sec = os::elapsedTime();
853 // Release the future to-space so that it is available for compaction into.
854 _g1->set_full_collection();
855 }
857 void G1CollectorPolicy::record_full_collection_end() {
858 // Consider this like a collection pause for the purposes of allocation
859 // since last pause.
860 double end_sec = os::elapsedTime();
861 double full_gc_time_sec = end_sec - _cur_collection_start_sec;
862 double full_gc_time_ms = full_gc_time_sec * 1000.0;
864 _all_full_gc_times_ms->add(full_gc_time_ms);
866 update_recent_gc_times(end_sec, full_gc_time_ms);
868 _g1->clear_full_collection();
870 // "Nuke" the heuristics that control the young/mixed GC
871 // transitions and make sure we start with young GCs after the Full GC.
872 set_gcs_are_young(true);
873 _last_young_gc = false;
874 _should_revert_to_young_gcs = false;
875 clear_initiate_conc_mark_if_possible();
876 clear_during_initial_mark_pause();
877 _known_garbage_bytes = 0;
878 _known_garbage_ratio = 0.0;
879 _in_marking_window = false;
880 _in_marking_window_im = false;
882 _short_lived_surv_rate_group->start_adding_regions();
883 // also call this on any additional surv rate groups
885 record_survivor_regions(0, NULL, NULL);
887 _free_regions_at_end_of_collection = _g1->free_regions();
888 // Reset survivors SurvRateGroup.
889 _survivor_surv_rate_group->reset();
890 update_young_list_target_length();
891 _collectionSetChooser->updateAfterFullCollection();
892 }
894 void G1CollectorPolicy::record_stop_world_start() {
895 _stop_world_start = os::elapsedTime();
896 }
898 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
899 size_t start_used) {
900 if (PrintGCDetails) {
901 gclog_or_tty->stamp(PrintGCTimeStamps);
902 gclog_or_tty->print("[GC pause");
903 gclog_or_tty->print(" (%s)", gcs_are_young() ? "young" : "mixed");
904 }
906 if (!during_initial_mark_pause()) {
907 // We only need to do this here as the policy will only be applied
908 // to the GC we're about to start. so, no point is calculating this
909 // every time we calculate / recalculate the target young length.
910 update_survivors_policy();
911 } else {
912 // The marking phase has a "we only copy implicitly live
913 // objects during marking" invariant. The easiest way to ensure it
914 // holds is not to allocate any survivor regions and tenure all
915 // objects. In the future we might change this and handle survivor
916 // regions specially during marking.
917 tenure_all_objects();
918 }
920 assert(_g1->used() == _g1->recalculate_used(),
921 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
922 _g1->used(), _g1->recalculate_used()));
924 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
925 _all_stop_world_times_ms->add(s_w_t_ms);
926 _stop_world_start = 0.0;
928 _cur_collection_start_sec = start_time_sec;
929 _cur_collection_pause_used_at_start_bytes = start_used;
930 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
931 _pending_cards = _g1->pending_card_num();
932 _max_pending_cards = _g1->max_pending_card_num();
934 _bytes_in_collection_set_before_gc = 0;
935 _bytes_copied_during_gc = 0;
937 YoungList* young_list = _g1->young_list();
938 _eden_bytes_before_gc = young_list->eden_used_bytes();
939 _survivor_bytes_before_gc = young_list->survivor_used_bytes();
940 _capacity_before_gc = _g1->capacity();
942 #ifdef DEBUG
943 // initialise these to something well known so that we can spot
944 // if they are not set properly
946 for (int i = 0; i < _parallel_gc_threads; ++i) {
947 _par_last_gc_worker_start_times_ms[i] = -1234.0;
948 _par_last_ext_root_scan_times_ms[i] = -1234.0;
949 _par_last_satb_filtering_times_ms[i] = -1234.0;
950 _par_last_update_rs_times_ms[i] = -1234.0;
951 _par_last_update_rs_processed_buffers[i] = -1234.0;
952 _par_last_scan_rs_times_ms[i] = -1234.0;
953 _par_last_obj_copy_times_ms[i] = -1234.0;
954 _par_last_termination_times_ms[i] = -1234.0;
955 _par_last_termination_attempts[i] = -1234.0;
956 _par_last_gc_worker_end_times_ms[i] = -1234.0;
957 _par_last_gc_worker_times_ms[i] = -1234.0;
958 _par_last_gc_worker_other_times_ms[i] = -1234.0;
959 }
960 #endif
962 for (int i = 0; i < _aux_num; ++i) {
963 _cur_aux_times_ms[i] = 0.0;
964 _cur_aux_times_set[i] = false;
965 }
967 // This is initialized to zero here and is set during
968 // the evacuation pause if marking is in progress.
969 _cur_satb_drain_time_ms = 0.0;
971 _last_gc_was_young = false;
973 // do that for any other surv rate groups
974 _short_lived_surv_rate_group->stop_adding_regions();
975 _survivors_age_table.clear();
977 assert( verify_young_ages(), "region age verification" );
978 }
980 void G1CollectorPolicy::record_concurrent_mark_init_end(double
981 mark_init_elapsed_time_ms) {
982 _during_marking = true;
983 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
984 clear_during_initial_mark_pause();
985 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
986 }
988 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
989 _mark_remark_start_sec = os::elapsedTime();
990 _during_marking = false;
991 }
993 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
994 double end_time_sec = os::elapsedTime();
995 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
996 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
997 _cur_mark_stop_world_time_ms += elapsed_time_ms;
998 _prev_collection_pause_end_ms += elapsed_time_ms;
1000 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
1001 }
1003 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
1004 _mark_cleanup_start_sec = os::elapsedTime();
1005 }
1007 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
1008 _should_revert_to_young_gcs = false;
1009 _last_young_gc = true;
1010 _in_marking_window = false;
1011 }
1013 void G1CollectorPolicy::record_concurrent_pause() {
1014 if (_stop_world_start > 0.0) {
1015 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
1016 _all_yield_times_ms->add(yield_ms);
1017 }
1018 }
1020 void G1CollectorPolicy::record_concurrent_pause_end() {
1021 }
1023 template<class T>
1024 T sum_of(T* sum_arr, int start, int n, int N) {
1025 T sum = (T)0;
1026 for (int i = 0; i < n; i++) {
1027 int j = (start + i) % N;
1028 sum += sum_arr[j];
1029 }
1030 return sum;
1031 }
1033 void G1CollectorPolicy::print_par_stats(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 (ms):", 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(" %3.1lf", val);
1048 }
1049 buf.append_and_print_cr("");
1050 double avg = total / (double) no_of_gc_threads();
1051 buf.append_and_print_cr(" Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf, Diff: %5.1lf]",
1052 avg, min, max, max - min);
1053 }
1055 void G1CollectorPolicy::print_par_sizes(int level,
1056 const char* str,
1057 double* data) {
1058 double min = data[0], max = data[0];
1059 double total = 0.0;
1060 LineBuffer buf(level);
1061 buf.append("[%s :", str);
1062 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1063 double val = data[i];
1064 if (val < min)
1065 min = val;
1066 if (val > max)
1067 max = val;
1068 total += val;
1069 buf.append(" %d", (int) val);
1070 }
1071 buf.append_and_print_cr("");
1072 double avg = total / (double) no_of_gc_threads();
1073 buf.append_and_print_cr(" Sum: %d, Avg: %d, Min: %d, Max: %d, Diff: %d]",
1074 (int)total, (int)avg, (int)min, (int)max, (int)max - (int)min);
1075 }
1077 void G1CollectorPolicy::print_stats(int level,
1078 const char* str,
1079 double value) {
1080 LineBuffer(level).append_and_print_cr("[%s: %5.1lf ms]", str, value);
1081 }
1083 void G1CollectorPolicy::print_stats(int level,
1084 const char* str,
1085 int value) {
1086 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value);
1087 }
1089 double G1CollectorPolicy::avg_value(double* data) {
1090 if (G1CollectedHeap::use_parallel_gc_threads()) {
1091 double ret = 0.0;
1092 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1093 ret += data[i];
1094 }
1095 return ret / (double) no_of_gc_threads();
1096 } else {
1097 return data[0];
1098 }
1099 }
1101 double G1CollectorPolicy::max_value(double* data) {
1102 if (G1CollectedHeap::use_parallel_gc_threads()) {
1103 double ret = data[0];
1104 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1105 if (data[i] > ret) {
1106 ret = data[i];
1107 }
1108 }
1109 return ret;
1110 } else {
1111 return data[0];
1112 }
1113 }
1115 double G1CollectorPolicy::sum_of_values(double* data) {
1116 if (G1CollectedHeap::use_parallel_gc_threads()) {
1117 double sum = 0.0;
1118 for (uint i = 0; i < no_of_gc_threads(); i++) {
1119 sum += data[i];
1120 }
1121 return sum;
1122 } else {
1123 return data[0];
1124 }
1125 }
1127 double G1CollectorPolicy::max_sum(double* data1, double* data2) {
1128 double ret = data1[0] + data2[0];
1130 if (G1CollectedHeap::use_parallel_gc_threads()) {
1131 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1132 double data = data1[i] + data2[i];
1133 if (data > ret) {
1134 ret = data;
1135 }
1136 }
1137 }
1138 return ret;
1139 }
1141 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source) {
1142 if (_g1->mark_in_progress()) {
1143 return false;
1144 }
1146 size_t marking_initiating_used_threshold =
1147 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
1148 size_t cur_used_bytes = _g1->non_young_capacity_bytes();
1150 if (cur_used_bytes > marking_initiating_used_threshold) {
1151 if (gcs_are_young()) {
1152 ergo_verbose4(ErgoConcCycles,
1153 "request concurrent cycle initiation",
1154 ergo_format_reason("occupancy higher than threshold")
1155 ergo_format_byte("occupancy")
1156 ergo_format_byte_perc("threshold")
1157 ergo_format_str("source"),
1158 cur_used_bytes,
1159 marking_initiating_used_threshold,
1160 (double) InitiatingHeapOccupancyPercent,
1161 source);
1162 return true;
1163 } else {
1164 ergo_verbose4(ErgoConcCycles,
1165 "do not request concurrent cycle initiation",
1166 ergo_format_reason("still doing mixed collections")
1167 ergo_format_byte("occupancy")
1168 ergo_format_byte_perc("threshold")
1169 ergo_format_str("source"),
1170 cur_used_bytes,
1171 marking_initiating_used_threshold,
1172 (double) InitiatingHeapOccupancyPercent,
1173 source);
1174 }
1175 }
1177 return false;
1178 }
1180 // Anything below that is considered to be zero
1181 #define MIN_TIMER_GRANULARITY 0.0000001
1183 void G1CollectorPolicy::record_collection_pause_end(int no_of_gc_threads) {
1184 double end_time_sec = os::elapsedTime();
1185 double elapsed_ms = _last_pause_time_ms;
1186 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1187 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(),
1188 "otherwise, the subtraction below does not make sense");
1189 size_t rs_size =
1190 _cur_collection_pause_used_regions_at_start - cset_region_length();
1191 size_t cur_used_bytes = _g1->used();
1192 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1193 bool last_pause_included_initial_mark = false;
1194 bool update_stats = !_g1->evacuation_failed();
1195 set_no_of_gc_threads(no_of_gc_threads);
1197 #ifndef PRODUCT
1198 if (G1YoungSurvRateVerbose) {
1199 gclog_or_tty->print_cr("");
1200 _short_lived_surv_rate_group->print();
1201 // do that for any other surv rate groups too
1202 }
1203 #endif // PRODUCT
1205 last_pause_included_initial_mark = during_initial_mark_pause();
1206 if (last_pause_included_initial_mark) {
1207 record_concurrent_mark_init_end(0.0);
1208 }
1210 if (!_last_young_gc && need_to_start_conc_mark("end of GC")) {
1211 // Note: this might have already been set, if during the last
1212 // pause we decided to start a cycle but at the beginning of
1213 // this pause we decided to postpone it. That's OK.
1214 set_initiate_conc_mark_if_possible();
1215 }
1217 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1218 end_time_sec, false);
1220 // This assert is exempted when we're doing parallel collection pauses,
1221 // because the fragmentation caused by the parallel GC allocation buffers
1222 // can lead to more memory being used during collection than was used
1223 // before. Best leave this out until the fragmentation problem is fixed.
1224 // Pauses in which evacuation failed can also lead to negative
1225 // collections, since no space is reclaimed from a region containing an
1226 // object whose evacuation failed.
1227 // Further, we're now always doing parallel collection. But I'm still
1228 // leaving this here as a placeholder for a more precise assertion later.
1229 // (DLD, 10/05.)
1230 assert((true || parallel) // Always using GC LABs now.
1231 || _g1->evacuation_failed()
1232 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1233 "Negative collection");
1235 size_t freed_bytes =
1236 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1237 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1239 double survival_fraction =
1240 (double)surviving_bytes/
1241 (double)_collection_set_bytes_used_before;
1243 // These values are used to update the summary information that is
1244 // displayed when TraceGen0Time is enabled, and are output as part
1245 // of the PrintGCDetails output, in the non-parallel case.
1247 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1248 double satb_filtering_time = avg_value(_par_last_satb_filtering_times_ms);
1249 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1250 double update_rs_processed_buffers =
1251 sum_of_values(_par_last_update_rs_processed_buffers);
1252 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1253 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1254 double termination_time = avg_value(_par_last_termination_times_ms);
1256 double known_time = ext_root_scan_time +
1257 satb_filtering_time +
1258 update_rs_time +
1259 scan_rs_time +
1260 obj_copy_time;
1262 double other_time_ms = elapsed_ms;
1264 // Subtract the SATB drain time. It's initialized to zero at the
1265 // start of the pause and is updated during the pause if marking
1266 // is in progress.
1267 other_time_ms -= _cur_satb_drain_time_ms;
1269 if (parallel) {
1270 other_time_ms -= _cur_collection_par_time_ms;
1271 } else {
1272 other_time_ms -= known_time;
1273 }
1275 // Subtract the time taken to clean the card table from the
1276 // current value of "other time"
1277 other_time_ms -= _cur_clear_ct_time_ms;
1279 // Subtract the time spent completing marking in the collection
1280 // set. Note if marking is not in progress during the pause
1281 // the value of _mark_closure_time_ms will be zero.
1282 other_time_ms -= _mark_closure_time_ms;
1284 // TraceGen0Time and TraceGen1Time summary info updating.
1285 _all_pause_times_ms->add(elapsed_ms);
1287 if (update_stats) {
1288 _summary->record_total_time_ms(elapsed_ms);
1289 _summary->record_other_time_ms(other_time_ms);
1291 MainBodySummary* body_summary = _summary->main_body_summary();
1292 assert(body_summary != NULL, "should not be null!");
1294 // This will be non-zero iff marking is currently in progress (i.e.
1295 // _g1->mark_in_progress() == true) and the currrent pause was not
1296 // an initial mark pause. Since the body_summary items are NumberSeqs,
1297 // however, they have to be consistent and updated in lock-step with
1298 // each other. Therefore we unconditionally record the SATB drain
1299 // time - even if it's zero.
1300 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms);
1302 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1303 body_summary->record_satb_filtering_time_ms(satb_filtering_time);
1304 body_summary->record_update_rs_time_ms(update_rs_time);
1305 body_summary->record_scan_rs_time_ms(scan_rs_time);
1306 body_summary->record_obj_copy_time_ms(obj_copy_time);
1308 if (parallel) {
1309 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1310 body_summary->record_termination_time_ms(termination_time);
1312 double parallel_known_time = known_time + termination_time;
1313 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time;
1314 body_summary->record_parallel_other_time_ms(parallel_other_time);
1315 }
1317 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms);
1318 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1320 // We exempt parallel collection from this check because Alloc Buffer
1321 // fragmentation can produce negative collections. Same with evac
1322 // failure.
1323 // Further, we're now always doing parallel collection. But I'm still
1324 // leaving this here as a placeholder for a more precise assertion later.
1325 // (DLD, 10/05.
1326 assert((true || parallel)
1327 || _g1->evacuation_failed()
1328 || surviving_bytes <= _collection_set_bytes_used_before,
1329 "Or else negative collection!");
1331 // this is where we update the allocation rate of the application
1332 double app_time_ms =
1333 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1334 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1335 // This usually happens due to the timer not having the required
1336 // granularity. Some Linuxes are the usual culprits.
1337 // We'll just set it to something (arbitrarily) small.
1338 app_time_ms = 1.0;
1339 }
1340 // We maintain the invariant that all objects allocated by mutator
1341 // threads will be allocated out of eden regions. So, we can use
1342 // the eden region number allocated since the previous GC to
1343 // calculate the application's allocate rate. The only exception
1344 // to that is humongous objects that are allocated separately. But
1345 // given that humongous object allocations do not really affect
1346 // either the pause's duration nor when the next pause will take
1347 // place we can safely ignore them here.
1348 size_t regions_allocated = eden_cset_region_length();
1349 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1350 _alloc_rate_ms_seq->add(alloc_rate_ms);
1352 double interval_ms =
1353 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1354 update_recent_gc_times(end_time_sec, elapsed_ms);
1355 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1356 if (recent_avg_pause_time_ratio() < 0.0 ||
1357 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1358 #ifndef PRODUCT
1359 // Dump info to allow post-facto debugging
1360 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1361 gclog_or_tty->print_cr("-------------------------------------------");
1362 gclog_or_tty->print_cr("Recent GC Times (ms):");
1363 _recent_gc_times_ms->dump();
1364 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1365 _recent_prev_end_times_for_all_gcs_sec->dump();
1366 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1367 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1368 // In debug mode, terminate the JVM if the user wants to debug at this point.
1369 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1370 #endif // !PRODUCT
1371 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1372 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1373 if (_recent_avg_pause_time_ratio < 0.0) {
1374 _recent_avg_pause_time_ratio = 0.0;
1375 } else {
1376 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1377 _recent_avg_pause_time_ratio = 1.0;
1378 }
1379 }
1380 }
1382 for (int i = 0; i < _aux_num; ++i) {
1383 if (_cur_aux_times_set[i]) {
1384 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1385 }
1386 }
1388 // PrintGCDetails output
1389 if (PrintGCDetails) {
1390 bool print_marking_info =
1391 _g1->mark_in_progress() && !last_pause_included_initial_mark;
1393 gclog_or_tty->print_cr("%s, %1.8lf secs]",
1394 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1395 elapsed_ms / 1000.0);
1397 if (parallel) {
1398 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1399 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms);
1400 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1401 if (print_marking_info) {
1402 print_par_stats(2, "SATB Filtering", _par_last_satb_filtering_times_ms);
1403 }
1404 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1405 print_par_sizes(3, "Processed Buffers", _par_last_update_rs_processed_buffers);
1406 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1407 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1408 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1409 print_par_sizes(3, "Termination Attempts", _par_last_termination_attempts);
1410 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms);
1412 for (int i = 0; i < _parallel_gc_threads; i++) {
1413 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] - _par_last_gc_worker_start_times_ms[i];
1415 double worker_known_time = _par_last_ext_root_scan_times_ms[i] +
1416 _par_last_satb_filtering_times_ms[i] +
1417 _par_last_update_rs_times_ms[i] +
1418 _par_last_scan_rs_times_ms[i] +
1419 _par_last_obj_copy_times_ms[i] +
1420 _par_last_termination_times_ms[i];
1422 _par_last_gc_worker_other_times_ms[i] = _cur_collection_par_time_ms - worker_known_time;
1423 }
1424 print_par_stats(2, "GC Worker", _par_last_gc_worker_times_ms);
1425 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms);
1426 } else {
1427 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1428 if (print_marking_info) {
1429 print_stats(1, "SATB Filtering", satb_filtering_time);
1430 }
1431 print_stats(1, "Update RS", update_rs_time);
1432 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers);
1433 print_stats(1, "Scan RS", scan_rs_time);
1434 print_stats(1, "Object Copying", obj_copy_time);
1435 }
1436 if (print_marking_info) {
1437 print_stats(1, "Complete CSet Marking", _mark_closure_time_ms);
1438 }
1439 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1440 #ifndef PRODUCT
1441 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1442 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1443 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1444 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1445 if (_num_cc_clears > 0) {
1446 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1447 }
1448 #endif
1449 print_stats(1, "Other", other_time_ms);
1450 print_stats(2, "Choose CSet",
1451 (_recorded_young_cset_choice_time_ms +
1452 _recorded_non_young_cset_choice_time_ms));
1453 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms);
1454 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms);
1455 print_stats(2, "Free CSet",
1456 (_recorded_young_free_cset_time_ms +
1457 _recorded_non_young_free_cset_time_ms));
1459 for (int i = 0; i < _aux_num; ++i) {
1460 if (_cur_aux_times_set[i]) {
1461 char buffer[96];
1462 sprintf(buffer, "Aux%d", i);
1463 print_stats(1, buffer, _cur_aux_times_ms[i]);
1464 }
1465 }
1466 }
1468 // Update the efficiency-since-mark vars.
1469 double proc_ms = elapsed_ms * (double) _parallel_gc_threads;
1470 if (elapsed_ms < MIN_TIMER_GRANULARITY) {
1471 // This usually happens due to the timer not having the required
1472 // granularity. Some Linuxes are the usual culprits.
1473 // We'll just set it to something (arbitrarily) small.
1474 proc_ms = 1.0;
1475 }
1476 double cur_efficiency = (double) freed_bytes / proc_ms;
1478 bool new_in_marking_window = _in_marking_window;
1479 bool new_in_marking_window_im = false;
1480 if (during_initial_mark_pause()) {
1481 new_in_marking_window = true;
1482 new_in_marking_window_im = true;
1483 }
1485 if (_last_young_gc) {
1486 if (!last_pause_included_initial_mark) {
1487 ergo_verbose2(ErgoMixedGCs,
1488 "start mixed GCs",
1489 ergo_format_byte_perc("known garbage"),
1490 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1491 set_gcs_are_young(false);
1492 } else {
1493 ergo_verbose0(ErgoMixedGCs,
1494 "do not start mixed GCs",
1495 ergo_format_reason("concurrent cycle is about to start"));
1496 }
1497 _last_young_gc = false;
1498 }
1500 if (!_last_gc_was_young) {
1501 if (_should_revert_to_young_gcs) {
1502 ergo_verbose2(ErgoMixedGCs,
1503 "end mixed GCs",
1504 ergo_format_reason("mixed GCs end requested")
1505 ergo_format_byte_perc("known garbage"),
1506 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1507 set_gcs_are_young(true);
1508 } else if (_known_garbage_ratio < 0.05) {
1509 ergo_verbose3(ErgoMixedGCs,
1510 "end mixed GCs",
1511 ergo_format_reason("known garbage percent lower than threshold")
1512 ergo_format_byte_perc("known garbage")
1513 ergo_format_perc("threshold"),
1514 _known_garbage_bytes, _known_garbage_ratio * 100.0,
1515 0.05 * 100.0);
1516 set_gcs_are_young(true);
1517 } else if (adaptive_young_list_length() &&
1518 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) {
1519 ergo_verbose5(ErgoMixedGCs,
1520 "end mixed GCs",
1521 ergo_format_reason("current GC efficiency lower than "
1522 "predicted young GC efficiency")
1523 ergo_format_double("GC efficiency factor")
1524 ergo_format_double("current GC efficiency")
1525 ergo_format_double("predicted young GC efficiency")
1526 ergo_format_byte_perc("known garbage"),
1527 get_gc_eff_factor(), cur_efficiency,
1528 predict_young_gc_eff(),
1529 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1530 set_gcs_are_young(true);
1531 }
1532 }
1533 _should_revert_to_young_gcs = false;
1535 if (_last_gc_was_young && !_during_marking) {
1536 _young_gc_eff_seq->add(cur_efficiency);
1537 }
1539 _short_lived_surv_rate_group->start_adding_regions();
1540 // do that for any other surv rate groupsx
1542 if (update_stats) {
1543 double pause_time_ms = elapsed_ms;
1545 size_t diff = 0;
1546 if (_max_pending_cards >= _pending_cards)
1547 diff = _max_pending_cards - _pending_cards;
1548 _pending_card_diff_seq->add((double) diff);
1550 double cost_per_card_ms = 0.0;
1551 if (_pending_cards > 0) {
1552 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1553 _cost_per_card_ms_seq->add(cost_per_card_ms);
1554 }
1556 size_t cards_scanned = _g1->cards_scanned();
1558 double cost_per_entry_ms = 0.0;
1559 if (cards_scanned > 10) {
1560 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1561 if (_last_gc_was_young) {
1562 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1563 } else {
1564 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1565 }
1566 }
1568 if (_max_rs_lengths > 0) {
1569 double cards_per_entry_ratio =
1570 (double) cards_scanned / (double) _max_rs_lengths;
1571 if (_last_gc_was_young) {
1572 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1573 } else {
1574 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1575 }
1576 }
1578 // This is defensive. For a while _max_rs_lengths could get
1579 // smaller than _recorded_rs_lengths which was causing
1580 // rs_length_diff to get very large and mess up the RSet length
1581 // predictions. The reason was unsafe concurrent updates to the
1582 // _inc_cset_recorded_rs_lengths field which the code below guards
1583 // against (see CR 7118202). This bug has now been fixed (see CR
1584 // 7119027). However, I'm still worried that
1585 // _inc_cset_recorded_rs_lengths might still end up somewhat
1586 // inaccurate. The concurrent refinement thread calculates an
1587 // RSet's length concurrently with other CR threads updating it
1588 // which might cause it to calculate the length incorrectly (if,
1589 // say, it's in mid-coarsening). So I'll leave in the defensive
1590 // conditional below just in case.
1591 size_t rs_length_diff = 0;
1592 if (_max_rs_lengths > _recorded_rs_lengths) {
1593 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1594 }
1595 _rs_length_diff_seq->add((double) rs_length_diff);
1597 size_t copied_bytes = surviving_bytes;
1598 double cost_per_byte_ms = 0.0;
1599 if (copied_bytes > 0) {
1600 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1601 if (_in_marking_window) {
1602 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1603 } else {
1604 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1605 }
1606 }
1608 double all_other_time_ms = pause_time_ms -
1609 (update_rs_time + scan_rs_time + obj_copy_time +
1610 _mark_closure_time_ms + termination_time);
1612 double young_other_time_ms = 0.0;
1613 if (young_cset_region_length() > 0) {
1614 young_other_time_ms =
1615 _recorded_young_cset_choice_time_ms +
1616 _recorded_young_free_cset_time_ms;
1617 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1618 (double) young_cset_region_length());
1619 }
1620 double non_young_other_time_ms = 0.0;
1621 if (old_cset_region_length() > 0) {
1622 non_young_other_time_ms =
1623 _recorded_non_young_cset_choice_time_ms +
1624 _recorded_non_young_free_cset_time_ms;
1626 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1627 (double) old_cset_region_length());
1628 }
1630 double constant_other_time_ms = all_other_time_ms -
1631 (young_other_time_ms + non_young_other_time_ms);
1632 _constant_other_time_ms_seq->add(constant_other_time_ms);
1634 double survival_ratio = 0.0;
1635 if (_bytes_in_collection_set_before_gc > 0) {
1636 survival_ratio = (double) _bytes_copied_during_gc /
1637 (double) _bytes_in_collection_set_before_gc;
1638 }
1640 _pending_cards_seq->add((double) _pending_cards);
1641 _rs_lengths_seq->add((double) _max_rs_lengths);
1643 double expensive_region_limit_ms =
1644 (double) MaxGCPauseMillis - predict_constant_other_time_ms();
1645 if (expensive_region_limit_ms < 0.0) {
1646 // this means that the other time was predicted to be longer than
1647 // than the max pause time
1648 expensive_region_limit_ms = (double) MaxGCPauseMillis;
1649 }
1650 _expensive_region_limit_ms = expensive_region_limit_ms;
1651 }
1653 _in_marking_window = new_in_marking_window;
1654 _in_marking_window_im = new_in_marking_window_im;
1655 _free_regions_at_end_of_collection = _g1->free_regions();
1656 update_young_list_target_length();
1658 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1659 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1660 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1662 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end.");
1663 }
1665 #define EXT_SIZE_FORMAT "%d%s"
1666 #define EXT_SIZE_PARAMS(bytes) \
1667 byte_size_in_proper_unit((bytes)), \
1668 proper_unit_for_byte_size((bytes))
1670 void G1CollectorPolicy::print_heap_transition() {
1671 if (PrintGCDetails) {
1672 YoungList* young_list = _g1->young_list();
1673 size_t eden_bytes = young_list->eden_used_bytes();
1674 size_t survivor_bytes = young_list->survivor_used_bytes();
1675 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes;
1676 size_t used = _g1->used();
1677 size_t capacity = _g1->capacity();
1678 size_t eden_capacity =
1679 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes;
1681 gclog_or_tty->print_cr(
1682 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1683 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1684 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1685 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1686 EXT_SIZE_PARAMS(_eden_bytes_before_gc),
1687 EXT_SIZE_PARAMS(_prev_eden_capacity),
1688 EXT_SIZE_PARAMS(eden_bytes),
1689 EXT_SIZE_PARAMS(eden_capacity),
1690 EXT_SIZE_PARAMS(_survivor_bytes_before_gc),
1691 EXT_SIZE_PARAMS(survivor_bytes),
1692 EXT_SIZE_PARAMS(used_before_gc),
1693 EXT_SIZE_PARAMS(_capacity_before_gc),
1694 EXT_SIZE_PARAMS(used),
1695 EXT_SIZE_PARAMS(capacity));
1697 _prev_eden_capacity = eden_capacity;
1698 } else if (PrintGC) {
1699 _g1->print_size_transition(gclog_or_tty,
1700 _cur_collection_pause_used_at_start_bytes,
1701 _g1->used(), _g1->capacity());
1702 }
1703 }
1705 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1706 double update_rs_processed_buffers,
1707 double goal_ms) {
1708 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1709 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1711 if (G1UseAdaptiveConcRefinement) {
1712 const int k_gy = 3, k_gr = 6;
1713 const double inc_k = 1.1, dec_k = 0.9;
1715 int g = cg1r->green_zone();
1716 if (update_rs_time > goal_ms) {
1717 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1718 } else {
1719 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1720 g = (int)MAX2(g * inc_k, g + 1.0);
1721 }
1722 }
1723 // Change the refinement threads params
1724 cg1r->set_green_zone(g);
1725 cg1r->set_yellow_zone(g * k_gy);
1726 cg1r->set_red_zone(g * k_gr);
1727 cg1r->reinitialize_threads();
1729 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1730 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1731 cg1r->yellow_zone());
1732 // Change the barrier params
1733 dcqs.set_process_completed_threshold(processing_threshold);
1734 dcqs.set_max_completed_queue(cg1r->red_zone());
1735 }
1737 int curr_queue_size = dcqs.completed_buffers_num();
1738 if (curr_queue_size >= cg1r->yellow_zone()) {
1739 dcqs.set_completed_queue_padding(curr_queue_size);
1740 } else {
1741 dcqs.set_completed_queue_padding(0);
1742 }
1743 dcqs.notify_if_necessary();
1744 }
1746 double
1747 G1CollectorPolicy::
1748 predict_young_collection_elapsed_time_ms(size_t adjustment) {
1749 guarantee( adjustment == 0 || adjustment == 1, "invariant" );
1751 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1752 size_t young_num = g1h->young_list()->length();
1753 if (young_num == 0)
1754 return 0.0;
1756 young_num += adjustment;
1757 size_t pending_cards = predict_pending_cards();
1758 size_t rs_lengths = g1h->young_list()->sampled_rs_lengths() +
1759 predict_rs_length_diff();
1760 size_t card_num;
1761 if (gcs_are_young()) {
1762 card_num = predict_young_card_num(rs_lengths);
1763 } else {
1764 card_num = predict_non_young_card_num(rs_lengths);
1765 }
1766 size_t young_byte_size = young_num * HeapRegion::GrainBytes;
1767 double accum_yg_surv_rate =
1768 _short_lived_surv_rate_group->accum_surv_rate(adjustment);
1770 size_t bytes_to_copy =
1771 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes);
1773 return
1774 predict_rs_update_time_ms(pending_cards) +
1775 predict_rs_scan_time_ms(card_num) +
1776 predict_object_copy_time_ms(bytes_to_copy) +
1777 predict_young_other_time_ms(young_num) +
1778 predict_constant_other_time_ms();
1779 }
1781 double
1782 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1783 size_t rs_length = predict_rs_length_diff();
1784 size_t card_num;
1785 if (gcs_are_young()) {
1786 card_num = predict_young_card_num(rs_length);
1787 } else {
1788 card_num = predict_non_young_card_num(rs_length);
1789 }
1790 return predict_base_elapsed_time_ms(pending_cards, card_num);
1791 }
1793 double
1794 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1795 size_t scanned_cards) {
1796 return
1797 predict_rs_update_time_ms(pending_cards) +
1798 predict_rs_scan_time_ms(scanned_cards) +
1799 predict_constant_other_time_ms();
1800 }
1802 double
1803 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1804 bool young) {
1805 size_t rs_length = hr->rem_set()->occupied();
1806 size_t card_num;
1807 if (gcs_are_young()) {
1808 card_num = predict_young_card_num(rs_length);
1809 } else {
1810 card_num = predict_non_young_card_num(rs_length);
1811 }
1812 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1814 double region_elapsed_time_ms =
1815 predict_rs_scan_time_ms(card_num) +
1816 predict_object_copy_time_ms(bytes_to_copy);
1818 if (young)
1819 region_elapsed_time_ms += predict_young_other_time_ms(1);
1820 else
1821 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1823 return region_elapsed_time_ms;
1824 }
1826 size_t
1827 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1828 size_t bytes_to_copy;
1829 if (hr->is_marked())
1830 bytes_to_copy = hr->max_live_bytes();
1831 else {
1832 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1,
1833 "invariant" );
1834 int age = hr->age_in_surv_rate_group();
1835 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1836 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1837 }
1839 return bytes_to_copy;
1840 }
1842 void
1843 G1CollectorPolicy::init_cset_region_lengths(size_t eden_cset_region_length,
1844 size_t survivor_cset_region_length) {
1845 _eden_cset_region_length = eden_cset_region_length;
1846 _survivor_cset_region_length = survivor_cset_region_length;
1847 _old_cset_region_length = 0;
1848 }
1850 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1851 _recorded_rs_lengths = rs_lengths;
1852 }
1854 void G1CollectorPolicy::check_if_region_is_too_expensive(double
1855 predicted_time_ms) {
1856 // I don't think we need to do this when in young GC mode since
1857 // marking will be initiated next time we hit the soft limit anyway...
1858 if (predicted_time_ms > _expensive_region_limit_ms) {
1859 ergo_verbose2(ErgoMixedGCs,
1860 "request mixed GCs end",
1861 ergo_format_reason("predicted region time higher than threshold")
1862 ergo_format_ms("predicted region time")
1863 ergo_format_ms("threshold"),
1864 predicted_time_ms, _expensive_region_limit_ms);
1865 // no point in doing another mixed GC
1866 _should_revert_to_young_gcs = true;
1867 }
1868 }
1870 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1871 double elapsed_ms) {
1872 _recent_gc_times_ms->add(elapsed_ms);
1873 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1874 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1875 }
1877 size_t G1CollectorPolicy::expansion_amount() {
1878 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1879 double threshold = _gc_overhead_perc;
1880 if (recent_gc_overhead > threshold) {
1881 // We will double the existing space, or take
1882 // G1ExpandByPercentOfAvailable % of the available expansion
1883 // space, whichever is smaller, bounded below by a minimum
1884 // expansion (unless that's all that's left.)
1885 const size_t min_expand_bytes = 1*M;
1886 size_t reserved_bytes = _g1->max_capacity();
1887 size_t committed_bytes = _g1->capacity();
1888 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1889 size_t expand_bytes;
1890 size_t expand_bytes_via_pct =
1891 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1892 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1893 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1894 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1896 ergo_verbose5(ErgoHeapSizing,
1897 "attempt heap expansion",
1898 ergo_format_reason("recent GC overhead higher than "
1899 "threshold after GC")
1900 ergo_format_perc("recent GC overhead")
1901 ergo_format_perc("threshold")
1902 ergo_format_byte("uncommitted")
1903 ergo_format_byte_perc("calculated expansion amount"),
1904 recent_gc_overhead, threshold,
1905 uncommitted_bytes,
1906 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
1908 return expand_bytes;
1909 } else {
1910 return 0;
1911 }
1912 }
1914 class CountCSClosure: public HeapRegionClosure {
1915 G1CollectorPolicy* _g1_policy;
1916 public:
1917 CountCSClosure(G1CollectorPolicy* g1_policy) :
1918 _g1_policy(g1_policy) {}
1919 bool doHeapRegion(HeapRegion* r) {
1920 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
1921 return false;
1922 }
1923 };
1925 void G1CollectorPolicy::count_CS_bytes_used() {
1926 CountCSClosure cs_closure(this);
1927 _g1->collection_set_iterate(&cs_closure);
1928 }
1930 void G1CollectorPolicy::print_summary(int level,
1931 const char* str,
1932 NumberSeq* seq) const {
1933 double sum = seq->sum();
1934 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
1935 str, sum / 1000.0, seq->avg());
1936 }
1938 void G1CollectorPolicy::print_summary_sd(int level,
1939 const char* str,
1940 NumberSeq* seq) const {
1941 print_summary(level, str, seq);
1942 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
1943 seq->num(), seq->sd(), seq->maximum());
1944 }
1946 void G1CollectorPolicy::check_other_times(int level,
1947 NumberSeq* other_times_ms,
1948 NumberSeq* calc_other_times_ms) const {
1949 bool should_print = false;
1950 LineBuffer buf(level + 2);
1952 double max_sum = MAX2(fabs(other_times_ms->sum()),
1953 fabs(calc_other_times_ms->sum()));
1954 double min_sum = MIN2(fabs(other_times_ms->sum()),
1955 fabs(calc_other_times_ms->sum()));
1956 double sum_ratio = max_sum / min_sum;
1957 if (sum_ratio > 1.1) {
1958 should_print = true;
1959 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
1960 }
1962 double max_avg = MAX2(fabs(other_times_ms->avg()),
1963 fabs(calc_other_times_ms->avg()));
1964 double min_avg = MIN2(fabs(other_times_ms->avg()),
1965 fabs(calc_other_times_ms->avg()));
1966 double avg_ratio = max_avg / min_avg;
1967 if (avg_ratio > 1.1) {
1968 should_print = true;
1969 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
1970 }
1972 if (other_times_ms->sum() < -0.01) {
1973 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
1974 }
1976 if (other_times_ms->avg() < -0.01) {
1977 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
1978 }
1980 if (calc_other_times_ms->sum() < -0.01) {
1981 should_print = true;
1982 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
1983 }
1985 if (calc_other_times_ms->avg() < -0.01) {
1986 should_print = true;
1987 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
1988 }
1990 if (should_print)
1991 print_summary(level, "Other(Calc)", calc_other_times_ms);
1992 }
1994 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
1995 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1996 MainBodySummary* body_summary = summary->main_body_summary();
1997 if (summary->get_total_seq()->num() > 0) {
1998 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
1999 if (body_summary != NULL) {
2000 if (parallel) {
2001 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
2002 print_summary(2, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
2003 print_summary(2, "SATB Filtering", body_summary->get_satb_filtering_seq());
2004 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
2005 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
2006 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
2007 print_summary(2, "Termination", body_summary->get_termination_seq());
2008 print_summary(2, "Parallel Other", body_summary->get_parallel_other_seq());
2009 {
2010 NumberSeq* other_parts[] = {
2011 body_summary->get_ext_root_scan_seq(),
2012 body_summary->get_satb_filtering_seq(),
2013 body_summary->get_update_rs_seq(),
2014 body_summary->get_scan_rs_seq(),
2015 body_summary->get_obj_copy_seq(),
2016 body_summary->get_termination_seq()
2017 };
2018 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
2019 6, other_parts);
2020 check_other_times(2, body_summary->get_parallel_other_seq(),
2021 &calc_other_times_ms);
2022 }
2023 } else {
2024 print_summary(1, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
2025 print_summary(1, "SATB Filtering", body_summary->get_satb_filtering_seq());
2026 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
2027 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
2028 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
2029 }
2030 }
2031 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq());
2032 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
2033 print_summary(1, "Other", summary->get_other_seq());
2034 {
2035 if (body_summary != NULL) {
2036 NumberSeq calc_other_times_ms;
2037 if (parallel) {
2038 // parallel
2039 NumberSeq* other_parts[] = {
2040 body_summary->get_satb_drain_seq(),
2041 body_summary->get_parallel_seq(),
2042 body_summary->get_clear_ct_seq()
2043 };
2044 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2045 3, other_parts);
2046 } else {
2047 // serial
2048 NumberSeq* other_parts[] = {
2049 body_summary->get_satb_drain_seq(),
2050 body_summary->get_update_rs_seq(),
2051 body_summary->get_ext_root_scan_seq(),
2052 body_summary->get_satb_filtering_seq(),
2053 body_summary->get_scan_rs_seq(),
2054 body_summary->get_obj_copy_seq()
2055 };
2056 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2057 6, other_parts);
2058 }
2059 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
2060 }
2061 }
2062 } else {
2063 LineBuffer(1).append_and_print_cr("none");
2064 }
2065 LineBuffer(0).append_and_print_cr("");
2066 }
2068 void G1CollectorPolicy::print_tracing_info() const {
2069 if (TraceGen0Time) {
2070 gclog_or_tty->print_cr("ALL PAUSES");
2071 print_summary_sd(0, "Total", _all_pause_times_ms);
2072 gclog_or_tty->print_cr("");
2073 gclog_or_tty->print_cr("");
2074 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
2075 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
2076 gclog_or_tty->print_cr("");
2078 gclog_or_tty->print_cr("EVACUATION PAUSES");
2079 print_summary(_summary);
2081 gclog_or_tty->print_cr("MISC");
2082 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
2083 print_summary_sd(0, "Yields", _all_yield_times_ms);
2084 for (int i = 0; i < _aux_num; ++i) {
2085 if (_all_aux_times_ms[i].num() > 0) {
2086 char buffer[96];
2087 sprintf(buffer, "Aux%d", i);
2088 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
2089 }
2090 }
2091 }
2092 if (TraceGen1Time) {
2093 if (_all_full_gc_times_ms->num() > 0) {
2094 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2095 _all_full_gc_times_ms->num(),
2096 _all_full_gc_times_ms->sum() / 1000.0);
2097 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
2098 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2099 _all_full_gc_times_ms->sd(),
2100 _all_full_gc_times_ms->maximum());
2101 }
2102 }
2103 }
2105 void G1CollectorPolicy::print_yg_surv_rate_info() const {
2106 #ifndef PRODUCT
2107 _short_lived_surv_rate_group->print_surv_rate_summary();
2108 // add this call for any other surv rate groups
2109 #endif // PRODUCT
2110 }
2112 #ifndef PRODUCT
2113 // for debugging, bit of a hack...
2114 static char*
2115 region_num_to_mbs(int length) {
2116 static char buffer[64];
2117 double bytes = (double) (length * HeapRegion::GrainBytes);
2118 double mbs = bytes / (double) (1024 * 1024);
2119 sprintf(buffer, "%7.2lfMB", mbs);
2120 return buffer;
2121 }
2122 #endif // PRODUCT
2124 size_t G1CollectorPolicy::max_regions(int purpose) {
2125 switch (purpose) {
2126 case GCAllocForSurvived:
2127 return _max_survivor_regions;
2128 case GCAllocForTenured:
2129 return REGIONS_UNLIMITED;
2130 default:
2131 ShouldNotReachHere();
2132 return REGIONS_UNLIMITED;
2133 };
2134 }
2136 void G1CollectorPolicy::update_max_gc_locker_expansion() {
2137 size_t expansion_region_num = 0;
2138 if (GCLockerEdenExpansionPercent > 0) {
2139 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
2140 double expansion_region_num_d = perc * (double) _young_list_target_length;
2141 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
2142 // less than 1.0) we'll get 1.
2143 expansion_region_num = (size_t) ceil(expansion_region_num_d);
2144 } else {
2145 assert(expansion_region_num == 0, "sanity");
2146 }
2147 _young_list_max_length = _young_list_target_length + expansion_region_num;
2148 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
2149 }
2151 // Calculates survivor space parameters.
2152 void G1CollectorPolicy::update_survivors_policy() {
2153 double max_survivor_regions_d =
2154 (double) _young_list_target_length / (double) SurvivorRatio;
2155 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
2156 // smaller than 1.0) we'll get 1.
2157 _max_survivor_regions = (size_t) ceil(max_survivor_regions_d);
2159 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2160 HeapRegion::GrainWords * _max_survivor_regions);
2161 }
2163 #ifndef PRODUCT
2164 class HRSortIndexIsOKClosure: public HeapRegionClosure {
2165 CollectionSetChooser* _chooser;
2166 public:
2167 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) :
2168 _chooser(chooser) {}
2170 bool doHeapRegion(HeapRegion* r) {
2171 if (!r->continuesHumongous()) {
2172 assert(_chooser->regionProperlyOrdered(r), "Ought to be.");
2173 }
2174 return false;
2175 }
2176 };
2178 bool G1CollectorPolicy::assertMarkedBytesDataOK() {
2179 HRSortIndexIsOKClosure cl(_collectionSetChooser);
2180 _g1->heap_region_iterate(&cl);
2181 return true;
2182 }
2183 #endif
2185 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
2186 GCCause::Cause gc_cause) {
2187 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2188 if (!during_cycle) {
2189 ergo_verbose1(ErgoConcCycles,
2190 "request concurrent cycle initiation",
2191 ergo_format_reason("requested by GC cause")
2192 ergo_format_str("GC cause"),
2193 GCCause::to_string(gc_cause));
2194 set_initiate_conc_mark_if_possible();
2195 return true;
2196 } else {
2197 ergo_verbose1(ErgoConcCycles,
2198 "do not request concurrent cycle initiation",
2199 ergo_format_reason("concurrent cycle already in progress")
2200 ergo_format_str("GC cause"),
2201 GCCause::to_string(gc_cause));
2202 return false;
2203 }
2204 }
2206 void
2207 G1CollectorPolicy::decide_on_conc_mark_initiation() {
2208 // We are about to decide on whether this pause will be an
2209 // initial-mark pause.
2211 // First, during_initial_mark_pause() should not be already set. We
2212 // will set it here if we have to. However, it should be cleared by
2213 // the end of the pause (it's only set for the duration of an
2214 // initial-mark pause).
2215 assert(!during_initial_mark_pause(), "pre-condition");
2217 if (initiate_conc_mark_if_possible()) {
2218 // We had noticed on a previous pause that the heap occupancy has
2219 // gone over the initiating threshold and we should start a
2220 // concurrent marking cycle. So we might initiate one.
2222 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2223 if (!during_cycle) {
2224 // The concurrent marking thread is not "during a cycle", i.e.,
2225 // it has completed the last one. So we can go ahead and
2226 // initiate a new cycle.
2228 set_during_initial_mark_pause();
2229 // We do not allow mixed GCs during marking.
2230 if (!gcs_are_young()) {
2231 set_gcs_are_young(true);
2232 ergo_verbose0(ErgoMixedGCs,
2233 "end mixed GCs",
2234 ergo_format_reason("concurrent cycle is about to start"));
2235 }
2237 // And we can now clear initiate_conc_mark_if_possible() as
2238 // we've already acted on it.
2239 clear_initiate_conc_mark_if_possible();
2241 ergo_verbose0(ErgoConcCycles,
2242 "initiate concurrent cycle",
2243 ergo_format_reason("concurrent cycle initiation requested"));
2244 } else {
2245 // The concurrent marking thread is still finishing up the
2246 // previous cycle. If we start one right now the two cycles
2247 // overlap. In particular, the concurrent marking thread might
2248 // be in the process of clearing the next marking bitmap (which
2249 // we will use for the next cycle if we start one). Starting a
2250 // cycle now will be bad given that parts of the marking
2251 // information might get cleared by the marking thread. And we
2252 // cannot wait for the marking thread to finish the cycle as it
2253 // periodically yields while clearing the next marking bitmap
2254 // and, if it's in a yield point, it's waiting for us to
2255 // finish. So, at this point we will not start a cycle and we'll
2256 // let the concurrent marking thread complete the last one.
2257 ergo_verbose0(ErgoConcCycles,
2258 "do not initiate concurrent cycle",
2259 ergo_format_reason("concurrent cycle already in progress"));
2260 }
2261 }
2262 }
2264 class KnownGarbageClosure: public HeapRegionClosure {
2265 CollectionSetChooser* _hrSorted;
2267 public:
2268 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2269 _hrSorted(hrSorted)
2270 {}
2272 bool doHeapRegion(HeapRegion* r) {
2273 // We only include humongous regions in collection
2274 // sets when concurrent mark shows that their contained object is
2275 // unreachable.
2277 // Do we have any marking information for this region?
2278 if (r->is_marked()) {
2279 // We don't include humongous regions in collection
2280 // sets because we collect them immediately at the end of a marking
2281 // cycle. We also don't include young regions because we *must*
2282 // include them in the next collection pause.
2283 if (!r->isHumongous() && !r->is_young()) {
2284 _hrSorted->addMarkedHeapRegion(r);
2285 }
2286 }
2287 return false;
2288 }
2289 };
2291 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2292 CollectionSetChooser* _hrSorted;
2293 jint _marked_regions_added;
2294 jint _chunk_size;
2295 jint _cur_chunk_idx;
2296 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2297 int _worker;
2298 int _invokes;
2300 void get_new_chunk() {
2301 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size);
2302 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2303 }
2304 void add_region(HeapRegion* r) {
2305 if (_cur_chunk_idx == _cur_chunk_end) {
2306 get_new_chunk();
2307 }
2308 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2309 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r);
2310 _marked_regions_added++;
2311 _cur_chunk_idx++;
2312 }
2314 public:
2315 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2316 jint chunk_size,
2317 int worker) :
2318 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker),
2319 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0),
2320 _invokes(0)
2321 {}
2323 bool doHeapRegion(HeapRegion* r) {
2324 // We only include humongous regions in collection
2325 // sets when concurrent mark shows that their contained object is
2326 // unreachable.
2327 _invokes++;
2329 // Do we have any marking information for this region?
2330 if (r->is_marked()) {
2331 // We don't include humongous regions in collection
2332 // sets because we collect them immediately at the end of a marking
2333 // cycle.
2334 // We also do not include young regions in collection sets
2335 if (!r->isHumongous() && !r->is_young()) {
2336 add_region(r);
2337 }
2338 }
2339 return false;
2340 }
2341 jint marked_regions_added() { return _marked_regions_added; }
2342 int invokes() { return _invokes; }
2343 };
2345 class ParKnownGarbageTask: public AbstractGangTask {
2346 CollectionSetChooser* _hrSorted;
2347 jint _chunk_size;
2348 G1CollectedHeap* _g1;
2349 public:
2350 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) :
2351 AbstractGangTask("ParKnownGarbageTask"),
2352 _hrSorted(hrSorted), _chunk_size(chunk_size),
2353 _g1(G1CollectedHeap::heap())
2354 {}
2356 void work(uint worker_id) {
2357 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted,
2358 _chunk_size,
2359 worker_id);
2360 // Back to zero for the claim value.
2361 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id,
2362 _g1->workers()->active_workers(),
2363 HeapRegion::InitialClaimValue);
2364 jint regions_added = parKnownGarbageCl.marked_regions_added();
2365 _hrSorted->incNumMarkedHeapRegions(regions_added);
2366 if (G1PrintParCleanupStats) {
2367 gclog_or_tty->print_cr(" Thread %d called %d times, added %d regions to list.",
2368 worker_id, parKnownGarbageCl.invokes(), regions_added);
2369 }
2370 }
2371 };
2373 void
2374 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) {
2375 double start_sec;
2376 if (G1PrintParCleanupStats) {
2377 start_sec = os::elapsedTime();
2378 }
2380 _collectionSetChooser->clearMarkedHeapRegions();
2381 double clear_marked_end_sec;
2382 if (G1PrintParCleanupStats) {
2383 clear_marked_end_sec = os::elapsedTime();
2384 gclog_or_tty->print_cr(" clear marked regions: %8.3f ms.",
2385 (clear_marked_end_sec - start_sec) * 1000.0);
2386 }
2388 if (G1CollectedHeap::use_parallel_gc_threads()) {
2389 const size_t OverpartitionFactor = 4;
2390 size_t WorkUnit;
2391 // The use of MinChunkSize = 8 in the original code
2392 // causes some assertion failures when the total number of
2393 // region is less than 8. The code here tries to fix that.
2394 // Should the original code also be fixed?
2395 if (no_of_gc_threads > 0) {
2396 const size_t MinWorkUnit =
2397 MAX2(_g1->n_regions() / no_of_gc_threads, (size_t) 1U);
2398 WorkUnit =
2399 MAX2(_g1->n_regions() / (no_of_gc_threads * OverpartitionFactor),
2400 MinWorkUnit);
2401 } else {
2402 assert(no_of_gc_threads > 0,
2403 "The active gc workers should be greater than 0");
2404 // In a product build do something reasonable to avoid a crash.
2405 const size_t MinWorkUnit =
2406 MAX2(_g1->n_regions() / ParallelGCThreads, (size_t) 1U);
2407 WorkUnit =
2408 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor),
2409 MinWorkUnit);
2410 }
2411 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(),
2412 WorkUnit);
2413 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2414 (int) WorkUnit);
2415 _g1->workers()->run_task(&parKnownGarbageTask);
2417 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2418 "sanity check");
2419 } else {
2420 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2421 _g1->heap_region_iterate(&knownGarbagecl);
2422 }
2423 double known_garbage_end_sec;
2424 if (G1PrintParCleanupStats) {
2425 known_garbage_end_sec = os::elapsedTime();
2426 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.",
2427 (known_garbage_end_sec - clear_marked_end_sec) * 1000.0);
2428 }
2430 _collectionSetChooser->sortMarkedHeapRegions();
2431 double end_sec = os::elapsedTime();
2432 if (G1PrintParCleanupStats) {
2433 gclog_or_tty->print_cr(" sorting: %8.3f ms.",
2434 (end_sec - known_garbage_end_sec) * 1000.0);
2435 }
2437 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
2438 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
2439 _cur_mark_stop_world_time_ms += elapsed_time_ms;
2440 _prev_collection_pause_end_ms += elapsed_time_ms;
2441 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
2442 }
2444 // Add the heap region at the head of the non-incremental collection set
2445 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
2446 assert(_inc_cset_build_state == Active, "Precondition");
2447 assert(!hr->is_young(), "non-incremental add of young region");
2449 assert(!hr->in_collection_set(), "should not already be in the CSet");
2450 hr->set_in_collection_set(true);
2451 hr->set_next_in_collection_set(_collection_set);
2452 _collection_set = hr;
2453 _collection_set_bytes_used_before += hr->used();
2454 _g1->register_region_with_in_cset_fast_test(hr);
2455 size_t rs_length = hr->rem_set()->occupied();
2456 _recorded_rs_lengths += rs_length;
2457 _old_cset_region_length += 1;
2458 }
2460 // Initialize the per-collection-set information
2461 void G1CollectorPolicy::start_incremental_cset_building() {
2462 assert(_inc_cset_build_state == Inactive, "Precondition");
2464 _inc_cset_head = NULL;
2465 _inc_cset_tail = NULL;
2466 _inc_cset_bytes_used_before = 0;
2468 _inc_cset_max_finger = 0;
2469 _inc_cset_recorded_rs_lengths = 0;
2470 _inc_cset_recorded_rs_lengths_diffs = 0;
2471 _inc_cset_predicted_elapsed_time_ms = 0.0;
2472 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
2473 _inc_cset_build_state = Active;
2474 }
2476 void G1CollectorPolicy::finalize_incremental_cset_building() {
2477 assert(_inc_cset_build_state == Active, "Precondition");
2478 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2480 // The two "main" fields, _inc_cset_recorded_rs_lengths and
2481 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread
2482 // that adds a new region to the CSet. Further updates by the
2483 // concurrent refinement thread that samples the young RSet lengths
2484 // are accumulated in the *_diffs fields. Here we add the diffs to
2485 // the "main" fields.
2487 if (_inc_cset_recorded_rs_lengths_diffs >= 0) {
2488 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs;
2489 } else {
2490 // This is defensive. The diff should in theory be always positive
2491 // as RSets can only grow between GCs. However, given that we
2492 // sample their size concurrently with other threads updating them
2493 // it's possible that we might get the wrong size back, which
2494 // could make the calculations somewhat inaccurate.
2495 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs);
2496 if (_inc_cset_recorded_rs_lengths >= diffs) {
2497 _inc_cset_recorded_rs_lengths -= diffs;
2498 } else {
2499 _inc_cset_recorded_rs_lengths = 0;
2500 }
2501 }
2502 _inc_cset_predicted_elapsed_time_ms +=
2503 _inc_cset_predicted_elapsed_time_ms_diffs;
2505 _inc_cset_recorded_rs_lengths_diffs = 0;
2506 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
2507 }
2509 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
2510 // This routine is used when:
2511 // * adding survivor regions to the incremental cset at the end of an
2512 // evacuation pause,
2513 // * adding the current allocation region to the incremental cset
2514 // when it is retired, and
2515 // * updating existing policy information for a region in the
2516 // incremental cset via young list RSet sampling.
2517 // Therefore this routine may be called at a safepoint by the
2518 // VM thread, or in-between safepoints by mutator threads (when
2519 // retiring the current allocation region) or a concurrent
2520 // refine thread (RSet sampling).
2522 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2523 size_t used_bytes = hr->used();
2524 _inc_cset_recorded_rs_lengths += rs_length;
2525 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
2526 _inc_cset_bytes_used_before += used_bytes;
2528 // Cache the values we have added to the aggregated informtion
2529 // in the heap region in case we have to remove this region from
2530 // the incremental collection set, or it is updated by the
2531 // rset sampling code
2532 hr->set_recorded_rs_length(rs_length);
2533 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
2534 }
2536 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr,
2537 size_t new_rs_length) {
2538 // Update the CSet information that is dependent on the new RS length
2539 assert(hr->is_young(), "Precondition");
2540 assert(!SafepointSynchronize::is_at_safepoint(),
2541 "should not be at a safepoint");
2543 // We could have updated _inc_cset_recorded_rs_lengths and
2544 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do
2545 // that atomically, as this code is executed by a concurrent
2546 // refinement thread, potentially concurrently with a mutator thread
2547 // allocating a new region and also updating the same fields. To
2548 // avoid the atomic operations we accumulate these updates on two
2549 // separate fields (*_diffs) and we'll just add them to the "main"
2550 // fields at the start of a GC.
2552 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length();
2553 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length;
2554 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff;
2556 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
2557 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2558 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms;
2559 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff;
2561 hr->set_recorded_rs_length(new_rs_length);
2562 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms);
2563 }
2565 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
2566 assert(hr->is_young(), "invariant");
2567 assert(hr->young_index_in_cset() > -1, "should have already been set");
2568 assert(_inc_cset_build_state == Active, "Precondition");
2570 // We need to clear and set the cached recorded/cached collection set
2571 // information in the heap region here (before the region gets added
2572 // to the collection set). An individual heap region's cached values
2573 // are calculated, aggregated with the policy collection set info,
2574 // and cached in the heap region here (initially) and (subsequently)
2575 // by the Young List sampling code.
2577 size_t rs_length = hr->rem_set()->occupied();
2578 add_to_incremental_cset_info(hr, rs_length);
2580 HeapWord* hr_end = hr->end();
2581 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
2583 assert(!hr->in_collection_set(), "invariant");
2584 hr->set_in_collection_set(true);
2585 assert( hr->next_in_collection_set() == NULL, "invariant");
2587 _g1->register_region_with_in_cset_fast_test(hr);
2588 }
2590 // Add the region at the RHS of the incremental cset
2591 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
2592 // We should only ever be appending survivors at the end of a pause
2593 assert( hr->is_survivor(), "Logic");
2595 // Do the 'common' stuff
2596 add_region_to_incremental_cset_common(hr);
2598 // Now add the region at the right hand side
2599 if (_inc_cset_tail == NULL) {
2600 assert(_inc_cset_head == NULL, "invariant");
2601 _inc_cset_head = hr;
2602 } else {
2603 _inc_cset_tail->set_next_in_collection_set(hr);
2604 }
2605 _inc_cset_tail = hr;
2606 }
2608 // Add the region to the LHS of the incremental cset
2609 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
2610 // Survivors should be added to the RHS at the end of a pause
2611 assert(!hr->is_survivor(), "Logic");
2613 // Do the 'common' stuff
2614 add_region_to_incremental_cset_common(hr);
2616 // Add the region at the left hand side
2617 hr->set_next_in_collection_set(_inc_cset_head);
2618 if (_inc_cset_head == NULL) {
2619 assert(_inc_cset_tail == NULL, "Invariant");
2620 _inc_cset_tail = hr;
2621 }
2622 _inc_cset_head = hr;
2623 }
2625 #ifndef PRODUCT
2626 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
2627 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
2629 st->print_cr("\nCollection_set:");
2630 HeapRegion* csr = list_head;
2631 while (csr != NULL) {
2632 HeapRegion* next = csr->next_in_collection_set();
2633 assert(csr->in_collection_set(), "bad CS");
2634 st->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
2635 "age: %4d, y: %d, surv: %d",
2636 csr->bottom(), csr->end(),
2637 csr->top(),
2638 csr->prev_top_at_mark_start(),
2639 csr->next_top_at_mark_start(),
2640 csr->top_at_conc_mark_count(),
2641 csr->age_in_surv_rate_group_cond(),
2642 csr->is_young(),
2643 csr->is_survivor());
2644 csr = next;
2645 }
2646 }
2647 #endif // !PRODUCT
2649 void G1CollectorPolicy::choose_collection_set(double target_pause_time_ms) {
2650 // Set this here - in case we're not doing young collections.
2651 double non_young_start_time_sec = os::elapsedTime();
2653 YoungList* young_list = _g1->young_list();
2654 finalize_incremental_cset_building();
2656 guarantee(target_pause_time_ms > 0.0,
2657 err_msg("target_pause_time_ms = %1.6lf should be positive",
2658 target_pause_time_ms));
2659 guarantee(_collection_set == NULL, "Precondition");
2661 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2662 double predicted_pause_time_ms = base_time_ms;
2664 double time_remaining_ms = target_pause_time_ms - base_time_ms;
2666 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2667 "start choosing CSet",
2668 ergo_format_ms("predicted base time")
2669 ergo_format_ms("remaining time")
2670 ergo_format_ms("target pause time"),
2671 base_time_ms, time_remaining_ms, target_pause_time_ms);
2673 // the 10% and 50% values are arbitrary...
2674 double threshold = 0.10 * target_pause_time_ms;
2675 if (time_remaining_ms < threshold) {
2676 double prev_time_remaining_ms = time_remaining_ms;
2677 time_remaining_ms = 0.50 * target_pause_time_ms;
2678 ergo_verbose3(ErgoCSetConstruction,
2679 "adjust remaining time",
2680 ergo_format_reason("remaining time lower than threshold")
2681 ergo_format_ms("remaining time")
2682 ergo_format_ms("threshold")
2683 ergo_format_ms("adjusted remaining time"),
2684 prev_time_remaining_ms, threshold, time_remaining_ms);
2685 }
2687 size_t expansion_bytes = _g1->expansion_regions() * HeapRegion::GrainBytes;
2689 HeapRegion* hr;
2690 double young_start_time_sec = os::elapsedTime();
2692 _collection_set_bytes_used_before = 0;
2693 _last_gc_was_young = gcs_are_young() ? true : false;
2695 if (_last_gc_was_young) {
2696 ++_young_pause_num;
2697 } else {
2698 ++_mixed_pause_num;
2699 }
2701 // The young list is laid with the survivor regions from the previous
2702 // pause are appended to the RHS of the young list, i.e.
2703 // [Newly Young Regions ++ Survivors from last pause].
2705 size_t survivor_region_length = young_list->survivor_length();
2706 size_t eden_region_length = young_list->length() - survivor_region_length;
2707 init_cset_region_lengths(eden_region_length, survivor_region_length);
2708 hr = young_list->first_survivor_region();
2709 while (hr != NULL) {
2710 assert(hr->is_survivor(), "badly formed young list");
2711 hr->set_young();
2712 hr = hr->get_next_young_region();
2713 }
2715 // Clear the fields that point to the survivor list - they are all young now.
2716 young_list->clear_survivors();
2718 _collection_set = _inc_cset_head;
2719 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2720 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms;
2721 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
2723 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2724 "add young regions to CSet",
2725 ergo_format_region("eden")
2726 ergo_format_region("survivors")
2727 ergo_format_ms("predicted young region time"),
2728 eden_region_length, survivor_region_length,
2729 _inc_cset_predicted_elapsed_time_ms);
2731 // The number of recorded young regions is the incremental
2732 // collection set's current size
2733 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2735 double young_end_time_sec = os::elapsedTime();
2736 _recorded_young_cset_choice_time_ms =
2737 (young_end_time_sec - young_start_time_sec) * 1000.0;
2739 // We are doing young collections so reset this.
2740 non_young_start_time_sec = young_end_time_sec;
2742 if (!gcs_are_young()) {
2743 bool should_continue = true;
2744 NumberSeq seq;
2745 double avg_prediction = 100000000000000000.0; // something very large
2747 double prev_predicted_pause_time_ms = predicted_pause_time_ms;
2748 do {
2749 // Note that add_old_region_to_cset() increments the
2750 // _old_cset_region_length field and cset_region_length() returns the
2751 // sum of _eden_cset_region_length, _survivor_cset_region_length, and
2752 // _old_cset_region_length. So, as old regions are added to the
2753 // CSet, _old_cset_region_length will be incremented and
2754 // cset_region_length(), which is used below, will always reflect
2755 // the the total number of regions added up to this point to the CSet.
2757 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms,
2758 avg_prediction);
2759 if (hr != NULL) {
2760 _g1->old_set_remove(hr);
2761 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
2762 time_remaining_ms -= predicted_time_ms;
2763 predicted_pause_time_ms += predicted_time_ms;
2764 add_old_region_to_cset(hr);
2765 seq.add(predicted_time_ms);
2766 avg_prediction = seq.avg() + seq.sd();
2767 }
2769 should_continue = true;
2770 if (hr == NULL) {
2771 // No need for an ergo verbose message here,
2772 // getNextMarkRegion() does this when it returns NULL.
2773 should_continue = false;
2774 } else {
2775 if (adaptive_young_list_length()) {
2776 if (time_remaining_ms < 0.0) {
2777 ergo_verbose1(ErgoCSetConstruction,
2778 "stop adding old regions to CSet",
2779 ergo_format_reason("remaining time is lower than 0")
2780 ergo_format_ms("remaining time"),
2781 time_remaining_ms);
2782 should_continue = false;
2783 }
2784 } else {
2785 if (cset_region_length() >= _young_list_fixed_length) {
2786 ergo_verbose2(ErgoCSetConstruction,
2787 "stop adding old regions to CSet",
2788 ergo_format_reason("CSet length reached target")
2789 ergo_format_region("CSet")
2790 ergo_format_region("young target"),
2791 cset_region_length(), _young_list_fixed_length);
2792 should_continue = false;
2793 }
2794 }
2795 }
2796 } while (should_continue);
2798 if (!adaptive_young_list_length() &&
2799 cset_region_length() < _young_list_fixed_length) {
2800 ergo_verbose2(ErgoCSetConstruction,
2801 "request mixed GCs end",
2802 ergo_format_reason("CSet length lower than target")
2803 ergo_format_region("CSet")
2804 ergo_format_region("young target"),
2805 cset_region_length(), _young_list_fixed_length);
2806 _should_revert_to_young_gcs = true;
2807 }
2809 ergo_verbose2(ErgoCSetConstruction | ErgoHigh,
2810 "add old regions to CSet",
2811 ergo_format_region("old")
2812 ergo_format_ms("predicted old region time"),
2813 old_cset_region_length(),
2814 predicted_pause_time_ms - prev_predicted_pause_time_ms);
2815 }
2817 stop_incremental_cset_building();
2819 count_CS_bytes_used();
2821 ergo_verbose5(ErgoCSetConstruction,
2822 "finish choosing CSet",
2823 ergo_format_region("eden")
2824 ergo_format_region("survivors")
2825 ergo_format_region("old")
2826 ergo_format_ms("predicted pause time")
2827 ergo_format_ms("target pause time"),
2828 eden_region_length, survivor_region_length,
2829 old_cset_region_length(),
2830 predicted_pause_time_ms, target_pause_time_ms);
2832 double non_young_end_time_sec = os::elapsedTime();
2833 _recorded_non_young_cset_choice_time_ms =
2834 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
2835 }