Tue, 10 Jan 2012 18:58:13 -0500
6888336: G1: avoid explicitly marking and pushing objects in survivor spaces
Summary: This change simplifies the interaction between GC and concurrent marking. By disabling survivor spaces during the initial-mark pause we don't need to propagate marks of objects we copy during each GC (since we never need to copy an explicitly marked object).
Reviewed-by: johnc, brutisso
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 _prev_collection_pause_used_at_end_bytes(0),
218 _eden_cset_region_length(0),
219 _survivor_cset_region_length(0),
220 _old_cset_region_length(0),
222 _collection_set(NULL),
223 _collection_set_bytes_used_before(0),
225 // Incremental CSet attributes
226 _inc_cset_build_state(Inactive),
227 _inc_cset_head(NULL),
228 _inc_cset_tail(NULL),
229 _inc_cset_bytes_used_before(0),
230 _inc_cset_max_finger(NULL),
231 _inc_cset_recorded_rs_lengths(0),
232 _inc_cset_recorded_rs_lengths_diffs(0),
233 _inc_cset_predicted_elapsed_time_ms(0.0),
234 _inc_cset_predicted_elapsed_time_ms_diffs(0.0),
236 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
237 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
238 #endif // _MSC_VER
240 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
241 G1YoungSurvRateNumRegionsSummary)),
242 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
243 G1YoungSurvRateNumRegionsSummary)),
244 // add here any more surv rate groups
245 _recorded_survivor_regions(0),
246 _recorded_survivor_head(NULL),
247 _recorded_survivor_tail(NULL),
248 _survivors_age_table(true),
250 _gc_overhead_perc(0.0) {
252 // Set up the region size and associated fields. Given that the
253 // policy is created before the heap, we have to set this up here,
254 // so it's done as soon as possible.
255 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
256 HeapRegionRemSet::setup_remset_size();
258 G1ErgoVerbose::initialize();
259 if (PrintAdaptiveSizePolicy) {
260 // Currently, we only use a single switch for all the heuristics.
261 G1ErgoVerbose::set_enabled(true);
262 // Given that we don't currently have a verboseness level
263 // parameter, we'll hardcode this to high. This can be easily
264 // changed in the future.
265 G1ErgoVerbose::set_level(ErgoHigh);
266 } else {
267 G1ErgoVerbose::set_enabled(false);
268 }
270 // Verify PLAB sizes
271 const size_t region_size = HeapRegion::GrainWords;
272 if (YoungPLABSize > region_size || OldPLABSize > region_size) {
273 char buffer[128];
274 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT,
275 OldPLABSize > region_size ? "Old" : "Young", region_size);
276 vm_exit_during_initialization(buffer);
277 }
279 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
280 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
282 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads];
283 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads];
284 _par_last_satb_filtering_times_ms = new double[_parallel_gc_threads];
286 _par_last_update_rs_times_ms = new double[_parallel_gc_threads];
287 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads];
289 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads];
291 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads];
293 _par_last_termination_times_ms = new double[_parallel_gc_threads];
294 _par_last_termination_attempts = new double[_parallel_gc_threads];
295 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads];
296 _par_last_gc_worker_times_ms = new double[_parallel_gc_threads];
297 _par_last_gc_worker_other_times_ms = new double[_parallel_gc_threads];
299 // start conservatively
300 _expensive_region_limit_ms = 0.5 * (double) MaxGCPauseMillis;
302 int index;
303 if (ParallelGCThreads == 0)
304 index = 0;
305 else if (ParallelGCThreads > 8)
306 index = 7;
307 else
308 index = ParallelGCThreads - 1;
310 _pending_card_diff_seq->add(0.0);
311 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
312 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
313 _young_cards_per_entry_ratio_seq->add(
314 young_cards_per_entry_ratio_defaults[index]);
315 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
316 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
317 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
318 _young_other_cost_per_region_ms_seq->add(
319 young_other_cost_per_region_ms_defaults[index]);
320 _non_young_other_cost_per_region_ms_seq->add(
321 non_young_other_cost_per_region_ms_defaults[index]);
323 // Below, we might need to calculate the pause time target based on
324 // the pause interval. When we do so we are going to give G1 maximum
325 // flexibility and allow it to do pauses when it needs to. So, we'll
326 // arrange that the pause interval to be pause time target + 1 to
327 // ensure that a) the pause time target is maximized with respect to
328 // the pause interval and b) we maintain the invariant that pause
329 // time target < pause interval. If the user does not want this
330 // maximum flexibility, they will have to set the pause interval
331 // explicitly.
333 // First make sure that, if either parameter is set, its value is
334 // reasonable.
335 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
336 if (MaxGCPauseMillis < 1) {
337 vm_exit_during_initialization("MaxGCPauseMillis should be "
338 "greater than 0");
339 }
340 }
341 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
342 if (GCPauseIntervalMillis < 1) {
343 vm_exit_during_initialization("GCPauseIntervalMillis should be "
344 "greater than 0");
345 }
346 }
348 // Then, if the pause time target parameter was not set, set it to
349 // the default value.
350 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
351 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
352 // The default pause time target in G1 is 200ms
353 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
354 } else {
355 // We do not allow the pause interval to be set without the
356 // pause time target
357 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
358 "without setting MaxGCPauseMillis");
359 }
360 }
362 // Then, if the interval parameter was not set, set it according to
363 // the pause time target (this will also deal with the case when the
364 // pause time target is the default value).
365 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
366 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
367 }
369 // Finally, make sure that the two parameters are consistent.
370 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
371 char buffer[256];
372 jio_snprintf(buffer, 256,
373 "MaxGCPauseMillis (%u) should be less than "
374 "GCPauseIntervalMillis (%u)",
375 MaxGCPauseMillis, GCPauseIntervalMillis);
376 vm_exit_during_initialization(buffer);
377 }
379 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
380 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
381 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
382 _sigma = (double) G1ConfidencePercent / 100.0;
384 // start conservatively (around 50ms is about right)
385 _concurrent_mark_remark_times_ms->add(0.05);
386 _concurrent_mark_cleanup_times_ms->add(0.20);
387 _tenuring_threshold = MaxTenuringThreshold;
388 // _max_survivor_regions will be calculated by
389 // update_young_list_target_length() during initialization.
390 _max_survivor_regions = 0;
392 assert(GCTimeRatio > 0,
393 "we should have set it to a default value set_g1_gc_flags() "
394 "if a user set it to 0");
395 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
397 uintx reserve_perc = G1ReservePercent;
398 // Put an artificial ceiling on this so that it's not set to a silly value.
399 if (reserve_perc > 50) {
400 reserve_perc = 50;
401 warning("G1ReservePercent is set to a value that is too large, "
402 "it's been updated to %u", reserve_perc);
403 }
404 _reserve_factor = (double) reserve_perc / 100.0;
405 // This will be set when the heap is expanded
406 // for the first time during initialization.
407 _reserve_regions = 0;
409 initialize_all();
410 _collectionSetChooser = new CollectionSetChooser();
411 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags
412 }
414 void G1CollectorPolicy::initialize_flags() {
415 set_min_alignment(HeapRegion::GrainBytes);
416 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
417 if (SurvivorRatio < 1) {
418 vm_exit_during_initialization("Invalid survivor ratio specified");
419 }
420 CollectorPolicy::initialize_flags();
421 }
423 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true) {
424 assert(G1DefaultMinNewGenPercent <= G1DefaultMaxNewGenPercent, "Min larger than max");
425 assert(G1DefaultMinNewGenPercent > 0 && G1DefaultMinNewGenPercent < 100, "Min out of bounds");
426 assert(G1DefaultMaxNewGenPercent > 0 && G1DefaultMaxNewGenPercent < 100, "Max out of bounds");
428 if (FLAG_IS_CMDLINE(NewRatio)) {
429 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
430 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
431 } else {
432 _sizer_kind = SizerNewRatio;
433 _adaptive_size = false;
434 return;
435 }
436 }
438 if (FLAG_IS_CMDLINE(NewSize)) {
439 _min_desired_young_length = MAX2((size_t) 1, NewSize / HeapRegion::GrainBytes);
440 if (FLAG_IS_CMDLINE(MaxNewSize)) {
441 _max_desired_young_length = MAX2((size_t) 1, MaxNewSize / HeapRegion::GrainBytes);
442 _sizer_kind = SizerMaxAndNewSize;
443 _adaptive_size = _min_desired_young_length == _max_desired_young_length;
444 } else {
445 _sizer_kind = SizerNewSizeOnly;
446 }
447 } else if (FLAG_IS_CMDLINE(MaxNewSize)) {
448 _max_desired_young_length = MAX2((size_t) 1, MaxNewSize / HeapRegion::GrainBytes);
449 _sizer_kind = SizerMaxNewSizeOnly;
450 }
451 }
453 size_t G1YoungGenSizer::calculate_default_min_length(size_t new_number_of_heap_regions) {
454 size_t default_value = (new_number_of_heap_regions * G1DefaultMinNewGenPercent) / 100;
455 return MAX2((size_t)1, default_value);
456 }
458 size_t G1YoungGenSizer::calculate_default_max_length(size_t new_number_of_heap_regions) {
459 size_t default_value = (new_number_of_heap_regions * G1DefaultMaxNewGenPercent) / 100;
460 return MAX2((size_t)1, default_value);
461 }
463 void G1YoungGenSizer::heap_size_changed(size_t new_number_of_heap_regions) {
464 assert(new_number_of_heap_regions > 0, "Heap must be initialized");
466 switch (_sizer_kind) {
467 case SizerDefaults:
468 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
469 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
470 break;
471 case SizerNewSizeOnly:
472 _max_desired_young_length = calculate_default_max_length(new_number_of_heap_regions);
473 _max_desired_young_length = MAX2(_min_desired_young_length, _max_desired_young_length);
474 break;
475 case SizerMaxNewSizeOnly:
476 _min_desired_young_length = calculate_default_min_length(new_number_of_heap_regions);
477 _min_desired_young_length = MIN2(_min_desired_young_length, _max_desired_young_length);
478 break;
479 case SizerMaxAndNewSize:
480 // Do nothing. Values set on the command line, don't update them at runtime.
481 break;
482 case SizerNewRatio:
483 _min_desired_young_length = new_number_of_heap_regions / (NewRatio + 1);
484 _max_desired_young_length = _min_desired_young_length;
485 break;
486 default:
487 ShouldNotReachHere();
488 }
490 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values");
491 }
493 void G1CollectorPolicy::init() {
494 // Set aside an initial future to_space.
495 _g1 = G1CollectedHeap::heap();
497 assert(Heap_lock->owned_by_self(), "Locking discipline.");
499 initialize_gc_policy_counters();
501 if (adaptive_young_list_length()) {
502 _young_list_fixed_length = 0;
503 } else {
504 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length();
505 }
506 _free_regions_at_end_of_collection = _g1->free_regions();
507 update_young_list_target_length();
508 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes;
510 // We may immediately start allocating regions and placing them on the
511 // collection set list. Initialize the per-collection set info
512 start_incremental_cset_building();
513 }
515 // Create the jstat counters for the policy.
516 void G1CollectorPolicy::initialize_gc_policy_counters() {
517 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
518 }
520 bool G1CollectorPolicy::predict_will_fit(size_t young_length,
521 double base_time_ms,
522 size_t base_free_regions,
523 double target_pause_time_ms) {
524 if (young_length >= base_free_regions) {
525 // end condition 1: not enough space for the young regions
526 return false;
527 }
529 double accum_surv_rate = accum_yg_surv_rate_pred((int)(young_length - 1));
530 size_t bytes_to_copy =
531 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
532 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
533 double young_other_time_ms = predict_young_other_time_ms(young_length);
534 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
535 if (pause_time_ms > target_pause_time_ms) {
536 // end condition 2: prediction is over the target pause time
537 return false;
538 }
540 size_t free_bytes =
541 (base_free_regions - young_length) * HeapRegion::GrainBytes;
542 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
543 // end condition 3: out-of-space (conservatively!)
544 return false;
545 }
547 // success!
548 return true;
549 }
551 void G1CollectorPolicy::record_new_heap_size(size_t new_number_of_regions) {
552 // re-calculate the necessary reserve
553 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
554 // We use ceiling so that if reserve_regions_d is > 0.0 (but
555 // smaller than 1.0) we'll get 1.
556 _reserve_regions = (size_t) ceil(reserve_regions_d);
558 _young_gen_sizer->heap_size_changed(new_number_of_regions);
559 }
561 size_t G1CollectorPolicy::calculate_young_list_desired_min_length(
562 size_t base_min_length) {
563 size_t desired_min_length = 0;
564 if (adaptive_young_list_length()) {
565 if (_alloc_rate_ms_seq->num() > 3) {
566 double now_sec = os::elapsedTime();
567 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
568 double alloc_rate_ms = predict_alloc_rate_ms();
569 desired_min_length = (size_t) ceil(alloc_rate_ms * when_ms);
570 } else {
571 // otherwise we don't have enough info to make the prediction
572 }
573 }
574 desired_min_length += base_min_length;
575 // make sure we don't go below any user-defined minimum bound
576 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length);
577 }
579 size_t G1CollectorPolicy::calculate_young_list_desired_max_length() {
580 // Here, we might want to also take into account any additional
581 // constraints (i.e., user-defined minimum bound). Currently, we
582 // effectively don't set this bound.
583 return _young_gen_sizer->max_desired_young_length();
584 }
586 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
587 if (rs_lengths == (size_t) -1) {
588 // if it's set to the default value (-1), we should predict it;
589 // otherwise, use the given value.
590 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
591 }
593 // Calculate the absolute and desired min bounds.
595 // This is how many young regions we already have (currently: the survivors).
596 size_t base_min_length = recorded_survivor_regions();
597 // This is the absolute minimum young length, which ensures that we
598 // can allocate one eden region in the worst-case.
599 size_t absolute_min_length = base_min_length + 1;
600 size_t desired_min_length =
601 calculate_young_list_desired_min_length(base_min_length);
602 if (desired_min_length < absolute_min_length) {
603 desired_min_length = absolute_min_length;
604 }
606 // Calculate the absolute and desired max bounds.
608 // We will try our best not to "eat" into the reserve.
609 size_t absolute_max_length = 0;
610 if (_free_regions_at_end_of_collection > _reserve_regions) {
611 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
612 }
613 size_t desired_max_length = calculate_young_list_desired_max_length();
614 if (desired_max_length > absolute_max_length) {
615 desired_max_length = absolute_max_length;
616 }
618 size_t young_list_target_length = 0;
619 if (adaptive_young_list_length()) {
620 if (gcs_are_young()) {
621 young_list_target_length =
622 calculate_young_list_target_length(rs_lengths,
623 base_min_length,
624 desired_min_length,
625 desired_max_length);
626 _rs_lengths_prediction = rs_lengths;
627 } else {
628 // Don't calculate anything and let the code below bound it to
629 // the desired_min_length, i.e., do the next GC as soon as
630 // possible to maximize how many old regions we can add to it.
631 }
632 } else {
633 if (gcs_are_young()) {
634 young_list_target_length = _young_list_fixed_length;
635 } else {
636 // A bit arbitrary: during mixed GCs we allocate half
637 // the young regions to try to add old regions to the CSet.
638 young_list_target_length = _young_list_fixed_length / 2;
639 // We choose to accept that we might go under the desired min
640 // length given that we intentionally ask for a smaller young gen.
641 desired_min_length = absolute_min_length;
642 }
643 }
645 // Make sure we don't go over the desired max length, nor under the
646 // desired min length. In case they clash, desired_min_length wins
647 // which is why that test is second.
648 if (young_list_target_length > desired_max_length) {
649 young_list_target_length = desired_max_length;
650 }
651 if (young_list_target_length < desired_min_length) {
652 young_list_target_length = desired_min_length;
653 }
655 assert(young_list_target_length > recorded_survivor_regions(),
656 "we should be able to allocate at least one eden region");
657 assert(young_list_target_length >= absolute_min_length, "post-condition");
658 _young_list_target_length = young_list_target_length;
660 update_max_gc_locker_expansion();
661 }
663 size_t
664 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
665 size_t base_min_length,
666 size_t desired_min_length,
667 size_t desired_max_length) {
668 assert(adaptive_young_list_length(), "pre-condition");
669 assert(gcs_are_young(), "only call this for young GCs");
671 // In case some edge-condition makes the desired max length too small...
672 if (desired_max_length <= desired_min_length) {
673 return desired_min_length;
674 }
676 // We'll adjust min_young_length and max_young_length not to include
677 // the already allocated young regions (i.e., so they reflect the
678 // min and max eden regions we'll allocate). The base_min_length
679 // will be reflected in the predictions by the
680 // survivor_regions_evac_time prediction.
681 assert(desired_min_length > base_min_length, "invariant");
682 size_t min_young_length = desired_min_length - base_min_length;
683 assert(desired_max_length > base_min_length, "invariant");
684 size_t max_young_length = desired_max_length - base_min_length;
686 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
687 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
688 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
689 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
690 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
691 double base_time_ms =
692 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
693 survivor_regions_evac_time;
694 size_t available_free_regions = _free_regions_at_end_of_collection;
695 size_t base_free_regions = 0;
696 if (available_free_regions > _reserve_regions) {
697 base_free_regions = available_free_regions - _reserve_regions;
698 }
700 // Here, we will make sure that the shortest young length that
701 // makes sense fits within the target pause time.
703 if (predict_will_fit(min_young_length, base_time_ms,
704 base_free_regions, target_pause_time_ms)) {
705 // The shortest young length will fit into the target pause time;
706 // we'll now check whether the absolute maximum number of young
707 // regions will fit in the target pause time. If not, we'll do
708 // a binary search between min_young_length and max_young_length.
709 if (predict_will_fit(max_young_length, base_time_ms,
710 base_free_regions, target_pause_time_ms)) {
711 // The maximum young length will fit into the target pause time.
712 // We are done so set min young length to the maximum length (as
713 // the result is assumed to be returned in min_young_length).
714 min_young_length = max_young_length;
715 } else {
716 // The maximum possible number of young regions will not fit within
717 // the target pause time so we'll search for the optimal
718 // length. The loop invariants are:
719 //
720 // min_young_length < max_young_length
721 // min_young_length is known to fit into the target pause time
722 // max_young_length is known not to fit into the target pause time
723 //
724 // Going into the loop we know the above hold as we've just
725 // checked them. Every time around the loop we check whether
726 // the middle value between min_young_length and
727 // max_young_length fits into the target pause time. If it
728 // does, it becomes the new min. If it doesn't, it becomes
729 // the new max. This way we maintain the loop invariants.
731 assert(min_young_length < max_young_length, "invariant");
732 size_t diff = (max_young_length - min_young_length) / 2;
733 while (diff > 0) {
734 size_t young_length = min_young_length + diff;
735 if (predict_will_fit(young_length, base_time_ms,
736 base_free_regions, target_pause_time_ms)) {
737 min_young_length = young_length;
738 } else {
739 max_young_length = young_length;
740 }
741 assert(min_young_length < max_young_length, "invariant");
742 diff = (max_young_length - min_young_length) / 2;
743 }
744 // The results is min_young_length which, according to the
745 // loop invariants, should fit within the target pause time.
747 // These are the post-conditions of the binary search above:
748 assert(min_young_length < max_young_length,
749 "otherwise we should have discovered that max_young_length "
750 "fits into the pause target and not done the binary search");
751 assert(predict_will_fit(min_young_length, base_time_ms,
752 base_free_regions, target_pause_time_ms),
753 "min_young_length, the result of the binary search, should "
754 "fit into the pause target");
755 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
756 base_free_regions, target_pause_time_ms),
757 "min_young_length, the result of the binary search, should be "
758 "optimal, so no larger length should fit into the pause target");
759 }
760 } else {
761 // Even the minimum length doesn't fit into the pause time
762 // target, return it as the result nevertheless.
763 }
764 return base_min_length + min_young_length;
765 }
767 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
768 double survivor_regions_evac_time = 0.0;
769 for (HeapRegion * r = _recorded_survivor_head;
770 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
771 r = r->get_next_young_region()) {
772 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
773 }
774 return survivor_regions_evac_time;
775 }
777 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
778 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
780 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
781 if (rs_lengths > _rs_lengths_prediction) {
782 // add 10% to avoid having to recalculate often
783 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
784 update_young_list_target_length(rs_lengths_prediction);
785 }
786 }
790 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
791 bool is_tlab,
792 bool* gc_overhead_limit_was_exceeded) {
793 guarantee(false, "Not using this policy feature yet.");
794 return NULL;
795 }
797 // This method controls how a collector handles one or more
798 // of its generations being fully allocated.
799 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
800 bool is_tlab) {
801 guarantee(false, "Not using this policy feature yet.");
802 return NULL;
803 }
806 #ifndef PRODUCT
807 bool G1CollectorPolicy::verify_young_ages() {
808 HeapRegion* head = _g1->young_list()->first_region();
809 return
810 verify_young_ages(head, _short_lived_surv_rate_group);
811 // also call verify_young_ages on any additional surv rate groups
812 }
814 bool
815 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
816 SurvRateGroup *surv_rate_group) {
817 guarantee( surv_rate_group != NULL, "pre-condition" );
819 const char* name = surv_rate_group->name();
820 bool ret = true;
821 int prev_age = -1;
823 for (HeapRegion* curr = head;
824 curr != NULL;
825 curr = curr->get_next_young_region()) {
826 SurvRateGroup* group = curr->surv_rate_group();
827 if (group == NULL && !curr->is_survivor()) {
828 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
829 ret = false;
830 }
832 if (surv_rate_group == group) {
833 int age = curr->age_in_surv_rate_group();
835 if (age < 0) {
836 gclog_or_tty->print_cr("## %s: encountered negative age", name);
837 ret = false;
838 }
840 if (age <= prev_age) {
841 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
842 "(%d, %d)", name, age, prev_age);
843 ret = false;
844 }
845 prev_age = age;
846 }
847 }
849 return ret;
850 }
851 #endif // PRODUCT
853 void G1CollectorPolicy::record_full_collection_start() {
854 _cur_collection_start_sec = os::elapsedTime();
855 // Release the future to-space so that it is available for compaction into.
856 _g1->set_full_collection();
857 }
859 void G1CollectorPolicy::record_full_collection_end() {
860 // Consider this like a collection pause for the purposes of allocation
861 // since last pause.
862 double end_sec = os::elapsedTime();
863 double full_gc_time_sec = end_sec - _cur_collection_start_sec;
864 double full_gc_time_ms = full_gc_time_sec * 1000.0;
866 _all_full_gc_times_ms->add(full_gc_time_ms);
868 update_recent_gc_times(end_sec, full_gc_time_ms);
870 _g1->clear_full_collection();
872 // "Nuke" the heuristics that control the young/mixed GC
873 // transitions and make sure we start with young GCs after the Full GC.
874 set_gcs_are_young(true);
875 _last_young_gc = false;
876 _should_revert_to_young_gcs = false;
877 clear_initiate_conc_mark_if_possible();
878 clear_during_initial_mark_pause();
879 _known_garbage_bytes = 0;
880 _known_garbage_ratio = 0.0;
881 _in_marking_window = false;
882 _in_marking_window_im = false;
884 _short_lived_surv_rate_group->start_adding_regions();
885 // also call this on any additional surv rate groups
887 record_survivor_regions(0, NULL, NULL);
889 _free_regions_at_end_of_collection = _g1->free_regions();
890 // Reset survivors SurvRateGroup.
891 _survivor_surv_rate_group->reset();
892 update_young_list_target_length();
893 _collectionSetChooser->updateAfterFullCollection();
894 }
896 void G1CollectorPolicy::record_stop_world_start() {
897 _stop_world_start = os::elapsedTime();
898 }
900 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
901 size_t start_used) {
902 if (PrintGCDetails) {
903 gclog_or_tty->stamp(PrintGCTimeStamps);
904 gclog_or_tty->print("[GC pause");
905 gclog_or_tty->print(" (%s)", gcs_are_young() ? "young" : "mixed");
906 }
908 if (!during_initial_mark_pause()) {
909 // We only need to do this here as the policy will only be applied
910 // to the GC we're about to start. so, no point is calculating this
911 // every time we calculate / recalculate the target young length.
912 update_survivors_policy();
913 } else {
914 // The marking phase has a "we only copy implicitly live
915 // objects during marking" invariant. The easiest way to ensure it
916 // holds is not to allocate any survivor regions and tenure all
917 // objects. In the future we might change this and handle survivor
918 // regions specially during marking.
919 tenure_all_objects();
920 }
922 assert(_g1->used() == _g1->recalculate_used(),
923 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
924 _g1->used(), _g1->recalculate_used()));
926 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
927 _all_stop_world_times_ms->add(s_w_t_ms);
928 _stop_world_start = 0.0;
930 _cur_collection_start_sec = start_time_sec;
931 _cur_collection_pause_used_at_start_bytes = start_used;
932 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
933 _pending_cards = _g1->pending_card_num();
934 _max_pending_cards = _g1->max_pending_card_num();
936 _bytes_in_collection_set_before_gc = 0;
937 _bytes_copied_during_gc = 0;
939 YoungList* young_list = _g1->young_list();
940 _eden_bytes_before_gc = young_list->eden_used_bytes();
941 _survivor_bytes_before_gc = young_list->survivor_used_bytes();
942 _capacity_before_gc = _g1->capacity();
944 #ifdef DEBUG
945 // initialise these to something well known so that we can spot
946 // if they are not set properly
948 for (int i = 0; i < _parallel_gc_threads; ++i) {
949 _par_last_gc_worker_start_times_ms[i] = -1234.0;
950 _par_last_ext_root_scan_times_ms[i] = -1234.0;
951 _par_last_satb_filtering_times_ms[i] = -1234.0;
952 _par_last_update_rs_times_ms[i] = -1234.0;
953 _par_last_update_rs_processed_buffers[i] = -1234.0;
954 _par_last_scan_rs_times_ms[i] = -1234.0;
955 _par_last_obj_copy_times_ms[i] = -1234.0;
956 _par_last_termination_times_ms[i] = -1234.0;
957 _par_last_termination_attempts[i] = -1234.0;
958 _par_last_gc_worker_end_times_ms[i] = -1234.0;
959 _par_last_gc_worker_times_ms[i] = -1234.0;
960 _par_last_gc_worker_other_times_ms[i] = -1234.0;
961 }
962 #endif
964 for (int i = 0; i < _aux_num; ++i) {
965 _cur_aux_times_ms[i] = 0.0;
966 _cur_aux_times_set[i] = false;
967 }
969 // This is initialized to zero here and is set during
970 // the evacuation pause if marking is in progress.
971 _cur_satb_drain_time_ms = 0.0;
973 _last_gc_was_young = false;
975 // do that for any other surv rate groups
976 _short_lived_surv_rate_group->stop_adding_regions();
977 _survivors_age_table.clear();
979 assert( verify_young_ages(), "region age verification" );
980 }
982 void G1CollectorPolicy::record_concurrent_mark_init_end(double
983 mark_init_elapsed_time_ms) {
984 _during_marking = true;
985 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
986 clear_during_initial_mark_pause();
987 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
988 }
990 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
991 _mark_remark_start_sec = os::elapsedTime();
992 _during_marking = false;
993 }
995 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
996 double end_time_sec = os::elapsedTime();
997 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
998 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
999 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1000 _prev_collection_pause_end_ms += elapsed_time_ms;
1002 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
1003 }
1005 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
1006 _mark_cleanup_start_sec = os::elapsedTime();
1007 }
1009 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
1010 _should_revert_to_young_gcs = false;
1011 _last_young_gc = true;
1012 _in_marking_window = false;
1013 }
1015 void G1CollectorPolicy::record_concurrent_pause() {
1016 if (_stop_world_start > 0.0) {
1017 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
1018 _all_yield_times_ms->add(yield_ms);
1019 }
1020 }
1022 void G1CollectorPolicy::record_concurrent_pause_end() {
1023 }
1025 template<class T>
1026 T sum_of(T* sum_arr, int start, int n, int N) {
1027 T sum = (T)0;
1028 for (int i = 0; i < n; i++) {
1029 int j = (start + i) % N;
1030 sum += sum_arr[j];
1031 }
1032 return sum;
1033 }
1035 void G1CollectorPolicy::print_par_stats(int level,
1036 const char* str,
1037 double* data) {
1038 double min = data[0], max = data[0];
1039 double total = 0.0;
1040 LineBuffer buf(level);
1041 buf.append("[%s (ms):", str);
1042 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1043 double val = data[i];
1044 if (val < min)
1045 min = val;
1046 if (val > max)
1047 max = val;
1048 total += val;
1049 buf.append(" %3.1lf", val);
1050 }
1051 buf.append_and_print_cr("");
1052 double avg = total / (double) no_of_gc_threads();
1053 buf.append_and_print_cr(" Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf, Diff: %5.1lf]",
1054 avg, min, max, max - min);
1055 }
1057 void G1CollectorPolicy::print_par_sizes(int level,
1058 const char* str,
1059 double* data) {
1060 double min = data[0], max = data[0];
1061 double total = 0.0;
1062 LineBuffer buf(level);
1063 buf.append("[%s :", str);
1064 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1065 double val = data[i];
1066 if (val < min)
1067 min = val;
1068 if (val > max)
1069 max = val;
1070 total += val;
1071 buf.append(" %d", (int) val);
1072 }
1073 buf.append_and_print_cr("");
1074 double avg = total / (double) no_of_gc_threads();
1075 buf.append_and_print_cr(" Sum: %d, Avg: %d, Min: %d, Max: %d, Diff: %d]",
1076 (int)total, (int)avg, (int)min, (int)max, (int)max - (int)min);
1077 }
1079 void G1CollectorPolicy::print_stats(int level,
1080 const char* str,
1081 double value) {
1082 LineBuffer(level).append_and_print_cr("[%s: %5.1lf ms]", str, value);
1083 }
1085 void G1CollectorPolicy::print_stats(int level,
1086 const char* str,
1087 int value) {
1088 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value);
1089 }
1091 double G1CollectorPolicy::avg_value(double* data) {
1092 if (G1CollectedHeap::use_parallel_gc_threads()) {
1093 double ret = 0.0;
1094 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1095 ret += data[i];
1096 }
1097 return ret / (double) no_of_gc_threads();
1098 } else {
1099 return data[0];
1100 }
1101 }
1103 double G1CollectorPolicy::max_value(double* data) {
1104 if (G1CollectedHeap::use_parallel_gc_threads()) {
1105 double ret = data[0];
1106 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1107 if (data[i] > ret) {
1108 ret = data[i];
1109 }
1110 }
1111 return ret;
1112 } else {
1113 return data[0];
1114 }
1115 }
1117 double G1CollectorPolicy::sum_of_values(double* data) {
1118 if (G1CollectedHeap::use_parallel_gc_threads()) {
1119 double sum = 0.0;
1120 for (uint i = 0; i < no_of_gc_threads(); i++) {
1121 sum += data[i];
1122 }
1123 return sum;
1124 } else {
1125 return data[0];
1126 }
1127 }
1129 double G1CollectorPolicy::max_sum(double* data1, double* data2) {
1130 double ret = data1[0] + data2[0];
1132 if (G1CollectedHeap::use_parallel_gc_threads()) {
1133 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1134 double data = data1[i] + data2[i];
1135 if (data > ret) {
1136 ret = data;
1137 }
1138 }
1139 }
1140 return ret;
1141 }
1143 // Anything below that is considered to be zero
1144 #define MIN_TIMER_GRANULARITY 0.0000001
1146 void G1CollectorPolicy::record_collection_pause_end(int no_of_gc_threads) {
1147 double end_time_sec = os::elapsedTime();
1148 double elapsed_ms = _last_pause_time_ms;
1149 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1150 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(),
1151 "otherwise, the subtraction below does not make sense");
1152 size_t rs_size =
1153 _cur_collection_pause_used_regions_at_start - cset_region_length();
1154 size_t cur_used_bytes = _g1->used();
1155 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1156 bool last_pause_included_initial_mark = false;
1157 bool update_stats = !_g1->evacuation_failed();
1158 set_no_of_gc_threads(no_of_gc_threads);
1160 #ifndef PRODUCT
1161 if (G1YoungSurvRateVerbose) {
1162 gclog_or_tty->print_cr("");
1163 _short_lived_surv_rate_group->print();
1164 // do that for any other surv rate groups too
1165 }
1166 #endif // PRODUCT
1168 last_pause_included_initial_mark = during_initial_mark_pause();
1169 if (last_pause_included_initial_mark)
1170 record_concurrent_mark_init_end(0.0);
1172 size_t marking_initiating_used_threshold =
1173 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
1175 if (!_g1->mark_in_progress() && !_last_young_gc) {
1176 assert(!last_pause_included_initial_mark, "invariant");
1177 if (cur_used_bytes > marking_initiating_used_threshold) {
1178 if (cur_used_bytes > _prev_collection_pause_used_at_end_bytes) {
1179 assert(!during_initial_mark_pause(), "we should not see this here");
1181 ergo_verbose3(ErgoConcCycles,
1182 "request concurrent cycle initiation",
1183 ergo_format_reason("occupancy higher than threshold")
1184 ergo_format_byte("occupancy")
1185 ergo_format_byte_perc("threshold"),
1186 cur_used_bytes,
1187 marking_initiating_used_threshold,
1188 (double) InitiatingHeapOccupancyPercent);
1190 // Note: this might have already been set, if during the last
1191 // pause we decided to start a cycle but at the beginning of
1192 // this pause we decided to postpone it. That's OK.
1193 set_initiate_conc_mark_if_possible();
1194 } else {
1195 ergo_verbose2(ErgoConcCycles,
1196 "do not request concurrent cycle initiation",
1197 ergo_format_reason("occupancy lower than previous occupancy")
1198 ergo_format_byte("occupancy")
1199 ergo_format_byte("previous occupancy"),
1200 cur_used_bytes,
1201 _prev_collection_pause_used_at_end_bytes);
1202 }
1203 }
1204 }
1206 _prev_collection_pause_used_at_end_bytes = cur_used_bytes;
1208 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1209 end_time_sec, false);
1211 // This assert is exempted when we're doing parallel collection pauses,
1212 // because the fragmentation caused by the parallel GC allocation buffers
1213 // can lead to more memory being used during collection than was used
1214 // before. Best leave this out until the fragmentation problem is fixed.
1215 // Pauses in which evacuation failed can also lead to negative
1216 // collections, since no space is reclaimed from a region containing an
1217 // object whose evacuation failed.
1218 // Further, we're now always doing parallel collection. But I'm still
1219 // leaving this here as a placeholder for a more precise assertion later.
1220 // (DLD, 10/05.)
1221 assert((true || parallel) // Always using GC LABs now.
1222 || _g1->evacuation_failed()
1223 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1224 "Negative collection");
1226 size_t freed_bytes =
1227 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1228 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1230 double survival_fraction =
1231 (double)surviving_bytes/
1232 (double)_collection_set_bytes_used_before;
1234 // These values are used to update the summary information that is
1235 // displayed when TraceGen0Time is enabled, and are output as part
1236 // of the PrintGCDetails output, in the non-parallel case.
1238 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1239 double satb_filtering_time = avg_value(_par_last_satb_filtering_times_ms);
1240 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1241 double update_rs_processed_buffers =
1242 sum_of_values(_par_last_update_rs_processed_buffers);
1243 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1244 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1245 double termination_time = avg_value(_par_last_termination_times_ms);
1247 double known_time = ext_root_scan_time +
1248 satb_filtering_time +
1249 update_rs_time +
1250 scan_rs_time +
1251 obj_copy_time;
1253 double other_time_ms = elapsed_ms;
1255 // Subtract the SATB drain time. It's initialized to zero at the
1256 // start of the pause and is updated during the pause if marking
1257 // is in progress.
1258 other_time_ms -= _cur_satb_drain_time_ms;
1260 if (parallel) {
1261 other_time_ms -= _cur_collection_par_time_ms;
1262 } else {
1263 other_time_ms -= known_time;
1264 }
1266 // Subtract the time taken to clean the card table from the
1267 // current value of "other time"
1268 other_time_ms -= _cur_clear_ct_time_ms;
1270 // Subtract the time spent completing marking in the collection
1271 // set. Note if marking is not in progress during the pause
1272 // the value of _mark_closure_time_ms will be zero.
1273 other_time_ms -= _mark_closure_time_ms;
1275 // TraceGen0Time and TraceGen1Time summary info updating.
1276 _all_pause_times_ms->add(elapsed_ms);
1278 if (update_stats) {
1279 _summary->record_total_time_ms(elapsed_ms);
1280 _summary->record_other_time_ms(other_time_ms);
1282 MainBodySummary* body_summary = _summary->main_body_summary();
1283 assert(body_summary != NULL, "should not be null!");
1285 // This will be non-zero iff marking is currently in progress (i.e.
1286 // _g1->mark_in_progress() == true) and the currrent pause was not
1287 // an initial mark pause. Since the body_summary items are NumberSeqs,
1288 // however, they have to be consistent and updated in lock-step with
1289 // each other. Therefore we unconditionally record the SATB drain
1290 // time - even if it's zero.
1291 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms);
1293 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1294 body_summary->record_satb_filtering_time_ms(satb_filtering_time);
1295 body_summary->record_update_rs_time_ms(update_rs_time);
1296 body_summary->record_scan_rs_time_ms(scan_rs_time);
1297 body_summary->record_obj_copy_time_ms(obj_copy_time);
1299 if (parallel) {
1300 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1301 body_summary->record_termination_time_ms(termination_time);
1303 double parallel_known_time = known_time + termination_time;
1304 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time;
1305 body_summary->record_parallel_other_time_ms(parallel_other_time);
1306 }
1308 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms);
1309 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1311 // We exempt parallel collection from this check because Alloc Buffer
1312 // fragmentation can produce negative collections. Same with evac
1313 // failure.
1314 // Further, we're now always doing parallel collection. But I'm still
1315 // leaving this here as a placeholder for a more precise assertion later.
1316 // (DLD, 10/05.
1317 assert((true || parallel)
1318 || _g1->evacuation_failed()
1319 || surviving_bytes <= _collection_set_bytes_used_before,
1320 "Or else negative collection!");
1322 // this is where we update the allocation rate of the application
1323 double app_time_ms =
1324 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1325 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1326 // This usually happens due to the timer not having the required
1327 // granularity. Some Linuxes are the usual culprits.
1328 // We'll just set it to something (arbitrarily) small.
1329 app_time_ms = 1.0;
1330 }
1331 // We maintain the invariant that all objects allocated by mutator
1332 // threads will be allocated out of eden regions. So, we can use
1333 // the eden region number allocated since the previous GC to
1334 // calculate the application's allocate rate. The only exception
1335 // to that is humongous objects that are allocated separately. But
1336 // given that humongous object allocations do not really affect
1337 // either the pause's duration nor when the next pause will take
1338 // place we can safely ignore them here.
1339 size_t regions_allocated = eden_cset_region_length();
1340 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1341 _alloc_rate_ms_seq->add(alloc_rate_ms);
1343 double interval_ms =
1344 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1345 update_recent_gc_times(end_time_sec, elapsed_ms);
1346 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1347 if (recent_avg_pause_time_ratio() < 0.0 ||
1348 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1349 #ifndef PRODUCT
1350 // Dump info to allow post-facto debugging
1351 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1352 gclog_or_tty->print_cr("-------------------------------------------");
1353 gclog_or_tty->print_cr("Recent GC Times (ms):");
1354 _recent_gc_times_ms->dump();
1355 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1356 _recent_prev_end_times_for_all_gcs_sec->dump();
1357 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1358 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1359 // In debug mode, terminate the JVM if the user wants to debug at this point.
1360 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1361 #endif // !PRODUCT
1362 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1363 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1364 if (_recent_avg_pause_time_ratio < 0.0) {
1365 _recent_avg_pause_time_ratio = 0.0;
1366 } else {
1367 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1368 _recent_avg_pause_time_ratio = 1.0;
1369 }
1370 }
1371 }
1373 for (int i = 0; i < _aux_num; ++i) {
1374 if (_cur_aux_times_set[i]) {
1375 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1376 }
1377 }
1379 // PrintGCDetails output
1380 if (PrintGCDetails) {
1381 bool print_marking_info =
1382 _g1->mark_in_progress() && !last_pause_included_initial_mark;
1384 gclog_or_tty->print_cr("%s, %1.8lf secs]",
1385 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1386 elapsed_ms / 1000.0);
1388 if (parallel) {
1389 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1390 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms);
1391 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1392 if (print_marking_info) {
1393 print_par_stats(2, "SATB Filtering", _par_last_satb_filtering_times_ms);
1394 }
1395 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1396 print_par_sizes(3, "Processed Buffers", _par_last_update_rs_processed_buffers);
1397 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1398 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1399 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1400 print_par_sizes(3, "Termination Attempts", _par_last_termination_attempts);
1401 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms);
1403 for (int i = 0; i < _parallel_gc_threads; i++) {
1404 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] - _par_last_gc_worker_start_times_ms[i];
1406 double worker_known_time = _par_last_ext_root_scan_times_ms[i] +
1407 _par_last_satb_filtering_times_ms[i] +
1408 _par_last_update_rs_times_ms[i] +
1409 _par_last_scan_rs_times_ms[i] +
1410 _par_last_obj_copy_times_ms[i] +
1411 _par_last_termination_times_ms[i];
1413 _par_last_gc_worker_other_times_ms[i] = _cur_collection_par_time_ms - worker_known_time;
1414 }
1415 print_par_stats(2, "GC Worker", _par_last_gc_worker_times_ms);
1416 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms);
1417 } else {
1418 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1419 if (print_marking_info) {
1420 print_stats(1, "SATB Filtering", satb_filtering_time);
1421 }
1422 print_stats(1, "Update RS", update_rs_time);
1423 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers);
1424 print_stats(1, "Scan RS", scan_rs_time);
1425 print_stats(1, "Object Copying", obj_copy_time);
1426 }
1427 if (print_marking_info) {
1428 print_stats(1, "Complete CSet Marking", _mark_closure_time_ms);
1429 }
1430 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1431 #ifndef PRODUCT
1432 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1433 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1434 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1435 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1436 if (_num_cc_clears > 0) {
1437 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1438 }
1439 #endif
1440 print_stats(1, "Other", other_time_ms);
1441 print_stats(2, "Choose CSet",
1442 (_recorded_young_cset_choice_time_ms +
1443 _recorded_non_young_cset_choice_time_ms));
1444 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms);
1445 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms);
1446 print_stats(2, "Free CSet",
1447 (_recorded_young_free_cset_time_ms +
1448 _recorded_non_young_free_cset_time_ms));
1450 for (int i = 0; i < _aux_num; ++i) {
1451 if (_cur_aux_times_set[i]) {
1452 char buffer[96];
1453 sprintf(buffer, "Aux%d", i);
1454 print_stats(1, buffer, _cur_aux_times_ms[i]);
1455 }
1456 }
1457 }
1459 // Update the efficiency-since-mark vars.
1460 double proc_ms = elapsed_ms * (double) _parallel_gc_threads;
1461 if (elapsed_ms < MIN_TIMER_GRANULARITY) {
1462 // This usually happens due to the timer not having the required
1463 // granularity. Some Linuxes are the usual culprits.
1464 // We'll just set it to something (arbitrarily) small.
1465 proc_ms = 1.0;
1466 }
1467 double cur_efficiency = (double) freed_bytes / proc_ms;
1469 bool new_in_marking_window = _in_marking_window;
1470 bool new_in_marking_window_im = false;
1471 if (during_initial_mark_pause()) {
1472 new_in_marking_window = true;
1473 new_in_marking_window_im = true;
1474 }
1476 if (_last_young_gc) {
1477 if (!last_pause_included_initial_mark) {
1478 ergo_verbose2(ErgoMixedGCs,
1479 "start mixed GCs",
1480 ergo_format_byte_perc("known garbage"),
1481 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1482 set_gcs_are_young(false);
1483 } else {
1484 ergo_verbose0(ErgoMixedGCs,
1485 "do not start mixed GCs",
1486 ergo_format_reason("concurrent cycle is about to start"));
1487 }
1488 _last_young_gc = false;
1489 }
1491 if (!_last_gc_was_young) {
1492 if (_should_revert_to_young_gcs) {
1493 ergo_verbose2(ErgoMixedGCs,
1494 "end mixed GCs",
1495 ergo_format_reason("mixed GCs end requested")
1496 ergo_format_byte_perc("known garbage"),
1497 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1498 set_gcs_are_young(true);
1499 } else if (_known_garbage_ratio < 0.05) {
1500 ergo_verbose3(ErgoMixedGCs,
1501 "end mixed GCs",
1502 ergo_format_reason("known garbage percent lower than threshold")
1503 ergo_format_byte_perc("known garbage")
1504 ergo_format_perc("threshold"),
1505 _known_garbage_bytes, _known_garbage_ratio * 100.0,
1506 0.05 * 100.0);
1507 set_gcs_are_young(true);
1508 } else if (adaptive_young_list_length() &&
1509 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) {
1510 ergo_verbose5(ErgoMixedGCs,
1511 "end mixed GCs",
1512 ergo_format_reason("current GC efficiency lower than "
1513 "predicted young GC efficiency")
1514 ergo_format_double("GC efficiency factor")
1515 ergo_format_double("current GC efficiency")
1516 ergo_format_double("predicted young GC efficiency")
1517 ergo_format_byte_perc("known garbage"),
1518 get_gc_eff_factor(), cur_efficiency,
1519 predict_young_gc_eff(),
1520 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1521 set_gcs_are_young(true);
1522 }
1523 }
1524 _should_revert_to_young_gcs = false;
1526 if (_last_gc_was_young && !_during_marking) {
1527 _young_gc_eff_seq->add(cur_efficiency);
1528 }
1530 _short_lived_surv_rate_group->start_adding_regions();
1531 // do that for any other surv rate groupsx
1533 if (update_stats) {
1534 double pause_time_ms = elapsed_ms;
1536 size_t diff = 0;
1537 if (_max_pending_cards >= _pending_cards)
1538 diff = _max_pending_cards - _pending_cards;
1539 _pending_card_diff_seq->add((double) diff);
1541 double cost_per_card_ms = 0.0;
1542 if (_pending_cards > 0) {
1543 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1544 _cost_per_card_ms_seq->add(cost_per_card_ms);
1545 }
1547 size_t cards_scanned = _g1->cards_scanned();
1549 double cost_per_entry_ms = 0.0;
1550 if (cards_scanned > 10) {
1551 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1552 if (_last_gc_was_young) {
1553 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1554 } else {
1555 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1556 }
1557 }
1559 if (_max_rs_lengths > 0) {
1560 double cards_per_entry_ratio =
1561 (double) cards_scanned / (double) _max_rs_lengths;
1562 if (_last_gc_was_young) {
1563 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1564 } else {
1565 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1566 }
1567 }
1569 // This is defensive. For a while _max_rs_lengths could get
1570 // smaller than _recorded_rs_lengths which was causing
1571 // rs_length_diff to get very large and mess up the RSet length
1572 // predictions. The reason was unsafe concurrent updates to the
1573 // _inc_cset_recorded_rs_lengths field which the code below guards
1574 // against (see CR 7118202). This bug has now been fixed (see CR
1575 // 7119027). However, I'm still worried that
1576 // _inc_cset_recorded_rs_lengths might still end up somewhat
1577 // inaccurate. The concurrent refinement thread calculates an
1578 // RSet's length concurrently with other CR threads updating it
1579 // which might cause it to calculate the length incorrectly (if,
1580 // say, it's in mid-coarsening). So I'll leave in the defensive
1581 // conditional below just in case.
1582 size_t rs_length_diff = 0;
1583 if (_max_rs_lengths > _recorded_rs_lengths) {
1584 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1585 }
1586 _rs_length_diff_seq->add((double) rs_length_diff);
1588 size_t copied_bytes = surviving_bytes;
1589 double cost_per_byte_ms = 0.0;
1590 if (copied_bytes > 0) {
1591 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1592 if (_in_marking_window) {
1593 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1594 } else {
1595 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1596 }
1597 }
1599 double all_other_time_ms = pause_time_ms -
1600 (update_rs_time + scan_rs_time + obj_copy_time +
1601 _mark_closure_time_ms + termination_time);
1603 double young_other_time_ms = 0.0;
1604 if (young_cset_region_length() > 0) {
1605 young_other_time_ms =
1606 _recorded_young_cset_choice_time_ms +
1607 _recorded_young_free_cset_time_ms;
1608 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1609 (double) young_cset_region_length());
1610 }
1611 double non_young_other_time_ms = 0.0;
1612 if (old_cset_region_length() > 0) {
1613 non_young_other_time_ms =
1614 _recorded_non_young_cset_choice_time_ms +
1615 _recorded_non_young_free_cset_time_ms;
1617 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1618 (double) old_cset_region_length());
1619 }
1621 double constant_other_time_ms = all_other_time_ms -
1622 (young_other_time_ms + non_young_other_time_ms);
1623 _constant_other_time_ms_seq->add(constant_other_time_ms);
1625 double survival_ratio = 0.0;
1626 if (_bytes_in_collection_set_before_gc > 0) {
1627 survival_ratio = (double) _bytes_copied_during_gc /
1628 (double) _bytes_in_collection_set_before_gc;
1629 }
1631 _pending_cards_seq->add((double) _pending_cards);
1632 _rs_lengths_seq->add((double) _max_rs_lengths);
1634 double expensive_region_limit_ms =
1635 (double) MaxGCPauseMillis - predict_constant_other_time_ms();
1636 if (expensive_region_limit_ms < 0.0) {
1637 // this means that the other time was predicted to be longer than
1638 // than the max pause time
1639 expensive_region_limit_ms = (double) MaxGCPauseMillis;
1640 }
1641 _expensive_region_limit_ms = expensive_region_limit_ms;
1642 }
1644 _in_marking_window = new_in_marking_window;
1645 _in_marking_window_im = new_in_marking_window_im;
1646 _free_regions_at_end_of_collection = _g1->free_regions();
1647 update_young_list_target_length();
1649 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1650 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1651 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1653 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end.");
1654 }
1656 #define EXT_SIZE_FORMAT "%d%s"
1657 #define EXT_SIZE_PARAMS(bytes) \
1658 byte_size_in_proper_unit((bytes)), \
1659 proper_unit_for_byte_size((bytes))
1661 void G1CollectorPolicy::print_heap_transition() {
1662 if (PrintGCDetails) {
1663 YoungList* young_list = _g1->young_list();
1664 size_t eden_bytes = young_list->eden_used_bytes();
1665 size_t survivor_bytes = young_list->survivor_used_bytes();
1666 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes;
1667 size_t used = _g1->used();
1668 size_t capacity = _g1->capacity();
1669 size_t eden_capacity =
1670 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes;
1672 gclog_or_tty->print_cr(
1673 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1674 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1675 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1676 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1677 EXT_SIZE_PARAMS(_eden_bytes_before_gc),
1678 EXT_SIZE_PARAMS(_prev_eden_capacity),
1679 EXT_SIZE_PARAMS(eden_bytes),
1680 EXT_SIZE_PARAMS(eden_capacity),
1681 EXT_SIZE_PARAMS(_survivor_bytes_before_gc),
1682 EXT_SIZE_PARAMS(survivor_bytes),
1683 EXT_SIZE_PARAMS(used_before_gc),
1684 EXT_SIZE_PARAMS(_capacity_before_gc),
1685 EXT_SIZE_PARAMS(used),
1686 EXT_SIZE_PARAMS(capacity));
1688 _prev_eden_capacity = eden_capacity;
1689 } else if (PrintGC) {
1690 _g1->print_size_transition(gclog_or_tty,
1691 _cur_collection_pause_used_at_start_bytes,
1692 _g1->used(), _g1->capacity());
1693 }
1694 }
1696 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1697 double update_rs_processed_buffers,
1698 double goal_ms) {
1699 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1700 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1702 if (G1UseAdaptiveConcRefinement) {
1703 const int k_gy = 3, k_gr = 6;
1704 const double inc_k = 1.1, dec_k = 0.9;
1706 int g = cg1r->green_zone();
1707 if (update_rs_time > goal_ms) {
1708 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1709 } else {
1710 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1711 g = (int)MAX2(g * inc_k, g + 1.0);
1712 }
1713 }
1714 // Change the refinement threads params
1715 cg1r->set_green_zone(g);
1716 cg1r->set_yellow_zone(g * k_gy);
1717 cg1r->set_red_zone(g * k_gr);
1718 cg1r->reinitialize_threads();
1720 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1721 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1722 cg1r->yellow_zone());
1723 // Change the barrier params
1724 dcqs.set_process_completed_threshold(processing_threshold);
1725 dcqs.set_max_completed_queue(cg1r->red_zone());
1726 }
1728 int curr_queue_size = dcqs.completed_buffers_num();
1729 if (curr_queue_size >= cg1r->yellow_zone()) {
1730 dcqs.set_completed_queue_padding(curr_queue_size);
1731 } else {
1732 dcqs.set_completed_queue_padding(0);
1733 }
1734 dcqs.notify_if_necessary();
1735 }
1737 double
1738 G1CollectorPolicy::
1739 predict_young_collection_elapsed_time_ms(size_t adjustment) {
1740 guarantee( adjustment == 0 || adjustment == 1, "invariant" );
1742 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1743 size_t young_num = g1h->young_list()->length();
1744 if (young_num == 0)
1745 return 0.0;
1747 young_num += adjustment;
1748 size_t pending_cards = predict_pending_cards();
1749 size_t rs_lengths = g1h->young_list()->sampled_rs_lengths() +
1750 predict_rs_length_diff();
1751 size_t card_num;
1752 if (gcs_are_young()) {
1753 card_num = predict_young_card_num(rs_lengths);
1754 } else {
1755 card_num = predict_non_young_card_num(rs_lengths);
1756 }
1757 size_t young_byte_size = young_num * HeapRegion::GrainBytes;
1758 double accum_yg_surv_rate =
1759 _short_lived_surv_rate_group->accum_surv_rate(adjustment);
1761 size_t bytes_to_copy =
1762 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes);
1764 return
1765 predict_rs_update_time_ms(pending_cards) +
1766 predict_rs_scan_time_ms(card_num) +
1767 predict_object_copy_time_ms(bytes_to_copy) +
1768 predict_young_other_time_ms(young_num) +
1769 predict_constant_other_time_ms();
1770 }
1772 double
1773 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1774 size_t rs_length = predict_rs_length_diff();
1775 size_t card_num;
1776 if (gcs_are_young()) {
1777 card_num = predict_young_card_num(rs_length);
1778 } else {
1779 card_num = predict_non_young_card_num(rs_length);
1780 }
1781 return predict_base_elapsed_time_ms(pending_cards, card_num);
1782 }
1784 double
1785 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1786 size_t scanned_cards) {
1787 return
1788 predict_rs_update_time_ms(pending_cards) +
1789 predict_rs_scan_time_ms(scanned_cards) +
1790 predict_constant_other_time_ms();
1791 }
1793 double
1794 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1795 bool young) {
1796 size_t rs_length = hr->rem_set()->occupied();
1797 size_t card_num;
1798 if (gcs_are_young()) {
1799 card_num = predict_young_card_num(rs_length);
1800 } else {
1801 card_num = predict_non_young_card_num(rs_length);
1802 }
1803 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1805 double region_elapsed_time_ms =
1806 predict_rs_scan_time_ms(card_num) +
1807 predict_object_copy_time_ms(bytes_to_copy);
1809 if (young)
1810 region_elapsed_time_ms += predict_young_other_time_ms(1);
1811 else
1812 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1814 return region_elapsed_time_ms;
1815 }
1817 size_t
1818 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1819 size_t bytes_to_copy;
1820 if (hr->is_marked())
1821 bytes_to_copy = hr->max_live_bytes();
1822 else {
1823 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1,
1824 "invariant" );
1825 int age = hr->age_in_surv_rate_group();
1826 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1827 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1828 }
1830 return bytes_to_copy;
1831 }
1833 void
1834 G1CollectorPolicy::init_cset_region_lengths(size_t eden_cset_region_length,
1835 size_t survivor_cset_region_length) {
1836 _eden_cset_region_length = eden_cset_region_length;
1837 _survivor_cset_region_length = survivor_cset_region_length;
1838 _old_cset_region_length = 0;
1839 }
1841 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1842 _recorded_rs_lengths = rs_lengths;
1843 }
1845 void G1CollectorPolicy::check_if_region_is_too_expensive(double
1846 predicted_time_ms) {
1847 // I don't think we need to do this when in young GC mode since
1848 // marking will be initiated next time we hit the soft limit anyway...
1849 if (predicted_time_ms > _expensive_region_limit_ms) {
1850 ergo_verbose2(ErgoMixedGCs,
1851 "request mixed GCs end",
1852 ergo_format_reason("predicted region time higher than threshold")
1853 ergo_format_ms("predicted region time")
1854 ergo_format_ms("threshold"),
1855 predicted_time_ms, _expensive_region_limit_ms);
1856 // no point in doing another mixed GC
1857 _should_revert_to_young_gcs = true;
1858 }
1859 }
1861 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1862 double elapsed_ms) {
1863 _recent_gc_times_ms->add(elapsed_ms);
1864 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1865 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1866 }
1868 size_t G1CollectorPolicy::expansion_amount() {
1869 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1870 double threshold = _gc_overhead_perc;
1871 if (recent_gc_overhead > threshold) {
1872 // We will double the existing space, or take
1873 // G1ExpandByPercentOfAvailable % of the available expansion
1874 // space, whichever is smaller, bounded below by a minimum
1875 // expansion (unless that's all that's left.)
1876 const size_t min_expand_bytes = 1*M;
1877 size_t reserved_bytes = _g1->max_capacity();
1878 size_t committed_bytes = _g1->capacity();
1879 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1880 size_t expand_bytes;
1881 size_t expand_bytes_via_pct =
1882 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1883 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1884 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1885 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1887 ergo_verbose5(ErgoHeapSizing,
1888 "attempt heap expansion",
1889 ergo_format_reason("recent GC overhead higher than "
1890 "threshold after GC")
1891 ergo_format_perc("recent GC overhead")
1892 ergo_format_perc("threshold")
1893 ergo_format_byte("uncommitted")
1894 ergo_format_byte_perc("calculated expansion amount"),
1895 recent_gc_overhead, threshold,
1896 uncommitted_bytes,
1897 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
1899 return expand_bytes;
1900 } else {
1901 return 0;
1902 }
1903 }
1905 class CountCSClosure: public HeapRegionClosure {
1906 G1CollectorPolicy* _g1_policy;
1907 public:
1908 CountCSClosure(G1CollectorPolicy* g1_policy) :
1909 _g1_policy(g1_policy) {}
1910 bool doHeapRegion(HeapRegion* r) {
1911 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
1912 return false;
1913 }
1914 };
1916 void G1CollectorPolicy::count_CS_bytes_used() {
1917 CountCSClosure cs_closure(this);
1918 _g1->collection_set_iterate(&cs_closure);
1919 }
1921 void G1CollectorPolicy::print_summary(int level,
1922 const char* str,
1923 NumberSeq* seq) const {
1924 double sum = seq->sum();
1925 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
1926 str, sum / 1000.0, seq->avg());
1927 }
1929 void G1CollectorPolicy::print_summary_sd(int level,
1930 const char* str,
1931 NumberSeq* seq) const {
1932 print_summary(level, str, seq);
1933 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
1934 seq->num(), seq->sd(), seq->maximum());
1935 }
1937 void G1CollectorPolicy::check_other_times(int level,
1938 NumberSeq* other_times_ms,
1939 NumberSeq* calc_other_times_ms) const {
1940 bool should_print = false;
1941 LineBuffer buf(level + 2);
1943 double max_sum = MAX2(fabs(other_times_ms->sum()),
1944 fabs(calc_other_times_ms->sum()));
1945 double min_sum = MIN2(fabs(other_times_ms->sum()),
1946 fabs(calc_other_times_ms->sum()));
1947 double sum_ratio = max_sum / min_sum;
1948 if (sum_ratio > 1.1) {
1949 should_print = true;
1950 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
1951 }
1953 double max_avg = MAX2(fabs(other_times_ms->avg()),
1954 fabs(calc_other_times_ms->avg()));
1955 double min_avg = MIN2(fabs(other_times_ms->avg()),
1956 fabs(calc_other_times_ms->avg()));
1957 double avg_ratio = max_avg / min_avg;
1958 if (avg_ratio > 1.1) {
1959 should_print = true;
1960 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
1961 }
1963 if (other_times_ms->sum() < -0.01) {
1964 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
1965 }
1967 if (other_times_ms->avg() < -0.01) {
1968 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
1969 }
1971 if (calc_other_times_ms->sum() < -0.01) {
1972 should_print = true;
1973 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
1974 }
1976 if (calc_other_times_ms->avg() < -0.01) {
1977 should_print = true;
1978 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
1979 }
1981 if (should_print)
1982 print_summary(level, "Other(Calc)", calc_other_times_ms);
1983 }
1985 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
1986 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1987 MainBodySummary* body_summary = summary->main_body_summary();
1988 if (summary->get_total_seq()->num() > 0) {
1989 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
1990 if (body_summary != NULL) {
1991 if (parallel) {
1992 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
1993 print_summary(2, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
1994 print_summary(2, "SATB Filtering", body_summary->get_satb_filtering_seq());
1995 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
1996 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
1997 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
1998 print_summary(2, "Termination", body_summary->get_termination_seq());
1999 print_summary(2, "Parallel Other", body_summary->get_parallel_other_seq());
2000 {
2001 NumberSeq* other_parts[] = {
2002 body_summary->get_ext_root_scan_seq(),
2003 body_summary->get_satb_filtering_seq(),
2004 body_summary->get_update_rs_seq(),
2005 body_summary->get_scan_rs_seq(),
2006 body_summary->get_obj_copy_seq(),
2007 body_summary->get_termination_seq()
2008 };
2009 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
2010 6, other_parts);
2011 check_other_times(2, body_summary->get_parallel_other_seq(),
2012 &calc_other_times_ms);
2013 }
2014 } else {
2015 print_summary(1, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
2016 print_summary(1, "SATB Filtering", body_summary->get_satb_filtering_seq());
2017 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
2018 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
2019 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
2020 }
2021 }
2022 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq());
2023 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
2024 print_summary(1, "Other", summary->get_other_seq());
2025 {
2026 if (body_summary != NULL) {
2027 NumberSeq calc_other_times_ms;
2028 if (parallel) {
2029 // parallel
2030 NumberSeq* other_parts[] = {
2031 body_summary->get_satb_drain_seq(),
2032 body_summary->get_parallel_seq(),
2033 body_summary->get_clear_ct_seq()
2034 };
2035 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2036 3, other_parts);
2037 } else {
2038 // serial
2039 NumberSeq* other_parts[] = {
2040 body_summary->get_satb_drain_seq(),
2041 body_summary->get_update_rs_seq(),
2042 body_summary->get_ext_root_scan_seq(),
2043 body_summary->get_satb_filtering_seq(),
2044 body_summary->get_scan_rs_seq(),
2045 body_summary->get_obj_copy_seq()
2046 };
2047 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2048 6, other_parts);
2049 }
2050 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
2051 }
2052 }
2053 } else {
2054 LineBuffer(1).append_and_print_cr("none");
2055 }
2056 LineBuffer(0).append_and_print_cr("");
2057 }
2059 void G1CollectorPolicy::print_tracing_info() const {
2060 if (TraceGen0Time) {
2061 gclog_or_tty->print_cr("ALL PAUSES");
2062 print_summary_sd(0, "Total", _all_pause_times_ms);
2063 gclog_or_tty->print_cr("");
2064 gclog_or_tty->print_cr("");
2065 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num);
2066 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num);
2067 gclog_or_tty->print_cr("");
2069 gclog_or_tty->print_cr("EVACUATION PAUSES");
2070 print_summary(_summary);
2072 gclog_or_tty->print_cr("MISC");
2073 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
2074 print_summary_sd(0, "Yields", _all_yield_times_ms);
2075 for (int i = 0; i < _aux_num; ++i) {
2076 if (_all_aux_times_ms[i].num() > 0) {
2077 char buffer[96];
2078 sprintf(buffer, "Aux%d", i);
2079 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
2080 }
2081 }
2082 }
2083 if (TraceGen1Time) {
2084 if (_all_full_gc_times_ms->num() > 0) {
2085 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2086 _all_full_gc_times_ms->num(),
2087 _all_full_gc_times_ms->sum() / 1000.0);
2088 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
2089 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2090 _all_full_gc_times_ms->sd(),
2091 _all_full_gc_times_ms->maximum());
2092 }
2093 }
2094 }
2096 void G1CollectorPolicy::print_yg_surv_rate_info() const {
2097 #ifndef PRODUCT
2098 _short_lived_surv_rate_group->print_surv_rate_summary();
2099 // add this call for any other surv rate groups
2100 #endif // PRODUCT
2101 }
2103 #ifndef PRODUCT
2104 // for debugging, bit of a hack...
2105 static char*
2106 region_num_to_mbs(int length) {
2107 static char buffer[64];
2108 double bytes = (double) (length * HeapRegion::GrainBytes);
2109 double mbs = bytes / (double) (1024 * 1024);
2110 sprintf(buffer, "%7.2lfMB", mbs);
2111 return buffer;
2112 }
2113 #endif // PRODUCT
2115 size_t G1CollectorPolicy::max_regions(int purpose) {
2116 switch (purpose) {
2117 case GCAllocForSurvived:
2118 return _max_survivor_regions;
2119 case GCAllocForTenured:
2120 return REGIONS_UNLIMITED;
2121 default:
2122 ShouldNotReachHere();
2123 return REGIONS_UNLIMITED;
2124 };
2125 }
2127 void G1CollectorPolicy::update_max_gc_locker_expansion() {
2128 size_t expansion_region_num = 0;
2129 if (GCLockerEdenExpansionPercent > 0) {
2130 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
2131 double expansion_region_num_d = perc * (double) _young_list_target_length;
2132 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
2133 // less than 1.0) we'll get 1.
2134 expansion_region_num = (size_t) ceil(expansion_region_num_d);
2135 } else {
2136 assert(expansion_region_num == 0, "sanity");
2137 }
2138 _young_list_max_length = _young_list_target_length + expansion_region_num;
2139 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
2140 }
2142 // Calculates survivor space parameters.
2143 void G1CollectorPolicy::update_survivors_policy() {
2144 double max_survivor_regions_d =
2145 (double) _young_list_target_length / (double) SurvivorRatio;
2146 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
2147 // smaller than 1.0) we'll get 1.
2148 _max_survivor_regions = (size_t) ceil(max_survivor_regions_d);
2150 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2151 HeapRegion::GrainWords * _max_survivor_regions);
2152 }
2154 #ifndef PRODUCT
2155 class HRSortIndexIsOKClosure: public HeapRegionClosure {
2156 CollectionSetChooser* _chooser;
2157 public:
2158 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) :
2159 _chooser(chooser) {}
2161 bool doHeapRegion(HeapRegion* r) {
2162 if (!r->continuesHumongous()) {
2163 assert(_chooser->regionProperlyOrdered(r), "Ought to be.");
2164 }
2165 return false;
2166 }
2167 };
2169 bool G1CollectorPolicy::assertMarkedBytesDataOK() {
2170 HRSortIndexIsOKClosure cl(_collectionSetChooser);
2171 _g1->heap_region_iterate(&cl);
2172 return true;
2173 }
2174 #endif
2176 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
2177 GCCause::Cause gc_cause) {
2178 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2179 if (!during_cycle) {
2180 ergo_verbose1(ErgoConcCycles,
2181 "request concurrent cycle initiation",
2182 ergo_format_reason("requested by GC cause")
2183 ergo_format_str("GC cause"),
2184 GCCause::to_string(gc_cause));
2185 set_initiate_conc_mark_if_possible();
2186 return true;
2187 } else {
2188 ergo_verbose1(ErgoConcCycles,
2189 "do not request concurrent cycle initiation",
2190 ergo_format_reason("concurrent cycle already in progress")
2191 ergo_format_str("GC cause"),
2192 GCCause::to_string(gc_cause));
2193 return false;
2194 }
2195 }
2197 void
2198 G1CollectorPolicy::decide_on_conc_mark_initiation() {
2199 // We are about to decide on whether this pause will be an
2200 // initial-mark pause.
2202 // First, during_initial_mark_pause() should not be already set. We
2203 // will set it here if we have to. However, it should be cleared by
2204 // the end of the pause (it's only set for the duration of an
2205 // initial-mark pause).
2206 assert(!during_initial_mark_pause(), "pre-condition");
2208 if (initiate_conc_mark_if_possible()) {
2209 // We had noticed on a previous pause that the heap occupancy has
2210 // gone over the initiating threshold and we should start a
2211 // concurrent marking cycle. So we might initiate one.
2213 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2214 if (!during_cycle) {
2215 // The concurrent marking thread is not "during a cycle", i.e.,
2216 // it has completed the last one. So we can go ahead and
2217 // initiate a new cycle.
2219 set_during_initial_mark_pause();
2220 // We do not allow mixed GCs during marking.
2221 if (!gcs_are_young()) {
2222 set_gcs_are_young(true);
2223 ergo_verbose0(ErgoMixedGCs,
2224 "end mixed GCs",
2225 ergo_format_reason("concurrent cycle is about to start"));
2226 }
2228 // And we can now clear initiate_conc_mark_if_possible() as
2229 // we've already acted on it.
2230 clear_initiate_conc_mark_if_possible();
2232 ergo_verbose0(ErgoConcCycles,
2233 "initiate concurrent cycle",
2234 ergo_format_reason("concurrent cycle initiation requested"));
2235 } else {
2236 // The concurrent marking thread is still finishing up the
2237 // previous cycle. If we start one right now the two cycles
2238 // overlap. In particular, the concurrent marking thread might
2239 // be in the process of clearing the next marking bitmap (which
2240 // we will use for the next cycle if we start one). Starting a
2241 // cycle now will be bad given that parts of the marking
2242 // information might get cleared by the marking thread. And we
2243 // cannot wait for the marking thread to finish the cycle as it
2244 // periodically yields while clearing the next marking bitmap
2245 // and, if it's in a yield point, it's waiting for us to
2246 // finish. So, at this point we will not start a cycle and we'll
2247 // let the concurrent marking thread complete the last one.
2248 ergo_verbose0(ErgoConcCycles,
2249 "do not initiate concurrent cycle",
2250 ergo_format_reason("concurrent cycle already in progress"));
2251 }
2252 }
2253 }
2255 class KnownGarbageClosure: public HeapRegionClosure {
2256 CollectionSetChooser* _hrSorted;
2258 public:
2259 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2260 _hrSorted(hrSorted)
2261 {}
2263 bool doHeapRegion(HeapRegion* r) {
2264 // We only include humongous regions in collection
2265 // sets when concurrent mark shows that their contained object is
2266 // unreachable.
2268 // Do we have any marking information for this region?
2269 if (r->is_marked()) {
2270 // We don't include humongous regions in collection
2271 // sets because we collect them immediately at the end of a marking
2272 // cycle. We also don't include young regions because we *must*
2273 // include them in the next collection pause.
2274 if (!r->isHumongous() && !r->is_young()) {
2275 _hrSorted->addMarkedHeapRegion(r);
2276 }
2277 }
2278 return false;
2279 }
2280 };
2282 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2283 CollectionSetChooser* _hrSorted;
2284 jint _marked_regions_added;
2285 jint _chunk_size;
2286 jint _cur_chunk_idx;
2287 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2288 int _worker;
2289 int _invokes;
2291 void get_new_chunk() {
2292 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size);
2293 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2294 }
2295 void add_region(HeapRegion* r) {
2296 if (_cur_chunk_idx == _cur_chunk_end) {
2297 get_new_chunk();
2298 }
2299 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2300 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r);
2301 _marked_regions_added++;
2302 _cur_chunk_idx++;
2303 }
2305 public:
2306 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2307 jint chunk_size,
2308 int worker) :
2309 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker),
2310 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0),
2311 _invokes(0)
2312 {}
2314 bool doHeapRegion(HeapRegion* r) {
2315 // We only include humongous regions in collection
2316 // sets when concurrent mark shows that their contained object is
2317 // unreachable.
2318 _invokes++;
2320 // Do we have any marking information for this region?
2321 if (r->is_marked()) {
2322 // We don't include humongous regions in collection
2323 // sets because we collect them immediately at the end of a marking
2324 // cycle.
2325 // We also do not include young regions in collection sets
2326 if (!r->isHumongous() && !r->is_young()) {
2327 add_region(r);
2328 }
2329 }
2330 return false;
2331 }
2332 jint marked_regions_added() { return _marked_regions_added; }
2333 int invokes() { return _invokes; }
2334 };
2336 class ParKnownGarbageTask: public AbstractGangTask {
2337 CollectionSetChooser* _hrSorted;
2338 jint _chunk_size;
2339 G1CollectedHeap* _g1;
2340 public:
2341 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) :
2342 AbstractGangTask("ParKnownGarbageTask"),
2343 _hrSorted(hrSorted), _chunk_size(chunk_size),
2344 _g1(G1CollectedHeap::heap())
2345 {}
2347 void work(uint worker_id) {
2348 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted,
2349 _chunk_size,
2350 worker_id);
2351 // Back to zero for the claim value.
2352 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id,
2353 _g1->workers()->active_workers(),
2354 HeapRegion::InitialClaimValue);
2355 jint regions_added = parKnownGarbageCl.marked_regions_added();
2356 _hrSorted->incNumMarkedHeapRegions(regions_added);
2357 if (G1PrintParCleanupStats) {
2358 gclog_or_tty->print_cr(" Thread %d called %d times, added %d regions to list.",
2359 worker_id, parKnownGarbageCl.invokes(), regions_added);
2360 }
2361 }
2362 };
2364 void
2365 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) {
2366 double start_sec;
2367 if (G1PrintParCleanupStats) {
2368 start_sec = os::elapsedTime();
2369 }
2371 _collectionSetChooser->clearMarkedHeapRegions();
2372 double clear_marked_end_sec;
2373 if (G1PrintParCleanupStats) {
2374 clear_marked_end_sec = os::elapsedTime();
2375 gclog_or_tty->print_cr(" clear marked regions: %8.3f ms.",
2376 (clear_marked_end_sec - start_sec) * 1000.0);
2377 }
2379 if (G1CollectedHeap::use_parallel_gc_threads()) {
2380 const size_t OverpartitionFactor = 4;
2381 size_t WorkUnit;
2382 // The use of MinChunkSize = 8 in the original code
2383 // causes some assertion failures when the total number of
2384 // region is less than 8. The code here tries to fix that.
2385 // Should the original code also be fixed?
2386 if (no_of_gc_threads > 0) {
2387 const size_t MinWorkUnit =
2388 MAX2(_g1->n_regions() / no_of_gc_threads, (size_t) 1U);
2389 WorkUnit =
2390 MAX2(_g1->n_regions() / (no_of_gc_threads * OverpartitionFactor),
2391 MinWorkUnit);
2392 } else {
2393 assert(no_of_gc_threads > 0,
2394 "The active gc workers should be greater than 0");
2395 // In a product build do something reasonable to avoid a crash.
2396 const size_t MinWorkUnit =
2397 MAX2(_g1->n_regions() / ParallelGCThreads, (size_t) 1U);
2398 WorkUnit =
2399 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor),
2400 MinWorkUnit);
2401 }
2402 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(),
2403 WorkUnit);
2404 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2405 (int) WorkUnit);
2406 _g1->workers()->run_task(&parKnownGarbageTask);
2408 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2409 "sanity check");
2410 } else {
2411 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2412 _g1->heap_region_iterate(&knownGarbagecl);
2413 }
2414 double known_garbage_end_sec;
2415 if (G1PrintParCleanupStats) {
2416 known_garbage_end_sec = os::elapsedTime();
2417 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.",
2418 (known_garbage_end_sec - clear_marked_end_sec) * 1000.0);
2419 }
2421 _collectionSetChooser->sortMarkedHeapRegions();
2422 double end_sec = os::elapsedTime();
2423 if (G1PrintParCleanupStats) {
2424 gclog_or_tty->print_cr(" sorting: %8.3f ms.",
2425 (end_sec - known_garbage_end_sec) * 1000.0);
2426 }
2428 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
2429 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
2430 _cur_mark_stop_world_time_ms += elapsed_time_ms;
2431 _prev_collection_pause_end_ms += elapsed_time_ms;
2432 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
2433 }
2435 // Add the heap region at the head of the non-incremental collection set
2436 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
2437 assert(_inc_cset_build_state == Active, "Precondition");
2438 assert(!hr->is_young(), "non-incremental add of young region");
2440 assert(!hr->in_collection_set(), "should not already be in the CSet");
2441 hr->set_in_collection_set(true);
2442 hr->set_next_in_collection_set(_collection_set);
2443 _collection_set = hr;
2444 _collection_set_bytes_used_before += hr->used();
2445 _g1->register_region_with_in_cset_fast_test(hr);
2446 size_t rs_length = hr->rem_set()->occupied();
2447 _recorded_rs_lengths += rs_length;
2448 _old_cset_region_length += 1;
2449 }
2451 // Initialize the per-collection-set information
2452 void G1CollectorPolicy::start_incremental_cset_building() {
2453 assert(_inc_cset_build_state == Inactive, "Precondition");
2455 _inc_cset_head = NULL;
2456 _inc_cset_tail = NULL;
2457 _inc_cset_bytes_used_before = 0;
2459 _inc_cset_max_finger = 0;
2460 _inc_cset_recorded_rs_lengths = 0;
2461 _inc_cset_recorded_rs_lengths_diffs = 0;
2462 _inc_cset_predicted_elapsed_time_ms = 0.0;
2463 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
2464 _inc_cset_build_state = Active;
2465 }
2467 void G1CollectorPolicy::finalize_incremental_cset_building() {
2468 assert(_inc_cset_build_state == Active, "Precondition");
2469 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2471 // The two "main" fields, _inc_cset_recorded_rs_lengths and
2472 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread
2473 // that adds a new region to the CSet. Further updates by the
2474 // concurrent refinement thread that samples the young RSet lengths
2475 // are accumulated in the *_diffs fields. Here we add the diffs to
2476 // the "main" fields.
2478 if (_inc_cset_recorded_rs_lengths_diffs >= 0) {
2479 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs;
2480 } else {
2481 // This is defensive. The diff should in theory be always positive
2482 // as RSets can only grow between GCs. However, given that we
2483 // sample their size concurrently with other threads updating them
2484 // it's possible that we might get the wrong size back, which
2485 // could make the calculations somewhat inaccurate.
2486 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs);
2487 if (_inc_cset_recorded_rs_lengths >= diffs) {
2488 _inc_cset_recorded_rs_lengths -= diffs;
2489 } else {
2490 _inc_cset_recorded_rs_lengths = 0;
2491 }
2492 }
2493 _inc_cset_predicted_elapsed_time_ms +=
2494 _inc_cset_predicted_elapsed_time_ms_diffs;
2496 _inc_cset_recorded_rs_lengths_diffs = 0;
2497 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0;
2498 }
2500 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
2501 // This routine is used when:
2502 // * adding survivor regions to the incremental cset at the end of an
2503 // evacuation pause,
2504 // * adding the current allocation region to the incremental cset
2505 // when it is retired, and
2506 // * updating existing policy information for a region in the
2507 // incremental cset via young list RSet sampling.
2508 // Therefore this routine may be called at a safepoint by the
2509 // VM thread, or in-between safepoints by mutator threads (when
2510 // retiring the current allocation region) or a concurrent
2511 // refine thread (RSet sampling).
2513 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2514 size_t used_bytes = hr->used();
2515 _inc_cset_recorded_rs_lengths += rs_length;
2516 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
2517 _inc_cset_bytes_used_before += used_bytes;
2519 // Cache the values we have added to the aggregated informtion
2520 // in the heap region in case we have to remove this region from
2521 // the incremental collection set, or it is updated by the
2522 // rset sampling code
2523 hr->set_recorded_rs_length(rs_length);
2524 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
2525 }
2527 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr,
2528 size_t new_rs_length) {
2529 // Update the CSet information that is dependent on the new RS length
2530 assert(hr->is_young(), "Precondition");
2531 assert(!SafepointSynchronize::is_at_safepoint(),
2532 "should not be at a safepoint");
2534 // We could have updated _inc_cset_recorded_rs_lengths and
2535 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do
2536 // that atomically, as this code is executed by a concurrent
2537 // refinement thread, potentially concurrently with a mutator thread
2538 // allocating a new region and also updating the same fields. To
2539 // avoid the atomic operations we accumulate these updates on two
2540 // separate fields (*_diffs) and we'll just add them to the "main"
2541 // fields at the start of a GC.
2543 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length();
2544 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length;
2545 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff;
2547 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
2548 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2549 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms;
2550 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff;
2552 hr->set_recorded_rs_length(new_rs_length);
2553 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms);
2554 }
2556 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
2557 assert(hr->is_young(), "invariant");
2558 assert(hr->young_index_in_cset() > -1, "should have already been set");
2559 assert(_inc_cset_build_state == Active, "Precondition");
2561 // We need to clear and set the cached recorded/cached collection set
2562 // information in the heap region here (before the region gets added
2563 // to the collection set). An individual heap region's cached values
2564 // are calculated, aggregated with the policy collection set info,
2565 // and cached in the heap region here (initially) and (subsequently)
2566 // by the Young List sampling code.
2568 size_t rs_length = hr->rem_set()->occupied();
2569 add_to_incremental_cset_info(hr, rs_length);
2571 HeapWord* hr_end = hr->end();
2572 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
2574 assert(!hr->in_collection_set(), "invariant");
2575 hr->set_in_collection_set(true);
2576 assert( hr->next_in_collection_set() == NULL, "invariant");
2578 _g1->register_region_with_in_cset_fast_test(hr);
2579 }
2581 // Add the region at the RHS of the incremental cset
2582 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
2583 // We should only ever be appending survivors at the end of a pause
2584 assert( hr->is_survivor(), "Logic");
2586 // Do the 'common' stuff
2587 add_region_to_incremental_cset_common(hr);
2589 // Now add the region at the right hand side
2590 if (_inc_cset_tail == NULL) {
2591 assert(_inc_cset_head == NULL, "invariant");
2592 _inc_cset_head = hr;
2593 } else {
2594 _inc_cset_tail->set_next_in_collection_set(hr);
2595 }
2596 _inc_cset_tail = hr;
2597 }
2599 // Add the region to the LHS of the incremental cset
2600 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
2601 // Survivors should be added to the RHS at the end of a pause
2602 assert(!hr->is_survivor(), "Logic");
2604 // Do the 'common' stuff
2605 add_region_to_incremental_cset_common(hr);
2607 // Add the region at the left hand side
2608 hr->set_next_in_collection_set(_inc_cset_head);
2609 if (_inc_cset_head == NULL) {
2610 assert(_inc_cset_tail == NULL, "Invariant");
2611 _inc_cset_tail = hr;
2612 }
2613 _inc_cset_head = hr;
2614 }
2616 #ifndef PRODUCT
2617 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
2618 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
2620 st->print_cr("\nCollection_set:");
2621 HeapRegion* csr = list_head;
2622 while (csr != NULL) {
2623 HeapRegion* next = csr->next_in_collection_set();
2624 assert(csr->in_collection_set(), "bad CS");
2625 st->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
2626 "age: %4d, y: %d, surv: %d",
2627 csr->bottom(), csr->end(),
2628 csr->top(),
2629 csr->prev_top_at_mark_start(),
2630 csr->next_top_at_mark_start(),
2631 csr->top_at_conc_mark_count(),
2632 csr->age_in_surv_rate_group_cond(),
2633 csr->is_young(),
2634 csr->is_survivor());
2635 csr = next;
2636 }
2637 }
2638 #endif // !PRODUCT
2640 void G1CollectorPolicy::choose_collection_set(double target_pause_time_ms) {
2641 // Set this here - in case we're not doing young collections.
2642 double non_young_start_time_sec = os::elapsedTime();
2644 YoungList* young_list = _g1->young_list();
2645 finalize_incremental_cset_building();
2647 guarantee(target_pause_time_ms > 0.0,
2648 err_msg("target_pause_time_ms = %1.6lf should be positive",
2649 target_pause_time_ms));
2650 guarantee(_collection_set == NULL, "Precondition");
2652 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2653 double predicted_pause_time_ms = base_time_ms;
2655 double time_remaining_ms = target_pause_time_ms - base_time_ms;
2657 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2658 "start choosing CSet",
2659 ergo_format_ms("predicted base time")
2660 ergo_format_ms("remaining time")
2661 ergo_format_ms("target pause time"),
2662 base_time_ms, time_remaining_ms, target_pause_time_ms);
2664 // the 10% and 50% values are arbitrary...
2665 double threshold = 0.10 * target_pause_time_ms;
2666 if (time_remaining_ms < threshold) {
2667 double prev_time_remaining_ms = time_remaining_ms;
2668 time_remaining_ms = 0.50 * target_pause_time_ms;
2669 ergo_verbose3(ErgoCSetConstruction,
2670 "adjust remaining time",
2671 ergo_format_reason("remaining time lower than threshold")
2672 ergo_format_ms("remaining time")
2673 ergo_format_ms("threshold")
2674 ergo_format_ms("adjusted remaining time"),
2675 prev_time_remaining_ms, threshold, time_remaining_ms);
2676 }
2678 size_t expansion_bytes = _g1->expansion_regions() * HeapRegion::GrainBytes;
2680 HeapRegion* hr;
2681 double young_start_time_sec = os::elapsedTime();
2683 _collection_set_bytes_used_before = 0;
2684 _last_gc_was_young = gcs_are_young() ? true : false;
2686 if (_last_gc_was_young) {
2687 ++_young_pause_num;
2688 } else {
2689 ++_mixed_pause_num;
2690 }
2692 // The young list is laid with the survivor regions from the previous
2693 // pause are appended to the RHS of the young list, i.e.
2694 // [Newly Young Regions ++ Survivors from last pause].
2696 size_t survivor_region_length = young_list->survivor_length();
2697 size_t eden_region_length = young_list->length() - survivor_region_length;
2698 init_cset_region_lengths(eden_region_length, survivor_region_length);
2699 hr = young_list->first_survivor_region();
2700 while (hr != NULL) {
2701 assert(hr->is_survivor(), "badly formed young list");
2702 hr->set_young();
2703 hr = hr->get_next_young_region();
2704 }
2706 // Clear the fields that point to the survivor list - they are all young now.
2707 young_list->clear_survivors();
2709 _collection_set = _inc_cset_head;
2710 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2711 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms;
2712 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
2714 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2715 "add young regions to CSet",
2716 ergo_format_region("eden")
2717 ergo_format_region("survivors")
2718 ergo_format_ms("predicted young region time"),
2719 eden_region_length, survivor_region_length,
2720 _inc_cset_predicted_elapsed_time_ms);
2722 // The number of recorded young regions is the incremental
2723 // collection set's current size
2724 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2726 double young_end_time_sec = os::elapsedTime();
2727 _recorded_young_cset_choice_time_ms =
2728 (young_end_time_sec - young_start_time_sec) * 1000.0;
2730 // We are doing young collections so reset this.
2731 non_young_start_time_sec = young_end_time_sec;
2733 if (!gcs_are_young()) {
2734 bool should_continue = true;
2735 NumberSeq seq;
2736 double avg_prediction = 100000000000000000.0; // something very large
2738 double prev_predicted_pause_time_ms = predicted_pause_time_ms;
2739 do {
2740 // Note that add_old_region_to_cset() increments the
2741 // _old_cset_region_length field and cset_region_length() returns the
2742 // sum of _eden_cset_region_length, _survivor_cset_region_length, and
2743 // _old_cset_region_length. So, as old regions are added to the
2744 // CSet, _old_cset_region_length will be incremented and
2745 // cset_region_length(), which is used below, will always reflect
2746 // the the total number of regions added up to this point to the CSet.
2748 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms,
2749 avg_prediction);
2750 if (hr != NULL) {
2751 _g1->old_set_remove(hr);
2752 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
2753 time_remaining_ms -= predicted_time_ms;
2754 predicted_pause_time_ms += predicted_time_ms;
2755 add_old_region_to_cset(hr);
2756 seq.add(predicted_time_ms);
2757 avg_prediction = seq.avg() + seq.sd();
2758 }
2760 should_continue = true;
2761 if (hr == NULL) {
2762 // No need for an ergo verbose message here,
2763 // getNextMarkRegion() does this when it returns NULL.
2764 should_continue = false;
2765 } else {
2766 if (adaptive_young_list_length()) {
2767 if (time_remaining_ms < 0.0) {
2768 ergo_verbose1(ErgoCSetConstruction,
2769 "stop adding old regions to CSet",
2770 ergo_format_reason("remaining time is lower than 0")
2771 ergo_format_ms("remaining time"),
2772 time_remaining_ms);
2773 should_continue = false;
2774 }
2775 } else {
2776 if (cset_region_length() >= _young_list_fixed_length) {
2777 ergo_verbose2(ErgoCSetConstruction,
2778 "stop adding old regions to CSet",
2779 ergo_format_reason("CSet length reached target")
2780 ergo_format_region("CSet")
2781 ergo_format_region("young target"),
2782 cset_region_length(), _young_list_fixed_length);
2783 should_continue = false;
2784 }
2785 }
2786 }
2787 } while (should_continue);
2789 if (!adaptive_young_list_length() &&
2790 cset_region_length() < _young_list_fixed_length) {
2791 ergo_verbose2(ErgoCSetConstruction,
2792 "request mixed GCs end",
2793 ergo_format_reason("CSet length lower than target")
2794 ergo_format_region("CSet")
2795 ergo_format_region("young target"),
2796 cset_region_length(), _young_list_fixed_length);
2797 _should_revert_to_young_gcs = true;
2798 }
2800 ergo_verbose2(ErgoCSetConstruction | ErgoHigh,
2801 "add old regions to CSet",
2802 ergo_format_region("old")
2803 ergo_format_ms("predicted old region time"),
2804 old_cset_region_length(),
2805 predicted_pause_time_ms - prev_predicted_pause_time_ms);
2806 }
2808 stop_incremental_cset_building();
2810 count_CS_bytes_used();
2812 ergo_verbose5(ErgoCSetConstruction,
2813 "finish choosing CSet",
2814 ergo_format_region("eden")
2815 ergo_format_region("survivors")
2816 ergo_format_region("old")
2817 ergo_format_ms("predicted pause time")
2818 ergo_format_ms("target pause time"),
2819 eden_region_length, survivor_region_length,
2820 old_cset_region_length(),
2821 predicted_pause_time_ms, target_pause_time_ms);
2823 double non_young_end_time_sec = os::elapsedTime();
2824 _recorded_non_young_cset_choice_time_ms =
2825 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
2826 }