|
1 /* |
|
2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved. |
|
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
|
4 * |
|
5 * This code is free software; you can redistribute it and/or modify it |
|
6 * under the terms of the GNU General Public License version 2 only, as |
|
7 * published by the Free Software Foundation. |
|
8 * |
|
9 * This code is distributed in the hope that it will be useful, but WITHOUT |
|
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
|
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
|
12 * version 2 for more details (a copy is included in the LICENSE file that |
|
13 * accompanied this code). |
|
14 * |
|
15 * You should have received a copy of the GNU General Public License version |
|
16 * 2 along with this work; if not, write to the Free Software Foundation, |
|
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
|
18 * |
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
|
20 * or visit www.oracle.com if you need additional information or have any |
|
21 * questions. |
|
22 * |
|
23 */ |
|
24 |
|
25 #ifndef __clang_major__ |
|
26 #define ATTRIBUTE_PRINTF(x,y) // FIXME, formats are a mess. |
|
27 #endif |
|
28 |
|
29 #include "precompiled.hpp" |
|
30 #include "gc_implementation/g1/concurrentG1Refine.hpp" |
|
31 #include "gc_implementation/g1/concurrentMark.hpp" |
|
32 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp" |
|
33 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" |
|
34 #include "gc_implementation/g1/g1CollectorPolicy.hpp" |
|
35 #include "gc_implementation/g1/g1ErgoVerbose.hpp" |
|
36 #include "gc_implementation/g1/g1GCPhaseTimes.hpp" |
|
37 #include "gc_implementation/g1/g1Log.hpp" |
|
38 #include "gc_implementation/g1/heapRegionRemSet.hpp" |
|
39 #include "gc_implementation/shared/gcPolicyCounters.hpp" |
|
40 #include "runtime/arguments.hpp" |
|
41 #include "runtime/java.hpp" |
|
42 #include "runtime/mutexLocker.hpp" |
|
43 #include "utilities/debug.hpp" |
|
44 |
|
45 // Different defaults for different number of GC threads |
|
46 // They were chosen by running GCOld and SPECjbb on debris with different |
|
47 // numbers of GC threads and choosing them based on the results |
|
48 |
|
49 // all the same |
|
50 static double rs_length_diff_defaults[] = { |
|
51 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 |
|
52 }; |
|
53 |
|
54 static double cost_per_card_ms_defaults[] = { |
|
55 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015 |
|
56 }; |
|
57 |
|
58 // all the same |
|
59 static double young_cards_per_entry_ratio_defaults[] = { |
|
60 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 |
|
61 }; |
|
62 |
|
63 static double cost_per_entry_ms_defaults[] = { |
|
64 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 |
|
65 }; |
|
66 |
|
67 static double cost_per_byte_ms_defaults[] = { |
|
68 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 |
|
69 }; |
|
70 |
|
71 // these should be pretty consistent |
|
72 static double constant_other_time_ms_defaults[] = { |
|
73 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 |
|
74 }; |
|
75 |
|
76 |
|
77 static double young_other_cost_per_region_ms_defaults[] = { |
|
78 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 |
|
79 }; |
|
80 |
|
81 static double non_young_other_cost_per_region_ms_defaults[] = { |
|
82 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 |
|
83 }; |
|
84 |
|
85 G1CollectorPolicy::G1CollectorPolicy() : |
|
86 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads() |
|
87 ? ParallelGCThreads : 1), |
|
88 |
|
89 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), |
|
90 _stop_world_start(0.0), |
|
91 |
|
92 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), |
|
93 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), |
|
94 |
|
95 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
96 _prev_collection_pause_end_ms(0.0), |
|
97 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
98 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
99 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
100 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
101 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
102 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
103 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
104 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
105 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
106 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
107 _non_young_other_cost_per_region_ms_seq( |
|
108 new TruncatedSeq(TruncatedSeqLength)), |
|
109 |
|
110 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
111 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), |
|
112 |
|
113 _pause_time_target_ms((double) MaxGCPauseMillis), |
|
114 |
|
115 _gcs_are_young(true), |
|
116 |
|
117 _during_marking(false), |
|
118 _in_marking_window(false), |
|
119 _in_marking_window_im(false), |
|
120 |
|
121 _recent_prev_end_times_for_all_gcs_sec( |
|
122 new TruncatedSeq(NumPrevPausesForHeuristics)), |
|
123 |
|
124 _recent_avg_pause_time_ratio(0.0), |
|
125 |
|
126 _initiate_conc_mark_if_possible(false), |
|
127 _during_initial_mark_pause(false), |
|
128 _last_young_gc(false), |
|
129 _last_gc_was_young(false), |
|
130 |
|
131 _eden_used_bytes_before_gc(0), |
|
132 _survivor_used_bytes_before_gc(0), |
|
133 _heap_used_bytes_before_gc(0), |
|
134 _metaspace_used_bytes_before_gc(0), |
|
135 _eden_capacity_bytes_before_gc(0), |
|
136 _heap_capacity_bytes_before_gc(0), |
|
137 |
|
138 _eden_cset_region_length(0), |
|
139 _survivor_cset_region_length(0), |
|
140 _old_cset_region_length(0), |
|
141 |
|
142 _collection_set(NULL), |
|
143 _collection_set_bytes_used_before(0), |
|
144 |
|
145 // Incremental CSet attributes |
|
146 _inc_cset_build_state(Inactive), |
|
147 _inc_cset_head(NULL), |
|
148 _inc_cset_tail(NULL), |
|
149 _inc_cset_bytes_used_before(0), |
|
150 _inc_cset_max_finger(NULL), |
|
151 _inc_cset_recorded_rs_lengths(0), |
|
152 _inc_cset_recorded_rs_lengths_diffs(0), |
|
153 _inc_cset_predicted_elapsed_time_ms(0.0), |
|
154 _inc_cset_predicted_elapsed_time_ms_diffs(0.0), |
|
155 |
|
156 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away |
|
157 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list |
|
158 #endif // _MSC_VER |
|
159 |
|
160 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived", |
|
161 G1YoungSurvRateNumRegionsSummary)), |
|
162 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor", |
|
163 G1YoungSurvRateNumRegionsSummary)), |
|
164 // add here any more surv rate groups |
|
165 _recorded_survivor_regions(0), |
|
166 _recorded_survivor_head(NULL), |
|
167 _recorded_survivor_tail(NULL), |
|
168 _survivors_age_table(true), |
|
169 |
|
170 _gc_overhead_perc(0.0) { |
|
171 |
|
172 // Set up the region size and associated fields. Given that the |
|
173 // policy is created before the heap, we have to set this up here, |
|
174 // so it's done as soon as possible. |
|
175 |
|
176 // It would have been natural to pass initial_heap_byte_size() and |
|
177 // max_heap_byte_size() to setup_heap_region_size() but those have |
|
178 // not been set up at this point since they should be aligned with |
|
179 // the region size. So, there is a circular dependency here. We base |
|
180 // the region size on the heap size, but the heap size should be |
|
181 // aligned with the region size. To get around this we use the |
|
182 // unaligned values for the heap. |
|
183 HeapRegion::setup_heap_region_size(InitialHeapSize, MaxHeapSize); |
|
184 HeapRegionRemSet::setup_remset_size(); |
|
185 |
|
186 G1ErgoVerbose::initialize(); |
|
187 if (PrintAdaptiveSizePolicy) { |
|
188 // Currently, we only use a single switch for all the heuristics. |
|
189 G1ErgoVerbose::set_enabled(true); |
|
190 // Given that we don't currently have a verboseness level |
|
191 // parameter, we'll hardcode this to high. This can be easily |
|
192 // changed in the future. |
|
193 G1ErgoVerbose::set_level(ErgoHigh); |
|
194 } else { |
|
195 G1ErgoVerbose::set_enabled(false); |
|
196 } |
|
197 |
|
198 // Verify PLAB sizes |
|
199 const size_t region_size = HeapRegion::GrainWords; |
|
200 if (YoungPLABSize > region_size || OldPLABSize > region_size) { |
|
201 char buffer[128]; |
|
202 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT, |
|
203 OldPLABSize > region_size ? "Old" : "Young", region_size); |
|
204 vm_exit_during_initialization(buffer); |
|
205 } |
|
206 |
|
207 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); |
|
208 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; |
|
209 |
|
210 _phase_times = new G1GCPhaseTimes(_parallel_gc_threads); |
|
211 |
|
212 int index = MIN2(_parallel_gc_threads - 1, 7); |
|
213 |
|
214 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); |
|
215 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); |
|
216 _young_cards_per_entry_ratio_seq->add( |
|
217 young_cards_per_entry_ratio_defaults[index]); |
|
218 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); |
|
219 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); |
|
220 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); |
|
221 _young_other_cost_per_region_ms_seq->add( |
|
222 young_other_cost_per_region_ms_defaults[index]); |
|
223 _non_young_other_cost_per_region_ms_seq->add( |
|
224 non_young_other_cost_per_region_ms_defaults[index]); |
|
225 |
|
226 // Below, we might need to calculate the pause time target based on |
|
227 // the pause interval. When we do so we are going to give G1 maximum |
|
228 // flexibility and allow it to do pauses when it needs to. So, we'll |
|
229 // arrange that the pause interval to be pause time target + 1 to |
|
230 // ensure that a) the pause time target is maximized with respect to |
|
231 // the pause interval and b) we maintain the invariant that pause |
|
232 // time target < pause interval. If the user does not want this |
|
233 // maximum flexibility, they will have to set the pause interval |
|
234 // explicitly. |
|
235 |
|
236 // First make sure that, if either parameter is set, its value is |
|
237 // reasonable. |
|
238 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) { |
|
239 if (MaxGCPauseMillis < 1) { |
|
240 vm_exit_during_initialization("MaxGCPauseMillis should be " |
|
241 "greater than 0"); |
|
242 } |
|
243 } |
|
244 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { |
|
245 if (GCPauseIntervalMillis < 1) { |
|
246 vm_exit_during_initialization("GCPauseIntervalMillis should be " |
|
247 "greater than 0"); |
|
248 } |
|
249 } |
|
250 |
|
251 // Then, if the pause time target parameter was not set, set it to |
|
252 // the default value. |
|
253 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) { |
|
254 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { |
|
255 // The default pause time target in G1 is 200ms |
|
256 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200); |
|
257 } else { |
|
258 // We do not allow the pause interval to be set without the |
|
259 // pause time target |
|
260 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set " |
|
261 "without setting MaxGCPauseMillis"); |
|
262 } |
|
263 } |
|
264 |
|
265 // Then, if the interval parameter was not set, set it according to |
|
266 // the pause time target (this will also deal with the case when the |
|
267 // pause time target is the default value). |
|
268 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) { |
|
269 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1); |
|
270 } |
|
271 |
|
272 // Finally, make sure that the two parameters are consistent. |
|
273 if (MaxGCPauseMillis >= GCPauseIntervalMillis) { |
|
274 char buffer[256]; |
|
275 jio_snprintf(buffer, 256, |
|
276 "MaxGCPauseMillis (%u) should be less than " |
|
277 "GCPauseIntervalMillis (%u)", |
|
278 MaxGCPauseMillis, GCPauseIntervalMillis); |
|
279 vm_exit_during_initialization(buffer); |
|
280 } |
|
281 |
|
282 double max_gc_time = (double) MaxGCPauseMillis / 1000.0; |
|
283 double time_slice = (double) GCPauseIntervalMillis / 1000.0; |
|
284 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); |
|
285 |
|
286 uintx confidence_perc = G1ConfidencePercent; |
|
287 // Put an artificial ceiling on this so that it's not set to a silly value. |
|
288 if (confidence_perc > 100) { |
|
289 confidence_perc = 100; |
|
290 warning("G1ConfidencePercent is set to a value that is too large, " |
|
291 "it's been updated to %u", confidence_perc); |
|
292 } |
|
293 _sigma = (double) confidence_perc / 100.0; |
|
294 |
|
295 // start conservatively (around 50ms is about right) |
|
296 _concurrent_mark_remark_times_ms->add(0.05); |
|
297 _concurrent_mark_cleanup_times_ms->add(0.20); |
|
298 _tenuring_threshold = MaxTenuringThreshold; |
|
299 // _max_survivor_regions will be calculated by |
|
300 // update_young_list_target_length() during initialization. |
|
301 _max_survivor_regions = 0; |
|
302 |
|
303 assert(GCTimeRatio > 0, |
|
304 "we should have set it to a default value set_g1_gc_flags() " |
|
305 "if a user set it to 0"); |
|
306 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio)); |
|
307 |
|
308 uintx reserve_perc = G1ReservePercent; |
|
309 // Put an artificial ceiling on this so that it's not set to a silly value. |
|
310 if (reserve_perc > 50) { |
|
311 reserve_perc = 50; |
|
312 warning("G1ReservePercent is set to a value that is too large, " |
|
313 "it's been updated to %u", reserve_perc); |
|
314 } |
|
315 _reserve_factor = (double) reserve_perc / 100.0; |
|
316 // This will be set when the heap is expanded |
|
317 // for the first time during initialization. |
|
318 _reserve_regions = 0; |
|
319 |
|
320 _collectionSetChooser = new CollectionSetChooser(); |
|
321 } |
|
322 |
|
323 void G1CollectorPolicy::initialize_alignments() { |
|
324 _space_alignment = HeapRegion::GrainBytes; |
|
325 size_t card_table_alignment = GenRemSet::max_alignment_constraint(GenRemSet::CardTable); |
|
326 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size(); |
|
327 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size); |
|
328 } |
|
329 |
|
330 void G1CollectorPolicy::initialize_flags() { |
|
331 if (G1HeapRegionSize != HeapRegion::GrainBytes) { |
|
332 FLAG_SET_ERGO(uintx, G1HeapRegionSize, HeapRegion::GrainBytes); |
|
333 } |
|
334 |
|
335 if (SurvivorRatio < 1) { |
|
336 vm_exit_during_initialization("Invalid survivor ratio specified"); |
|
337 } |
|
338 CollectorPolicy::initialize_flags(); |
|
339 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags |
|
340 } |
|
341 |
|
342 void G1CollectorPolicy::post_heap_initialize() { |
|
343 uintx max_regions = G1CollectedHeap::heap()->max_regions(); |
|
344 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes; |
|
345 if (max_young_size != MaxNewSize) { |
|
346 FLAG_SET_ERGO(uintx, MaxNewSize, max_young_size); |
|
347 } |
|
348 } |
|
349 |
|
350 G1YoungGenSizer::G1YoungGenSizer() : _sizer_kind(SizerDefaults), _adaptive_size(true), |
|
351 _min_desired_young_length(0), _max_desired_young_length(0) { |
|
352 if (FLAG_IS_CMDLINE(NewRatio)) { |
|
353 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) { |
|
354 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio"); |
|
355 } else { |
|
356 _sizer_kind = SizerNewRatio; |
|
357 _adaptive_size = false; |
|
358 return; |
|
359 } |
|
360 } |
|
361 |
|
362 if (NewSize > MaxNewSize) { |
|
363 if (FLAG_IS_CMDLINE(MaxNewSize)) { |
|
364 warning("NewSize (" SIZE_FORMAT "k) is greater than the MaxNewSize (" SIZE_FORMAT "k). " |
|
365 "A new max generation size of " SIZE_FORMAT "k will be used.", |
|
366 NewSize/K, MaxNewSize/K, NewSize/K); |
|
367 } |
|
368 MaxNewSize = NewSize; |
|
369 } |
|
370 |
|
371 if (FLAG_IS_CMDLINE(NewSize)) { |
|
372 _min_desired_young_length = MAX2((uint) (NewSize / HeapRegion::GrainBytes), |
|
373 1U); |
|
374 if (FLAG_IS_CMDLINE(MaxNewSize)) { |
|
375 _max_desired_young_length = |
|
376 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), |
|
377 1U); |
|
378 _sizer_kind = SizerMaxAndNewSize; |
|
379 _adaptive_size = _min_desired_young_length == _max_desired_young_length; |
|
380 } else { |
|
381 _sizer_kind = SizerNewSizeOnly; |
|
382 } |
|
383 } else if (FLAG_IS_CMDLINE(MaxNewSize)) { |
|
384 _max_desired_young_length = |
|
385 MAX2((uint) (MaxNewSize / HeapRegion::GrainBytes), |
|
386 1U); |
|
387 _sizer_kind = SizerMaxNewSizeOnly; |
|
388 } |
|
389 } |
|
390 |
|
391 uint G1YoungGenSizer::calculate_default_min_length(uint new_number_of_heap_regions) { |
|
392 uint default_value = (new_number_of_heap_regions * G1NewSizePercent) / 100; |
|
393 return MAX2(1U, default_value); |
|
394 } |
|
395 |
|
396 uint G1YoungGenSizer::calculate_default_max_length(uint new_number_of_heap_regions) { |
|
397 uint default_value = (new_number_of_heap_regions * G1MaxNewSizePercent) / 100; |
|
398 return MAX2(1U, default_value); |
|
399 } |
|
400 |
|
401 void G1YoungGenSizer::recalculate_min_max_young_length(uint number_of_heap_regions, uint* min_young_length, uint* max_young_length) { |
|
402 assert(number_of_heap_regions > 0, "Heap must be initialized"); |
|
403 |
|
404 switch (_sizer_kind) { |
|
405 case SizerDefaults: |
|
406 *min_young_length = calculate_default_min_length(number_of_heap_regions); |
|
407 *max_young_length = calculate_default_max_length(number_of_heap_regions); |
|
408 break; |
|
409 case SizerNewSizeOnly: |
|
410 *max_young_length = calculate_default_max_length(number_of_heap_regions); |
|
411 *max_young_length = MAX2(*min_young_length, *max_young_length); |
|
412 break; |
|
413 case SizerMaxNewSizeOnly: |
|
414 *min_young_length = calculate_default_min_length(number_of_heap_regions); |
|
415 *min_young_length = MIN2(*min_young_length, *max_young_length); |
|
416 break; |
|
417 case SizerMaxAndNewSize: |
|
418 // Do nothing. Values set on the command line, don't update them at runtime. |
|
419 break; |
|
420 case SizerNewRatio: |
|
421 *min_young_length = number_of_heap_regions / (NewRatio + 1); |
|
422 *max_young_length = *min_young_length; |
|
423 break; |
|
424 default: |
|
425 ShouldNotReachHere(); |
|
426 } |
|
427 |
|
428 assert(*min_young_length <= *max_young_length, "Invalid min/max young gen size values"); |
|
429 } |
|
430 |
|
431 uint G1YoungGenSizer::max_young_length(uint number_of_heap_regions) { |
|
432 // We need to pass the desired values because recalculation may not update these |
|
433 // values in some cases. |
|
434 uint temp = _min_desired_young_length; |
|
435 uint result = _max_desired_young_length; |
|
436 recalculate_min_max_young_length(number_of_heap_regions, &temp, &result); |
|
437 return result; |
|
438 } |
|
439 |
|
440 void G1YoungGenSizer::heap_size_changed(uint new_number_of_heap_regions) { |
|
441 recalculate_min_max_young_length(new_number_of_heap_regions, &_min_desired_young_length, |
|
442 &_max_desired_young_length); |
|
443 } |
|
444 |
|
445 void G1CollectorPolicy::init() { |
|
446 // Set aside an initial future to_space. |
|
447 _g1 = G1CollectedHeap::heap(); |
|
448 |
|
449 assert(Heap_lock->owned_by_self(), "Locking discipline."); |
|
450 |
|
451 initialize_gc_policy_counters(); |
|
452 |
|
453 if (adaptive_young_list_length()) { |
|
454 _young_list_fixed_length = 0; |
|
455 } else { |
|
456 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length(); |
|
457 } |
|
458 _free_regions_at_end_of_collection = _g1->free_regions(); |
|
459 update_young_list_target_length(); |
|
460 |
|
461 // We may immediately start allocating regions and placing them on the |
|
462 // collection set list. Initialize the per-collection set info |
|
463 start_incremental_cset_building(); |
|
464 } |
|
465 |
|
466 // Create the jstat counters for the policy. |
|
467 void G1CollectorPolicy::initialize_gc_policy_counters() { |
|
468 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3); |
|
469 } |
|
470 |
|
471 bool G1CollectorPolicy::predict_will_fit(uint young_length, |
|
472 double base_time_ms, |
|
473 uint base_free_regions, |
|
474 double target_pause_time_ms) { |
|
475 if (young_length >= base_free_regions) { |
|
476 // end condition 1: not enough space for the young regions |
|
477 return false; |
|
478 } |
|
479 |
|
480 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1); |
|
481 size_t bytes_to_copy = |
|
482 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); |
|
483 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); |
|
484 double young_other_time_ms = predict_young_other_time_ms(young_length); |
|
485 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms; |
|
486 if (pause_time_ms > target_pause_time_ms) { |
|
487 // end condition 2: prediction is over the target pause time |
|
488 return false; |
|
489 } |
|
490 |
|
491 size_t free_bytes = |
|
492 (base_free_regions - young_length) * HeapRegion::GrainBytes; |
|
493 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) { |
|
494 // end condition 3: out-of-space (conservatively!) |
|
495 return false; |
|
496 } |
|
497 |
|
498 // success! |
|
499 return true; |
|
500 } |
|
501 |
|
502 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) { |
|
503 // re-calculate the necessary reserve |
|
504 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor; |
|
505 // We use ceiling so that if reserve_regions_d is > 0.0 (but |
|
506 // smaller than 1.0) we'll get 1. |
|
507 _reserve_regions = (uint) ceil(reserve_regions_d); |
|
508 |
|
509 _young_gen_sizer->heap_size_changed(new_number_of_regions); |
|
510 } |
|
511 |
|
512 uint G1CollectorPolicy::calculate_young_list_desired_min_length( |
|
513 uint base_min_length) { |
|
514 uint desired_min_length = 0; |
|
515 if (adaptive_young_list_length()) { |
|
516 if (_alloc_rate_ms_seq->num() > 3) { |
|
517 double now_sec = os::elapsedTime(); |
|
518 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; |
|
519 double alloc_rate_ms = predict_alloc_rate_ms(); |
|
520 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms); |
|
521 } else { |
|
522 // otherwise we don't have enough info to make the prediction |
|
523 } |
|
524 } |
|
525 desired_min_length += base_min_length; |
|
526 // make sure we don't go below any user-defined minimum bound |
|
527 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length); |
|
528 } |
|
529 |
|
530 uint G1CollectorPolicy::calculate_young_list_desired_max_length() { |
|
531 // Here, we might want to also take into account any additional |
|
532 // constraints (i.e., user-defined minimum bound). Currently, we |
|
533 // effectively don't set this bound. |
|
534 return _young_gen_sizer->max_desired_young_length(); |
|
535 } |
|
536 |
|
537 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) { |
|
538 if (rs_lengths == (size_t) -1) { |
|
539 // if it's set to the default value (-1), we should predict it; |
|
540 // otherwise, use the given value. |
|
541 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq); |
|
542 } |
|
543 |
|
544 // Calculate the absolute and desired min bounds. |
|
545 |
|
546 // This is how many young regions we already have (currently: the survivors). |
|
547 uint base_min_length = recorded_survivor_regions(); |
|
548 // This is the absolute minimum young length, which ensures that we |
|
549 // can allocate one eden region in the worst-case. |
|
550 uint absolute_min_length = base_min_length + 1; |
|
551 uint desired_min_length = |
|
552 calculate_young_list_desired_min_length(base_min_length); |
|
553 if (desired_min_length < absolute_min_length) { |
|
554 desired_min_length = absolute_min_length; |
|
555 } |
|
556 |
|
557 // Calculate the absolute and desired max bounds. |
|
558 |
|
559 // We will try our best not to "eat" into the reserve. |
|
560 uint absolute_max_length = 0; |
|
561 if (_free_regions_at_end_of_collection > _reserve_regions) { |
|
562 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions; |
|
563 } |
|
564 uint desired_max_length = calculate_young_list_desired_max_length(); |
|
565 if (desired_max_length > absolute_max_length) { |
|
566 desired_max_length = absolute_max_length; |
|
567 } |
|
568 |
|
569 uint young_list_target_length = 0; |
|
570 if (adaptive_young_list_length()) { |
|
571 if (gcs_are_young()) { |
|
572 young_list_target_length = |
|
573 calculate_young_list_target_length(rs_lengths, |
|
574 base_min_length, |
|
575 desired_min_length, |
|
576 desired_max_length); |
|
577 _rs_lengths_prediction = rs_lengths; |
|
578 } else { |
|
579 // Don't calculate anything and let the code below bound it to |
|
580 // the desired_min_length, i.e., do the next GC as soon as |
|
581 // possible to maximize how many old regions we can add to it. |
|
582 } |
|
583 } else { |
|
584 // The user asked for a fixed young gen so we'll fix the young gen |
|
585 // whether the next GC is young or mixed. |
|
586 young_list_target_length = _young_list_fixed_length; |
|
587 } |
|
588 |
|
589 // Make sure we don't go over the desired max length, nor under the |
|
590 // desired min length. In case they clash, desired_min_length wins |
|
591 // which is why that test is second. |
|
592 if (young_list_target_length > desired_max_length) { |
|
593 young_list_target_length = desired_max_length; |
|
594 } |
|
595 if (young_list_target_length < desired_min_length) { |
|
596 young_list_target_length = desired_min_length; |
|
597 } |
|
598 |
|
599 assert(young_list_target_length > recorded_survivor_regions(), |
|
600 "we should be able to allocate at least one eden region"); |
|
601 assert(young_list_target_length >= absolute_min_length, "post-condition"); |
|
602 _young_list_target_length = young_list_target_length; |
|
603 |
|
604 update_max_gc_locker_expansion(); |
|
605 } |
|
606 |
|
607 uint |
|
608 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths, |
|
609 uint base_min_length, |
|
610 uint desired_min_length, |
|
611 uint desired_max_length) { |
|
612 assert(adaptive_young_list_length(), "pre-condition"); |
|
613 assert(gcs_are_young(), "only call this for young GCs"); |
|
614 |
|
615 // In case some edge-condition makes the desired max length too small... |
|
616 if (desired_max_length <= desired_min_length) { |
|
617 return desired_min_length; |
|
618 } |
|
619 |
|
620 // We'll adjust min_young_length and max_young_length not to include |
|
621 // the already allocated young regions (i.e., so they reflect the |
|
622 // min and max eden regions we'll allocate). The base_min_length |
|
623 // will be reflected in the predictions by the |
|
624 // survivor_regions_evac_time prediction. |
|
625 assert(desired_min_length > base_min_length, "invariant"); |
|
626 uint min_young_length = desired_min_length - base_min_length; |
|
627 assert(desired_max_length > base_min_length, "invariant"); |
|
628 uint max_young_length = desired_max_length - base_min_length; |
|
629 |
|
630 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; |
|
631 double survivor_regions_evac_time = predict_survivor_regions_evac_time(); |
|
632 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); |
|
633 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); |
|
634 size_t scanned_cards = predict_young_card_num(adj_rs_lengths); |
|
635 double base_time_ms = |
|
636 predict_base_elapsed_time_ms(pending_cards, scanned_cards) + |
|
637 survivor_regions_evac_time; |
|
638 uint available_free_regions = _free_regions_at_end_of_collection; |
|
639 uint base_free_regions = 0; |
|
640 if (available_free_regions > _reserve_regions) { |
|
641 base_free_regions = available_free_regions - _reserve_regions; |
|
642 } |
|
643 |
|
644 // Here, we will make sure that the shortest young length that |
|
645 // makes sense fits within the target pause time. |
|
646 |
|
647 if (predict_will_fit(min_young_length, base_time_ms, |
|
648 base_free_regions, target_pause_time_ms)) { |
|
649 // The shortest young length will fit into the target pause time; |
|
650 // we'll now check whether the absolute maximum number of young |
|
651 // regions will fit in the target pause time. If not, we'll do |
|
652 // a binary search between min_young_length and max_young_length. |
|
653 if (predict_will_fit(max_young_length, base_time_ms, |
|
654 base_free_regions, target_pause_time_ms)) { |
|
655 // The maximum young length will fit into the target pause time. |
|
656 // We are done so set min young length to the maximum length (as |
|
657 // the result is assumed to be returned in min_young_length). |
|
658 min_young_length = max_young_length; |
|
659 } else { |
|
660 // The maximum possible number of young regions will not fit within |
|
661 // the target pause time so we'll search for the optimal |
|
662 // length. The loop invariants are: |
|
663 // |
|
664 // min_young_length < max_young_length |
|
665 // min_young_length is known to fit into the target pause time |
|
666 // max_young_length is known not to fit into the target pause time |
|
667 // |
|
668 // Going into the loop we know the above hold as we've just |
|
669 // checked them. Every time around the loop we check whether |
|
670 // the middle value between min_young_length and |
|
671 // max_young_length fits into the target pause time. If it |
|
672 // does, it becomes the new min. If it doesn't, it becomes |
|
673 // the new max. This way we maintain the loop invariants. |
|
674 |
|
675 assert(min_young_length < max_young_length, "invariant"); |
|
676 uint diff = (max_young_length - min_young_length) / 2; |
|
677 while (diff > 0) { |
|
678 uint young_length = min_young_length + diff; |
|
679 if (predict_will_fit(young_length, base_time_ms, |
|
680 base_free_regions, target_pause_time_ms)) { |
|
681 min_young_length = young_length; |
|
682 } else { |
|
683 max_young_length = young_length; |
|
684 } |
|
685 assert(min_young_length < max_young_length, "invariant"); |
|
686 diff = (max_young_length - min_young_length) / 2; |
|
687 } |
|
688 // The results is min_young_length which, according to the |
|
689 // loop invariants, should fit within the target pause time. |
|
690 |
|
691 // These are the post-conditions of the binary search above: |
|
692 assert(min_young_length < max_young_length, |
|
693 "otherwise we should have discovered that max_young_length " |
|
694 "fits into the pause target and not done the binary search"); |
|
695 assert(predict_will_fit(min_young_length, base_time_ms, |
|
696 base_free_regions, target_pause_time_ms), |
|
697 "min_young_length, the result of the binary search, should " |
|
698 "fit into the pause target"); |
|
699 assert(!predict_will_fit(min_young_length + 1, base_time_ms, |
|
700 base_free_regions, target_pause_time_ms), |
|
701 "min_young_length, the result of the binary search, should be " |
|
702 "optimal, so no larger length should fit into the pause target"); |
|
703 } |
|
704 } else { |
|
705 // Even the minimum length doesn't fit into the pause time |
|
706 // target, return it as the result nevertheless. |
|
707 } |
|
708 return base_min_length + min_young_length; |
|
709 } |
|
710 |
|
711 double G1CollectorPolicy::predict_survivor_regions_evac_time() { |
|
712 double survivor_regions_evac_time = 0.0; |
|
713 for (HeapRegion * r = _recorded_survivor_head; |
|
714 r != NULL && r != _recorded_survivor_tail->get_next_young_region(); |
|
715 r = r->get_next_young_region()) { |
|
716 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, gcs_are_young()); |
|
717 } |
|
718 return survivor_regions_evac_time; |
|
719 } |
|
720 |
|
721 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() { |
|
722 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); |
|
723 |
|
724 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths(); |
|
725 if (rs_lengths > _rs_lengths_prediction) { |
|
726 // add 10% to avoid having to recalculate often |
|
727 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; |
|
728 update_young_list_target_length(rs_lengths_prediction); |
|
729 } |
|
730 } |
|
731 |
|
732 |
|
733 |
|
734 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, |
|
735 bool is_tlab, |
|
736 bool* gc_overhead_limit_was_exceeded) { |
|
737 guarantee(false, "Not using this policy feature yet."); |
|
738 return NULL; |
|
739 } |
|
740 |
|
741 // This method controls how a collector handles one or more |
|
742 // of its generations being fully allocated. |
|
743 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, |
|
744 bool is_tlab) { |
|
745 guarantee(false, "Not using this policy feature yet."); |
|
746 return NULL; |
|
747 } |
|
748 |
|
749 |
|
750 #ifndef PRODUCT |
|
751 bool G1CollectorPolicy::verify_young_ages() { |
|
752 HeapRegion* head = _g1->young_list()->first_region(); |
|
753 return |
|
754 verify_young_ages(head, _short_lived_surv_rate_group); |
|
755 // also call verify_young_ages on any additional surv rate groups |
|
756 } |
|
757 |
|
758 bool |
|
759 G1CollectorPolicy::verify_young_ages(HeapRegion* head, |
|
760 SurvRateGroup *surv_rate_group) { |
|
761 guarantee( surv_rate_group != NULL, "pre-condition" ); |
|
762 |
|
763 const char* name = surv_rate_group->name(); |
|
764 bool ret = true; |
|
765 int prev_age = -1; |
|
766 |
|
767 for (HeapRegion* curr = head; |
|
768 curr != NULL; |
|
769 curr = curr->get_next_young_region()) { |
|
770 SurvRateGroup* group = curr->surv_rate_group(); |
|
771 if (group == NULL && !curr->is_survivor()) { |
|
772 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name); |
|
773 ret = false; |
|
774 } |
|
775 |
|
776 if (surv_rate_group == group) { |
|
777 int age = curr->age_in_surv_rate_group(); |
|
778 |
|
779 if (age < 0) { |
|
780 gclog_or_tty->print_cr("## %s: encountered negative age", name); |
|
781 ret = false; |
|
782 } |
|
783 |
|
784 if (age <= prev_age) { |
|
785 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing " |
|
786 "(%d, %d)", name, age, prev_age); |
|
787 ret = false; |
|
788 } |
|
789 prev_age = age; |
|
790 } |
|
791 } |
|
792 |
|
793 return ret; |
|
794 } |
|
795 #endif // PRODUCT |
|
796 |
|
797 void G1CollectorPolicy::record_full_collection_start() { |
|
798 _full_collection_start_sec = os::elapsedTime(); |
|
799 record_heap_size_info_at_start(true /* full */); |
|
800 // Release the future to-space so that it is available for compaction into. |
|
801 _g1->set_full_collection(); |
|
802 } |
|
803 |
|
804 void G1CollectorPolicy::record_full_collection_end() { |
|
805 // Consider this like a collection pause for the purposes of allocation |
|
806 // since last pause. |
|
807 double end_sec = os::elapsedTime(); |
|
808 double full_gc_time_sec = end_sec - _full_collection_start_sec; |
|
809 double full_gc_time_ms = full_gc_time_sec * 1000.0; |
|
810 |
|
811 _trace_gen1_time_data.record_full_collection(full_gc_time_ms); |
|
812 |
|
813 update_recent_gc_times(end_sec, full_gc_time_ms); |
|
814 |
|
815 _g1->clear_full_collection(); |
|
816 |
|
817 // "Nuke" the heuristics that control the young/mixed GC |
|
818 // transitions and make sure we start with young GCs after the Full GC. |
|
819 set_gcs_are_young(true); |
|
820 _last_young_gc = false; |
|
821 clear_initiate_conc_mark_if_possible(); |
|
822 clear_during_initial_mark_pause(); |
|
823 _in_marking_window = false; |
|
824 _in_marking_window_im = false; |
|
825 |
|
826 _short_lived_surv_rate_group->start_adding_regions(); |
|
827 // also call this on any additional surv rate groups |
|
828 |
|
829 record_survivor_regions(0, NULL, NULL); |
|
830 |
|
831 _free_regions_at_end_of_collection = _g1->free_regions(); |
|
832 // Reset survivors SurvRateGroup. |
|
833 _survivor_surv_rate_group->reset(); |
|
834 update_young_list_target_length(); |
|
835 _collectionSetChooser->clear(); |
|
836 } |
|
837 |
|
838 void G1CollectorPolicy::record_stop_world_start() { |
|
839 _stop_world_start = os::elapsedTime(); |
|
840 } |
|
841 |
|
842 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) { |
|
843 // We only need to do this here as the policy will only be applied |
|
844 // to the GC we're about to start. so, no point is calculating this |
|
845 // every time we calculate / recalculate the target young length. |
|
846 update_survivors_policy(); |
|
847 |
|
848 assert(_g1->used() == _g1->recalculate_used(), |
|
849 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT, |
|
850 _g1->used(), _g1->recalculate_used())); |
|
851 |
|
852 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; |
|
853 _trace_gen0_time_data.record_start_collection(s_w_t_ms); |
|
854 _stop_world_start = 0.0; |
|
855 |
|
856 record_heap_size_info_at_start(false /* full */); |
|
857 |
|
858 phase_times()->record_cur_collection_start_sec(start_time_sec); |
|
859 _pending_cards = _g1->pending_card_num(); |
|
860 |
|
861 _collection_set_bytes_used_before = 0; |
|
862 _bytes_copied_during_gc = 0; |
|
863 |
|
864 _last_gc_was_young = false; |
|
865 |
|
866 // do that for any other surv rate groups |
|
867 _short_lived_surv_rate_group->stop_adding_regions(); |
|
868 _survivors_age_table.clear(); |
|
869 |
|
870 assert( verify_young_ages(), "region age verification" ); |
|
871 } |
|
872 |
|
873 void G1CollectorPolicy::record_concurrent_mark_init_end(double |
|
874 mark_init_elapsed_time_ms) { |
|
875 _during_marking = true; |
|
876 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now"); |
|
877 clear_during_initial_mark_pause(); |
|
878 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; |
|
879 } |
|
880 |
|
881 void G1CollectorPolicy::record_concurrent_mark_remark_start() { |
|
882 _mark_remark_start_sec = os::elapsedTime(); |
|
883 _during_marking = false; |
|
884 } |
|
885 |
|
886 void G1CollectorPolicy::record_concurrent_mark_remark_end() { |
|
887 double end_time_sec = os::elapsedTime(); |
|
888 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; |
|
889 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); |
|
890 _cur_mark_stop_world_time_ms += elapsed_time_ms; |
|
891 _prev_collection_pause_end_ms += elapsed_time_ms; |
|
892 |
|
893 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true); |
|
894 } |
|
895 |
|
896 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { |
|
897 _mark_cleanup_start_sec = os::elapsedTime(); |
|
898 } |
|
899 |
|
900 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { |
|
901 _last_young_gc = true; |
|
902 _in_marking_window = false; |
|
903 } |
|
904 |
|
905 void G1CollectorPolicy::record_concurrent_pause() { |
|
906 if (_stop_world_start > 0.0) { |
|
907 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; |
|
908 _trace_gen0_time_data.record_yield_time(yield_ms); |
|
909 } |
|
910 } |
|
911 |
|
912 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) { |
|
913 if (_g1->concurrent_mark()->cmThread()->during_cycle()) { |
|
914 return false; |
|
915 } |
|
916 |
|
917 size_t marking_initiating_used_threshold = |
|
918 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent; |
|
919 size_t cur_used_bytes = _g1->non_young_capacity_bytes(); |
|
920 size_t alloc_byte_size = alloc_word_size * HeapWordSize; |
|
921 |
|
922 if ((cur_used_bytes + alloc_byte_size) > marking_initiating_used_threshold) { |
|
923 if (gcs_are_young() && !_last_young_gc) { |
|
924 ergo_verbose5(ErgoConcCycles, |
|
925 "request concurrent cycle initiation", |
|
926 ergo_format_reason("occupancy higher than threshold") |
|
927 ergo_format_byte("occupancy") |
|
928 ergo_format_byte("allocation request") |
|
929 ergo_format_byte_perc("threshold") |
|
930 ergo_format_str("source"), |
|
931 cur_used_bytes, |
|
932 alloc_byte_size, |
|
933 marking_initiating_used_threshold, |
|
934 (double) InitiatingHeapOccupancyPercent, |
|
935 source); |
|
936 return true; |
|
937 } else { |
|
938 ergo_verbose5(ErgoConcCycles, |
|
939 "do not request concurrent cycle initiation", |
|
940 ergo_format_reason("still doing mixed collections") |
|
941 ergo_format_byte("occupancy") |
|
942 ergo_format_byte("allocation request") |
|
943 ergo_format_byte_perc("threshold") |
|
944 ergo_format_str("source"), |
|
945 cur_used_bytes, |
|
946 alloc_byte_size, |
|
947 marking_initiating_used_threshold, |
|
948 (double) InitiatingHeapOccupancyPercent, |
|
949 source); |
|
950 } |
|
951 } |
|
952 |
|
953 return false; |
|
954 } |
|
955 |
|
956 // Anything below that is considered to be zero |
|
957 #define MIN_TIMER_GRANULARITY 0.0000001 |
|
958 |
|
959 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, EvacuationInfo& evacuation_info) { |
|
960 double end_time_sec = os::elapsedTime(); |
|
961 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(), |
|
962 "otherwise, the subtraction below does not make sense"); |
|
963 size_t rs_size = |
|
964 _cur_collection_pause_used_regions_at_start - cset_region_length(); |
|
965 size_t cur_used_bytes = _g1->used(); |
|
966 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); |
|
967 bool last_pause_included_initial_mark = false; |
|
968 bool update_stats = !_g1->evacuation_failed(); |
|
969 |
|
970 #ifndef PRODUCT |
|
971 if (G1YoungSurvRateVerbose) { |
|
972 gclog_or_tty->cr(); |
|
973 _short_lived_surv_rate_group->print(); |
|
974 // do that for any other surv rate groups too |
|
975 } |
|
976 #endif // PRODUCT |
|
977 |
|
978 last_pause_included_initial_mark = during_initial_mark_pause(); |
|
979 if (last_pause_included_initial_mark) { |
|
980 record_concurrent_mark_init_end(0.0); |
|
981 } else if (need_to_start_conc_mark("end of GC")) { |
|
982 // Note: this might have already been set, if during the last |
|
983 // pause we decided to start a cycle but at the beginning of |
|
984 // this pause we decided to postpone it. That's OK. |
|
985 set_initiate_conc_mark_if_possible(); |
|
986 } |
|
987 |
|
988 _mmu_tracker->add_pause(end_time_sec - pause_time_ms/1000.0, |
|
989 end_time_sec, false); |
|
990 |
|
991 evacuation_info.set_collectionset_used_before(_collection_set_bytes_used_before); |
|
992 evacuation_info.set_bytes_copied(_bytes_copied_during_gc); |
|
993 |
|
994 if (update_stats) { |
|
995 _trace_gen0_time_data.record_end_collection(pause_time_ms, phase_times()); |
|
996 // this is where we update the allocation rate of the application |
|
997 double app_time_ms = |
|
998 (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms); |
|
999 if (app_time_ms < MIN_TIMER_GRANULARITY) { |
|
1000 // This usually happens due to the timer not having the required |
|
1001 // granularity. Some Linuxes are the usual culprits. |
|
1002 // We'll just set it to something (arbitrarily) small. |
|
1003 app_time_ms = 1.0; |
|
1004 } |
|
1005 // We maintain the invariant that all objects allocated by mutator |
|
1006 // threads will be allocated out of eden regions. So, we can use |
|
1007 // the eden region number allocated since the previous GC to |
|
1008 // calculate the application's allocate rate. The only exception |
|
1009 // to that is humongous objects that are allocated separately. But |
|
1010 // given that humongous object allocations do not really affect |
|
1011 // either the pause's duration nor when the next pause will take |
|
1012 // place we can safely ignore them here. |
|
1013 uint regions_allocated = eden_cset_region_length(); |
|
1014 double alloc_rate_ms = (double) regions_allocated / app_time_ms; |
|
1015 _alloc_rate_ms_seq->add(alloc_rate_ms); |
|
1016 |
|
1017 double interval_ms = |
|
1018 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; |
|
1019 update_recent_gc_times(end_time_sec, pause_time_ms); |
|
1020 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; |
|
1021 if (recent_avg_pause_time_ratio() < 0.0 || |
|
1022 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) { |
|
1023 #ifndef PRODUCT |
|
1024 // Dump info to allow post-facto debugging |
|
1025 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds"); |
|
1026 gclog_or_tty->print_cr("-------------------------------------------"); |
|
1027 gclog_or_tty->print_cr("Recent GC Times (ms):"); |
|
1028 _recent_gc_times_ms->dump(); |
|
1029 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec); |
|
1030 _recent_prev_end_times_for_all_gcs_sec->dump(); |
|
1031 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f", |
|
1032 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio()); |
|
1033 // In debug mode, terminate the JVM if the user wants to debug at this point. |
|
1034 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above"); |
|
1035 #endif // !PRODUCT |
|
1036 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in |
|
1037 // CR 6902692 by redoing the manner in which the ratio is incrementally computed. |
|
1038 if (_recent_avg_pause_time_ratio < 0.0) { |
|
1039 _recent_avg_pause_time_ratio = 0.0; |
|
1040 } else { |
|
1041 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant"); |
|
1042 _recent_avg_pause_time_ratio = 1.0; |
|
1043 } |
|
1044 } |
|
1045 } |
|
1046 |
|
1047 bool new_in_marking_window = _in_marking_window; |
|
1048 bool new_in_marking_window_im = false; |
|
1049 if (during_initial_mark_pause()) { |
|
1050 new_in_marking_window = true; |
|
1051 new_in_marking_window_im = true; |
|
1052 } |
|
1053 |
|
1054 if (_last_young_gc) { |
|
1055 // This is supposed to to be the "last young GC" before we start |
|
1056 // doing mixed GCs. Here we decide whether to start mixed GCs or not. |
|
1057 |
|
1058 if (!last_pause_included_initial_mark) { |
|
1059 if (next_gc_should_be_mixed("start mixed GCs", |
|
1060 "do not start mixed GCs")) { |
|
1061 set_gcs_are_young(false); |
|
1062 } |
|
1063 } else { |
|
1064 ergo_verbose0(ErgoMixedGCs, |
|
1065 "do not start mixed GCs", |
|
1066 ergo_format_reason("concurrent cycle is about to start")); |
|
1067 } |
|
1068 _last_young_gc = false; |
|
1069 } |
|
1070 |
|
1071 if (!_last_gc_was_young) { |
|
1072 // This is a mixed GC. Here we decide whether to continue doing |
|
1073 // mixed GCs or not. |
|
1074 |
|
1075 if (!next_gc_should_be_mixed("continue mixed GCs", |
|
1076 "do not continue mixed GCs")) { |
|
1077 set_gcs_are_young(true); |
|
1078 } |
|
1079 } |
|
1080 |
|
1081 _short_lived_surv_rate_group->start_adding_regions(); |
|
1082 // do that for any other surv rate groupsx |
|
1083 |
|
1084 if (update_stats) { |
|
1085 double cost_per_card_ms = 0.0; |
|
1086 if (_pending_cards > 0) { |
|
1087 cost_per_card_ms = phase_times()->average_last_update_rs_time() / (double) _pending_cards; |
|
1088 _cost_per_card_ms_seq->add(cost_per_card_ms); |
|
1089 } |
|
1090 |
|
1091 size_t cards_scanned = _g1->cards_scanned(); |
|
1092 |
|
1093 double cost_per_entry_ms = 0.0; |
|
1094 if (cards_scanned > 10) { |
|
1095 cost_per_entry_ms = phase_times()->average_last_scan_rs_time() / (double) cards_scanned; |
|
1096 if (_last_gc_was_young) { |
|
1097 _cost_per_entry_ms_seq->add(cost_per_entry_ms); |
|
1098 } else { |
|
1099 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms); |
|
1100 } |
|
1101 } |
|
1102 |
|
1103 if (_max_rs_lengths > 0) { |
|
1104 double cards_per_entry_ratio = |
|
1105 (double) cards_scanned / (double) _max_rs_lengths; |
|
1106 if (_last_gc_was_young) { |
|
1107 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); |
|
1108 } else { |
|
1109 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); |
|
1110 } |
|
1111 } |
|
1112 |
|
1113 // This is defensive. For a while _max_rs_lengths could get |
|
1114 // smaller than _recorded_rs_lengths which was causing |
|
1115 // rs_length_diff to get very large and mess up the RSet length |
|
1116 // predictions. The reason was unsafe concurrent updates to the |
|
1117 // _inc_cset_recorded_rs_lengths field which the code below guards |
|
1118 // against (see CR 7118202). This bug has now been fixed (see CR |
|
1119 // 7119027). However, I'm still worried that |
|
1120 // _inc_cset_recorded_rs_lengths might still end up somewhat |
|
1121 // inaccurate. The concurrent refinement thread calculates an |
|
1122 // RSet's length concurrently with other CR threads updating it |
|
1123 // which might cause it to calculate the length incorrectly (if, |
|
1124 // say, it's in mid-coarsening). So I'll leave in the defensive |
|
1125 // conditional below just in case. |
|
1126 size_t rs_length_diff = 0; |
|
1127 if (_max_rs_lengths > _recorded_rs_lengths) { |
|
1128 rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; |
|
1129 } |
|
1130 _rs_length_diff_seq->add((double) rs_length_diff); |
|
1131 |
|
1132 size_t freed_bytes = _heap_used_bytes_before_gc - cur_used_bytes; |
|
1133 size_t copied_bytes = _collection_set_bytes_used_before - freed_bytes; |
|
1134 double cost_per_byte_ms = 0.0; |
|
1135 |
|
1136 if (copied_bytes > 0) { |
|
1137 cost_per_byte_ms = phase_times()->average_last_obj_copy_time() / (double) copied_bytes; |
|
1138 if (_in_marking_window) { |
|
1139 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); |
|
1140 } else { |
|
1141 _cost_per_byte_ms_seq->add(cost_per_byte_ms); |
|
1142 } |
|
1143 } |
|
1144 |
|
1145 double all_other_time_ms = pause_time_ms - |
|
1146 (phase_times()->average_last_update_rs_time() + phase_times()->average_last_scan_rs_time() |
|
1147 + phase_times()->average_last_obj_copy_time() + phase_times()->average_last_termination_time()); |
|
1148 |
|
1149 double young_other_time_ms = 0.0; |
|
1150 if (young_cset_region_length() > 0) { |
|
1151 young_other_time_ms = |
|
1152 phase_times()->young_cset_choice_time_ms() + |
|
1153 phase_times()->young_free_cset_time_ms(); |
|
1154 _young_other_cost_per_region_ms_seq->add(young_other_time_ms / |
|
1155 (double) young_cset_region_length()); |
|
1156 } |
|
1157 double non_young_other_time_ms = 0.0; |
|
1158 if (old_cset_region_length() > 0) { |
|
1159 non_young_other_time_ms = |
|
1160 phase_times()->non_young_cset_choice_time_ms() + |
|
1161 phase_times()->non_young_free_cset_time_ms(); |
|
1162 |
|
1163 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms / |
|
1164 (double) old_cset_region_length()); |
|
1165 } |
|
1166 |
|
1167 double constant_other_time_ms = all_other_time_ms - |
|
1168 (young_other_time_ms + non_young_other_time_ms); |
|
1169 _constant_other_time_ms_seq->add(constant_other_time_ms); |
|
1170 |
|
1171 double survival_ratio = 0.0; |
|
1172 if (_collection_set_bytes_used_before > 0) { |
|
1173 survival_ratio = (double) _bytes_copied_during_gc / |
|
1174 (double) _collection_set_bytes_used_before; |
|
1175 } |
|
1176 |
|
1177 _pending_cards_seq->add((double) _pending_cards); |
|
1178 _rs_lengths_seq->add((double) _max_rs_lengths); |
|
1179 } |
|
1180 |
|
1181 _in_marking_window = new_in_marking_window; |
|
1182 _in_marking_window_im = new_in_marking_window_im; |
|
1183 _free_regions_at_end_of_collection = _g1->free_regions(); |
|
1184 update_young_list_target_length(); |
|
1185 |
|
1186 // Note that _mmu_tracker->max_gc_time() returns the time in seconds. |
|
1187 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0; |
|
1188 adjust_concurrent_refinement(phase_times()->average_last_update_rs_time(), |
|
1189 phase_times()->sum_last_update_rs_processed_buffers(), update_rs_time_goal_ms); |
|
1190 |
|
1191 _collectionSetChooser->verify(); |
|
1192 } |
|
1193 |
|
1194 #define EXT_SIZE_FORMAT "%.1f%s" |
|
1195 #define EXT_SIZE_PARAMS(bytes) \ |
|
1196 byte_size_in_proper_unit((double)(bytes)), \ |
|
1197 proper_unit_for_byte_size((bytes)) |
|
1198 |
|
1199 void G1CollectorPolicy::record_heap_size_info_at_start(bool full) { |
|
1200 YoungList* young_list = _g1->young_list(); |
|
1201 _eden_used_bytes_before_gc = young_list->eden_used_bytes(); |
|
1202 _survivor_used_bytes_before_gc = young_list->survivor_used_bytes(); |
|
1203 _heap_capacity_bytes_before_gc = _g1->capacity(); |
|
1204 _heap_used_bytes_before_gc = _g1->used(); |
|
1205 _cur_collection_pause_used_regions_at_start = _g1->used_regions(); |
|
1206 |
|
1207 _eden_capacity_bytes_before_gc = |
|
1208 (_young_list_target_length * HeapRegion::GrainBytes) - _survivor_used_bytes_before_gc; |
|
1209 |
|
1210 if (full) { |
|
1211 _metaspace_used_bytes_before_gc = MetaspaceAux::used_bytes(); |
|
1212 } |
|
1213 } |
|
1214 |
|
1215 void G1CollectorPolicy::print_heap_transition() { |
|
1216 _g1->print_size_transition(gclog_or_tty, |
|
1217 _heap_used_bytes_before_gc, |
|
1218 _g1->used(), |
|
1219 _g1->capacity()); |
|
1220 } |
|
1221 |
|
1222 void G1CollectorPolicy::print_detailed_heap_transition(bool full) { |
|
1223 YoungList* young_list = _g1->young_list(); |
|
1224 |
|
1225 size_t eden_used_bytes_after_gc = young_list->eden_used_bytes(); |
|
1226 size_t survivor_used_bytes_after_gc = young_list->survivor_used_bytes(); |
|
1227 size_t heap_used_bytes_after_gc = _g1->used(); |
|
1228 |
|
1229 size_t heap_capacity_bytes_after_gc = _g1->capacity(); |
|
1230 size_t eden_capacity_bytes_after_gc = |
|
1231 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_used_bytes_after_gc; |
|
1232 |
|
1233 gclog_or_tty->print( |
|
1234 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") " |
|
1235 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" " |
|
1236 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->" |
|
1237 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]", |
|
1238 EXT_SIZE_PARAMS(_eden_used_bytes_before_gc), |
|
1239 EXT_SIZE_PARAMS(_eden_capacity_bytes_before_gc), |
|
1240 EXT_SIZE_PARAMS(eden_used_bytes_after_gc), |
|
1241 EXT_SIZE_PARAMS(eden_capacity_bytes_after_gc), |
|
1242 EXT_SIZE_PARAMS(_survivor_used_bytes_before_gc), |
|
1243 EXT_SIZE_PARAMS(survivor_used_bytes_after_gc), |
|
1244 EXT_SIZE_PARAMS(_heap_used_bytes_before_gc), |
|
1245 EXT_SIZE_PARAMS(_heap_capacity_bytes_before_gc), |
|
1246 EXT_SIZE_PARAMS(heap_used_bytes_after_gc), |
|
1247 EXT_SIZE_PARAMS(heap_capacity_bytes_after_gc)); |
|
1248 |
|
1249 if (full) { |
|
1250 MetaspaceAux::print_metaspace_change(_metaspace_used_bytes_before_gc); |
|
1251 } |
|
1252 |
|
1253 gclog_or_tty->cr(); |
|
1254 } |
|
1255 |
|
1256 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time, |
|
1257 double update_rs_processed_buffers, |
|
1258 double goal_ms) { |
|
1259 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); |
|
1260 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine(); |
|
1261 |
|
1262 if (G1UseAdaptiveConcRefinement) { |
|
1263 const int k_gy = 3, k_gr = 6; |
|
1264 const double inc_k = 1.1, dec_k = 0.9; |
|
1265 |
|
1266 int g = cg1r->green_zone(); |
|
1267 if (update_rs_time > goal_ms) { |
|
1268 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing. |
|
1269 } else { |
|
1270 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) { |
|
1271 g = (int)MAX2(g * inc_k, g + 1.0); |
|
1272 } |
|
1273 } |
|
1274 // Change the refinement threads params |
|
1275 cg1r->set_green_zone(g); |
|
1276 cg1r->set_yellow_zone(g * k_gy); |
|
1277 cg1r->set_red_zone(g * k_gr); |
|
1278 cg1r->reinitialize_threads(); |
|
1279 |
|
1280 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1); |
|
1281 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta, |
|
1282 cg1r->yellow_zone()); |
|
1283 // Change the barrier params |
|
1284 dcqs.set_process_completed_threshold(processing_threshold); |
|
1285 dcqs.set_max_completed_queue(cg1r->red_zone()); |
|
1286 } |
|
1287 |
|
1288 int curr_queue_size = dcqs.completed_buffers_num(); |
|
1289 if (curr_queue_size >= cg1r->yellow_zone()) { |
|
1290 dcqs.set_completed_queue_padding(curr_queue_size); |
|
1291 } else { |
|
1292 dcqs.set_completed_queue_padding(0); |
|
1293 } |
|
1294 dcqs.notify_if_necessary(); |
|
1295 } |
|
1296 |
|
1297 double |
|
1298 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, |
|
1299 size_t scanned_cards) { |
|
1300 return |
|
1301 predict_rs_update_time_ms(pending_cards) + |
|
1302 predict_rs_scan_time_ms(scanned_cards) + |
|
1303 predict_constant_other_time_ms(); |
|
1304 } |
|
1305 |
|
1306 double |
|
1307 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) { |
|
1308 size_t rs_length = predict_rs_length_diff(); |
|
1309 size_t card_num; |
|
1310 if (gcs_are_young()) { |
|
1311 card_num = predict_young_card_num(rs_length); |
|
1312 } else { |
|
1313 card_num = predict_non_young_card_num(rs_length); |
|
1314 } |
|
1315 return predict_base_elapsed_time_ms(pending_cards, card_num); |
|
1316 } |
|
1317 |
|
1318 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) { |
|
1319 size_t bytes_to_copy; |
|
1320 if (hr->is_marked()) |
|
1321 bytes_to_copy = hr->max_live_bytes(); |
|
1322 else { |
|
1323 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant"); |
|
1324 int age = hr->age_in_surv_rate_group(); |
|
1325 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group()); |
|
1326 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate); |
|
1327 } |
|
1328 return bytes_to_copy; |
|
1329 } |
|
1330 |
|
1331 double |
|
1332 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, |
|
1333 bool for_young_gc) { |
|
1334 size_t rs_length = hr->rem_set()->occupied(); |
|
1335 size_t card_num; |
|
1336 |
|
1337 // Predicting the number of cards is based on which type of GC |
|
1338 // we're predicting for. |
|
1339 if (for_young_gc) { |
|
1340 card_num = predict_young_card_num(rs_length); |
|
1341 } else { |
|
1342 card_num = predict_non_young_card_num(rs_length); |
|
1343 } |
|
1344 size_t bytes_to_copy = predict_bytes_to_copy(hr); |
|
1345 |
|
1346 double region_elapsed_time_ms = |
|
1347 predict_rs_scan_time_ms(card_num) + |
|
1348 predict_object_copy_time_ms(bytes_to_copy); |
|
1349 |
|
1350 // The prediction of the "other" time for this region is based |
|
1351 // upon the region type and NOT the GC type. |
|
1352 if (hr->is_young()) { |
|
1353 region_elapsed_time_ms += predict_young_other_time_ms(1); |
|
1354 } else { |
|
1355 region_elapsed_time_ms += predict_non_young_other_time_ms(1); |
|
1356 } |
|
1357 return region_elapsed_time_ms; |
|
1358 } |
|
1359 |
|
1360 void |
|
1361 G1CollectorPolicy::init_cset_region_lengths(uint eden_cset_region_length, |
|
1362 uint survivor_cset_region_length) { |
|
1363 _eden_cset_region_length = eden_cset_region_length; |
|
1364 _survivor_cset_region_length = survivor_cset_region_length; |
|
1365 _old_cset_region_length = 0; |
|
1366 } |
|
1367 |
|
1368 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) { |
|
1369 _recorded_rs_lengths = rs_lengths; |
|
1370 } |
|
1371 |
|
1372 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, |
|
1373 double elapsed_ms) { |
|
1374 _recent_gc_times_ms->add(elapsed_ms); |
|
1375 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); |
|
1376 _prev_collection_pause_end_ms = end_time_sec * 1000.0; |
|
1377 } |
|
1378 |
|
1379 size_t G1CollectorPolicy::expansion_amount() { |
|
1380 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0; |
|
1381 double threshold = _gc_overhead_perc; |
|
1382 if (recent_gc_overhead > threshold) { |
|
1383 // We will double the existing space, or take |
|
1384 // G1ExpandByPercentOfAvailable % of the available expansion |
|
1385 // space, whichever is smaller, bounded below by a minimum |
|
1386 // expansion (unless that's all that's left.) |
|
1387 const size_t min_expand_bytes = 1*M; |
|
1388 size_t reserved_bytes = _g1->max_capacity(); |
|
1389 size_t committed_bytes = _g1->capacity(); |
|
1390 size_t uncommitted_bytes = reserved_bytes - committed_bytes; |
|
1391 size_t expand_bytes; |
|
1392 size_t expand_bytes_via_pct = |
|
1393 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100; |
|
1394 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); |
|
1395 expand_bytes = MAX2(expand_bytes, min_expand_bytes); |
|
1396 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); |
|
1397 |
|
1398 ergo_verbose5(ErgoHeapSizing, |
|
1399 "attempt heap expansion", |
|
1400 ergo_format_reason("recent GC overhead higher than " |
|
1401 "threshold after GC") |
|
1402 ergo_format_perc("recent GC overhead") |
|
1403 ergo_format_perc("threshold") |
|
1404 ergo_format_byte("uncommitted") |
|
1405 ergo_format_byte_perc("calculated expansion amount"), |
|
1406 recent_gc_overhead, threshold, |
|
1407 uncommitted_bytes, |
|
1408 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable); |
|
1409 |
|
1410 return expand_bytes; |
|
1411 } else { |
|
1412 return 0; |
|
1413 } |
|
1414 } |
|
1415 |
|
1416 void G1CollectorPolicy::print_tracing_info() const { |
|
1417 _trace_gen0_time_data.print(); |
|
1418 _trace_gen1_time_data.print(); |
|
1419 } |
|
1420 |
|
1421 void G1CollectorPolicy::print_yg_surv_rate_info() const { |
|
1422 #ifndef PRODUCT |
|
1423 _short_lived_surv_rate_group->print_surv_rate_summary(); |
|
1424 // add this call for any other surv rate groups |
|
1425 #endif // PRODUCT |
|
1426 } |
|
1427 |
|
1428 uint G1CollectorPolicy::max_regions(int purpose) { |
|
1429 switch (purpose) { |
|
1430 case GCAllocForSurvived: |
|
1431 return _max_survivor_regions; |
|
1432 case GCAllocForTenured: |
|
1433 return REGIONS_UNLIMITED; |
|
1434 default: |
|
1435 ShouldNotReachHere(); |
|
1436 return REGIONS_UNLIMITED; |
|
1437 }; |
|
1438 } |
|
1439 |
|
1440 void G1CollectorPolicy::update_max_gc_locker_expansion() { |
|
1441 uint expansion_region_num = 0; |
|
1442 if (GCLockerEdenExpansionPercent > 0) { |
|
1443 double perc = (double) GCLockerEdenExpansionPercent / 100.0; |
|
1444 double expansion_region_num_d = perc * (double) _young_list_target_length; |
|
1445 // We use ceiling so that if expansion_region_num_d is > 0.0 (but |
|
1446 // less than 1.0) we'll get 1. |
|
1447 expansion_region_num = (uint) ceil(expansion_region_num_d); |
|
1448 } else { |
|
1449 assert(expansion_region_num == 0, "sanity"); |
|
1450 } |
|
1451 _young_list_max_length = _young_list_target_length + expansion_region_num; |
|
1452 assert(_young_list_target_length <= _young_list_max_length, "post-condition"); |
|
1453 } |
|
1454 |
|
1455 // Calculates survivor space parameters. |
|
1456 void G1CollectorPolicy::update_survivors_policy() { |
|
1457 double max_survivor_regions_d = |
|
1458 (double) _young_list_target_length / (double) SurvivorRatio; |
|
1459 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but |
|
1460 // smaller than 1.0) we'll get 1. |
|
1461 _max_survivor_regions = (uint) ceil(max_survivor_regions_d); |
|
1462 |
|
1463 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold( |
|
1464 HeapRegion::GrainWords * _max_survivor_regions); |
|
1465 } |
|
1466 |
|
1467 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle( |
|
1468 GCCause::Cause gc_cause) { |
|
1469 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); |
|
1470 if (!during_cycle) { |
|
1471 ergo_verbose1(ErgoConcCycles, |
|
1472 "request concurrent cycle initiation", |
|
1473 ergo_format_reason("requested by GC cause") |
|
1474 ergo_format_str("GC cause"), |
|
1475 GCCause::to_string(gc_cause)); |
|
1476 set_initiate_conc_mark_if_possible(); |
|
1477 return true; |
|
1478 } else { |
|
1479 ergo_verbose1(ErgoConcCycles, |
|
1480 "do not request concurrent cycle initiation", |
|
1481 ergo_format_reason("concurrent cycle already in progress") |
|
1482 ergo_format_str("GC cause"), |
|
1483 GCCause::to_string(gc_cause)); |
|
1484 return false; |
|
1485 } |
|
1486 } |
|
1487 |
|
1488 void |
|
1489 G1CollectorPolicy::decide_on_conc_mark_initiation() { |
|
1490 // We are about to decide on whether this pause will be an |
|
1491 // initial-mark pause. |
|
1492 |
|
1493 // First, during_initial_mark_pause() should not be already set. We |
|
1494 // will set it here if we have to. However, it should be cleared by |
|
1495 // the end of the pause (it's only set for the duration of an |
|
1496 // initial-mark pause). |
|
1497 assert(!during_initial_mark_pause(), "pre-condition"); |
|
1498 |
|
1499 if (initiate_conc_mark_if_possible()) { |
|
1500 // We had noticed on a previous pause that the heap occupancy has |
|
1501 // gone over the initiating threshold and we should start a |
|
1502 // concurrent marking cycle. So we might initiate one. |
|
1503 |
|
1504 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle(); |
|
1505 if (!during_cycle) { |
|
1506 // The concurrent marking thread is not "during a cycle", i.e., |
|
1507 // it has completed the last one. So we can go ahead and |
|
1508 // initiate a new cycle. |
|
1509 |
|
1510 set_during_initial_mark_pause(); |
|
1511 // We do not allow mixed GCs during marking. |
|
1512 if (!gcs_are_young()) { |
|
1513 set_gcs_are_young(true); |
|
1514 ergo_verbose0(ErgoMixedGCs, |
|
1515 "end mixed GCs", |
|
1516 ergo_format_reason("concurrent cycle is about to start")); |
|
1517 } |
|
1518 |
|
1519 // And we can now clear initiate_conc_mark_if_possible() as |
|
1520 // we've already acted on it. |
|
1521 clear_initiate_conc_mark_if_possible(); |
|
1522 |
|
1523 ergo_verbose0(ErgoConcCycles, |
|
1524 "initiate concurrent cycle", |
|
1525 ergo_format_reason("concurrent cycle initiation requested")); |
|
1526 } else { |
|
1527 // The concurrent marking thread is still finishing up the |
|
1528 // previous cycle. If we start one right now the two cycles |
|
1529 // overlap. In particular, the concurrent marking thread might |
|
1530 // be in the process of clearing the next marking bitmap (which |
|
1531 // we will use for the next cycle if we start one). Starting a |
|
1532 // cycle now will be bad given that parts of the marking |
|
1533 // information might get cleared by the marking thread. And we |
|
1534 // cannot wait for the marking thread to finish the cycle as it |
|
1535 // periodically yields while clearing the next marking bitmap |
|
1536 // and, if it's in a yield point, it's waiting for us to |
|
1537 // finish. So, at this point we will not start a cycle and we'll |
|
1538 // let the concurrent marking thread complete the last one. |
|
1539 ergo_verbose0(ErgoConcCycles, |
|
1540 "do not initiate concurrent cycle", |
|
1541 ergo_format_reason("concurrent cycle already in progress")); |
|
1542 } |
|
1543 } |
|
1544 } |
|
1545 |
|
1546 class KnownGarbageClosure: public HeapRegionClosure { |
|
1547 G1CollectedHeap* _g1h; |
|
1548 CollectionSetChooser* _hrSorted; |
|
1549 |
|
1550 public: |
|
1551 KnownGarbageClosure(CollectionSetChooser* hrSorted) : |
|
1552 _g1h(G1CollectedHeap::heap()), _hrSorted(hrSorted) { } |
|
1553 |
|
1554 bool doHeapRegion(HeapRegion* r) { |
|
1555 // We only include humongous regions in collection |
|
1556 // sets when concurrent mark shows that their contained object is |
|
1557 // unreachable. |
|
1558 |
|
1559 // Do we have any marking information for this region? |
|
1560 if (r->is_marked()) { |
|
1561 // We will skip any region that's currently used as an old GC |
|
1562 // alloc region (we should not consider those for collection |
|
1563 // before we fill them up). |
|
1564 if (_hrSorted->should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { |
|
1565 _hrSorted->add_region(r); |
|
1566 } |
|
1567 } |
|
1568 return false; |
|
1569 } |
|
1570 }; |
|
1571 |
|
1572 class ParKnownGarbageHRClosure: public HeapRegionClosure { |
|
1573 G1CollectedHeap* _g1h; |
|
1574 CSetChooserParUpdater _cset_updater; |
|
1575 |
|
1576 public: |
|
1577 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, |
|
1578 uint chunk_size) : |
|
1579 _g1h(G1CollectedHeap::heap()), |
|
1580 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } |
|
1581 |
|
1582 bool doHeapRegion(HeapRegion* r) { |
|
1583 // Do we have any marking information for this region? |
|
1584 if (r->is_marked()) { |
|
1585 // We will skip any region that's currently used as an old GC |
|
1586 // alloc region (we should not consider those for collection |
|
1587 // before we fill them up). |
|
1588 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { |
|
1589 _cset_updater.add_region(r); |
|
1590 } |
|
1591 } |
|
1592 return false; |
|
1593 } |
|
1594 }; |
|
1595 |
|
1596 class ParKnownGarbageTask: public AbstractGangTask { |
|
1597 CollectionSetChooser* _hrSorted; |
|
1598 uint _chunk_size; |
|
1599 G1CollectedHeap* _g1; |
|
1600 public: |
|
1601 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size) : |
|
1602 AbstractGangTask("ParKnownGarbageTask"), |
|
1603 _hrSorted(hrSorted), _chunk_size(chunk_size), |
|
1604 _g1(G1CollectedHeap::heap()) { } |
|
1605 |
|
1606 void work(uint worker_id) { |
|
1607 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size); |
|
1608 |
|
1609 // Back to zero for the claim value. |
|
1610 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, worker_id, |
|
1611 _g1->workers()->active_workers(), |
|
1612 HeapRegion::InitialClaimValue); |
|
1613 } |
|
1614 }; |
|
1615 |
|
1616 void |
|
1617 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) { |
|
1618 _collectionSetChooser->clear(); |
|
1619 |
|
1620 uint region_num = _g1->n_regions(); |
|
1621 if (G1CollectedHeap::use_parallel_gc_threads()) { |
|
1622 const uint OverpartitionFactor = 4; |
|
1623 uint WorkUnit; |
|
1624 // The use of MinChunkSize = 8 in the original code |
|
1625 // causes some assertion failures when the total number of |
|
1626 // region is less than 8. The code here tries to fix that. |
|
1627 // Should the original code also be fixed? |
|
1628 if (no_of_gc_threads > 0) { |
|
1629 const uint MinWorkUnit = MAX2(region_num / no_of_gc_threads, 1U); |
|
1630 WorkUnit = MAX2(region_num / (no_of_gc_threads * OverpartitionFactor), |
|
1631 MinWorkUnit); |
|
1632 } else { |
|
1633 assert(no_of_gc_threads > 0, |
|
1634 "The active gc workers should be greater than 0"); |
|
1635 // In a product build do something reasonable to avoid a crash. |
|
1636 const uint MinWorkUnit = MAX2(region_num / (uint) ParallelGCThreads, 1U); |
|
1637 WorkUnit = |
|
1638 MAX2(region_num / (uint) (ParallelGCThreads * OverpartitionFactor), |
|
1639 MinWorkUnit); |
|
1640 } |
|
1641 _collectionSetChooser->prepare_for_par_region_addition(_g1->n_regions(), |
|
1642 WorkUnit); |
|
1643 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser, |
|
1644 (int) WorkUnit); |
|
1645 _g1->workers()->run_task(&parKnownGarbageTask); |
|
1646 |
|
1647 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue), |
|
1648 "sanity check"); |
|
1649 } else { |
|
1650 KnownGarbageClosure knownGarbagecl(_collectionSetChooser); |
|
1651 _g1->heap_region_iterate(&knownGarbagecl); |
|
1652 } |
|
1653 |
|
1654 _collectionSetChooser->sort_regions(); |
|
1655 |
|
1656 double end_sec = os::elapsedTime(); |
|
1657 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0; |
|
1658 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); |
|
1659 _cur_mark_stop_world_time_ms += elapsed_time_ms; |
|
1660 _prev_collection_pause_end_ms += elapsed_time_ms; |
|
1661 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true); |
|
1662 } |
|
1663 |
|
1664 // Add the heap region at the head of the non-incremental collection set |
|
1665 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) { |
|
1666 assert(_inc_cset_build_state == Active, "Precondition"); |
|
1667 assert(!hr->is_young(), "non-incremental add of young region"); |
|
1668 |
|
1669 assert(!hr->in_collection_set(), "should not already be in the CSet"); |
|
1670 hr->set_in_collection_set(true); |
|
1671 hr->set_next_in_collection_set(_collection_set); |
|
1672 _collection_set = hr; |
|
1673 _collection_set_bytes_used_before += hr->used(); |
|
1674 _g1->register_region_with_in_cset_fast_test(hr); |
|
1675 size_t rs_length = hr->rem_set()->occupied(); |
|
1676 _recorded_rs_lengths += rs_length; |
|
1677 _old_cset_region_length += 1; |
|
1678 } |
|
1679 |
|
1680 // Initialize the per-collection-set information |
|
1681 void G1CollectorPolicy::start_incremental_cset_building() { |
|
1682 assert(_inc_cset_build_state == Inactive, "Precondition"); |
|
1683 |
|
1684 _inc_cset_head = NULL; |
|
1685 _inc_cset_tail = NULL; |
|
1686 _inc_cset_bytes_used_before = 0; |
|
1687 |
|
1688 _inc_cset_max_finger = 0; |
|
1689 _inc_cset_recorded_rs_lengths = 0; |
|
1690 _inc_cset_recorded_rs_lengths_diffs = 0; |
|
1691 _inc_cset_predicted_elapsed_time_ms = 0.0; |
|
1692 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; |
|
1693 _inc_cset_build_state = Active; |
|
1694 } |
|
1695 |
|
1696 void G1CollectorPolicy::finalize_incremental_cset_building() { |
|
1697 assert(_inc_cset_build_state == Active, "Precondition"); |
|
1698 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); |
|
1699 |
|
1700 // The two "main" fields, _inc_cset_recorded_rs_lengths and |
|
1701 // _inc_cset_predicted_elapsed_time_ms, are updated by the thread |
|
1702 // that adds a new region to the CSet. Further updates by the |
|
1703 // concurrent refinement thread that samples the young RSet lengths |
|
1704 // are accumulated in the *_diffs fields. Here we add the diffs to |
|
1705 // the "main" fields. |
|
1706 |
|
1707 if (_inc_cset_recorded_rs_lengths_diffs >= 0) { |
|
1708 _inc_cset_recorded_rs_lengths += _inc_cset_recorded_rs_lengths_diffs; |
|
1709 } else { |
|
1710 // This is defensive. The diff should in theory be always positive |
|
1711 // as RSets can only grow between GCs. However, given that we |
|
1712 // sample their size concurrently with other threads updating them |
|
1713 // it's possible that we might get the wrong size back, which |
|
1714 // could make the calculations somewhat inaccurate. |
|
1715 size_t diffs = (size_t) (-_inc_cset_recorded_rs_lengths_diffs); |
|
1716 if (_inc_cset_recorded_rs_lengths >= diffs) { |
|
1717 _inc_cset_recorded_rs_lengths -= diffs; |
|
1718 } else { |
|
1719 _inc_cset_recorded_rs_lengths = 0; |
|
1720 } |
|
1721 } |
|
1722 _inc_cset_predicted_elapsed_time_ms += |
|
1723 _inc_cset_predicted_elapsed_time_ms_diffs; |
|
1724 |
|
1725 _inc_cset_recorded_rs_lengths_diffs = 0; |
|
1726 _inc_cset_predicted_elapsed_time_ms_diffs = 0.0; |
|
1727 } |
|
1728 |
|
1729 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) { |
|
1730 // This routine is used when: |
|
1731 // * adding survivor regions to the incremental cset at the end of an |
|
1732 // evacuation pause, |
|
1733 // * adding the current allocation region to the incremental cset |
|
1734 // when it is retired, and |
|
1735 // * updating existing policy information for a region in the |
|
1736 // incremental cset via young list RSet sampling. |
|
1737 // Therefore this routine may be called at a safepoint by the |
|
1738 // VM thread, or in-between safepoints by mutator threads (when |
|
1739 // retiring the current allocation region) or a concurrent |
|
1740 // refine thread (RSet sampling). |
|
1741 |
|
1742 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young()); |
|
1743 size_t used_bytes = hr->used(); |
|
1744 _inc_cset_recorded_rs_lengths += rs_length; |
|
1745 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms; |
|
1746 _inc_cset_bytes_used_before += used_bytes; |
|
1747 |
|
1748 // Cache the values we have added to the aggregated informtion |
|
1749 // in the heap region in case we have to remove this region from |
|
1750 // the incremental collection set, or it is updated by the |
|
1751 // rset sampling code |
|
1752 hr->set_recorded_rs_length(rs_length); |
|
1753 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms); |
|
1754 } |
|
1755 |
|
1756 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, |
|
1757 size_t new_rs_length) { |
|
1758 // Update the CSet information that is dependent on the new RS length |
|
1759 assert(hr->is_young(), "Precondition"); |
|
1760 assert(!SafepointSynchronize::is_at_safepoint(), |
|
1761 "should not be at a safepoint"); |
|
1762 |
|
1763 // We could have updated _inc_cset_recorded_rs_lengths and |
|
1764 // _inc_cset_predicted_elapsed_time_ms directly but we'd need to do |
|
1765 // that atomically, as this code is executed by a concurrent |
|
1766 // refinement thread, potentially concurrently with a mutator thread |
|
1767 // allocating a new region and also updating the same fields. To |
|
1768 // avoid the atomic operations we accumulate these updates on two |
|
1769 // separate fields (*_diffs) and we'll just add them to the "main" |
|
1770 // fields at the start of a GC. |
|
1771 |
|
1772 ssize_t old_rs_length = (ssize_t) hr->recorded_rs_length(); |
|
1773 ssize_t rs_lengths_diff = (ssize_t) new_rs_length - old_rs_length; |
|
1774 _inc_cset_recorded_rs_lengths_diffs += rs_lengths_diff; |
|
1775 |
|
1776 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms(); |
|
1777 double new_region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young()); |
|
1778 double elapsed_ms_diff = new_region_elapsed_time_ms - old_elapsed_time_ms; |
|
1779 _inc_cset_predicted_elapsed_time_ms_diffs += elapsed_ms_diff; |
|
1780 |
|
1781 hr->set_recorded_rs_length(new_rs_length); |
|
1782 hr->set_predicted_elapsed_time_ms(new_region_elapsed_time_ms); |
|
1783 } |
|
1784 |
|
1785 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) { |
|
1786 assert(hr->is_young(), "invariant"); |
|
1787 assert(hr->young_index_in_cset() > -1, "should have already been set"); |
|
1788 assert(_inc_cset_build_state == Active, "Precondition"); |
|
1789 |
|
1790 // We need to clear and set the cached recorded/cached collection set |
|
1791 // information in the heap region here (before the region gets added |
|
1792 // to the collection set). An individual heap region's cached values |
|
1793 // are calculated, aggregated with the policy collection set info, |
|
1794 // and cached in the heap region here (initially) and (subsequently) |
|
1795 // by the Young List sampling code. |
|
1796 |
|
1797 size_t rs_length = hr->rem_set()->occupied(); |
|
1798 add_to_incremental_cset_info(hr, rs_length); |
|
1799 |
|
1800 HeapWord* hr_end = hr->end(); |
|
1801 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end); |
|
1802 |
|
1803 assert(!hr->in_collection_set(), "invariant"); |
|
1804 hr->set_in_collection_set(true); |
|
1805 assert( hr->next_in_collection_set() == NULL, "invariant"); |
|
1806 |
|
1807 _g1->register_region_with_in_cset_fast_test(hr); |
|
1808 } |
|
1809 |
|
1810 // Add the region at the RHS of the incremental cset |
|
1811 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) { |
|
1812 // We should only ever be appending survivors at the end of a pause |
|
1813 assert( hr->is_survivor(), "Logic"); |
|
1814 |
|
1815 // Do the 'common' stuff |
|
1816 add_region_to_incremental_cset_common(hr); |
|
1817 |
|
1818 // Now add the region at the right hand side |
|
1819 if (_inc_cset_tail == NULL) { |
|
1820 assert(_inc_cset_head == NULL, "invariant"); |
|
1821 _inc_cset_head = hr; |
|
1822 } else { |
|
1823 _inc_cset_tail->set_next_in_collection_set(hr); |
|
1824 } |
|
1825 _inc_cset_tail = hr; |
|
1826 } |
|
1827 |
|
1828 // Add the region to the LHS of the incremental cset |
|
1829 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) { |
|
1830 // Survivors should be added to the RHS at the end of a pause |
|
1831 assert(!hr->is_survivor(), "Logic"); |
|
1832 |
|
1833 // Do the 'common' stuff |
|
1834 add_region_to_incremental_cset_common(hr); |
|
1835 |
|
1836 // Add the region at the left hand side |
|
1837 hr->set_next_in_collection_set(_inc_cset_head); |
|
1838 if (_inc_cset_head == NULL) { |
|
1839 assert(_inc_cset_tail == NULL, "Invariant"); |
|
1840 _inc_cset_tail = hr; |
|
1841 } |
|
1842 _inc_cset_head = hr; |
|
1843 } |
|
1844 |
|
1845 #ifndef PRODUCT |
|
1846 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) { |
|
1847 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be"); |
|
1848 |
|
1849 st->print_cr("\nCollection_set:"); |
|
1850 HeapRegion* csr = list_head; |
|
1851 while (csr != NULL) { |
|
1852 HeapRegion* next = csr->next_in_collection_set(); |
|
1853 assert(csr->in_collection_set(), "bad CS"); |
|
1854 st->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d", |
|
1855 HR_FORMAT_PARAMS(csr), |
|
1856 csr->prev_top_at_mark_start(), csr->next_top_at_mark_start(), |
|
1857 csr->age_in_surv_rate_group_cond()); |
|
1858 csr = next; |
|
1859 } |
|
1860 } |
|
1861 #endif // !PRODUCT |
|
1862 |
|
1863 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) { |
|
1864 // Returns the given amount of reclaimable bytes (that represents |
|
1865 // the amount of reclaimable space still to be collected) as a |
|
1866 // percentage of the current heap capacity. |
|
1867 size_t capacity_bytes = _g1->capacity(); |
|
1868 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes; |
|
1869 } |
|
1870 |
|
1871 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str, |
|
1872 const char* false_action_str) { |
|
1873 CollectionSetChooser* cset_chooser = _collectionSetChooser; |
|
1874 if (cset_chooser->is_empty()) { |
|
1875 ergo_verbose0(ErgoMixedGCs, |
|
1876 false_action_str, |
|
1877 ergo_format_reason("candidate old regions not available")); |
|
1878 return false; |
|
1879 } |
|
1880 |
|
1881 // Is the amount of uncollected reclaimable space above G1HeapWastePercent? |
|
1882 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); |
|
1883 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); |
|
1884 double threshold = (double) G1HeapWastePercent; |
|
1885 if (reclaimable_perc <= threshold) { |
|
1886 ergo_verbose4(ErgoMixedGCs, |
|
1887 false_action_str, |
|
1888 ergo_format_reason("reclaimable percentage not over threshold") |
|
1889 ergo_format_region("candidate old regions") |
|
1890 ergo_format_byte_perc("reclaimable") |
|
1891 ergo_format_perc("threshold"), |
|
1892 cset_chooser->remaining_regions(), |
|
1893 reclaimable_bytes, |
|
1894 reclaimable_perc, threshold); |
|
1895 return false; |
|
1896 } |
|
1897 |
|
1898 ergo_verbose4(ErgoMixedGCs, |
|
1899 true_action_str, |
|
1900 ergo_format_reason("candidate old regions available") |
|
1901 ergo_format_region("candidate old regions") |
|
1902 ergo_format_byte_perc("reclaimable") |
|
1903 ergo_format_perc("threshold"), |
|
1904 cset_chooser->remaining_regions(), |
|
1905 reclaimable_bytes, |
|
1906 reclaimable_perc, threshold); |
|
1907 return true; |
|
1908 } |
|
1909 |
|
1910 uint G1CollectorPolicy::calc_min_old_cset_length() { |
|
1911 // The min old CSet region bound is based on the maximum desired |
|
1912 // number of mixed GCs after a cycle. I.e., even if some old regions |
|
1913 // look expensive, we should add them to the CSet anyway to make |
|
1914 // sure we go through the available old regions in no more than the |
|
1915 // maximum desired number of mixed GCs. |
|
1916 // |
|
1917 // The calculation is based on the number of marked regions we added |
|
1918 // to the CSet chooser in the first place, not how many remain, so |
|
1919 // that the result is the same during all mixed GCs that follow a cycle. |
|
1920 |
|
1921 const size_t region_num = (size_t) _collectionSetChooser->length(); |
|
1922 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1); |
|
1923 size_t result = region_num / gc_num; |
|
1924 // emulate ceiling |
|
1925 if (result * gc_num < region_num) { |
|
1926 result += 1; |
|
1927 } |
|
1928 return (uint) result; |
|
1929 } |
|
1930 |
|
1931 uint G1CollectorPolicy::calc_max_old_cset_length() { |
|
1932 // The max old CSet region bound is based on the threshold expressed |
|
1933 // as a percentage of the heap size. I.e., it should bound the |
|
1934 // number of old regions added to the CSet irrespective of how many |
|
1935 // of them are available. |
|
1936 |
|
1937 G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
|
1938 const size_t region_num = g1h->n_regions(); |
|
1939 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent; |
|
1940 size_t result = region_num * perc / 100; |
|
1941 // emulate ceiling |
|
1942 if (100 * result < region_num * perc) { |
|
1943 result += 1; |
|
1944 } |
|
1945 return (uint) result; |
|
1946 } |
|
1947 |
|
1948 |
|
1949 void G1CollectorPolicy::finalize_cset(double target_pause_time_ms, EvacuationInfo& evacuation_info) { |
|
1950 double young_start_time_sec = os::elapsedTime(); |
|
1951 |
|
1952 YoungList* young_list = _g1->young_list(); |
|
1953 finalize_incremental_cset_building(); |
|
1954 |
|
1955 guarantee(target_pause_time_ms > 0.0, |
|
1956 err_msg("target_pause_time_ms = %1.6lf should be positive", |
|
1957 target_pause_time_ms)); |
|
1958 guarantee(_collection_set == NULL, "Precondition"); |
|
1959 |
|
1960 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); |
|
1961 double predicted_pause_time_ms = base_time_ms; |
|
1962 double time_remaining_ms = MAX2(target_pause_time_ms - base_time_ms, 0.0); |
|
1963 |
|
1964 ergo_verbose4(ErgoCSetConstruction | ErgoHigh, |
|
1965 "start choosing CSet", |
|
1966 ergo_format_size("_pending_cards") |
|
1967 ergo_format_ms("predicted base time") |
|
1968 ergo_format_ms("remaining time") |
|
1969 ergo_format_ms("target pause time"), |
|
1970 _pending_cards, base_time_ms, time_remaining_ms, target_pause_time_ms); |
|
1971 |
|
1972 _last_gc_was_young = gcs_are_young() ? true : false; |
|
1973 |
|
1974 if (_last_gc_was_young) { |
|
1975 _trace_gen0_time_data.increment_young_collection_count(); |
|
1976 } else { |
|
1977 _trace_gen0_time_data.increment_mixed_collection_count(); |
|
1978 } |
|
1979 |
|
1980 // The young list is laid with the survivor regions from the previous |
|
1981 // pause are appended to the RHS of the young list, i.e. |
|
1982 // [Newly Young Regions ++ Survivors from last pause]. |
|
1983 |
|
1984 uint survivor_region_length = young_list->survivor_length(); |
|
1985 uint eden_region_length = young_list->length() - survivor_region_length; |
|
1986 init_cset_region_lengths(eden_region_length, survivor_region_length); |
|
1987 |
|
1988 HeapRegion* hr = young_list->first_survivor_region(); |
|
1989 while (hr != NULL) { |
|
1990 assert(hr->is_survivor(), "badly formed young list"); |
|
1991 hr->set_young(); |
|
1992 hr = hr->get_next_young_region(); |
|
1993 } |
|
1994 |
|
1995 // Clear the fields that point to the survivor list - they are all young now. |
|
1996 young_list->clear_survivors(); |
|
1997 |
|
1998 _collection_set = _inc_cset_head; |
|
1999 _collection_set_bytes_used_before = _inc_cset_bytes_used_before; |
|
2000 time_remaining_ms = MAX2(time_remaining_ms - _inc_cset_predicted_elapsed_time_ms, 0.0); |
|
2001 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms; |
|
2002 |
|
2003 ergo_verbose3(ErgoCSetConstruction | ErgoHigh, |
|
2004 "add young regions to CSet", |
|
2005 ergo_format_region("eden") |
|
2006 ergo_format_region("survivors") |
|
2007 ergo_format_ms("predicted young region time"), |
|
2008 eden_region_length, survivor_region_length, |
|
2009 _inc_cset_predicted_elapsed_time_ms); |
|
2010 |
|
2011 // The number of recorded young regions is the incremental |
|
2012 // collection set's current size |
|
2013 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths); |
|
2014 |
|
2015 double young_end_time_sec = os::elapsedTime(); |
|
2016 phase_times()->record_young_cset_choice_time_ms((young_end_time_sec - young_start_time_sec) * 1000.0); |
|
2017 |
|
2018 // Set the start of the non-young choice time. |
|
2019 double non_young_start_time_sec = young_end_time_sec; |
|
2020 |
|
2021 if (!gcs_are_young()) { |
|
2022 CollectionSetChooser* cset_chooser = _collectionSetChooser; |
|
2023 cset_chooser->verify(); |
|
2024 const uint min_old_cset_length = calc_min_old_cset_length(); |
|
2025 const uint max_old_cset_length = calc_max_old_cset_length(); |
|
2026 |
|
2027 uint expensive_region_num = 0; |
|
2028 bool check_time_remaining = adaptive_young_list_length(); |
|
2029 |
|
2030 HeapRegion* hr = cset_chooser->peek(); |
|
2031 while (hr != NULL) { |
|
2032 if (old_cset_region_length() >= max_old_cset_length) { |
|
2033 // Added maximum number of old regions to the CSet. |
|
2034 ergo_verbose2(ErgoCSetConstruction, |
|
2035 "finish adding old regions to CSet", |
|
2036 ergo_format_reason("old CSet region num reached max") |
|
2037 ergo_format_region("old") |
|
2038 ergo_format_region("max"), |
|
2039 old_cset_region_length(), max_old_cset_length); |
|
2040 break; |
|
2041 } |
|
2042 |
|
2043 |
|
2044 // Stop adding regions if the remaining reclaimable space is |
|
2045 // not above G1HeapWastePercent. |
|
2046 size_t reclaimable_bytes = cset_chooser->remaining_reclaimable_bytes(); |
|
2047 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes); |
|
2048 double threshold = (double) G1HeapWastePercent; |
|
2049 if (reclaimable_perc <= threshold) { |
|
2050 // We've added enough old regions that the amount of uncollected |
|
2051 // reclaimable space is at or below the waste threshold. Stop |
|
2052 // adding old regions to the CSet. |
|
2053 ergo_verbose5(ErgoCSetConstruction, |
|
2054 "finish adding old regions to CSet", |
|
2055 ergo_format_reason("reclaimable percentage not over threshold") |
|
2056 ergo_format_region("old") |
|
2057 ergo_format_region("max") |
|
2058 ergo_format_byte_perc("reclaimable") |
|
2059 ergo_format_perc("threshold"), |
|
2060 old_cset_region_length(), |
|
2061 max_old_cset_length, |
|
2062 reclaimable_bytes, |
|
2063 reclaimable_perc, threshold); |
|
2064 break; |
|
2065 } |
|
2066 |
|
2067 double predicted_time_ms = predict_region_elapsed_time_ms(hr, gcs_are_young()); |
|
2068 if (check_time_remaining) { |
|
2069 if (predicted_time_ms > time_remaining_ms) { |
|
2070 // Too expensive for the current CSet. |
|
2071 |
|
2072 if (old_cset_region_length() >= min_old_cset_length) { |
|
2073 // We have added the minimum number of old regions to the CSet, |
|
2074 // we are done with this CSet. |
|
2075 ergo_verbose4(ErgoCSetConstruction, |
|
2076 "finish adding old regions to CSet", |
|
2077 ergo_format_reason("predicted time is too high") |
|
2078 ergo_format_ms("predicted time") |
|
2079 ergo_format_ms("remaining time") |
|
2080 ergo_format_region("old") |
|
2081 ergo_format_region("min"), |
|
2082 predicted_time_ms, time_remaining_ms, |
|
2083 old_cset_region_length(), min_old_cset_length); |
|
2084 break; |
|
2085 } |
|
2086 |
|
2087 // We'll add it anyway given that we haven't reached the |
|
2088 // minimum number of old regions. |
|
2089 expensive_region_num += 1; |
|
2090 } |
|
2091 } else { |
|
2092 if (old_cset_region_length() >= min_old_cset_length) { |
|
2093 // In the non-auto-tuning case, we'll finish adding regions |
|
2094 // to the CSet if we reach the minimum. |
|
2095 ergo_verbose2(ErgoCSetConstruction, |
|
2096 "finish adding old regions to CSet", |
|
2097 ergo_format_reason("old CSet region num reached min") |
|
2098 ergo_format_region("old") |
|
2099 ergo_format_region("min"), |
|
2100 old_cset_region_length(), min_old_cset_length); |
|
2101 break; |
|
2102 } |
|
2103 } |
|
2104 |
|
2105 // We will add this region to the CSet. |
|
2106 time_remaining_ms = MAX2(time_remaining_ms - predicted_time_ms, 0.0); |
|
2107 predicted_pause_time_ms += predicted_time_ms; |
|
2108 cset_chooser->remove_and_move_to_next(hr); |
|
2109 _g1->old_set_remove(hr); |
|
2110 add_old_region_to_cset(hr); |
|
2111 |
|
2112 hr = cset_chooser->peek(); |
|
2113 } |
|
2114 if (hr == NULL) { |
|
2115 ergo_verbose0(ErgoCSetConstruction, |
|
2116 "finish adding old regions to CSet", |
|
2117 ergo_format_reason("candidate old regions not available")); |
|
2118 } |
|
2119 |
|
2120 if (expensive_region_num > 0) { |
|
2121 // We print the information once here at the end, predicated on |
|
2122 // whether we added any apparently expensive regions or not, to |
|
2123 // avoid generating output per region. |
|
2124 ergo_verbose4(ErgoCSetConstruction, |
|
2125 "added expensive regions to CSet", |
|
2126 ergo_format_reason("old CSet region num not reached min") |
|
2127 ergo_format_region("old") |
|
2128 ergo_format_region("expensive") |
|
2129 ergo_format_region("min") |
|
2130 ergo_format_ms("remaining time"), |
|
2131 old_cset_region_length(), |
|
2132 expensive_region_num, |
|
2133 min_old_cset_length, |
|
2134 time_remaining_ms); |
|
2135 } |
|
2136 |
|
2137 cset_chooser->verify(); |
|
2138 } |
|
2139 |
|
2140 stop_incremental_cset_building(); |
|
2141 |
|
2142 ergo_verbose5(ErgoCSetConstruction, |
|
2143 "finish choosing CSet", |
|
2144 ergo_format_region("eden") |
|
2145 ergo_format_region("survivors") |
|
2146 ergo_format_region("old") |
|
2147 ergo_format_ms("predicted pause time") |
|
2148 ergo_format_ms("target pause time"), |
|
2149 eden_region_length, survivor_region_length, |
|
2150 old_cset_region_length(), |
|
2151 predicted_pause_time_ms, target_pause_time_ms); |
|
2152 |
|
2153 double non_young_end_time_sec = os::elapsedTime(); |
|
2154 phase_times()->record_non_young_cset_choice_time_ms((non_young_end_time_sec - non_young_start_time_sec) * 1000.0); |
|
2155 evacuation_info.set_collectionset_regions(cset_region_length()); |
|
2156 } |
|
2157 |
|
2158 void TraceGen0TimeData::record_start_collection(double time_to_stop_the_world_ms) { |
|
2159 if(TraceGen0Time) { |
|
2160 _all_stop_world_times_ms.add(time_to_stop_the_world_ms); |
|
2161 } |
|
2162 } |
|
2163 |
|
2164 void TraceGen0TimeData::record_yield_time(double yield_time_ms) { |
|
2165 if(TraceGen0Time) { |
|
2166 _all_yield_times_ms.add(yield_time_ms); |
|
2167 } |
|
2168 } |
|
2169 |
|
2170 void TraceGen0TimeData::record_end_collection(double pause_time_ms, G1GCPhaseTimes* phase_times) { |
|
2171 if(TraceGen0Time) { |
|
2172 _total.add(pause_time_ms); |
|
2173 _other.add(pause_time_ms - phase_times->accounted_time_ms()); |
|
2174 _root_region_scan_wait.add(phase_times->root_region_scan_wait_time_ms()); |
|
2175 _parallel.add(phase_times->cur_collection_par_time_ms()); |
|
2176 _ext_root_scan.add(phase_times->average_last_ext_root_scan_time()); |
|
2177 _satb_filtering.add(phase_times->average_last_satb_filtering_times_ms()); |
|
2178 _update_rs.add(phase_times->average_last_update_rs_time()); |
|
2179 _scan_rs.add(phase_times->average_last_scan_rs_time()); |
|
2180 _obj_copy.add(phase_times->average_last_obj_copy_time()); |
|
2181 _termination.add(phase_times->average_last_termination_time()); |
|
2182 |
|
2183 double parallel_known_time = phase_times->average_last_ext_root_scan_time() + |
|
2184 phase_times->average_last_satb_filtering_times_ms() + |
|
2185 phase_times->average_last_update_rs_time() + |
|
2186 phase_times->average_last_scan_rs_time() + |
|
2187 phase_times->average_last_obj_copy_time() + |
|
2188 + phase_times->average_last_termination_time(); |
|
2189 |
|
2190 double parallel_other_time = phase_times->cur_collection_par_time_ms() - parallel_known_time; |
|
2191 _parallel_other.add(parallel_other_time); |
|
2192 _clear_ct.add(phase_times->cur_clear_ct_time_ms()); |
|
2193 } |
|
2194 } |
|
2195 |
|
2196 void TraceGen0TimeData::increment_young_collection_count() { |
|
2197 if(TraceGen0Time) { |
|
2198 ++_young_pause_num; |
|
2199 } |
|
2200 } |
|
2201 |
|
2202 void TraceGen0TimeData::increment_mixed_collection_count() { |
|
2203 if(TraceGen0Time) { |
|
2204 ++_mixed_pause_num; |
|
2205 } |
|
2206 } |
|
2207 |
|
2208 void TraceGen0TimeData::print_summary(const char* str, |
|
2209 const NumberSeq* seq) const { |
|
2210 double sum = seq->sum(); |
|
2211 gclog_or_tty->print_cr("%-27s = %8.2lf s (avg = %8.2lf ms)", |
|
2212 str, sum / 1000.0, seq->avg()); |
|
2213 } |
|
2214 |
|
2215 void TraceGen0TimeData::print_summary_sd(const char* str, |
|
2216 const NumberSeq* seq) const { |
|
2217 print_summary(str, seq); |
|
2218 gclog_or_tty->print_cr("%+45s = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", |
|
2219 "(num", seq->num(), seq->sd(), seq->maximum()); |
|
2220 } |
|
2221 |
|
2222 void TraceGen0TimeData::print() const { |
|
2223 if (!TraceGen0Time) { |
|
2224 return; |
|
2225 } |
|
2226 |
|
2227 gclog_or_tty->print_cr("ALL PAUSES"); |
|
2228 print_summary_sd(" Total", &_total); |
|
2229 gclog_or_tty->cr(); |
|
2230 gclog_or_tty->cr(); |
|
2231 gclog_or_tty->print_cr(" Young GC Pauses: %8d", _young_pause_num); |
|
2232 gclog_or_tty->print_cr(" Mixed GC Pauses: %8d", _mixed_pause_num); |
|
2233 gclog_or_tty->cr(); |
|
2234 |
|
2235 gclog_or_tty->print_cr("EVACUATION PAUSES"); |
|
2236 |
|
2237 if (_young_pause_num == 0 && _mixed_pause_num == 0) { |
|
2238 gclog_or_tty->print_cr("none"); |
|
2239 } else { |
|
2240 print_summary_sd(" Evacuation Pauses", &_total); |
|
2241 print_summary(" Root Region Scan Wait", &_root_region_scan_wait); |
|
2242 print_summary(" Parallel Time", &_parallel); |
|
2243 print_summary(" Ext Root Scanning", &_ext_root_scan); |
|
2244 print_summary(" SATB Filtering", &_satb_filtering); |
|
2245 print_summary(" Update RS", &_update_rs); |
|
2246 print_summary(" Scan RS", &_scan_rs); |
|
2247 print_summary(" Object Copy", &_obj_copy); |
|
2248 print_summary(" Termination", &_termination); |
|
2249 print_summary(" Parallel Other", &_parallel_other); |
|
2250 print_summary(" Clear CT", &_clear_ct); |
|
2251 print_summary(" Other", &_other); |
|
2252 } |
|
2253 gclog_or_tty->cr(); |
|
2254 |
|
2255 gclog_or_tty->print_cr("MISC"); |
|
2256 print_summary_sd(" Stop World", &_all_stop_world_times_ms); |
|
2257 print_summary_sd(" Yields", &_all_yield_times_ms); |
|
2258 } |
|
2259 |
|
2260 void TraceGen1TimeData::record_full_collection(double full_gc_time_ms) { |
|
2261 if (TraceGen1Time) { |
|
2262 _all_full_gc_times.add(full_gc_time_ms); |
|
2263 } |
|
2264 } |
|
2265 |
|
2266 void TraceGen1TimeData::print() const { |
|
2267 if (!TraceGen1Time) { |
|
2268 return; |
|
2269 } |
|
2270 |
|
2271 if (_all_full_gc_times.num() > 0) { |
|
2272 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", |
|
2273 _all_full_gc_times.num(), |
|
2274 _all_full_gc_times.sum() / 1000.0); |
|
2275 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times.avg()); |
|
2276 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", |
|
2277 _all_full_gc_times.sd(), |
|
2278 _all_full_gc_times.maximum()); |
|
2279 } |
|
2280 } |