1 /*
2 * Copyright (c) 2018, 2019, Red Hat, Inc. 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 #include "precompiled.hpp"
26
27 #include "gc/shenandoah/shenandoahFreeSet.hpp"
28 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
29 #include "gc/shenandoah/shenandoahPacer.hpp"
30 #include "runtime/atomic.hpp"
31 #include "runtime/mutexLocker.hpp"
32
33 /*
34 * In normal concurrent cycle, we have to pace the application to let GC finish.
35 *
36 * Here, we do not know how large would be the collection set, and what are the
37 * relative performances of the each stage in the concurrent cycle, and so we have to
38 * make some assumptions.
39 *
40 * For concurrent mark, there is no clear notion of progress. The moderately accurate
41 * and easy to get metric is the amount of live objects the mark had encountered. But,
42 * that does directly correlate with the used heap, because the heap might be fully
43 * dead or fully alive. We cannot assume either of the extremes: we would either allow
44 * application to run out of memory if we assume heap is fully dead but it is not, and,
45 * conversely, we would pacify application excessively if we assume heap is fully alive
46 * but it is not. So we need to guesstimate the particular expected value for heap liveness.
47 * The best way to do this is apparently recording the past history.
48 *
49 * For concurrent evac and update-refs, we are walking the heap per-region, and so the
50 * notion of progress is clear: we get reported the "used" size from the processed regions
51 * and use the global heap-used as the baseline.
52 *
53 * The allocatable space when GC is running is "free" at the start of phase, but the
54 * accounted budget is based on "used". So, we need to adjust the tax knowing that.
55 */
56
57 void ShenandoahPacer::setup_for_mark() {
58 assert(ShenandoahPacing, "Only be here when pacing is enabled");
59
60 size_t live = update_and_get_progress_history();
61 size_t free = _heap->free_set()->available();
62
63 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
64 size_t taxable = free - non_taxable;
65
66 double tax = 1.0 * live / taxable; // base tax for available free space
67 tax *= 1; // mark can succeed with immediate garbage, claim all available space
68 tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap
69
70 restart_with(non_taxable, tax);
71
72 log_info(gc, ergo)("Pacer for Mark. Expected Live: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, "
73 "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
74 byte_size_in_proper_unit(live), proper_unit_for_byte_size(live),
75 byte_size_in_proper_unit(free), proper_unit_for_byte_size(free),
76 byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable),
77 tax);
78 }
79
80 void ShenandoahPacer::setup_for_evac() {
81 assert(ShenandoahPacing, "Only be here when pacing is enabled");
82
83 size_t used = _heap->collection_set()->used();
84 size_t free = _heap->free_set()->available();
85
86 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
87 size_t taxable = free - non_taxable;
88
89 double tax = 1.0 * used / taxable; // base tax for available free space
90 tax *= 2; // evac is followed by update-refs, claim 1/2 of remaining free
91 tax = MAX2<double>(1, tax); // never allocate more than GC processes during the phase
92 tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap
93
94 restart_with(non_taxable, tax);
95
96 log_info(gc, ergo)("Pacer for Evacuation. Used CSet: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, "
97 "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
98 byte_size_in_proper_unit(used), proper_unit_for_byte_size(used),
99 byte_size_in_proper_unit(free), proper_unit_for_byte_size(free),
100 byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable),
101 tax);
102 }
103
104 void ShenandoahPacer::setup_for_updaterefs() {
105 assert(ShenandoahPacing, "Only be here when pacing is enabled");
106
107 size_t used = _heap->used();
108 size_t free = _heap->free_set()->available();
109
110 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
111 size_t taxable = free - non_taxable;
112
113 double tax = 1.0 * used / taxable; // base tax for available free space
114 tax *= 1; // update-refs is the last phase, claim the remaining free
115 tax = MAX2<double>(1, tax); // never allocate more than GC processes during the phase
116 tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap
117
118 restart_with(non_taxable, tax);
119
120 log_info(gc, ergo)("Pacer for Update Refs. Used: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, "
121 "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
122 byte_size_in_proper_unit(used), proper_unit_for_byte_size(used),
123 byte_size_in_proper_unit(free), proper_unit_for_byte_size(free),
124 byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable),
125 tax);
126 }
127
128 /*
129 * In idle phase, we have to pace the application to let control thread react with GC start.
130 *
131 * Here, we have rendezvous with concurrent thread that adds up the budget as it acknowledges
132 * it had seen recent allocations. It will naturally pace the allocations if control thread is
133 * not catching up. To bootstrap this feedback cycle, we need to start with some initial budget
134 * for applications to allocate at.
135 */
136
137 void ShenandoahPacer::setup_for_idle() {
138 assert(ShenandoahPacing, "Only be here when pacing is enabled");
139
140 size_t initial = _heap->max_capacity() / 100 * ShenandoahPacingIdleSlack;
141 double tax = 1;
142
143 restart_with(initial, tax);
144
145 log_info(gc, ergo)("Pacer for Idle. Initial: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
146 byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial),
147 tax);
148 }
149
150 /*
151 * There is no useful notion of progress for these operations. To avoid stalling
152 * the allocators unnecessarily, allow them to run unimpeded.
153 */
154
155 void ShenandoahPacer::setup_for_preclean() {
156 assert(ShenandoahPacing, "Only be here when pacing is enabled");
157
158 size_t initial = _heap->max_capacity();
159 restart_with(initial, 1.0);
160
161 log_info(gc, ergo)("Pacer for Precleaning. Non-Taxable: " SIZE_FORMAT "%s",
162 byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial));
163 }
164
165 void ShenandoahPacer::setup_for_reset() {
166 assert(ShenandoahPacing, "Only be here when pacing is enabled");
167
168 size_t initial = _heap->max_capacity();
169 restart_with(initial, 1.0);
170
171 log_info(gc, ergo)("Pacer for Reset. Non-Taxable: " SIZE_FORMAT "%s",
172 byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial));
173 }
174
175 size_t ShenandoahPacer::update_and_get_progress_history() {
176 if (_progress == -1) {
177 // First initialization, report some prior
178 Atomic::store(&_progress, (intptr_t)PACING_PROGRESS_ZERO);
179 return (size_t) (_heap->max_capacity() * 0.1);
180 } else {
181 // Record history, and reply historical data
182 _progress_history->add(_progress);
183 Atomic::store(&_progress, (intptr_t)PACING_PROGRESS_ZERO);
184 return (size_t) (_progress_history->avg() * HeapWordSize);
185 }
186 }
187
188 void ShenandoahPacer::restart_with(size_t non_taxable_bytes, double tax_rate) {
189 size_t initial = (size_t)(non_taxable_bytes * tax_rate) >> LogHeapWordSize;
190 STATIC_ASSERT(sizeof(size_t) <= sizeof(intptr_t));
191 Atomic::xchg(&_budget, (intptr_t)initial);
192 Atomic::store(&_tax_rate, tax_rate);
193 Atomic::inc(&_epoch);
194 }
195
196 bool ShenandoahPacer::claim_for_alloc(size_t words, bool force) {
197 assert(ShenandoahPacing, "Only be here when pacing is enabled");
198
199 intptr_t tax = MAX2<intptr_t>(1, words * Atomic::load(&_tax_rate));
200
201 intptr_t cur = 0;
202 intptr_t new_val = 0;
203 do {
204 cur = Atomic::load(&_budget);
205 if (cur < tax && !force) {
206 // Progress depleted, alas.
207 return false;
208 }
209 new_val = cur - tax;
210 } while (Atomic::cmpxchg(&_budget, cur, new_val) != cur);
211 return true;
212 }
213
214 void ShenandoahPacer::unpace_for_alloc(intptr_t epoch, size_t words) {
215 assert(ShenandoahPacing, "Only be here when pacing is enabled");
216
217 if (_epoch != epoch) {
218 // Stale ticket, no need to unpace.
219 return;
220 }
221
222 intptr_t tax = MAX2<intptr_t>(1, words * Atomic::load(&_tax_rate));
223 Atomic::add(&_budget, tax);
224 }
225
226 intptr_t ShenandoahPacer::epoch() {
227 return Atomic::load(&_epoch);
228 }
229
230 void ShenandoahPacer::pace_for_alloc(size_t words) {
231 assert(ShenandoahPacing, "Only be here when pacing is enabled");
232
233 // Fast path: try to allocate right away
234 if (claim_for_alloc(words, false)) {
235 return;
236 }
237
238 // Threads that are attaching should not block at all: they are not
239 // fully initialized yet. Blocking them would be awkward.
240 // This is probably the path that allocates the thread oop itself.
241 // Forcefully claim without waiting.
242 if (JavaThread::current()->is_attaching_via_jni()) {
243 claim_for_alloc(words, true);
244 return;
245 }
246
247 size_t max = ShenandoahPacingMaxDelay;
248 double start = os::elapsedTime();
249
250 size_t total = 0;
251 size_t cur = 0;
252
253 while (true) {
254 // We could instead assist GC, but this would suffice for now.
255 // This code should also participate in safepointing.
256 // Perform the exponential backoff, limited by max.
257
258 cur = cur * 2;
259 if (total + cur > max) {
260 cur = (max > total) ? (max - total) : 0;
261 }
262 cur = MAX2<size_t>(1, cur);
263
264 wait(cur);
265
266 double end = os::elapsedTime();
267 total = (size_t)((end - start) * 1000);
268
269 if (total > max) {
270 // Spent local time budget to wait for enough GC progress.
271 // Breaking out and allocating anyway, which may mean we outpace GC,
272 // and start Degenerated GC cycle.
273 _delays.add(total);
274
275 // Forcefully claim the budget: it may go negative at this point, and
276 // GC should replenish for this and subsequent allocations
277 claim_for_alloc(words, true);
278 break;
279 }
280
281 if (claim_for_alloc(words, false)) {
282 // Acquired enough permit, nice. Can allocate now.
283 _delays.add(total);
284 break;
285 }
286 }
287 }
288
289 void ShenandoahPacer::wait(long time_ms) {
290 // Perform timed wait. It works like like sleep(), except without modifying
291 // the thread interruptible status. MonitorLocker also checks for safepoints.
292 assert(time_ms > 0, "Should not call this with zero argument, as it would stall until notify");
293 MonitorLocker locker(_wait_monitor);
294 _wait_monitor->wait(time_ms);
295 }
296
297 void ShenandoahPacer::notify_waiters() {
298 MonitorLocker locker(_wait_monitor);
299 _wait_monitor->notify_all();
300 }
301
302 void ShenandoahPacer::print_on(outputStream* out) const {
303 out->print_cr("ALLOCATION PACING:");
304 out->cr();
305
306 out->print_cr("Max pacing delay is set for " UINTX_FORMAT " ms.", ShenandoahPacingMaxDelay);
307 out->cr();
308
309 out->print_cr("Higher delay would prevent application outpacing the GC, but it will hide the GC latencies");
310 out->print_cr("from the STW pause times. Pacing affects the individual threads, and so it would also be");
311 out->print_cr("invisible to the usual profiling tools, but would add up to end-to-end application latency.");
312 out->print_cr("Raise max pacing delay with care.");
313 out->cr();
314
315 out->print_cr("Actual pacing delays histogram:");
316 out->cr();
317
318 out->print_cr("%10s - %10s %12s%12s", "From", "To", "Count", "Sum");
319
320 size_t total_count = 0;
321 size_t total_sum = 0;
322 for (int c = _delays.min_level(); c <= _delays.max_level(); c++) {
323 int l = (c == 0) ? 0 : 1 << (c - 1);
324 int r = 1 << c;
325 size_t count = _delays.level(c);
326 size_t sum = count * (r - l) / 2;
327 total_count += count;
328 total_sum += sum;
329
330 out->print_cr("%7d ms - %7d ms: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", l, r, count, sum);
331 }
332 out->print_cr("%23s: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", "Total", total_count, total_sum);
333 out->cr();
334 out->print_cr("Pacing delays are measured from entering the pacing code till exiting it. Therefore,");
335 out->print_cr("observed pacing delays may be higher than the threshold when paced thread spent more");
336 out->print_cr("time in the pacing code. It usually happens when thread is de-scheduled while paced,");
337 out->print_cr("OS takes longer to unblock the thread, or JVM experiences an STW pause.");
338 out->cr();
339 }