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
32 /*
33 * In normal concurrent cycle, we have to pace the application to let GC finish.
34 *
35 * Here, we do not know how large would be the collection set, and what are the
36 * relative performances of the each stage in the concurrent cycle, and so we have to
37 * make some assumptions.
38 *
39 * For concurrent mark, there is no clear notion of progress. The moderately accurate
40 * and easy to get metric is the amount of live objects the mark had encountered. But,
41 * that does directly correlate with the used heap, because the heap might be fully
42 * dead or fully alive. We cannot assume either of the extremes: we would either allow
43 * application to run out of memory if we assume heap is fully dead but it is not, and,
44 * conversely, we would pacify application excessively if we assume heap is fully alive
45 * but it is not. So we need to guesstimate the particular expected value for heap liveness.
46 * The best way to do this is apparently recording the past history.
47 *
48 * For concurrent evac and update-refs, we are walking the heap per-region, and so the
49 * notion of progress is clear: we get reported the "used" size from the processed regions
50 * and use the global heap-used as the baseline.
51 *
52 * The allocatable space when GC is running is "free" at the start of cycle, but the
53 * accounted budget is based on "used". So, we need to adjust the tax knowing that.
54 * Also, since we effectively count the used space three times (mark, evac, update-refs),
55 * we need to multiply the tax by 3. Example: for 10 MB free and 90 MB used, GC would
56 * come back with 3*90 MB budget, and thus for each 1 MB of allocation, we have to pay
57 * 3*90 / 10 MBs. In the end, we would pay back the entire budget.
58 */
59
60 void ShenandoahPacer::setup_for_mark() {
61 assert(ShenandoahPacing, "Only be here when pacing is enabled");
62
63 size_t live = update_and_get_progress_history();
64 size_t free = _heap->free_set()->available();
65
66 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
67 size_t taxable = free - non_taxable;
68
69 double tax = 1.0 * live / taxable; // base tax for available free space
70 tax *= 3; // mark is phase 1 of 3, claim 1/3 of free for it
71 tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap
72
73 restart_with(non_taxable, tax);
74
75 log_info(gc, ergo)("Pacer for Mark. Expected Live: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, "
76 "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
77 byte_size_in_proper_unit(live), proper_unit_for_byte_size(live),
78 byte_size_in_proper_unit(free), proper_unit_for_byte_size(free),
79 byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable),
80 tax);
81 }
82
83 void ShenandoahPacer::setup_for_evac() {
84 assert(ShenandoahPacing, "Only be here when pacing is enabled");
85
86 size_t used = _heap->collection_set()->used();
87 size_t free = _heap->free_set()->available();
88
89 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
90 size_t taxable = free - non_taxable;
91
92 double tax = 1.0 * used / taxable; // base tax for available free space
93 tax *= 2; // evac is phase 2 of 3, claim 1/2 of remaining free
94 tax = MAX2<double>(1, tax); // never allocate more than GC processes during the phase
95 tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap
96
97 restart_with(non_taxable, tax);
98
99 log_info(gc, ergo)("Pacer for Evacuation. Used CSet: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, "
100 "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
101 byte_size_in_proper_unit(used), proper_unit_for_byte_size(used),
102 byte_size_in_proper_unit(free), proper_unit_for_byte_size(free),
103 byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable),
104 tax);
105 }
106
107 void ShenandoahPacer::setup_for_updaterefs() {
108 assert(ShenandoahPacing, "Only be here when pacing is enabled");
109
110 size_t used = _heap->used();
111 size_t free = _heap->free_set()->available();
112
113 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
114 size_t taxable = free - non_taxable;
115
116 double tax = 1.0 * used / taxable; // base tax for available free space
117 tax *= 1; // update-refs is phase 3 of 3, claim the remaining free
118 tax = MAX2<double>(1, tax); // never allocate more than GC processes during the phase
119 tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap
120
121 restart_with(non_taxable, tax);
122
123 log_info(gc, ergo)("Pacer for Update Refs. Used: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, "
124 "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
125 byte_size_in_proper_unit(used), proper_unit_for_byte_size(used),
126 byte_size_in_proper_unit(free), proper_unit_for_byte_size(free),
127 byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable),
128 tax);
129 }
130
131 /*
132 * In idle phase, we have to pace the application to let control thread react with GC start.
133 *
134 * Here, we have rendezvous with concurrent thread that adds up the budget as it acknowledges
135 * it had seen recent allocations. It will naturally pace the allocations if control thread is
136 * not catching up. To bootstrap this feedback cycle, we need to start with some initial budget
137 * for applications to allocate at.
138 */
139
140 void ShenandoahPacer::setup_for_idle() {
141 assert(ShenandoahPacing, "Only be here when pacing is enabled");
142
143 size_t initial = _heap->max_capacity() / 100 * ShenandoahPacingIdleSlack;
144 double tax = 1;
145
146 restart_with(initial, tax);
147
148 log_info(gc, ergo)("Pacer for Idle. Initial: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
149 byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial),
150 tax);
151 }
152
153 /*
154 * There is no useful notion of progress for these operations. To avoid stalling
155 * the allocators unnecessarily, allow them to run unimpeded.
156 */
157
158 void ShenandoahPacer::setup_for_preclean() {
159 assert(ShenandoahPacing, "Only be here when pacing is enabled");
160
161 size_t initial = _heap->max_capacity();
162 restart_with(initial, 1.0);
163
164 log_info(gc, ergo)("Pacer for Precleaning. Non-Taxable: " SIZE_FORMAT "%s",
165 byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial));
166 }
167
168 void ShenandoahPacer::setup_for_reset() {
169 assert(ShenandoahPacing, "Only be here when pacing is enabled");
170
171 size_t initial = _heap->max_capacity();
172 restart_with(initial, 1.0);
173
174 log_info(gc, ergo)("Pacer for Reset. Non-Taxable: " SIZE_FORMAT "%s",
175 byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial));
176 }
177
178 size_t ShenandoahPacer::update_and_get_progress_history() {
179 if (_progress == -1) {
180 // First initialization, report some prior
181 Atomic::store(&_progress, (intptr_t)PACING_PROGRESS_ZERO);
182 return (size_t) (_heap->max_capacity() * 0.1);
183 } else {
184 // Record history, and reply historical data
185 _progress_history->add(_progress);
186 Atomic::store(&_progress, (intptr_t)PACING_PROGRESS_ZERO);
187 return (size_t) (_progress_history->avg() * HeapWordSize);
188 }
189 }
190
191 void ShenandoahPacer::restart_with(size_t non_taxable_bytes, double tax_rate) {
192 size_t initial = (size_t)(non_taxable_bytes * tax_rate) >> LogHeapWordSize;
193 STATIC_ASSERT(sizeof(size_t) <= sizeof(intptr_t));
194 Atomic::xchg(&_budget, (intptr_t)initial);
195 Atomic::store(&_tax_rate, tax_rate);
196 Atomic::inc(&_epoch);
197 }
198
199 bool ShenandoahPacer::claim_for_alloc(size_t words, bool force) {
200 assert(ShenandoahPacing, "Only be here when pacing is enabled");
201
202 intptr_t tax = MAX2<intptr_t>(1, words * Atomic::load(&_tax_rate));
203
204 intptr_t cur = 0;
205 intptr_t new_val = 0;
206 do {
207 cur = Atomic::load(&_budget);
208 if (cur < tax && !force) {
209 // Progress depleted, alas.
210 return false;
211 }
212 new_val = cur - tax;
213 } while (Atomic::cmpxchg(&_budget, cur, new_val) != cur);
214 return true;
215 }
216
217 void ShenandoahPacer::unpace_for_alloc(intptr_t epoch, size_t words) {
218 assert(ShenandoahPacing, "Only be here when pacing is enabled");
219
220 if (_epoch != epoch) {
221 // Stale ticket, no need to unpace.
222 return;
223 }
224
225 intptr_t tax = MAX2<intptr_t>(1, words * Atomic::load(&_tax_rate));
226 Atomic::add(&_budget, tax);
227 }
228
229 intptr_t ShenandoahPacer::epoch() {
230 return Atomic::load(&_epoch);
231 }
232
233 void ShenandoahPacer::pace_for_alloc(size_t words) {
234 assert(ShenandoahPacing, "Only be here when pacing is enabled");
235
236 // Fast path: try to allocate right away
237 if (claim_for_alloc(words, false)) {
238 return;
239 }
240
241 // Threads that are attaching should not block at all: they are not
242 // fully initialized yet. Calling sleep() on them would be awkward.
243 // This is probably the path that allocates the thread oop itself.
244 // Forcefully claim without waiting.
245 if (JavaThread::current()->is_attaching_via_jni()) {
246 claim_for_alloc(words, true);
247 return;
248 }
249
250 size_t max = ShenandoahPacingMaxDelay;
251 double start = os::elapsedTime();
252
253 size_t total = 0;
254 size_t cur = 0;
255
256 while (true) {
257 // We could instead assist GC, but this would suffice for now.
258 // This code should also participate in safepointing.
259 // Perform the exponential backoff, limited by max.
260
261 cur = cur * 2;
262 if (total + cur > max) {
263 cur = (max > total) ? (max - total) : 0;
264 }
265 cur = MAX2<size_t>(1, cur);
266
267 JavaThread::current()->sleep(cur);
268
269 double end = os::elapsedTime();
270 total = (size_t)((end - start) * 1000);
271
272 if (total > max) {
273 // Spent local time budget to wait for enough GC progress.
274 // Breaking out and allocating anyway, which may mean we outpace GC,
275 // and start Degenerated GC cycle.
276 _delays.add(total);
277
278 // Forcefully claim the budget: it may go negative at this point, and
279 // GC should replenish for this and subsequent allocations
280 claim_for_alloc(words, true);
281 break;
282 }
283
284 if (claim_for_alloc(words, false)) {
285 // Acquired enough permit, nice. Can allocate now.
286 _delays.add(total);
287 break;
288 }
289 }
290 }
291
292 void ShenandoahPacer::print_on(outputStream* out) const {
293 out->print_cr("ALLOCATION PACING:");
294 out->cr();
295
296 out->print_cr("Max pacing delay is set for " UINTX_FORMAT " ms.", ShenandoahPacingMaxDelay);
297 out->cr();
298
299 out->print_cr("Higher delay would prevent application outpacing the GC, but it will hide the GC latencies");
300 out->print_cr("from the STW pause times. Pacing affects the individual threads, and so it would also be");
301 out->print_cr("invisible to the usual profiling tools, but would add up to end-to-end application latency.");
302 out->print_cr("Raise max pacing delay with care.");
303 out->cr();
304
305 out->print_cr("Actual pacing delays histogram:");
306 out->cr();
307
308 out->print_cr("%10s - %10s %12s%12s", "From", "To", "Count", "Sum");
309
310 size_t total_count = 0;
311 size_t total_sum = 0;
312 for (int c = _delays.min_level(); c <= _delays.max_level(); c++) {
313 int l = (c == 0) ? 0 : 1 << (c - 1);
314 int r = 1 << c;
315 size_t count = _delays.level(c);
316 size_t sum = count * (r - l) / 2;
317 total_count += count;
318 total_sum += sum;
319
320 out->print_cr("%7d ms - %7d ms: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", l, r, count, sum);
321 }
322 out->print_cr("%23s: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", "Total", total_count, total_sum);
323 out->cr();
324 out->print_cr("Pacing delays are measured from entering the pacing code till exiting it. Therefore,");
325 out->print_cr("observed pacing delays may be higher than the threshold when paced thread spent more");
326 out->print_cr("time in the pacing code. It usually happens when thread is de-scheduled while paced,");
327 out->print_cr("OS takes longer to unblock the thread, or JVM experiences an STW pause.");
328 out->cr();
329 }