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 "gc/shenandoah/shenandoahPhaseTimings.hpp"
 31 #include "runtime/atomic.hpp"
 32 #include "runtime/javaThread.inline.hpp"
 33 #include "runtime/mutexLocker.hpp"
 34 #include "runtime/threadSMR.hpp"
 35 
 36 /*
 37  * In normal concurrent cycle, we have to pace the application to let GC finish.
 38  *
 39  * Here, we do not know how large would be the collection set, and what are the
 40  * relative performances of the each stage in the concurrent cycle, and so we have to
 41  * make some assumptions.
 42  *
 43  * For concurrent mark, there is no clear notion of progress. The moderately accurate
 44  * and easy to get metric is the amount of live objects the mark had encountered. But,
 45  * that does directly correlate with the used heap, because the heap might be fully
 46  * dead or fully alive. We cannot assume either of the extremes: we would either allow
 47  * application to run out of memory if we assume heap is fully dead but it is not, and,
 48  * conversely, we would pacify application excessively if we assume heap is fully alive
 49  * but it is not. So we need to guesstimate the particular expected value for heap liveness.
 50  * The best way to do this is apparently recording the past history.
 51  *
 52  * For concurrent evac and update-refs, we are walking the heap per-region, and so the
 53  * notion of progress is clear: we get reported the "used" size from the processed regions
 54  * and use the global heap-used as the baseline.
 55  *
 56  * The allocatable space when GC is running is "free" at the start of phase, but the
 57  * accounted budget is based on "used". So, we need to adjust the tax knowing that.
 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 *= 1;                          // mark can succeed with immediate garbage, claim all available space
 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 followed by update-refs, 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 the last phase, 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_reset() {
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 Reset. Non-Taxable: " SIZE_FORMAT "%s",
165                      byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial));
166 }
167 
168 size_t ShenandoahPacer::update_and_get_progress_history() {
169   if (_progress == -1) {
170     // First initialization, report some prior
171     Atomic::store(&_progress, (intptr_t)PACING_PROGRESS_ZERO);
172     return (size_t) (_heap->max_capacity() * 0.1);
173   } else {
174     // Record history, and reply historical data
175     _progress_history->add(_progress);
176     Atomic::store(&_progress, (intptr_t)PACING_PROGRESS_ZERO);
177     return (size_t) (_progress_history->avg() * HeapWordSize);
178   }
179 }
180 
181 void ShenandoahPacer::restart_with(size_t non_taxable_bytes, double tax_rate) {
182   size_t initial = (size_t)(non_taxable_bytes * tax_rate) >> LogHeapWordSize;
183   STATIC_ASSERT(sizeof(size_t) <= sizeof(intptr_t));
184   Atomic::xchg(&_budget, (intptr_t)initial, memory_order_relaxed);
185   Atomic::store(&_tax_rate, tax_rate);
186   Atomic::inc(&_epoch);
187 
188   // Shake up stalled waiters after budget update.
189   _need_notify_waiters.try_set();
190 }
191 
192 bool ShenandoahPacer::claim_for_alloc(size_t words, bool force) {
193   assert(ShenandoahPacing, "Only be here when pacing is enabled");
194 
195   intptr_t tax = MAX2<intptr_t>(1, words * Atomic::load(&_tax_rate));
196 
197   intptr_t cur = 0;
198   intptr_t new_val = 0;
199   do {
200     cur = Atomic::load(&_budget);
201     if (cur < tax && !force) {
202       // Progress depleted, alas.
203       return false;
204     }
205     new_val = cur - tax;
206   } while (Atomic::cmpxchg(&_budget, cur, new_val, memory_order_relaxed) != cur);
207   return true;
208 }
209 
210 void ShenandoahPacer::unpace_for_alloc(intptr_t epoch, size_t words) {
211   assert(ShenandoahPacing, "Only be here when pacing is enabled");
212 
213   if (Atomic::load(&_epoch) != epoch) {
214     // Stale ticket, no need to unpace.
215     return;
216   }
217 
218   size_t tax = MAX2<size_t>(1, words * Atomic::load(&_tax_rate));
219   add_budget(tax);
220 }
221 
222 intptr_t ShenandoahPacer::epoch() {
223   return Atomic::load(&_epoch);
224 }
225 
226 void ShenandoahPacer::pace_for_alloc(size_t words) {
227   assert(ShenandoahPacing, "Only be here when pacing is enabled");
228 
229   // Fast path: try to allocate right away
230   bool claimed = claim_for_alloc(words, false);
231   if (claimed) {
232     return;
233   }
234 
235   // Forcefully claim the budget: it may go negative at this point, and
236   // GC should replenish for this and subsequent allocations. After this claim,
237   // we would wait a bit until our claim is matched by additional progress,
238   // or the time budget depletes.
239   claimed = claim_for_alloc(words, true);
240   assert(claimed, "Should always succeed");
241 
242   // Threads that are attaching should not block at all: they are not
243   // fully initialized yet. Blocking them would be awkward.
244   // This is probably the path that allocates the thread oop itself.
245   //
246   // Thread which is not an active Java thread should also not block.
247   // This can happen during VM init when main thread is still not an
248   // active Java thread.
249   JavaThread* current = JavaThread::current();
250   if (current->is_attaching_via_jni() ||
251       !current->is_active_Java_thread()) {
252     return;
253   }
254 
255   double start = os::elapsedTime();
256 
257   size_t max_ms = ShenandoahPacingMaxDelay;
258   size_t total_ms = 0;
259 
260   while (true) {
261     // We could instead assist GC, but this would suffice for now.
262     size_t cur_ms = (max_ms > total_ms) ? (max_ms - total_ms) : 1;
263     wait(cur_ms);
264 
265     double end = os::elapsedTime();
266     total_ms = (size_t)((end - start) * 1000);
267 
268     if (total_ms > max_ms || Atomic::load(&_budget) >= 0) {
269       // Exiting if either:
270       //  a) 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       //  b) The budget had been replenished, which means our claim is satisfied.
274       ShenandoahThreadLocalData::add_paced_time(JavaThread::current(), end - start);
275       break;
276     }
277   }
278 }
279 
280 void ShenandoahPacer::wait(size_t time_ms) {
281   // Perform timed wait. It works like like sleep(), except without modifying
282   // the thread interruptible status. MonitorLocker also checks for safepoints.
283   assert(time_ms > 0, "Should not call this with zero argument, as it would stall until notify");
284   assert(time_ms <= LONG_MAX, "Sanity");
285   MonitorLocker locker(_wait_monitor);
286   _wait_monitor->wait((long)time_ms);
287 }
288 
289 void ShenandoahPacer::notify_waiters() {
290   if (_need_notify_waiters.try_unset()) {
291     MonitorLocker locker(_wait_monitor);
292     _wait_monitor->notify_all();
293   }
294 }
295 
296 void ShenandoahPacer::flush_stats_to_cycle() {
297   double sum = 0;
298   for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
299     sum += ShenandoahThreadLocalData::paced_time(t);
300   }
301   ShenandoahHeap::heap()->phase_timings()->record_phase_time(ShenandoahPhaseTimings::pacing, sum);
302 }
303 
304 void ShenandoahPacer::print_cycle_on(outputStream* out) {
305   MutexLocker lock(Threads_lock);
306 
307   double now = os::elapsedTime();
308   double total = now - _last_time;
309   _last_time = now;
310 
311   out->cr();
312   out->print_cr("Allocation pacing accrued:");
313 
314   size_t threads_total = 0;
315   size_t threads_nz = 0;
316   double sum = 0;
317   for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
318     double d = ShenandoahThreadLocalData::paced_time(t);
319     if (d > 0) {
320       threads_nz++;
321       sum += d;
322       out->print_cr("  %5.0f of %5.0f ms (%5.1f%%): %s",
323               d * 1000, total * 1000, d/total*100, t->name());
324     }
325     threads_total++;
326     ShenandoahThreadLocalData::reset_paced_time(t);
327   }
328   out->print_cr("  %5.0f of %5.0f ms (%5.1f%%): <total>",
329           sum * 1000, total * 1000, sum/total*100);
330 
331   if (threads_total > 0) {
332     out->print_cr("  %5.0f of %5.0f ms (%5.1f%%): <average total>",
333             sum / threads_total * 1000, total * 1000, sum / threads_total / total * 100);
334   }
335   if (threads_nz > 0) {
336     out->print_cr("  %5.0f of %5.0f ms (%5.1f%%): <average non-zero>",
337             sum / threads_nz * 1000, total * 1000, sum / threads_nz / total * 100);
338   }
339   out->cr();
340 }