1 /*
   2  * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.
   3  *
   4  * This code is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License version 2 only, as
   6  * published by the Free Software Foundation.
   7  *
   8  * This code is distributed in the hope that it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  11  * version 2 for more details (a copy is included in the LICENSE file that
  12  * accompanied this code).
  13  *
  14  * You should have received a copy of the GNU General Public License version
  15  * 2 along with this work; if not, write to the Free Software Foundation,
  16  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  17  *
  18  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  19  * or visit www.oracle.com if you need additional information or have any
  20  * questions.
  21  *
  22  */
  23 
  24 #include "precompiled.hpp"
  25 
  26 #include "gc/shenandoah/shenandoahFreeSet.hpp"
  27 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
  28 #include "gc/shenandoah/shenandoahPacer.hpp"
  29 #include "runtime/mutexLocker.hpp"
  30 
  31 /*
  32  * In normal concurrent cycle, we have to pace the application to let GC finish.
  33  *
  34  * Here, we do not know how large would be the collection set, and what are the
  35  * relative performances of the each stage in the concurrent cycle, and so we have to
  36  * make some assumptions.
  37  *
  38  * For concurrent mark, there is no clear notion of progress. The moderately accurate
  39  * and easy to get metric is the amount of live objects the mark had encountered. But,
  40  * that does directly correlate with the used heap, because the heap might be fully
  41  * dead or fully alive. We cannot assume either of the extremes: we would either allow
  42  * application to run out of memory if we assume heap is fully dead but it is not, and,
  43  * conversely, we would pacify application excessively if we assume heap is fully alive
  44  * but it is not. So we need to guesstimate the particular expected value for heap liveness.
  45  * The best way to do this is apparently recording the past history.
  46  *
  47  * For concurrent evac and update-refs, we are walking the heap per-region, and so the
  48  * notion of progress is clear: we get reported the "used" size from the processed regions
  49  * and use the global heap-used as the baseline.
  50  *
  51  * The allocatable space when GC is running is "free" at the start of phase, but the
  52  * accounted budget is based on "used". So, we need to adjust the tax knowing that.
  53  */
  54 
  55 void ShenandoahPacer::setup_for_mark() {
  56   assert(ShenandoahPacing, "Only be here when pacing is enabled");
  57 
  58   size_t live = update_and_get_progress_history();
  59   size_t free = _heap->free_set()->available();
  60 
  61   size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
  62   size_t taxable = free - non_taxable;
  63 
  64   double tax = 1.0 * live / taxable; // base tax for available free space
  65   tax *= 1;                          // mark can succeed with immediate garbage, claim all available space
  66   tax *= ShenandoahPacingSurcharge;  // additional surcharge to help unclutter heap
  67 
  68   restart_with(non_taxable, tax);
  69 
  70   log_info(gc, ergo)("Pacer for Mark. Expected Live: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, "
  71                      "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
  72                      byte_size_in_proper_unit(live),        proper_unit_for_byte_size(live),
  73                      byte_size_in_proper_unit(free),        proper_unit_for_byte_size(free),
  74                      byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable),
  75                      tax);
  76 }
  77 
  78 void ShenandoahPacer::setup_for_evac() {
  79   assert(ShenandoahPacing, "Only be here when pacing is enabled");
  80 
  81   size_t used = _heap->collection_set()->used();
  82   size_t free = _heap->free_set()->available();
  83 
  84   size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
  85   size_t taxable = free - non_taxable;
  86 
  87   double tax = 1.0 * used / taxable; // base tax for available free space
  88   tax *= 2;                          // evac is followed by update-refs, claim 1/2 of remaining free
  89   tax = MAX2<double>(1, tax);        // never allocate more than GC processes during the phase
  90   tax *= ShenandoahPacingSurcharge;  // additional surcharge to help unclutter heap
  91 
  92   restart_with(non_taxable, tax);
  93 
  94   log_info(gc, ergo)("Pacer for Evacuation. Used CSet: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, "
  95                      "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
  96                      byte_size_in_proper_unit(used),        proper_unit_for_byte_size(used),
  97                      byte_size_in_proper_unit(free),        proper_unit_for_byte_size(free),
  98                      byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable),
  99                      tax);
 100 }
 101 
 102 void ShenandoahPacer::setup_for_updaterefs() {
 103   assert(ShenandoahPacing, "Only be here when pacing is enabled");
 104 
 105   size_t used = _heap->used();
 106   size_t free = _heap->free_set()->available();
 107 
 108   size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
 109   size_t taxable = free - non_taxable;
 110 
 111   double tax = 1.0 * used / taxable; // base tax for available free space
 112   tax *= 1;                          // update-refs is the last phase, claim the remaining free
 113   tax = MAX2<double>(1, tax);        // never allocate more than GC processes during the phase
 114   tax *= ShenandoahPacingSurcharge;  // additional surcharge to help unclutter heap
 115 
 116   restart_with(non_taxable, tax);
 117 
 118   log_info(gc, ergo)("Pacer for Update Refs. Used: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, "
 119                      "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
 120                      byte_size_in_proper_unit(used),        proper_unit_for_byte_size(used),
 121                      byte_size_in_proper_unit(free),        proper_unit_for_byte_size(free),
 122                      byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable),
 123                      tax);
 124 }
 125 
 126 /*
 127  * In idle phase, we have to pace the application to let control thread react with GC start.
 128  *
 129  * Here, we have rendezvous with concurrent thread that adds up the budget as it acknowledges
 130  * it had seen recent allocations. It will naturally pace the allocations if control thread is
 131  * not catching up. To bootstrap this feedback cycle, we need to start with some initial budget
 132  * for applications to allocate at.
 133  */
 134 
 135 void ShenandoahPacer::setup_for_idle() {
 136   assert(ShenandoahPacing, "Only be here when pacing is enabled");
 137 
 138   size_t initial = _heap->max_capacity() / 100 * ShenandoahPacingIdleSlack;
 139   double tax = 1;
 140 
 141   restart_with(initial, tax);
 142 
 143   log_info(gc, ergo)("Pacer for Idle. Initial: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx",
 144                      byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial),
 145                      tax);
 146 }
 147 
 148 /*
 149  * There is no useful notion of progress for these operations. To avoid stalling
 150  * the allocators unnecessarily, allow them to run unimpeded.
 151  */
 152 
 153 void ShenandoahPacer::setup_for_preclean() {
 154   assert(ShenandoahPacing, "Only be here when pacing is enabled");
 155 
 156   size_t initial = _heap->max_capacity();
 157   restart_with(initial, 1.0);
 158 
 159   log_info(gc, ergo)("Pacer for Precleaning. Non-Taxable: " SIZE_FORMAT "%s",
 160                      byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial));
 161 }
 162 
 163 void ShenandoahPacer::setup_for_reset() {
 164   assert(ShenandoahPacing, "Only be here when pacing is enabled");
 165 
 166   size_t initial = _heap->max_capacity();
 167   restart_with(initial, 1.0);
 168 
 169   log_info(gc, ergo)("Pacer for Reset. Non-Taxable: " SIZE_FORMAT "%s",
 170                      byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial));
 171 }
 172 
 173 size_t ShenandoahPacer::update_and_get_progress_history() {
 174   if (_progress == -1) {
 175     // First initialization, report some prior
 176     Atomic::store((intptr_t)PACING_PROGRESS_ZERO, &_progress);
 177     return (size_t) (_heap->max_capacity() * 0.1);
 178   } else {
 179     // Record history, and reply historical data
 180     _progress_history->add(_progress);
 181     Atomic::store((intptr_t)PACING_PROGRESS_ZERO, &_progress);
 182     return (size_t) (_progress_history->avg() * HeapWordSize);
 183   }
 184 }
 185 
 186 void ShenandoahPacer::restart_with(size_t non_taxable_bytes, double tax_rate) {
 187   size_t initial = (size_t)(non_taxable_bytes * tax_rate) >> LogHeapWordSize;
 188   STATIC_ASSERT(sizeof(size_t) <= sizeof(intptr_t));
 189   Atomic::xchg((intptr_t)initial, &_budget);
 190   Atomic::store(tax_rate, &_tax_rate);
 191   Atomic::inc(&_epoch);
 192 }
 193 
 194 bool ShenandoahPacer::claim_for_alloc(size_t words, bool force) {
 195   assert(ShenandoahPacing, "Only be here when pacing is enabled");
 196 
 197   intptr_t tax = MAX2<intptr_t>(1, words * Atomic::load(&_tax_rate));
 198 
 199   intptr_t cur = 0;
 200   intptr_t new_val = 0;
 201   do {
 202     cur = Atomic::load(&_budget);
 203     if (cur < tax && !force) {
 204       // Progress depleted, alas.
 205       return false;
 206     }
 207     new_val = cur - tax;
 208   } while (Atomic::cmpxchg(new_val, &_budget, cur) != cur);
 209   return true;
 210 }
 211 
 212 void ShenandoahPacer::unpace_for_alloc(intptr_t epoch, size_t words) {
 213   assert(ShenandoahPacing, "Only be here when pacing is enabled");
 214 
 215   if (_epoch != epoch) {
 216     // Stale ticket, no need to unpace.
 217     return;
 218   }
 219 
 220   intptr_t tax = MAX2<intptr_t>(1, words * Atomic::load(&_tax_rate));
 221   Atomic::add(tax, &_budget);
 222 }
 223 
 224 intptr_t ShenandoahPacer::epoch() {
 225   return Atomic::load(&_epoch);
 226 }
 227 
 228 void ShenandoahPacer::pace_for_alloc(size_t words) {
 229   assert(ShenandoahPacing, "Only be here when pacing is enabled");
 230 
 231   // Fast path: try to allocate right away
 232   if (claim_for_alloc(words, false)) {
 233     return;
 234   }
 235 
 236   // Threads that are attaching should not block at all: they are not
 237   // fully initialized yet. Blocking them would be awkward.
 238   // This is probably the path that allocates the thread oop itself.
 239   // Forcefully claim without waiting.
 240   if (JavaThread::current()->is_attaching_via_jni()) {
 241     claim_for_alloc(words, true);
 242     return;
 243   }
 244 
 245   size_t max = ShenandoahPacingMaxDelay;
 246   double start = os::elapsedTime();
 247 
 248   size_t total = 0;
 249   size_t cur = 0;
 250 
 251   while (true) {
 252     // We could instead assist GC, but this would suffice for now.
 253     // This code should also participate in safepointing.
 254     // Perform the exponential backoff, limited by max.
 255 
 256     cur = cur * 2;
 257     if (total + cur > max) {
 258       cur = (max > total) ? (max - total) : 0;
 259     }
 260     cur = MAX2<size_t>(1, cur);
 261 
 262     wait(cur);
 263 
 264     double end = os::elapsedTime();
 265     total = (size_t)((end - start) * 1000);
 266 
 267     if (total > max) {
 268       // Spent local time budget to wait for enough GC progress.
 269       // Breaking out and allocating anyway, which may mean we outpace GC,
 270       // and start Degenerated GC cycle.
 271       _delays.add(total);
 272 
 273       // Forcefully claim the budget: it may go negative at this point, and
 274       // GC should replenish for this and subsequent allocations
 275       claim_for_alloc(words, true);
 276       break;
 277     }
 278 
 279     if (claim_for_alloc(words, false)) {
 280       // Acquired enough permit, nice. Can allocate now.
 281       _delays.add(total);
 282       break;
 283     }
 284   }
 285 }
 286 
 287 void ShenandoahPacer::wait(size_t time_ms) {
 288   // Perform timed wait. It works like like sleep(), except without modifying
 289   // the thread interruptible status. MonitorLocker also checks for safepoints.
 290   assert(time_ms > 0, "Should not call this with zero argument, as it would stall until notify");
 291   assert(time_ms <= LONG_MAX, "Sanity");
 292   MonitorLockerEx locker(_wait_monitor);
 293   _wait_monitor->wait(!Mutex::_no_safepoint_check_flag, (long)time_ms);
 294 }
 295 
 296 void ShenandoahPacer::notify_waiters() {
 297   MonitorLockerEx locker(_wait_monitor);
 298   _wait_monitor->notify_all();
 299 }
 300 
 301 void ShenandoahPacer::print_on(outputStream* out) const {
 302   out->print_cr("ALLOCATION PACING:");
 303   out->cr();
 304 
 305   out->print_cr("Max pacing delay is set for " UINTX_FORMAT " ms.", ShenandoahPacingMaxDelay);
 306   out->cr();
 307 
 308   out->print_cr("Higher delay would prevent application outpacing the GC, but it will hide the GC latencies");
 309   out->print_cr("from the STW pause times. Pacing affects the individual threads, and so it would also be");
 310   out->print_cr("invisible to the usual profiling tools, but would add up to end-to-end application latency.");
 311   out->print_cr("Raise max pacing delay with care.");
 312   out->cr();
 313 
 314   out->print_cr("Actual pacing delays histogram:");
 315   out->cr();
 316 
 317   out->print_cr("%10s - %10s  %12s%12s", "From", "To", "Count", "Sum");
 318 
 319   size_t total_count = 0;
 320   size_t total_sum = 0;
 321   for (int c = _delays.min_level(); c <= _delays.max_level(); c++) {
 322     int l = (c == 0) ? 0 : 1 << (c - 1);
 323     int r = 1 << c;
 324     size_t count = _delays.level(c);
 325     size_t sum   = count * (r - l) / 2;
 326     total_count += count;
 327     total_sum   += sum;
 328 
 329     out->print_cr("%7d ms - %7d ms: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", l, r, count, sum);
 330   }
 331   out->print_cr("%23s: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", "Total", total_count, total_sum);
 332   out->cr();
 333   out->print_cr("Pacing delays are measured from entering the pacing code till exiting it. Therefore,");
 334   out->print_cr("observed pacing delays may be higher than the threshold when paced thread spent more");
 335   out->print_cr("time in the pacing code. It usually happens when thread is de-scheduled while paced,");
 336   out->print_cr("OS takes longer to unblock the thread, or JVM experiences an STW pause.");
 337   out->cr();
 338 }