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/heuristics/shenandoahTraversalHeuristics.hpp"
  27 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
  28 #include "gc/shenandoah/shenandoahFreeSet.hpp"
  29 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
  30 #include "gc/shenandoah/shenandoahHeuristics.hpp"
  31 #include "gc/shenandoah/shenandoahTraversalGC.hpp"
  32 #include "logging/log.hpp"
  33 #include "logging/logTag.hpp"
  34 #include "utilities/quickSort.hpp"
  35 
  36 ShenandoahTraversalHeuristics::ShenandoahTraversalHeuristics() : ShenandoahHeuristics(),
  37   _last_cset_select(0) {}
  38 
  39 bool ShenandoahTraversalHeuristics::is_experimental() {
  40   return false;
  41 }
  42 
  43 bool ShenandoahTraversalHeuristics::is_diagnostic() {
  44   return false;
  45 }
  46 
  47 const char* ShenandoahTraversalHeuristics::name() {
  48   return "traversal";
  49 }
  50 
  51 void ShenandoahTraversalHeuristics::choose_collection_set(ShenandoahCollectionSet* collection_set) {
  52   ShenandoahHeap* heap = ShenandoahHeap::heap();
  53 
  54   ShenandoahTraversalGC* traversal_gc = heap->traversal_gc();
  55 
  56   ShenandoahHeapRegionSet* traversal_set = traversal_gc->traversal_set();
  57   traversal_set->clear();
  58 
  59   RegionData *data = get_region_data_cache(heap->num_regions());
  60   size_t cnt = 0;
  61 
  62   // About to choose the collection set, make sure we have pinned regions in correct state
  63   heap->assert_pinned_region_status();
  64 
  65   // Step 0. Prepare all regions
  66 
  67   for (size_t i = 0; i < heap->num_regions(); i++) {
  68     ShenandoahHeapRegion* r = heap->get_region(i);
  69     if (r->used() > 0) {
  70       if (r->is_regular()) {
  71         data[cnt]._region = r;
  72         data[cnt]._garbage = r->garbage();
  73         data[cnt]._seqnum_last_alloc = r->seqnum_last_alloc_mutator();
  74         cnt++;
  75       }
  76       traversal_set->add_region(r);
  77     }
  78   }
  79 
  80   // The logic for cset selection is similar to that of adaptive:
  81   //
  82   //   1. We cannot get cset larger than available free space. Otherwise we guarantee OOME
  83   //      during evacuation, and thus guarantee full GC. In practice, we also want to let
  84   //      application to allocate something. This is why we limit CSet to some fraction of
  85   //      available space. In non-overloaded heap, max_cset would contain all plausible candidates
  86   //      over garbage threshold.
  87   //
  88   //   2. We should not get cset too low so that free threshold would not be met right
  89   //      after the cycle. Otherwise we get back-to-back cycles for no reason if heap is
  90   //      too fragmented. In non-overloaded non-fragmented heap min_garbage would be around zero.
  91   //
  92   // Therefore, we start by sorting the regions by garbage. Then we unconditionally add the best candidates
  93   // before we meet min_garbage. Then we add all candidates that fit with a garbage threshold before
  94   // we hit max_cset. When max_cset is hit, we terminate the cset selection. Note that in this scheme,
  95   // ShenandoahGarbageThreshold is the soft threshold which would be ignored until min_garbage is hit.
  96   //
  97   // The significant complication is that liveness data was collected at the previous cycle, and only
  98   // for those regions that were allocated before previous cycle started.
  99 
 100   size_t capacity    = heap->max_capacity();
 101   size_t actual_free = heap->free_set()->available();
 102   size_t free_target = capacity / 100 * ShenandoahMinFreeThreshold;
 103   size_t min_garbage = free_target > actual_free ? (free_target - actual_free) : 0;
 104   size_t max_cset    = (size_t)((1.0 * capacity / 100 * ShenandoahEvacReserve) / ShenandoahEvacWaste);
 105 
 106   log_info(gc, ergo)("Adaptive CSet Selection. Target Free: " SIZE_FORMAT "%s, Actual Free: "
 107                      SIZE_FORMAT "%s, Max CSet: " SIZE_FORMAT "%s, Min Garbage: " SIZE_FORMAT "%s",
 108                      byte_size_in_proper_unit(free_target), proper_unit_for_byte_size(free_target),
 109                      byte_size_in_proper_unit(actual_free), proper_unit_for_byte_size(actual_free),
 110                      byte_size_in_proper_unit(max_cset),    proper_unit_for_byte_size(max_cset),
 111                      byte_size_in_proper_unit(min_garbage), proper_unit_for_byte_size(min_garbage));
 112 
 113   // Better select garbage-first regions, and then older ones
 114   QuickSort::sort<RegionData>(data, (int) cnt, compare_by_garbage_then_alloc_seq_ascending, false);
 115 
 116   size_t cur_cset = 0;
 117   size_t cur_garbage = 0;
 118 
 119   size_t garbage_threshold = ShenandoahHeapRegion::region_size_bytes() / 100 * ShenandoahGarbageThreshold;
 120 
 121   // Step 1. Add trustworthy regions to collection set.
 122   //
 123   // We can trust live/garbage data from regions that were fully traversed during
 124   // previous cycle. Even if actual liveness is different now, we can only have _less_
 125   // live objects, because dead objects are not resurrected. Which means we can undershoot
 126   // the collection set, but not overshoot it.
 127 
 128   for (size_t i = 0; i < cnt; i++) {
 129     if (data[i]._seqnum_last_alloc > _last_cset_select) continue;
 130 
 131     ShenandoahHeapRegion* r = data[i]._region;
 132     assert (r->is_regular(), "should have been filtered before");
 133 
 134     size_t new_garbage = cur_garbage + r->garbage();
 135     size_t new_cset    = cur_cset    + r->get_live_data_bytes();
 136 
 137     if (new_cset > max_cset) {
 138       break;
 139     }
 140 
 141     if ((new_garbage < min_garbage) || (r->garbage() > garbage_threshold)) {
 142       assert(!collection_set->is_in(r), "must not yet be in cset");
 143       collection_set->add_region(r);
 144       cur_cset = new_cset;
 145       cur_garbage = new_garbage;
 146     }
 147   }
 148 
 149   // Step 2. Try to catch some recently allocated regions for evacuation ride.
 150   //
 151   // Pessimistically assume we are going to evacuate the entire region. While this
 152   // is very pessimistic and in most cases undershoots the collection set when regions
 153   // are mostly dead, it also provides more safety against running into allocation
 154   // failure when newly allocated regions are fully live.
 155 
 156   for (size_t i = 0; i < cnt; i++) {
 157     if (data[i]._seqnum_last_alloc <= _last_cset_select) continue;
 158 
 159     ShenandoahHeapRegion* r = data[i]._region;
 160     assert (r->is_regular(), "should have been filtered before");
 161 
 162     // size_t new_garbage = cur_garbage + 0; (implied)
 163     size_t new_cset = cur_cset + r->used();
 164 
 165     if (new_cset > max_cset) {
 166       break;
 167     }
 168 
 169     assert(!collection_set->is_in(r), "must not yet be in cset");
 170     collection_set->add_region(r);
 171     cur_cset = new_cset;
 172   }
 173 
 174   // Step 3. Clear liveness data
 175   // TODO: Merge it with step 0, but save live data in RegionData before.
 176   for (size_t i = 0; i < heap->num_regions(); i++) {
 177     ShenandoahHeapRegion* r = heap->get_region(i);
 178     if (r->used() > 0) {
 179       r->clear_live_data();
 180     }
 181   }
 182 
 183   collection_set->update_region_status();
 184 
 185   _last_cset_select = ShenandoahHeapRegion::seqnum_current_alloc();
 186 }
 187 
 188 bool ShenandoahTraversalHeuristics::should_start_gc() const {
 189   ShenandoahHeap* heap = ShenandoahHeap::heap();
 190   assert(!heap->has_forwarded_objects(), "no forwarded objects here");
 191 
 192   size_t capacity = heap->max_capacity();
 193   size_t available = heap->free_set()->available();
 194 
 195   // Check if we are falling below the worst limit, time to trigger the GC, regardless of
 196   // anything else.
 197   size_t min_threshold = capacity / 100 * ShenandoahMinFreeThreshold;
 198   if (available < min_threshold) {
 199     log_info(gc)("Trigger: Free (" SIZE_FORMAT "%s) is below minimum threshold (" SIZE_FORMAT "%s)",
 200                  byte_size_in_proper_unit(available),     proper_unit_for_byte_size(available),
 201                  byte_size_in_proper_unit(min_threshold), proper_unit_for_byte_size(min_threshold));
 202     return true;
 203   }
 204 
 205   // Check if are need to learn a bit about the application
 206   const size_t max_learn = ShenandoahLearningSteps;
 207   if (_gc_times_learned < max_learn) {
 208     size_t init_threshold = capacity / 100 * ShenandoahInitFreeThreshold;
 209     if (available < init_threshold) {
 210       log_info(gc)("Trigger: Learning " SIZE_FORMAT " of " SIZE_FORMAT ". Free (" SIZE_FORMAT "%s) is below initial threshold (" SIZE_FORMAT "%s)",
 211                    _gc_times_learned + 1, max_learn,
 212                    byte_size_in_proper_unit(available),      proper_unit_for_byte_size(available),
 213                    byte_size_in_proper_unit(init_threshold), proper_unit_for_byte_size(init_threshold));
 214       return true;
 215     }
 216   }
 217 
 218   // Check if allocation headroom is still okay. This also factors in:
 219   //   1. Some space to absorb allocation spikes
 220   //   2. Accumulated penalties from Degenerated and Full GC
 221 
 222   size_t allocation_headroom = available;
 223 
 224   size_t spike_headroom = capacity / 100 * ShenandoahAllocSpikeFactor;
 225   size_t penalties      = capacity / 100 * _gc_time_penalties;
 226 
 227   allocation_headroom -= MIN2(allocation_headroom, spike_headroom);
 228   allocation_headroom -= MIN2(allocation_headroom, penalties);
 229 
 230   double average_gc = _gc_time_history->avg();
 231   double time_since_last = time_since_last_gc();
 232   double allocation_rate = heap->bytes_allocated_since_gc_start() / time_since_last;
 233 
 234   if (average_gc > allocation_headroom / allocation_rate) {
 235     log_info(gc)("Trigger: Average GC time (%.2f ms) is above the time for allocation rate (%.0f %sB/s) to deplete free headroom (" SIZE_FORMAT "%s)",
 236                  average_gc * 1000,
 237                  byte_size_in_proper_unit(allocation_rate),     proper_unit_for_byte_size(allocation_rate),
 238                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom));
 239     log_info(gc, ergo)("Free headroom: " SIZE_FORMAT "%s (free) - " SIZE_FORMAT "%s (spike) - " SIZE_FORMAT "%s (penalties) = " SIZE_FORMAT "%s",
 240                  byte_size_in_proper_unit(available),           proper_unit_for_byte_size(available),
 241                  byte_size_in_proper_unit(spike_headroom),      proper_unit_for_byte_size(spike_headroom),
 242                  byte_size_in_proper_unit(penalties),           proper_unit_for_byte_size(penalties),
 243                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom));
 244     return true;
 245   } else if (ShenandoahHeuristics::should_start_gc()) {
 246     return true;
 247   }
 248 
 249   return false;
 250 }
 251 
 252 void ShenandoahTraversalHeuristics::choose_collection_set_from_regiondata(ShenandoahCollectionSet* set,
 253                                                                           RegionData* data, size_t data_size,
 254                                                                           size_t free) {
 255   ShouldNotReachHere();
 256 }