rev 57223 : imported patch 8225484-changes-to-survivor-calculation

   1 /*
   2  * Copyright (c) 2014, 2019, Oracle and/or its affiliates. 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 #include "gc/g1/g1Allocator.inline.hpp"
  27 #include "gc/g1/g1CollectedHeap.inline.hpp"
  28 #include "gc/g1/g1CollectionSet.hpp"
  29 #include "gc/g1/g1OopClosures.inline.hpp"
  30 #include "gc/g1/g1ParScanThreadState.inline.hpp"
  31 #include "gc/g1/g1RootClosures.hpp"
  32 #include "gc/g1/g1StringDedup.hpp"
  33 #include "gc/g1/g1Trace.hpp"
  34 #include "gc/shared/taskqueue.inline.hpp"
  35 #include "memory/allocation.inline.hpp"
  36 #include "oops/access.inline.hpp"
  37 #include "oops/oop.inline.hpp"
  38 #include "runtime/prefetch.inline.hpp"
  39 
  40 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h,
  41                                            G1RedirtyCardsQueueSet* rdcqs,
  42                                            uint worker_id,
  43                                            size_t young_cset_length,
  44                                            size_t optional_cset_length)
  45   : _g1h(g1h),
  46     _refs(g1h->task_queue(worker_id)),
  47     _rdcq(rdcqs),
  48     _ct(g1h->card_table()),
  49     _closures(NULL),
  50     _plab_allocator(NULL),
  51     _age_table(false),
  52     _tenuring_threshold(g1h->policy()->tenuring_threshold()),
  53     _scanner(g1h, this),
  54     _worker_id(worker_id),
  55     _last_enqueued_card(SIZE_MAX),
  56     _stack_trim_upper_threshold(GCDrainStackTargetSize * 2 + 1),
  57     _stack_trim_lower_threshold(GCDrainStackTargetSize),
  58     _trim_ticks(),
  59     _surviving_survivor_words(0),
  60     _surviving_young_words_base(NULL),
  61     _surviving_young_words(NULL),
  62     _surviving_words_length(young_cset_length + 1),
  63     _old_gen_is_full(false),
  64     _num_optional_regions(optional_cset_length),
  65     _numa(g1h->numa()),
  66     _obj_alloc_stat(NULL)
  67 {
  68   // We allocate number of young gen regions in the collection set plus one
  69   // entries, since entry 0 keeps track of surviving bytes for non-young regions.
  70   // We also add a few elements at the beginning and at the end in
  71   // an attempt to eliminate cache contention
  72   size_t array_length = PADDING_ELEM_NUM + _surviving_words_length + PADDING_ELEM_NUM;
  73   _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
  74   _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
  75   memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
  76 
  77   _plab_allocator = new G1PLABAllocator(_g1h->allocator());
  78 
  79   // The dest for Young is used when the objects are aged enough to
  80   // need to be moved to the next space.
  81   _dest[G1HeapRegionAttr::Young] = G1HeapRegionAttr::Old;
  82   _dest[G1HeapRegionAttr::Old]   = G1HeapRegionAttr::Old;
  83 
  84   _closures = G1EvacuationRootClosures::create_root_closures(this, _g1h);
  85 
  86   _oops_into_optional_regions = new G1OopStarChunkedList[_num_optional_regions];
  87 
  88   initialize_numa_stats();
  89 }
  90 
  91 size_t G1ParScanThreadState::flush(size_t* surviving_young_words) {
  92   _rdcq.flush();
  93   flush_numa_stats();
  94   // Update allocation statistics.
  95   _plab_allocator->flush_and_retire_stats();
  96   _g1h->policy()->record_age_table(&_age_table);
  97   _g1h->policy()->record_surviving_survivor_words(_surviving_survivor_words);
  98 
  99   size_t sum = 0;
 100   for (uint i = 0; i < _surviving_words_length; i++) {
 101     surviving_young_words[i] += _surviving_young_words[i];
 102     sum += _surviving_young_words[i];
 103   }
 104   return sum;
 105 }
 106 
 107 G1ParScanThreadState::~G1ParScanThreadState() {
 108   delete _plab_allocator;
 109   delete _closures;
 110   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
 111   delete[] _oops_into_optional_regions;
 112   FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
 113 }
 114 
 115 size_t G1ParScanThreadState::lab_waste_words() const {
 116   return _plab_allocator->waste();
 117 }
 118 
 119 size_t G1ParScanThreadState::lab_undo_waste_words() const {
 120   return _plab_allocator->undo_waste();
 121 }
 122 
 123 #ifdef ASSERT
 124 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
 125   assert(ref != NULL, "invariant");
 126   assert(UseCompressedOops, "sanity");
 127   assert(!has_partial_array_mask(ref), "ref=" PTR_FORMAT, p2i(ref));
 128   oop p = RawAccess<>::oop_load(ref);
 129   assert(_g1h->is_in_g1_reserved(p),
 130          "ref=" PTR_FORMAT " p=" PTR_FORMAT, p2i(ref), p2i(p));
 131   return true;
 132 }
 133 
 134 bool G1ParScanThreadState::verify_ref(oop* ref) const {
 135   assert(ref != NULL, "invariant");
 136   if (has_partial_array_mask(ref)) {
 137     // Must be in the collection set--it's already been copied.
 138     oop p = clear_partial_array_mask(ref);
 139     assert(_g1h->is_in_cset(p),
 140            "ref=" PTR_FORMAT " p=" PTR_FORMAT, p2i(ref), p2i(p));
 141   } else {
 142     oop p = RawAccess<>::oop_load(ref);
 143     assert(_g1h->is_in_g1_reserved(p),
 144            "ref=" PTR_FORMAT " p=" PTR_FORMAT, p2i(ref), p2i(p));
 145   }
 146   return true;
 147 }
 148 
 149 bool G1ParScanThreadState::verify_task(StarTask ref) const {
 150   if (ref.is_narrow()) {
 151     return verify_ref((narrowOop*) ref);
 152   } else {
 153     return verify_ref((oop*) ref);
 154   }
 155 }
 156 #endif // ASSERT
 157 
 158 void G1ParScanThreadState::trim_queue() {
 159   StarTask ref;
 160   do {
 161     // Fully drain the queue.
 162     trim_queue_to_threshold(0);
 163   } while (!_refs->is_empty());
 164 }
 165 
 166 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
 167                                                       size_t word_sz,
 168                                                       bool previous_plab_refill_failed,
 169                                                       uint node_index) {
 170 
 171   assert(dest->is_in_cset_or_humongous(), "Unexpected dest: %s region attr", dest->get_type_str());
 172 
 173   // Right now we only have two types of regions (young / old) so
 174   // let's keep the logic here simple. We can generalize it when necessary.
 175   if (dest->is_young()) {
 176     bool plab_refill_in_old_failed = false;
 177     HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
 178                                                         word_sz,
 179                                                         &plab_refill_in_old_failed,
 180                                                         node_index);
 181     // Make sure that we won't attempt to copy any other objects out
 182     // of a survivor region (given that apparently we cannot allocate
 183     // any new ones) to avoid coming into this slow path again and again.
 184     // Only consider failed PLAB refill here: failed inline allocations are
 185     // typically large, so not indicative of remaining space.
 186     if (previous_plab_refill_failed) {
 187       _tenuring_threshold = 0;
 188     }
 189 
 190     if (obj_ptr != NULL) {
 191       dest->set_old();
 192     } else {
 193       // We just failed to allocate in old gen. The same idea as explained above
 194       // for making survivor gen unavailable for allocation applies for old gen.
 195       _old_gen_is_full = plab_refill_in_old_failed;
 196     }
 197     return obj_ptr;
 198   } else {
 199     _old_gen_is_full = previous_plab_refill_failed;
 200     assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
 201     // no other space to try.
 202     return NULL;
 203   }
 204 }
 205 
 206 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
 207   if (region_attr.is_young()) {
 208     age = !m.has_displaced_mark_helper() ? m.age()
 209                                          : m.displaced_mark_helper().age();
 210     if (age < _tenuring_threshold) {
 211       return region_attr;
 212     }
 213   }
 214   return dest(region_attr);
 215 }
 216 
 217 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
 218                                                   oop const old, size_t word_sz, uint age,
 219                                                   HeapWord * const obj_ptr, uint node_index) const {
 220   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
 221   if (alloc_buf->contains(obj_ptr)) {
 222     _g1h->_gc_tracer_stw->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age,
 223                                                              dest_attr.type() == G1HeapRegionAttr::Old,
 224                                                              alloc_buf->word_sz() * HeapWordSize);
 225   } else {
 226     _g1h->_gc_tracer_stw->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age,
 227                                                               dest_attr.type() == G1HeapRegionAttr::Old);
 228   }
 229 }
 230 
 231 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr const region_attr,
 232                                                  oop const old,
 233                                                  markWord const old_mark) {
 234   const size_t word_sz = old->size();
 235 
 236   uint age = 0;
 237   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
 238   // The second clause is to prevent premature evacuation failure in case there
 239   // is still space in survivor, but old gen is full.
 240   if (_old_gen_is_full && dest_attr.is_old()) {
 241     return handle_evacuation_failure_par(old, old_mark);
 242   }
 243   HeapRegion* const from_region = _g1h->heap_region_containing(old);
 244   uint node_index = from_region->node_index();
 245 
 246   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
 247 
 248   // PLAB allocations should succeed most of the time, so we'll
 249   // normally check against NULL once and that's it.
 250   if (obj_ptr == NULL) {
 251     bool plab_refill_failed = false;
 252     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(dest_attr, word_sz, &plab_refill_failed, node_index);
 253     if (obj_ptr == NULL) {
 254       assert(region_attr.is_in_cset(), "Unexpected region attr type: %s", region_attr.get_type_str());
 255       obj_ptr = allocate_in_next_plab(&dest_attr, word_sz, plab_refill_failed, node_index);
 256       if (obj_ptr == NULL) {
 257         // This will either forward-to-self, or detect that someone else has
 258         // installed a forwarding pointer.
 259         return handle_evacuation_failure_par(old, old_mark);
 260       }
 261     }
 262     update_numa_stats(node_index);
 263 
 264     if (_g1h->_gc_tracer_stw->should_report_promotion_events()) {
 265       // The events are checked individually as part of the actual commit
 266       report_promotion_event(dest_attr, old, word_sz, age, obj_ptr, node_index);
 267     }
 268   }
 269 
 270   assert(obj_ptr != NULL, "when we get here, allocation should have succeeded");
 271   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
 272 
 273 #ifndef PRODUCT
 274   // Should this evacuation fail?
 275   if (_g1h->evacuation_should_fail()) {
 276     // Doing this after all the allocation attempts also tests the
 277     // undo_allocation() method too.
 278     _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
 279     return handle_evacuation_failure_par(old, old_mark);
 280   }
 281 #endif // !PRODUCT
 282 
 283   // We're going to allocate linearly, so might as well prefetch ahead.
 284   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
 285 
 286   const oop obj = oop(obj_ptr);
 287   const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
 288   if (forward_ptr == NULL) {
 289     Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
 290 
 291     const uint young_index = from_region->young_index_in_cset();
 292 
 293     assert(from_region->is_young() ||
 294            (!from_region->is_young() && young_index == 0), "region %u %s index %u", from_region->hrm_index(), from_region->get_short_type_str(), young_index);
 295 
 296     if (age > 0 && region_attr.is_young()) {
 297       _surviving_survivor_words += word_sz;
 298     }
 299 
 300     if (dest_attr.is_young()) {
 301       if (age < markWord::max_age) {
 302         age++;
 303       }
 304       if (old_mark.has_displaced_mark_helper()) {
 305         // In this case, we have to install the mark word first,
 306         // otherwise obj looks to be forwarded (the old mark word,
 307         // which contains the forward pointer, was copied)
 308         obj->set_mark_raw(old_mark);
 309         markWord new_mark = old_mark.displaced_mark_helper().set_age(age);
 310         old_mark.set_displaced_mark_helper(new_mark);
 311       } else {
 312         obj->set_mark_raw(old_mark.set_age(age));
 313       }
 314       _age_table.add(age, word_sz);
 315     } else {
 316       obj->set_mark_raw(old_mark);
 317     }
 318 
 319     if (G1StringDedup::is_enabled()) {
 320       const bool is_from_young = region_attr.is_young();
 321       const bool is_to_young = dest_attr.is_young();
 322       assert(is_from_young == from_region->is_young(),
 323              "sanity");
 324       assert(is_to_young == _g1h->heap_region_containing(obj)->is_young(),
 325              "sanity");
 326       G1StringDedup::enqueue_from_evacuation(is_from_young,
 327                                              is_to_young,
 328                                              _worker_id,
 329                                              obj);
 330     }
 331 
 332     _surviving_young_words[young_index] += word_sz;
 333 
 334     if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
 335       // We keep track of the next start index in the length field of
 336       // the to-space object. The actual length can be found in the
 337       // length field of the from-space object.
 338       arrayOop(obj)->set_length(0);
 339       oop* old_p = set_partial_array_mask(old);
 340       do_oop_partial_array(old_p);
 341     } else {
 342       G1ScanInYoungSetter x(&_scanner, dest_attr.is_young());
 343       obj->oop_iterate_backwards(&_scanner);
 344     }
 345     return obj;
 346   } else {
 347     _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
 348     return forward_ptr;
 349   }
 350 }
 351 
 352 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
 353   assert(worker_id < _n_workers, "out of bounds access");
 354   if (_states[worker_id] == NULL) {
 355     _states[worker_id] =
 356       new G1ParScanThreadState(_g1h, _rdcqs, worker_id, _young_cset_length, _optional_cset_length);
 357   }
 358   return _states[worker_id];
 359 }
 360 
 361 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
 362   assert(_flushed, "thread local state from the per thread states should have been flushed");
 363   return _surviving_young_words_total;
 364 }
 365 
 366 void G1ParScanThreadStateSet::flush() {
 367   assert(!_flushed, "thread local state from the per thread states should be flushed once");
 368 
 369   for (uint worker_id = 0; worker_id < _n_workers; ++worker_id) {
 370     G1ParScanThreadState* pss = _states[worker_id];
 371 
 372     if (pss == NULL) {
 373       continue;
 374     }
 375 
 376     G1GCPhaseTimes* p = _g1h->phase_times();
 377 
 378     // Need to get the following two before the call to G1ParThreadScanState::flush()
 379     // because it resets the PLAB allocator where we get this info from.
 380     size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
 381     size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
 382     size_t copied_bytes = pss->flush(_surviving_young_words_total) * HeapWordSize;
 383 
 384     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
 385     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
 386     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
 387 
 388     delete pss;
 389     _states[worker_id] = NULL;
 390   }
 391   _flushed = true;
 392 }
 393 
 394 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
 395   for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) {
 396     G1ParScanThreadState* pss = _states[worker_index];
 397 
 398     if (pss == NULL) {
 399       continue;
 400     }
 401 
 402     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
 403     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
 404   }
 405 }
 406 
 407 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m) {
 408   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
 409 
 410   oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed);
 411   if (forward_ptr == NULL) {
 412     // Forward-to-self succeeded. We are the "owner" of the object.
 413     HeapRegion* r = _g1h->heap_region_containing(old);
 414 
 415     if (!r->evacuation_failed()) {
 416       r->set_evacuation_failed(true);
 417      _g1h->hr_printer()->evac_failure(r);
 418     }
 419 
 420     _g1h->preserve_mark_during_evac_failure(_worker_id, old, m);
 421 
 422     G1ScanInYoungSetter x(&_scanner, r->is_young());
 423     old->oop_iterate_backwards(&_scanner);
 424 
 425     return old;
 426   } else {
 427     // Forward-to-self failed. Either someone else managed to allocate
 428     // space for this object (old != forward_ptr) or they beat us in
 429     // self-forwarding it (old == forward_ptr).
 430     assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
 431            "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
 432            "should not be in the CSet",
 433            p2i(old), p2i(forward_ptr));
 434     return forward_ptr;
 435   }
 436 }
 437 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
 438                                                  G1RedirtyCardsQueueSet* rdcqs,
 439                                                  uint n_workers,
 440                                                  size_t young_cset_length,
 441                                                  size_t optional_cset_length) :
 442     _g1h(g1h),
 443     _rdcqs(rdcqs),
 444     _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, n_workers, mtGC)),
 445     _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, young_cset_length + 1, mtGC)),
 446     _young_cset_length(young_cset_length),
 447     _optional_cset_length(optional_cset_length),
 448     _n_workers(n_workers),
 449     _flushed(false) {
 450   for (uint i = 0; i < n_workers; ++i) {
 451     _states[i] = NULL;
 452   }
 453   memset(_surviving_young_words_total, 0, (young_cset_length + 1) * sizeof(size_t));
 454 }
 455 
 456 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
 457   assert(_flushed, "thread local state from the per thread states should have been flushed");
 458   FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
 459   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
 460 }
--- EOF ---