1 /*
   2  * Copyright (c) 2014, 2020, 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     _task_queue(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_young_words_base(NULL),
  60     _surviving_young_words(NULL),
  61     _surviving_words_length(young_cset_length + 1),
  62     _old_gen_is_full(false),
  63     _num_optional_regions(optional_cset_length),
  64     _numa(g1h->numa()),
  65     _obj_alloc_stat(NULL)
  66 {
  67   // We allocate number of young gen regions in the collection set plus one
  68   // entries, since entry 0 keeps track of surviving bytes for non-young regions.
  69   // We also add a few elements at the beginning and at the end in
  70   // an attempt to eliminate cache contention
  71   size_t array_length = PADDING_ELEM_NUM + _surviving_words_length + PADDING_ELEM_NUM;
  72   _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
  73   _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
  74   memset(_surviving_young_words, 0, _surviving_words_length * sizeof(size_t));
  75 
  76   _plab_allocator = new G1PLABAllocator(_g1h->allocator());
  77 
  78   // The dest for Young is used when the objects are aged enough to
  79   // need to be moved to the next space.
  80   _dest[G1HeapRegionAttr::Young] = G1HeapRegionAttr::Old;
  81   _dest[G1HeapRegionAttr::Old]   = G1HeapRegionAttr::Old;
  82 
  83   _closures = G1EvacuationRootClosures::create_root_closures(this, _g1h);
  84 
  85   _oops_into_optional_regions = new G1OopStarChunkedList[_num_optional_regions];
  86 
  87   initialize_numa_stats();
  88 }
  89 
  90 size_t G1ParScanThreadState::flush(size_t* surviving_young_words) {
  91   _rdcq.flush();
  92   flush_numa_stats();
  93   // Update allocation statistics.
  94   _plab_allocator->flush_and_retire_stats();
  95   _g1h->policy()->record_age_table(&_age_table);
  96 
  97   size_t sum = 0;
  98   for (uint i = 0; i < _surviving_words_length; i++) {
  99     surviving_young_words[i] += _surviving_young_words[i];
 100     sum += _surviving_young_words[i];
 101   }
 102   return sum;
 103 }
 104 
 105 G1ParScanThreadState::~G1ParScanThreadState() {
 106   delete _plab_allocator;
 107   delete _closures;
 108   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_base);
 109   delete[] _oops_into_optional_regions;
 110   FREE_C_HEAP_ARRAY(size_t, _obj_alloc_stat);
 111 }
 112 
 113 size_t G1ParScanThreadState::lab_waste_words() const {
 114   return _plab_allocator->waste();
 115 }
 116 
 117 size_t G1ParScanThreadState::lab_undo_waste_words() const {
 118   return _plab_allocator->undo_waste();
 119 }
 120 
 121 #ifdef ASSERT
 122 void G1ParScanThreadState::verify_task(narrowOop* task) const {
 123   assert(task != NULL, "invariant");
 124   assert(UseCompressedOops, "sanity");
 125   oop p = RawAccess<>::oop_load(task);
 126   assert(_g1h->is_in_g1_reserved(p),
 127          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
 128 }
 129 
 130 void G1ParScanThreadState::verify_task(oop* task) const {
 131   assert(task != NULL, "invariant");
 132   oop p = RawAccess<>::oop_load(task);
 133   assert(_g1h->is_in_g1_reserved(p),
 134          "task=" PTR_FORMAT " p=" PTR_FORMAT, p2i(task), p2i(p));
 135 }
 136 
 137 void G1ParScanThreadState::verify_task(PartialArrayScanTask task) const {
 138   // Must be in the collection set--it's already been copied.
 139   oop p = task.to_source_array();
 140   assert(_g1h->is_in_cset(p), "p=" PTR_FORMAT, p2i(p));
 141 }
 142 
 143 void G1ParScanThreadState::verify_task(ScannerTask task) const {
 144   if (task.is_narrow_oop_ptr()) {
 145     verify_task(task.to_narrow_oop_ptr());
 146   } else if (task.is_oop_ptr()) {
 147     verify_task(task.to_oop_ptr());
 148   } else if (task.is_partial_array_task()) {
 149     verify_task(task.to_partial_array_task());
 150   } else {
 151     ShouldNotReachHere();
 152   }
 153 }
 154 #endif // ASSERT
 155 
 156 void G1ParScanThreadState::trim_queue() {
 157   do {
 158     // Fully drain the queue.
 159     trim_queue_to_threshold(0);
 160   } while (!_task_queue->is_empty());
 161 }
 162 
 163 HeapWord* G1ParScanThreadState::allocate_in_next_plab(G1HeapRegionAttr* dest,
 164                                                       size_t word_sz,
 165                                                       bool previous_plab_refill_failed,
 166                                                       uint node_index) {
 167 
 168   assert(dest->is_in_cset_or_humongous(), "Unexpected dest: %s region attr", dest->get_type_str());
 169 
 170   // Right now we only have two types of regions (young / old) so
 171   // let's keep the logic here simple. We can generalize it when necessary.
 172   if (dest->is_young()) {
 173     bool plab_refill_in_old_failed = false;
 174     HeapWord* const obj_ptr = _plab_allocator->allocate(G1HeapRegionAttr::Old,
 175                                                         word_sz,
 176                                                         &plab_refill_in_old_failed,
 177                                                         node_index);
 178     // Make sure that we won't attempt to copy any other objects out
 179     // of a survivor region (given that apparently we cannot allocate
 180     // any new ones) to avoid coming into this slow path again and again.
 181     // Only consider failed PLAB refill here: failed inline allocations are
 182     // typically large, so not indicative of remaining space.
 183     if (previous_plab_refill_failed) {
 184       _tenuring_threshold = 0;
 185     }
 186 
 187     if (obj_ptr != NULL) {
 188       dest->set_old();
 189     } else {
 190       // We just failed to allocate in old gen. The same idea as explained above
 191       // for making survivor gen unavailable for allocation applies for old gen.
 192       _old_gen_is_full = plab_refill_in_old_failed;
 193     }
 194     return obj_ptr;
 195   } else {
 196     _old_gen_is_full = previous_plab_refill_failed;
 197     assert(dest->is_old(), "Unexpected dest region attr: %s", dest->get_type_str());
 198     // no other space to try.
 199     return NULL;
 200   }
 201 }
 202 
 203 G1HeapRegionAttr G1ParScanThreadState::next_region_attr(G1HeapRegionAttr const region_attr, markWord const m, uint& age) {
 204   if (region_attr.is_young()) {
 205     age = !m.has_displaced_mark_helper() ? m.age()
 206                                          : m.displaced_mark_helper().age();
 207     if (age < _tenuring_threshold) {
 208       return region_attr;
 209     }
 210   }
 211   return dest(region_attr);
 212 }
 213 
 214 void G1ParScanThreadState::report_promotion_event(G1HeapRegionAttr const dest_attr,
 215                                                   oop const old, size_t word_sz, uint age,
 216                                                   HeapWord * const obj_ptr, uint node_index) const {
 217   PLAB* alloc_buf = _plab_allocator->alloc_buffer(dest_attr, node_index);
 218   if (alloc_buf->contains(obj_ptr)) {
 219     _g1h->_gc_tracer_stw->report_promotion_in_new_plab_event(old->klass(), word_sz * HeapWordSize, age,
 220                                                              dest_attr.type() == G1HeapRegionAttr::Old,
 221                                                              alloc_buf->word_sz() * HeapWordSize);
 222   } else {
 223     _g1h->_gc_tracer_stw->report_promotion_outside_plab_event(old->klass(), word_sz * HeapWordSize, age,
 224                                                               dest_attr.type() == G1HeapRegionAttr::Old);
 225   }
 226 }
 227 
 228 oop G1ParScanThreadState::copy_to_survivor_space(G1HeapRegionAttr const region_attr,
 229                                                  oop const old,
 230                                                  markWord const old_mark) {
 231   const size_t word_sz = old->size();
 232 
 233   uint age = 0;
 234   G1HeapRegionAttr dest_attr = next_region_attr(region_attr, old_mark, age);
 235   // The second clause is to prevent premature evacuation failure in case there
 236   // is still space in survivor, but old gen is full.
 237   if (_old_gen_is_full && dest_attr.is_old()) {
 238     return handle_evacuation_failure_par(old, old_mark);
 239   }
 240   HeapRegion* const from_region = _g1h->heap_region_containing(old);
 241   uint node_index = from_region->node_index();
 242 
 243   HeapWord* obj_ptr = _plab_allocator->plab_allocate(dest_attr, word_sz, node_index);
 244 
 245   // PLAB allocations should succeed most of the time, so we'll
 246   // normally check against NULL once and that's it.
 247   if (obj_ptr == NULL) {
 248     bool plab_refill_failed = false;
 249     obj_ptr = _plab_allocator->allocate_direct_or_new_plab(dest_attr, word_sz, &plab_refill_failed, node_index);
 250     if (obj_ptr == NULL) {
 251       assert(region_attr.is_in_cset(), "Unexpected region attr type: %s", region_attr.get_type_str());
 252       obj_ptr = allocate_in_next_plab(&dest_attr, word_sz, plab_refill_failed, node_index);
 253       if (obj_ptr == NULL) {
 254         // This will either forward-to-self, or detect that someone else has
 255         // installed a forwarding pointer.
 256         return handle_evacuation_failure_par(old, old_mark);
 257       }
 258     }
 259     update_numa_stats(node_index);
 260 
 261     if (_g1h->_gc_tracer_stw->should_report_promotion_events()) {
 262       // The events are checked individually as part of the actual commit
 263       report_promotion_event(dest_attr, old, word_sz, age, obj_ptr, node_index);
 264     }
 265   }
 266 
 267   assert(obj_ptr != NULL, "when we get here, allocation should have succeeded");
 268   assert(_g1h->is_in_reserved(obj_ptr), "Allocated memory should be in the heap");
 269 
 270 #ifndef PRODUCT
 271   // Should this evacuation fail?
 272   if (_g1h->evacuation_should_fail()) {
 273     // Doing this after all the allocation attempts also tests the
 274     // undo_allocation() method too.
 275     _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
 276     return handle_evacuation_failure_par(old, old_mark);
 277   }
 278 #endif // !PRODUCT
 279 
 280   // We're going to allocate linearly, so might as well prefetch ahead.
 281   Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
 282 
 283   const oop obj = oop(obj_ptr);
 284   const oop forward_ptr = old->forward_to_atomic(obj, old_mark, memory_order_relaxed);
 285   if (forward_ptr == NULL) {
 286     Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(old), obj_ptr, word_sz);
 287 
 288     const uint young_index = from_region->young_index_in_cset();
 289 
 290     assert((from_region->is_young() && young_index >  0) ||
 291            (!from_region->is_young() && young_index == 0), "invariant" );
 292 
 293     if (dest_attr.is_young()) {
 294       if (age < markWord::max_age) {
 295         age++;
 296       }
 297       if (old_mark.has_displaced_mark_helper()) {
 298         // In this case, we have to install the mark word first,
 299         // otherwise obj looks to be forwarded (the old mark word,
 300         // which contains the forward pointer, was copied)
 301         obj->set_mark_raw(old_mark);
 302         markWord new_mark = old_mark.displaced_mark_helper().set_age(age);
 303         old_mark.set_displaced_mark_helper(new_mark);
 304       } else {
 305         obj->set_mark_raw(old_mark.set_age(age));
 306       }
 307       _age_table.add(age, word_sz);
 308     } else {
 309       obj->set_mark_raw(old_mark);
 310     }
 311 
 312     if (G1StringDedup::is_enabled()) {
 313       const bool is_from_young = region_attr.is_young();
 314       const bool is_to_young = dest_attr.is_young();
 315       assert(is_from_young == from_region->is_young(),
 316              "sanity");
 317       assert(is_to_young == _g1h->heap_region_containing(obj)->is_young(),
 318              "sanity");
 319       G1StringDedup::enqueue_from_evacuation(is_from_young,
 320                                              is_to_young,
 321                                              _worker_id,
 322                                              obj);
 323     }
 324 
 325     _surviving_young_words[young_index] += word_sz;
 326 
 327     if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
 328       // We keep track of the next start index in the length field of
 329       // the to-space object. The actual length can be found in the
 330       // length field of the from-space object.
 331       arrayOop(obj)->set_length(0);
 332       do_partial_array(PartialArrayScanTask(old));
 333     } else {
 334       G1ScanInYoungSetter x(&_scanner, dest_attr.is_young());
 335       obj->oop_iterate_backwards(&_scanner);
 336     }
 337     return obj;
 338   } else {
 339     _plab_allocator->undo_allocation(dest_attr, obj_ptr, word_sz, node_index);
 340     return forward_ptr;
 341   }
 342 }
 343 
 344 G1ParScanThreadState* G1ParScanThreadStateSet::state_for_worker(uint worker_id) {
 345   assert(worker_id < _n_workers, "out of bounds access");
 346   if (_states[worker_id] == NULL) {
 347     _states[worker_id] =
 348       new G1ParScanThreadState(_g1h, _rdcqs, worker_id, _young_cset_length, _optional_cset_length);
 349   }
 350   return _states[worker_id];
 351 }
 352 
 353 const size_t* G1ParScanThreadStateSet::surviving_young_words() const {
 354   assert(_flushed, "thread local state from the per thread states should have been flushed");
 355   return _surviving_young_words_total;
 356 }
 357 
 358 void G1ParScanThreadStateSet::flush() {
 359   assert(!_flushed, "thread local state from the per thread states should be flushed once");
 360 
 361   for (uint worker_id = 0; worker_id < _n_workers; ++worker_id) {
 362     G1ParScanThreadState* pss = _states[worker_id];
 363 
 364     if (pss == NULL) {
 365       continue;
 366     }
 367 
 368     G1GCPhaseTimes* p = _g1h->phase_times();
 369 
 370     // Need to get the following two before the call to G1ParThreadScanState::flush()
 371     // because it resets the PLAB allocator where we get this info from.
 372     size_t lab_waste_bytes = pss->lab_waste_words() * HeapWordSize;
 373     size_t lab_undo_waste_bytes = pss->lab_undo_waste_words() * HeapWordSize;
 374     size_t copied_bytes = pss->flush(_surviving_young_words_total) * HeapWordSize;
 375 
 376     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, copied_bytes, G1GCPhaseTimes::MergePSSCopiedBytes);
 377     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_waste_bytes, G1GCPhaseTimes::MergePSSLABWasteBytes);
 378     p->record_or_add_thread_work_item(G1GCPhaseTimes::MergePSS, worker_id, lab_undo_waste_bytes, G1GCPhaseTimes::MergePSSLABUndoWasteBytes);
 379 
 380     delete pss;
 381     _states[worker_id] = NULL;
 382   }
 383   _flushed = true;
 384 }
 385 
 386 void G1ParScanThreadStateSet::record_unused_optional_region(HeapRegion* hr) {
 387   for (uint worker_index = 0; worker_index < _n_workers; ++worker_index) {
 388     G1ParScanThreadState* pss = _states[worker_index];
 389 
 390     if (pss == NULL) {
 391       continue;
 392     }
 393 
 394     size_t used_memory = pss->oops_into_optional_region(hr)->used_memory();
 395     _g1h->phase_times()->record_or_add_thread_work_item(G1GCPhaseTimes::OptScanHR, worker_index, used_memory, G1GCPhaseTimes::ScanHRUsedMemory);
 396   }
 397 }
 398 
 399 oop G1ParScanThreadState::handle_evacuation_failure_par(oop old, markWord m) {
 400   assert(_g1h->is_in_cset(old), "Object " PTR_FORMAT " should be in the CSet", p2i(old));
 401 
 402   oop forward_ptr = old->forward_to_atomic(old, m, memory_order_relaxed);
 403   if (forward_ptr == NULL) {
 404     // Forward-to-self succeeded. We are the "owner" of the object.
 405     HeapRegion* r = _g1h->heap_region_containing(old);
 406 
 407     if (!r->evacuation_failed()) {
 408       r->set_evacuation_failed(true);
 409      _g1h->hr_printer()->evac_failure(r);
 410     }
 411 
 412     _g1h->preserve_mark_during_evac_failure(_worker_id, old, m);
 413 
 414     G1ScanInYoungSetter x(&_scanner, r->is_young());
 415     old->oop_iterate_backwards(&_scanner);
 416 
 417     return old;
 418   } else {
 419     // Forward-to-self failed. Either someone else managed to allocate
 420     // space for this object (old != forward_ptr) or they beat us in
 421     // self-forwarding it (old == forward_ptr).
 422     assert(old == forward_ptr || !_g1h->is_in_cset(forward_ptr),
 423            "Object " PTR_FORMAT " forwarded to: " PTR_FORMAT " "
 424            "should not be in the CSet",
 425            p2i(old), p2i(forward_ptr));
 426     return forward_ptr;
 427   }
 428 }
 429 G1ParScanThreadStateSet::G1ParScanThreadStateSet(G1CollectedHeap* g1h,
 430                                                  G1RedirtyCardsQueueSet* rdcqs,
 431                                                  uint n_workers,
 432                                                  size_t young_cset_length,
 433                                                  size_t optional_cset_length) :
 434     _g1h(g1h),
 435     _rdcqs(rdcqs),
 436     _states(NEW_C_HEAP_ARRAY(G1ParScanThreadState*, n_workers, mtGC)),
 437     _surviving_young_words_total(NEW_C_HEAP_ARRAY(size_t, young_cset_length + 1, mtGC)),
 438     _young_cset_length(young_cset_length),
 439     _optional_cset_length(optional_cset_length),
 440     _n_workers(n_workers),
 441     _flushed(false) {
 442   for (uint i = 0; i < n_workers; ++i) {
 443     _states[i] = NULL;
 444   }
 445   memset(_surviving_young_words_total, 0, (young_cset_length + 1) * sizeof(size_t));
 446 }
 447 
 448 G1ParScanThreadStateSet::~G1ParScanThreadStateSet() {
 449   assert(_flushed, "thread local state from the per thread states should have been flushed");
 450   FREE_C_HEAP_ARRAY(G1ParScanThreadState*, _states);
 451   FREE_C_HEAP_ARRAY(size_t, _surviving_young_words_total);
 452 }