1 /*
   2  * Copyright (c) 2014, 2021, 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 "compiler/oopMap.hpp"
  28 #include "gc/shared/continuationGCSupport.hpp"
  29 #include "gc/shared/gcTraceTime.inline.hpp"
  30 #include "gc/shared/preservedMarks.inline.hpp"
  31 #include "gc/shared/tlab_globals.hpp"
  32 #include "gc/shared/workerThread.hpp"
  33 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
  34 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
  35 #include "gc/shenandoah/shenandoahConcurrentGC.hpp"
  36 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
  37 #include "gc/shenandoah/shenandoahFreeSet.hpp"
  38 #include "gc/shenandoah/shenandoahFullGC.hpp"
  39 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
  40 #include "gc/shenandoah/shenandoahMark.inline.hpp"
  41 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
  42 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
  43 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
  44 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
  45 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
  46 #include "gc/shenandoah/shenandoahMetrics.hpp"
  47 #include "gc/shenandoah/shenandoahOopClosures.inline.hpp"
  48 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
  49 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
  50 #include "gc/shenandoah/shenandoahSTWMark.hpp"
  51 #include "gc/shenandoah/shenandoahUtils.hpp"
  52 #include "gc/shenandoah/shenandoahVerifier.hpp"
  53 #include "gc/shenandoah/shenandoahVMOperations.hpp"
  54 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
  55 #include "memory/metaspaceUtils.hpp"
  56 #include "memory/universe.hpp"
  57 #include "oops/compressedOops.inline.hpp"
  58 #include "oops/oop.inline.hpp"
  59 #include "runtime/javaThread.hpp"
  60 #include "runtime/orderAccess.hpp"
  61 #include "runtime/vmThread.hpp"
  62 #include "utilities/copy.hpp"
  63 #include "utilities/events.hpp"
  64 #include "utilities/growableArray.hpp"
  65 
  66 ShenandoahFullGC::ShenandoahFullGC() :
  67   _gc_timer(ShenandoahHeap::heap()->gc_timer()),
  68   _preserved_marks(new PreservedMarksSet(true)) {}
  69 
  70 ShenandoahFullGC::~ShenandoahFullGC() {
  71   delete _preserved_marks;
  72 }
  73 
  74 bool ShenandoahFullGC::collect(GCCause::Cause cause) {
  75   vmop_entry_full(cause);
  76   // Always success
  77   return true;
  78 }
  79 
  80 void ShenandoahFullGC::vmop_entry_full(GCCause::Cause cause) {
  81   ShenandoahHeap* const heap = ShenandoahHeap::heap();
  82   TraceCollectorStats tcs(heap->monitoring_support()->full_stw_collection_counters());
  83   ShenandoahTimingsTracker timing(ShenandoahPhaseTimings::full_gc_gross);
  84 
  85   heap->try_inject_alloc_failure();
  86   VM_ShenandoahFullGC op(cause, this);
  87   VMThread::execute(&op);
  88 }
  89 
  90 void ShenandoahFullGC::entry_full(GCCause::Cause cause) {
  91   static const char* msg = "Pause Full";
  92   ShenandoahPausePhase gc_phase(msg, ShenandoahPhaseTimings::full_gc, true /* log_heap_usage */);
  93   EventMark em("%s", msg);
  94 
  95   ShenandoahWorkerScope scope(ShenandoahHeap::heap()->workers(),
  96                               ShenandoahWorkerPolicy::calc_workers_for_fullgc(),
  97                               "full gc");
  98 
  99   op_full(cause);
 100 }
 101 
 102 void ShenandoahFullGC::op_full(GCCause::Cause cause) {
 103   ShenandoahMetricsSnapshot metrics;
 104   metrics.snap_before();
 105 
 106   // Perform full GC
 107   do_it(cause);
 108 
 109   ShenandoahHeap* const heap = ShenandoahHeap::heap();
 110 
 111   metrics.snap_after();
 112 
 113   if (metrics.is_good_progress()) {
 114     heap->notify_gc_progress();
 115   } else {
 116     // Nothing to do. Tell the allocation path that we have failed to make
 117     // progress, and it can finally fail.
 118     heap->notify_gc_no_progress();
 119   }
 120 
 121   // Regardless if progress was made, we record that we completed a "successful" full GC.
 122   heap->heuristics()->record_success_full();
 123   heap->shenandoah_policy()->record_success_full();
 124 }
 125 
 126 void ShenandoahFullGC::do_it(GCCause::Cause gc_cause) {
 127   ShenandoahHeap* heap = ShenandoahHeap::heap();
 128 
 129   if (ShenandoahVerify) {
 130     heap->verifier()->verify_before_fullgc();
 131   }
 132 
 133   if (VerifyBeforeGC) {
 134     Universe::verify();
 135   }
 136 
 137   // Degenerated GC may carry concurrent root flags when upgrading to
 138   // full GC. We need to reset it before mutators resume.
 139   heap->set_concurrent_strong_root_in_progress(false);
 140   heap->set_concurrent_weak_root_in_progress(false);
 141 
 142   heap->set_full_gc_in_progress(true);
 143 
 144   assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at a safepoint");
 145   assert(Thread::current()->is_VM_thread(), "Do full GC only while world is stopped");
 146 
 147   {
 148     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_pre);
 149     heap->pre_full_gc_dump(_gc_timer);
 150   }
 151 
 152   {
 153     ShenandoahGCPhase prepare_phase(ShenandoahPhaseTimings::full_gc_prepare);
 154     // Full GC is supposed to recover from any GC state:
 155 
 156     // a0. Remember if we have forwarded objects
 157     bool has_forwarded_objects = heap->has_forwarded_objects();
 158 
 159     // a1. Cancel evacuation, if in progress
 160     if (heap->is_evacuation_in_progress()) {
 161       heap->set_evacuation_in_progress(false);
 162     }
 163     assert(!heap->is_evacuation_in_progress(), "sanity");
 164 
 165     // a2. Cancel update-refs, if in progress
 166     if (heap->is_update_refs_in_progress()) {
 167       heap->set_update_refs_in_progress(false);
 168     }
 169     assert(!heap->is_update_refs_in_progress(), "sanity");
 170 
 171     // b. Cancel concurrent mark, if in progress
 172     if (heap->is_concurrent_mark_in_progress()) {
 173       ShenandoahConcurrentGC::cancel();
 174       heap->set_concurrent_mark_in_progress(false);
 175     }
 176     assert(!heap->is_concurrent_mark_in_progress(), "sanity");
 177 
 178     // c. Update roots if this full GC is due to evac-oom, which may carry from-space pointers in roots.
 179     if (has_forwarded_objects) {
 180       update_roots(true /*full_gc*/);
 181     }
 182 
 183     // d. Reset the bitmaps for new marking
 184     heap->reset_mark_bitmap();
 185     assert(heap->marking_context()->is_bitmap_clear(), "sanity");
 186     assert(!heap->marking_context()->is_complete(), "sanity");
 187 
 188     // e. Abandon reference discovery and clear all discovered references.
 189     ShenandoahReferenceProcessor* rp = heap->ref_processor();
 190     rp->abandon_partial_discovery();
 191 
 192     // f. Sync pinned region status from the CP marks
 193     heap->sync_pinned_region_status();
 194 
 195     // The rest of prologue:
 196     _preserved_marks->init(heap->workers()->active_workers());
 197 
 198     assert(heap->has_forwarded_objects() == has_forwarded_objects, "This should not change");
 199   }
 200 
 201   if (UseTLAB) {
 202     heap->gclabs_retire(ResizeTLAB);
 203     heap->tlabs_retire(ResizeTLAB);
 204   }
 205 
 206   OrderAccess::fence();
 207 
 208   phase1_mark_heap();
 209 
 210   // Once marking is done, which may have fixed up forwarded objects, we can drop it.
 211   // Coming out of Full GC, we would not have any forwarded objects.
 212   // This also prevents resolves with fwdptr from kicking in while adjusting pointers in phase3.
 213   heap->set_has_forwarded_objects(false);
 214 
 215   heap->set_full_gc_move_in_progress(true);
 216 
 217   // Setup workers for the rest
 218   OrderAccess::fence();
 219 
 220   // Initialize worker slices
 221   ShenandoahHeapRegionSet** worker_slices = NEW_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, heap->max_workers(), mtGC);
 222   for (uint i = 0; i < heap->max_workers(); i++) {
 223     worker_slices[i] = new ShenandoahHeapRegionSet();
 224   }
 225 
 226   {
 227     // The rest of code performs region moves, where region status is undefined
 228     // until all phases run together.
 229     ShenandoahHeapLocker lock(heap->lock());
 230 
 231     phase2_calculate_target_addresses(worker_slices);
 232 
 233     OrderAccess::fence();
 234 
 235     phase3_update_references();
 236 
 237     phase4_compact_objects(worker_slices);
 238 
 239     phase5_epilog();
 240   }
 241 
 242   // Resize metaspace
 243   MetaspaceGC::compute_new_size();
 244 
 245   // Free worker slices
 246   for (uint i = 0; i < heap->max_workers(); i++) {
 247     delete worker_slices[i];
 248   }
 249   FREE_C_HEAP_ARRAY(ShenandoahHeapRegionSet*, worker_slices);
 250 
 251   heap->set_full_gc_move_in_progress(false);
 252   heap->set_full_gc_in_progress(false);
 253 
 254   if (ShenandoahVerify) {
 255     heap->verifier()->verify_after_fullgc();
 256   }
 257 
 258   if (VerifyAfterGC) {
 259     Universe::verify();
 260   }
 261 
 262   {
 263     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_heapdump_post);
 264     heap->post_full_gc_dump(_gc_timer);
 265   }
 266 }
 267 
 268 class ShenandoahPrepareForMarkClosure: public ShenandoahHeapRegionClosure {
 269 private:
 270   ShenandoahMarkingContext* const _ctx;
 271 
 272 public:
 273   ShenandoahPrepareForMarkClosure() : _ctx(ShenandoahHeap::heap()->marking_context()) {}
 274 
 275   void heap_region_do(ShenandoahHeapRegion *r) {
 276     _ctx->capture_top_at_mark_start(r);
 277     r->clear_live_data();
 278   }
 279 
 280   bool is_thread_safe() { return true; }
 281 };
 282 
 283 void ShenandoahFullGC::phase1_mark_heap() {
 284   GCTraceTime(Info, gc, phases) time("Phase 1: Mark live objects", _gc_timer);
 285   ShenandoahGCPhase mark_phase(ShenandoahPhaseTimings::full_gc_mark);
 286 
 287   ShenandoahHeap* heap = ShenandoahHeap::heap();
 288 
 289   ShenandoahPrepareForMarkClosure cl;
 290   heap->parallel_heap_region_iterate(&cl);
 291 
 292   heap->set_unload_classes(heap->heuristics()->can_unload_classes());
 293 
 294   ShenandoahReferenceProcessor* rp = heap->ref_processor();
 295   // enable ("weak") refs discovery
 296   rp->set_soft_reference_policy(true); // forcefully purge all soft references
 297 
 298   ShenandoahSTWMark mark(true /*full_gc*/);
 299   mark.mark();
 300   heap->parallel_cleaning(true /* full_gc */);
 301 }
 302 
 303 class ShenandoahPrepareForCompactionObjectClosure : public ObjectClosure {
 304 private:
 305   PreservedMarks*          const _preserved_marks;
 306   ShenandoahHeap*          const _heap;
 307   GrowableArray<ShenandoahHeapRegion*>& _empty_regions;
 308   int _empty_regions_pos;
 309   ShenandoahHeapRegion*          _to_region;
 310   ShenandoahHeapRegion*          _from_region;
 311   HeapWord* _compact_point;
 312 
 313 public:
 314   ShenandoahPrepareForCompactionObjectClosure(PreservedMarks* preserved_marks,
 315                                               GrowableArray<ShenandoahHeapRegion*>& empty_regions,
 316                                               ShenandoahHeapRegion* to_region) :
 317     _preserved_marks(preserved_marks),
 318     _heap(ShenandoahHeap::heap()),
 319     _empty_regions(empty_regions),
 320     _empty_regions_pos(0),
 321     _to_region(to_region),
 322     _from_region(nullptr),
 323     _compact_point(to_region->bottom()) {}
 324 
 325   void set_from_region(ShenandoahHeapRegion* from_region) {
 326     _from_region = from_region;
 327   }
 328 
 329   void finish() {
 330     assert(_to_region != nullptr, "should not happen");
 331     _to_region->set_new_top(_compact_point);
 332   }
 333 
 334   bool is_compact_same_region() {
 335     return _from_region == _to_region;
 336   }
 337 
 338   int empty_regions_pos() {
 339     return _empty_regions_pos;
 340   }
 341 
 342   void do_object(oop p) {
 343     assert(_from_region != nullptr, "must set before work");
 344     assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
 345     assert(!_heap->complete_marking_context()->allocated_after_mark_start(p), "must be truly marked");
 346 
 347     size_t obj_size = p->size();
 348     if (_compact_point + obj_size > _to_region->end()) {
 349       finish();
 350 
 351       // Object doesn't fit. Pick next empty region and start compacting there.
 352       ShenandoahHeapRegion* new_to_region;
 353       if (_empty_regions_pos < _empty_regions.length()) {
 354         new_to_region = _empty_regions.at(_empty_regions_pos);
 355         _empty_regions_pos++;
 356       } else {
 357         // Out of empty region? Compact within the same region.
 358         new_to_region = _from_region;
 359       }
 360 
 361       assert(new_to_region != _to_region, "must not reuse same to-region");
 362       assert(new_to_region != nullptr, "must not be null");
 363       _to_region = new_to_region;
 364       _compact_point = _to_region->bottom();
 365     }
 366 
 367     // Object fits into current region, record new location:
 368     assert(_compact_point + obj_size <= _to_region->end(), "must fit");
 369     shenandoah_assert_not_forwarded(nullptr, p);
 370     _preserved_marks->push_if_necessary(p, p->mark());
 371     p->forward_to(cast_to_oop(_compact_point));
 372     _compact_point += obj_size;
 373   }
 374 };
 375 
 376 class ShenandoahPrepareForCompactionTask : public WorkerTask {
 377 private:
 378   PreservedMarksSet*        const _preserved_marks;
 379   ShenandoahHeap*           const _heap;
 380   ShenandoahHeapRegionSet** const _worker_slices;
 381 
 382 public:
 383   ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
 384     WorkerTask("Shenandoah Prepare For Compaction"),
 385     _preserved_marks(preserved_marks),
 386     _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
 387   }
 388 
 389   static bool is_candidate_region(ShenandoahHeapRegion* r) {
 390     // Empty region: get it into the slice to defragment the slice itself.
 391     // We could have skipped this without violating correctness, but we really
 392     // want to compact all live regions to the start of the heap, which sometimes
 393     // means moving them into the fully empty regions.
 394     if (r->is_empty()) return true;
 395 
 396     // Can move the region, and this is not the humongous region. Humongous
 397     // moves are special cased here, because their moves are handled separately.
 398     return r->is_stw_move_allowed() && !r->is_humongous();
 399   }
 400 
 401   void work(uint worker_id) override;
 402 private:
 403   template<typename ClosureType>
 404   void prepare_for_compaction(ClosureType& cl,
 405                               GrowableArray<ShenandoahHeapRegion*>& empty_regions,
 406                               ShenandoahHeapRegionSetIterator& it,
 407                               ShenandoahHeapRegion* from_region);
 408 };
 409 
 410 void ShenandoahPrepareForCompactionTask::work(uint worker_id) {
 411   ShenandoahParallelWorkerSession worker_session(worker_id);
 412   ShenandoahHeapRegionSet* slice = _worker_slices[worker_id];
 413   ShenandoahHeapRegionSetIterator it(slice);
 414   ShenandoahHeapRegion* from_region = it.next();
 415   // No work?
 416   if (from_region == nullptr) {
 417     return;
 418   }
 419 
 420   // Sliding compaction. Walk all regions in the slice, and compact them.
 421   // Remember empty regions and reuse them as needed.
 422   ResourceMark rm;
 423 
 424   GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions());
 425 
 426   ShenandoahPrepareForCompactionObjectClosure cl(_preserved_marks->get(worker_id), empty_regions, from_region);
 427   prepare_for_compaction(cl, empty_regions, it, from_region);
 428 }
 429 
 430 template<typename ClosureType>
 431 void ShenandoahPrepareForCompactionTask::prepare_for_compaction(ClosureType& cl,
 432                                                                 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
 433                                                                 ShenandoahHeapRegionSetIterator& it,
 434                                                                 ShenandoahHeapRegion* from_region) {
 435   while (from_region != nullptr) {
 436     assert(is_candidate_region(from_region), "Sanity");
 437     cl.set_from_region(from_region);
 438     if (from_region->has_live()) {
 439       _heap->marked_object_iterate(from_region, &cl);
 440     }
 441 
 442     // Compacted the region to somewhere else? From-region is empty then.
 443     if (!cl.is_compact_same_region()) {
 444       empty_regions.append(from_region);
 445     }
 446     from_region = it.next();
 447   }
 448   cl.finish();
 449 
 450   // Mark all remaining regions as empty
 451   for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
 452     ShenandoahHeapRegion* r = empty_regions.at(pos);
 453     r->set_new_top(r->bottom());
 454   }
 455 }
 456 
 457 void ShenandoahFullGC::calculate_target_humongous_objects() {
 458   ShenandoahHeap* heap = ShenandoahHeap::heap();
 459 
 460   // Compute the new addresses for humongous objects. We need to do this after addresses
 461   // for regular objects are calculated, and we know what regions in heap suffix are
 462   // available for humongous moves.
 463   //
 464   // Scan the heap backwards, because we are compacting humongous regions towards the end.
 465   // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
 466   // humongous start there.
 467   //
 468   // The complication is potential non-movable regions during the scan. If such region is
 469   // detected, then sliding restarts towards that non-movable region.
 470 
 471   size_t to_begin = heap->num_regions();
 472   size_t to_end = heap->num_regions();
 473 
 474   for (size_t c = heap->num_regions(); c > 0; c--) {
 475     ShenandoahHeapRegion *r = heap->get_region(c - 1);
 476     if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
 477       // To-region candidate: record this, and continue scan
 478       to_begin = r->index();
 479       continue;
 480     }
 481 
 482     if (r->is_humongous_start() && r->is_stw_move_allowed()) {
 483       // From-region candidate: movable humongous region
 484       oop old_obj = cast_to_oop(r->bottom());
 485       size_t words_size = old_obj->size();
 486       size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
 487 
 488       size_t start = to_end - num_regions;
 489 
 490       if (start >= to_begin && start != r->index()) {
 491         // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
 492         _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
 493         old_obj->forward_to(cast_to_oop(heap->get_region(start)->bottom()));
 494         to_end = start;
 495         continue;
 496       }
 497     }
 498 
 499     // Failed to fit. Scan starting from current region.
 500     to_begin = r->index();
 501     to_end = r->index();
 502   }
 503 }
 504 
 505 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
 506 private:
 507   ShenandoahHeap* const _heap;
 508 
 509 public:
 510   ShenandoahEnsureHeapActiveClosure() : _heap(ShenandoahHeap::heap()) {}
 511   void heap_region_do(ShenandoahHeapRegion* r) {
 512     if (r->is_trash()) {
 513       r->recycle();
 514     }
 515     if (r->is_cset()) {
 516       r->make_regular_bypass();
 517     }
 518     if (r->is_empty_uncommitted()) {
 519       r->make_committed_bypass();
 520     }
 521     assert (r->is_committed(), "only committed regions in heap now, see region " SIZE_FORMAT, r->index());
 522 
 523     // Record current region occupancy: this communicates empty regions are free
 524     // to the rest of Full GC code.
 525     r->set_new_top(r->top());
 526   }
 527 };
 528 
 529 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
 530 private:
 531   ShenandoahHeap* const _heap;
 532   ShenandoahMarkingContext* const _ctx;
 533 
 534 public:
 535   ShenandoahTrashImmediateGarbageClosure() :
 536     _heap(ShenandoahHeap::heap()),
 537     _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
 538 
 539   void heap_region_do(ShenandoahHeapRegion* r) {
 540     if (r->is_humongous_start()) {
 541       oop humongous_obj = cast_to_oop(r->bottom());
 542       if (!_ctx->is_marked(humongous_obj)) {
 543         assert(!r->has_live(),
 544                "Region " SIZE_FORMAT " is not marked, should not have live", r->index());
 545         _heap->trash_humongous_region_at(r);
 546       } else {
 547         assert(r->has_live(),
 548                "Region " SIZE_FORMAT " should have live", r->index());
 549       }
 550     } else if (r->is_humongous_continuation()) {
 551       // If we hit continuation, the non-live humongous starts should have been trashed already
 552       assert(r->humongous_start_region()->has_live(),
 553              "Region " SIZE_FORMAT " should have live", r->index());
 554     } else if (r->is_regular()) {
 555       if (!r->has_live()) {
 556         r->make_trash_immediate();
 557       }
 558     }
 559   }
 560 };
 561 
 562 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
 563   ShenandoahHeap* heap = ShenandoahHeap::heap();
 564 
 565   uint n_workers = heap->workers()->active_workers();
 566   size_t n_regions = heap->num_regions();
 567 
 568   // What we want to accomplish: have the dense prefix of data, while still balancing
 569   // out the parallel work.
 570   //
 571   // Assuming the amount of work is driven by the live data that needs moving, we can slice
 572   // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
 573   // thread takes all regions in its prefix subset, and then it takes some regions from
 574   // the tail.
 575   //
 576   // Tail region selection becomes interesting.
 577   //
 578   // First, we want to distribute the regions fairly between the workers, and those regions
 579   // might have different amount of live data. So, until we sure no workers need live data,
 580   // we need to only take what the worker needs.
 581   //
 582   // Second, since we slide everything to the left in each slice, the most busy regions
 583   // would be the ones on the left. Which means we want to have all workers have their after-tail
 584   // regions as close to the left as possible.
 585   //
 586   // The easiest way to do this is to distribute after-tail regions in round-robin between
 587   // workers that still need live data.
 588   //
 589   // Consider parallel workers A, B, C, then the target slice layout would be:
 590   //
 591   //  AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
 592   //
 593   //  (.....dense-prefix.....) (.....................tail...................)
 594   //  [all regions fully live] [left-most regions are fuller that right-most]
 595   //
 596 
 597   // Compute how much live data is there. This would approximate the size of dense prefix
 598   // we target to create.
 599   size_t total_live = 0;
 600   for (size_t idx = 0; idx < n_regions; idx++) {
 601     ShenandoahHeapRegion *r = heap->get_region(idx);
 602     if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
 603       total_live += r->get_live_data_words();
 604     }
 605   }
 606 
 607   // Estimate the size for the dense prefix. Note that we specifically count only the
 608   // "full" regions, so there would be some non-full regions in the slice tail.
 609   size_t live_per_worker = total_live / n_workers;
 610   size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
 611   size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
 612   prefix_regions_total = MIN2(prefix_regions_total, n_regions);
 613   assert(prefix_regions_total <= n_regions, "Sanity");
 614 
 615   // There might be non-candidate regions in the prefix. To compute where the tail actually
 616   // ends up being, we need to account those as well.
 617   size_t prefix_end = prefix_regions_total;
 618   for (size_t idx = 0; idx < prefix_regions_total; idx++) {
 619     ShenandoahHeapRegion *r = heap->get_region(idx);
 620     if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
 621       prefix_end++;
 622     }
 623   }
 624   prefix_end = MIN2(prefix_end, n_regions);
 625   assert(prefix_end <= n_regions, "Sanity");
 626 
 627   // Distribute prefix regions per worker: each thread definitely gets its own same-sized
 628   // subset of dense prefix.
 629   size_t prefix_idx = 0;
 630 
 631   size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
 632 
 633   for (size_t wid = 0; wid < n_workers; wid++) {
 634     ShenandoahHeapRegionSet* slice = worker_slices[wid];
 635 
 636     live[wid] = 0;
 637     size_t regs = 0;
 638 
 639     // Add all prefix regions for this worker
 640     while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
 641       ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
 642       if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
 643         slice->add_region(r);
 644         live[wid] += r->get_live_data_words();
 645         regs++;
 646       }
 647       prefix_idx++;
 648     }
 649   }
 650 
 651   // Distribute the tail among workers in round-robin fashion.
 652   size_t wid = n_workers - 1;
 653 
 654   for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
 655     ShenandoahHeapRegion *r = heap->get_region(tail_idx);
 656     if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
 657       assert(wid < n_workers, "Sanity");
 658 
 659       size_t live_region = r->get_live_data_words();
 660 
 661       // Select next worker that still needs live data.
 662       size_t old_wid = wid;
 663       do {
 664         wid++;
 665         if (wid == n_workers) wid = 0;
 666       } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
 667 
 668       if (old_wid == wid) {
 669         // Circled back to the same worker? This means liveness data was
 670         // miscalculated. Bump the live_per_worker limit so that
 671         // everyone gets a piece of the leftover work.
 672         live_per_worker += ShenandoahHeapRegion::region_size_words();
 673       }
 674 
 675       worker_slices[wid]->add_region(r);
 676       live[wid] += live_region;
 677     }
 678   }
 679 
 680   FREE_C_HEAP_ARRAY(size_t, live);
 681 
 682 #ifdef ASSERT
 683   ResourceBitMap map(n_regions);
 684   for (size_t wid = 0; wid < n_workers; wid++) {
 685     ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
 686     ShenandoahHeapRegion* r = it.next();
 687     while (r != nullptr) {
 688       size_t idx = r->index();
 689       assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: " SIZE_FORMAT, idx);
 690       assert(!map.at(idx), "No region distributed twice: " SIZE_FORMAT, idx);
 691       map.at_put(idx, true);
 692       r = it.next();
 693     }
 694   }
 695 
 696   for (size_t rid = 0; rid < n_regions; rid++) {
 697     bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
 698     bool is_distributed = map.at(rid);
 699     assert(is_distributed || !is_candidate, "All candidates are distributed: " SIZE_FORMAT, rid);
 700   }
 701 #endif
 702 }
 703 
 704 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
 705   GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
 706   ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
 707 
 708   ShenandoahHeap* heap = ShenandoahHeap::heap();
 709 
 710   // About to figure out which regions can be compacted, make sure pinning status
 711   // had been updated in GC prologue.
 712   heap->assert_pinned_region_status();
 713 
 714   {
 715     // Trash the immediately collectible regions before computing addresses
 716     ShenandoahTrashImmediateGarbageClosure tigcl;
 717     heap->heap_region_iterate(&tigcl);
 718 
 719     // Make sure regions are in good state: committed, active, clean.
 720     // This is needed because we are potentially sliding the data through them.
 721     ShenandoahEnsureHeapActiveClosure ecl;
 722     heap->heap_region_iterate(&ecl);
 723   }
 724 
 725   // Compute the new addresses for regular objects
 726   {
 727     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
 728 
 729     distribute_slices(worker_slices);
 730 
 731     ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
 732     heap->workers()->run_task(&task);
 733   }
 734 
 735   // Compute the new addresses for humongous objects
 736   {
 737     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
 738     calculate_target_humongous_objects();
 739   }
 740 }
 741 
 742 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
 743 private:
 744   ShenandoahHeap* const _heap;
 745   ShenandoahMarkingContext* const _ctx;
 746 
 747   template <class T>
 748   inline void do_oop_work(T* p) {
 749     T o = RawAccess<>::oop_load(p);
 750     if (!CompressedOops::is_null(o)) {
 751       oop obj = CompressedOops::decode_not_null(o);
 752       assert(_ctx->is_marked(obj), "must be marked");
 753       if (obj->is_forwarded()) {
 754         oop forw = obj->forwardee();
 755         RawAccess<IS_NOT_NULL>::oop_store(p, forw);
 756       }
 757     }
 758   }
 759 
 760 public:
 761   ShenandoahAdjustPointersClosure() :
 762     _heap(ShenandoahHeap::heap()),
 763     _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
 764 
 765   void do_oop(oop* p)       { do_oop_work(p); }
 766   void do_oop(narrowOop* p) { do_oop_work(p); }
 767   void do_method(Method* m) {}
 768   void do_nmethod(nmethod* nm) {}
 769 };
 770 
 771 class ShenandoahAdjustPointersObjectClosure : public ObjectClosure {
 772 private:
 773   ShenandoahHeap* const _heap;
 774   ShenandoahAdjustPointersClosure _cl;
 775 
 776 public:
 777   ShenandoahAdjustPointersObjectClosure() :
 778     _heap(ShenandoahHeap::heap()) {
 779   }
 780   void do_object(oop p) {
 781     assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
 782     p->oop_iterate(&_cl);
 783   }
 784 };
 785 
 786 class ShenandoahAdjustPointersTask : public WorkerTask {
 787 private:
 788   ShenandoahHeap*          const _heap;
 789   ShenandoahRegionIterator       _regions;
 790 
 791 public:
 792   ShenandoahAdjustPointersTask() :
 793     WorkerTask("Shenandoah Adjust Pointers"),
 794     _heap(ShenandoahHeap::heap()) {
 795   }
 796 
 797   void work(uint worker_id) {
 798     ShenandoahParallelWorkerSession worker_session(worker_id);
 799     ShenandoahAdjustPointersObjectClosure obj_cl;
 800     ShenandoahHeapRegion* r = _regions.next();
 801     while (r != nullptr) {
 802       if (!r->is_humongous_continuation() && r->has_live()) {
 803         _heap->marked_object_iterate(r, &obj_cl);
 804       }
 805       r = _regions.next();
 806     }
 807   }
 808 };
 809 
 810 class ShenandoahAdjustRootPointersTask : public WorkerTask {
 811 private:
 812   ShenandoahRootAdjuster* _rp;
 813   PreservedMarksSet* _preserved_marks;
 814 public:
 815   ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
 816     WorkerTask("Shenandoah Adjust Root Pointers"),
 817     _rp(rp),
 818     _preserved_marks(preserved_marks) {}
 819 
 820   void work(uint worker_id) {
 821     ShenandoahParallelWorkerSession worker_session(worker_id);
 822     ShenandoahAdjustPointersClosure cl;
 823     _rp->roots_do(worker_id, &cl);
 824     _preserved_marks->get(worker_id)->adjust_during_full_gc();
 825   }
 826 };
 827 
 828 void ShenandoahFullGC::phase3_update_references() {
 829   GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer);
 830   ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers);
 831 
 832   ShenandoahHeap* heap = ShenandoahHeap::heap();
 833 
 834   WorkerThreads* workers = heap->workers();
 835   uint nworkers = workers->active_workers();
 836   {
 837 #if COMPILER2_OR_JVMCI
 838     DerivedPointerTable::clear();
 839 #endif
 840     ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
 841     ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
 842     workers->run_task(&task);
 843 #if COMPILER2_OR_JVMCI
 844     DerivedPointerTable::update_pointers();
 845 #endif
 846   }
 847 
 848   ShenandoahAdjustPointersTask adjust_pointers_task;
 849   workers->run_task(&adjust_pointers_task);
 850 }
 851 
 852 class ShenandoahCompactObjectsClosure : public ObjectClosure {
 853 private:
 854   ShenandoahHeap* const _heap;
 855   uint            const _worker_id;
 856 
 857 public:
 858   ShenandoahCompactObjectsClosure(uint worker_id) :
 859     _heap(ShenandoahHeap::heap()), _worker_id(worker_id) {}
 860 
 861   void do_object(oop p) {
 862     assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
 863     size_t size = p->size();
 864     if (p->is_forwarded()) {
 865       HeapWord* compact_from = cast_from_oop<HeapWord*>(p);
 866       HeapWord* compact_to = cast_from_oop<HeapWord*>(p->forwardee());
 867       Copy::aligned_conjoint_words(compact_from, compact_to, size);
 868       oop new_obj = cast_to_oop(compact_to);
 869 
 870       ContinuationGCSupport::relativize_stack_chunk(new_obj);
 871       new_obj->init_mark();
 872     }
 873   }
 874 };
 875 
 876 class ShenandoahCompactObjectsTask : public WorkerTask {
 877 private:
 878   ShenandoahHeap* const _heap;
 879   ShenandoahHeapRegionSet** const _worker_slices;
 880 
 881 public:
 882   ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) :
 883     WorkerTask("Shenandoah Compact Objects"),
 884     _heap(ShenandoahHeap::heap()),
 885     _worker_slices(worker_slices) {
 886   }
 887 
 888   void work(uint worker_id) {
 889     ShenandoahParallelWorkerSession worker_session(worker_id);
 890     ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]);
 891 
 892     ShenandoahCompactObjectsClosure cl(worker_id);
 893     ShenandoahHeapRegion* r = slice.next();
 894     while (r != nullptr) {
 895       assert(!r->is_humongous(), "must not get humongous regions here");
 896       if (r->has_live()) {
 897         _heap->marked_object_iterate(r, &cl);
 898       }
 899       r->set_top(r->new_top());
 900       r = slice.next();
 901     }
 902   }
 903 };
 904 
 905 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure {
 906 private:
 907   ShenandoahHeap* const _heap;
 908   size_t _live;
 909 
 910 public:
 911   ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()), _live(0) {
 912     _heap->free_set()->clear();
 913   }
 914 
 915   void heap_region_do(ShenandoahHeapRegion* r) {
 916     assert (!r->is_cset(), "cset regions should have been demoted already");
 917 
 918     // Need to reset the complete-top-at-mark-start pointer here because
 919     // the complete marking bitmap is no longer valid. This ensures
 920     // size-based iteration in marked_object_iterate().
 921     // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip
 922     // pinned regions.
 923     if (!r->is_pinned()) {
 924       _heap->complete_marking_context()->reset_top_at_mark_start(r);
 925     }
 926 
 927     size_t live = r->used();
 928 
 929     // Make empty regions that have been allocated into regular
 930     if (r->is_empty() && live > 0) {
 931       r->make_regular_bypass();
 932       if (ZapUnusedHeapArea) {
 933         SpaceMangler::mangle_region(MemRegion(r->top(), r->end()));
 934       }
 935     }
 936 
 937     // Reclaim regular regions that became empty
 938     if (r->is_regular() && live == 0) {
 939       r->make_trash();
 940     }
 941 
 942     // Recycle all trash regions
 943     if (r->is_trash()) {
 944       live = 0;
 945       r->recycle();
 946     }
 947 
 948     r->set_live_data(live);
 949     r->reset_alloc_metadata();
 950     _live += live;
 951   }
 952 
 953   size_t get_live() {
 954     return _live;
 955   }
 956 };
 957 
 958 void ShenandoahFullGC::compact_humongous_objects() {
 959   // Compact humongous regions, based on their fwdptr objects.
 960   //
 961   // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
 962   // humongous regions are already compacted, and do not require further moves, which alleviates
 963   // sliding costs. We may consider doing this in parallel in the future.
 964 
 965   ShenandoahHeap* heap = ShenandoahHeap::heap();
 966 
 967   for (size_t c = heap->num_regions(); c > 0; c--) {
 968     ShenandoahHeapRegion* r = heap->get_region(c - 1);
 969     if (r->is_humongous_start()) {
 970       oop old_obj = cast_to_oop(r->bottom());
 971       if (!old_obj->is_forwarded()) {
 972         // No need to move the object, it stays at the same slot
 973         continue;
 974       }
 975       size_t words_size = old_obj->size();
 976       size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
 977 
 978       size_t old_start = r->index();
 979       size_t old_end   = old_start + num_regions - 1;
 980       size_t new_start = heap->heap_region_index_containing(old_obj->forwardee());
 981       size_t new_end   = new_start + num_regions - 1;
 982       assert(old_start != new_start, "must be real move");
 983       assert(r->is_stw_move_allowed(), "Region " SIZE_FORMAT " should be movable", r->index());
 984 
 985       Copy::aligned_conjoint_words(r->bottom(), heap->get_region(new_start)->bottom(), words_size);
 986       ContinuationGCSupport::relativize_stack_chunk(cast_to_oop<HeapWord*>(r->bottom()));
 987 
 988       oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom());
 989       new_obj->init_mark();
 990 
 991       {
 992         for (size_t c = old_start; c <= old_end; c++) {
 993           ShenandoahHeapRegion* r = heap->get_region(c);
 994           r->make_regular_bypass();
 995           r->set_top(r->bottom());
 996         }
 997 
 998         for (size_t c = new_start; c <= new_end; c++) {
 999           ShenandoahHeapRegion* r = heap->get_region(c);
1000           if (c == new_start) {
1001             r->make_humongous_start_bypass();
1002           } else {
1003             r->make_humongous_cont_bypass();
1004           }
1005 
1006           // Trailing region may be non-full, record the remainder there
1007           size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
1008           if ((c == new_end) && (remainder != 0)) {
1009             r->set_top(r->bottom() + remainder);
1010           } else {
1011             r->set_top(r->end());
1012           }
1013 
1014           r->reset_alloc_metadata();
1015         }
1016       }
1017     }
1018   }
1019 }
1020 
1021 // This is slightly different to ShHeap::reset_next_mark_bitmap:
1022 // we need to remain able to walk pinned regions.
1023 // Since pinned region do not move and don't get compacted, we will get holes with
1024 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
1025 // cannot be iterated over using oop->size(). The only way to safely iterate over those is using
1026 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
1027 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
1028 class ShenandoahMCResetCompleteBitmapTask : public WorkerTask {
1029 private:
1030   ShenandoahRegionIterator _regions;
1031 
1032 public:
1033   ShenandoahMCResetCompleteBitmapTask() :
1034     WorkerTask("Shenandoah Reset Bitmap") {
1035   }
1036 
1037   void work(uint worker_id) {
1038     ShenandoahParallelWorkerSession worker_session(worker_id);
1039     ShenandoahHeapRegion* region = _regions.next();
1040     ShenandoahHeap* heap = ShenandoahHeap::heap();
1041     ShenandoahMarkingContext* const ctx = heap->complete_marking_context();
1042     while (region != nullptr) {
1043       if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
1044         ctx->clear_bitmap(region);
1045       }
1046       region = _regions.next();
1047     }
1048   }
1049 };
1050 
1051 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
1052   GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
1053   ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
1054 
1055   ShenandoahHeap* heap = ShenandoahHeap::heap();
1056 
1057   // Compact regular objects first
1058   {
1059     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
1060     ShenandoahCompactObjectsTask compact_task(worker_slices);
1061     heap->workers()->run_task(&compact_task);
1062   }
1063 
1064   // Compact humongous objects after regular object moves
1065   {
1066     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
1067     compact_humongous_objects();
1068   }
1069 }
1070 
1071 void ShenandoahFullGC::phase5_epilog() {
1072   GCTraceTime(Info, gc, phases) time("Phase 5: Full GC epilog", _gc_timer);
1073   ShenandoahHeap* heap = ShenandoahHeap::heap();
1074 
1075   // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1076   // and must ensure the bitmap is in sync.
1077   {
1078     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1079     ShenandoahMCResetCompleteBitmapTask task;
1080     heap->workers()->run_task(&task);
1081   }
1082 
1083   // Bring regions in proper states after the collection, and set heap properties.
1084   {
1085     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1086     ShenandoahPostCompactClosure post_compact;
1087     heap->heap_region_iterate(&post_compact);
1088     heap->set_used(post_compact.get_live());
1089 
1090     heap->collection_set()->clear();
1091     heap->free_set()->rebuild();
1092     heap->clear_cancelled_gc();
1093   }
1094 
1095   _preserved_marks->restore(heap->workers());
1096   _preserved_marks->reclaim();
1097 }