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, if object does not move:
 368     assert(_compact_point + obj_size <= _to_region->end(), "must fit");
 369     shenandoah_assert_not_forwarded(nullptr, p);
 370     if (_compact_point != cast_from_oop<HeapWord*>(p)) {
 371       _preserved_marks->push_if_necessary(p, p->mark());
 372       p->forward_to(cast_to_oop(_compact_point));
 373     }
 374     _compact_point += obj_size;
 375   }
 376 };
 377 
 378 class ShenandoahPrepareForCompactionTask : public WorkerTask {
 379 private:
 380   PreservedMarksSet*        const _preserved_marks;
 381   ShenandoahHeap*           const _heap;
 382   ShenandoahHeapRegionSet** const _worker_slices;
 383 
 384 public:
 385   ShenandoahPrepareForCompactionTask(PreservedMarksSet *preserved_marks, ShenandoahHeapRegionSet **worker_slices) :
 386     WorkerTask("Shenandoah Prepare For Compaction"),
 387     _preserved_marks(preserved_marks),
 388     _heap(ShenandoahHeap::heap()), _worker_slices(worker_slices) {
 389   }
 390 
 391   static bool is_candidate_region(ShenandoahHeapRegion* r) {
 392     // Empty region: get it into the slice to defragment the slice itself.
 393     // We could have skipped this without violating correctness, but we really
 394     // want to compact all live regions to the start of the heap, which sometimes
 395     // means moving them into the fully empty regions.
 396     if (r->is_empty()) return true;
 397 
 398     // Can move the region, and this is not the humongous region. Humongous
 399     // moves are special cased here, because their moves are handled separately.
 400     return r->is_stw_move_allowed() && !r->is_humongous();
 401   }
 402 
 403   void work(uint worker_id) override;
 404 private:
 405   template<typename ClosureType>
 406   void prepare_for_compaction(ClosureType& cl,
 407                               GrowableArray<ShenandoahHeapRegion*>& empty_regions,
 408                               ShenandoahHeapRegionSetIterator& it,
 409                               ShenandoahHeapRegion* from_region);
 410 };
 411 
 412 void ShenandoahPrepareForCompactionTask::work(uint worker_id) {
 413   ShenandoahParallelWorkerSession worker_session(worker_id);
 414   ShenandoahHeapRegionSet* slice = _worker_slices[worker_id];
 415   ShenandoahHeapRegionSetIterator it(slice);
 416   ShenandoahHeapRegion* from_region = it.next();
 417   // No work?
 418   if (from_region == nullptr) {
 419     return;
 420   }
 421 
 422   // Sliding compaction. Walk all regions in the slice, and compact them.
 423   // Remember empty regions and reuse them as needed.
 424   ResourceMark rm;
 425 
 426   GrowableArray<ShenandoahHeapRegion*> empty_regions((int)_heap->num_regions());
 427 
 428   ShenandoahPrepareForCompactionObjectClosure cl(_preserved_marks->get(worker_id), empty_regions, from_region);
 429   prepare_for_compaction(cl, empty_regions, it, from_region);
 430 }
 431 
 432 template<typename ClosureType>
 433 void ShenandoahPrepareForCompactionTask::prepare_for_compaction(ClosureType& cl,
 434                                                                 GrowableArray<ShenandoahHeapRegion*>& empty_regions,
 435                                                                 ShenandoahHeapRegionSetIterator& it,
 436                                                                 ShenandoahHeapRegion* from_region) {
 437   while (from_region != nullptr) {
 438     assert(is_candidate_region(from_region), "Sanity");
 439     cl.set_from_region(from_region);
 440     if (from_region->has_live()) {
 441       _heap->marked_object_iterate(from_region, &cl);
 442     }
 443 
 444     // Compacted the region to somewhere else? From-region is empty then.
 445     if (!cl.is_compact_same_region()) {
 446       empty_regions.append(from_region);
 447     }
 448     from_region = it.next();
 449   }
 450   cl.finish();
 451 
 452   // Mark all remaining regions as empty
 453   for (int pos = cl.empty_regions_pos(); pos < empty_regions.length(); ++pos) {
 454     ShenandoahHeapRegion* r = empty_regions.at(pos);
 455     r->set_new_top(r->bottom());
 456   }
 457 }
 458 
 459 void ShenandoahFullGC::calculate_target_humongous_objects() {
 460   ShenandoahHeap* heap = ShenandoahHeap::heap();
 461 
 462   // Compute the new addresses for humongous objects. We need to do this after addresses
 463   // for regular objects are calculated, and we know what regions in heap suffix are
 464   // available for humongous moves.
 465   //
 466   // Scan the heap backwards, because we are compacting humongous regions towards the end.
 467   // Maintain the contiguous compaction window in [to_begin; to_end), so that we can slide
 468   // humongous start there.
 469   //
 470   // The complication is potential non-movable regions during the scan. If such region is
 471   // detected, then sliding restarts towards that non-movable region.
 472 
 473   size_t to_begin = heap->num_regions();
 474   size_t to_end = heap->num_regions();
 475 
 476   for (size_t c = heap->num_regions(); c > 0; c--) {
 477     ShenandoahHeapRegion *r = heap->get_region(c - 1);
 478     if (r->is_humongous_continuation() || (r->new_top() == r->bottom())) {
 479       // To-region candidate: record this, and continue scan
 480       to_begin = r->index();
 481       continue;
 482     }
 483 
 484     if (r->is_humongous_start() && r->is_stw_move_allowed()) {
 485       // From-region candidate: movable humongous region
 486       oop old_obj = cast_to_oop(r->bottom());
 487       size_t words_size = old_obj->size();
 488       size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
 489 
 490       size_t start = to_end - num_regions;
 491 
 492       if (start >= to_begin && start != r->index()) {
 493         // Fits into current window, and the move is non-trivial. Record the move then, and continue scan.
 494         _preserved_marks->get(0)->push_if_necessary(old_obj, old_obj->mark());
 495         old_obj->forward_to(cast_to_oop(heap->get_region(start)->bottom()));
 496         to_end = start;
 497         continue;
 498       }
 499     }
 500 
 501     // Failed to fit. Scan starting from current region.
 502     to_begin = r->index();
 503     to_end = r->index();
 504   }
 505 }
 506 
 507 class ShenandoahEnsureHeapActiveClosure: public ShenandoahHeapRegionClosure {
 508 private:
 509   ShenandoahHeap* const _heap;
 510 
 511 public:
 512   ShenandoahEnsureHeapActiveClosure() : _heap(ShenandoahHeap::heap()) {}
 513   void heap_region_do(ShenandoahHeapRegion* r) {
 514     if (r->is_trash()) {
 515       r->recycle();
 516     }
 517     if (r->is_cset()) {
 518       r->make_regular_bypass();
 519     }
 520     if (r->is_empty_uncommitted()) {
 521       r->make_committed_bypass();
 522     }
 523     assert (r->is_committed(), "only committed regions in heap now, see region " SIZE_FORMAT, r->index());
 524 
 525     // Record current region occupancy: this communicates empty regions are free
 526     // to the rest of Full GC code.
 527     r->set_new_top(r->top());
 528   }
 529 };
 530 
 531 class ShenandoahTrashImmediateGarbageClosure: public ShenandoahHeapRegionClosure {
 532 private:
 533   ShenandoahHeap* const _heap;
 534   ShenandoahMarkingContext* const _ctx;
 535 
 536 public:
 537   ShenandoahTrashImmediateGarbageClosure() :
 538     _heap(ShenandoahHeap::heap()),
 539     _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
 540 
 541   void heap_region_do(ShenandoahHeapRegion* r) {
 542     if (r->is_humongous_start()) {
 543       oop humongous_obj = cast_to_oop(r->bottom());
 544       if (!_ctx->is_marked(humongous_obj)) {
 545         assert(!r->has_live(),
 546                "Region " SIZE_FORMAT " is not marked, should not have live", r->index());
 547         _heap->trash_humongous_region_at(r);
 548       } else {
 549         assert(r->has_live(),
 550                "Region " SIZE_FORMAT " should have live", r->index());
 551       }
 552     } else if (r->is_humongous_continuation()) {
 553       // If we hit continuation, the non-live humongous starts should have been trashed already
 554       assert(r->humongous_start_region()->has_live(),
 555              "Region " SIZE_FORMAT " should have live", r->index());
 556     } else if (r->is_regular()) {
 557       if (!r->has_live()) {
 558         r->make_trash_immediate();
 559       }
 560     }
 561   }
 562 };
 563 
 564 void ShenandoahFullGC::distribute_slices(ShenandoahHeapRegionSet** worker_slices) {
 565   ShenandoahHeap* heap = ShenandoahHeap::heap();
 566 
 567   uint n_workers = heap->workers()->active_workers();
 568   size_t n_regions = heap->num_regions();
 569 
 570   // What we want to accomplish: have the dense prefix of data, while still balancing
 571   // out the parallel work.
 572   //
 573   // Assuming the amount of work is driven by the live data that needs moving, we can slice
 574   // the entire heap into equal-live-sized prefix slices, and compact into them. So, each
 575   // thread takes all regions in its prefix subset, and then it takes some regions from
 576   // the tail.
 577   //
 578   // Tail region selection becomes interesting.
 579   //
 580   // First, we want to distribute the regions fairly between the workers, and those regions
 581   // might have different amount of live data. So, until we sure no workers need live data,
 582   // we need to only take what the worker needs.
 583   //
 584   // Second, since we slide everything to the left in each slice, the most busy regions
 585   // would be the ones on the left. Which means we want to have all workers have their after-tail
 586   // regions as close to the left as possible.
 587   //
 588   // The easiest way to do this is to distribute after-tail regions in round-robin between
 589   // workers that still need live data.
 590   //
 591   // Consider parallel workers A, B, C, then the target slice layout would be:
 592   //
 593   //  AAAAAAAABBBBBBBBCCCCCCCC|ABCABCABCABCABCABCABCABABABABABABABABABABAAAAA
 594   //
 595   //  (.....dense-prefix.....) (.....................tail...................)
 596   //  [all regions fully live] [left-most regions are fuller that right-most]
 597   //
 598 
 599   // Compute how much live data is there. This would approximate the size of dense prefix
 600   // we target to create.
 601   size_t total_live = 0;
 602   for (size_t idx = 0; idx < n_regions; idx++) {
 603     ShenandoahHeapRegion *r = heap->get_region(idx);
 604     if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
 605       total_live += r->get_live_data_words();
 606     }
 607   }
 608 
 609   // Estimate the size for the dense prefix. Note that we specifically count only the
 610   // "full" regions, so there would be some non-full regions in the slice tail.
 611   size_t live_per_worker = total_live / n_workers;
 612   size_t prefix_regions_per_worker = live_per_worker / ShenandoahHeapRegion::region_size_words();
 613   size_t prefix_regions_total = prefix_regions_per_worker * n_workers;
 614   prefix_regions_total = MIN2(prefix_regions_total, n_regions);
 615   assert(prefix_regions_total <= n_regions, "Sanity");
 616 
 617   // There might be non-candidate regions in the prefix. To compute where the tail actually
 618   // ends up being, we need to account those as well.
 619   size_t prefix_end = prefix_regions_total;
 620   for (size_t idx = 0; idx < prefix_regions_total; idx++) {
 621     ShenandoahHeapRegion *r = heap->get_region(idx);
 622     if (!ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
 623       prefix_end++;
 624     }
 625   }
 626   prefix_end = MIN2(prefix_end, n_regions);
 627   assert(prefix_end <= n_regions, "Sanity");
 628 
 629   // Distribute prefix regions per worker: each thread definitely gets its own same-sized
 630   // subset of dense prefix.
 631   size_t prefix_idx = 0;
 632 
 633   size_t* live = NEW_C_HEAP_ARRAY(size_t, n_workers, mtGC);
 634 
 635   for (size_t wid = 0; wid < n_workers; wid++) {
 636     ShenandoahHeapRegionSet* slice = worker_slices[wid];
 637 
 638     live[wid] = 0;
 639     size_t regs = 0;
 640 
 641     // Add all prefix regions for this worker
 642     while (prefix_idx < prefix_end && regs < prefix_regions_per_worker) {
 643       ShenandoahHeapRegion *r = heap->get_region(prefix_idx);
 644       if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
 645         slice->add_region(r);
 646         live[wid] += r->get_live_data_words();
 647         regs++;
 648       }
 649       prefix_idx++;
 650     }
 651   }
 652 
 653   // Distribute the tail among workers in round-robin fashion.
 654   size_t wid = n_workers - 1;
 655 
 656   for (size_t tail_idx = prefix_end; tail_idx < n_regions; tail_idx++) {
 657     ShenandoahHeapRegion *r = heap->get_region(tail_idx);
 658     if (ShenandoahPrepareForCompactionTask::is_candidate_region(r)) {
 659       assert(wid < n_workers, "Sanity");
 660 
 661       size_t live_region = r->get_live_data_words();
 662 
 663       // Select next worker that still needs live data.
 664       size_t old_wid = wid;
 665       do {
 666         wid++;
 667         if (wid == n_workers) wid = 0;
 668       } while (live[wid] + live_region >= live_per_worker && old_wid != wid);
 669 
 670       if (old_wid == wid) {
 671         // Circled back to the same worker? This means liveness data was
 672         // miscalculated. Bump the live_per_worker limit so that
 673         // everyone gets a piece of the leftover work.
 674         live_per_worker += ShenandoahHeapRegion::region_size_words();
 675       }
 676 
 677       worker_slices[wid]->add_region(r);
 678       live[wid] += live_region;
 679     }
 680   }
 681 
 682   FREE_C_HEAP_ARRAY(size_t, live);
 683 
 684 #ifdef ASSERT
 685   ResourceBitMap map(n_regions);
 686   for (size_t wid = 0; wid < n_workers; wid++) {
 687     ShenandoahHeapRegionSetIterator it(worker_slices[wid]);
 688     ShenandoahHeapRegion* r = it.next();
 689     while (r != nullptr) {
 690       size_t idx = r->index();
 691       assert(ShenandoahPrepareForCompactionTask::is_candidate_region(r), "Sanity: " SIZE_FORMAT, idx);
 692       assert(!map.at(idx), "No region distributed twice: " SIZE_FORMAT, idx);
 693       map.at_put(idx, true);
 694       r = it.next();
 695     }
 696   }
 697 
 698   for (size_t rid = 0; rid < n_regions; rid++) {
 699     bool is_candidate = ShenandoahPrepareForCompactionTask::is_candidate_region(heap->get_region(rid));
 700     bool is_distributed = map.at(rid);
 701     assert(is_distributed || !is_candidate, "All candidates are distributed: " SIZE_FORMAT, rid);
 702   }
 703 #endif
 704 }
 705 
 706 void ShenandoahFullGC::phase2_calculate_target_addresses(ShenandoahHeapRegionSet** worker_slices) {
 707   GCTraceTime(Info, gc, phases) time("Phase 2: Compute new object addresses", _gc_timer);
 708   ShenandoahGCPhase calculate_address_phase(ShenandoahPhaseTimings::full_gc_calculate_addresses);
 709 
 710   ShenandoahHeap* heap = ShenandoahHeap::heap();
 711 
 712   // About to figure out which regions can be compacted, make sure pinning status
 713   // had been updated in GC prologue.
 714   heap->assert_pinned_region_status();
 715 
 716   {
 717     // Trash the immediately collectible regions before computing addresses
 718     ShenandoahTrashImmediateGarbageClosure tigcl;
 719     heap->heap_region_iterate(&tigcl);
 720 
 721     // Make sure regions are in good state: committed, active, clean.
 722     // This is needed because we are potentially sliding the data through them.
 723     ShenandoahEnsureHeapActiveClosure ecl;
 724     heap->heap_region_iterate(&ecl);
 725   }
 726 
 727   // Compute the new addresses for regular objects
 728   {
 729     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_regular);
 730 
 731     distribute_slices(worker_slices);
 732 
 733     ShenandoahPrepareForCompactionTask task(_preserved_marks, worker_slices);
 734     heap->workers()->run_task(&task);
 735   }
 736 
 737   // Compute the new addresses for humongous objects
 738   {
 739     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_calculate_addresses_humong);
 740     calculate_target_humongous_objects();
 741   }
 742 }
 743 
 744 class ShenandoahAdjustPointersClosure : public MetadataVisitingOopIterateClosure {
 745 private:
 746   ShenandoahHeap* const _heap;
 747   ShenandoahMarkingContext* const _ctx;
 748 
 749   template <class T>
 750   inline void do_oop_work(T* p) {
 751     T o = RawAccess<>::oop_load(p);
 752     if (!CompressedOops::is_null(o)) {
 753       oop obj = CompressedOops::decode_not_null(o);
 754       assert(_ctx->is_marked(obj), "must be marked");
 755       if (obj->is_forwarded()) {
 756         oop forw = obj->forwardee();
 757         RawAccess<IS_NOT_NULL>::oop_store(p, forw);
 758       }
 759     }
 760   }
 761 
 762 public:
 763   ShenandoahAdjustPointersClosure() :
 764     _heap(ShenandoahHeap::heap()),
 765     _ctx(ShenandoahHeap::heap()->complete_marking_context()) {}
 766 
 767   void do_oop(oop* p)       { do_oop_work(p); }
 768   void do_oop(narrowOop* p) { do_oop_work(p); }
 769   void do_method(Method* m) {}
 770   void do_nmethod(nmethod* nm) {}
 771 };
 772 
 773 class ShenandoahAdjustPointersObjectClosure : public ObjectClosure {
 774 private:
 775   ShenandoahHeap* const _heap;
 776   ShenandoahAdjustPointersClosure _cl;
 777 
 778 public:
 779   ShenandoahAdjustPointersObjectClosure() :
 780     _heap(ShenandoahHeap::heap()) {
 781   }
 782   void do_object(oop p) {
 783     assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
 784     p->oop_iterate(&_cl);
 785   }
 786 };
 787 
 788 class ShenandoahAdjustPointersTask : public WorkerTask {
 789 private:
 790   ShenandoahHeap*          const _heap;
 791   ShenandoahRegionIterator       _regions;
 792 
 793 public:
 794   ShenandoahAdjustPointersTask() :
 795     WorkerTask("Shenandoah Adjust Pointers"),
 796     _heap(ShenandoahHeap::heap()) {
 797   }
 798 
 799   void work(uint worker_id) {
 800     ShenandoahParallelWorkerSession worker_session(worker_id);
 801     ShenandoahAdjustPointersObjectClosure obj_cl;
 802     ShenandoahHeapRegion* r = _regions.next();
 803     while (r != nullptr) {
 804       if (!r->is_humongous_continuation() && r->has_live()) {
 805         _heap->marked_object_iterate(r, &obj_cl);
 806       }
 807       r = _regions.next();
 808     }
 809   }
 810 };
 811 
 812 class ShenandoahAdjustRootPointersTask : public WorkerTask {
 813 private:
 814   ShenandoahRootAdjuster* _rp;
 815   PreservedMarksSet* _preserved_marks;
 816 public:
 817   ShenandoahAdjustRootPointersTask(ShenandoahRootAdjuster* rp, PreservedMarksSet* preserved_marks) :
 818     WorkerTask("Shenandoah Adjust Root Pointers"),
 819     _rp(rp),
 820     _preserved_marks(preserved_marks) {}
 821 
 822   void work(uint worker_id) {
 823     ShenandoahParallelWorkerSession worker_session(worker_id);
 824     ShenandoahAdjustPointersClosure cl;
 825     _rp->roots_do(worker_id, &cl);
 826     _preserved_marks->get(worker_id)->adjust_during_full_gc();
 827   }
 828 };
 829 
 830 void ShenandoahFullGC::phase3_update_references() {
 831   GCTraceTime(Info, gc, phases) time("Phase 3: Adjust pointers", _gc_timer);
 832   ShenandoahGCPhase adjust_pointer_phase(ShenandoahPhaseTimings::full_gc_adjust_pointers);
 833 
 834   ShenandoahHeap* heap = ShenandoahHeap::heap();
 835 
 836   WorkerThreads* workers = heap->workers();
 837   uint nworkers = workers->active_workers();
 838   {
 839 #if COMPILER2_OR_JVMCI
 840     DerivedPointerTable::clear();
 841 #endif
 842     ShenandoahRootAdjuster rp(nworkers, ShenandoahPhaseTimings::full_gc_adjust_roots);
 843     ShenandoahAdjustRootPointersTask task(&rp, _preserved_marks);
 844     workers->run_task(&task);
 845 #if COMPILER2_OR_JVMCI
 846     DerivedPointerTable::update_pointers();
 847 #endif
 848   }
 849 
 850   ShenandoahAdjustPointersTask adjust_pointers_task;
 851   workers->run_task(&adjust_pointers_task);
 852 }
 853 
 854 class ShenandoahCompactObjectsClosure : public ObjectClosure {
 855 private:
 856   ShenandoahHeap* const _heap;
 857   uint            const _worker_id;
 858 
 859 public:
 860   ShenandoahCompactObjectsClosure(uint worker_id) :
 861     _heap(ShenandoahHeap::heap()), _worker_id(worker_id) {}
 862 
 863   void do_object(oop p) {
 864     assert(_heap->complete_marking_context()->is_marked(p), "must be marked");
 865     size_t size = p->size();
 866     if (p->is_forwarded()) {
 867       HeapWord* compact_from = cast_from_oop<HeapWord*>(p);
 868       HeapWord* compact_to = cast_from_oop<HeapWord*>(p->forwardee());
 869       assert(compact_from != compact_to, "Forwarded object should move");
 870       Copy::aligned_conjoint_words(compact_from, compact_to, size);
 871       oop new_obj = cast_to_oop(compact_to);
 872 
 873       ContinuationGCSupport::relativize_stack_chunk(new_obj);
 874       new_obj->init_mark();
 875     }
 876   }
 877 };
 878 
 879 class ShenandoahCompactObjectsTask : public WorkerTask {
 880 private:
 881   ShenandoahHeap* const _heap;
 882   ShenandoahHeapRegionSet** const _worker_slices;
 883 
 884 public:
 885   ShenandoahCompactObjectsTask(ShenandoahHeapRegionSet** worker_slices) :
 886     WorkerTask("Shenandoah Compact Objects"),
 887     _heap(ShenandoahHeap::heap()),
 888     _worker_slices(worker_slices) {
 889   }
 890 
 891   void work(uint worker_id) {
 892     ShenandoahParallelWorkerSession worker_session(worker_id);
 893     ShenandoahHeapRegionSetIterator slice(_worker_slices[worker_id]);
 894 
 895     ShenandoahCompactObjectsClosure cl(worker_id);
 896     ShenandoahHeapRegion* r = slice.next();
 897     while (r != nullptr) {
 898       assert(!r->is_humongous(), "must not get humongous regions here");
 899       if (r->has_live()) {
 900         _heap->marked_object_iterate(r, &cl);
 901       }
 902       r->set_top(r->new_top());
 903       r = slice.next();
 904     }
 905   }
 906 };
 907 
 908 class ShenandoahPostCompactClosure : public ShenandoahHeapRegionClosure {
 909 private:
 910   ShenandoahHeap* const _heap;
 911   size_t _live;
 912 
 913 public:
 914   ShenandoahPostCompactClosure() : _heap(ShenandoahHeap::heap()), _live(0) {
 915   }
 916 
 917   void heap_region_do(ShenandoahHeapRegion* r) {
 918     assert (!r->is_cset(), "cset regions should have been demoted already");
 919 
 920     // Need to reset the complete-top-at-mark-start pointer here because
 921     // the complete marking bitmap is no longer valid. This ensures
 922     // size-based iteration in marked_object_iterate().
 923     // NOTE: See blurb at ShenandoahMCResetCompleteBitmapTask on why we need to skip
 924     // pinned regions.
 925     if (!r->is_pinned()) {
 926       _heap->complete_marking_context()->reset_top_at_mark_start(r);
 927     }
 928 
 929     size_t live = r->used();
 930 
 931     // Make empty regions that have been allocated into regular
 932     if (r->is_empty() && live > 0) {
 933       r->make_regular_bypass();
 934       if (ZapUnusedHeapArea) {
 935         SpaceMangler::mangle_region(MemRegion(r->top(), r->end()));
 936       }
 937     }
 938 
 939     // Reclaim regular regions that became empty
 940     if (r->is_regular() && live == 0) {
 941       r->make_trash();
 942     }
 943 
 944     // Recycle all trash regions
 945     if (r->is_trash()) {
 946       live = 0;
 947       r->recycle();
 948     }
 949 
 950     r->set_live_data(live);
 951     r->reset_alloc_metadata();
 952     _live += live;
 953   }
 954 
 955   size_t get_live() {
 956     return _live;
 957   }
 958 };
 959 
 960 void ShenandoahFullGC::compact_humongous_objects() {
 961   // Compact humongous regions, based on their fwdptr objects.
 962   //
 963   // This code is serial, because doing the in-slice parallel sliding is tricky. In most cases,
 964   // humongous regions are already compacted, and do not require further moves, which alleviates
 965   // sliding costs. We may consider doing this in parallel in the future.
 966 
 967   ShenandoahHeap* heap = ShenandoahHeap::heap();
 968 
 969   for (size_t c = heap->num_regions(); c > 0; c--) {
 970     ShenandoahHeapRegion* r = heap->get_region(c - 1);
 971     if (r->is_humongous_start()) {
 972       oop old_obj = cast_to_oop(r->bottom());
 973       if (!old_obj->is_forwarded()) {
 974         // No need to move the object, it stays at the same slot
 975         continue;
 976       }
 977       size_t words_size = old_obj->size();
 978       size_t num_regions = ShenandoahHeapRegion::required_regions(words_size * HeapWordSize);
 979 
 980       size_t old_start = r->index();
 981       size_t old_end   = old_start + num_regions - 1;
 982       size_t new_start = heap->heap_region_index_containing(old_obj->forwardee());
 983       size_t new_end   = new_start + num_regions - 1;
 984       assert(old_start != new_start, "must be real move");
 985       assert(r->is_stw_move_allowed(), "Region " SIZE_FORMAT " should be movable", r->index());
 986 
 987       Copy::aligned_conjoint_words(r->bottom(), heap->get_region(new_start)->bottom(), words_size);
 988       ContinuationGCSupport::relativize_stack_chunk(cast_to_oop<HeapWord*>(r->bottom()));
 989 
 990       oop new_obj = cast_to_oop(heap->get_region(new_start)->bottom());
 991       new_obj->init_mark();
 992 
 993       {
 994         for (size_t c = old_start; c <= old_end; c++) {
 995           ShenandoahHeapRegion* r = heap->get_region(c);
 996           r->make_regular_bypass();
 997           r->set_top(r->bottom());
 998         }
 999 
1000         for (size_t c = new_start; c <= new_end; c++) {
1001           ShenandoahHeapRegion* r = heap->get_region(c);
1002           if (c == new_start) {
1003             r->make_humongous_start_bypass();
1004           } else {
1005             r->make_humongous_cont_bypass();
1006           }
1007 
1008           // Trailing region may be non-full, record the remainder there
1009           size_t remainder = words_size & ShenandoahHeapRegion::region_size_words_mask();
1010           if ((c == new_end) && (remainder != 0)) {
1011             r->set_top(r->bottom() + remainder);
1012           } else {
1013             r->set_top(r->end());
1014           }
1015 
1016           r->reset_alloc_metadata();
1017         }
1018       }
1019     }
1020   }
1021 }
1022 
1023 // This is slightly different to ShHeap::reset_next_mark_bitmap:
1024 // we need to remain able to walk pinned regions.
1025 // Since pinned region do not move and don't get compacted, we will get holes with
1026 // unreachable objects in them (which may have pointers to unloaded Klasses and thus
1027 // cannot be iterated over using oop->size(). The only way to safely iterate over those is using
1028 // a valid marking bitmap and valid TAMS pointer. This class only resets marking
1029 // bitmaps for un-pinned regions, and later we only reset TAMS for unpinned regions.
1030 class ShenandoahMCResetCompleteBitmapTask : public WorkerTask {
1031 private:
1032   ShenandoahRegionIterator _regions;
1033 
1034 public:
1035   ShenandoahMCResetCompleteBitmapTask() :
1036     WorkerTask("Shenandoah Reset Bitmap") {
1037   }
1038 
1039   void work(uint worker_id) {
1040     ShenandoahParallelWorkerSession worker_session(worker_id);
1041     ShenandoahHeapRegion* region = _regions.next();
1042     ShenandoahHeap* heap = ShenandoahHeap::heap();
1043     ShenandoahMarkingContext* const ctx = heap->complete_marking_context();
1044     while (region != nullptr) {
1045       if (heap->is_bitmap_slice_committed(region) && !region->is_pinned() && region->has_live()) {
1046         ctx->clear_bitmap(region);
1047       }
1048       region = _regions.next();
1049     }
1050   }
1051 };
1052 
1053 void ShenandoahFullGC::phase4_compact_objects(ShenandoahHeapRegionSet** worker_slices) {
1054   GCTraceTime(Info, gc, phases) time("Phase 4: Move objects", _gc_timer);
1055   ShenandoahGCPhase compaction_phase(ShenandoahPhaseTimings::full_gc_copy_objects);
1056 
1057   ShenandoahHeap* heap = ShenandoahHeap::heap();
1058 
1059   // Compact regular objects first
1060   {
1061     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_regular);
1062     ShenandoahCompactObjectsTask compact_task(worker_slices);
1063     heap->workers()->run_task(&compact_task);
1064   }
1065 
1066   // Compact humongous objects after regular object moves
1067   {
1068     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_humong);
1069     compact_humongous_objects();
1070   }
1071 }
1072 
1073 void ShenandoahFullGC::phase5_epilog() {
1074   GCTraceTime(Info, gc, phases) time("Phase 5: Full GC epilog", _gc_timer);
1075   ShenandoahHeap* heap = ShenandoahHeap::heap();
1076 
1077   // Reset complete bitmap. We're about to reset the complete-top-at-mark-start pointer
1078   // and must ensure the bitmap is in sync.
1079   {
1080     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_reset_complete);
1081     ShenandoahMCResetCompleteBitmapTask task;
1082     heap->workers()->run_task(&task);
1083   }
1084 
1085   // Bring regions in proper states after the collection, and set heap properties.
1086   {
1087     ShenandoahGCPhase phase(ShenandoahPhaseTimings::full_gc_copy_objects_rebuild);
1088     ShenandoahPostCompactClosure post_compact;
1089     heap->heap_region_iterate(&post_compact);
1090     heap->set_used(post_compact.get_live());
1091 
1092     heap->collection_set()->clear();
1093     heap->free_set()->rebuild();
1094     heap->clear_cancelled_gc();
1095   }
1096 
1097   _preserved_marks->restore(heap->workers());
1098   _preserved_marks->reclaim();
1099 }