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