mark_sweep.cc revision 8016bdee7ca1a066221a5d2fe5e60890de950a5b 
1/*
2 * Copyright (C) 2011 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 *      http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17#include "mark_sweep.h"
18
19#include <atomic>
20#include <functional>
21#include <numeric>
22#include <climits>
23#include <vector>
24
25#include "base/bounded_fifo.h"
26#include "base/logging.h"
27#include "base/macros.h"
28#include "base/mutex-inl.h"
29#include "base/systrace.h"
30#include "base/time_utils.h"
31#include "base/timing_logger.h"
32#include "gc/accounting/card_table-inl.h"
33#include "gc/accounting/heap_bitmap-inl.h"
34#include "gc/accounting/mod_union_table.h"
35#include "gc/accounting/space_bitmap-inl.h"
36#include "gc/heap.h"
37#include "gc/reference_processor.h"
38#include "gc/space/large_object_space.h"
39#include "gc/space/space-inl.h"
40#include "mark_sweep-inl.h"
41#include "mirror/object-inl.h"
42#include "runtime.h"
43#include "scoped_thread_state_change.h"
44#include "thread-inl.h"
45#include "thread_list.h"
46
47namespace art {
48namespace gc {
49namespace collector {
50
51// Performance options.
52static constexpr bool kUseRecursiveMark = false;
53static constexpr bool kUseMarkStackPrefetch = true;
54static constexpr size_t kSweepArrayChunkFreeSize = 1024;
55static constexpr bool kPreCleanCards = true;
56
57// Parallelism options.
58static constexpr bool kParallelCardScan = true;
59static constexpr bool kParallelRecursiveMark = true;
60// Don't attempt to parallelize mark stack processing unless the mark stack is at least n
61// elements. This is temporary until we reduce the overhead caused by allocating tasks, etc.. Not
62// having this can add overhead in ProcessReferences since we may end up doing many calls of
63// ProcessMarkStack with very small mark stacks.
64static constexpr size_t kMinimumParallelMarkStackSize = 128;
65static constexpr bool kParallelProcessMarkStack = true;
66
67// Profiling and information flags.
68static constexpr bool kProfileLargeObjects = false;
69static constexpr bool kMeasureOverhead = false;
70static constexpr bool kCountTasks = false;
71static constexpr bool kCountMarkedObjects = false;
72
73// Turn off kCheckLocks when profiling the GC since it slows the GC down by up to 40%.
74static constexpr bool kCheckLocks = kDebugLocking;
75static constexpr bool kVerifyRootsMarked = kIsDebugBuild;
76
77// If true, revoke the rosalloc thread-local buffers at the
78// checkpoint, as opposed to during the pause.
79static constexpr bool kRevokeRosAllocThreadLocalBuffersAtCheckpoint = true;
80
81void MarkSweep::BindBitmaps() {
82  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
83  WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_);
84  // Mark all of the spaces we never collect as immune.
85  for (const auto& space : GetHeap()->GetContinuousSpaces()) {
86    if (space->GetGcRetentionPolicy() == space::kGcRetentionPolicyNeverCollect) {
87      immune_spaces_.AddSpace(space);
88    }
89  }
90}
91
92MarkSweep::MarkSweep(Heap* heap, bool is_concurrent, const std::string& name_prefix)
93    : GarbageCollector(heap,
94                       name_prefix +
95                       (is_concurrent ? "concurrent mark sweep": "mark sweep")),
96      current_space_bitmap_(nullptr),
97      mark_bitmap_(nullptr),
98      mark_stack_(nullptr),
99      gc_barrier_(new Barrier(0)),
100      mark_stack_lock_("mark sweep mark stack lock", kMarkSweepMarkStackLock),
101      is_concurrent_(is_concurrent),
102      live_stack_freeze_size_(0) {
103  std::string error_msg;
104  MemMap* mem_map = MemMap::MapAnonymous(
105      "mark sweep sweep array free buffer", nullptr,
106      RoundUp(kSweepArrayChunkFreeSize * sizeof(mirror::Object*), kPageSize),
107      PROT_READ | PROT_WRITE, false, false, &error_msg);
108  CHECK(mem_map != nullptr) << "Couldn't allocate sweep array free buffer: " << error_msg;
109  sweep_array_free_buffer_mem_map_.reset(mem_map);
110}
111
112void MarkSweep::InitializePhase() {
113  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
114  mark_stack_ = heap_->GetMarkStack();
115  DCHECK(mark_stack_ != nullptr);
116  immune_spaces_.Reset();
117  no_reference_class_count_.StoreRelaxed(0);
118  normal_count_.StoreRelaxed(0);
119  class_count_.StoreRelaxed(0);
120  object_array_count_.StoreRelaxed(0);
121  other_count_.StoreRelaxed(0);
122  reference_count_.StoreRelaxed(0);
123  large_object_test_.StoreRelaxed(0);
124  large_object_mark_.StoreRelaxed(0);
125  overhead_time_ .StoreRelaxed(0);
126  work_chunks_created_.StoreRelaxed(0);
127  work_chunks_deleted_.StoreRelaxed(0);
128  mark_null_count_.StoreRelaxed(0);
129  mark_immune_count_.StoreRelaxed(0);
130  mark_fastpath_count_.StoreRelaxed(0);
131  mark_slowpath_count_.StoreRelaxed(0);
132  {
133    // TODO: I don't think we should need heap bitmap lock to Get the mark bitmap.
134    ReaderMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_);
135    mark_bitmap_ = heap_->GetMarkBitmap();
136  }
137  if (!GetCurrentIteration()->GetClearSoftReferences()) {
138    // Always clear soft references if a non-sticky collection.
139    GetCurrentIteration()->SetClearSoftReferences(GetGcType() != collector::kGcTypeSticky);
140  }
141}
142
143void MarkSweep::RunPhases() {
144  Thread* self = Thread::Current();
145  InitializePhase();
146  Locks::mutator_lock_->AssertNotHeld(self);
147  if (IsConcurrent()) {
148    GetHeap()->PreGcVerification(this);
149    {
150      ReaderMutexLock mu(self, *Locks::mutator_lock_);
151      MarkingPhase();
152    }
153    ScopedPause pause(this);
154    GetHeap()->PrePauseRosAllocVerification(this);
155    PausePhase();
156    RevokeAllThreadLocalBuffers();
157  } else {
158    ScopedPause pause(this);
159    GetHeap()->PreGcVerificationPaused(this);
160    MarkingPhase();
161    GetHeap()->PrePauseRosAllocVerification(this);
162    PausePhase();
163    RevokeAllThreadLocalBuffers();
164  }
165  {
166    // Sweeping always done concurrently, even for non concurrent mark sweep.
167    ReaderMutexLock mu(self, *Locks::mutator_lock_);
168    ReclaimPhase();
169  }
170  GetHeap()->PostGcVerification(this);
171  FinishPhase();
172}
173
174void MarkSweep::ProcessReferences(Thread* self) {
175  WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
176  GetHeap()->GetReferenceProcessor()->ProcessReferences(
177      true,
178      GetTimings(),
179      GetCurrentIteration()->GetClearSoftReferences(),
180      this);
181}
182
183void MarkSweep::PausePhase() {
184  TimingLogger::ScopedTiming t("(Paused)PausePhase", GetTimings());
185  Thread* self = Thread::Current();
186  Locks::mutator_lock_->AssertExclusiveHeld(self);
187  if (IsConcurrent()) {
188    // Handle the dirty objects if we are a concurrent GC.
189    WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
190    // Re-mark root set.
191    ReMarkRoots();
192    // Scan dirty objects, this is only required if we are not doing concurrent GC.
193    RecursiveMarkDirtyObjects(true, accounting::CardTable::kCardDirty);
194  }
195  {
196    TimingLogger::ScopedTiming t2("SwapStacks", GetTimings());
197    WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
198    heap_->SwapStacks();
199    live_stack_freeze_size_ = heap_->GetLiveStack()->Size();
200    // Need to revoke all the thread local allocation stacks since we just swapped the allocation
201    // stacks and don't want anybody to allocate into the live stack.
202    RevokeAllThreadLocalAllocationStacks(self);
203  }
204  heap_->PreSweepingGcVerification(this);
205  // Disallow new system weaks to prevent a race which occurs when someone adds a new system
206  // weak before we sweep them. Since this new system weak may not be marked, the GC may
207  // incorrectly sweep it. This also fixes a race where interning may attempt to return a strong
208  // reference to a string that is about to be swept.
209  Runtime::Current()->DisallowNewSystemWeaks();
210  // Enable the reference processing slow path, needs to be done with mutators paused since there
211  // is no lock in the GetReferent fast path.
212  GetHeap()->GetReferenceProcessor()->EnableSlowPath();
213}
214
215void MarkSweep::PreCleanCards() {
216  // Don't do this for non concurrent GCs since they don't have any dirty cards.
217  if (kPreCleanCards && IsConcurrent()) {
218    TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
219    Thread* self = Thread::Current();
220    CHECK(!Locks::mutator_lock_->IsExclusiveHeld(self));
221    // Process dirty cards and add dirty cards to mod union tables, also ages cards.
222    heap_->ProcessCards(GetTimings(), false, true, false);
223    // The checkpoint root marking is required to avoid a race condition which occurs if the
224    // following happens during a reference write:
225    // 1. mutator dirties the card (write barrier)
226    // 2. GC ages the card (the above ProcessCards call)
227    // 3. GC scans the object (the RecursiveMarkDirtyObjects call below)
228    // 4. mutator writes the value (corresponding to the write barrier in 1.)
229    // This causes the GC to age the card but not necessarily mark the reference which the mutator
230    // wrote into the object stored in the card.
231    // Having the checkpoint fixes this issue since it ensures that the card mark and the
232    // reference write are visible to the GC before the card is scanned (this is due to locks being
233    // acquired / released in the checkpoint code).
234    // The other roots are also marked to help reduce the pause.
235    MarkRootsCheckpoint(self, false);
236    MarkNonThreadRoots();
237    MarkConcurrentRoots(
238        static_cast<VisitRootFlags>(kVisitRootFlagClearRootLog | kVisitRootFlagNewRoots));
239    // Process the newly aged cards.
240    RecursiveMarkDirtyObjects(false, accounting::CardTable::kCardDirty - 1);
241    // TODO: Empty allocation stack to reduce the number of objects we need to test / mark as live
242    // in the next GC.
243  }
244}
245
246void MarkSweep::RevokeAllThreadLocalAllocationStacks(Thread* self) {
247  if (kUseThreadLocalAllocationStack) {
248    TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
249    Locks::mutator_lock_->AssertExclusiveHeld(self);
250    heap_->RevokeAllThreadLocalAllocationStacks(self);
251  }
252}
253
254void MarkSweep::MarkingPhase() {
255  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
256  Thread* self = Thread::Current();
257  BindBitmaps();
258  FindDefaultSpaceBitmap();
259  // Process dirty cards and add dirty cards to mod union tables.
260  // If the GC type is non sticky, then we just clear the cards instead of ageing them.
261  heap_->ProcessCards(GetTimings(), false, true, GetGcType() != kGcTypeSticky);
262  WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
263  MarkRoots(self);
264  MarkReachableObjects();
265  // Pre-clean dirtied cards to reduce pauses.
266  PreCleanCards();
267}
268
269class MarkSweep::ScanObjectVisitor {
270 public:
271  explicit ScanObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE
272      : mark_sweep_(mark_sweep) {}
273
274  void operator()(mirror::Object* obj) const
275      ALWAYS_INLINE
276      REQUIRES(Locks::heap_bitmap_lock_)
277      SHARED_REQUIRES(Locks::mutator_lock_) {
278    if (kCheckLocks) {
279      Locks::mutator_lock_->AssertSharedHeld(Thread::Current());
280      Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current());
281    }
282    mark_sweep_->ScanObject(obj);
283  }
284
285 private:
286  MarkSweep* const mark_sweep_;
287};
288
289void MarkSweep::UpdateAndMarkModUnion() {
290  for (const auto& space : immune_spaces_.GetSpaces()) {
291    const char* name = space->IsZygoteSpace()
292        ? "UpdateAndMarkZygoteModUnionTable"
293        : "UpdateAndMarkImageModUnionTable";
294    DCHECK(space->IsZygoteSpace() || space->IsImageSpace()) << *space;
295    TimingLogger::ScopedTiming t(name, GetTimings());
296    accounting::ModUnionTable* mod_union_table = heap_->FindModUnionTableFromSpace(space);
297    if (mod_union_table != nullptr) {
298      mod_union_table->UpdateAndMarkReferences(this);
299    } else {
300      // No mod-union table, scan all the live bits. This can only occur for app images.
301      space->GetLiveBitmap()->VisitMarkedRange(reinterpret_cast<uintptr_t>(space->Begin()),
302                                               reinterpret_cast<uintptr_t>(space->End()),
303                                               ScanObjectVisitor(this));
304    }
305  }
306}
307
308void MarkSweep::MarkReachableObjects() {
309  UpdateAndMarkModUnion();
310  // Recursively mark all the non-image bits set in the mark bitmap.
311  RecursiveMark();
312}
313
314void MarkSweep::ReclaimPhase() {
315  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
316  Thread* const self = Thread::Current();
317  // Process the references concurrently.
318  ProcessReferences(self);
319  SweepSystemWeaks(self);
320  Runtime* const runtime = Runtime::Current();
321  runtime->AllowNewSystemWeaks();
322  // Clean up class loaders after system weaks are swept since that is how we know if class
323  // unloading occurred.
324  runtime->GetClassLinker()->CleanupClassLoaders();
325  {
326    WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
327    GetHeap()->RecordFreeRevoke();
328    // Reclaim unmarked objects.
329    Sweep(false);
330    // Swap the live and mark bitmaps for each space which we modified space. This is an
331    // optimization that enables us to not clear live bits inside of the sweep. Only swaps unbound
332    // bitmaps.
333    SwapBitmaps();
334    // Unbind the live and mark bitmaps.
335    GetHeap()->UnBindBitmaps();
336  }
337}
338
339void MarkSweep::FindDefaultSpaceBitmap() {
340  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
341  for (const auto& space : GetHeap()->GetContinuousSpaces()) {
342    accounting::ContinuousSpaceBitmap* bitmap = space->GetMarkBitmap();
343    // We want to have the main space instead of non moving if possible.
344    if (bitmap != nullptr &&
345        space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) {
346      current_space_bitmap_ = bitmap;
347      // If we are not the non moving space exit the loop early since this will be good enough.
348      if (space != heap_->GetNonMovingSpace()) {
349        break;
350      }
351    }
352  }
353  CHECK(current_space_bitmap_ != nullptr) << "Could not find a default mark bitmap\n"
354      << heap_->DumpSpaces();
355}
356
357void MarkSweep::ExpandMarkStack() {
358  ResizeMarkStack(mark_stack_->Capacity() * 2);
359}
360
361void MarkSweep::ResizeMarkStack(size_t new_size) {
362  // Rare case, no need to have Thread::Current be a parameter.
363  if (UNLIKELY(mark_stack_->Size() < mark_stack_->Capacity())) {
364    // Someone else acquired the lock and expanded the mark stack before us.
365    return;
366  }
367  std::vector<StackReference<mirror::Object>> temp(mark_stack_->Begin(), mark_stack_->End());
368  CHECK_LE(mark_stack_->Size(), new_size);
369  mark_stack_->Resize(new_size);
370  for (auto& obj : temp) {
371    mark_stack_->PushBack(obj.AsMirrorPtr());
372  }
373}
374
375mirror::Object* MarkSweep::MarkObject(mirror::Object* obj) {
376  MarkObject(obj, nullptr, MemberOffset(0));
377  return obj;
378}
379
380inline void MarkSweep::MarkObjectNonNullParallel(mirror::Object* obj) {
381  DCHECK(obj != nullptr);
382  if (MarkObjectParallel(obj)) {
383    MutexLock mu(Thread::Current(), mark_stack_lock_);
384    if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) {
385      ExpandMarkStack();
386    }
387    // The object must be pushed on to the mark stack.
388    mark_stack_->PushBack(obj);
389  }
390}
391
392bool MarkSweep::IsMarkedHeapReference(mirror::HeapReference<mirror::Object>* ref) {
393  return IsMarked(ref->AsMirrorPtr());
394}
395
396class MarkSweep::MarkObjectSlowPath {
397 public:
398  explicit MarkObjectSlowPath(MarkSweep* mark_sweep,
399                              mirror::Object* holder = nullptr,
400                              MemberOffset offset = MemberOffset(0))
401      : mark_sweep_(mark_sweep),
402        holder_(holder),
403        offset_(offset) {}
404
405  void operator()(const mirror::Object* obj) const NO_THREAD_SAFETY_ANALYSIS {
406    if (kProfileLargeObjects) {
407      // TODO: Differentiate between marking and testing somehow.
408      ++mark_sweep_->large_object_test_;
409      ++mark_sweep_->large_object_mark_;
410    }
411    space::LargeObjectSpace* large_object_space = mark_sweep_->GetHeap()->GetLargeObjectsSpace();
412    if (UNLIKELY(obj == nullptr || !IsAligned<kPageSize>(obj) ||
413                 (kIsDebugBuild && large_object_space != nullptr &&
414                     !large_object_space->Contains(obj)))) {
415      LOG(INTERNAL_FATAL) << "Tried to mark " << obj << " not contained by any spaces";
416      if (holder_ != nullptr) {
417        size_t holder_size = holder_->SizeOf();
418        ArtField* field = holder_->FindFieldByOffset(offset_);
419        LOG(INTERNAL_FATAL) << "Field info: "
420                            << " holder=" << holder_
421                            << " holder is "
422                            << (mark_sweep_->GetHeap()->IsLiveObjectLocked(holder_)
423                                ? "alive" : "dead")
424                            << " holder_size=" << holder_size
425                            << " holder_type=" << PrettyTypeOf(holder_)
426                            << " offset=" << offset_.Uint32Value()
427                            << " field=" << (field != nullptr ? field->GetName() : "nullptr")
428                            << " field_type="
429                            << (field != nullptr ? field->GetTypeDescriptor() : "")
430                            << " first_ref_field_offset="
431                            << (holder_->IsClass()
432                                ? holder_->AsClass()->GetFirstReferenceStaticFieldOffset(
433                                    sizeof(void*))
434                                : holder_->GetClass()->GetFirstReferenceInstanceFieldOffset())
435                            << " num_of_ref_fields="
436                            << (holder_->IsClass()
437                                ? holder_->AsClass()->NumReferenceStaticFields()
438                                : holder_->GetClass()->NumReferenceInstanceFields())
439                            << "\n";
440        // Print the memory content of the holder.
441        for (size_t i = 0; i < holder_size / sizeof(uint32_t); ++i) {
442          uint32_t* p = reinterpret_cast<uint32_t*>(holder_);
443          LOG(INTERNAL_FATAL) << &p[i] << ": " << "holder+" << (i * sizeof(uint32_t)) << " = "
444                              << std::hex << p[i];
445        }
446      }
447      PrintFileToLog("/proc/self/maps", LogSeverity::INTERNAL_FATAL);
448      MemMap::DumpMaps(LOG(INTERNAL_FATAL), true);
449      {
450        LOG(INTERNAL_FATAL) << "Attempting see if it's a bad root";
451        Thread* self = Thread::Current();
452        if (Locks::mutator_lock_->IsExclusiveHeld(self)) {
453          mark_sweep_->VerifyRoots();
454        } else {
455          const bool heap_bitmap_exclusive_locked =
456              Locks::heap_bitmap_lock_->IsExclusiveHeld(self);
457          if (heap_bitmap_exclusive_locked) {
458            Locks::heap_bitmap_lock_->ExclusiveUnlock(self);
459          }
460          {
461            ScopedThreadSuspension(self, kSuspended);
462            ScopedSuspendAll ssa(__FUNCTION__);
463            mark_sweep_->VerifyRoots();
464          }
465          if (heap_bitmap_exclusive_locked) {
466            Locks::heap_bitmap_lock_->ExclusiveLock(self);
467          }
468        }
469      }
470      LOG(FATAL) << "Can't mark invalid object";
471    }
472  }
473
474 private:
475  MarkSweep* const mark_sweep_;
476  mirror::Object* const holder_;
477  MemberOffset offset_;
478};
479
480inline void MarkSweep::MarkObjectNonNull(mirror::Object* obj,
481                                         mirror::Object* holder,
482                                         MemberOffset offset) {
483  DCHECK(obj != nullptr);
484  if (kUseBakerOrBrooksReadBarrier) {
485    // Verify all the objects have the correct pointer installed.
486    obj->AssertReadBarrierPointer();
487  }
488  if (immune_spaces_.IsInImmuneRegion(obj)) {
489    if (kCountMarkedObjects) {
490      ++mark_immune_count_;
491    }
492    DCHECK(mark_bitmap_->Test(obj));
493  } else if (LIKELY(current_space_bitmap_->HasAddress(obj))) {
494    if (kCountMarkedObjects) {
495      ++mark_fastpath_count_;
496    }
497    if (UNLIKELY(!current_space_bitmap_->Set(obj))) {
498      PushOnMarkStack(obj);  // This object was not previously marked.
499    }
500  } else {
501    if (kCountMarkedObjects) {
502      ++mark_slowpath_count_;
503    }
504    MarkObjectSlowPath visitor(this, holder, offset);
505    // TODO: We already know that the object is not in the current_space_bitmap_ but MarkBitmap::Set
506    // will check again.
507    if (!mark_bitmap_->Set(obj, visitor)) {
508      PushOnMarkStack(obj);  // Was not already marked, push.
509    }
510  }
511}
512
513inline void MarkSweep::PushOnMarkStack(mirror::Object* obj) {
514  if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) {
515    // Lock is not needed but is here anyways to please annotalysis.
516    MutexLock mu(Thread::Current(), mark_stack_lock_);
517    ExpandMarkStack();
518  }
519  // The object must be pushed on to the mark stack.
520  mark_stack_->PushBack(obj);
521}
522
523inline bool MarkSweep::MarkObjectParallel(mirror::Object* obj) {
524  DCHECK(obj != nullptr);
525  if (kUseBakerOrBrooksReadBarrier) {
526    // Verify all the objects have the correct pointer installed.
527    obj->AssertReadBarrierPointer();
528  }
529  if (immune_spaces_.IsInImmuneRegion(obj)) {
530    DCHECK(IsMarked(obj) != nullptr);
531    return false;
532  }
533  // Try to take advantage of locality of references within a space, failing this find the space
534  // the hard way.
535  accounting::ContinuousSpaceBitmap* object_bitmap = current_space_bitmap_;
536  if (LIKELY(object_bitmap->HasAddress(obj))) {
537    return !object_bitmap->AtomicTestAndSet(obj);
538  }
539  MarkObjectSlowPath visitor(this);
540  return !mark_bitmap_->AtomicTestAndSet(obj, visitor);
541}
542
543void MarkSweep::MarkHeapReference(mirror::HeapReference<mirror::Object>* ref) {
544  MarkObject(ref->AsMirrorPtr(), nullptr, MemberOffset(0));
545}
546
547// Used to mark objects when processing the mark stack. If an object is null, it is not marked.
548inline void MarkSweep::MarkObject(mirror::Object* obj,
549                                  mirror::Object* holder,
550                                  MemberOffset offset) {
551  if (obj != nullptr) {
552    MarkObjectNonNull(obj, holder, offset);
553  } else if (kCountMarkedObjects) {
554    ++mark_null_count_;
555  }
556}
557
558class MarkSweep::VerifyRootMarkedVisitor : public SingleRootVisitor {
559 public:
560  explicit VerifyRootMarkedVisitor(MarkSweep* collector) : collector_(collector) { }
561
562  void VisitRoot(mirror::Object* root, const RootInfo& info) OVERRIDE
563      SHARED_REQUIRES(Locks::mutator_lock_, Locks::heap_bitmap_lock_) {
564    CHECK(collector_->IsMarked(root) != nullptr) << info.ToString();
565  }
566
567 private:
568  MarkSweep* const collector_;
569};
570
571void MarkSweep::VisitRoots(mirror::Object*** roots,
572                           size_t count,
573                           const RootInfo& info ATTRIBUTE_UNUSED) {
574  for (size_t i = 0; i < count; ++i) {
575    MarkObjectNonNull(*roots[i]);
576  }
577}
578
579void MarkSweep::VisitRoots(mirror::CompressedReference<mirror::Object>** roots,
580                           size_t count,
581                           const RootInfo& info ATTRIBUTE_UNUSED) {
582  for (size_t i = 0; i < count; ++i) {
583    MarkObjectNonNull(roots[i]->AsMirrorPtr());
584  }
585}
586
587class MarkSweep::VerifyRootVisitor : public SingleRootVisitor {
588 public:
589  void VisitRoot(mirror::Object* root, const RootInfo& info) OVERRIDE
590      SHARED_REQUIRES(Locks::mutator_lock_, Locks::heap_bitmap_lock_) {
591    // See if the root is on any space bitmap.
592    auto* heap = Runtime::Current()->GetHeap();
593    if (heap->GetLiveBitmap()->GetContinuousSpaceBitmap(root) == nullptr) {
594      space::LargeObjectSpace* large_object_space = heap->GetLargeObjectsSpace();
595      if (large_object_space != nullptr && !large_object_space->Contains(root)) {
596        LOG(INTERNAL_FATAL) << "Found invalid root: " << root << " " << info;
597      }
598    }
599  }
600};
601
602void MarkSweep::VerifyRoots() {
603  VerifyRootVisitor visitor;
604  Runtime::Current()->GetThreadList()->VisitRoots(&visitor);
605}
606
607void MarkSweep::MarkRoots(Thread* self) {
608  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
609  if (Locks::mutator_lock_->IsExclusiveHeld(self)) {
610    // If we exclusively hold the mutator lock, all threads must be suspended.
611    Runtime::Current()->VisitRoots(this);
612    RevokeAllThreadLocalAllocationStacks(self);
613  } else {
614    MarkRootsCheckpoint(self, kRevokeRosAllocThreadLocalBuffersAtCheckpoint);
615    // At this point the live stack should no longer have any mutators which push into it.
616    MarkNonThreadRoots();
617    MarkConcurrentRoots(
618        static_cast<VisitRootFlags>(kVisitRootFlagAllRoots | kVisitRootFlagStartLoggingNewRoots));
619  }
620}
621
622void MarkSweep::MarkNonThreadRoots() {
623  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
624  Runtime::Current()->VisitNonThreadRoots(this);
625}
626
627void MarkSweep::MarkConcurrentRoots(VisitRootFlags flags) {
628  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
629  // Visit all runtime roots and clear dirty flags.
630  Runtime::Current()->VisitConcurrentRoots(
631      this, static_cast<VisitRootFlags>(flags | kVisitRootFlagNonMoving));
632}
633
634class MarkSweep::DelayReferenceReferentVisitor {
635 public:
636  explicit DelayReferenceReferentVisitor(MarkSweep* collector) : collector_(collector) {}
637
638  void operator()(mirror::Class* klass, mirror::Reference* ref) const
639      REQUIRES(Locks::heap_bitmap_lock_)
640      SHARED_REQUIRES(Locks::mutator_lock_) {
641    collector_->DelayReferenceReferent(klass, ref);
642  }
643
644 private:
645  MarkSweep* const collector_;
646};
647
648template <bool kUseFinger = false>
649class MarkSweep::MarkStackTask : public Task {
650 public:
651  MarkStackTask(ThreadPool* thread_pool,
652                MarkSweep* mark_sweep,
653                size_t mark_stack_size,
654                StackReference<mirror::Object>* mark_stack)
655      : mark_sweep_(mark_sweep),
656        thread_pool_(thread_pool),
657        mark_stack_pos_(mark_stack_size) {
658    // We may have to copy part of an existing mark stack when another mark stack overflows.
659    if (mark_stack_size != 0) {
660      DCHECK(mark_stack != nullptr);
661      // TODO: Check performance?
662      std::copy(mark_stack, mark_stack + mark_stack_size, mark_stack_);
663    }
664    if (kCountTasks) {
665      ++mark_sweep_->work_chunks_created_;
666    }
667  }
668
669  static const size_t kMaxSize = 1 * KB;
670
671 protected:
672  class MarkObjectParallelVisitor {
673   public:
674    ALWAYS_INLINE MarkObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task,
675                                            MarkSweep* mark_sweep)
676        : chunk_task_(chunk_task), mark_sweep_(mark_sweep) {}
677
678    ALWAYS_INLINE void operator()(mirror::Object* obj,
679                    MemberOffset offset,
680                    bool is_static ATTRIBUTE_UNUSED) const
681        SHARED_REQUIRES(Locks::mutator_lock_) {
682      Mark(obj->GetFieldObject<mirror::Object>(offset));
683    }
684
685    void VisitRootIfNonNull(mirror::CompressedReference<mirror::Object>* root) const
686        SHARED_REQUIRES(Locks::mutator_lock_) {
687      if (!root->IsNull()) {
688        VisitRoot(root);
689      }
690    }
691
692    void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const
693        SHARED_REQUIRES(Locks::mutator_lock_) {
694      if (kCheckLocks) {
695        Locks::mutator_lock_->AssertSharedHeld(Thread::Current());
696        Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current());
697      }
698      Mark(root->AsMirrorPtr());
699    }
700
701   private:
702    ALWAYS_INLINE void Mark(mirror::Object* ref) const SHARED_REQUIRES(Locks::mutator_lock_) {
703      if (ref != nullptr && mark_sweep_->MarkObjectParallel(ref)) {
704        if (kUseFinger) {
705          std::atomic_thread_fence(std::memory_order_seq_cst);
706          if (reinterpret_cast<uintptr_t>(ref) >=
707              static_cast<uintptr_t>(mark_sweep_->atomic_finger_.LoadRelaxed())) {
708            return;
709          }
710        }
711        chunk_task_->MarkStackPush(ref);
712      }
713    }
714
715    MarkStackTask<kUseFinger>* const chunk_task_;
716    MarkSweep* const mark_sweep_;
717  };
718
719  class ScanObjectParallelVisitor {
720   public:
721    ALWAYS_INLINE explicit ScanObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task)
722        : chunk_task_(chunk_task) {}
723
724    // No thread safety analysis since multiple threads will use this visitor.
725    void operator()(mirror::Object* obj) const
726        REQUIRES(Locks::heap_bitmap_lock_)
727        SHARED_REQUIRES(Locks::mutator_lock_) {
728      MarkSweep* const mark_sweep = chunk_task_->mark_sweep_;
729      MarkObjectParallelVisitor mark_visitor(chunk_task_, mark_sweep);
730      DelayReferenceReferentVisitor ref_visitor(mark_sweep);
731      mark_sweep->ScanObjectVisit(obj, mark_visitor, ref_visitor);
732    }
733
734   private:
735    MarkStackTask<kUseFinger>* const chunk_task_;
736  };
737
738  virtual ~MarkStackTask() {
739    // Make sure that we have cleared our mark stack.
740    DCHECK_EQ(mark_stack_pos_, 0U);
741    if (kCountTasks) {
742      ++mark_sweep_->work_chunks_deleted_;
743    }
744  }
745
746  MarkSweep* const mark_sweep_;
747  ThreadPool* const thread_pool_;
748  // Thread local mark stack for this task.
749  StackReference<mirror::Object> mark_stack_[kMaxSize];
750  // Mark stack position.
751  size_t mark_stack_pos_;
752
753  ALWAYS_INLINE void MarkStackPush(mirror::Object* obj)
754      SHARED_REQUIRES(Locks::mutator_lock_) {
755    if (UNLIKELY(mark_stack_pos_ == kMaxSize)) {
756      // Mark stack overflow, give 1/2 the stack to the thread pool as a new work task.
757      mark_stack_pos_ /= 2;
758      auto* task = new MarkStackTask(thread_pool_,
759                                     mark_sweep_,
760                                     kMaxSize - mark_stack_pos_,
761                                     mark_stack_ + mark_stack_pos_);
762      thread_pool_->AddTask(Thread::Current(), task);
763    }
764    DCHECK(obj != nullptr);
765    DCHECK_LT(mark_stack_pos_, kMaxSize);
766    mark_stack_[mark_stack_pos_++].Assign(obj);
767  }
768
769  virtual void Finalize() {
770    delete this;
771  }
772
773  // Scans all of the objects
774  virtual void Run(Thread* self ATTRIBUTE_UNUSED)
775      REQUIRES(Locks::heap_bitmap_lock_)
776      SHARED_REQUIRES(Locks::mutator_lock_) {
777    ScanObjectParallelVisitor visitor(this);
778    // TODO: Tune this.
779    static const size_t kFifoSize = 4;
780    BoundedFifoPowerOfTwo<mirror::Object*, kFifoSize> prefetch_fifo;
781    for (;;) {
782      mirror::Object* obj = nullptr;
783      if (kUseMarkStackPrefetch) {
784        while (mark_stack_pos_ != 0 && prefetch_fifo.size() < kFifoSize) {
785          mirror::Object* const mark_stack_obj = mark_stack_[--mark_stack_pos_].AsMirrorPtr();
786          DCHECK(mark_stack_obj != nullptr);
787          __builtin_prefetch(mark_stack_obj);
788          prefetch_fifo.push_back(mark_stack_obj);
789        }
790        if (UNLIKELY(prefetch_fifo.empty())) {
791          break;
792        }
793        obj = prefetch_fifo.front();
794        prefetch_fifo.pop_front();
795      } else {
796        if (UNLIKELY(mark_stack_pos_ == 0)) {
797          break;
798        }
799        obj = mark_stack_[--mark_stack_pos_].AsMirrorPtr();
800      }
801      DCHECK(obj != nullptr);
802      visitor(obj);
803    }
804  }
805};
806
807class MarkSweep::CardScanTask : public MarkStackTask<false> {
808 public:
809  CardScanTask(ThreadPool* thread_pool,
810               MarkSweep* mark_sweep,
811               accounting::ContinuousSpaceBitmap* bitmap,
812               uint8_t* begin,
813               uint8_t* end,
814               uint8_t minimum_age,
815               size_t mark_stack_size,
816               StackReference<mirror::Object>* mark_stack_obj,
817               bool clear_card)
818      : MarkStackTask<false>(thread_pool, mark_sweep, mark_stack_size, mark_stack_obj),
819        bitmap_(bitmap),
820        begin_(begin),
821        end_(end),
822        minimum_age_(minimum_age),
823        clear_card_(clear_card) {}
824
825 protected:
826  accounting::ContinuousSpaceBitmap* const bitmap_;
827  uint8_t* const begin_;
828  uint8_t* const end_;
829  const uint8_t minimum_age_;
830  const bool clear_card_;
831
832  virtual void Finalize() {
833    delete this;
834  }
835
836  virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS {
837    ScanObjectParallelVisitor visitor(this);
838    accounting::CardTable* card_table = mark_sweep_->GetHeap()->GetCardTable();
839    size_t cards_scanned = clear_card_
840        ? card_table->Scan<true>(bitmap_, begin_, end_, visitor, minimum_age_)
841        : card_table->Scan<false>(bitmap_, begin_, end_, visitor, minimum_age_);
842    VLOG(heap) << "Parallel scanning cards " << reinterpret_cast<void*>(begin_) << " - "
843        << reinterpret_cast<void*>(end_) << " = " << cards_scanned;
844    // Finish by emptying our local mark stack.
845    MarkStackTask::Run(self);
846  }
847};
848
849size_t MarkSweep::GetThreadCount(bool paused) const {
850  // Use less threads if we are in a background state (non jank perceptible) since we want to leave
851  // more CPU time for the foreground apps.
852  if (heap_->GetThreadPool() == nullptr || !Runtime::Current()->InJankPerceptibleProcessState()) {
853    return 1;
854  }
855  return (paused ? heap_->GetParallelGCThreadCount() : heap_->GetConcGCThreadCount()) + 1;
856}
857
858void MarkSweep::ScanGrayObjects(bool paused, uint8_t minimum_age) {
859  accounting::CardTable* card_table = GetHeap()->GetCardTable();
860  ThreadPool* thread_pool = GetHeap()->GetThreadPool();
861  size_t thread_count = GetThreadCount(paused);
862  // The parallel version with only one thread is faster for card scanning, TODO: fix.
863  if (kParallelCardScan && thread_count > 1) {
864    Thread* self = Thread::Current();
865    // Can't have a different split for each space since multiple spaces can have their cards being
866    // scanned at the same time.
867    TimingLogger::ScopedTiming t(paused ? "(Paused)ScanGrayObjects" : __FUNCTION__,
868        GetTimings());
869    // Try to take some of the mark stack since we can pass this off to the worker tasks.
870    StackReference<mirror::Object>* mark_stack_begin = mark_stack_->Begin();
871    StackReference<mirror::Object>* mark_stack_end = mark_stack_->End();
872    const size_t mark_stack_size = mark_stack_end - mark_stack_begin;
873    // Estimated number of work tasks we will create.
874    const size_t mark_stack_tasks = GetHeap()->GetContinuousSpaces().size() * thread_count;
875    DCHECK_NE(mark_stack_tasks, 0U);
876    const size_t mark_stack_delta = std::min(CardScanTask::kMaxSize / 2,
877                                             mark_stack_size / mark_stack_tasks + 1);
878    for (const auto& space : GetHeap()->GetContinuousSpaces()) {
879      if (space->GetMarkBitmap() == nullptr) {
880        continue;
881      }
882      uint8_t* card_begin = space->Begin();
883      uint8_t* card_end = space->End();
884      // Align up the end address. For example, the image space's end
885      // may not be card-size-aligned.
886      card_end = AlignUp(card_end, accounting::CardTable::kCardSize);
887      DCHECK_ALIGNED(card_begin, accounting::CardTable::kCardSize);
888      DCHECK_ALIGNED(card_end, accounting::CardTable::kCardSize);
889      // Calculate how many bytes of heap we will scan,
890      const size_t address_range = card_end - card_begin;
891      // Calculate how much address range each task gets.
892      const size_t card_delta = RoundUp(address_range / thread_count + 1,
893                                        accounting::CardTable::kCardSize);
894      // If paused and the space is neither zygote nor image space, we could clear the dirty
895      // cards to avoid accumulating them to increase card scanning load in the following GC
896      // cycles. We need to keep dirty cards of image space and zygote space in order to track
897      // references to the other spaces.
898      bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace();
899      // Create the worker tasks for this space.
900      while (card_begin != card_end) {
901        // Add a range of cards.
902        size_t addr_remaining = card_end - card_begin;
903        size_t card_increment = std::min(card_delta, addr_remaining);
904        // Take from the back of the mark stack.
905        size_t mark_stack_remaining = mark_stack_end - mark_stack_begin;
906        size_t mark_stack_increment = std::min(mark_stack_delta, mark_stack_remaining);
907        mark_stack_end -= mark_stack_increment;
908        mark_stack_->PopBackCount(static_cast<int32_t>(mark_stack_increment));
909        DCHECK_EQ(mark_stack_end, mark_stack_->End());
910        // Add the new task to the thread pool.
911        auto* task = new CardScanTask(thread_pool,
912                                      this,
913                                      space->GetMarkBitmap(),
914                                      card_begin,
915                                      card_begin + card_increment,
916                                      minimum_age,
917                                      mark_stack_increment,
918                                      mark_stack_end,
919                                      clear_card);
920        thread_pool->AddTask(self, task);
921        card_begin += card_increment;
922      }
923    }
924
925    // Note: the card scan below may dirty new cards (and scan them)
926    // as a side effect when a Reference object is encountered and
927    // queued during the marking. See b/11465268.
928    thread_pool->SetMaxActiveWorkers(thread_count - 1);
929    thread_pool->StartWorkers(self);
930    thread_pool->Wait(self, true, true);
931    thread_pool->StopWorkers(self);
932  } else {
933    for (const auto& space : GetHeap()->GetContinuousSpaces()) {
934      if (space->GetMarkBitmap() != nullptr) {
935        // Image spaces are handled properly since live == marked for them.
936        const char* name = nullptr;
937        switch (space->GetGcRetentionPolicy()) {
938        case space::kGcRetentionPolicyNeverCollect:
939          name = paused ? "(Paused)ScanGrayImageSpaceObjects" : "ScanGrayImageSpaceObjects";
940          break;
941        case space::kGcRetentionPolicyFullCollect:
942          name = paused ? "(Paused)ScanGrayZygoteSpaceObjects" : "ScanGrayZygoteSpaceObjects";
943          break;
944        case space::kGcRetentionPolicyAlwaysCollect:
945          name = paused ? "(Paused)ScanGrayAllocSpaceObjects" : "ScanGrayAllocSpaceObjects";
946          break;
947        default:
948          LOG(FATAL) << "Unreachable";
949          UNREACHABLE();
950        }
951        TimingLogger::ScopedTiming t(name, GetTimings());
952        ScanObjectVisitor visitor(this);
953        bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace();
954        if (clear_card) {
955          card_table->Scan<true>(space->GetMarkBitmap(),
956                                 space->Begin(),
957                                 space->End(),
958                                 visitor,
959                                 minimum_age);
960        } else {
961          card_table->Scan<false>(space->GetMarkBitmap(),
962                                  space->Begin(),
963                                  space->End(),
964                                  visitor,
965                                  minimum_age);
966        }
967      }
968    }
969  }
970}
971
972class MarkSweep::RecursiveMarkTask : public MarkStackTask<false> {
973 public:
974  RecursiveMarkTask(ThreadPool* thread_pool,
975                    MarkSweep* mark_sweep,
976                    accounting::ContinuousSpaceBitmap* bitmap,
977                    uintptr_t begin,
978                    uintptr_t end)
979      : MarkStackTask<false>(thread_pool, mark_sweep, 0, nullptr),
980        bitmap_(bitmap),
981        begin_(begin),
982        end_(end) {}
983
984 protected:
985  accounting::ContinuousSpaceBitmap* const bitmap_;
986  const uintptr_t begin_;
987  const uintptr_t end_;
988
989  virtual void Finalize() {
990    delete this;
991  }
992
993  // Scans all of the objects
994  virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS {
995    ScanObjectParallelVisitor visitor(this);
996    bitmap_->VisitMarkedRange(begin_, end_, visitor);
997    // Finish by emptying our local mark stack.
998    MarkStackTask::Run(self);
999  }
1000};
1001
1002// Populates the mark stack based on the set of marked objects and
1003// recursively marks until the mark stack is emptied.
1004void MarkSweep::RecursiveMark() {
1005  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1006  // RecursiveMark will build the lists of known instances of the Reference classes. See
1007  // DelayReferenceReferent for details.
1008  if (kUseRecursiveMark) {
1009    const bool partial = GetGcType() == kGcTypePartial;
1010    ScanObjectVisitor scan_visitor(this);
1011    auto* self = Thread::Current();
1012    ThreadPool* thread_pool = heap_->GetThreadPool();
1013    size_t thread_count = GetThreadCount(false);
1014    const bool parallel = kParallelRecursiveMark && thread_count > 1;
1015    mark_stack_->Reset();
1016    for (const auto& space : GetHeap()->GetContinuousSpaces()) {
1017      if ((space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) ||
1018          (!partial && space->GetGcRetentionPolicy() == space::kGcRetentionPolicyFullCollect)) {
1019        current_space_bitmap_ = space->GetMarkBitmap();
1020        if (current_space_bitmap_ == nullptr) {
1021          continue;
1022        }
1023        if (parallel) {
1024          // We will use the mark stack the future.
1025          // CHECK(mark_stack_->IsEmpty());
1026          // This function does not handle heap end increasing, so we must use the space end.
1027          uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin());
1028          uintptr_t end = reinterpret_cast<uintptr_t>(space->End());
1029          atomic_finger_.StoreRelaxed(AtomicInteger::MaxValue());
1030
1031          // Create a few worker tasks.
1032          const size_t n = thread_count * 2;
1033          while (begin != end) {
1034            uintptr_t start = begin;
1035            uintptr_t delta = (end - begin) / n;
1036            delta = RoundUp(delta, KB);
1037            if (delta < 16 * KB) delta = end - begin;
1038            begin += delta;
1039            auto* task = new RecursiveMarkTask(thread_pool,
1040                                               this,
1041                                               current_space_bitmap_,
1042                                               start,
1043                                               begin);
1044            thread_pool->AddTask(self, task);
1045          }
1046          thread_pool->SetMaxActiveWorkers(thread_count - 1);
1047          thread_pool->StartWorkers(self);
1048          thread_pool->Wait(self, true, true);
1049          thread_pool->StopWorkers(self);
1050        } else {
1051          // This function does not handle heap end increasing, so we must use the space end.
1052          uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin());
1053          uintptr_t end = reinterpret_cast<uintptr_t>(space->End());
1054          current_space_bitmap_->VisitMarkedRange(begin, end, scan_visitor);
1055        }
1056      }
1057    }
1058  }
1059  ProcessMarkStack(false);
1060}
1061
1062void MarkSweep::RecursiveMarkDirtyObjects(bool paused, uint8_t minimum_age) {
1063  ScanGrayObjects(paused, minimum_age);
1064  ProcessMarkStack(paused);
1065}
1066
1067void MarkSweep::ReMarkRoots() {
1068  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1069  Locks::mutator_lock_->AssertExclusiveHeld(Thread::Current());
1070  Runtime::Current()->VisitRoots(this, static_cast<VisitRootFlags>(
1071      kVisitRootFlagNewRoots | kVisitRootFlagStopLoggingNewRoots | kVisitRootFlagClearRootLog));
1072  if (kVerifyRootsMarked) {
1073    TimingLogger::ScopedTiming t2("(Paused)VerifyRoots", GetTimings());
1074    VerifyRootMarkedVisitor visitor(this);
1075    Runtime::Current()->VisitRoots(&visitor);
1076  }
1077}
1078
1079void MarkSweep::SweepSystemWeaks(Thread* self) {
1080  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1081  ReaderMutexLock mu(self, *Locks::heap_bitmap_lock_);
1082  Runtime::Current()->SweepSystemWeaks(this);
1083}
1084
1085class MarkSweep::VerifySystemWeakVisitor : public IsMarkedVisitor {
1086 public:
1087  explicit VerifySystemWeakVisitor(MarkSweep* mark_sweep) : mark_sweep_(mark_sweep) {}
1088
1089  virtual mirror::Object* IsMarked(mirror::Object* obj)
1090      OVERRIDE
1091      SHARED_REQUIRES(Locks::mutator_lock_, Locks::heap_bitmap_lock_) {
1092    mark_sweep_->VerifyIsLive(obj);
1093    return obj;
1094  }
1095
1096  MarkSweep* const mark_sweep_;
1097};
1098
1099void MarkSweep::VerifyIsLive(const mirror::Object* obj) {
1100  if (!heap_->GetLiveBitmap()->Test(obj)) {
1101    // TODO: Consider live stack? Has this code bitrotted?
1102    CHECK(!heap_->allocation_stack_->Contains(obj))
1103        << "Found dead object " << obj << "\n" << heap_->DumpSpaces();
1104  }
1105}
1106
1107void MarkSweep::VerifySystemWeaks() {
1108  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1109  // Verify system weaks, uses a special object visitor which returns the input object.
1110  VerifySystemWeakVisitor visitor(this);
1111  Runtime::Current()->SweepSystemWeaks(&visitor);
1112}
1113
1114class MarkSweep::CheckpointMarkThreadRoots : public Closure, public RootVisitor {
1115 public:
1116  CheckpointMarkThreadRoots(MarkSweep* mark_sweep,
1117                            bool revoke_ros_alloc_thread_local_buffers_at_checkpoint)
1118      : mark_sweep_(mark_sweep),
1119        revoke_ros_alloc_thread_local_buffers_at_checkpoint_(
1120            revoke_ros_alloc_thread_local_buffers_at_checkpoint) {
1121  }
1122
1123  void VisitRoots(mirror::Object*** roots, size_t count, const RootInfo& info ATTRIBUTE_UNUSED)
1124      OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_)
1125      REQUIRES(Locks::heap_bitmap_lock_) {
1126    for (size_t i = 0; i < count; ++i) {
1127      mark_sweep_->MarkObjectNonNullParallel(*roots[i]);
1128    }
1129  }
1130
1131  void VisitRoots(mirror::CompressedReference<mirror::Object>** roots,
1132                  size_t count,
1133                  const RootInfo& info ATTRIBUTE_UNUSED)
1134      OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_)
1135      REQUIRES(Locks::heap_bitmap_lock_) {
1136    for (size_t i = 0; i < count; ++i) {
1137      mark_sweep_->MarkObjectNonNullParallel(roots[i]->AsMirrorPtr());
1138    }
1139  }
1140
1141  virtual void Run(Thread* thread) OVERRIDE NO_THREAD_SAFETY_ANALYSIS {
1142    ScopedTrace trace("Marking thread roots");
1143    // Note: self is not necessarily equal to thread since thread may be suspended.
1144    Thread* const self = Thread::Current();
1145    CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc)
1146        << thread->GetState() << " thread " << thread << " self " << self;
1147    thread->VisitRoots(this);
1148    if (revoke_ros_alloc_thread_local_buffers_at_checkpoint_) {
1149      ScopedTrace trace2("RevokeRosAllocThreadLocalBuffers");
1150      mark_sweep_->GetHeap()->RevokeRosAllocThreadLocalBuffers(thread);
1151    }
1152    // If thread is a running mutator, then act on behalf of the garbage collector.
1153    // See the code in ThreadList::RunCheckpoint.
1154    mark_sweep_->GetBarrier().Pass(self);
1155  }
1156
1157 private:
1158  MarkSweep* const mark_sweep_;
1159  const bool revoke_ros_alloc_thread_local_buffers_at_checkpoint_;
1160};
1161
1162void MarkSweep::MarkRootsCheckpoint(Thread* self,
1163                                    bool revoke_ros_alloc_thread_local_buffers_at_checkpoint) {
1164  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1165  CheckpointMarkThreadRoots check_point(this, revoke_ros_alloc_thread_local_buffers_at_checkpoint);
1166  ThreadList* thread_list = Runtime::Current()->GetThreadList();
1167  // Request the check point is run on all threads returning a count of the threads that must
1168  // run through the barrier including self.
1169  size_t barrier_count = thread_list->RunCheckpoint(&check_point);
1170  // Release locks then wait for all mutator threads to pass the barrier.
1171  // If there are no threads to wait which implys that all the checkpoint functions are finished,
1172  // then no need to release locks.
1173  if (barrier_count == 0) {
1174    return;
1175  }
1176  Locks::heap_bitmap_lock_->ExclusiveUnlock(self);
1177  Locks::mutator_lock_->SharedUnlock(self);
1178  {
1179    ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
1180    gc_barrier_->Increment(self, barrier_count);
1181  }
1182  Locks::mutator_lock_->SharedLock(self);
1183  Locks::heap_bitmap_lock_->ExclusiveLock(self);
1184}
1185
1186void MarkSweep::SweepArray(accounting::ObjectStack* allocations, bool swap_bitmaps) {
1187  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1188  Thread* self = Thread::Current();
1189  mirror::Object** chunk_free_buffer = reinterpret_cast<mirror::Object**>(
1190      sweep_array_free_buffer_mem_map_->BaseBegin());
1191  size_t chunk_free_pos = 0;
1192  ObjectBytePair freed;
1193  ObjectBytePair freed_los;
1194  // How many objects are left in the array, modified after each space is swept.
1195  StackReference<mirror::Object>* objects = allocations->Begin();
1196  size_t count = allocations->Size();
1197  // Change the order to ensure that the non-moving space last swept as an optimization.
1198  std::vector<space::ContinuousSpace*> sweep_spaces;
1199  space::ContinuousSpace* non_moving_space = nullptr;
1200  for (space::ContinuousSpace* space : heap_->GetContinuousSpaces()) {
1201    if (space->IsAllocSpace() &&
1202        !immune_spaces_.ContainsSpace(space) &&
1203        space->GetLiveBitmap() != nullptr) {
1204      if (space == heap_->GetNonMovingSpace()) {
1205        non_moving_space = space;
1206      } else {
1207        sweep_spaces.push_back(space);
1208      }
1209    }
1210  }
1211  // Unlikely to sweep a significant amount of non_movable objects, so we do these after the after
1212  // the other alloc spaces as an optimization.
1213  if (non_moving_space != nullptr) {
1214    sweep_spaces.push_back(non_moving_space);
1215  }
1216  // Start by sweeping the continuous spaces.
1217  for (space::ContinuousSpace* space : sweep_spaces) {
1218    space::AllocSpace* alloc_space = space->AsAllocSpace();
1219    accounting::ContinuousSpaceBitmap* live_bitmap = space->GetLiveBitmap();
1220    accounting::ContinuousSpaceBitmap* mark_bitmap = space->GetMarkBitmap();
1221    if (swap_bitmaps) {
1222      std::swap(live_bitmap, mark_bitmap);
1223    }
1224    StackReference<mirror::Object>* out = objects;
1225    for (size_t i = 0; i < count; ++i) {
1226      mirror::Object* const obj = objects[i].AsMirrorPtr();
1227      if (kUseThreadLocalAllocationStack && obj == nullptr) {
1228        continue;
1229      }
1230      if (space->HasAddress(obj)) {
1231        // This object is in the space, remove it from the array and add it to the sweep buffer
1232        // if needed.
1233        if (!mark_bitmap->Test(obj)) {
1234          if (chunk_free_pos >= kSweepArrayChunkFreeSize) {
1235            TimingLogger::ScopedTiming t2("FreeList", GetTimings());
1236            freed.objects += chunk_free_pos;
1237            freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer);
1238            chunk_free_pos = 0;
1239          }
1240          chunk_free_buffer[chunk_free_pos++] = obj;
1241        }
1242      } else {
1243        (out++)->Assign(obj);
1244      }
1245    }
1246    if (chunk_free_pos > 0) {
1247      TimingLogger::ScopedTiming t2("FreeList", GetTimings());
1248      freed.objects += chunk_free_pos;
1249      freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer);
1250      chunk_free_pos = 0;
1251    }
1252    // All of the references which space contained are no longer in the allocation stack, update
1253    // the count.
1254    count = out - objects;
1255  }
1256  // Handle the large object space.
1257  space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace();
1258  if (large_object_space != nullptr) {
1259    accounting::LargeObjectBitmap* large_live_objects = large_object_space->GetLiveBitmap();
1260    accounting::LargeObjectBitmap* large_mark_objects = large_object_space->GetMarkBitmap();
1261    if (swap_bitmaps) {
1262      std::swap(large_live_objects, large_mark_objects);
1263    }
1264    for (size_t i = 0; i < count; ++i) {
1265      mirror::Object* const obj = objects[i].AsMirrorPtr();
1266      // Handle large objects.
1267      if (kUseThreadLocalAllocationStack && obj == nullptr) {
1268        continue;
1269      }
1270      if (!large_mark_objects->Test(obj)) {
1271        ++freed_los.objects;
1272        freed_los.bytes += large_object_space->Free(self, obj);
1273      }
1274    }
1275  }
1276  {
1277    TimingLogger::ScopedTiming t2("RecordFree", GetTimings());
1278    RecordFree(freed);
1279    RecordFreeLOS(freed_los);
1280    t2.NewTiming("ResetStack");
1281    allocations->Reset();
1282  }
1283  sweep_array_free_buffer_mem_map_->MadviseDontNeedAndZero();
1284}
1285
1286void MarkSweep::Sweep(bool swap_bitmaps) {
1287  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1288  // Ensure that nobody inserted items in the live stack after we swapped the stacks.
1289  CHECK_GE(live_stack_freeze_size_, GetHeap()->GetLiveStack()->Size());
1290  {
1291    TimingLogger::ScopedTiming t2("MarkAllocStackAsLive", GetTimings());
1292    // Mark everything allocated since the last as GC live so that we can sweep concurrently,
1293    // knowing that new allocations won't be marked as live.
1294    accounting::ObjectStack* live_stack = heap_->GetLiveStack();
1295    heap_->MarkAllocStackAsLive(live_stack);
1296    live_stack->Reset();
1297    DCHECK(mark_stack_->IsEmpty());
1298  }
1299  for (const auto& space : GetHeap()->GetContinuousSpaces()) {
1300    if (space->IsContinuousMemMapAllocSpace()) {
1301      space::ContinuousMemMapAllocSpace* alloc_space = space->AsContinuousMemMapAllocSpace();
1302      TimingLogger::ScopedTiming split(
1303          alloc_space->IsZygoteSpace() ? "SweepZygoteSpace" : "SweepMallocSpace",
1304          GetTimings());
1305      RecordFree(alloc_space->Sweep(swap_bitmaps));
1306    }
1307  }
1308  SweepLargeObjects(swap_bitmaps);
1309}
1310
1311void MarkSweep::SweepLargeObjects(bool swap_bitmaps) {
1312  space::LargeObjectSpace* los = heap_->GetLargeObjectsSpace();
1313  if (los != nullptr) {
1314    TimingLogger::ScopedTiming split(__FUNCTION__, GetTimings());
1315    RecordFreeLOS(los->Sweep(swap_bitmaps));
1316  }
1317}
1318
1319// Process the "referent" field in a java.lang.ref.Reference.  If the referent has not yet been
1320// marked, put it on the appropriate list in the heap for later processing.
1321void MarkSweep::DelayReferenceReferent(mirror::Class* klass, mirror::Reference* ref) {
1322  heap_->GetReferenceProcessor()->DelayReferenceReferent(klass, ref, this);
1323}
1324
1325class MarkVisitor {
1326 public:
1327  ALWAYS_INLINE explicit MarkVisitor(MarkSweep* const mark_sweep) : mark_sweep_(mark_sweep) {}
1328
1329  ALWAYS_INLINE void operator()(mirror::Object* obj,
1330                                MemberOffset offset,
1331                                bool is_static ATTRIBUTE_UNUSED) const
1332      REQUIRES(Locks::heap_bitmap_lock_)
1333      SHARED_REQUIRES(Locks::mutator_lock_) {
1334    if (kCheckLocks) {
1335      Locks::mutator_lock_->AssertSharedHeld(Thread::Current());
1336      Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current());
1337    }
1338    mark_sweep_->MarkObject(obj->GetFieldObject<mirror::Object>(offset), obj, offset);
1339  }
1340
1341  void VisitRootIfNonNull(mirror::CompressedReference<mirror::Object>* root) const
1342      REQUIRES(Locks::heap_bitmap_lock_)
1343      SHARED_REQUIRES(Locks::mutator_lock_) {
1344    if (!root->IsNull()) {
1345      VisitRoot(root);
1346    }
1347  }
1348
1349  void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const
1350      REQUIRES(Locks::heap_bitmap_lock_)
1351      SHARED_REQUIRES(Locks::mutator_lock_) {
1352    if (kCheckLocks) {
1353      Locks::mutator_lock_->AssertSharedHeld(Thread::Current());
1354      Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current());
1355    }
1356    mark_sweep_->MarkObject(root->AsMirrorPtr());
1357  }
1358
1359 private:
1360  MarkSweep* const mark_sweep_;
1361};
1362
1363// Scans an object reference.  Determines the type of the reference
1364// and dispatches to a specialized scanning routine.
1365void MarkSweep::ScanObject(mirror::Object* obj) {
1366  MarkVisitor mark_visitor(this);
1367  DelayReferenceReferentVisitor ref_visitor(this);
1368  ScanObjectVisit(obj, mark_visitor, ref_visitor);
1369}
1370
1371void MarkSweep::ProcessMarkStackParallel(size_t thread_count) {
1372  Thread* self = Thread::Current();
1373  ThreadPool* thread_pool = GetHeap()->GetThreadPool();
1374  const size_t chunk_size = std::min(mark_stack_->Size() / thread_count + 1,
1375                                     static_cast<size_t>(MarkStackTask<false>::kMaxSize));
1376  CHECK_GT(chunk_size, 0U);
1377  // Split the current mark stack up into work tasks.
1378  for (auto* it = mark_stack_->Begin(), *end = mark_stack_->End(); it < end; ) {
1379    const size_t delta = std::min(static_cast<size_t>(end - it), chunk_size);
1380    thread_pool->AddTask(self, new MarkStackTask<false>(thread_pool, this, delta, it));
1381    it += delta;
1382  }
1383  thread_pool->SetMaxActiveWorkers(thread_count - 1);
1384  thread_pool->StartWorkers(self);
1385  thread_pool->Wait(self, true, true);
1386  thread_pool->StopWorkers(self);
1387  mark_stack_->Reset();
1388  CHECK_EQ(work_chunks_created_.LoadSequentiallyConsistent(),
1389           work_chunks_deleted_.LoadSequentiallyConsistent())
1390      << " some of the work chunks were leaked";
1391}
1392
1393// Scan anything that's on the mark stack.
1394void MarkSweep::ProcessMarkStack(bool paused) {
1395  TimingLogger::ScopedTiming t(paused ? "(Paused)ProcessMarkStack" : __FUNCTION__, GetTimings());
1396  size_t thread_count = GetThreadCount(paused);
1397  if (kParallelProcessMarkStack && thread_count > 1 &&
1398      mark_stack_->Size() >= kMinimumParallelMarkStackSize) {
1399    ProcessMarkStackParallel(thread_count);
1400  } else {
1401    // TODO: Tune this.
1402    static const size_t kFifoSize = 4;
1403    BoundedFifoPowerOfTwo<mirror::Object*, kFifoSize> prefetch_fifo;
1404    for (;;) {
1405      mirror::Object* obj = nullptr;
1406      if (kUseMarkStackPrefetch) {
1407        while (!mark_stack_->IsEmpty() && prefetch_fifo.size() < kFifoSize) {
1408          mirror::Object* mark_stack_obj = mark_stack_->PopBack();
1409          DCHECK(mark_stack_obj != nullptr);
1410          __builtin_prefetch(mark_stack_obj);
1411          prefetch_fifo.push_back(mark_stack_obj);
1412        }
1413        if (prefetch_fifo.empty()) {
1414          break;
1415        }
1416        obj = prefetch_fifo.front();
1417        prefetch_fifo.pop_front();
1418      } else {
1419        if (mark_stack_->IsEmpty()) {
1420          break;
1421        }
1422        obj = mark_stack_->PopBack();
1423      }
1424      DCHECK(obj != nullptr);
1425      ScanObject(obj);
1426    }
1427  }
1428}
1429
1430inline mirror::Object* MarkSweep::IsMarked(mirror::Object* object) {
1431  if (immune_spaces_.IsInImmuneRegion(object)) {
1432    return object;
1433  }
1434  if (current_space_bitmap_->HasAddress(object)) {
1435    return current_space_bitmap_->Test(object) ? object : nullptr;
1436  }
1437  return mark_bitmap_->Test(object) ? object : nullptr;
1438}
1439
1440void MarkSweep::FinishPhase() {
1441  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1442  if (kCountScannedTypes) {
1443    VLOG(gc)
1444        << "MarkSweep scanned"
1445        << " no reference objects=" << no_reference_class_count_.LoadRelaxed()
1446        << " normal objects=" << normal_count_.LoadRelaxed()
1447        << " classes=" << class_count_.LoadRelaxed()
1448        << " object arrays=" << object_array_count_.LoadRelaxed()
1449        << " references=" << reference_count_.LoadRelaxed()
1450        << " other=" << other_count_.LoadRelaxed();
1451  }
1452  if (kCountTasks) {
1453    VLOG(gc) << "Total number of work chunks allocated: " << work_chunks_created_.LoadRelaxed();
1454  }
1455  if (kMeasureOverhead) {
1456    VLOG(gc) << "Overhead time " << PrettyDuration(overhead_time_.LoadRelaxed());
1457  }
1458  if (kProfileLargeObjects) {
1459    VLOG(gc) << "Large objects tested " << large_object_test_.LoadRelaxed()
1460        << " marked " << large_object_mark_.LoadRelaxed();
1461  }
1462  if (kCountMarkedObjects) {
1463    VLOG(gc) << "Marked: null=" << mark_null_count_.LoadRelaxed()
1464        << " immune=" <<  mark_immune_count_.LoadRelaxed()
1465        << " fastpath=" << mark_fastpath_count_.LoadRelaxed()
1466        << " slowpath=" << mark_slowpath_count_.LoadRelaxed();
1467  }
1468  CHECK(mark_stack_->IsEmpty());  // Ensure that the mark stack is empty.
1469  mark_stack_->Reset();
1470  Thread* const self = Thread::Current();
1471  ReaderMutexLock mu(self, *Locks::mutator_lock_);
1472  WriterMutexLock mu2(self, *Locks::heap_bitmap_lock_);
1473  heap_->ClearMarkedObjects();
1474}
1475
1476void MarkSweep::RevokeAllThreadLocalBuffers() {
1477  if (kRevokeRosAllocThreadLocalBuffersAtCheckpoint && IsConcurrent()) {
1478    // If concurrent, rosalloc thread-local buffers are revoked at the
1479    // thread checkpoint. Bump pointer space thread-local buffers must
1480    // not be in use.
1481    GetHeap()->AssertAllBumpPointerSpaceThreadLocalBuffersAreRevoked();
1482  } else {
1483    TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1484    GetHeap()->RevokeAllThreadLocalBuffers();
1485  }
1486}
1487
1488}  // namespace collector
1489}  // namespace gc
1490}  // namespace art
1491