mark_sweep.cc revision f8cb1781a4e4be5df1f845206f7b37ed89092b64 
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 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 MarkSweepMarkObjectSlowPath {
397 public:
398  explicit MarkSweepMarkObjectSlowPath(MarkSweep* mark_sweep,
399                                       mirror::Object* holder = nullptr,
400                                       MemberOffset offset = MemberOffset(0))
401      : mark_sweep_(mark_sweep), holder_(holder), offset_(offset) {}
402
403  void operator()(const mirror::Object* obj) const NO_THREAD_SAFETY_ANALYSIS {
404    if (kProfileLargeObjects) {
405      // TODO: Differentiate between marking and testing somehow.
406      ++mark_sweep_->large_object_test_;
407      ++mark_sweep_->large_object_mark_;
408    }
409    space::LargeObjectSpace* large_object_space = mark_sweep_->GetHeap()->GetLargeObjectsSpace();
410    if (UNLIKELY(obj == nullptr || !IsAligned<kPageSize>(obj) ||
411                 (kIsDebugBuild && large_object_space != nullptr &&
412                     !large_object_space->Contains(obj)))) {
413      LOG(INTERNAL_FATAL) << "Tried to mark " << obj << " not contained by any spaces";
414      if (holder_ != nullptr) {
415        size_t holder_size = holder_->SizeOf();
416        ArtField* field = holder_->FindFieldByOffset(offset_);
417        LOG(INTERNAL_FATAL) << "Field info: "
418                            << " holder=" << holder_
419                            << " holder is "
420                            << (mark_sweep_->GetHeap()->IsLiveObjectLocked(holder_)
421                                ? "alive" : "dead")
422                            << " holder_size=" << holder_size
423                            << " holder_type=" << PrettyTypeOf(holder_)
424                            << " offset=" << offset_.Uint32Value()
425                            << " field=" << (field != nullptr ? field->GetName() : "nullptr")
426                            << " field_type="
427                            << (field != nullptr ? field->GetTypeDescriptor() : "")
428                            << " first_ref_field_offset="
429                            << (holder_->IsClass()
430                                ? holder_->AsClass()->GetFirstReferenceStaticFieldOffset(
431                                    sizeof(void*))
432                                : holder_->GetClass()->GetFirstReferenceInstanceFieldOffset())
433                            << " num_of_ref_fields="
434                            << (holder_->IsClass()
435                                ? holder_->AsClass()->NumReferenceStaticFields()
436                                : holder_->GetClass()->NumReferenceInstanceFields())
437                            << "\n";
438        // Print the memory content of the holder.
439        for (size_t i = 0; i < holder_size / sizeof(uint32_t); ++i) {
440          uint32_t* p = reinterpret_cast<uint32_t*>(holder_);
441          LOG(INTERNAL_FATAL) << &p[i] << ": " << "holder+" << (i * sizeof(uint32_t)) << " = "
442                              << std::hex << p[i];
443        }
444      }
445      PrintFileToLog("/proc/self/maps", LogSeverity::INTERNAL_FATAL);
446      MemMap::DumpMaps(LOG(INTERNAL_FATAL), true);
447      {
448        LOG(INTERNAL_FATAL) << "Attempting see if it's a bad root";
449        Thread* self = Thread::Current();
450        if (Locks::mutator_lock_->IsExclusiveHeld(self)) {
451          mark_sweep_->VerifyRoots();
452        } else {
453          const bool heap_bitmap_exclusive_locked =
454              Locks::heap_bitmap_lock_->IsExclusiveHeld(self);
455          if (heap_bitmap_exclusive_locked) {
456            Locks::heap_bitmap_lock_->ExclusiveUnlock(self);
457          }
458          {
459            ScopedThreadSuspension(self, kSuspended);
460            ScopedSuspendAll ssa(__FUNCTION__);
461            mark_sweep_->VerifyRoots();
462          }
463          if (heap_bitmap_exclusive_locked) {
464            Locks::heap_bitmap_lock_->ExclusiveLock(self);
465          }
466        }
467      }
468      LOG(FATAL) << "Can't mark invalid object";
469    }
470  }
471
472 private:
473  MarkSweep* const mark_sweep_;
474  mirror::Object* const holder_;
475  MemberOffset offset_;
476};
477
478inline void MarkSweep::MarkObjectNonNull(mirror::Object* obj,
479                                         mirror::Object* holder,
480                                         MemberOffset offset) {
481  DCHECK(obj != nullptr);
482  if (kUseBakerOrBrooksReadBarrier) {
483    // Verify all the objects have the correct pointer installed.
484    obj->AssertReadBarrierPointer();
485  }
486  if (immune_spaces_.IsInImmuneRegion(obj)) {
487    if (kCountMarkedObjects) {
488      ++mark_immune_count_;
489    }
490    DCHECK(mark_bitmap_->Test(obj));
491  } else if (LIKELY(current_space_bitmap_->HasAddress(obj))) {
492    if (kCountMarkedObjects) {
493      ++mark_fastpath_count_;
494    }
495    if (UNLIKELY(!current_space_bitmap_->Set(obj))) {
496      PushOnMarkStack(obj);  // This object was not previously marked.
497    }
498  } else {
499    if (kCountMarkedObjects) {
500      ++mark_slowpath_count_;
501    }
502    MarkSweepMarkObjectSlowPath visitor(this, holder, offset);
503    // TODO: We already know that the object is not in the current_space_bitmap_ but MarkBitmap::Set
504    // will check again.
505    if (!mark_bitmap_->Set(obj, visitor)) {
506      PushOnMarkStack(obj);  // Was not already marked, push.
507    }
508  }
509}
510
511inline void MarkSweep::PushOnMarkStack(mirror::Object* obj) {
512  if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) {
513    // Lock is not needed but is here anyways to please annotalysis.
514    MutexLock mu(Thread::Current(), mark_stack_lock_);
515    ExpandMarkStack();
516  }
517  // The object must be pushed on to the mark stack.
518  mark_stack_->PushBack(obj);
519}
520
521inline bool MarkSweep::MarkObjectParallel(mirror::Object* obj) {
522  DCHECK(obj != nullptr);
523  if (kUseBakerOrBrooksReadBarrier) {
524    // Verify all the objects have the correct pointer installed.
525    obj->AssertReadBarrierPointer();
526  }
527  if (immune_spaces_.IsInImmuneRegion(obj)) {
528    DCHECK(IsMarked(obj) != nullptr);
529    return false;
530  }
531  // Try to take advantage of locality of references within a space, failing this find the space
532  // the hard way.
533  accounting::ContinuousSpaceBitmap* object_bitmap = current_space_bitmap_;
534  if (LIKELY(object_bitmap->HasAddress(obj))) {
535    return !object_bitmap->AtomicTestAndSet(obj);
536  }
537  MarkSweepMarkObjectSlowPath visitor(this);
538  return !mark_bitmap_->AtomicTestAndSet(obj, visitor);
539}
540
541void MarkSweep::MarkHeapReference(mirror::HeapReference<mirror::Object>* ref) {
542  MarkObject(ref->AsMirrorPtr(), nullptr, MemberOffset(0));
543}
544
545// Used to mark objects when processing the mark stack. If an object is null, it is not marked.
546inline void MarkSweep::MarkObject(mirror::Object* obj,
547                                  mirror::Object* holder,
548                                  MemberOffset offset) {
549  if (obj != nullptr) {
550    MarkObjectNonNull(obj, holder, offset);
551  } else if (kCountMarkedObjects) {
552    ++mark_null_count_;
553  }
554}
555
556class VerifyRootMarkedVisitor : public SingleRootVisitor {
557 public:
558  explicit VerifyRootMarkedVisitor(MarkSweep* collector) : collector_(collector) { }
559
560  void VisitRoot(mirror::Object* root, const RootInfo& info) OVERRIDE
561      SHARED_REQUIRES(Locks::mutator_lock_, Locks::heap_bitmap_lock_) {
562    CHECK(collector_->IsMarked(root) != nullptr) << info.ToString();
563  }
564
565 private:
566  MarkSweep* const collector_;
567};
568
569void MarkSweep::VisitRoots(mirror::Object*** roots,
570                           size_t count,
571                           const RootInfo& info ATTRIBUTE_UNUSED) {
572  for (size_t i = 0; i < count; ++i) {
573    MarkObjectNonNull(*roots[i]);
574  }
575}
576
577void MarkSweep::VisitRoots(mirror::CompressedReference<mirror::Object>** roots,
578                           size_t count,
579                           const RootInfo& info ATTRIBUTE_UNUSED) {
580  for (size_t i = 0; i < count; ++i) {
581    MarkObjectNonNull(roots[i]->AsMirrorPtr());
582  }
583}
584
585class VerifyRootVisitor : public SingleRootVisitor {
586 public:
587  void VisitRoot(mirror::Object* root, const RootInfo& info) OVERRIDE
588      SHARED_REQUIRES(Locks::mutator_lock_, Locks::heap_bitmap_lock_) {
589    // See if the root is on any space bitmap.
590    auto* heap = Runtime::Current()->GetHeap();
591    if (heap->GetLiveBitmap()->GetContinuousSpaceBitmap(root) == nullptr) {
592      space::LargeObjectSpace* large_object_space = heap->GetLargeObjectsSpace();
593      if (large_object_space != nullptr && !large_object_space->Contains(root)) {
594        LOG(INTERNAL_FATAL) << "Found invalid root: " << root << " " << info;
595      }
596    }
597  }
598};
599
600void MarkSweep::VerifyRoots() {
601  VerifyRootVisitor visitor;
602  Runtime::Current()->GetThreadList()->VisitRoots(&visitor);
603}
604
605void MarkSweep::MarkRoots(Thread* self) {
606  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
607  if (Locks::mutator_lock_->IsExclusiveHeld(self)) {
608    // If we exclusively hold the mutator lock, all threads must be suspended.
609    Runtime::Current()->VisitRoots(this);
610    RevokeAllThreadLocalAllocationStacks(self);
611  } else {
612    MarkRootsCheckpoint(self, kRevokeRosAllocThreadLocalBuffersAtCheckpoint);
613    // At this point the live stack should no longer have any mutators which push into it.
614    MarkNonThreadRoots();
615    MarkConcurrentRoots(
616        static_cast<VisitRootFlags>(kVisitRootFlagAllRoots | kVisitRootFlagStartLoggingNewRoots));
617  }
618}
619
620void MarkSweep::MarkNonThreadRoots() {
621  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
622  Runtime::Current()->VisitNonThreadRoots(this);
623}
624
625void MarkSweep::MarkConcurrentRoots(VisitRootFlags flags) {
626  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
627  // Visit all runtime roots and clear dirty flags.
628  Runtime::Current()->VisitConcurrentRoots(
629      this, static_cast<VisitRootFlags>(flags | kVisitRootFlagNonMoving));
630}
631
632class DelayReferenceReferentVisitor {
633 public:
634  explicit DelayReferenceReferentVisitor(MarkSweep* collector) : collector_(collector) {}
635
636  void operator()(mirror::Class* klass, mirror::Reference* ref) const
637      REQUIRES(Locks::heap_bitmap_lock_)
638      SHARED_REQUIRES(Locks::mutator_lock_) {
639    collector_->DelayReferenceReferent(klass, ref);
640  }
641
642 private:
643  MarkSweep* const collector_;
644};
645
646template <bool kUseFinger = false>
647class MarkStackTask : public Task {
648 public:
649  MarkStackTask(ThreadPool* thread_pool,
650                MarkSweep* mark_sweep,
651                size_t mark_stack_size,
652                StackReference<mirror::Object>* mark_stack)
653      : mark_sweep_(mark_sweep),
654        thread_pool_(thread_pool),
655        mark_stack_pos_(mark_stack_size) {
656    // We may have to copy part of an existing mark stack when another mark stack overflows.
657    if (mark_stack_size != 0) {
658      DCHECK(mark_stack != nullptr);
659      // TODO: Check performance?
660      std::copy(mark_stack, mark_stack + mark_stack_size, mark_stack_);
661    }
662    if (kCountTasks) {
663      ++mark_sweep_->work_chunks_created_;
664    }
665  }
666
667  static const size_t kMaxSize = 1 * KB;
668
669 protected:
670  class MarkObjectParallelVisitor {
671   public:
672    ALWAYS_INLINE MarkObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task,
673                                            MarkSweep* mark_sweep)
674        : chunk_task_(chunk_task), mark_sweep_(mark_sweep) {}
675
676    ALWAYS_INLINE void operator()(mirror::Object* obj,
677                    MemberOffset offset,
678                    bool is_static ATTRIBUTE_UNUSED) const
679        SHARED_REQUIRES(Locks::mutator_lock_) {
680      Mark(obj->GetFieldObject<mirror::Object>(offset));
681    }
682
683    void VisitRootIfNonNull(mirror::CompressedReference<mirror::Object>* root) const
684        SHARED_REQUIRES(Locks::mutator_lock_) {
685      if (!root->IsNull()) {
686        VisitRoot(root);
687      }
688    }
689
690    void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const
691        SHARED_REQUIRES(Locks::mutator_lock_) {
692      if (kCheckLocks) {
693        Locks::mutator_lock_->AssertSharedHeld(Thread::Current());
694        Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current());
695      }
696      Mark(root->AsMirrorPtr());
697    }
698
699   private:
700    ALWAYS_INLINE void Mark(mirror::Object* ref) const SHARED_REQUIRES(Locks::mutator_lock_) {
701      if (ref != nullptr && mark_sweep_->MarkObjectParallel(ref)) {
702        if (kUseFinger) {
703          std::atomic_thread_fence(std::memory_order_seq_cst);
704          if (reinterpret_cast<uintptr_t>(ref) >=
705              static_cast<uintptr_t>(mark_sweep_->atomic_finger_.LoadRelaxed())) {
706            return;
707          }
708        }
709        chunk_task_->MarkStackPush(ref);
710      }
711    }
712
713    MarkStackTask<kUseFinger>* const chunk_task_;
714    MarkSweep* const mark_sweep_;
715  };
716
717  class ScanObjectParallelVisitor {
718   public:
719    ALWAYS_INLINE explicit ScanObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task)
720        : chunk_task_(chunk_task) {}
721
722    // No thread safety analysis since multiple threads will use this visitor.
723    void operator()(mirror::Object* obj) const
724        REQUIRES(Locks::heap_bitmap_lock_)
725        SHARED_REQUIRES(Locks::mutator_lock_) {
726      MarkSweep* const mark_sweep = chunk_task_->mark_sweep_;
727      MarkObjectParallelVisitor mark_visitor(chunk_task_, mark_sweep);
728      DelayReferenceReferentVisitor ref_visitor(mark_sweep);
729      mark_sweep->ScanObjectVisit(obj, mark_visitor, ref_visitor);
730    }
731
732   private:
733    MarkStackTask<kUseFinger>* const chunk_task_;
734  };
735
736  virtual ~MarkStackTask() {
737    // Make sure that we have cleared our mark stack.
738    DCHECK_EQ(mark_stack_pos_, 0U);
739    if (kCountTasks) {
740      ++mark_sweep_->work_chunks_deleted_;
741    }
742  }
743
744  MarkSweep* const mark_sweep_;
745  ThreadPool* const thread_pool_;
746  // Thread local mark stack for this task.
747  StackReference<mirror::Object> mark_stack_[kMaxSize];
748  // Mark stack position.
749  size_t mark_stack_pos_;
750
751  ALWAYS_INLINE void MarkStackPush(mirror::Object* obj)
752      SHARED_REQUIRES(Locks::mutator_lock_) {
753    if (UNLIKELY(mark_stack_pos_ == kMaxSize)) {
754      // Mark stack overflow, give 1/2 the stack to the thread pool as a new work task.
755      mark_stack_pos_ /= 2;
756      auto* task = new MarkStackTask(thread_pool_,
757                                     mark_sweep_,
758                                     kMaxSize - mark_stack_pos_,
759                                     mark_stack_ + mark_stack_pos_);
760      thread_pool_->AddTask(Thread::Current(), task);
761    }
762    DCHECK(obj != nullptr);
763    DCHECK_LT(mark_stack_pos_, kMaxSize);
764    mark_stack_[mark_stack_pos_++].Assign(obj);
765  }
766
767  virtual void Finalize() {
768    delete this;
769  }
770
771  // Scans all of the objects
772  virtual void Run(Thread* self ATTRIBUTE_UNUSED)
773      REQUIRES(Locks::heap_bitmap_lock_)
774      SHARED_REQUIRES(Locks::mutator_lock_) {
775    ScanObjectParallelVisitor visitor(this);
776    // TODO: Tune this.
777    static const size_t kFifoSize = 4;
778    BoundedFifoPowerOfTwo<mirror::Object*, kFifoSize> prefetch_fifo;
779    for (;;) {
780      mirror::Object* obj = nullptr;
781      if (kUseMarkStackPrefetch) {
782        while (mark_stack_pos_ != 0 && prefetch_fifo.size() < kFifoSize) {
783          mirror::Object* const mark_stack_obj = mark_stack_[--mark_stack_pos_].AsMirrorPtr();
784          DCHECK(mark_stack_obj != nullptr);
785          __builtin_prefetch(mark_stack_obj);
786          prefetch_fifo.push_back(mark_stack_obj);
787        }
788        if (UNLIKELY(prefetch_fifo.empty())) {
789          break;
790        }
791        obj = prefetch_fifo.front();
792        prefetch_fifo.pop_front();
793      } else {
794        if (UNLIKELY(mark_stack_pos_ == 0)) {
795          break;
796        }
797        obj = mark_stack_[--mark_stack_pos_].AsMirrorPtr();
798      }
799      DCHECK(obj != nullptr);
800      visitor(obj);
801    }
802  }
803};
804
805class CardScanTask : public MarkStackTask<false> {
806 public:
807  CardScanTask(ThreadPool* thread_pool,
808               MarkSweep* mark_sweep,
809               accounting::ContinuousSpaceBitmap* bitmap,
810               uint8_t* begin,
811               uint8_t* end,
812               uint8_t minimum_age,
813               size_t mark_stack_size,
814               StackReference<mirror::Object>* mark_stack_obj,
815               bool clear_card)
816      : MarkStackTask<false>(thread_pool, mark_sweep, mark_stack_size, mark_stack_obj),
817        bitmap_(bitmap),
818        begin_(begin),
819        end_(end),
820        minimum_age_(minimum_age),
821        clear_card_(clear_card) {}
822
823 protected:
824  accounting::ContinuousSpaceBitmap* const bitmap_;
825  uint8_t* const begin_;
826  uint8_t* const end_;
827  const uint8_t minimum_age_;
828  const bool clear_card_;
829
830  virtual void Finalize() {
831    delete this;
832  }
833
834  virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS {
835    ScanObjectParallelVisitor visitor(this);
836    accounting::CardTable* card_table = mark_sweep_->GetHeap()->GetCardTable();
837    size_t cards_scanned = clear_card_
838        ? card_table->Scan<true>(bitmap_, begin_, end_, visitor, minimum_age_)
839        : card_table->Scan<false>(bitmap_, begin_, end_, visitor, minimum_age_);
840    VLOG(heap) << "Parallel scanning cards " << reinterpret_cast<void*>(begin_) << " - "
841        << reinterpret_cast<void*>(end_) << " = " << cards_scanned;
842    // Finish by emptying our local mark stack.
843    MarkStackTask::Run(self);
844  }
845};
846
847size_t MarkSweep::GetThreadCount(bool paused) const {
848  // Use less threads if we are in a background state (non jank perceptible) since we want to leave
849  // more CPU time for the foreground apps.
850  if (heap_->GetThreadPool() == nullptr || !Runtime::Current()->InJankPerceptibleProcessState()) {
851    return 1;
852  }
853  return (paused ? heap_->GetParallelGCThreadCount() : heap_->GetConcGCThreadCount()) + 1;
854}
855
856void MarkSweep::ScanGrayObjects(bool paused, uint8_t minimum_age) {
857  accounting::CardTable* card_table = GetHeap()->GetCardTable();
858  ThreadPool* thread_pool = GetHeap()->GetThreadPool();
859  size_t thread_count = GetThreadCount(paused);
860  // The parallel version with only one thread is faster for card scanning, TODO: fix.
861  if (kParallelCardScan && thread_count > 1) {
862    Thread* self = Thread::Current();
863    // Can't have a different split for each space since multiple spaces can have their cards being
864    // scanned at the same time.
865    TimingLogger::ScopedTiming t(paused ? "(Paused)ScanGrayObjects" : __FUNCTION__,
866        GetTimings());
867    // Try to take some of the mark stack since we can pass this off to the worker tasks.
868    StackReference<mirror::Object>* mark_stack_begin = mark_stack_->Begin();
869    StackReference<mirror::Object>* mark_stack_end = mark_stack_->End();
870    const size_t mark_stack_size = mark_stack_end - mark_stack_begin;
871    // Estimated number of work tasks we will create.
872    const size_t mark_stack_tasks = GetHeap()->GetContinuousSpaces().size() * thread_count;
873    DCHECK_NE(mark_stack_tasks, 0U);
874    const size_t mark_stack_delta = std::min(CardScanTask::kMaxSize / 2,
875                                             mark_stack_size / mark_stack_tasks + 1);
876    for (const auto& space : GetHeap()->GetContinuousSpaces()) {
877      if (space->GetMarkBitmap() == nullptr) {
878        continue;
879      }
880      uint8_t* card_begin = space->Begin();
881      uint8_t* card_end = space->End();
882      // Align up the end address. For example, the image space's end
883      // may not be card-size-aligned.
884      card_end = AlignUp(card_end, accounting::CardTable::kCardSize);
885      DCHECK_ALIGNED(card_begin, accounting::CardTable::kCardSize);
886      DCHECK_ALIGNED(card_end, accounting::CardTable::kCardSize);
887      // Calculate how many bytes of heap we will scan,
888      const size_t address_range = card_end - card_begin;
889      // Calculate how much address range each task gets.
890      const size_t card_delta = RoundUp(address_range / thread_count + 1,
891                                        accounting::CardTable::kCardSize);
892      // If paused and the space is neither zygote nor image space, we could clear the dirty
893      // cards to avoid accumulating them to increase card scanning load in the following GC
894      // cycles. We need to keep dirty cards of image space and zygote space in order to track
895      // references to the other spaces.
896      bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace();
897      // Create the worker tasks for this space.
898      while (card_begin != card_end) {
899        // Add a range of cards.
900        size_t addr_remaining = card_end - card_begin;
901        size_t card_increment = std::min(card_delta, addr_remaining);
902        // Take from the back of the mark stack.
903        size_t mark_stack_remaining = mark_stack_end - mark_stack_begin;
904        size_t mark_stack_increment = std::min(mark_stack_delta, mark_stack_remaining);
905        mark_stack_end -= mark_stack_increment;
906        mark_stack_->PopBackCount(static_cast<int32_t>(mark_stack_increment));
907        DCHECK_EQ(mark_stack_end, mark_stack_->End());
908        // Add the new task to the thread pool.
909        auto* task = new CardScanTask(thread_pool,
910                                      this,
911                                      space->GetMarkBitmap(),
912                                      card_begin,
913                                      card_begin + card_increment,
914                                      minimum_age,
915                                      mark_stack_increment,
916                                      mark_stack_end,
917                                      clear_card);
918        thread_pool->AddTask(self, task);
919        card_begin += card_increment;
920      }
921    }
922
923    // Note: the card scan below may dirty new cards (and scan them)
924    // as a side effect when a Reference object is encountered and
925    // queued during the marking. See b/11465268.
926    thread_pool->SetMaxActiveWorkers(thread_count - 1);
927    thread_pool->StartWorkers(self);
928    thread_pool->Wait(self, true, true);
929    thread_pool->StopWorkers(self);
930  } else {
931    for (const auto& space : GetHeap()->GetContinuousSpaces()) {
932      if (space->GetMarkBitmap() != nullptr) {
933        // Image spaces are handled properly since live == marked for them.
934        const char* name = nullptr;
935        switch (space->GetGcRetentionPolicy()) {
936        case space::kGcRetentionPolicyNeverCollect:
937          name = paused ? "(Paused)ScanGrayImageSpaceObjects" : "ScanGrayImageSpaceObjects";
938          break;
939        case space::kGcRetentionPolicyFullCollect:
940          name = paused ? "(Paused)ScanGrayZygoteSpaceObjects" : "ScanGrayZygoteSpaceObjects";
941          break;
942        case space::kGcRetentionPolicyAlwaysCollect:
943          name = paused ? "(Paused)ScanGrayAllocSpaceObjects" : "ScanGrayAllocSpaceObjects";
944          break;
945        default:
946          LOG(FATAL) << "Unreachable";
947          UNREACHABLE();
948        }
949        TimingLogger::ScopedTiming t(name, GetTimings());
950        ScanObjectVisitor visitor(this);
951        bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace();
952        if (clear_card) {
953          card_table->Scan<true>(space->GetMarkBitmap(),
954                                 space->Begin(),
955                                 space->End(),
956                                 visitor,
957                                 minimum_age);
958        } else {
959          card_table->Scan<false>(space->GetMarkBitmap(),
960                                  space->Begin(),
961                                  space->End(),
962                                  visitor,
963                                  minimum_age);
964        }
965      }
966    }
967  }
968}
969
970class RecursiveMarkTask : public MarkStackTask<false> {
971 public:
972  RecursiveMarkTask(ThreadPool* thread_pool,
973                    MarkSweep* mark_sweep,
974                    accounting::ContinuousSpaceBitmap* bitmap,
975                    uintptr_t begin,
976                    uintptr_t end)
977      : MarkStackTask<false>(thread_pool, mark_sweep, 0, nullptr),
978        bitmap_(bitmap),
979        begin_(begin),
980        end_(end) {}
981
982 protected:
983  accounting::ContinuousSpaceBitmap* const bitmap_;
984  const uintptr_t begin_;
985  const uintptr_t end_;
986
987  virtual void Finalize() {
988    delete this;
989  }
990
991  // Scans all of the objects
992  virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS {
993    ScanObjectParallelVisitor visitor(this);
994    bitmap_->VisitMarkedRange(begin_, end_, visitor);
995    // Finish by emptying our local mark stack.
996    MarkStackTask::Run(self);
997  }
998};
999
1000// Populates the mark stack based on the set of marked objects and
1001// recursively marks until the mark stack is emptied.
1002void MarkSweep::RecursiveMark() {
1003  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1004  // RecursiveMark will build the lists of known instances of the Reference classes. See
1005  // DelayReferenceReferent for details.
1006  if (kUseRecursiveMark) {
1007    const bool partial = GetGcType() == kGcTypePartial;
1008    ScanObjectVisitor scan_visitor(this);
1009    auto* self = Thread::Current();
1010    ThreadPool* thread_pool = heap_->GetThreadPool();
1011    size_t thread_count = GetThreadCount(false);
1012    const bool parallel = kParallelRecursiveMark && thread_count > 1;
1013    mark_stack_->Reset();
1014    for (const auto& space : GetHeap()->GetContinuousSpaces()) {
1015      if ((space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) ||
1016          (!partial && space->GetGcRetentionPolicy() == space::kGcRetentionPolicyFullCollect)) {
1017        current_space_bitmap_ = space->GetMarkBitmap();
1018        if (current_space_bitmap_ == nullptr) {
1019          continue;
1020        }
1021        if (parallel) {
1022          // We will use the mark stack the future.
1023          // CHECK(mark_stack_->IsEmpty());
1024          // This function does not handle heap end increasing, so we must use the space end.
1025          uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin());
1026          uintptr_t end = reinterpret_cast<uintptr_t>(space->End());
1027          atomic_finger_.StoreRelaxed(AtomicInteger::MaxValue());
1028
1029          // Create a few worker tasks.
1030          const size_t n = thread_count * 2;
1031          while (begin != end) {
1032            uintptr_t start = begin;
1033            uintptr_t delta = (end - begin) / n;
1034            delta = RoundUp(delta, KB);
1035            if (delta < 16 * KB) delta = end - begin;
1036            begin += delta;
1037            auto* task = new RecursiveMarkTask(thread_pool,
1038                                               this,
1039                                               current_space_bitmap_,
1040                                               start,
1041                                               begin);
1042            thread_pool->AddTask(self, task);
1043          }
1044          thread_pool->SetMaxActiveWorkers(thread_count - 1);
1045          thread_pool->StartWorkers(self);
1046          thread_pool->Wait(self, true, true);
1047          thread_pool->StopWorkers(self);
1048        } else {
1049          // This function does not handle heap end increasing, so we must use the space end.
1050          uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin());
1051          uintptr_t end = reinterpret_cast<uintptr_t>(space->End());
1052          current_space_bitmap_->VisitMarkedRange(begin, end, scan_visitor);
1053        }
1054      }
1055    }
1056  }
1057  ProcessMarkStack(false);
1058}
1059
1060void MarkSweep::RecursiveMarkDirtyObjects(bool paused, uint8_t minimum_age) {
1061  ScanGrayObjects(paused, minimum_age);
1062  ProcessMarkStack(paused);
1063}
1064
1065void MarkSweep::ReMarkRoots() {
1066  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1067  Locks::mutator_lock_->AssertExclusiveHeld(Thread::Current());
1068  Runtime::Current()->VisitRoots(this, static_cast<VisitRootFlags>(
1069      kVisitRootFlagNewRoots | kVisitRootFlagStopLoggingNewRoots | kVisitRootFlagClearRootLog));
1070  if (kVerifyRootsMarked) {
1071    TimingLogger::ScopedTiming t2("(Paused)VerifyRoots", GetTimings());
1072    VerifyRootMarkedVisitor visitor(this);
1073    Runtime::Current()->VisitRoots(&visitor);
1074  }
1075}
1076
1077void MarkSweep::SweepSystemWeaks(Thread* self) {
1078  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1079  ReaderMutexLock mu(self, *Locks::heap_bitmap_lock_);
1080  Runtime::Current()->SweepSystemWeaks(this);
1081}
1082
1083class VerifySystemWeakVisitor : public IsMarkedVisitor {
1084 public:
1085  explicit VerifySystemWeakVisitor(MarkSweep* mark_sweep) : mark_sweep_(mark_sweep) {}
1086
1087  virtual mirror::Object* IsMarked(mirror::Object* obj)
1088      OVERRIDE
1089      SHARED_REQUIRES(Locks::mutator_lock_, Locks::heap_bitmap_lock_) {
1090    mark_sweep_->VerifyIsLive(obj);
1091    return obj;
1092  }
1093
1094  MarkSweep* const mark_sweep_;
1095};
1096
1097void MarkSweep::VerifyIsLive(const mirror::Object* obj) {
1098  if (!heap_->GetLiveBitmap()->Test(obj)) {
1099    // TODO: Consider live stack? Has this code bitrotted?
1100    CHECK(!heap_->allocation_stack_->Contains(obj))
1101        << "Found dead object " << obj << "\n" << heap_->DumpSpaces();
1102  }
1103}
1104
1105void MarkSweep::VerifySystemWeaks() {
1106  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1107  // Verify system weaks, uses a special object visitor which returns the input object.
1108  VerifySystemWeakVisitor visitor(this);
1109  Runtime::Current()->SweepSystemWeaks(&visitor);
1110}
1111
1112class CheckpointMarkThreadRoots : public Closure, public RootVisitor {
1113 public:
1114  CheckpointMarkThreadRoots(MarkSweep* mark_sweep,
1115                            bool revoke_ros_alloc_thread_local_buffers_at_checkpoint)
1116      : mark_sweep_(mark_sweep),
1117        revoke_ros_alloc_thread_local_buffers_at_checkpoint_(
1118            revoke_ros_alloc_thread_local_buffers_at_checkpoint) {
1119  }
1120
1121  void VisitRoots(mirror::Object*** roots, size_t count, const RootInfo& info ATTRIBUTE_UNUSED)
1122      OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_)
1123      REQUIRES(Locks::heap_bitmap_lock_) {
1124    for (size_t i = 0; i < count; ++i) {
1125      mark_sweep_->MarkObjectNonNullParallel(*roots[i]);
1126    }
1127  }
1128
1129  void VisitRoots(mirror::CompressedReference<mirror::Object>** roots,
1130                  size_t count,
1131                  const RootInfo& info ATTRIBUTE_UNUSED)
1132      OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_)
1133      REQUIRES(Locks::heap_bitmap_lock_) {
1134    for (size_t i = 0; i < count; ++i) {
1135      mark_sweep_->MarkObjectNonNullParallel(roots[i]->AsMirrorPtr());
1136    }
1137  }
1138
1139  virtual void Run(Thread* thread) OVERRIDE NO_THREAD_SAFETY_ANALYSIS {
1140    ScopedTrace trace("Marking thread roots");
1141    // Note: self is not necessarily equal to thread since thread may be suspended.
1142    Thread* const self = Thread::Current();
1143    CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc)
1144        << thread->GetState() << " thread " << thread << " self " << self;
1145    thread->VisitRoots(this);
1146    if (revoke_ros_alloc_thread_local_buffers_at_checkpoint_) {
1147      ScopedTrace trace2("RevokeRosAllocThreadLocalBuffers");
1148      mark_sweep_->GetHeap()->RevokeRosAllocThreadLocalBuffers(thread);
1149    }
1150    // If thread is a running mutator, then act on behalf of the garbage collector.
1151    // See the code in ThreadList::RunCheckpoint.
1152    mark_sweep_->GetBarrier().Pass(self);
1153  }
1154
1155 private:
1156  MarkSweep* const mark_sweep_;
1157  const bool revoke_ros_alloc_thread_local_buffers_at_checkpoint_;
1158};
1159
1160void MarkSweep::MarkRootsCheckpoint(Thread* self,
1161                                    bool revoke_ros_alloc_thread_local_buffers_at_checkpoint) {
1162  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1163  CheckpointMarkThreadRoots check_point(this, revoke_ros_alloc_thread_local_buffers_at_checkpoint);
1164  ThreadList* thread_list = Runtime::Current()->GetThreadList();
1165  // Request the check point is run on all threads returning a count of the threads that must
1166  // run through the barrier including self.
1167  size_t barrier_count = thread_list->RunCheckpoint(&check_point);
1168  // Release locks then wait for all mutator threads to pass the barrier.
1169  // If there are no threads to wait which implys that all the checkpoint functions are finished,
1170  // then no need to release locks.
1171  if (barrier_count == 0) {
1172    return;
1173  }
1174  Locks::heap_bitmap_lock_->ExclusiveUnlock(self);
1175  Locks::mutator_lock_->SharedUnlock(self);
1176  {
1177    ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
1178    gc_barrier_->Increment(self, barrier_count);
1179  }
1180  Locks::mutator_lock_->SharedLock(self);
1181  Locks::heap_bitmap_lock_->ExclusiveLock(self);
1182}
1183
1184void MarkSweep::SweepArray(accounting::ObjectStack* allocations, bool swap_bitmaps) {
1185  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1186  Thread* self = Thread::Current();
1187  mirror::Object** chunk_free_buffer = reinterpret_cast<mirror::Object**>(
1188      sweep_array_free_buffer_mem_map_->BaseBegin());
1189  size_t chunk_free_pos = 0;
1190  ObjectBytePair freed;
1191  ObjectBytePair freed_los;
1192  // How many objects are left in the array, modified after each space is swept.
1193  StackReference<mirror::Object>* objects = allocations->Begin();
1194  size_t count = allocations->Size();
1195  // Change the order to ensure that the non-moving space last swept as an optimization.
1196  std::vector<space::ContinuousSpace*> sweep_spaces;
1197  space::ContinuousSpace* non_moving_space = nullptr;
1198  for (space::ContinuousSpace* space : heap_->GetContinuousSpaces()) {
1199    if (space->IsAllocSpace() &&
1200        !immune_spaces_.ContainsSpace(space) &&
1201        space->GetLiveBitmap() != nullptr) {
1202      if (space == heap_->GetNonMovingSpace()) {
1203        non_moving_space = space;
1204      } else {
1205        sweep_spaces.push_back(space);
1206      }
1207    }
1208  }
1209  // Unlikely to sweep a significant amount of non_movable objects, so we do these after the after
1210  // the other alloc spaces as an optimization.
1211  if (non_moving_space != nullptr) {
1212    sweep_spaces.push_back(non_moving_space);
1213  }
1214  // Start by sweeping the continuous spaces.
1215  for (space::ContinuousSpace* space : sweep_spaces) {
1216    space::AllocSpace* alloc_space = space->AsAllocSpace();
1217    accounting::ContinuousSpaceBitmap* live_bitmap = space->GetLiveBitmap();
1218    accounting::ContinuousSpaceBitmap* mark_bitmap = space->GetMarkBitmap();
1219    if (swap_bitmaps) {
1220      std::swap(live_bitmap, mark_bitmap);
1221    }
1222    StackReference<mirror::Object>* out = objects;
1223    for (size_t i = 0; i < count; ++i) {
1224      mirror::Object* const obj = objects[i].AsMirrorPtr();
1225      if (kUseThreadLocalAllocationStack && obj == nullptr) {
1226        continue;
1227      }
1228      if (space->HasAddress(obj)) {
1229        // This object is in the space, remove it from the array and add it to the sweep buffer
1230        // if needed.
1231        if (!mark_bitmap->Test(obj)) {
1232          if (chunk_free_pos >= kSweepArrayChunkFreeSize) {
1233            TimingLogger::ScopedTiming t2("FreeList", GetTimings());
1234            freed.objects += chunk_free_pos;
1235            freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer);
1236            chunk_free_pos = 0;
1237          }
1238          chunk_free_buffer[chunk_free_pos++] = obj;
1239        }
1240      } else {
1241        (out++)->Assign(obj);
1242      }
1243    }
1244    if (chunk_free_pos > 0) {
1245      TimingLogger::ScopedTiming t2("FreeList", GetTimings());
1246      freed.objects += chunk_free_pos;
1247      freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer);
1248      chunk_free_pos = 0;
1249    }
1250    // All of the references which space contained are no longer in the allocation stack, update
1251    // the count.
1252    count = out - objects;
1253  }
1254  // Handle the large object space.
1255  space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace();
1256  if (large_object_space != nullptr) {
1257    accounting::LargeObjectBitmap* large_live_objects = large_object_space->GetLiveBitmap();
1258    accounting::LargeObjectBitmap* large_mark_objects = large_object_space->GetMarkBitmap();
1259    if (swap_bitmaps) {
1260      std::swap(large_live_objects, large_mark_objects);
1261    }
1262    for (size_t i = 0; i < count; ++i) {
1263      mirror::Object* const obj = objects[i].AsMirrorPtr();
1264      // Handle large objects.
1265      if (kUseThreadLocalAllocationStack && obj == nullptr) {
1266        continue;
1267      }
1268      if (!large_mark_objects->Test(obj)) {
1269        ++freed_los.objects;
1270        freed_los.bytes += large_object_space->Free(self, obj);
1271      }
1272    }
1273  }
1274  {
1275    TimingLogger::ScopedTiming t2("RecordFree", GetTimings());
1276    RecordFree(freed);
1277    RecordFreeLOS(freed_los);
1278    t2.NewTiming("ResetStack");
1279    allocations->Reset();
1280  }
1281  sweep_array_free_buffer_mem_map_->MadviseDontNeedAndZero();
1282}
1283
1284void MarkSweep::Sweep(bool swap_bitmaps) {
1285  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1286  // Ensure that nobody inserted items in the live stack after we swapped the stacks.
1287  CHECK_GE(live_stack_freeze_size_, GetHeap()->GetLiveStack()->Size());
1288  {
1289    TimingLogger::ScopedTiming t2("MarkAllocStackAsLive", GetTimings());
1290    // Mark everything allocated since the last as GC live so that we can sweep concurrently,
1291    // knowing that new allocations won't be marked as live.
1292    accounting::ObjectStack* live_stack = heap_->GetLiveStack();
1293    heap_->MarkAllocStackAsLive(live_stack);
1294    live_stack->Reset();
1295    DCHECK(mark_stack_->IsEmpty());
1296  }
1297  for (const auto& space : GetHeap()->GetContinuousSpaces()) {
1298    if (space->IsContinuousMemMapAllocSpace()) {
1299      space::ContinuousMemMapAllocSpace* alloc_space = space->AsContinuousMemMapAllocSpace();
1300      TimingLogger::ScopedTiming split(
1301          alloc_space->IsZygoteSpace() ? "SweepZygoteSpace" : "SweepMallocSpace",
1302          GetTimings());
1303      RecordFree(alloc_space->Sweep(swap_bitmaps));
1304    }
1305  }
1306  SweepLargeObjects(swap_bitmaps);
1307}
1308
1309void MarkSweep::SweepLargeObjects(bool swap_bitmaps) {
1310  space::LargeObjectSpace* los = heap_->GetLargeObjectsSpace();
1311  if (los != nullptr) {
1312    TimingLogger::ScopedTiming split(__FUNCTION__, GetTimings());
1313    RecordFreeLOS(los->Sweep(swap_bitmaps));
1314  }
1315}
1316
1317// Process the "referent" field in a java.lang.ref.Reference.  If the referent has not yet been
1318// marked, put it on the appropriate list in the heap for later processing.
1319void MarkSweep::DelayReferenceReferent(mirror::Class* klass, mirror::Reference* ref) {
1320  heap_->GetReferenceProcessor()->DelayReferenceReferent(klass, ref, this);
1321}
1322
1323class MarkVisitor {
1324 public:
1325  ALWAYS_INLINE explicit MarkVisitor(MarkSweep* const mark_sweep) : mark_sweep_(mark_sweep) {}
1326
1327  ALWAYS_INLINE void operator()(mirror::Object* obj,
1328                                MemberOffset offset,
1329                                bool is_static ATTRIBUTE_UNUSED) const
1330      REQUIRES(Locks::heap_bitmap_lock_)
1331      SHARED_REQUIRES(Locks::mutator_lock_) {
1332    if (kCheckLocks) {
1333      Locks::mutator_lock_->AssertSharedHeld(Thread::Current());
1334      Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current());
1335    }
1336    mark_sweep_->MarkObject(obj->GetFieldObject<mirror::Object>(offset), obj, offset);
1337  }
1338
1339  void VisitRootIfNonNull(mirror::CompressedReference<mirror::Object>* root) const
1340      REQUIRES(Locks::heap_bitmap_lock_)
1341      SHARED_REQUIRES(Locks::mutator_lock_) {
1342    if (!root->IsNull()) {
1343      VisitRoot(root);
1344    }
1345  }
1346
1347  void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const
1348      REQUIRES(Locks::heap_bitmap_lock_)
1349      SHARED_REQUIRES(Locks::mutator_lock_) {
1350    if (kCheckLocks) {
1351      Locks::mutator_lock_->AssertSharedHeld(Thread::Current());
1352      Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current());
1353    }
1354    mark_sweep_->MarkObject(root->AsMirrorPtr());
1355  }
1356
1357 private:
1358  MarkSweep* const mark_sweep_;
1359};
1360
1361// Scans an object reference.  Determines the type of the reference
1362// and dispatches to a specialized scanning routine.
1363void MarkSweep::ScanObject(mirror::Object* obj) {
1364  MarkVisitor mark_visitor(this);
1365  DelayReferenceReferentVisitor ref_visitor(this);
1366  ScanObjectVisit(obj, mark_visitor, ref_visitor);
1367}
1368
1369void MarkSweep::ProcessMarkStackParallel(size_t thread_count) {
1370  Thread* self = Thread::Current();
1371  ThreadPool* thread_pool = GetHeap()->GetThreadPool();
1372  const size_t chunk_size = std::min(mark_stack_->Size() / thread_count + 1,
1373                                     static_cast<size_t>(MarkStackTask<false>::kMaxSize));
1374  CHECK_GT(chunk_size, 0U);
1375  // Split the current mark stack up into work tasks.
1376  for (auto* it = mark_stack_->Begin(), *end = mark_stack_->End(); it < end; ) {
1377    const size_t delta = std::min(static_cast<size_t>(end - it), chunk_size);
1378    thread_pool->AddTask(self, new MarkStackTask<false>(thread_pool, this, delta, it));
1379    it += delta;
1380  }
1381  thread_pool->SetMaxActiveWorkers(thread_count - 1);
1382  thread_pool->StartWorkers(self);
1383  thread_pool->Wait(self, true, true);
1384  thread_pool->StopWorkers(self);
1385  mark_stack_->Reset();
1386  CHECK_EQ(work_chunks_created_.LoadSequentiallyConsistent(),
1387           work_chunks_deleted_.LoadSequentiallyConsistent())
1388      << " some of the work chunks were leaked";
1389}
1390
1391// Scan anything that's on the mark stack.
1392void MarkSweep::ProcessMarkStack(bool paused) {
1393  TimingLogger::ScopedTiming t(paused ? "(Paused)ProcessMarkStack" : __FUNCTION__, GetTimings());
1394  size_t thread_count = GetThreadCount(paused);
1395  if (kParallelProcessMarkStack && thread_count > 1 &&
1396      mark_stack_->Size() >= kMinimumParallelMarkStackSize) {
1397    ProcessMarkStackParallel(thread_count);
1398  } else {
1399    // TODO: Tune this.
1400    static const size_t kFifoSize = 4;
1401    BoundedFifoPowerOfTwo<mirror::Object*, kFifoSize> prefetch_fifo;
1402    for (;;) {
1403      mirror::Object* obj = nullptr;
1404      if (kUseMarkStackPrefetch) {
1405        while (!mark_stack_->IsEmpty() && prefetch_fifo.size() < kFifoSize) {
1406          mirror::Object* mark_stack_obj = mark_stack_->PopBack();
1407          DCHECK(mark_stack_obj != nullptr);
1408          __builtin_prefetch(mark_stack_obj);
1409          prefetch_fifo.push_back(mark_stack_obj);
1410        }
1411        if (prefetch_fifo.empty()) {
1412          break;
1413        }
1414        obj = prefetch_fifo.front();
1415        prefetch_fifo.pop_front();
1416      } else {
1417        if (mark_stack_->IsEmpty()) {
1418          break;
1419        }
1420        obj = mark_stack_->PopBack();
1421      }
1422      DCHECK(obj != nullptr);
1423      ScanObject(obj);
1424    }
1425  }
1426}
1427
1428inline mirror::Object* MarkSweep::IsMarked(mirror::Object* object) {
1429  if (immune_spaces_.IsInImmuneRegion(object)) {
1430    return object;
1431  }
1432  if (current_space_bitmap_->HasAddress(object)) {
1433    return current_space_bitmap_->Test(object) ? object : nullptr;
1434  }
1435  return mark_bitmap_->Test(object) ? object : nullptr;
1436}
1437
1438void MarkSweep::FinishPhase() {
1439  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1440  if (kCountScannedTypes) {
1441    VLOG(gc)
1442        << "MarkSweep scanned"
1443        << " no reference objects=" << no_reference_class_count_.LoadRelaxed()
1444        << " normal objects=" << normal_count_.LoadRelaxed()
1445        << " classes=" << class_count_.LoadRelaxed()
1446        << " object arrays=" << object_array_count_.LoadRelaxed()
1447        << " references=" << reference_count_.LoadRelaxed()
1448        << " other=" << other_count_.LoadRelaxed();
1449  }
1450  if (kCountTasks) {
1451    VLOG(gc) << "Total number of work chunks allocated: " << work_chunks_created_.LoadRelaxed();
1452  }
1453  if (kMeasureOverhead) {
1454    VLOG(gc) << "Overhead time " << PrettyDuration(overhead_time_.LoadRelaxed());
1455  }
1456  if (kProfileLargeObjects) {
1457    VLOG(gc) << "Large objects tested " << large_object_test_.LoadRelaxed()
1458        << " marked " << large_object_mark_.LoadRelaxed();
1459  }
1460  if (kCountMarkedObjects) {
1461    VLOG(gc) << "Marked: null=" << mark_null_count_.LoadRelaxed()
1462        << " immune=" <<  mark_immune_count_.LoadRelaxed()
1463        << " fastpath=" << mark_fastpath_count_.LoadRelaxed()
1464        << " slowpath=" << mark_slowpath_count_.LoadRelaxed();
1465  }
1466  CHECK(mark_stack_->IsEmpty());  // Ensure that the mark stack is empty.
1467  mark_stack_->Reset();
1468  Thread* const self = Thread::Current();
1469  ReaderMutexLock mu(self, *Locks::mutator_lock_);
1470  WriterMutexLock mu2(self, *Locks::heap_bitmap_lock_);
1471  heap_->ClearMarkedObjects();
1472}
1473
1474void MarkSweep::RevokeAllThreadLocalBuffers() {
1475  if (kRevokeRosAllocThreadLocalBuffersAtCheckpoint && IsConcurrent()) {
1476    // If concurrent, rosalloc thread-local buffers are revoked at the
1477    // thread checkpoint. Bump pointer space thread-local buffers must
1478    // not be in use.
1479    GetHeap()->AssertAllBumpPointerSpaceThreadLocalBuffersAreRevoked();
1480  } else {
1481    TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1482    GetHeap()->RevokeAllThreadLocalBuffers();
1483  }
1484}
1485
1486}  // namespace collector
1487}  // namespace gc
1488}  // namespace art
1489