mark_sweep.cc revision 3d21bdf8894e780d349c481e5c9e29fe1556051c 
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#define ATRACE_TAG ATRACE_TAG_DALVIK
26#include "cutils/trace.h"
27
28#include "base/bounded_fifo.h"
29#include "base/logging.h"
30#include "base/macros.h"
31#include "base/mutex-inl.h"
32#include "base/time_utils.h"
33#include "base/timing_logger.h"
34#include "gc/accounting/card_table-inl.h"
35#include "gc/accounting/heap_bitmap-inl.h"
36#include "gc/accounting/mod_union_table.h"
37#include "gc/accounting/space_bitmap-inl.h"
38#include "gc/heap.h"
39#include "gc/reference_processor.h"
40#include "gc/space/image_space.h"
41#include "gc/space/large_object_space.h"
42#include "gc/space/space-inl.h"
43#include "mark_sweep-inl.h"
44#include "mirror/object-inl.h"
45#include "runtime.h"
46#include "scoped_thread_state_change.h"
47#include "thread-inl.h"
48#include "thread_list.h"
49
50using ::art::mirror::Object;
51
52namespace art {
53namespace gc {
54namespace collector {
55
56// Performance options.
57static constexpr bool kUseRecursiveMark = false;
58static constexpr bool kUseMarkStackPrefetch = true;
59static constexpr size_t kSweepArrayChunkFreeSize = 1024;
60static constexpr bool kPreCleanCards = true;
61
62// Parallelism options.
63static constexpr bool kParallelCardScan = true;
64static constexpr bool kParallelRecursiveMark = true;
65// Don't attempt to parallelize mark stack processing unless the mark stack is at least n
66// elements. This is temporary until we reduce the overhead caused by allocating tasks, etc.. Not
67// having this can add overhead in ProcessReferences since we may end up doing many calls of
68// ProcessMarkStack with very small mark stacks.
69static constexpr size_t kMinimumParallelMarkStackSize = 128;
70static constexpr bool kParallelProcessMarkStack = true;
71
72// Profiling and information flags.
73static constexpr bool kProfileLargeObjects = false;
74static constexpr bool kMeasureOverhead = false;
75static constexpr bool kCountTasks = false;
76static constexpr bool kCountJavaLangRefs = false;
77static constexpr bool kCountMarkedObjects = false;
78
79// Turn off kCheckLocks when profiling the GC since it slows the GC down by up to 40%.
80static constexpr bool kCheckLocks = kDebugLocking;
81static constexpr bool kVerifyRootsMarked = kIsDebugBuild;
82
83// If true, revoke the rosalloc thread-local buffers at the
84// checkpoint, as opposed to during the pause.
85static constexpr bool kRevokeRosAllocThreadLocalBuffersAtCheckpoint = true;
86
87void MarkSweep::BindBitmaps() {
88  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
89  WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_);
90  // Mark all of the spaces we never collect as immune.
91  for (const auto& space : GetHeap()->GetContinuousSpaces()) {
92    if (space->GetGcRetentionPolicy() == space::kGcRetentionPolicyNeverCollect) {
93      CHECK(immune_region_.AddContinuousSpace(space)) << "Failed to add space " << *space;
94    }
95  }
96}
97
98MarkSweep::MarkSweep(Heap* heap, bool is_concurrent, const std::string& name_prefix)
99    : GarbageCollector(heap,
100                       name_prefix +
101                       (is_concurrent ? "concurrent mark sweep": "mark sweep")),
102      current_space_bitmap_(nullptr), mark_bitmap_(nullptr), mark_stack_(nullptr),
103      gc_barrier_(new Barrier(0)),
104      mark_stack_lock_("mark sweep mark stack lock", kMarkSweepMarkStackLock),
105      is_concurrent_(is_concurrent), live_stack_freeze_size_(0) {
106  std::string error_msg;
107  MemMap* mem_map = MemMap::MapAnonymous(
108      "mark sweep sweep array free buffer", nullptr,
109      RoundUp(kSweepArrayChunkFreeSize * sizeof(mirror::Object*), kPageSize),
110      PROT_READ | PROT_WRITE, false, false, &error_msg);
111  CHECK(mem_map != nullptr) << "Couldn't allocate sweep array free buffer: " << error_msg;
112  sweep_array_free_buffer_mem_map_.reset(mem_map);
113}
114
115void MarkSweep::InitializePhase() {
116  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
117  mark_stack_ = heap_->GetMarkStack();
118  DCHECK(mark_stack_ != nullptr);
119  immune_region_.Reset();
120  class_count_.StoreRelaxed(0);
121  array_count_.StoreRelaxed(0);
122  other_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  reference_count_.StoreRelaxed(0);
129  mark_null_count_.StoreRelaxed(0);
130  mark_immune_count_.StoreRelaxed(0);
131  mark_fastpath_count_.StoreRelaxed(0);
132  mark_slowpath_count_.StoreRelaxed(0);
133  {
134    // TODO: I don't think we should need heap bitmap lock to Get the mark bitmap.
135    ReaderMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_);
136    mark_bitmap_ = heap_->GetMarkBitmap();
137  }
138  if (!GetCurrentIteration()->GetClearSoftReferences()) {
139    // Always clear soft references if a non-sticky collection.
140    GetCurrentIteration()->SetClearSoftReferences(GetGcType() != collector::kGcTypeSticky);
141  }
142}
143
144void MarkSweep::RunPhases() {
145  Thread* self = Thread::Current();
146  InitializePhase();
147  Locks::mutator_lock_->AssertNotHeld(self);
148  if (IsConcurrent()) {
149    GetHeap()->PreGcVerification(this);
150    {
151      ReaderMutexLock mu(self, *Locks::mutator_lock_);
152      MarkingPhase();
153    }
154    ScopedPause pause(this);
155    GetHeap()->PrePauseRosAllocVerification(this);
156    PausePhase();
157    RevokeAllThreadLocalBuffers();
158  } else {
159    ScopedPause pause(this);
160    GetHeap()->PreGcVerificationPaused(this);
161    MarkingPhase();
162    GetHeap()->PrePauseRosAllocVerification(this);
163    PausePhase();
164    RevokeAllThreadLocalBuffers();
165  }
166  {
167    // Sweeping always done concurrently, even for non concurrent mark sweep.
168    ReaderMutexLock mu(self, *Locks::mutator_lock_);
169    ReclaimPhase();
170  }
171  GetHeap()->PostGcVerification(this);
172  FinishPhase();
173}
174
175void MarkSweep::ProcessReferences(Thread* self) {
176  WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
177  GetHeap()->GetReferenceProcessor()->ProcessReferences(
178      true, GetTimings(), GetCurrentIteration()->GetClearSoftReferences(),
179      &HeapReferenceMarkedCallback, &MarkObjectCallback, &ProcessMarkStackCallback, this);
180}
181
182void MarkSweep::PausePhase() {
183  TimingLogger::ScopedTiming t("(Paused)PausePhase", GetTimings());
184  Thread* self = Thread::Current();
185  Locks::mutator_lock_->AssertExclusiveHeld(self);
186  if (IsConcurrent()) {
187    // Handle the dirty objects if we are a concurrent GC.
188    WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
189    // Re-mark root set.
190    ReMarkRoots();
191    // Scan dirty objects, this is only required if we are not doing concurrent GC.
192    RecursiveMarkDirtyObjects(true, accounting::CardTable::kCardDirty);
193  }
194  {
195    TimingLogger::ScopedTiming t2("SwapStacks", GetTimings());
196    WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
197    heap_->SwapStacks(self);
198    live_stack_freeze_size_ = heap_->GetLiveStack()->Size();
199    // Need to revoke all the thread local allocation stacks since we just swapped the allocation
200    // stacks and don't want anybody to allocate into the live stack.
201    RevokeAllThreadLocalAllocationStacks(self);
202  }
203  heap_->PreSweepingGcVerification(this);
204  // Disallow new system weaks to prevent a race which occurs when someone adds a new system
205  // weak before we sweep them. Since this new system weak may not be marked, the GC may
206  // incorrectly sweep it. This also fixes a race where interning may attempt to return a strong
207  // reference to a string that is about to be swept.
208  Runtime::Current()->DisallowNewSystemWeaks();
209  // Enable the reference processing slow path, needs to be done with mutators paused since there
210  // is no lock in the GetReferent fast path.
211  GetHeap()->GetReferenceProcessor()->EnableSlowPath();
212}
213
214void MarkSweep::PreCleanCards() {
215  // Don't do this for non concurrent GCs since they don't have any dirty cards.
216  if (kPreCleanCards && IsConcurrent()) {
217    TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
218    Thread* self = Thread::Current();
219    CHECK(!Locks::mutator_lock_->IsExclusiveHeld(self));
220    // Process dirty cards and add dirty cards to mod union tables, also ages cards.
221    heap_->ProcessCards(GetTimings(), false, true, false);
222    // The checkpoint root marking is required to avoid a race condition which occurs if the
223    // following happens during a reference write:
224    // 1. mutator dirties the card (write barrier)
225    // 2. GC ages the card (the above ProcessCards call)
226    // 3. GC scans the object (the RecursiveMarkDirtyObjects call below)
227    // 4. mutator writes the value (corresponding to the write barrier in 1.)
228    // This causes the GC to age the card but not necessarily mark the reference which the mutator
229    // wrote into the object stored in the card.
230    // Having the checkpoint fixes this issue since it ensures that the card mark and the
231    // reference write are visible to the GC before the card is scanned (this is due to locks being
232    // acquired / released in the checkpoint code).
233    // The other roots are also marked to help reduce the pause.
234    MarkRootsCheckpoint(self, false);
235    MarkNonThreadRoots();
236    MarkConcurrentRoots(
237        static_cast<VisitRootFlags>(kVisitRootFlagClearRootLog | kVisitRootFlagNewRoots));
238    // Process the newly aged cards.
239    RecursiveMarkDirtyObjects(false, accounting::CardTable::kCardDirty - 1);
240    // TODO: Empty allocation stack to reduce the number of objects we need to test / mark as live
241    // in the next GC.
242  }
243}
244
245void MarkSweep::RevokeAllThreadLocalAllocationStacks(Thread* self) {
246  if (kUseThreadLocalAllocationStack) {
247    TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
248    Locks::mutator_lock_->AssertExclusiveHeld(self);
249    heap_->RevokeAllThreadLocalAllocationStacks(self);
250  }
251}
252
253void MarkSweep::MarkingPhase() {
254  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
255  Thread* self = Thread::Current();
256  BindBitmaps();
257  FindDefaultSpaceBitmap();
258  // Process dirty cards and add dirty cards to mod union tables.
259  // If the GC type is non sticky, then we just clear the cards instead of ageing them.
260  heap_->ProcessCards(GetTimings(), false, true, GetGcType() != kGcTypeSticky);
261  WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
262  MarkRoots(self);
263  MarkReachableObjects();
264  // Pre-clean dirtied cards to reduce pauses.
265  PreCleanCards();
266}
267
268void MarkSweep::UpdateAndMarkModUnion() {
269  for (const auto& space : heap_->GetContinuousSpaces()) {
270    if (immune_region_.ContainsSpace(space)) {
271      const char* name = space->IsZygoteSpace() ? "UpdateAndMarkZygoteModUnionTable" :
272          "UpdateAndMarkImageModUnionTable";
273      TimingLogger::ScopedTiming t(name, GetTimings());
274      accounting::ModUnionTable* mod_union_table = heap_->FindModUnionTableFromSpace(space);
275      CHECK(mod_union_table != nullptr);
276      mod_union_table->UpdateAndMarkReferences(MarkHeapReferenceCallback, this);
277    }
278  }
279}
280
281void MarkSweep::MarkReachableObjects() {
282  UpdateAndMarkModUnion();
283  // Recursively mark all the non-image bits set in the mark bitmap.
284  RecursiveMark();
285}
286
287void MarkSweep::ReclaimPhase() {
288  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
289  Thread* self = Thread::Current();
290  // Process the references concurrently.
291  ProcessReferences(self);
292  SweepSystemWeaks(self);
293  Runtime::Current()->AllowNewSystemWeaks();
294  {
295    WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
296    GetHeap()->RecordFreeRevoke();
297    // Reclaim unmarked objects.
298    Sweep(false);
299    // Swap the live and mark bitmaps for each space which we modified space. This is an
300    // optimization that enables us to not clear live bits inside of the sweep. Only swaps unbound
301    // bitmaps.
302    SwapBitmaps();
303    // Unbind the live and mark bitmaps.
304    GetHeap()->UnBindBitmaps();
305  }
306}
307
308void MarkSweep::FindDefaultSpaceBitmap() {
309  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
310  for (const auto& space : GetHeap()->GetContinuousSpaces()) {
311    accounting::ContinuousSpaceBitmap* bitmap = space->GetMarkBitmap();
312    // We want to have the main space instead of non moving if possible.
313    if (bitmap != nullptr &&
314        space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) {
315      current_space_bitmap_ = bitmap;
316      // If we are not the non moving space exit the loop early since this will be good enough.
317      if (space != heap_->GetNonMovingSpace()) {
318        break;
319      }
320    }
321  }
322  CHECK(current_space_bitmap_ != nullptr) << "Could not find a default mark bitmap\n"
323      << heap_->DumpSpaces();
324}
325
326void MarkSweep::ExpandMarkStack() {
327  ResizeMarkStack(mark_stack_->Capacity() * 2);
328}
329
330void MarkSweep::ResizeMarkStack(size_t new_size) {
331  // Rare case, no need to have Thread::Current be a parameter.
332  if (UNLIKELY(mark_stack_->Size() < mark_stack_->Capacity())) {
333    // Someone else acquired the lock and expanded the mark stack before us.
334    return;
335  }
336  std::vector<StackReference<Object>> temp(mark_stack_->Begin(), mark_stack_->End());
337  CHECK_LE(mark_stack_->Size(), new_size);
338  mark_stack_->Resize(new_size);
339  for (auto& obj : temp) {
340    mark_stack_->PushBack(obj.AsMirrorPtr());
341  }
342}
343
344inline void MarkSweep::MarkObjectNonNullParallel(Object* obj) {
345  DCHECK(obj != nullptr);
346  if (MarkObjectParallel(obj)) {
347    MutexLock mu(Thread::Current(), mark_stack_lock_);
348    if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) {
349      ExpandMarkStack();
350    }
351    // The object must be pushed on to the mark stack.
352    mark_stack_->PushBack(obj);
353  }
354}
355
356mirror::Object* MarkSweep::MarkObjectCallback(mirror::Object* obj, void* arg) {
357  MarkSweep* mark_sweep = reinterpret_cast<MarkSweep*>(arg);
358  mark_sweep->MarkObject(obj);
359  return obj;
360}
361
362void MarkSweep::MarkHeapReferenceCallback(mirror::HeapReference<mirror::Object>* ref, void* arg) {
363  reinterpret_cast<MarkSweep*>(arg)->MarkObject(ref->AsMirrorPtr());
364}
365
366bool MarkSweep::HeapReferenceMarkedCallback(mirror::HeapReference<mirror::Object>* ref, void* arg) {
367  return reinterpret_cast<MarkSweep*>(arg)->IsMarked(ref->AsMirrorPtr());
368}
369
370class MarkSweepMarkObjectSlowPath {
371 public:
372  explicit MarkSweepMarkObjectSlowPath(MarkSweep* mark_sweep, Object* holder = nullptr,
373                                       MemberOffset offset = MemberOffset(0))
374      : mark_sweep_(mark_sweep), holder_(holder), offset_(offset) {
375  }
376
377  void operator()(const Object* obj) const ALWAYS_INLINE NO_THREAD_SAFETY_ANALYSIS {
378    if (kProfileLargeObjects) {
379      // TODO: Differentiate between marking and testing somehow.
380      ++mark_sweep_->large_object_test_;
381      ++mark_sweep_->large_object_mark_;
382    }
383    space::LargeObjectSpace* large_object_space = mark_sweep_->GetHeap()->GetLargeObjectsSpace();
384    if (UNLIKELY(obj == nullptr || !IsAligned<kPageSize>(obj) ||
385                 (kIsDebugBuild && large_object_space != nullptr &&
386                     !large_object_space->Contains(obj)))) {
387      LOG(INTERNAL_FATAL) << "Tried to mark " << obj << " not contained by any spaces";
388      if (holder_ != nullptr) {
389        size_t holder_size = holder_->SizeOf();
390        ArtField* field = holder_->FindFieldByOffset(offset_);
391        LOG(INTERNAL_FATAL) << "Field info: "
392                            << " holder=" << holder_
393                            << " holder is "
394                            << (mark_sweep_->GetHeap()->IsLiveObjectLocked(holder_)
395                                ? "alive" : "dead")
396                            << " holder_size=" << holder_size
397                            << " holder_type=" << PrettyTypeOf(holder_)
398                            << " offset=" << offset_.Uint32Value()
399                            << " field=" << (field != nullptr ? field->GetName() : "nullptr")
400                            << " field_type="
401                            << (field != nullptr ? field->GetTypeDescriptor() : "")
402                            << " first_ref_field_offset="
403                            << (holder_->IsClass()
404                                ? holder_->AsClass()->GetFirstReferenceStaticFieldOffset(
405                                    sizeof(void*))
406                                : holder_->GetClass()->GetFirstReferenceInstanceFieldOffset())
407                            << " num_of_ref_fields="
408                            << (holder_->IsClass()
409                                ? holder_->AsClass()->NumReferenceStaticFields()
410                                : holder_->GetClass()->NumReferenceInstanceFields())
411                            << "\n";
412        // Print the memory content of the holder.
413        for (size_t i = 0; i < holder_size / sizeof(uint32_t); ++i) {
414          uint32_t* p = reinterpret_cast<uint32_t*>(holder_);
415          LOG(INTERNAL_FATAL) << &p[i] << ": " << "holder+" << (i * sizeof(uint32_t)) << " = "
416                              << std::hex << p[i];
417        }
418      }
419      PrintFileToLog("/proc/self/maps", LogSeverity::INTERNAL_FATAL);
420      MemMap::DumpMaps(LOG(INTERNAL_FATAL), true);
421      {
422        LOG(INTERNAL_FATAL) << "Attempting see if it's a bad root";
423        Thread* self = Thread::Current();
424        if (Locks::mutator_lock_->IsExclusiveHeld(self)) {
425          mark_sweep_->VerifyRoots();
426        } else {
427          const bool heap_bitmap_exclusive_locked =
428              Locks::heap_bitmap_lock_->IsExclusiveHeld(self);
429          if (heap_bitmap_exclusive_locked) {
430            Locks::heap_bitmap_lock_->ExclusiveUnlock(self);
431          }
432          Locks::mutator_lock_->SharedUnlock(self);
433          ThreadList* tl = Runtime::Current()->GetThreadList();
434          tl->SuspendAll(__FUNCTION__);
435          mark_sweep_->VerifyRoots();
436          tl->ResumeAll();
437          Locks::mutator_lock_->SharedLock(self);
438          if (heap_bitmap_exclusive_locked) {
439            Locks::heap_bitmap_lock_->ExclusiveLock(self);
440          }
441        }
442      }
443      LOG(FATAL) << "Can't mark invalid object";
444    }
445  }
446
447 private:
448  MarkSweep* const mark_sweep_;
449  mirror::Object* const holder_;
450  MemberOffset offset_;
451};
452
453inline void MarkSweep::MarkObjectNonNull(Object* obj, Object* holder, MemberOffset offset) {
454  DCHECK(obj != nullptr);
455  if (kUseBakerOrBrooksReadBarrier) {
456    // Verify all the objects have the correct pointer installed.
457    obj->AssertReadBarrierPointer();
458  }
459  if (immune_region_.ContainsObject(obj)) {
460    if (kCountMarkedObjects) {
461      ++mark_immune_count_;
462    }
463    DCHECK(mark_bitmap_->Test(obj));
464  } else if (LIKELY(current_space_bitmap_->HasAddress(obj))) {
465    if (kCountMarkedObjects) {
466      ++mark_fastpath_count_;
467    }
468    if (UNLIKELY(!current_space_bitmap_->Set(obj))) {
469      PushOnMarkStack(obj);  // This object was not previously marked.
470    }
471  } else {
472    if (kCountMarkedObjects) {
473      ++mark_slowpath_count_;
474    }
475    MarkSweepMarkObjectSlowPath visitor(this, holder, offset);
476    // TODO: We already know that the object is not in the current_space_bitmap_ but MarkBitmap::Set
477    // will check again.
478    if (!mark_bitmap_->Set(obj, visitor)) {
479      PushOnMarkStack(obj);  // Was not already marked, push.
480    }
481  }
482}
483
484inline void MarkSweep::PushOnMarkStack(Object* obj) {
485  if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) {
486    // Lock is not needed but is here anyways to please annotalysis.
487    MutexLock mu(Thread::Current(), mark_stack_lock_);
488    ExpandMarkStack();
489  }
490  // The object must be pushed on to the mark stack.
491  mark_stack_->PushBack(obj);
492}
493
494inline bool MarkSweep::MarkObjectParallel(const Object* obj) {
495  DCHECK(obj != nullptr);
496  if (kUseBakerOrBrooksReadBarrier) {
497    // Verify all the objects have the correct pointer installed.
498    obj->AssertReadBarrierPointer();
499  }
500  if (immune_region_.ContainsObject(obj)) {
501    DCHECK(IsMarked(obj));
502    return false;
503  }
504  // Try to take advantage of locality of references within a space, failing this find the space
505  // the hard way.
506  accounting::ContinuousSpaceBitmap* object_bitmap = current_space_bitmap_;
507  if (LIKELY(object_bitmap->HasAddress(obj))) {
508    return !object_bitmap->AtomicTestAndSet(obj);
509  }
510  MarkSweepMarkObjectSlowPath visitor(this);
511  return !mark_bitmap_->AtomicTestAndSet(obj, visitor);
512}
513
514// Used to mark objects when processing the mark stack. If an object is null, it is not marked.
515inline void MarkSweep::MarkObject(Object* obj, Object* holder, MemberOffset offset) {
516  if (obj != nullptr) {
517    MarkObjectNonNull(obj, holder, offset);
518  } else if (kCountMarkedObjects) {
519    ++mark_null_count_;
520  }
521}
522
523class VerifyRootMarkedVisitor : public SingleRootVisitor {
524 public:
525  explicit VerifyRootMarkedVisitor(MarkSweep* collector) : collector_(collector) { }
526
527  void VisitRoot(mirror::Object* root, const RootInfo& info) OVERRIDE
528      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_, Locks::heap_bitmap_lock_) {
529    CHECK(collector_->IsMarked(root)) << info.ToString();
530  }
531
532 private:
533  MarkSweep* const collector_;
534};
535
536void MarkSweep::VisitRoots(mirror::Object*** roots, size_t count,
537                           const RootInfo& info ATTRIBUTE_UNUSED) {
538  for (size_t i = 0; i < count; ++i) {
539    MarkObjectNonNull(*roots[i]);
540  }
541}
542
543void MarkSweep::VisitRoots(mirror::CompressedReference<mirror::Object>** roots, size_t count,
544                           const RootInfo& info ATTRIBUTE_UNUSED) {
545  for (size_t i = 0; i < count; ++i) {
546    MarkObjectNonNull(roots[i]->AsMirrorPtr());
547  }
548}
549
550class VerifyRootVisitor : public SingleRootVisitor {
551 public:
552  void VisitRoot(mirror::Object* root, const RootInfo& info) OVERRIDE
553      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_, Locks::heap_bitmap_lock_) {
554    // See if the root is on any space bitmap.
555    auto* heap = Runtime::Current()->GetHeap();
556    if (heap->GetLiveBitmap()->GetContinuousSpaceBitmap(root) == nullptr) {
557      space::LargeObjectSpace* large_object_space = heap->GetLargeObjectsSpace();
558      if (large_object_space != nullptr && !large_object_space->Contains(root)) {
559        LOG(INTERNAL_FATAL) << "Found invalid root: " << root << " " << info;
560      }
561    }
562  }
563};
564
565void MarkSweep::VerifyRoots() {
566  VerifyRootVisitor visitor;
567  Runtime::Current()->GetThreadList()->VisitRoots(&visitor);
568}
569
570void MarkSweep::MarkRoots(Thread* self) {
571  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
572  if (Locks::mutator_lock_->IsExclusiveHeld(self)) {
573    // If we exclusively hold the mutator lock, all threads must be suspended.
574    Runtime::Current()->VisitRoots(this);
575    RevokeAllThreadLocalAllocationStacks(self);
576  } else {
577    MarkRootsCheckpoint(self, kRevokeRosAllocThreadLocalBuffersAtCheckpoint);
578    // At this point the live stack should no longer have any mutators which push into it.
579    MarkNonThreadRoots();
580    MarkConcurrentRoots(
581        static_cast<VisitRootFlags>(kVisitRootFlagAllRoots | kVisitRootFlagStartLoggingNewRoots));
582  }
583}
584
585void MarkSweep::MarkNonThreadRoots() {
586  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
587  Runtime::Current()->VisitNonThreadRoots(this);
588}
589
590void MarkSweep::MarkConcurrentRoots(VisitRootFlags flags) {
591  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
592  // Visit all runtime roots and clear dirty flags.
593  Runtime::Current()->VisitConcurrentRoots(
594      this, static_cast<VisitRootFlags>(flags | kVisitRootFlagNonMoving));
595}
596
597class ScanObjectVisitor {
598 public:
599  explicit ScanObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE
600      : mark_sweep_(mark_sweep) {}
601
602  void operator()(Object* obj) const ALWAYS_INLINE SHARED_LOCKS_REQUIRED(Locks::mutator_lock_)
603      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) {
604    if (kCheckLocks) {
605      Locks::mutator_lock_->AssertSharedHeld(Thread::Current());
606      Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current());
607    }
608    mark_sweep_->ScanObject(obj);
609  }
610
611 private:
612  MarkSweep* const mark_sweep_;
613};
614
615class DelayReferenceReferentVisitor {
616 public:
617  explicit DelayReferenceReferentVisitor(MarkSweep* collector) : collector_(collector) {
618  }
619
620  void operator()(mirror::Class* klass, mirror::Reference* ref) const
621      SHARED_LOCKS_REQUIRED(Locks::mutator_lock_)
622      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) {
623    collector_->DelayReferenceReferent(klass, ref);
624  }
625
626 private:
627  MarkSweep* const collector_;
628};
629
630template <bool kUseFinger = false>
631class MarkStackTask : public Task {
632 public:
633  MarkStackTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, size_t mark_stack_size,
634                StackReference<Object>* mark_stack)
635      : mark_sweep_(mark_sweep),
636        thread_pool_(thread_pool),
637        mark_stack_pos_(mark_stack_size) {
638    // We may have to copy part of an existing mark stack when another mark stack overflows.
639    if (mark_stack_size != 0) {
640      DCHECK(mark_stack != nullptr);
641      // TODO: Check performance?
642      std::copy(mark_stack, mark_stack + mark_stack_size, mark_stack_);
643    }
644    if (kCountTasks) {
645      ++mark_sweep_->work_chunks_created_;
646    }
647  }
648
649  static const size_t kMaxSize = 1 * KB;
650
651 protected:
652  class MarkObjectParallelVisitor {
653   public:
654    explicit MarkObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task,
655                                       MarkSweep* mark_sweep) ALWAYS_INLINE
656            : chunk_task_(chunk_task), mark_sweep_(mark_sweep) {}
657
658    void operator()(Object* obj, MemberOffset offset, bool /* static */) const ALWAYS_INLINE
659        SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
660      mirror::Object* ref = obj->GetFieldObject<mirror::Object>(offset);
661      if (ref != nullptr && mark_sweep_->MarkObjectParallel(ref)) {
662        if (kUseFinger) {
663          std::atomic_thread_fence(std::memory_order_seq_cst);
664          if (reinterpret_cast<uintptr_t>(ref) >=
665              static_cast<uintptr_t>(mark_sweep_->atomic_finger_.LoadRelaxed())) {
666            return;
667          }
668        }
669        chunk_task_->MarkStackPush(ref);
670      }
671    }
672
673   private:
674    MarkStackTask<kUseFinger>* const chunk_task_;
675    MarkSweep* const mark_sweep_;
676  };
677
678  class ScanObjectParallelVisitor {
679   public:
680    explicit ScanObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task) ALWAYS_INLINE
681        : chunk_task_(chunk_task) {}
682
683    // No thread safety analysis since multiple threads will use this visitor.
684    void operator()(Object* obj) const SHARED_LOCKS_REQUIRED(Locks::mutator_lock_)
685        EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) {
686      MarkSweep* const mark_sweep = chunk_task_->mark_sweep_;
687      MarkObjectParallelVisitor mark_visitor(chunk_task_, mark_sweep);
688      DelayReferenceReferentVisitor ref_visitor(mark_sweep);
689      mark_sweep->ScanObjectVisit(obj, mark_visitor, ref_visitor);
690    }
691
692   private:
693    MarkStackTask<kUseFinger>* const chunk_task_;
694  };
695
696  virtual ~MarkStackTask() {
697    // Make sure that we have cleared our mark stack.
698    DCHECK_EQ(mark_stack_pos_, 0U);
699    if (kCountTasks) {
700      ++mark_sweep_->work_chunks_deleted_;
701    }
702  }
703
704  MarkSweep* const mark_sweep_;
705  ThreadPool* const thread_pool_;
706  // Thread local mark stack for this task.
707  StackReference<Object> mark_stack_[kMaxSize];
708  // Mark stack position.
709  size_t mark_stack_pos_;
710
711  ALWAYS_INLINE void MarkStackPush(Object* obj) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
712    if (UNLIKELY(mark_stack_pos_ == kMaxSize)) {
713      // Mark stack overflow, give 1/2 the stack to the thread pool as a new work task.
714      mark_stack_pos_ /= 2;
715      auto* task = new MarkStackTask(thread_pool_, mark_sweep_, kMaxSize - mark_stack_pos_,
716                                     mark_stack_ + mark_stack_pos_);
717      thread_pool_->AddTask(Thread::Current(), task);
718    }
719    DCHECK(obj != nullptr);
720    DCHECK_LT(mark_stack_pos_, kMaxSize);
721    mark_stack_[mark_stack_pos_++].Assign(obj);
722  }
723
724  virtual void Finalize() {
725    delete this;
726  }
727
728  // Scans all of the objects
729  virtual void Run(Thread* self) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_)
730      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) {
731    UNUSED(self);
732    ScanObjectParallelVisitor visitor(this);
733    // TODO: Tune this.
734    static const size_t kFifoSize = 4;
735    BoundedFifoPowerOfTwo<Object*, kFifoSize> prefetch_fifo;
736    for (;;) {
737      Object* obj = nullptr;
738      if (kUseMarkStackPrefetch) {
739        while (mark_stack_pos_ != 0 && prefetch_fifo.size() < kFifoSize) {
740          Object* const mark_stack_obj = mark_stack_[--mark_stack_pos_].AsMirrorPtr();
741          DCHECK(mark_stack_obj != nullptr);
742          __builtin_prefetch(mark_stack_obj);
743          prefetch_fifo.push_back(mark_stack_obj);
744        }
745        if (UNLIKELY(prefetch_fifo.empty())) {
746          break;
747        }
748        obj = prefetch_fifo.front();
749        prefetch_fifo.pop_front();
750      } else {
751        if (UNLIKELY(mark_stack_pos_ == 0)) {
752          break;
753        }
754        obj = mark_stack_[--mark_stack_pos_].AsMirrorPtr();
755      }
756      DCHECK(obj != nullptr);
757      visitor(obj);
758    }
759  }
760};
761
762class CardScanTask : public MarkStackTask<false> {
763 public:
764  CardScanTask(ThreadPool* thread_pool, MarkSweep* mark_sweep,
765               accounting::ContinuousSpaceBitmap* bitmap,
766               uint8_t* begin, uint8_t* end, uint8_t minimum_age, size_t mark_stack_size,
767               StackReference<Object>* mark_stack_obj, bool clear_card)
768      : MarkStackTask<false>(thread_pool, mark_sweep, mark_stack_size, mark_stack_obj),
769        bitmap_(bitmap),
770        begin_(begin),
771        end_(end),
772        minimum_age_(minimum_age), clear_card_(clear_card) {
773  }
774
775 protected:
776  accounting::ContinuousSpaceBitmap* const bitmap_;
777  uint8_t* const begin_;
778  uint8_t* const end_;
779  const uint8_t minimum_age_;
780  const bool clear_card_;
781
782  virtual void Finalize() {
783    delete this;
784  }
785
786  virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS {
787    ScanObjectParallelVisitor visitor(this);
788    accounting::CardTable* card_table = mark_sweep_->GetHeap()->GetCardTable();
789    size_t cards_scanned = clear_card_ ?
790                           card_table->Scan<true>(bitmap_, begin_, end_, visitor, minimum_age_) :
791                           card_table->Scan<false>(bitmap_, begin_, end_, visitor, minimum_age_);
792    VLOG(heap) << "Parallel scanning cards " << reinterpret_cast<void*>(begin_) << " - "
793        << reinterpret_cast<void*>(end_) << " = " << cards_scanned;
794    // Finish by emptying our local mark stack.
795    MarkStackTask::Run(self);
796  }
797};
798
799size_t MarkSweep::GetThreadCount(bool paused) const {
800  if (heap_->GetThreadPool() == nullptr || !heap_->CareAboutPauseTimes()) {
801    return 1;
802  }
803  return (paused ? heap_->GetParallelGCThreadCount() : heap_->GetConcGCThreadCount()) + 1;
804}
805
806void MarkSweep::ScanGrayObjects(bool paused, uint8_t minimum_age) {
807  accounting::CardTable* card_table = GetHeap()->GetCardTable();
808  ThreadPool* thread_pool = GetHeap()->GetThreadPool();
809  size_t thread_count = GetThreadCount(paused);
810  // The parallel version with only one thread is faster for card scanning, TODO: fix.
811  if (kParallelCardScan && thread_count > 1) {
812    Thread* self = Thread::Current();
813    // Can't have a different split for each space since multiple spaces can have their cards being
814    // scanned at the same time.
815    TimingLogger::ScopedTiming t(paused ? "(Paused)ScanGrayObjects" : __FUNCTION__,
816        GetTimings());
817    // Try to take some of the mark stack since we can pass this off to the worker tasks.
818    StackReference<Object>* mark_stack_begin = mark_stack_->Begin();
819    StackReference<Object>* mark_stack_end = mark_stack_->End();
820    const size_t mark_stack_size = mark_stack_end - mark_stack_begin;
821    // Estimated number of work tasks we will create.
822    const size_t mark_stack_tasks = GetHeap()->GetContinuousSpaces().size() * thread_count;
823    DCHECK_NE(mark_stack_tasks, 0U);
824    const size_t mark_stack_delta = std::min(CardScanTask::kMaxSize / 2,
825                                             mark_stack_size / mark_stack_tasks + 1);
826    for (const auto& space : GetHeap()->GetContinuousSpaces()) {
827      if (space->GetMarkBitmap() == nullptr) {
828        continue;
829      }
830      uint8_t* card_begin = space->Begin();
831      uint8_t* card_end = space->End();
832      // Align up the end address. For example, the image space's end
833      // may not be card-size-aligned.
834      card_end = AlignUp(card_end, accounting::CardTable::kCardSize);
835      DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_begin));
836      DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_end));
837      // Calculate how many bytes of heap we will scan,
838      const size_t address_range = card_end - card_begin;
839      // Calculate how much address range each task gets.
840      const size_t card_delta = RoundUp(address_range / thread_count + 1,
841                                        accounting::CardTable::kCardSize);
842      // If paused and the space is neither zygote nor image space, we could clear the dirty
843      // cards to avoid accumulating them to increase card scanning load in the following GC
844      // cycles. We need to keep dirty cards of image space and zygote space in order to track
845      // references to the other spaces.
846      bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace();
847      // Create the worker tasks for this space.
848      while (card_begin != card_end) {
849        // Add a range of cards.
850        size_t addr_remaining = card_end - card_begin;
851        size_t card_increment = std::min(card_delta, addr_remaining);
852        // Take from the back of the mark stack.
853        size_t mark_stack_remaining = mark_stack_end - mark_stack_begin;
854        size_t mark_stack_increment = std::min(mark_stack_delta, mark_stack_remaining);
855        mark_stack_end -= mark_stack_increment;
856        mark_stack_->PopBackCount(static_cast<int32_t>(mark_stack_increment));
857        DCHECK_EQ(mark_stack_end, mark_stack_->End());
858        // Add the new task to the thread pool.
859        auto* task = new CardScanTask(thread_pool, this, space->GetMarkBitmap(), card_begin,
860                                      card_begin + card_increment, minimum_age,
861                                      mark_stack_increment, mark_stack_end, clear_card);
862        thread_pool->AddTask(self, task);
863        card_begin += card_increment;
864      }
865    }
866
867    // Note: the card scan below may dirty new cards (and scan them)
868    // as a side effect when a Reference object is encountered and
869    // queued during the marking. See b/11465268.
870    thread_pool->SetMaxActiveWorkers(thread_count - 1);
871    thread_pool->StartWorkers(self);
872    thread_pool->Wait(self, true, true);
873    thread_pool->StopWorkers(self);
874  } else {
875    for (const auto& space : GetHeap()->GetContinuousSpaces()) {
876      if (space->GetMarkBitmap() != nullptr) {
877        // Image spaces are handled properly since live == marked for them.
878        const char* name = nullptr;
879        switch (space->GetGcRetentionPolicy()) {
880        case space::kGcRetentionPolicyNeverCollect:
881          name = paused ? "(Paused)ScanGrayImageSpaceObjects" : "ScanGrayImageSpaceObjects";
882          break;
883        case space::kGcRetentionPolicyFullCollect:
884          name = paused ? "(Paused)ScanGrayZygoteSpaceObjects" : "ScanGrayZygoteSpaceObjects";
885          break;
886        case space::kGcRetentionPolicyAlwaysCollect:
887          name = paused ? "(Paused)ScanGrayAllocSpaceObjects" : "ScanGrayAllocSpaceObjects";
888          break;
889        default:
890          LOG(FATAL) << "Unreachable";
891          UNREACHABLE();
892        }
893        TimingLogger::ScopedTiming t(name, GetTimings());
894        ScanObjectVisitor visitor(this);
895        bool clear_card = paused && !space->IsZygoteSpace() && !space->IsImageSpace();
896        if (clear_card) {
897          card_table->Scan<true>(space->GetMarkBitmap(), space->Begin(), space->End(), visitor,
898                                 minimum_age);
899        } else {
900          card_table->Scan<false>(space->GetMarkBitmap(), space->Begin(), space->End(), visitor,
901                                  minimum_age);
902        }
903      }
904    }
905  }
906}
907
908class RecursiveMarkTask : public MarkStackTask<false> {
909 public:
910  RecursiveMarkTask(ThreadPool* thread_pool, MarkSweep* mark_sweep,
911                    accounting::ContinuousSpaceBitmap* bitmap, uintptr_t begin, uintptr_t end)
912      : MarkStackTask<false>(thread_pool, mark_sweep, 0, nullptr), bitmap_(bitmap), begin_(begin),
913        end_(end) {
914  }
915
916 protected:
917  accounting::ContinuousSpaceBitmap* const bitmap_;
918  const uintptr_t begin_;
919  const uintptr_t end_;
920
921  virtual void Finalize() {
922    delete this;
923  }
924
925  // Scans all of the objects
926  virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS {
927    ScanObjectParallelVisitor visitor(this);
928    bitmap_->VisitMarkedRange(begin_, end_, visitor);
929    // Finish by emptying our local mark stack.
930    MarkStackTask::Run(self);
931  }
932};
933
934// Populates the mark stack based on the set of marked objects and
935// recursively marks until the mark stack is emptied.
936void MarkSweep::RecursiveMark() {
937  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
938  // RecursiveMark will build the lists of known instances of the Reference classes. See
939  // DelayReferenceReferent for details.
940  if (kUseRecursiveMark) {
941    const bool partial = GetGcType() == kGcTypePartial;
942    ScanObjectVisitor scan_visitor(this);
943    auto* self = Thread::Current();
944    ThreadPool* thread_pool = heap_->GetThreadPool();
945    size_t thread_count = GetThreadCount(false);
946    const bool parallel = kParallelRecursiveMark && thread_count > 1;
947    mark_stack_->Reset();
948    for (const auto& space : GetHeap()->GetContinuousSpaces()) {
949      if ((space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) ||
950          (!partial && space->GetGcRetentionPolicy() == space::kGcRetentionPolicyFullCollect)) {
951        current_space_bitmap_ = space->GetMarkBitmap();
952        if (current_space_bitmap_ == nullptr) {
953          continue;
954        }
955        if (parallel) {
956          // We will use the mark stack the future.
957          // CHECK(mark_stack_->IsEmpty());
958          // This function does not handle heap end increasing, so we must use the space end.
959          uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin());
960          uintptr_t end = reinterpret_cast<uintptr_t>(space->End());
961          atomic_finger_.StoreRelaxed(AtomicInteger::MaxValue());
962
963          // Create a few worker tasks.
964          const size_t n = thread_count * 2;
965          while (begin != end) {
966            uintptr_t start = begin;
967            uintptr_t delta = (end - begin) / n;
968            delta = RoundUp(delta, KB);
969            if (delta < 16 * KB) delta = end - begin;
970            begin += delta;
971            auto* task = new RecursiveMarkTask(thread_pool, this, current_space_bitmap_, start,
972                                               begin);
973            thread_pool->AddTask(self, task);
974          }
975          thread_pool->SetMaxActiveWorkers(thread_count - 1);
976          thread_pool->StartWorkers(self);
977          thread_pool->Wait(self, true, true);
978          thread_pool->StopWorkers(self);
979        } else {
980          // This function does not handle heap end increasing, so we must use the space end.
981          uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin());
982          uintptr_t end = reinterpret_cast<uintptr_t>(space->End());
983          current_space_bitmap_->VisitMarkedRange(begin, end, scan_visitor);
984        }
985      }
986    }
987  }
988  ProcessMarkStack(false);
989}
990
991mirror::Object* MarkSweep::IsMarkedCallback(mirror::Object* object, void* arg) {
992  if (reinterpret_cast<MarkSweep*>(arg)->IsMarked(object)) {
993    return object;
994  }
995  return nullptr;
996}
997
998void MarkSweep::RecursiveMarkDirtyObjects(bool paused, uint8_t minimum_age) {
999  ScanGrayObjects(paused, minimum_age);
1000  ProcessMarkStack(paused);
1001}
1002
1003void MarkSweep::ReMarkRoots() {
1004  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1005  Locks::mutator_lock_->AssertExclusiveHeld(Thread::Current());
1006  Runtime::Current()->VisitRoots(this, static_cast<VisitRootFlags>(
1007      kVisitRootFlagNewRoots | kVisitRootFlagStopLoggingNewRoots | kVisitRootFlagClearRootLog));
1008  if (kVerifyRootsMarked) {
1009    TimingLogger::ScopedTiming t2("(Paused)VerifyRoots", GetTimings());
1010    VerifyRootMarkedVisitor visitor(this);
1011    Runtime::Current()->VisitRoots(&visitor);
1012  }
1013}
1014
1015void MarkSweep::SweepSystemWeaks(Thread* self) {
1016  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1017  WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
1018  Runtime::Current()->SweepSystemWeaks(IsMarkedCallback, this);
1019}
1020
1021mirror::Object* MarkSweep::VerifySystemWeakIsLiveCallback(Object* obj, void* arg) {
1022  reinterpret_cast<MarkSweep*>(arg)->VerifyIsLive(obj);
1023  // We don't actually want to sweep the object, so lets return "marked"
1024  return obj;
1025}
1026
1027void MarkSweep::VerifyIsLive(const Object* obj) {
1028  if (!heap_->GetLiveBitmap()->Test(obj)) {
1029    // TODO: Consider live stack? Has this code bitrotted?
1030    CHECK(!heap_->allocation_stack_->Contains(obj))
1031        << "Found dead object " << obj << "\n" << heap_->DumpSpaces();
1032  }
1033}
1034
1035void MarkSweep::VerifySystemWeaks() {
1036  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1037  // Verify system weaks, uses a special object visitor which returns the input object.
1038  Runtime::Current()->SweepSystemWeaks(VerifySystemWeakIsLiveCallback, this);
1039}
1040
1041class CheckpointMarkThreadRoots : public Closure, public RootVisitor {
1042 public:
1043  explicit CheckpointMarkThreadRoots(MarkSweep* mark_sweep,
1044                                     bool revoke_ros_alloc_thread_local_buffers_at_checkpoint)
1045      : mark_sweep_(mark_sweep),
1046        revoke_ros_alloc_thread_local_buffers_at_checkpoint_(
1047            revoke_ros_alloc_thread_local_buffers_at_checkpoint) {
1048  }
1049
1050  void VisitRoots(mirror::Object*** roots, size_t count, const RootInfo& info ATTRIBUTE_UNUSED)
1051      OVERRIDE SHARED_LOCKS_REQUIRED(Locks::mutator_lock_)
1052      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) {
1053    for (size_t i = 0; i < count; ++i) {
1054      mark_sweep_->MarkObjectNonNullParallel(*roots[i]);
1055    }
1056  }
1057
1058  void VisitRoots(mirror::CompressedReference<mirror::Object>** roots, size_t count,
1059                  const RootInfo& info ATTRIBUTE_UNUSED)
1060      OVERRIDE SHARED_LOCKS_REQUIRED(Locks::mutator_lock_)
1061      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) {
1062    for (size_t i = 0; i < count; ++i) {
1063      mark_sweep_->MarkObjectNonNullParallel(roots[i]->AsMirrorPtr());
1064    }
1065  }
1066
1067  virtual void Run(Thread* thread) OVERRIDE NO_THREAD_SAFETY_ANALYSIS {
1068    ATRACE_BEGIN("Marking thread roots");
1069    // Note: self is not necessarily equal to thread since thread may be suspended.
1070    Thread* const self = Thread::Current();
1071    CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc)
1072        << thread->GetState() << " thread " << thread << " self " << self;
1073    thread->VisitRoots(this);
1074    ATRACE_END();
1075    if (revoke_ros_alloc_thread_local_buffers_at_checkpoint_) {
1076      ATRACE_BEGIN("RevokeRosAllocThreadLocalBuffers");
1077      mark_sweep_->GetHeap()->RevokeRosAllocThreadLocalBuffers(thread);
1078      ATRACE_END();
1079    }
1080    // If thread is a running mutator, then act on behalf of the garbage collector.
1081    // See the code in ThreadList::RunCheckpoint.
1082    if (thread->GetState() == kRunnable) {
1083      mark_sweep_->GetBarrier().Pass(self);
1084    }
1085  }
1086
1087 private:
1088  MarkSweep* const mark_sweep_;
1089  const bool revoke_ros_alloc_thread_local_buffers_at_checkpoint_;
1090};
1091
1092void MarkSweep::MarkRootsCheckpoint(Thread* self,
1093                                    bool revoke_ros_alloc_thread_local_buffers_at_checkpoint) {
1094  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1095  CheckpointMarkThreadRoots check_point(this, revoke_ros_alloc_thread_local_buffers_at_checkpoint);
1096  ThreadList* thread_list = Runtime::Current()->GetThreadList();
1097  // Request the check point is run on all threads returning a count of the threads that must
1098  // run through the barrier including self.
1099  size_t barrier_count = thread_list->RunCheckpoint(&check_point);
1100  // Release locks then wait for all mutator threads to pass the barrier.
1101  // If there are no threads to wait which implys that all the checkpoint functions are finished,
1102  // then no need to release locks.
1103  if (barrier_count == 0) {
1104    return;
1105  }
1106  Locks::heap_bitmap_lock_->ExclusiveUnlock(self);
1107  Locks::mutator_lock_->SharedUnlock(self);
1108  {
1109    ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
1110    gc_barrier_->Increment(self, barrier_count);
1111  }
1112  Locks::mutator_lock_->SharedLock(self);
1113  Locks::heap_bitmap_lock_->ExclusiveLock(self);
1114}
1115
1116void MarkSweep::SweepArray(accounting::ObjectStack* allocations, bool swap_bitmaps) {
1117  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1118  Thread* self = Thread::Current();
1119  mirror::Object** chunk_free_buffer = reinterpret_cast<mirror::Object**>(
1120      sweep_array_free_buffer_mem_map_->BaseBegin());
1121  size_t chunk_free_pos = 0;
1122  ObjectBytePair freed;
1123  ObjectBytePair freed_los;
1124  // How many objects are left in the array, modified after each space is swept.
1125  StackReference<Object>* objects = allocations->Begin();
1126  size_t count = allocations->Size();
1127  // Change the order to ensure that the non-moving space last swept as an optimization.
1128  std::vector<space::ContinuousSpace*> sweep_spaces;
1129  space::ContinuousSpace* non_moving_space = nullptr;
1130  for (space::ContinuousSpace* space : heap_->GetContinuousSpaces()) {
1131    if (space->IsAllocSpace() && !immune_region_.ContainsSpace(space) &&
1132        space->GetLiveBitmap() != nullptr) {
1133      if (space == heap_->GetNonMovingSpace()) {
1134        non_moving_space = space;
1135      } else {
1136        sweep_spaces.push_back(space);
1137      }
1138    }
1139  }
1140  // Unlikely to sweep a significant amount of non_movable objects, so we do these after the after
1141  // the other alloc spaces as an optimization.
1142  if (non_moving_space != nullptr) {
1143    sweep_spaces.push_back(non_moving_space);
1144  }
1145  // Start by sweeping the continuous spaces.
1146  for (space::ContinuousSpace* space : sweep_spaces) {
1147    space::AllocSpace* alloc_space = space->AsAllocSpace();
1148    accounting::ContinuousSpaceBitmap* live_bitmap = space->GetLiveBitmap();
1149    accounting::ContinuousSpaceBitmap* mark_bitmap = space->GetMarkBitmap();
1150    if (swap_bitmaps) {
1151      std::swap(live_bitmap, mark_bitmap);
1152    }
1153    StackReference<Object>* out = objects;
1154    for (size_t i = 0; i < count; ++i) {
1155      Object* const obj = objects[i].AsMirrorPtr();
1156      if (kUseThreadLocalAllocationStack && obj == nullptr) {
1157        continue;
1158      }
1159      if (space->HasAddress(obj)) {
1160        // This object is in the space, remove it from the array and add it to the sweep buffer
1161        // if needed.
1162        if (!mark_bitmap->Test(obj)) {
1163          if (chunk_free_pos >= kSweepArrayChunkFreeSize) {
1164            TimingLogger::ScopedTiming t2("FreeList", GetTimings());
1165            freed.objects += chunk_free_pos;
1166            freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer);
1167            chunk_free_pos = 0;
1168          }
1169          chunk_free_buffer[chunk_free_pos++] = obj;
1170        }
1171      } else {
1172        (out++)->Assign(obj);
1173      }
1174    }
1175    if (chunk_free_pos > 0) {
1176      TimingLogger::ScopedTiming t2("FreeList", GetTimings());
1177      freed.objects += chunk_free_pos;
1178      freed.bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer);
1179      chunk_free_pos = 0;
1180    }
1181    // All of the references which space contained are no longer in the allocation stack, update
1182    // the count.
1183    count = out - objects;
1184  }
1185  // Handle the large object space.
1186  space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace();
1187  if (large_object_space != nullptr) {
1188    accounting::LargeObjectBitmap* large_live_objects = large_object_space->GetLiveBitmap();
1189    accounting::LargeObjectBitmap* large_mark_objects = large_object_space->GetMarkBitmap();
1190    if (swap_bitmaps) {
1191      std::swap(large_live_objects, large_mark_objects);
1192    }
1193    for (size_t i = 0; i < count; ++i) {
1194      Object* const obj = objects[i].AsMirrorPtr();
1195      // Handle large objects.
1196      if (kUseThreadLocalAllocationStack && obj == nullptr) {
1197        continue;
1198      }
1199      if (!large_mark_objects->Test(obj)) {
1200        ++freed_los.objects;
1201        freed_los.bytes += large_object_space->Free(self, obj);
1202      }
1203    }
1204  }
1205  {
1206    TimingLogger::ScopedTiming t2("RecordFree", GetTimings());
1207    RecordFree(freed);
1208    RecordFreeLOS(freed_los);
1209    t2.NewTiming("ResetStack");
1210    allocations->Reset();
1211  }
1212  sweep_array_free_buffer_mem_map_->MadviseDontNeedAndZero();
1213}
1214
1215void MarkSweep::Sweep(bool swap_bitmaps) {
1216  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1217  // Ensure that nobody inserted items in the live stack after we swapped the stacks.
1218  CHECK_GE(live_stack_freeze_size_, GetHeap()->GetLiveStack()->Size());
1219  {
1220    TimingLogger::ScopedTiming t2("MarkAllocStackAsLive", GetTimings());
1221    // Mark everything allocated since the last as GC live so that we can sweep concurrently,
1222    // knowing that new allocations won't be marked as live.
1223    accounting::ObjectStack* live_stack = heap_->GetLiveStack();
1224    heap_->MarkAllocStackAsLive(live_stack);
1225    live_stack->Reset();
1226    DCHECK(mark_stack_->IsEmpty());
1227  }
1228  for (const auto& space : GetHeap()->GetContinuousSpaces()) {
1229    if (space->IsContinuousMemMapAllocSpace()) {
1230      space::ContinuousMemMapAllocSpace* alloc_space = space->AsContinuousMemMapAllocSpace();
1231      TimingLogger::ScopedTiming split(
1232          alloc_space->IsZygoteSpace() ? "SweepZygoteSpace" : "SweepMallocSpace", GetTimings());
1233      RecordFree(alloc_space->Sweep(swap_bitmaps));
1234    }
1235  }
1236  SweepLargeObjects(swap_bitmaps);
1237}
1238
1239void MarkSweep::SweepLargeObjects(bool swap_bitmaps) {
1240  space::LargeObjectSpace* los = heap_->GetLargeObjectsSpace();
1241  if (los != nullptr) {
1242    TimingLogger::ScopedTiming split(__FUNCTION__, GetTimings());
1243    RecordFreeLOS(los->Sweep(swap_bitmaps));
1244  }
1245}
1246
1247// Process the "referent" field in a java.lang.ref.Reference.  If the referent has not yet been
1248// marked, put it on the appropriate list in the heap for later processing.
1249void MarkSweep::DelayReferenceReferent(mirror::Class* klass, mirror::Reference* ref) {
1250  if (kCountJavaLangRefs) {
1251    ++reference_count_;
1252  }
1253  heap_->GetReferenceProcessor()->DelayReferenceReferent(klass, ref, &HeapReferenceMarkedCallback,
1254                                                         this);
1255}
1256
1257class MarkObjectVisitor {
1258 public:
1259  explicit MarkObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE : mark_sweep_(mark_sweep) {
1260  }
1261
1262  void operator()(Object* obj, MemberOffset offset, bool /* is_static */) const
1263      ALWAYS_INLINE SHARED_LOCKS_REQUIRED(Locks::mutator_lock_)
1264      EXCLUSIVE_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) {
1265    if (kCheckLocks) {
1266      Locks::mutator_lock_->AssertSharedHeld(Thread::Current());
1267      Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current());
1268    }
1269    mark_sweep_->MarkObject(obj->GetFieldObject<mirror::Object>(offset), obj, offset);
1270  }
1271
1272 private:
1273  MarkSweep* const mark_sweep_;
1274};
1275
1276// Scans an object reference.  Determines the type of the reference
1277// and dispatches to a specialized scanning routine.
1278void MarkSweep::ScanObject(Object* obj) {
1279  MarkObjectVisitor mark_visitor(this);
1280  DelayReferenceReferentVisitor ref_visitor(this);
1281  ScanObjectVisit(obj, mark_visitor, ref_visitor);
1282}
1283
1284void MarkSweep::ProcessMarkStackCallback(void* arg) {
1285  reinterpret_cast<MarkSweep*>(arg)->ProcessMarkStack(false);
1286}
1287
1288void MarkSweep::ProcessMarkStackParallel(size_t thread_count) {
1289  Thread* self = Thread::Current();
1290  ThreadPool* thread_pool = GetHeap()->GetThreadPool();
1291  const size_t chunk_size = std::min(mark_stack_->Size() / thread_count + 1,
1292                                     static_cast<size_t>(MarkStackTask<false>::kMaxSize));
1293  CHECK_GT(chunk_size, 0U);
1294  // Split the current mark stack up into work tasks.
1295  for (auto* it = mark_stack_->Begin(), *end = mark_stack_->End(); it < end; ) {
1296    const size_t delta = std::min(static_cast<size_t>(end - it), chunk_size);
1297    thread_pool->AddTask(self, new MarkStackTask<false>(thread_pool, this, delta, it));
1298    it += delta;
1299  }
1300  thread_pool->SetMaxActiveWorkers(thread_count - 1);
1301  thread_pool->StartWorkers(self);
1302  thread_pool->Wait(self, true, true);
1303  thread_pool->StopWorkers(self);
1304  mark_stack_->Reset();
1305  CHECK_EQ(work_chunks_created_.LoadSequentiallyConsistent(),
1306           work_chunks_deleted_.LoadSequentiallyConsistent())
1307      << " some of the work chunks were leaked";
1308}
1309
1310// Scan anything that's on the mark stack.
1311void MarkSweep::ProcessMarkStack(bool paused) {
1312  TimingLogger::ScopedTiming t(paused ? "(Paused)ProcessMarkStack" : __FUNCTION__, GetTimings());
1313  size_t thread_count = GetThreadCount(paused);
1314  if (kParallelProcessMarkStack && thread_count > 1 &&
1315      mark_stack_->Size() >= kMinimumParallelMarkStackSize) {
1316    ProcessMarkStackParallel(thread_count);
1317  } else {
1318    // TODO: Tune this.
1319    static const size_t kFifoSize = 4;
1320    BoundedFifoPowerOfTwo<Object*, kFifoSize> prefetch_fifo;
1321    for (;;) {
1322      Object* obj = nullptr;
1323      if (kUseMarkStackPrefetch) {
1324        while (!mark_stack_->IsEmpty() && prefetch_fifo.size() < kFifoSize) {
1325          Object* mark_stack_obj = mark_stack_->PopBack();
1326          DCHECK(mark_stack_obj != nullptr);
1327          __builtin_prefetch(mark_stack_obj);
1328          prefetch_fifo.push_back(mark_stack_obj);
1329        }
1330        if (prefetch_fifo.empty()) {
1331          break;
1332        }
1333        obj = prefetch_fifo.front();
1334        prefetch_fifo.pop_front();
1335      } else {
1336        if (mark_stack_->IsEmpty()) {
1337          break;
1338        }
1339        obj = mark_stack_->PopBack();
1340      }
1341      DCHECK(obj != nullptr);
1342      ScanObject(obj);
1343    }
1344  }
1345}
1346
1347inline bool MarkSweep::IsMarked(const Object* object) const {
1348  if (immune_region_.ContainsObject(object)) {
1349    return true;
1350  }
1351  if (current_space_bitmap_->HasAddress(object)) {
1352    return current_space_bitmap_->Test(object);
1353  }
1354  return mark_bitmap_->Test(object);
1355}
1356
1357void MarkSweep::FinishPhase() {
1358  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1359  if (kCountScannedTypes) {
1360    VLOG(gc) << "MarkSweep scanned classes=" << class_count_.LoadRelaxed()
1361        << " arrays=" << array_count_.LoadRelaxed() << " other=" << other_count_.LoadRelaxed();
1362  }
1363  if (kCountTasks) {
1364    VLOG(gc) << "Total number of work chunks allocated: " << work_chunks_created_.LoadRelaxed();
1365  }
1366  if (kMeasureOverhead) {
1367    VLOG(gc) << "Overhead time " << PrettyDuration(overhead_time_.LoadRelaxed());
1368  }
1369  if (kProfileLargeObjects) {
1370    VLOG(gc) << "Large objects tested " << large_object_test_.LoadRelaxed()
1371        << " marked " << large_object_mark_.LoadRelaxed();
1372  }
1373  if (kCountJavaLangRefs) {
1374    VLOG(gc) << "References scanned " << reference_count_.LoadRelaxed();
1375  }
1376  if (kCountMarkedObjects) {
1377    VLOG(gc) << "Marked: null=" << mark_null_count_.LoadRelaxed()
1378        << " immune=" <<  mark_immune_count_.LoadRelaxed()
1379        << " fastpath=" << mark_fastpath_count_.LoadRelaxed()
1380        << " slowpath=" << mark_slowpath_count_.LoadRelaxed();
1381  }
1382  CHECK(mark_stack_->IsEmpty());  // Ensure that the mark stack is empty.
1383  mark_stack_->Reset();
1384  WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_);
1385  heap_->ClearMarkedObjects();
1386}
1387
1388void MarkSweep::RevokeAllThreadLocalBuffers() {
1389  if (kRevokeRosAllocThreadLocalBuffersAtCheckpoint && IsConcurrent()) {
1390    // If concurrent, rosalloc thread-local buffers are revoked at the
1391    // thread checkpoint. Bump pointer space thread-local buffers must
1392    // not be in use.
1393    GetHeap()->AssertAllBumpPointerSpaceThreadLocalBuffersAreRevoked();
1394  } else {
1395    TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
1396    GetHeap()->RevokeAllThreadLocalBuffers();
1397  }
1398}
1399
1400}  // namespace collector
1401}  // namespace gc
1402}  // namespace art
1403