concurrent_copying.cc revision 5496f69c0a4c2cc357a065f57b7f4ff5d9ad2fa9
1/*
2 * Copyright (C) 2014 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 "concurrent_copying.h"
18
19#include "art_field-inl.h"
20#include "base/stl_util.h"
21#include "debugger.h"
22#include "gc/accounting/heap_bitmap-inl.h"
23#include "gc/accounting/space_bitmap-inl.h"
24#include "gc/reference_processor.h"
25#include "gc/space/image_space.h"
26#include "gc/space/space-inl.h"
27#include "image-inl.h"
28#include "intern_table.h"
29#include "mirror/class-inl.h"
30#include "mirror/object-inl.h"
31#include "scoped_thread_state_change.h"
32#include "thread-inl.h"
33#include "thread_list.h"
34#include "well_known_classes.h"
35
36namespace art {
37namespace gc {
38namespace collector {
39
40static constexpr size_t kDefaultGcMarkStackSize = 2 * MB;
41
42ConcurrentCopying::ConcurrentCopying(Heap* heap, const std::string& name_prefix)
43    : GarbageCollector(heap,
44                       name_prefix + (name_prefix.empty() ? "" : " ") +
45                       "concurrent copying + mark sweep"),
46      region_space_(nullptr), gc_barrier_(new Barrier(0)),
47      gc_mark_stack_(accounting::ObjectStack::Create("concurrent copying gc mark stack",
48                                                     kDefaultGcMarkStackSize,
49                                                     kDefaultGcMarkStackSize)),
50      mark_stack_lock_("concurrent copying mark stack lock", kMarkSweepMarkStackLock),
51      thread_running_gc_(nullptr),
52      is_marking_(false), is_active_(false), is_asserting_to_space_invariant_(false),
53      heap_mark_bitmap_(nullptr), live_stack_freeze_size_(0), mark_stack_mode_(kMarkStackModeOff),
54      weak_ref_access_enabled_(true),
55      skipped_blocks_lock_("concurrent copying bytes blocks lock", kMarkSweepMarkStackLock),
56      rb_table_(heap_->GetReadBarrierTable()),
57      force_evacuate_all_(false) {
58  static_assert(space::RegionSpace::kRegionSize == accounting::ReadBarrierTable::kRegionSize,
59                "The region space size and the read barrier table region size must match");
60  cc_heap_bitmap_.reset(new accounting::HeapBitmap(heap));
61  Thread* self = Thread::Current();
62  {
63    ReaderMutexLock mu(self, *Locks::heap_bitmap_lock_);
64    // Cache this so that we won't have to lock heap_bitmap_lock_ in
65    // Mark() which could cause a nested lock on heap_bitmap_lock_
66    // when GC causes a RB while doing GC or a lock order violation
67    // (class_linker_lock_ and heap_bitmap_lock_).
68    heap_mark_bitmap_ = heap->GetMarkBitmap();
69  }
70  {
71    MutexLock mu(self, mark_stack_lock_);
72    for (size_t i = 0; i < kMarkStackPoolSize; ++i) {
73      accounting::AtomicStack<mirror::Object>* mark_stack =
74          accounting::AtomicStack<mirror::Object>::Create(
75              "thread local mark stack", kMarkStackSize, kMarkStackSize);
76      pooled_mark_stacks_.push_back(mark_stack);
77    }
78  }
79}
80
81void ConcurrentCopying::MarkHeapReference(mirror::HeapReference<mirror::Object>* from_ref) {
82  // Used for preserving soft references, should be OK to not have a CAS here since there should be
83  // no other threads which can trigger read barriers on the same referent during reference
84  // processing.
85  from_ref->Assign(Mark(from_ref->AsMirrorPtr()));
86  DCHECK(!from_ref->IsNull());
87}
88
89ConcurrentCopying::~ConcurrentCopying() {
90  STLDeleteElements(&pooled_mark_stacks_);
91}
92
93void ConcurrentCopying::RunPhases() {
94  CHECK(kUseBakerReadBarrier || kUseTableLookupReadBarrier);
95  CHECK(!is_active_);
96  is_active_ = true;
97  Thread* self = Thread::Current();
98  thread_running_gc_ = self;
99  Locks::mutator_lock_->AssertNotHeld(self);
100  {
101    ReaderMutexLock mu(self, *Locks::mutator_lock_);
102    InitializePhase();
103  }
104  FlipThreadRoots();
105  {
106    ReaderMutexLock mu(self, *Locks::mutator_lock_);
107    MarkingPhase();
108  }
109  // Verify no from space refs. This causes a pause.
110  if (kEnableNoFromSpaceRefsVerification || kIsDebugBuild) {
111    TimingLogger::ScopedTiming split("(Paused)VerifyNoFromSpaceReferences", GetTimings());
112    ScopedPause pause(this);
113    CheckEmptyMarkStack();
114    if (kVerboseMode) {
115      LOG(INFO) << "Verifying no from-space refs";
116    }
117    VerifyNoFromSpaceReferences();
118    if (kVerboseMode) {
119      LOG(INFO) << "Done verifying no from-space refs";
120    }
121    CheckEmptyMarkStack();
122  }
123  {
124    ReaderMutexLock mu(self, *Locks::mutator_lock_);
125    ReclaimPhase();
126  }
127  FinishPhase();
128  CHECK(is_active_);
129  is_active_ = false;
130  thread_running_gc_ = nullptr;
131}
132
133void ConcurrentCopying::BindBitmaps() {
134  Thread* self = Thread::Current();
135  WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
136  // Mark all of the spaces we never collect as immune.
137  for (const auto& space : heap_->GetContinuousSpaces()) {
138    if (space->GetGcRetentionPolicy() == space::kGcRetentionPolicyNeverCollect ||
139        space->GetGcRetentionPolicy() == space::kGcRetentionPolicyFullCollect) {
140      CHECK(space->IsZygoteSpace() || space->IsImageSpace());
141      immune_spaces_.AddSpace(space);
142      const char* bitmap_name = space->IsImageSpace() ? "cc image space bitmap" :
143          "cc zygote space bitmap";
144      // TODO: try avoiding using bitmaps for image/zygote to save space.
145      accounting::ContinuousSpaceBitmap* bitmap =
146          accounting::ContinuousSpaceBitmap::Create(bitmap_name, space->Begin(), space->Capacity());
147      cc_heap_bitmap_->AddContinuousSpaceBitmap(bitmap);
148      cc_bitmaps_.push_back(bitmap);
149    } else if (space == region_space_) {
150      accounting::ContinuousSpaceBitmap* bitmap =
151          accounting::ContinuousSpaceBitmap::Create("cc region space bitmap",
152                                                    space->Begin(), space->Capacity());
153      cc_heap_bitmap_->AddContinuousSpaceBitmap(bitmap);
154      cc_bitmaps_.push_back(bitmap);
155      region_space_bitmap_ = bitmap;
156    }
157  }
158}
159
160void ConcurrentCopying::InitializePhase() {
161  TimingLogger::ScopedTiming split("InitializePhase", GetTimings());
162  if (kVerboseMode) {
163    LOG(INFO) << "GC InitializePhase";
164    LOG(INFO) << "Region-space : " << reinterpret_cast<void*>(region_space_->Begin()) << "-"
165              << reinterpret_cast<void*>(region_space_->Limit());
166  }
167  CheckEmptyMarkStack();
168  immune_spaces_.Reset();
169  bytes_moved_.StoreRelaxed(0);
170  objects_moved_.StoreRelaxed(0);
171  if (GetCurrentIteration()->GetGcCause() == kGcCauseExplicit ||
172      GetCurrentIteration()->GetGcCause() == kGcCauseForNativeAlloc ||
173      GetCurrentIteration()->GetClearSoftReferences()) {
174    force_evacuate_all_ = true;
175  } else {
176    force_evacuate_all_ = false;
177  }
178  BindBitmaps();
179  if (kVerboseMode) {
180    LOG(INFO) << "force_evacuate_all=" << force_evacuate_all_;
181    LOG(INFO) << "Largest immune region: " << immune_spaces_.GetLargestImmuneRegion().Begin()
182              << "-" << immune_spaces_.GetLargestImmuneRegion().End();
183    for (space::ContinuousSpace* space : immune_spaces_.GetSpaces()) {
184      LOG(INFO) << "Immune space: " << *space;
185    }
186    LOG(INFO) << "GC end of InitializePhase";
187  }
188}
189
190// Used to switch the thread roots of a thread from from-space refs to to-space refs.
191class ThreadFlipVisitor : public Closure {
192 public:
193  ThreadFlipVisitor(ConcurrentCopying* concurrent_copying, bool use_tlab)
194      : concurrent_copying_(concurrent_copying), use_tlab_(use_tlab) {
195  }
196
197  virtual void Run(Thread* thread) OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_) {
198    // Note: self is not necessarily equal to thread since thread may be suspended.
199    Thread* self = Thread::Current();
200    CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc)
201        << thread->GetState() << " thread " << thread << " self " << self;
202    thread->SetIsGcMarking(true);
203    if (use_tlab_ && thread->HasTlab()) {
204      if (ConcurrentCopying::kEnableFromSpaceAccountingCheck) {
205        // This must come before the revoke.
206        size_t thread_local_objects = thread->GetThreadLocalObjectsAllocated();
207        concurrent_copying_->region_space_->RevokeThreadLocalBuffers(thread);
208        reinterpret_cast<Atomic<size_t>*>(&concurrent_copying_->from_space_num_objects_at_first_pause_)->
209            FetchAndAddSequentiallyConsistent(thread_local_objects);
210      } else {
211        concurrent_copying_->region_space_->RevokeThreadLocalBuffers(thread);
212      }
213    }
214    if (kUseThreadLocalAllocationStack) {
215      thread->RevokeThreadLocalAllocationStack();
216    }
217    ReaderMutexLock mu(self, *Locks::heap_bitmap_lock_);
218    thread->VisitRoots(concurrent_copying_);
219    concurrent_copying_->GetBarrier().Pass(self);
220  }
221
222 private:
223  ConcurrentCopying* const concurrent_copying_;
224  const bool use_tlab_;
225};
226
227// Called back from Runtime::FlipThreadRoots() during a pause.
228class FlipCallback : public Closure {
229 public:
230  explicit FlipCallback(ConcurrentCopying* concurrent_copying)
231      : concurrent_copying_(concurrent_copying) {
232  }
233
234  virtual void Run(Thread* thread) OVERRIDE REQUIRES(Locks::mutator_lock_) {
235    ConcurrentCopying* cc = concurrent_copying_;
236    TimingLogger::ScopedTiming split("(Paused)FlipCallback", cc->GetTimings());
237    // Note: self is not necessarily equal to thread since thread may be suspended.
238    Thread* self = Thread::Current();
239    CHECK(thread == self);
240    Locks::mutator_lock_->AssertExclusiveHeld(self);
241    cc->region_space_->SetFromSpace(cc->rb_table_, cc->force_evacuate_all_);
242    cc->SwapStacks();
243    if (ConcurrentCopying::kEnableFromSpaceAccountingCheck) {
244      cc->RecordLiveStackFreezeSize(self);
245      cc->from_space_num_objects_at_first_pause_ = cc->region_space_->GetObjectsAllocated();
246      cc->from_space_num_bytes_at_first_pause_ = cc->region_space_->GetBytesAllocated();
247    }
248    cc->is_marking_ = true;
249    cc->mark_stack_mode_.StoreRelaxed(ConcurrentCopying::kMarkStackModeThreadLocal);
250    if (UNLIKELY(Runtime::Current()->IsActiveTransaction())) {
251      CHECK(Runtime::Current()->IsAotCompiler());
252      TimingLogger::ScopedTiming split2("(Paused)VisitTransactionRoots", cc->GetTimings());
253      Runtime::Current()->VisitTransactionRoots(cc);
254    }
255  }
256
257 private:
258  ConcurrentCopying* const concurrent_copying_;
259};
260
261// Switch threads that from from-space to to-space refs. Forward/mark the thread roots.
262void ConcurrentCopying::FlipThreadRoots() {
263  TimingLogger::ScopedTiming split("FlipThreadRoots", GetTimings());
264  if (kVerboseMode) {
265    LOG(INFO) << "time=" << region_space_->Time();
266    region_space_->DumpNonFreeRegions(LOG(INFO));
267  }
268  Thread* self = Thread::Current();
269  Locks::mutator_lock_->AssertNotHeld(self);
270  gc_barrier_->Init(self, 0);
271  ThreadFlipVisitor thread_flip_visitor(this, heap_->use_tlab_);
272  FlipCallback flip_callback(this);
273  heap_->ThreadFlipBegin(self);  // Sync with JNI critical calls.
274  size_t barrier_count = Runtime::Current()->FlipThreadRoots(
275      &thread_flip_visitor, &flip_callback, this);
276  heap_->ThreadFlipEnd(self);
277  {
278    ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
279    gc_barrier_->Increment(self, barrier_count);
280  }
281  is_asserting_to_space_invariant_ = true;
282  QuasiAtomic::ThreadFenceForConstructor();
283  if (kVerboseMode) {
284    LOG(INFO) << "time=" << region_space_->Time();
285    region_space_->DumpNonFreeRegions(LOG(INFO));
286    LOG(INFO) << "GC end of FlipThreadRoots";
287  }
288}
289
290void ConcurrentCopying::SwapStacks() {
291  heap_->SwapStacks();
292}
293
294void ConcurrentCopying::RecordLiveStackFreezeSize(Thread* self) {
295  WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
296  live_stack_freeze_size_ = heap_->GetLiveStack()->Size();
297}
298
299// Used to visit objects in the immune spaces.
300class ConcurrentCopyingImmuneSpaceObjVisitor {
301 public:
302  explicit ConcurrentCopyingImmuneSpaceObjVisitor(ConcurrentCopying* cc)
303      : collector_(cc) {}
304
305  void operator()(mirror::Object* obj) const SHARED_REQUIRES(Locks::mutator_lock_)
306      SHARED_REQUIRES(Locks::heap_bitmap_lock_) {
307    DCHECK(obj != nullptr);
308    DCHECK(collector_->immune_spaces_.ContainsObject(obj));
309    accounting::ContinuousSpaceBitmap* cc_bitmap =
310        collector_->cc_heap_bitmap_->GetContinuousSpaceBitmap(obj);
311    DCHECK(cc_bitmap != nullptr)
312        << "An immune space object must have a bitmap";
313    if (kIsDebugBuild) {
314      DCHECK(collector_->heap_->GetMarkBitmap()->Test(obj))
315          << "Immune space object must be already marked";
316    }
317    // This may or may not succeed, which is ok.
318    if (kUseBakerReadBarrier) {
319      obj->AtomicSetReadBarrierPointer(ReadBarrier::WhitePtr(), ReadBarrier::GrayPtr());
320    }
321    if (cc_bitmap->AtomicTestAndSet(obj)) {
322      // Already marked. Do nothing.
323    } else {
324      // Newly marked. Set the gray bit and push it onto the mark stack.
325      CHECK(!kUseBakerReadBarrier || obj->GetReadBarrierPointer() == ReadBarrier::GrayPtr());
326      collector_->PushOntoMarkStack(obj);
327    }
328  }
329
330 private:
331  ConcurrentCopying* const collector_;
332};
333
334class EmptyCheckpoint : public Closure {
335 public:
336  explicit EmptyCheckpoint(ConcurrentCopying* concurrent_copying)
337      : concurrent_copying_(concurrent_copying) {
338  }
339
340  virtual void Run(Thread* thread) OVERRIDE NO_THREAD_SAFETY_ANALYSIS {
341    // Note: self is not necessarily equal to thread since thread may be suspended.
342    Thread* self = Thread::Current();
343    CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc)
344        << thread->GetState() << " thread " << thread << " self " << self;
345    // If thread is a running mutator, then act on behalf of the garbage collector.
346    // See the code in ThreadList::RunCheckpoint.
347    concurrent_copying_->GetBarrier().Pass(self);
348  }
349
350 private:
351  ConcurrentCopying* const concurrent_copying_;
352};
353
354// Concurrently mark roots that are guarded by read barriers and process the mark stack.
355void ConcurrentCopying::MarkingPhase() {
356  TimingLogger::ScopedTiming split("MarkingPhase", GetTimings());
357  if (kVerboseMode) {
358    LOG(INFO) << "GC MarkingPhase";
359  }
360  CHECK(weak_ref_access_enabled_);
361  {
362    // Mark the image root. The WB-based collectors do not need to
363    // scan the image objects from roots by relying on the card table,
364    // but it's necessary for the RB to-space invariant to hold.
365    TimingLogger::ScopedTiming split1("VisitImageRoots", GetTimings());
366    for (space::ContinuousSpace* space : heap_->GetContinuousSpaces()) {
367      if (space->IsImageSpace()) {
368        gc::space::ImageSpace* image = space->AsImageSpace();
369        if (image != nullptr) {
370          mirror::ObjectArray<mirror::Object>* image_root = image->GetImageHeader().GetImageRoots();
371          mirror::Object* marked_image_root = Mark(image_root);
372          CHECK_EQ(image_root, marked_image_root) << "An image object does not move";
373          if (ReadBarrier::kEnableToSpaceInvariantChecks) {
374            AssertToSpaceInvariant(nullptr, MemberOffset(0), marked_image_root);
375          }
376        }
377      }
378    }
379  }
380  {
381    TimingLogger::ScopedTiming split2("VisitConcurrentRoots", GetTimings());
382    Runtime::Current()->VisitConcurrentRoots(this, kVisitRootFlagAllRoots);
383  }
384  {
385    // TODO: don't visit the transaction roots if it's not active.
386    TimingLogger::ScopedTiming split5("VisitNonThreadRoots", GetTimings());
387    Runtime::Current()->VisitNonThreadRoots(this);
388  }
389
390  // Immune spaces.
391  for (auto& space : immune_spaces_.GetSpaces()) {
392    DCHECK(space->IsImageSpace() || space->IsZygoteSpace());
393    accounting::ContinuousSpaceBitmap* live_bitmap = space->GetLiveBitmap();
394    ConcurrentCopyingImmuneSpaceObjVisitor visitor(this);
395    live_bitmap->VisitMarkedRange(reinterpret_cast<uintptr_t>(space->Begin()),
396                                  reinterpret_cast<uintptr_t>(space->Limit()),
397                                  visitor);
398  }
399
400  Thread* self = Thread::Current();
401  {
402    TimingLogger::ScopedTiming split7("ProcessMarkStack", GetTimings());
403    // We transition through three mark stack modes (thread-local, shared, GC-exclusive). The
404    // primary reasons are the fact that we need to use a checkpoint to process thread-local mark
405    // stacks, but after we disable weak refs accesses, we can't use a checkpoint due to a deadlock
406    // issue because running threads potentially blocking at WaitHoldingLocks, and that once we
407    // reach the point where we process weak references, we can avoid using a lock when accessing
408    // the GC mark stack, which makes mark stack processing more efficient.
409
410    // Process the mark stack once in the thread local stack mode. This marks most of the live
411    // objects, aside from weak ref accesses with read barriers (Reference::GetReferent() and system
412    // weaks) that may happen concurrently while we processing the mark stack and newly mark/gray
413    // objects and push refs on the mark stack.
414    ProcessMarkStack();
415    // Switch to the shared mark stack mode. That is, revoke and process thread-local mark stacks
416    // for the last time before transitioning to the shared mark stack mode, which would process new
417    // refs that may have been concurrently pushed onto the mark stack during the ProcessMarkStack()
418    // call above. At the same time, disable weak ref accesses using a per-thread flag. It's
419    // important to do these together in a single checkpoint so that we can ensure that mutators
420    // won't newly gray objects and push new refs onto the mark stack due to weak ref accesses and
421    // mutators safely transition to the shared mark stack mode (without leaving unprocessed refs on
422    // the thread-local mark stacks), without a race. This is why we use a thread-local weak ref
423    // access flag Thread::tls32_.weak_ref_access_enabled_ instead of the global ones.
424    SwitchToSharedMarkStackMode();
425    CHECK(!self->GetWeakRefAccessEnabled());
426    // Now that weak refs accesses are disabled, once we exhaust the shared mark stack again here
427    // (which may be non-empty if there were refs found on thread-local mark stacks during the above
428    // SwitchToSharedMarkStackMode() call), we won't have new refs to process, that is, mutators
429    // (via read barriers) have no way to produce any more refs to process. Marking converges once
430    // before we process weak refs below.
431    ProcessMarkStack();
432    CheckEmptyMarkStack();
433    // Switch to the GC exclusive mark stack mode so that we can process the mark stack without a
434    // lock from this point on.
435    SwitchToGcExclusiveMarkStackMode();
436    CheckEmptyMarkStack();
437    if (kVerboseMode) {
438      LOG(INFO) << "ProcessReferences";
439    }
440    // Process weak references. This may produce new refs to process and have them processed via
441    // ProcessMarkStack (in the GC exclusive mark stack mode).
442    ProcessReferences(self);
443    CheckEmptyMarkStack();
444    if (kVerboseMode) {
445      LOG(INFO) << "SweepSystemWeaks";
446    }
447    SweepSystemWeaks(self);
448    if (kVerboseMode) {
449      LOG(INFO) << "SweepSystemWeaks done";
450    }
451    // Process the mark stack here one last time because the above SweepSystemWeaks() call may have
452    // marked some objects (strings alive) as hash_set::Erase() can call the hash function for
453    // arbitrary elements in the weak intern table in InternTable::Table::SweepWeaks().
454    ProcessMarkStack();
455    CheckEmptyMarkStack();
456    // Re-enable weak ref accesses.
457    ReenableWeakRefAccess(self);
458    // Free data for class loaders that we unloaded.
459    Runtime::Current()->GetClassLinker()->CleanupClassLoaders();
460    // Marking is done. Disable marking.
461    DisableMarking();
462    CheckEmptyMarkStack();
463  }
464
465  CHECK(weak_ref_access_enabled_);
466  if (kVerboseMode) {
467    LOG(INFO) << "GC end of MarkingPhase";
468  }
469}
470
471void ConcurrentCopying::ReenableWeakRefAccess(Thread* self) {
472  if (kVerboseMode) {
473    LOG(INFO) << "ReenableWeakRefAccess";
474  }
475  weak_ref_access_enabled_.StoreRelaxed(true);  // This is for new threads.
476  QuasiAtomic::ThreadFenceForConstructor();
477  // Iterate all threads (don't need to or can't use a checkpoint) and re-enable weak ref access.
478  {
479    MutexLock mu(self, *Locks::thread_list_lock_);
480    std::list<Thread*> thread_list = Runtime::Current()->GetThreadList()->GetList();
481    for (Thread* thread : thread_list) {
482      thread->SetWeakRefAccessEnabled(true);
483    }
484  }
485  // Unblock blocking threads.
486  GetHeap()->GetReferenceProcessor()->BroadcastForSlowPath(self);
487  Runtime::Current()->BroadcastForNewSystemWeaks();
488}
489
490class DisableMarkingCheckpoint : public Closure {
491 public:
492  explicit DisableMarkingCheckpoint(ConcurrentCopying* concurrent_copying)
493      : concurrent_copying_(concurrent_copying) {
494  }
495
496  void Run(Thread* thread) OVERRIDE NO_THREAD_SAFETY_ANALYSIS {
497    // Note: self is not necessarily equal to thread since thread may be suspended.
498    Thread* self = Thread::Current();
499    DCHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc)
500        << thread->GetState() << " thread " << thread << " self " << self;
501    // Disable the thread-local is_gc_marking flag.
502    // Note a thread that has just started right before this checkpoint may have already this flag
503    // set to false, which is ok.
504    thread->SetIsGcMarking(false);
505    // If thread is a running mutator, then act on behalf of the garbage collector.
506    // See the code in ThreadList::RunCheckpoint.
507    concurrent_copying_->GetBarrier().Pass(self);
508  }
509
510 private:
511  ConcurrentCopying* const concurrent_copying_;
512};
513
514void ConcurrentCopying::IssueDisableMarkingCheckpoint() {
515  Thread* self = Thread::Current();
516  DisableMarkingCheckpoint check_point(this);
517  ThreadList* thread_list = Runtime::Current()->GetThreadList();
518  gc_barrier_->Init(self, 0);
519  size_t barrier_count = thread_list->RunCheckpoint(&check_point);
520  // If there are no threads to wait which implies that all the checkpoint functions are finished,
521  // then no need to release the mutator lock.
522  if (barrier_count == 0) {
523    return;
524  }
525  // Release locks then wait for all mutator threads to pass the barrier.
526  Locks::mutator_lock_->SharedUnlock(self);
527  {
528    ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
529    gc_barrier_->Increment(self, barrier_count);
530  }
531  Locks::mutator_lock_->SharedLock(self);
532}
533
534void ConcurrentCopying::DisableMarking() {
535  // Change the global is_marking flag to false. Do a fence before doing a checkpoint to update the
536  // thread-local flags so that a new thread starting up will get the correct is_marking flag.
537  is_marking_ = false;
538  QuasiAtomic::ThreadFenceForConstructor();
539  // Use a checkpoint to turn off the thread-local is_gc_marking flags and to ensure no threads are
540  // still in the middle of a read barrier which may have a from-space ref cached in a local
541  // variable.
542  IssueDisableMarkingCheckpoint();
543  if (kUseTableLookupReadBarrier) {
544    heap_->rb_table_->ClearAll();
545    DCHECK(heap_->rb_table_->IsAllCleared());
546  }
547  is_mark_stack_push_disallowed_.StoreSequentiallyConsistent(1);
548  mark_stack_mode_.StoreSequentiallyConsistent(kMarkStackModeOff);
549}
550
551void ConcurrentCopying::IssueEmptyCheckpoint() {
552  Thread* self = Thread::Current();
553  EmptyCheckpoint check_point(this);
554  ThreadList* thread_list = Runtime::Current()->GetThreadList();
555  gc_barrier_->Init(self, 0);
556  size_t barrier_count = thread_list->RunCheckpoint(&check_point);
557  // If there are no threads to wait which implys that all the checkpoint functions are finished,
558  // then no need to release the mutator lock.
559  if (barrier_count == 0) {
560    return;
561  }
562  // Release locks then wait for all mutator threads to pass the barrier.
563  Locks::mutator_lock_->SharedUnlock(self);
564  {
565    ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
566    gc_barrier_->Increment(self, barrier_count);
567  }
568  Locks::mutator_lock_->SharedLock(self);
569}
570
571void ConcurrentCopying::ExpandGcMarkStack() {
572  DCHECK(gc_mark_stack_->IsFull());
573  const size_t new_size = gc_mark_stack_->Capacity() * 2;
574  std::vector<StackReference<mirror::Object>> temp(gc_mark_stack_->Begin(),
575                                                   gc_mark_stack_->End());
576  gc_mark_stack_->Resize(new_size);
577  for (auto& ref : temp) {
578    gc_mark_stack_->PushBack(ref.AsMirrorPtr());
579  }
580  DCHECK(!gc_mark_stack_->IsFull());
581}
582
583void ConcurrentCopying::PushOntoMarkStack(mirror::Object* to_ref) {
584  CHECK_EQ(is_mark_stack_push_disallowed_.LoadRelaxed(), 0)
585      << " " << to_ref << " " << PrettyTypeOf(to_ref);
586  Thread* self = Thread::Current();  // TODO: pass self as an argument from call sites?
587  CHECK(thread_running_gc_ != nullptr);
588  MarkStackMode mark_stack_mode = mark_stack_mode_.LoadRelaxed();
589  if (LIKELY(mark_stack_mode == kMarkStackModeThreadLocal)) {
590    if (LIKELY(self == thread_running_gc_)) {
591      // If GC-running thread, use the GC mark stack instead of a thread-local mark stack.
592      CHECK(self->GetThreadLocalMarkStack() == nullptr);
593      if (UNLIKELY(gc_mark_stack_->IsFull())) {
594        ExpandGcMarkStack();
595      }
596      gc_mark_stack_->PushBack(to_ref);
597    } else {
598      // Otherwise, use a thread-local mark stack.
599      accounting::AtomicStack<mirror::Object>* tl_mark_stack = self->GetThreadLocalMarkStack();
600      if (UNLIKELY(tl_mark_stack == nullptr || tl_mark_stack->IsFull())) {
601        MutexLock mu(self, mark_stack_lock_);
602        // Get a new thread local mark stack.
603        accounting::AtomicStack<mirror::Object>* new_tl_mark_stack;
604        if (!pooled_mark_stacks_.empty()) {
605          // Use a pooled mark stack.
606          new_tl_mark_stack = pooled_mark_stacks_.back();
607          pooled_mark_stacks_.pop_back();
608        } else {
609          // None pooled. Create a new one.
610          new_tl_mark_stack =
611              accounting::AtomicStack<mirror::Object>::Create(
612                  "thread local mark stack", 4 * KB, 4 * KB);
613        }
614        DCHECK(new_tl_mark_stack != nullptr);
615        DCHECK(new_tl_mark_stack->IsEmpty());
616        new_tl_mark_stack->PushBack(to_ref);
617        self->SetThreadLocalMarkStack(new_tl_mark_stack);
618        if (tl_mark_stack != nullptr) {
619          // Store the old full stack into a vector.
620          revoked_mark_stacks_.push_back(tl_mark_stack);
621        }
622      } else {
623        tl_mark_stack->PushBack(to_ref);
624      }
625    }
626  } else if (mark_stack_mode == kMarkStackModeShared) {
627    // Access the shared GC mark stack with a lock.
628    MutexLock mu(self, mark_stack_lock_);
629    if (UNLIKELY(gc_mark_stack_->IsFull())) {
630      ExpandGcMarkStack();
631    }
632    gc_mark_stack_->PushBack(to_ref);
633  } else {
634    CHECK_EQ(static_cast<uint32_t>(mark_stack_mode),
635             static_cast<uint32_t>(kMarkStackModeGcExclusive))
636        << "ref=" << to_ref
637        << " self->gc_marking=" << self->GetIsGcMarking()
638        << " cc->is_marking=" << is_marking_;
639    CHECK(self == thread_running_gc_)
640        << "Only GC-running thread should access the mark stack "
641        << "in the GC exclusive mark stack mode";
642    // Access the GC mark stack without a lock.
643    if (UNLIKELY(gc_mark_stack_->IsFull())) {
644      ExpandGcMarkStack();
645    }
646    gc_mark_stack_->PushBack(to_ref);
647  }
648}
649
650accounting::ObjectStack* ConcurrentCopying::GetAllocationStack() {
651  return heap_->allocation_stack_.get();
652}
653
654accounting::ObjectStack* ConcurrentCopying::GetLiveStack() {
655  return heap_->live_stack_.get();
656}
657
658// The following visitors are that used to verify that there's no
659// references to the from-space left after marking.
660class ConcurrentCopyingVerifyNoFromSpaceRefsVisitor : public SingleRootVisitor {
661 public:
662  explicit ConcurrentCopyingVerifyNoFromSpaceRefsVisitor(ConcurrentCopying* collector)
663      : collector_(collector) {}
664
665  void operator()(mirror::Object* ref) const
666      SHARED_REQUIRES(Locks::mutator_lock_) ALWAYS_INLINE {
667    if (ref == nullptr) {
668      // OK.
669      return;
670    }
671    collector_->AssertToSpaceInvariant(nullptr, MemberOffset(0), ref);
672    if (kUseBakerReadBarrier) {
673      if (collector_->RegionSpace()->IsInToSpace(ref)) {
674        CHECK(ref->GetReadBarrierPointer() == nullptr)
675            << "To-space ref " << ref << " " << PrettyTypeOf(ref)
676            << " has non-white rb_ptr " << ref->GetReadBarrierPointer();
677      } else {
678        CHECK(ref->GetReadBarrierPointer() == ReadBarrier::BlackPtr() ||
679              (ref->GetReadBarrierPointer() == ReadBarrier::WhitePtr() &&
680               collector_->IsOnAllocStack(ref)))
681            << "Non-moving/unevac from space ref " << ref << " " << PrettyTypeOf(ref)
682            << " has non-black rb_ptr " << ref->GetReadBarrierPointer()
683            << " but isn't on the alloc stack (and has white rb_ptr)."
684            << " Is it in the non-moving space="
685            << (collector_->GetHeap()->GetNonMovingSpace()->HasAddress(ref));
686      }
687    }
688  }
689
690  void VisitRoot(mirror::Object* root, const RootInfo& info ATTRIBUTE_UNUSED)
691      OVERRIDE SHARED_REQUIRES(Locks::mutator_lock_) {
692    DCHECK(root != nullptr);
693    operator()(root);
694  }
695
696 private:
697  ConcurrentCopying* const collector_;
698};
699
700class ConcurrentCopyingVerifyNoFromSpaceRefsFieldVisitor {
701 public:
702  explicit ConcurrentCopyingVerifyNoFromSpaceRefsFieldVisitor(ConcurrentCopying* collector)
703      : collector_(collector) {}
704
705  void operator()(mirror::Object* obj, MemberOffset offset, bool is_static ATTRIBUTE_UNUSED) const
706      SHARED_REQUIRES(Locks::mutator_lock_) ALWAYS_INLINE {
707    mirror::Object* ref =
708        obj->GetFieldObject<mirror::Object, kDefaultVerifyFlags, kWithoutReadBarrier>(offset);
709    ConcurrentCopyingVerifyNoFromSpaceRefsVisitor visitor(collector_);
710    visitor(ref);
711  }
712  void operator()(mirror::Class* klass, mirror::Reference* ref) const
713      SHARED_REQUIRES(Locks::mutator_lock_) ALWAYS_INLINE {
714    CHECK(klass->IsTypeOfReferenceClass());
715    this->operator()(ref, mirror::Reference::ReferentOffset(), false);
716  }
717
718  void VisitRootIfNonNull(mirror::CompressedReference<mirror::Object>* root) const
719      SHARED_REQUIRES(Locks::mutator_lock_) {
720    if (!root->IsNull()) {
721      VisitRoot(root);
722    }
723  }
724
725  void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const
726      SHARED_REQUIRES(Locks::mutator_lock_) {
727    ConcurrentCopyingVerifyNoFromSpaceRefsVisitor visitor(collector_);
728    visitor(root->AsMirrorPtr());
729  }
730
731 private:
732  ConcurrentCopying* const collector_;
733};
734
735class ConcurrentCopyingVerifyNoFromSpaceRefsObjectVisitor {
736 public:
737  explicit ConcurrentCopyingVerifyNoFromSpaceRefsObjectVisitor(ConcurrentCopying* collector)
738      : collector_(collector) {}
739  void operator()(mirror::Object* obj) const
740      SHARED_REQUIRES(Locks::mutator_lock_) {
741    ObjectCallback(obj, collector_);
742  }
743  static void ObjectCallback(mirror::Object* obj, void *arg)
744      SHARED_REQUIRES(Locks::mutator_lock_) {
745    CHECK(obj != nullptr);
746    ConcurrentCopying* collector = reinterpret_cast<ConcurrentCopying*>(arg);
747    space::RegionSpace* region_space = collector->RegionSpace();
748    CHECK(!region_space->IsInFromSpace(obj)) << "Scanning object " << obj << " in from space";
749    ConcurrentCopyingVerifyNoFromSpaceRefsFieldVisitor visitor(collector);
750    obj->VisitReferences(visitor, visitor);
751    if (kUseBakerReadBarrier) {
752      if (collector->RegionSpace()->IsInToSpace(obj)) {
753        CHECK(obj->GetReadBarrierPointer() == nullptr)
754            << "obj=" << obj << " non-white rb_ptr " << obj->GetReadBarrierPointer();
755      } else {
756        CHECK(obj->GetReadBarrierPointer() == ReadBarrier::BlackPtr() ||
757              (obj->GetReadBarrierPointer() == ReadBarrier::WhitePtr() &&
758               collector->IsOnAllocStack(obj)))
759            << "Non-moving space/unevac from space ref " << obj << " " << PrettyTypeOf(obj)
760            << " has non-black rb_ptr " << obj->GetReadBarrierPointer()
761            << " but isn't on the alloc stack (and has white rb_ptr). Is it in the non-moving space="
762            << (collector->GetHeap()->GetNonMovingSpace()->HasAddress(obj));
763      }
764    }
765  }
766
767 private:
768  ConcurrentCopying* const collector_;
769};
770
771// Verify there's no from-space references left after the marking phase.
772void ConcurrentCopying::VerifyNoFromSpaceReferences() {
773  Thread* self = Thread::Current();
774  DCHECK(Locks::mutator_lock_->IsExclusiveHeld(self));
775  // Verify all threads have is_gc_marking to be false
776  {
777    MutexLock mu(self, *Locks::thread_list_lock_);
778    std::list<Thread*> thread_list = Runtime::Current()->GetThreadList()->GetList();
779    for (Thread* thread : thread_list) {
780      CHECK(!thread->GetIsGcMarking());
781    }
782  }
783  ConcurrentCopyingVerifyNoFromSpaceRefsObjectVisitor visitor(this);
784  // Roots.
785  {
786    ReaderMutexLock mu(self, *Locks::heap_bitmap_lock_);
787    ConcurrentCopyingVerifyNoFromSpaceRefsVisitor ref_visitor(this);
788    Runtime::Current()->VisitRoots(&ref_visitor);
789  }
790  // The to-space.
791  region_space_->WalkToSpace(ConcurrentCopyingVerifyNoFromSpaceRefsObjectVisitor::ObjectCallback,
792                             this);
793  // Non-moving spaces.
794  {
795    WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
796    heap_->GetMarkBitmap()->Visit(visitor);
797  }
798  // The alloc stack.
799  {
800    ConcurrentCopyingVerifyNoFromSpaceRefsVisitor ref_visitor(this);
801    for (auto* it = heap_->allocation_stack_->Begin(), *end = heap_->allocation_stack_->End();
802        it < end; ++it) {
803      mirror::Object* const obj = it->AsMirrorPtr();
804      if (obj != nullptr && obj->GetClass() != nullptr) {
805        // TODO: need to call this only if obj is alive?
806        ref_visitor(obj);
807        visitor(obj);
808      }
809    }
810  }
811  // TODO: LOS. But only refs in LOS are classes.
812}
813
814// The following visitors are used to assert the to-space invariant.
815class ConcurrentCopyingAssertToSpaceInvariantRefsVisitor {
816 public:
817  explicit ConcurrentCopyingAssertToSpaceInvariantRefsVisitor(ConcurrentCopying* collector)
818      : collector_(collector) {}
819
820  void operator()(mirror::Object* ref) const
821      SHARED_REQUIRES(Locks::mutator_lock_) ALWAYS_INLINE {
822    if (ref == nullptr) {
823      // OK.
824      return;
825    }
826    collector_->AssertToSpaceInvariant(nullptr, MemberOffset(0), ref);
827  }
828
829 private:
830  ConcurrentCopying* const collector_;
831};
832
833class ConcurrentCopyingAssertToSpaceInvariantFieldVisitor {
834 public:
835  explicit ConcurrentCopyingAssertToSpaceInvariantFieldVisitor(ConcurrentCopying* collector)
836      : collector_(collector) {}
837
838  void operator()(mirror::Object* obj, MemberOffset offset, bool is_static ATTRIBUTE_UNUSED) const
839      SHARED_REQUIRES(Locks::mutator_lock_) ALWAYS_INLINE {
840    mirror::Object* ref =
841        obj->GetFieldObject<mirror::Object, kDefaultVerifyFlags, kWithoutReadBarrier>(offset);
842    ConcurrentCopyingAssertToSpaceInvariantRefsVisitor visitor(collector_);
843    visitor(ref);
844  }
845  void operator()(mirror::Class* klass, mirror::Reference* ref ATTRIBUTE_UNUSED) const
846      SHARED_REQUIRES(Locks::mutator_lock_) ALWAYS_INLINE {
847    CHECK(klass->IsTypeOfReferenceClass());
848  }
849
850  void VisitRootIfNonNull(mirror::CompressedReference<mirror::Object>* root) const
851      SHARED_REQUIRES(Locks::mutator_lock_) {
852    if (!root->IsNull()) {
853      VisitRoot(root);
854    }
855  }
856
857  void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const
858      SHARED_REQUIRES(Locks::mutator_lock_) {
859    ConcurrentCopyingAssertToSpaceInvariantRefsVisitor visitor(collector_);
860    visitor(root->AsMirrorPtr());
861  }
862
863 private:
864  ConcurrentCopying* const collector_;
865};
866
867class ConcurrentCopyingAssertToSpaceInvariantObjectVisitor {
868 public:
869  explicit ConcurrentCopyingAssertToSpaceInvariantObjectVisitor(ConcurrentCopying* collector)
870      : collector_(collector) {}
871  void operator()(mirror::Object* obj) const
872      SHARED_REQUIRES(Locks::mutator_lock_) {
873    ObjectCallback(obj, collector_);
874  }
875  static void ObjectCallback(mirror::Object* obj, void *arg)
876      SHARED_REQUIRES(Locks::mutator_lock_) {
877    CHECK(obj != nullptr);
878    ConcurrentCopying* collector = reinterpret_cast<ConcurrentCopying*>(arg);
879    space::RegionSpace* region_space = collector->RegionSpace();
880    CHECK(!region_space->IsInFromSpace(obj)) << "Scanning object " << obj << " in from space";
881    collector->AssertToSpaceInvariant(nullptr, MemberOffset(0), obj);
882    ConcurrentCopyingAssertToSpaceInvariantFieldVisitor visitor(collector);
883    obj->VisitReferences(visitor, visitor);
884  }
885
886 private:
887  ConcurrentCopying* const collector_;
888};
889
890class RevokeThreadLocalMarkStackCheckpoint : public Closure {
891 public:
892  RevokeThreadLocalMarkStackCheckpoint(ConcurrentCopying* concurrent_copying,
893                                       bool disable_weak_ref_access)
894      : concurrent_copying_(concurrent_copying),
895        disable_weak_ref_access_(disable_weak_ref_access) {
896  }
897
898  virtual void Run(Thread* thread) OVERRIDE NO_THREAD_SAFETY_ANALYSIS {
899    // Note: self is not necessarily equal to thread since thread may be suspended.
900    Thread* self = Thread::Current();
901    CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc)
902        << thread->GetState() << " thread " << thread << " self " << self;
903    // Revoke thread local mark stacks.
904    accounting::AtomicStack<mirror::Object>* tl_mark_stack = thread->GetThreadLocalMarkStack();
905    if (tl_mark_stack != nullptr) {
906      MutexLock mu(self, concurrent_copying_->mark_stack_lock_);
907      concurrent_copying_->revoked_mark_stacks_.push_back(tl_mark_stack);
908      thread->SetThreadLocalMarkStack(nullptr);
909    }
910    // Disable weak ref access.
911    if (disable_weak_ref_access_) {
912      thread->SetWeakRefAccessEnabled(false);
913    }
914    // If thread is a running mutator, then act on behalf of the garbage collector.
915    // See the code in ThreadList::RunCheckpoint.
916    concurrent_copying_->GetBarrier().Pass(self);
917  }
918
919 private:
920  ConcurrentCopying* const concurrent_copying_;
921  const bool disable_weak_ref_access_;
922};
923
924void ConcurrentCopying::RevokeThreadLocalMarkStacks(bool disable_weak_ref_access) {
925  Thread* self = Thread::Current();
926  RevokeThreadLocalMarkStackCheckpoint check_point(this, disable_weak_ref_access);
927  ThreadList* thread_list = Runtime::Current()->GetThreadList();
928  gc_barrier_->Init(self, 0);
929  size_t barrier_count = thread_list->RunCheckpoint(&check_point);
930  // If there are no threads to wait which implys that all the checkpoint functions are finished,
931  // then no need to release the mutator lock.
932  if (barrier_count == 0) {
933    return;
934  }
935  Locks::mutator_lock_->SharedUnlock(self);
936  {
937    ScopedThreadStateChange tsc(self, kWaitingForCheckPointsToRun);
938    gc_barrier_->Increment(self, barrier_count);
939  }
940  Locks::mutator_lock_->SharedLock(self);
941}
942
943void ConcurrentCopying::RevokeThreadLocalMarkStack(Thread* thread) {
944  Thread* self = Thread::Current();
945  CHECK_EQ(self, thread);
946  accounting::AtomicStack<mirror::Object>* tl_mark_stack = thread->GetThreadLocalMarkStack();
947  if (tl_mark_stack != nullptr) {
948    CHECK(is_marking_);
949    MutexLock mu(self, mark_stack_lock_);
950    revoked_mark_stacks_.push_back(tl_mark_stack);
951    thread->SetThreadLocalMarkStack(nullptr);
952  }
953}
954
955void ConcurrentCopying::ProcessMarkStack() {
956  if (kVerboseMode) {
957    LOG(INFO) << "ProcessMarkStack. ";
958  }
959  bool empty_prev = false;
960  while (true) {
961    bool empty = ProcessMarkStackOnce();
962    if (empty_prev && empty) {
963      // Saw empty mark stack for a second time, done.
964      break;
965    }
966    empty_prev = empty;
967  }
968}
969
970bool ConcurrentCopying::ProcessMarkStackOnce() {
971  Thread* self = Thread::Current();
972  CHECK(thread_running_gc_ != nullptr);
973  CHECK(self == thread_running_gc_);
974  CHECK(self->GetThreadLocalMarkStack() == nullptr);
975  size_t count = 0;
976  MarkStackMode mark_stack_mode = mark_stack_mode_.LoadRelaxed();
977  if (mark_stack_mode == kMarkStackModeThreadLocal) {
978    // Process the thread-local mark stacks and the GC mark stack.
979    count += ProcessThreadLocalMarkStacks(false);
980    while (!gc_mark_stack_->IsEmpty()) {
981      mirror::Object* to_ref = gc_mark_stack_->PopBack();
982      ProcessMarkStackRef(to_ref);
983      ++count;
984    }
985    gc_mark_stack_->Reset();
986  } else if (mark_stack_mode == kMarkStackModeShared) {
987    // Process the shared GC mark stack with a lock.
988    {
989      MutexLock mu(self, mark_stack_lock_);
990      CHECK(revoked_mark_stacks_.empty());
991    }
992    while (true) {
993      std::vector<mirror::Object*> refs;
994      {
995        // Copy refs with lock. Note the number of refs should be small.
996        MutexLock mu(self, mark_stack_lock_);
997        if (gc_mark_stack_->IsEmpty()) {
998          break;
999        }
1000        for (StackReference<mirror::Object>* p = gc_mark_stack_->Begin();
1001             p != gc_mark_stack_->End(); ++p) {
1002          refs.push_back(p->AsMirrorPtr());
1003        }
1004        gc_mark_stack_->Reset();
1005      }
1006      for (mirror::Object* ref : refs) {
1007        ProcessMarkStackRef(ref);
1008        ++count;
1009      }
1010    }
1011  } else {
1012    CHECK_EQ(static_cast<uint32_t>(mark_stack_mode),
1013             static_cast<uint32_t>(kMarkStackModeGcExclusive));
1014    {
1015      MutexLock mu(self, mark_stack_lock_);
1016      CHECK(revoked_mark_stacks_.empty());
1017    }
1018    // Process the GC mark stack in the exclusive mode. No need to take the lock.
1019    while (!gc_mark_stack_->IsEmpty()) {
1020      mirror::Object* to_ref = gc_mark_stack_->PopBack();
1021      ProcessMarkStackRef(to_ref);
1022      ++count;
1023    }
1024    gc_mark_stack_->Reset();
1025  }
1026
1027  // Return true if the stack was empty.
1028  return count == 0;
1029}
1030
1031size_t ConcurrentCopying::ProcessThreadLocalMarkStacks(bool disable_weak_ref_access) {
1032  // Run a checkpoint to collect all thread local mark stacks and iterate over them all.
1033  RevokeThreadLocalMarkStacks(disable_weak_ref_access);
1034  size_t count = 0;
1035  std::vector<accounting::AtomicStack<mirror::Object>*> mark_stacks;
1036  {
1037    MutexLock mu(Thread::Current(), mark_stack_lock_);
1038    // Make a copy of the mark stack vector.
1039    mark_stacks = revoked_mark_stacks_;
1040    revoked_mark_stacks_.clear();
1041  }
1042  for (accounting::AtomicStack<mirror::Object>* mark_stack : mark_stacks) {
1043    for (StackReference<mirror::Object>* p = mark_stack->Begin(); p != mark_stack->End(); ++p) {
1044      mirror::Object* to_ref = p->AsMirrorPtr();
1045      ProcessMarkStackRef(to_ref);
1046      ++count;
1047    }
1048    {
1049      MutexLock mu(Thread::Current(), mark_stack_lock_);
1050      if (pooled_mark_stacks_.size() >= kMarkStackPoolSize) {
1051        // The pool has enough. Delete it.
1052        delete mark_stack;
1053      } else {
1054        // Otherwise, put it into the pool for later reuse.
1055        mark_stack->Reset();
1056        pooled_mark_stacks_.push_back(mark_stack);
1057      }
1058    }
1059  }
1060  return count;
1061}
1062
1063inline void ConcurrentCopying::ProcessMarkStackRef(mirror::Object* to_ref) {
1064  DCHECK(!region_space_->IsInFromSpace(to_ref));
1065  if (kUseBakerReadBarrier) {
1066    DCHECK(to_ref->GetReadBarrierPointer() == ReadBarrier::GrayPtr())
1067        << " " << to_ref << " " << to_ref->GetReadBarrierPointer()
1068        << " is_marked=" << IsMarked(to_ref);
1069  }
1070  // Scan ref fields.
1071  Scan(to_ref);
1072  // Mark the gray ref as white or black.
1073  if (kUseBakerReadBarrier) {
1074    DCHECK(to_ref->GetReadBarrierPointer() == ReadBarrier::GrayPtr())
1075        << " " << to_ref << " " << to_ref->GetReadBarrierPointer()
1076        << " is_marked=" << IsMarked(to_ref);
1077  }
1078#ifdef USE_BAKER_OR_BROOKS_READ_BARRIER
1079  if (UNLIKELY((to_ref->GetClass<kVerifyNone, kWithoutReadBarrier>()->IsTypeOfReferenceClass() &&
1080                to_ref->AsReference()->GetReferent<kWithoutReadBarrier>() != nullptr &&
1081                !IsInToSpace(to_ref->AsReference()->GetReferent<kWithoutReadBarrier>())))) {
1082    // Leave this Reference gray in the queue so that GetReferent() will trigger a read barrier. We
1083    // will change it to black or white later in ReferenceQueue::DequeuePendingReference().
1084    DCHECK(to_ref->AsReference()->GetPendingNext() != nullptr) << "Left unenqueued ref gray " << to_ref;
1085  } else {
1086    // We may occasionally leave a Reference black or white in the queue if its referent happens to
1087    // be concurrently marked after the Scan() call above has enqueued the Reference, in which case
1088    // the above IsInToSpace() evaluates to true and we change the color from gray to black or white
1089    // here in this else block.
1090    if (kUseBakerReadBarrier) {
1091      if (region_space_->IsInToSpace(to_ref)) {
1092        // If to-space, change from gray to white.
1093        bool success = to_ref->AtomicSetReadBarrierPointer</*kCasRelease*/true>(
1094            ReadBarrier::GrayPtr(),
1095            ReadBarrier::WhitePtr());
1096        DCHECK(success) << "Must succeed as we won the race.";
1097        DCHECK(to_ref->GetReadBarrierPointer() == ReadBarrier::WhitePtr());
1098      } else {
1099        // If non-moving space/unevac from space, change from gray
1100        // to black. We can't change gray to white because it's not
1101        // safe to use CAS if two threads change values in opposite
1102        // directions (A->B and B->A). So, we change it to black to
1103        // indicate non-moving objects that have been marked
1104        // through. Note we'd need to change from black to white
1105        // later (concurrently).
1106        bool success = to_ref->AtomicSetReadBarrierPointer</*kCasRelease*/true>(
1107            ReadBarrier::GrayPtr(),
1108            ReadBarrier::BlackPtr());
1109        DCHECK(success) << "Must succeed as we won the race.";
1110        DCHECK(to_ref->GetReadBarrierPointer() == ReadBarrier::BlackPtr());
1111      }
1112    }
1113  }
1114#else
1115  DCHECK(!kUseBakerReadBarrier);
1116#endif
1117  if (ReadBarrier::kEnableToSpaceInvariantChecks || kIsDebugBuild) {
1118    ConcurrentCopyingAssertToSpaceInvariantObjectVisitor visitor(this);
1119    visitor(to_ref);
1120  }
1121}
1122
1123void ConcurrentCopying::SwitchToSharedMarkStackMode() {
1124  Thread* self = Thread::Current();
1125  CHECK(thread_running_gc_ != nullptr);
1126  CHECK_EQ(self, thread_running_gc_);
1127  CHECK(self->GetThreadLocalMarkStack() == nullptr);
1128  MarkStackMode before_mark_stack_mode = mark_stack_mode_.LoadRelaxed();
1129  CHECK_EQ(static_cast<uint32_t>(before_mark_stack_mode),
1130           static_cast<uint32_t>(kMarkStackModeThreadLocal));
1131  mark_stack_mode_.StoreRelaxed(kMarkStackModeShared);
1132  CHECK(weak_ref_access_enabled_.LoadRelaxed());
1133  weak_ref_access_enabled_.StoreRelaxed(false);
1134  QuasiAtomic::ThreadFenceForConstructor();
1135  // Process the thread local mark stacks one last time after switching to the shared mark stack
1136  // mode and disable weak ref accesses.
1137  ProcessThreadLocalMarkStacks(true);
1138  if (kVerboseMode) {
1139    LOG(INFO) << "Switched to shared mark stack mode and disabled weak ref access";
1140  }
1141}
1142
1143void ConcurrentCopying::SwitchToGcExclusiveMarkStackMode() {
1144  Thread* self = Thread::Current();
1145  CHECK(thread_running_gc_ != nullptr);
1146  CHECK_EQ(self, thread_running_gc_);
1147  CHECK(self->GetThreadLocalMarkStack() == nullptr);
1148  MarkStackMode before_mark_stack_mode = mark_stack_mode_.LoadRelaxed();
1149  CHECK_EQ(static_cast<uint32_t>(before_mark_stack_mode),
1150           static_cast<uint32_t>(kMarkStackModeShared));
1151  mark_stack_mode_.StoreRelaxed(kMarkStackModeGcExclusive);
1152  QuasiAtomic::ThreadFenceForConstructor();
1153  if (kVerboseMode) {
1154    LOG(INFO) << "Switched to GC exclusive mark stack mode";
1155  }
1156}
1157
1158void ConcurrentCopying::CheckEmptyMarkStack() {
1159  Thread* self = Thread::Current();
1160  CHECK(thread_running_gc_ != nullptr);
1161  CHECK_EQ(self, thread_running_gc_);
1162  CHECK(self->GetThreadLocalMarkStack() == nullptr);
1163  MarkStackMode mark_stack_mode = mark_stack_mode_.LoadRelaxed();
1164  if (mark_stack_mode == kMarkStackModeThreadLocal) {
1165    // Thread-local mark stack mode.
1166    RevokeThreadLocalMarkStacks(false);
1167    MutexLock mu(Thread::Current(), mark_stack_lock_);
1168    if (!revoked_mark_stacks_.empty()) {
1169      for (accounting::AtomicStack<mirror::Object>* mark_stack : revoked_mark_stacks_) {
1170        while (!mark_stack->IsEmpty()) {
1171          mirror::Object* obj = mark_stack->PopBack();
1172          if (kUseBakerReadBarrier) {
1173            mirror::Object* rb_ptr = obj->GetReadBarrierPointer();
1174            LOG(INFO) << "On mark queue : " << obj << " " << PrettyTypeOf(obj) << " rb_ptr=" << rb_ptr
1175                      << " is_marked=" << IsMarked(obj);
1176          } else {
1177            LOG(INFO) << "On mark queue : " << obj << " " << PrettyTypeOf(obj)
1178                      << " is_marked=" << IsMarked(obj);
1179          }
1180        }
1181      }
1182      LOG(FATAL) << "mark stack is not empty";
1183    }
1184  } else {
1185    // Shared, GC-exclusive, or off.
1186    MutexLock mu(Thread::Current(), mark_stack_lock_);
1187    CHECK(gc_mark_stack_->IsEmpty());
1188    CHECK(revoked_mark_stacks_.empty());
1189  }
1190}
1191
1192void ConcurrentCopying::SweepSystemWeaks(Thread* self) {
1193  TimingLogger::ScopedTiming split("SweepSystemWeaks", GetTimings());
1194  ReaderMutexLock mu(self, *Locks::heap_bitmap_lock_);
1195  Runtime::Current()->SweepSystemWeaks(this);
1196}
1197
1198void ConcurrentCopying::Sweep(bool swap_bitmaps) {
1199  {
1200    TimingLogger::ScopedTiming t("MarkStackAsLive", GetTimings());
1201    accounting::ObjectStack* live_stack = heap_->GetLiveStack();
1202    if (kEnableFromSpaceAccountingCheck) {
1203      CHECK_GE(live_stack_freeze_size_, live_stack->Size());
1204    }
1205    heap_->MarkAllocStackAsLive(live_stack);
1206    live_stack->Reset();
1207  }
1208  CheckEmptyMarkStack();
1209  TimingLogger::ScopedTiming split("Sweep", GetTimings());
1210  for (const auto& space : GetHeap()->GetContinuousSpaces()) {
1211    if (space->IsContinuousMemMapAllocSpace()) {
1212      space::ContinuousMemMapAllocSpace* alloc_space = space->AsContinuousMemMapAllocSpace();
1213      if (space == region_space_ || immune_spaces_.ContainsSpace(space)) {
1214        continue;
1215      }
1216      TimingLogger::ScopedTiming split2(
1217          alloc_space->IsZygoteSpace() ? "SweepZygoteSpace" : "SweepAllocSpace", GetTimings());
1218      RecordFree(alloc_space->Sweep(swap_bitmaps));
1219    }
1220  }
1221  SweepLargeObjects(swap_bitmaps);
1222}
1223
1224void ConcurrentCopying::SweepLargeObjects(bool swap_bitmaps) {
1225  TimingLogger::ScopedTiming split("SweepLargeObjects", GetTimings());
1226  RecordFreeLOS(heap_->GetLargeObjectsSpace()->Sweep(swap_bitmaps));
1227}
1228
1229class ConcurrentCopyingClearBlackPtrsVisitor {
1230 public:
1231  explicit ConcurrentCopyingClearBlackPtrsVisitor(ConcurrentCopying* cc)
1232      : collector_(cc) {}
1233  void operator()(mirror::Object* obj) const SHARED_REQUIRES(Locks::mutator_lock_)
1234      SHARED_REQUIRES(Locks::heap_bitmap_lock_) {
1235    DCHECK(obj != nullptr);
1236    DCHECK(collector_->heap_->GetMarkBitmap()->Test(obj)) << obj;
1237    DCHECK_EQ(obj->GetReadBarrierPointer(), ReadBarrier::BlackPtr()) << obj;
1238    obj->AtomicSetReadBarrierPointer(ReadBarrier::BlackPtr(), ReadBarrier::WhitePtr());
1239    DCHECK_EQ(obj->GetReadBarrierPointer(), ReadBarrier::WhitePtr()) << obj;
1240  }
1241
1242 private:
1243  ConcurrentCopying* const collector_;
1244};
1245
1246// Clear the black ptrs in non-moving objects back to white.
1247void ConcurrentCopying::ClearBlackPtrs() {
1248  CHECK(kUseBakerReadBarrier);
1249  TimingLogger::ScopedTiming split("ClearBlackPtrs", GetTimings());
1250  ConcurrentCopyingClearBlackPtrsVisitor visitor(this);
1251  for (auto& space : heap_->GetContinuousSpaces()) {
1252    if (space == region_space_) {
1253      continue;
1254    }
1255    accounting::ContinuousSpaceBitmap* mark_bitmap = space->GetMarkBitmap();
1256    if (kVerboseMode) {
1257      LOG(INFO) << "ClearBlackPtrs: " << *space << " bitmap: " << *mark_bitmap;
1258    }
1259    mark_bitmap->VisitMarkedRange(reinterpret_cast<uintptr_t>(space->Begin()),
1260                                  reinterpret_cast<uintptr_t>(space->Limit()),
1261                                  visitor);
1262  }
1263  space::LargeObjectSpace* large_object_space = heap_->GetLargeObjectsSpace();
1264  large_object_space->GetMarkBitmap()->VisitMarkedRange(
1265      reinterpret_cast<uintptr_t>(large_object_space->Begin()),
1266      reinterpret_cast<uintptr_t>(large_object_space->End()),
1267      visitor);
1268  // Objects on the allocation stack?
1269  if (ReadBarrier::kEnableReadBarrierInvariantChecks || kIsDebugBuild) {
1270    size_t count = GetAllocationStack()->Size();
1271    auto* it = GetAllocationStack()->Begin();
1272    auto* end = GetAllocationStack()->End();
1273    for (size_t i = 0; i < count; ++i, ++it) {
1274      CHECK_LT(it, end);
1275      mirror::Object* obj = it->AsMirrorPtr();
1276      if (obj != nullptr) {
1277        // Must have been cleared above.
1278        CHECK_EQ(obj->GetReadBarrierPointer(), ReadBarrier::WhitePtr()) << obj;
1279      }
1280    }
1281  }
1282}
1283
1284void ConcurrentCopying::ReclaimPhase() {
1285  TimingLogger::ScopedTiming split("ReclaimPhase", GetTimings());
1286  if (kVerboseMode) {
1287    LOG(INFO) << "GC ReclaimPhase";
1288  }
1289  Thread* self = Thread::Current();
1290
1291  {
1292    // Double-check that the mark stack is empty.
1293    // Note: need to set this after VerifyNoFromSpaceRef().
1294    is_asserting_to_space_invariant_ = false;
1295    QuasiAtomic::ThreadFenceForConstructor();
1296    if (kVerboseMode) {
1297      LOG(INFO) << "Issue an empty check point. ";
1298    }
1299    IssueEmptyCheckpoint();
1300    // Disable the check.
1301    is_mark_stack_push_disallowed_.StoreSequentiallyConsistent(0);
1302    CheckEmptyMarkStack();
1303  }
1304
1305  {
1306    // Record freed objects.
1307    TimingLogger::ScopedTiming split2("RecordFree", GetTimings());
1308    // Don't include thread-locals that are in the to-space.
1309    uint64_t from_bytes = region_space_->GetBytesAllocatedInFromSpace();
1310    uint64_t from_objects = region_space_->GetObjectsAllocatedInFromSpace();
1311    uint64_t unevac_from_bytes = region_space_->GetBytesAllocatedInUnevacFromSpace();
1312    uint64_t unevac_from_objects = region_space_->GetObjectsAllocatedInUnevacFromSpace();
1313    uint64_t to_bytes = bytes_moved_.LoadSequentiallyConsistent();
1314    uint64_t to_objects = objects_moved_.LoadSequentiallyConsistent();
1315    if (kEnableFromSpaceAccountingCheck) {
1316      CHECK_EQ(from_space_num_objects_at_first_pause_, from_objects + unevac_from_objects);
1317      CHECK_EQ(from_space_num_bytes_at_first_pause_, from_bytes + unevac_from_bytes);
1318    }
1319    CHECK_LE(to_objects, from_objects);
1320    CHECK_LE(to_bytes, from_bytes);
1321    int64_t freed_bytes = from_bytes - to_bytes;
1322    int64_t freed_objects = from_objects - to_objects;
1323    if (kVerboseMode) {
1324      LOG(INFO) << "RecordFree:"
1325                << " from_bytes=" << from_bytes << " from_objects=" << from_objects
1326                << " unevac_from_bytes=" << unevac_from_bytes << " unevac_from_objects=" << unevac_from_objects
1327                << " to_bytes=" << to_bytes << " to_objects=" << to_objects
1328                << " freed_bytes=" << freed_bytes << " freed_objects=" << freed_objects
1329                << " from_space size=" << region_space_->FromSpaceSize()
1330                << " unevac_from_space size=" << region_space_->UnevacFromSpaceSize()
1331                << " to_space size=" << region_space_->ToSpaceSize();
1332      LOG(INFO) << "(before) num_bytes_allocated=" << heap_->num_bytes_allocated_.LoadSequentiallyConsistent();
1333    }
1334    RecordFree(ObjectBytePair(freed_objects, freed_bytes));
1335    if (kVerboseMode) {
1336      LOG(INFO) << "(after) num_bytes_allocated=" << heap_->num_bytes_allocated_.LoadSequentiallyConsistent();
1337    }
1338  }
1339
1340  {
1341    TimingLogger::ScopedTiming split3("ComputeUnevacFromSpaceLiveRatio", GetTimings());
1342    ComputeUnevacFromSpaceLiveRatio();
1343  }
1344
1345  {
1346    TimingLogger::ScopedTiming split4("ClearFromSpace", GetTimings());
1347    region_space_->ClearFromSpace();
1348  }
1349
1350  {
1351    WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
1352    if (kUseBakerReadBarrier) {
1353      ClearBlackPtrs();
1354    }
1355    Sweep(false);
1356    SwapBitmaps();
1357    heap_->UnBindBitmaps();
1358
1359    // Remove bitmaps for the immune spaces.
1360    while (!cc_bitmaps_.empty()) {
1361      accounting::ContinuousSpaceBitmap* cc_bitmap = cc_bitmaps_.back();
1362      cc_heap_bitmap_->RemoveContinuousSpaceBitmap(cc_bitmap);
1363      delete cc_bitmap;
1364      cc_bitmaps_.pop_back();
1365    }
1366    region_space_bitmap_ = nullptr;
1367  }
1368
1369  CheckEmptyMarkStack();
1370
1371  if (kVerboseMode) {
1372    LOG(INFO) << "GC end of ReclaimPhase";
1373  }
1374}
1375
1376class ConcurrentCopyingComputeUnevacFromSpaceLiveRatioVisitor {
1377 public:
1378  explicit ConcurrentCopyingComputeUnevacFromSpaceLiveRatioVisitor(ConcurrentCopying* cc)
1379      : collector_(cc) {}
1380  void operator()(mirror::Object* ref) const SHARED_REQUIRES(Locks::mutator_lock_)
1381      SHARED_REQUIRES(Locks::heap_bitmap_lock_) {
1382    DCHECK(ref != nullptr);
1383    DCHECK(collector_->region_space_bitmap_->Test(ref)) << ref;
1384    DCHECK(collector_->region_space_->IsInUnevacFromSpace(ref)) << ref;
1385    if (kUseBakerReadBarrier) {
1386      DCHECK_EQ(ref->GetReadBarrierPointer(), ReadBarrier::BlackPtr()) << ref;
1387      // Clear the black ptr.
1388      ref->AtomicSetReadBarrierPointer(ReadBarrier::BlackPtr(), ReadBarrier::WhitePtr());
1389      DCHECK_EQ(ref->GetReadBarrierPointer(), ReadBarrier::WhitePtr()) << ref;
1390    }
1391    size_t obj_size = ref->SizeOf();
1392    size_t alloc_size = RoundUp(obj_size, space::RegionSpace::kAlignment);
1393    collector_->region_space_->AddLiveBytes(ref, alloc_size);
1394  }
1395
1396 private:
1397  ConcurrentCopying* const collector_;
1398};
1399
1400// Compute how much live objects are left in regions.
1401void ConcurrentCopying::ComputeUnevacFromSpaceLiveRatio() {
1402  region_space_->AssertAllRegionLiveBytesZeroOrCleared();
1403  ConcurrentCopyingComputeUnevacFromSpaceLiveRatioVisitor visitor(this);
1404  region_space_bitmap_->VisitMarkedRange(reinterpret_cast<uintptr_t>(region_space_->Begin()),
1405                                         reinterpret_cast<uintptr_t>(region_space_->Limit()),
1406                                         visitor);
1407}
1408
1409// Assert the to-space invariant.
1410void ConcurrentCopying::AssertToSpaceInvariant(mirror::Object* obj, MemberOffset offset,
1411                                               mirror::Object* ref) {
1412  CHECK(heap_->collector_type_ == kCollectorTypeCC) << static_cast<size_t>(heap_->collector_type_);
1413  if (is_asserting_to_space_invariant_) {
1414    if (region_space_->IsInToSpace(ref)) {
1415      // OK.
1416      return;
1417    } else if (region_space_->IsInUnevacFromSpace(ref)) {
1418      CHECK(region_space_bitmap_->Test(ref)) << ref;
1419    } else if (region_space_->IsInFromSpace(ref)) {
1420      // Not OK. Do extra logging.
1421      if (obj != nullptr) {
1422        LogFromSpaceRefHolder(obj, offset);
1423      }
1424      ref->GetLockWord(false).Dump(LOG(INTERNAL_FATAL));
1425      CHECK(false) << "Found from-space ref " << ref << " " << PrettyTypeOf(ref);
1426    } else {
1427      AssertToSpaceInvariantInNonMovingSpace(obj, ref);
1428    }
1429  }
1430}
1431
1432class RootPrinter {
1433 public:
1434  RootPrinter() { }
1435
1436  template <class MirrorType>
1437  ALWAYS_INLINE void VisitRootIfNonNull(mirror::CompressedReference<MirrorType>* root)
1438      SHARED_REQUIRES(Locks::mutator_lock_) {
1439    if (!root->IsNull()) {
1440      VisitRoot(root);
1441    }
1442  }
1443
1444  template <class MirrorType>
1445  void VisitRoot(mirror::Object** root)
1446      SHARED_REQUIRES(Locks::mutator_lock_) {
1447    LOG(INTERNAL_FATAL) << "root=" << root << " ref=" << *root;
1448  }
1449
1450  template <class MirrorType>
1451  void VisitRoot(mirror::CompressedReference<MirrorType>* root)
1452      SHARED_REQUIRES(Locks::mutator_lock_) {
1453    LOG(INTERNAL_FATAL) << "root=" << root << " ref=" << root->AsMirrorPtr();
1454  }
1455};
1456
1457void ConcurrentCopying::AssertToSpaceInvariant(GcRootSource* gc_root_source,
1458                                               mirror::Object* ref) {
1459  CHECK(heap_->collector_type_ == kCollectorTypeCC) << static_cast<size_t>(heap_->collector_type_);
1460  if (is_asserting_to_space_invariant_) {
1461    if (region_space_->IsInToSpace(ref)) {
1462      // OK.
1463      return;
1464    } else if (region_space_->IsInUnevacFromSpace(ref)) {
1465      CHECK(region_space_bitmap_->Test(ref)) << ref;
1466    } else if (region_space_->IsInFromSpace(ref)) {
1467      // Not OK. Do extra logging.
1468      if (gc_root_source == nullptr) {
1469        // No info.
1470      } else if (gc_root_source->HasArtField()) {
1471        ArtField* field = gc_root_source->GetArtField();
1472        LOG(INTERNAL_FATAL) << "gc root in field " << field << " " << PrettyField(field);
1473        RootPrinter root_printer;
1474        field->VisitRoots(root_printer);
1475      } else if (gc_root_source->HasArtMethod()) {
1476        ArtMethod* method = gc_root_source->GetArtMethod();
1477        LOG(INTERNAL_FATAL) << "gc root in method " << method << " " << PrettyMethod(method);
1478        RootPrinter root_printer;
1479        method->VisitRoots(root_printer, sizeof(void*));
1480      }
1481      ref->GetLockWord(false).Dump(LOG(INTERNAL_FATAL));
1482      region_space_->DumpNonFreeRegions(LOG(INTERNAL_FATAL));
1483      PrintFileToLog("/proc/self/maps", LogSeverity::INTERNAL_FATAL);
1484      MemMap::DumpMaps(LOG(INTERNAL_FATAL), true);
1485      CHECK(false) << "Found from-space ref " << ref << " " << PrettyTypeOf(ref);
1486    } else {
1487      AssertToSpaceInvariantInNonMovingSpace(nullptr, ref);
1488    }
1489  }
1490}
1491
1492void ConcurrentCopying::LogFromSpaceRefHolder(mirror::Object* obj, MemberOffset offset) {
1493  if (kUseBakerReadBarrier) {
1494    LOG(INFO) << "holder=" << obj << " " << PrettyTypeOf(obj)
1495              << " holder rb_ptr=" << obj->GetReadBarrierPointer();
1496  } else {
1497    LOG(INFO) << "holder=" << obj << " " << PrettyTypeOf(obj);
1498  }
1499  if (region_space_->IsInFromSpace(obj)) {
1500    LOG(INFO) << "holder is in the from-space.";
1501  } else if (region_space_->IsInToSpace(obj)) {
1502    LOG(INFO) << "holder is in the to-space.";
1503  } else if (region_space_->IsInUnevacFromSpace(obj)) {
1504    LOG(INFO) << "holder is in the unevac from-space.";
1505    if (region_space_bitmap_->Test(obj)) {
1506      LOG(INFO) << "holder is marked in the region space bitmap.";
1507    } else {
1508      LOG(INFO) << "holder is not marked in the region space bitmap.";
1509    }
1510  } else {
1511    // In a non-moving space.
1512    if (immune_spaces_.ContainsObject(obj)) {
1513      LOG(INFO) << "holder is in an immune image or the zygote space.";
1514      accounting::ContinuousSpaceBitmap* cc_bitmap =
1515          cc_heap_bitmap_->GetContinuousSpaceBitmap(obj);
1516      CHECK(cc_bitmap != nullptr)
1517          << "An immune space object must have a bitmap.";
1518      if (cc_bitmap->Test(obj)) {
1519        LOG(INFO) << "holder is marked in the bit map.";
1520      } else {
1521        LOG(INFO) << "holder is NOT marked in the bit map.";
1522      }
1523    } else {
1524      LOG(INFO) << "holder is in a non-immune, non-moving (or main) space.";
1525      accounting::ContinuousSpaceBitmap* mark_bitmap =
1526          heap_mark_bitmap_->GetContinuousSpaceBitmap(obj);
1527      accounting::LargeObjectBitmap* los_bitmap =
1528          heap_mark_bitmap_->GetLargeObjectBitmap(obj);
1529      CHECK(los_bitmap != nullptr) << "LOS bitmap covers the entire address range";
1530      bool is_los = mark_bitmap == nullptr;
1531      if (!is_los && mark_bitmap->Test(obj)) {
1532        LOG(INFO) << "holder is marked in the mark bit map.";
1533      } else if (is_los && los_bitmap->Test(obj)) {
1534        LOG(INFO) << "holder is marked in the los bit map.";
1535      } else {
1536        // If ref is on the allocation stack, then it is considered
1537        // mark/alive (but not necessarily on the live stack.)
1538        if (IsOnAllocStack(obj)) {
1539          LOG(INFO) << "holder is on the alloc stack.";
1540        } else {
1541          LOG(INFO) << "holder is not marked or on the alloc stack.";
1542        }
1543      }
1544    }
1545  }
1546  LOG(INFO) << "offset=" << offset.SizeValue();
1547}
1548
1549void ConcurrentCopying::AssertToSpaceInvariantInNonMovingSpace(mirror::Object* obj,
1550                                                               mirror::Object* ref) {
1551  // In a non-moving spaces. Check that the ref is marked.
1552  if (immune_spaces_.ContainsObject(ref)) {
1553    accounting::ContinuousSpaceBitmap* cc_bitmap =
1554        cc_heap_bitmap_->GetContinuousSpaceBitmap(ref);
1555    CHECK(cc_bitmap != nullptr)
1556        << "An immune space ref must have a bitmap. " << ref;
1557    if (kUseBakerReadBarrier) {
1558      CHECK(cc_bitmap->Test(ref))
1559          << "Unmarked immune space ref. obj=" << obj << " rb_ptr="
1560          << obj->GetReadBarrierPointer() << " ref=" << ref;
1561    } else {
1562      CHECK(cc_bitmap->Test(ref))
1563          << "Unmarked immune space ref. obj=" << obj << " ref=" << ref;
1564    }
1565  } else {
1566    accounting::ContinuousSpaceBitmap* mark_bitmap =
1567        heap_mark_bitmap_->GetContinuousSpaceBitmap(ref);
1568    accounting::LargeObjectBitmap* los_bitmap =
1569        heap_mark_bitmap_->GetLargeObjectBitmap(ref);
1570    CHECK(los_bitmap != nullptr) << "LOS bitmap covers the entire address range";
1571    bool is_los = mark_bitmap == nullptr;
1572    if ((!is_los && mark_bitmap->Test(ref)) ||
1573        (is_los && los_bitmap->Test(ref))) {
1574      // OK.
1575    } else {
1576      // If ref is on the allocation stack, then it may not be
1577      // marked live, but considered marked/alive (but not
1578      // necessarily on the live stack).
1579      CHECK(IsOnAllocStack(ref)) << "Unmarked ref that's not on the allocation stack. "
1580                                 << "obj=" << obj << " ref=" << ref;
1581    }
1582  }
1583}
1584
1585// Used to scan ref fields of an object.
1586class ConcurrentCopyingRefFieldsVisitor {
1587 public:
1588  explicit ConcurrentCopyingRefFieldsVisitor(ConcurrentCopying* collector)
1589      : collector_(collector) {}
1590
1591  void operator()(mirror::Object* obj, MemberOffset offset, bool /* is_static */)
1592      const ALWAYS_INLINE SHARED_REQUIRES(Locks::mutator_lock_)
1593      SHARED_REQUIRES(Locks::heap_bitmap_lock_) {
1594    collector_->Process(obj, offset);
1595  }
1596
1597  void operator()(mirror::Class* klass, mirror::Reference* ref) const
1598      SHARED_REQUIRES(Locks::mutator_lock_) ALWAYS_INLINE {
1599    CHECK(klass->IsTypeOfReferenceClass());
1600    collector_->DelayReferenceReferent(klass, ref);
1601  }
1602
1603  void VisitRootIfNonNull(mirror::CompressedReference<mirror::Object>* root) const
1604      ALWAYS_INLINE
1605      SHARED_REQUIRES(Locks::mutator_lock_) {
1606    if (!root->IsNull()) {
1607      VisitRoot(root);
1608    }
1609  }
1610
1611  void VisitRoot(mirror::CompressedReference<mirror::Object>* root) const
1612      ALWAYS_INLINE
1613      SHARED_REQUIRES(Locks::mutator_lock_) {
1614    collector_->MarkRoot(root);
1615  }
1616
1617 private:
1618  ConcurrentCopying* const collector_;
1619};
1620
1621// Scan ref fields of an object.
1622inline void ConcurrentCopying::Scan(mirror::Object* to_ref) {
1623  DCHECK(!region_space_->IsInFromSpace(to_ref));
1624  ConcurrentCopyingRefFieldsVisitor visitor(this);
1625  // Disable the read barrier for a performance reason.
1626  to_ref->VisitReferences</*kVisitNativeRoots*/true, kDefaultVerifyFlags, kWithoutReadBarrier>(
1627      visitor, visitor);
1628}
1629
1630// Process a field.
1631inline void ConcurrentCopying::Process(mirror::Object* obj, MemberOffset offset) {
1632  mirror::Object* ref = obj->GetFieldObject<
1633      mirror::Object, kVerifyNone, kWithoutReadBarrier, false>(offset);
1634  mirror::Object* to_ref = Mark(ref);
1635  if (to_ref == ref) {
1636    return;
1637  }
1638  // This may fail if the mutator writes to the field at the same time. But it's ok.
1639  mirror::Object* expected_ref = ref;
1640  mirror::Object* new_ref = to_ref;
1641  do {
1642    if (expected_ref !=
1643        obj->GetFieldObject<mirror::Object, kVerifyNone, kWithoutReadBarrier, false>(offset)) {
1644      // It was updated by the mutator.
1645      break;
1646    }
1647  } while (!obj->CasFieldWeakRelaxedObjectWithoutWriteBarrier<
1648      false, false, kVerifyNone>(offset, expected_ref, new_ref));
1649}
1650
1651// Process some roots.
1652inline void ConcurrentCopying::VisitRoots(
1653    mirror::Object*** roots, size_t count, const RootInfo& info ATTRIBUTE_UNUSED) {
1654  for (size_t i = 0; i < count; ++i) {
1655    mirror::Object** root = roots[i];
1656    mirror::Object* ref = *root;
1657    mirror::Object* to_ref = Mark(ref);
1658    if (to_ref == ref) {
1659      continue;
1660    }
1661    Atomic<mirror::Object*>* addr = reinterpret_cast<Atomic<mirror::Object*>*>(root);
1662    mirror::Object* expected_ref = ref;
1663    mirror::Object* new_ref = to_ref;
1664    do {
1665      if (expected_ref != addr->LoadRelaxed()) {
1666        // It was updated by the mutator.
1667        break;
1668      }
1669    } while (!addr->CompareExchangeWeakRelaxed(expected_ref, new_ref));
1670  }
1671}
1672
1673inline void ConcurrentCopying::MarkRoot(mirror::CompressedReference<mirror::Object>* root) {
1674  DCHECK(!root->IsNull());
1675  mirror::Object* const ref = root->AsMirrorPtr();
1676  mirror::Object* to_ref = Mark(ref);
1677  if (to_ref != ref) {
1678    auto* addr = reinterpret_cast<Atomic<mirror::CompressedReference<mirror::Object>>*>(root);
1679    auto expected_ref = mirror::CompressedReference<mirror::Object>::FromMirrorPtr(ref);
1680    auto new_ref = mirror::CompressedReference<mirror::Object>::FromMirrorPtr(to_ref);
1681    // If the cas fails, then it was updated by the mutator.
1682    do {
1683      if (ref != addr->LoadRelaxed().AsMirrorPtr()) {
1684        // It was updated by the mutator.
1685        break;
1686      }
1687    } while (!addr->CompareExchangeWeakRelaxed(expected_ref, new_ref));
1688  }
1689}
1690
1691inline void ConcurrentCopying::VisitRoots(
1692    mirror::CompressedReference<mirror::Object>** roots, size_t count,
1693    const RootInfo& info ATTRIBUTE_UNUSED) {
1694  for (size_t i = 0; i < count; ++i) {
1695    mirror::CompressedReference<mirror::Object>* const root = roots[i];
1696    if (!root->IsNull()) {
1697      MarkRoot(root);
1698    }
1699  }
1700}
1701
1702// Fill the given memory block with a dummy object. Used to fill in a
1703// copy of objects that was lost in race.
1704void ConcurrentCopying::FillWithDummyObject(mirror::Object* dummy_obj, size_t byte_size) {
1705  CHECK_ALIGNED(byte_size, kObjectAlignment);
1706  memset(dummy_obj, 0, byte_size);
1707  mirror::Class* int_array_class = mirror::IntArray::GetArrayClass();
1708  CHECK(int_array_class != nullptr);
1709  AssertToSpaceInvariant(nullptr, MemberOffset(0), int_array_class);
1710  size_t component_size = int_array_class->GetComponentSize();
1711  CHECK_EQ(component_size, sizeof(int32_t));
1712  size_t data_offset = mirror::Array::DataOffset(component_size).SizeValue();
1713  if (data_offset > byte_size) {
1714    // An int array is too big. Use java.lang.Object.
1715    mirror::Class* java_lang_Object = WellKnownClasses::ToClass(WellKnownClasses::java_lang_Object);
1716    AssertToSpaceInvariant(nullptr, MemberOffset(0), java_lang_Object);
1717    CHECK_EQ(byte_size, java_lang_Object->GetObjectSize());
1718    dummy_obj->SetClass(java_lang_Object);
1719    CHECK_EQ(byte_size, dummy_obj->SizeOf());
1720  } else {
1721    // Use an int array.
1722    dummy_obj->SetClass(int_array_class);
1723    CHECK(dummy_obj->IsArrayInstance());
1724    int32_t length = (byte_size - data_offset) / component_size;
1725    dummy_obj->AsArray()->SetLength(length);
1726    CHECK_EQ(dummy_obj->AsArray()->GetLength(), length)
1727        << "byte_size=" << byte_size << " length=" << length
1728        << " component_size=" << component_size << " data_offset=" << data_offset;
1729    CHECK_EQ(byte_size, dummy_obj->SizeOf())
1730        << "byte_size=" << byte_size << " length=" << length
1731        << " component_size=" << component_size << " data_offset=" << data_offset;
1732  }
1733}
1734
1735// Reuse the memory blocks that were copy of objects that were lost in race.
1736mirror::Object* ConcurrentCopying::AllocateInSkippedBlock(size_t alloc_size) {
1737  // Try to reuse the blocks that were unused due to CAS failures.
1738  CHECK_ALIGNED(alloc_size, space::RegionSpace::kAlignment);
1739  Thread* self = Thread::Current();
1740  size_t min_object_size = RoundUp(sizeof(mirror::Object), space::RegionSpace::kAlignment);
1741  MutexLock mu(self, skipped_blocks_lock_);
1742  auto it = skipped_blocks_map_.lower_bound(alloc_size);
1743  if (it == skipped_blocks_map_.end()) {
1744    // Not found.
1745    return nullptr;
1746  }
1747  {
1748    size_t byte_size = it->first;
1749    CHECK_GE(byte_size, alloc_size);
1750    if (byte_size > alloc_size && byte_size - alloc_size < min_object_size) {
1751      // If remainder would be too small for a dummy object, retry with a larger request size.
1752      it = skipped_blocks_map_.lower_bound(alloc_size + min_object_size);
1753      if (it == skipped_blocks_map_.end()) {
1754        // Not found.
1755        return nullptr;
1756      }
1757      CHECK_ALIGNED(it->first - alloc_size, space::RegionSpace::kAlignment);
1758      CHECK_GE(it->first - alloc_size, min_object_size)
1759          << "byte_size=" << byte_size << " it->first=" << it->first << " alloc_size=" << alloc_size;
1760    }
1761  }
1762  // Found a block.
1763  CHECK(it != skipped_blocks_map_.end());
1764  size_t byte_size = it->first;
1765  uint8_t* addr = it->second;
1766  CHECK_GE(byte_size, alloc_size);
1767  CHECK(region_space_->IsInToSpace(reinterpret_cast<mirror::Object*>(addr)));
1768  CHECK_ALIGNED(byte_size, space::RegionSpace::kAlignment);
1769  if (kVerboseMode) {
1770    LOG(INFO) << "Reusing skipped bytes : " << reinterpret_cast<void*>(addr) << ", " << byte_size;
1771  }
1772  skipped_blocks_map_.erase(it);
1773  memset(addr, 0, byte_size);
1774  if (byte_size > alloc_size) {
1775    // Return the remainder to the map.
1776    CHECK_ALIGNED(byte_size - alloc_size, space::RegionSpace::kAlignment);
1777    CHECK_GE(byte_size - alloc_size, min_object_size);
1778    FillWithDummyObject(reinterpret_cast<mirror::Object*>(addr + alloc_size),
1779                        byte_size - alloc_size);
1780    CHECK(region_space_->IsInToSpace(reinterpret_cast<mirror::Object*>(addr + alloc_size)));
1781    skipped_blocks_map_.insert(std::make_pair(byte_size - alloc_size, addr + alloc_size));
1782  }
1783  return reinterpret_cast<mirror::Object*>(addr);
1784}
1785
1786mirror::Object* ConcurrentCopying::Copy(mirror::Object* from_ref) {
1787  DCHECK(region_space_->IsInFromSpace(from_ref));
1788  // No read barrier to avoid nested RB that might violate the to-space
1789  // invariant. Note that from_ref is a from space ref so the SizeOf()
1790  // call will access the from-space meta objects, but it's ok and necessary.
1791  size_t obj_size = from_ref->SizeOf<kDefaultVerifyFlags, kWithoutReadBarrier>();
1792  size_t region_space_alloc_size = RoundUp(obj_size, space::RegionSpace::kAlignment);
1793  size_t region_space_bytes_allocated = 0U;
1794  size_t non_moving_space_bytes_allocated = 0U;
1795  size_t bytes_allocated = 0U;
1796  size_t dummy;
1797  mirror::Object* to_ref = region_space_->AllocNonvirtual<true>(
1798      region_space_alloc_size, &region_space_bytes_allocated, nullptr, &dummy);
1799  bytes_allocated = region_space_bytes_allocated;
1800  if (to_ref != nullptr) {
1801    DCHECK_EQ(region_space_alloc_size, region_space_bytes_allocated);
1802  }
1803  bool fall_back_to_non_moving = false;
1804  if (UNLIKELY(to_ref == nullptr)) {
1805    // Failed to allocate in the region space. Try the skipped blocks.
1806    to_ref = AllocateInSkippedBlock(region_space_alloc_size);
1807    if (to_ref != nullptr) {
1808      // Succeeded to allocate in a skipped block.
1809      if (heap_->use_tlab_) {
1810        // This is necessary for the tlab case as it's not accounted in the space.
1811        region_space_->RecordAlloc(to_ref);
1812      }
1813      bytes_allocated = region_space_alloc_size;
1814    } else {
1815      // Fall back to the non-moving space.
1816      fall_back_to_non_moving = true;
1817      if (kVerboseMode) {
1818        LOG(INFO) << "Out of memory in the to-space. Fall back to non-moving. skipped_bytes="
1819                  << to_space_bytes_skipped_.LoadSequentiallyConsistent()
1820                  << " skipped_objects=" << to_space_objects_skipped_.LoadSequentiallyConsistent();
1821      }
1822      fall_back_to_non_moving = true;
1823      to_ref = heap_->non_moving_space_->Alloc(Thread::Current(), obj_size,
1824                                               &non_moving_space_bytes_allocated, nullptr, &dummy);
1825      CHECK(to_ref != nullptr) << "Fall-back non-moving space allocation failed";
1826      bytes_allocated = non_moving_space_bytes_allocated;
1827      // Mark it in the mark bitmap.
1828      accounting::ContinuousSpaceBitmap* mark_bitmap =
1829          heap_mark_bitmap_->GetContinuousSpaceBitmap(to_ref);
1830      CHECK(mark_bitmap != nullptr);
1831      CHECK(!mark_bitmap->AtomicTestAndSet(to_ref));
1832    }
1833  }
1834  DCHECK(to_ref != nullptr);
1835
1836  // Attempt to install the forward pointer. This is in a loop as the
1837  // lock word atomic write can fail.
1838  while (true) {
1839    // Copy the object. TODO: copy only the lockword in the second iteration and on?
1840    memcpy(to_ref, from_ref, obj_size);
1841
1842    LockWord old_lock_word = to_ref->GetLockWord(false);
1843
1844    if (old_lock_word.GetState() == LockWord::kForwardingAddress) {
1845      // Lost the race. Another thread (either GC or mutator) stored
1846      // the forwarding pointer first. Make the lost copy (to_ref)
1847      // look like a valid but dead (dummy) object and keep it for
1848      // future reuse.
1849      FillWithDummyObject(to_ref, bytes_allocated);
1850      if (!fall_back_to_non_moving) {
1851        DCHECK(region_space_->IsInToSpace(to_ref));
1852        if (bytes_allocated > space::RegionSpace::kRegionSize) {
1853          // Free the large alloc.
1854          region_space_->FreeLarge(to_ref, bytes_allocated);
1855        } else {
1856          // Record the lost copy for later reuse.
1857          heap_->num_bytes_allocated_.FetchAndAddSequentiallyConsistent(bytes_allocated);
1858          to_space_bytes_skipped_.FetchAndAddSequentiallyConsistent(bytes_allocated);
1859          to_space_objects_skipped_.FetchAndAddSequentiallyConsistent(1);
1860          MutexLock mu(Thread::Current(), skipped_blocks_lock_);
1861          skipped_blocks_map_.insert(std::make_pair(bytes_allocated,
1862                                                    reinterpret_cast<uint8_t*>(to_ref)));
1863        }
1864      } else {
1865        DCHECK(heap_->non_moving_space_->HasAddress(to_ref));
1866        DCHECK_EQ(bytes_allocated, non_moving_space_bytes_allocated);
1867        // Free the non-moving-space chunk.
1868        accounting::ContinuousSpaceBitmap* mark_bitmap =
1869            heap_mark_bitmap_->GetContinuousSpaceBitmap(to_ref);
1870        CHECK(mark_bitmap != nullptr);
1871        CHECK(mark_bitmap->Clear(to_ref));
1872        heap_->non_moving_space_->Free(Thread::Current(), to_ref);
1873      }
1874
1875      // Get the winner's forward ptr.
1876      mirror::Object* lost_fwd_ptr = to_ref;
1877      to_ref = reinterpret_cast<mirror::Object*>(old_lock_word.ForwardingAddress());
1878      CHECK(to_ref != nullptr);
1879      CHECK_NE(to_ref, lost_fwd_ptr);
1880      CHECK(region_space_->IsInToSpace(to_ref) || heap_->non_moving_space_->HasAddress(to_ref));
1881      CHECK_NE(to_ref->GetLockWord(false).GetState(), LockWord::kForwardingAddress);
1882      return to_ref;
1883    }
1884
1885    // Set the gray ptr.
1886    if (kUseBakerReadBarrier) {
1887      to_ref->SetReadBarrierPointer(ReadBarrier::GrayPtr());
1888    }
1889
1890    LockWord new_lock_word = LockWord::FromForwardingAddress(reinterpret_cast<size_t>(to_ref));
1891
1892    // Try to atomically write the fwd ptr.
1893    bool success = from_ref->CasLockWordWeakSequentiallyConsistent(old_lock_word, new_lock_word);
1894    if (LIKELY(success)) {
1895      // The CAS succeeded.
1896      objects_moved_.FetchAndAddSequentiallyConsistent(1);
1897      bytes_moved_.FetchAndAddSequentiallyConsistent(region_space_alloc_size);
1898      if (LIKELY(!fall_back_to_non_moving)) {
1899        DCHECK(region_space_->IsInToSpace(to_ref));
1900      } else {
1901        DCHECK(heap_->non_moving_space_->HasAddress(to_ref));
1902        DCHECK_EQ(bytes_allocated, non_moving_space_bytes_allocated);
1903      }
1904      if (kUseBakerReadBarrier) {
1905        DCHECK(to_ref->GetReadBarrierPointer() == ReadBarrier::GrayPtr());
1906      }
1907      DCHECK(GetFwdPtr(from_ref) == to_ref);
1908      CHECK_NE(to_ref->GetLockWord(false).GetState(), LockWord::kForwardingAddress);
1909      PushOntoMarkStack(to_ref);
1910      return to_ref;
1911    } else {
1912      // The CAS failed. It may have lost the race or may have failed
1913      // due to monitor/hashcode ops. Either way, retry.
1914    }
1915  }
1916}
1917
1918mirror::Object* ConcurrentCopying::IsMarked(mirror::Object* from_ref) {
1919  DCHECK(from_ref != nullptr);
1920  space::RegionSpace::RegionType rtype = region_space_->GetRegionType(from_ref);
1921  if (rtype == space::RegionSpace::RegionType::kRegionTypeToSpace) {
1922    // It's already marked.
1923    return from_ref;
1924  }
1925  mirror::Object* to_ref;
1926  if (rtype == space::RegionSpace::RegionType::kRegionTypeFromSpace) {
1927    to_ref = GetFwdPtr(from_ref);
1928    DCHECK(to_ref == nullptr || region_space_->IsInToSpace(to_ref) ||
1929           heap_->non_moving_space_->HasAddress(to_ref))
1930        << "from_ref=" << from_ref << " to_ref=" << to_ref;
1931  } else if (rtype == space::RegionSpace::RegionType::kRegionTypeUnevacFromSpace) {
1932    if (region_space_bitmap_->Test(from_ref)) {
1933      to_ref = from_ref;
1934    } else {
1935      to_ref = nullptr;
1936    }
1937  } else {
1938    // from_ref is in a non-moving space.
1939    if (immune_spaces_.ContainsObject(from_ref)) {
1940      accounting::ContinuousSpaceBitmap* cc_bitmap =
1941          cc_heap_bitmap_->GetContinuousSpaceBitmap(from_ref);
1942      DCHECK(cc_bitmap != nullptr)
1943          << "An immune space object must have a bitmap";
1944      if (kIsDebugBuild) {
1945        DCHECK(heap_mark_bitmap_->GetContinuousSpaceBitmap(from_ref)->Test(from_ref))
1946            << "Immune space object must be already marked";
1947      }
1948      if (cc_bitmap->Test(from_ref)) {
1949        // Already marked.
1950        to_ref = from_ref;
1951      } else {
1952        // Newly marked.
1953        to_ref = nullptr;
1954      }
1955    } else {
1956      // Non-immune non-moving space. Use the mark bitmap.
1957      accounting::ContinuousSpaceBitmap* mark_bitmap =
1958          heap_mark_bitmap_->GetContinuousSpaceBitmap(from_ref);
1959      accounting::LargeObjectBitmap* los_bitmap =
1960          heap_mark_bitmap_->GetLargeObjectBitmap(from_ref);
1961      CHECK(los_bitmap != nullptr) << "LOS bitmap covers the entire address range";
1962      bool is_los = mark_bitmap == nullptr;
1963      if (!is_los && mark_bitmap->Test(from_ref)) {
1964        // Already marked.
1965        to_ref = from_ref;
1966      } else if (is_los && los_bitmap->Test(from_ref)) {
1967        // Already marked in LOS.
1968        to_ref = from_ref;
1969      } else {
1970        // Not marked.
1971        if (IsOnAllocStack(from_ref)) {
1972          // If on the allocation stack, it's considered marked.
1973          to_ref = from_ref;
1974        } else {
1975          // Not marked.
1976          to_ref = nullptr;
1977        }
1978      }
1979    }
1980  }
1981  return to_ref;
1982}
1983
1984bool ConcurrentCopying::IsOnAllocStack(mirror::Object* ref) {
1985  QuasiAtomic::ThreadFenceAcquire();
1986  accounting::ObjectStack* alloc_stack = GetAllocationStack();
1987  return alloc_stack->Contains(ref);
1988}
1989
1990mirror::Object* ConcurrentCopying::MarkNonMoving(mirror::Object* ref) {
1991  // ref is in a non-moving space (from_ref == to_ref).
1992  DCHECK(!region_space_->HasAddress(ref)) << ref;
1993  if (immune_spaces_.ContainsObject(ref)) {
1994    accounting::ContinuousSpaceBitmap* cc_bitmap =
1995        cc_heap_bitmap_->GetContinuousSpaceBitmap(ref);
1996    DCHECK(cc_bitmap != nullptr)
1997        << "An immune space object must have a bitmap";
1998    if (kIsDebugBuild) {
1999      DCHECK(heap_mark_bitmap_->GetContinuousSpaceBitmap(ref)->Test(ref))
2000          << "Immune space object must be already marked";
2001    }
2002    // This may or may not succeed, which is ok.
2003    if (kUseBakerReadBarrier) {
2004      ref->AtomicSetReadBarrierPointer(ReadBarrier::WhitePtr(), ReadBarrier::GrayPtr());
2005    }
2006    if (cc_bitmap->AtomicTestAndSet(ref)) {
2007      // Already marked.
2008    } else {
2009      // Newly marked.
2010      if (kUseBakerReadBarrier) {
2011        DCHECK_EQ(ref->GetReadBarrierPointer(), ReadBarrier::GrayPtr());
2012      }
2013      PushOntoMarkStack(ref);
2014    }
2015  } else {
2016    // Use the mark bitmap.
2017    accounting::ContinuousSpaceBitmap* mark_bitmap =
2018        heap_mark_bitmap_->GetContinuousSpaceBitmap(ref);
2019    accounting::LargeObjectBitmap* los_bitmap =
2020        heap_mark_bitmap_->GetLargeObjectBitmap(ref);
2021    CHECK(los_bitmap != nullptr) << "LOS bitmap covers the entire address range";
2022    bool is_los = mark_bitmap == nullptr;
2023    if (!is_los && mark_bitmap->Test(ref)) {
2024      // Already marked.
2025      if (kUseBakerReadBarrier) {
2026        DCHECK(ref->GetReadBarrierPointer() == ReadBarrier::GrayPtr() ||
2027               ref->GetReadBarrierPointer() == ReadBarrier::BlackPtr());
2028      }
2029    } else if (is_los && los_bitmap->Test(ref)) {
2030      // Already marked in LOS.
2031      if (kUseBakerReadBarrier) {
2032        DCHECK(ref->GetReadBarrierPointer() == ReadBarrier::GrayPtr() ||
2033               ref->GetReadBarrierPointer() == ReadBarrier::BlackPtr());
2034      }
2035    } else {
2036      // Not marked.
2037      if (IsOnAllocStack(ref)) {
2038        // If it's on the allocation stack, it's considered marked. Keep it white.
2039        // Objects on the allocation stack need not be marked.
2040        if (!is_los) {
2041          DCHECK(!mark_bitmap->Test(ref));
2042        } else {
2043          DCHECK(!los_bitmap->Test(ref));
2044        }
2045        if (kUseBakerReadBarrier) {
2046          DCHECK_EQ(ref->GetReadBarrierPointer(), ReadBarrier::WhitePtr());
2047        }
2048      } else {
2049        // Not marked or on the allocation stack. Try to mark it.
2050        // This may or may not succeed, which is ok.
2051        if (kUseBakerReadBarrier) {
2052          ref->AtomicSetReadBarrierPointer(ReadBarrier::WhitePtr(), ReadBarrier::GrayPtr());
2053        }
2054        if (!is_los && mark_bitmap->AtomicTestAndSet(ref)) {
2055          // Already marked.
2056        } else if (is_los && los_bitmap->AtomicTestAndSet(ref)) {
2057          // Already marked in LOS.
2058        } else {
2059          // Newly marked.
2060          if (kUseBakerReadBarrier) {
2061            DCHECK_EQ(ref->GetReadBarrierPointer(), ReadBarrier::GrayPtr());
2062          }
2063          PushOntoMarkStack(ref);
2064        }
2065      }
2066    }
2067  }
2068  return ref;
2069}
2070
2071void ConcurrentCopying::FinishPhase() {
2072  Thread* const self = Thread::Current();
2073  {
2074    MutexLock mu(self, mark_stack_lock_);
2075    CHECK_EQ(pooled_mark_stacks_.size(), kMarkStackPoolSize);
2076  }
2077  region_space_ = nullptr;
2078  {
2079    MutexLock mu(Thread::Current(), skipped_blocks_lock_);
2080    skipped_blocks_map_.clear();
2081  }
2082  ReaderMutexLock mu(self, *Locks::mutator_lock_);
2083  WriterMutexLock mu2(self, *Locks::heap_bitmap_lock_);
2084  heap_->ClearMarkedObjects();
2085}
2086
2087bool ConcurrentCopying::IsMarkedHeapReference(mirror::HeapReference<mirror::Object>* field) {
2088  mirror::Object* from_ref = field->AsMirrorPtr();
2089  mirror::Object* to_ref = IsMarked(from_ref);
2090  if (to_ref == nullptr) {
2091    return false;
2092  }
2093  if (from_ref != to_ref) {
2094    QuasiAtomic::ThreadFenceRelease();
2095    field->Assign(to_ref);
2096    QuasiAtomic::ThreadFenceSequentiallyConsistent();
2097  }
2098  return true;
2099}
2100
2101mirror::Object* ConcurrentCopying::MarkObject(mirror::Object* from_ref) {
2102  return Mark(from_ref);
2103}
2104
2105void ConcurrentCopying::DelayReferenceReferent(mirror::Class* klass, mirror::Reference* reference) {
2106  heap_->GetReferenceProcessor()->DelayReferenceReferent(klass, reference, this);
2107}
2108
2109void ConcurrentCopying::ProcessReferences(Thread* self) {
2110  TimingLogger::ScopedTiming split("ProcessReferences", GetTimings());
2111  // We don't really need to lock the heap bitmap lock as we use CAS to mark in bitmaps.
2112  WriterMutexLock mu(self, *Locks::heap_bitmap_lock_);
2113  GetHeap()->GetReferenceProcessor()->ProcessReferences(
2114      true /*concurrent*/, GetTimings(), GetCurrentIteration()->GetClearSoftReferences(), this);
2115}
2116
2117void ConcurrentCopying::RevokeAllThreadLocalBuffers() {
2118  TimingLogger::ScopedTiming t(__FUNCTION__, GetTimings());
2119  region_space_->RevokeAllThreadLocalBuffers();
2120}
2121
2122}  // namespace collector
2123}  // namespace gc
2124}  // namespace art
2125