mark_sweep.cc revision 3e41780cb3bcade3b724908e00443a9caf6977ef
1/* 2 * Copyright (C) 2011 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17#include "mark_sweep.h" 18 19#include <functional> 20#include <numeric> 21#include <climits> 22#include <vector> 23 24#include "base/bounded_fifo.h" 25#include "base/logging.h" 26#include "base/macros.h" 27#include "base/mutex-inl.h" 28#include "base/timing_logger.h" 29#include "gc/accounting/card_table-inl.h" 30#include "gc/accounting/heap_bitmap.h" 31#include "gc/accounting/mod_union_table.h" 32#include "gc/accounting/space_bitmap-inl.h" 33#include "gc/heap.h" 34#include "gc/space/image_space.h" 35#include "gc/space/large_object_space.h" 36#include "gc/space/space-inl.h" 37#include "indirect_reference_table.h" 38#include "intern_table.h" 39#include "jni_internal.h" 40#include "monitor.h" 41#include "mark_sweep-inl.h" 42#include "mirror/art_field.h" 43#include "mirror/art_field-inl.h" 44#include "mirror/class-inl.h" 45#include "mirror/class_loader.h" 46#include "mirror/dex_cache.h" 47#include "mirror/reference-inl.h" 48#include "mirror/object-inl.h" 49#include "mirror/object_array.h" 50#include "mirror/object_array-inl.h" 51#include "runtime.h" 52#include "thread-inl.h" 53#include "thread_list.h" 54#include "verifier/method_verifier.h" 55 56using ::art::mirror::ArtField; 57using ::art::mirror::Class; 58using ::art::mirror::Object; 59using ::art::mirror::ObjectArray; 60 61namespace art { 62namespace gc { 63namespace collector { 64 65// Performance options. 66static constexpr bool kUseRecursiveMark = false; 67static constexpr bool kUseMarkStackPrefetch = true; 68static constexpr size_t kSweepArrayChunkFreeSize = 1024; 69static constexpr bool kPreCleanCards = true; 70 71// Parallelism options. 72static constexpr bool kParallelCardScan = true; 73static constexpr bool kParallelRecursiveMark = true; 74// Don't attempt to parallelize mark stack processing unless the mark stack is at least n 75// elements. This is temporary until we reduce the overhead caused by allocating tasks, etc.. Not 76// having this can add overhead in ProcessReferences since we may end up doing many calls of 77// ProcessMarkStack with very small mark stacks. 78static constexpr size_t kMinimumParallelMarkStackSize = 128; 79static constexpr bool kParallelProcessMarkStack = true; 80 81// Profiling and information flags. 82static constexpr bool kCountClassesMarked = false; 83static constexpr bool kProfileLargeObjects = false; 84static constexpr bool kMeasureOverhead = false; 85static constexpr bool kCountTasks = false; 86static constexpr bool kCountJavaLangRefs = false; 87 88// Turn off kCheckLocks when profiling the GC since it slows the GC down by up to 40%. 89static constexpr bool kCheckLocks = kDebugLocking; 90static constexpr bool kVerifyRoots = kIsDebugBuild; 91 92void MarkSweep::BindBitmaps() { 93 timings_.StartSplit("BindBitmaps"); 94 WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 95 // Mark all of the spaces we never collect as immune. 96 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 97 if (space->GetGcRetentionPolicy() == space::kGcRetentionPolicyNeverCollect) { 98 CHECK(immune_region_.AddContinuousSpace(space)) << "Failed to add space " << *space; 99 } 100 } 101 timings_.EndSplit(); 102} 103 104MarkSweep::MarkSweep(Heap* heap, bool is_concurrent, const std::string& name_prefix) 105 : GarbageCollector(heap, 106 name_prefix + 107 (is_concurrent ? "concurrent mark sweep": "mark sweep")), 108 current_mark_bitmap_(NULL), 109 mark_stack_(NULL), 110 live_stack_freeze_size_(0), 111 gc_barrier_(new Barrier(0)), 112 large_object_lock_("mark sweep large object lock", kMarkSweepLargeObjectLock), 113 mark_stack_lock_("mark sweep mark stack lock", kMarkSweepMarkStackLock), 114 is_concurrent_(is_concurrent) { 115} 116 117void MarkSweep::InitializePhase() { 118 timings_.Reset(); 119 TimingLogger::ScopedSplit split("InitializePhase", &timings_); 120 mark_stack_ = heap_->mark_stack_.get(); 121 DCHECK(mark_stack_ != nullptr); 122 immune_region_.Reset(); 123 class_count_ = 0; 124 array_count_ = 0; 125 other_count_ = 0; 126 large_object_test_ = 0; 127 large_object_mark_ = 0; 128 classes_marked_ = 0; 129 overhead_time_ = 0; 130 work_chunks_created_ = 0; 131 work_chunks_deleted_ = 0; 132 reference_count_ = 0; 133 134 FindDefaultMarkBitmap(); 135 136 // Do any pre GC verification. 137 timings_.NewSplit("PreGcVerification"); 138 heap_->PreGcVerification(this); 139} 140 141void MarkSweep::ProcessReferences(Thread* self) { 142 TimingLogger::ScopedSplit split("ProcessReferences", &timings_); 143 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 144 GetHeap()->ProcessReferences(timings_, clear_soft_references_, &IsMarkedCallback, 145 &MarkObjectCallback, &ProcessMarkStackPausedCallback, this); 146} 147 148void MarkSweep::PreProcessReferences(Thread* self) { 149 timings_.NewSplit("PreProcessReferences"); 150 GetHeap()->ProcessSoftReferences(timings_, clear_soft_references_, &IsMarkedCallback, 151 &MarkObjectCallback, &ProcessMarkStackPausedCallback, this); 152} 153 154void MarkSweep::HandleDirtyObjectsPhase() { 155 TimingLogger::ScopedSplit split("(Paused)HandleDirtyObjectsPhase", &timings_); 156 Thread* self = Thread::Current(); 157 Locks::mutator_lock_->AssertExclusiveHeld(self); 158 159 { 160 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 161 162 // Re-mark root set. 163 ReMarkRoots(); 164 165 // Scan dirty objects, this is only required if we are not doing concurrent GC. 166 RecursiveMarkDirtyObjects(true, accounting::CardTable::kCardDirty); 167 } 168 169 ProcessReferences(self); 170 171 // Only need to do this if we have the card mark verification on, and only during concurrent GC. 172 if (GetHeap()->verify_missing_card_marks_ || GetHeap()->verify_pre_gc_heap_|| 173 GetHeap()->verify_post_gc_heap_) { 174 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 175 // This second sweep makes sure that we don't have any objects in the live stack which point to 176 // freed objects. These cause problems since their references may be previously freed objects. 177 SweepArray(GetHeap()->allocation_stack_.get(), false); 178 // Since SweepArray() above resets the (active) allocation 179 // stack. Need to revoke the thread-local allocation stacks that 180 // point into it. 181 RevokeAllThreadLocalAllocationStacks(self); 182 } 183 184 timings_.StartSplit("PreSweepingGcVerification"); 185 heap_->PreSweepingGcVerification(this); 186 timings_.EndSplit(); 187 188 // Ensure that nobody inserted items in the live stack after we swapped the stacks. 189 ReaderMutexLock mu(self, *Locks::heap_bitmap_lock_); 190 CHECK_GE(live_stack_freeze_size_, GetHeap()->GetLiveStack()->Size()); 191 192 // Disallow new system weaks to prevent a race which occurs when someone adds a new system 193 // weak before we sweep them. Since this new system weak may not be marked, the GC may 194 // incorrectly sweep it. This also fixes a race where interning may attempt to return a strong 195 // reference to a string that is about to be swept. 196 Runtime::Current()->DisallowNewSystemWeaks(); 197} 198 199void MarkSweep::PreCleanCards() { 200 // Don't do this for non concurrent GCs since they don't have any dirty cards. 201 if (kPreCleanCards && IsConcurrent()) { 202 Thread* self = Thread::Current(); 203 CHECK(!Locks::mutator_lock_->IsExclusiveHeld(self)); 204 // Process dirty cards and add dirty cards to mod union tables, also ages cards. 205 heap_->ProcessCards(timings_, false); 206 // The checkpoint root marking is required to avoid a race condition which occurs if the 207 // following happens during a reference write: 208 // 1. mutator dirties the card (write barrier) 209 // 2. GC ages the card (the above ProcessCards call) 210 // 3. GC scans the object (the RecursiveMarkDirtyObjects call below) 211 // 4. mutator writes the value (corresponding to the write barrier in 1.) 212 // This causes the GC to age the card but not necessarily mark the reference which the mutator 213 // wrote into the object stored in the card. 214 // Having the checkpoint fixes this issue since it ensures that the card mark and the 215 // reference write are visible to the GC before the card is scanned (this is due to locks being 216 // acquired / released in the checkpoint code). 217 // The other roots are also marked to help reduce the pause. 218 MarkThreadRoots(self); 219 // TODO: Only mark the dirty roots. 220 MarkNonThreadRoots(); 221 MarkConcurrentRoots( 222 static_cast<VisitRootFlags>(kVisitRootFlagClearRootLog | kVisitRootFlagNewRoots)); 223 // Process the newly aged cards. 224 RecursiveMarkDirtyObjects(false, accounting::CardTable::kCardDirty - 1); 225 // TODO: Empty allocation stack to reduce the number of objects we need to test / mark as live 226 // in the next GC. 227 } 228} 229 230void MarkSweep::RevokeAllThreadLocalAllocationStacks(Thread* self) { 231 if (kUseThreadLocalAllocationStack) { 232 Locks::mutator_lock_->AssertExclusiveHeld(self); 233 heap_->RevokeAllThreadLocalAllocationStacks(self); 234 } 235} 236 237void MarkSweep::MarkingPhase() { 238 TimingLogger::ScopedSplit split("MarkingPhase", &timings_); 239 Thread* self = Thread::Current(); 240 241 BindBitmaps(); 242 FindDefaultMarkBitmap(); 243 244 // Process dirty cards and add dirty cards to mod union tables. 245 heap_->ProcessCards(timings_, false); 246 247 // Need to do this before the checkpoint since we don't want any threads to add references to 248 // the live stack during the recursive mark. 249 timings_.NewSplit("SwapStacks"); 250 heap_->SwapStacks(self); 251 252 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 253 MarkRoots(self); 254 live_stack_freeze_size_ = heap_->GetLiveStack()->Size(); 255 UpdateAndMarkModUnion(); 256 MarkReachableObjects(); 257 // Pre-clean dirtied cards to reduce pauses. 258 PreCleanCards(); 259 if (IsConcurrent()) { 260 // No reason to do this for non-concurrent GC since pre processing soft references only helps 261 // pauses. 262 PreProcessReferences(self); 263 } 264} 265 266void MarkSweep::UpdateAndMarkModUnion() { 267 for (const auto& space : heap_->GetContinuousSpaces()) { 268 if (immune_region_.ContainsSpace(space)) { 269 const char* name = space->IsZygoteSpace() ? "UpdateAndMarkZygoteModUnionTable" : 270 "UpdateAndMarkImageModUnionTable"; 271 TimingLogger::ScopedSplit split(name, &timings_); 272 accounting::ModUnionTable* mod_union_table = heap_->FindModUnionTableFromSpace(space); 273 CHECK(mod_union_table != nullptr); 274 mod_union_table->UpdateAndMarkReferences(MarkObjectCallback, this); 275 } 276 } 277} 278 279void MarkSweep::MarkThreadRoots(Thread* self) { 280 MarkRootsCheckpoint(self); 281} 282 283void MarkSweep::MarkReachableObjects() { 284 // Mark everything allocated since the last as GC live so that we can sweep concurrently, 285 // knowing that new allocations won't be marked as live. 286 timings_.StartSplit("MarkStackAsLive"); 287 accounting::ObjectStack* live_stack = heap_->GetLiveStack(); 288 heap_->MarkAllocStackAsLive(live_stack); 289 live_stack->Reset(); 290 timings_.EndSplit(); 291 // Recursively mark all the non-image bits set in the mark bitmap. 292 RecursiveMark(); 293} 294 295void MarkSweep::ReclaimPhase() { 296 TimingLogger::ScopedSplit split("ReclaimPhase", &timings_); 297 Thread* self = Thread::Current(); 298 299 if (!IsConcurrent()) { 300 ProcessReferences(self); 301 } 302 303 { 304 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 305 SweepSystemWeaks(); 306 } 307 308 if (IsConcurrent()) { 309 Runtime::Current()->AllowNewSystemWeaks(); 310 311 TimingLogger::ScopedSplit split("UnMarkAllocStack", &timings_); 312 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 313 accounting::ObjectStack* allocation_stack = GetHeap()->allocation_stack_.get(); 314 if (!kPreCleanCards) { 315 // The allocation stack contains things allocated since the start of the GC. These may have 316 // been marked during this GC meaning they won't be eligible for reclaiming in the next 317 // sticky GC. Unmark these objects so that they are eligible for reclaiming in the next 318 // sticky GC. 319 // There is a race here which is safely handled. Another thread such as the hprof could 320 // have flushed the alloc stack after we resumed the threads. This is safe however, since 321 // reseting the allocation stack zeros it out with madvise. This means that we will either 322 // read NULLs or attempt to unmark a newly allocated object which will not be marked in the 323 // first place. 324 // We can't do this if we pre-clean cards since we will unmark objects which are no longer on 325 // a dirty card since we aged cards during the pre-cleaning process. 326 mirror::Object** end = allocation_stack->End(); 327 for (mirror::Object** it = allocation_stack->Begin(); it != end; ++it) { 328 const Object* obj = *it; 329 if (obj != nullptr) { 330 UnMarkObjectNonNull(obj); 331 } 332 } 333 } 334 } 335 336 { 337 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 338 339 // Reclaim unmarked objects. 340 Sweep(false); 341 342 // Swap the live and mark bitmaps for each space which we modified space. This is an 343 // optimization that enables us to not clear live bits inside of the sweep. Only swaps unbound 344 // bitmaps. 345 timings_.StartSplit("SwapBitmaps"); 346 SwapBitmaps(); 347 timings_.EndSplit(); 348 349 // Unbind the live and mark bitmaps. 350 TimingLogger::ScopedSplit split("UnBindBitmaps", &timings_); 351 GetHeap()->UnBindBitmaps(); 352 } 353} 354 355void MarkSweep::FindDefaultMarkBitmap() { 356 TimingLogger::ScopedSplit split("FindDefaultMarkBitmap", &timings_); 357 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 358 accounting::SpaceBitmap* bitmap = space->GetMarkBitmap(); 359 if (bitmap != nullptr && 360 space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) { 361 current_mark_bitmap_ = bitmap; 362 CHECK(current_mark_bitmap_ != NULL); 363 return; 364 } 365 } 366 GetHeap()->DumpSpaces(); 367 LOG(FATAL) << "Could not find a default mark bitmap"; 368} 369 370void MarkSweep::ExpandMarkStack() { 371 ResizeMarkStack(mark_stack_->Capacity() * 2); 372} 373 374void MarkSweep::ResizeMarkStack(size_t new_size) { 375 // Rare case, no need to have Thread::Current be a parameter. 376 if (UNLIKELY(mark_stack_->Size() < mark_stack_->Capacity())) { 377 // Someone else acquired the lock and expanded the mark stack before us. 378 return; 379 } 380 std::vector<Object*> temp(mark_stack_->Begin(), mark_stack_->End()); 381 CHECK_LE(mark_stack_->Size(), new_size); 382 mark_stack_->Resize(new_size); 383 for (const auto& obj : temp) { 384 mark_stack_->PushBack(obj); 385 } 386} 387 388inline void MarkSweep::MarkObjectNonNullParallel(const Object* obj) { 389 DCHECK(obj != NULL); 390 if (MarkObjectParallel(obj)) { 391 MutexLock mu(Thread::Current(), mark_stack_lock_); 392 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 393 ExpandMarkStack(); 394 } 395 // The object must be pushed on to the mark stack. 396 mark_stack_->PushBack(const_cast<Object*>(obj)); 397 } 398} 399 400mirror::Object* MarkSweep::MarkObjectCallback(mirror::Object* obj, void* arg) { 401 MarkSweep* mark_sweep = reinterpret_cast<MarkSweep*>(arg); 402 mark_sweep->MarkObject(obj); 403 return obj; 404} 405 406inline void MarkSweep::UnMarkObjectNonNull(const Object* obj) { 407 DCHECK(!immune_region_.ContainsObject(obj)); 408 409 if (kUseBrooksPointer) { 410 // Verify all the objects have the correct Brooks pointer installed. 411 obj->AssertSelfBrooksPointer(); 412 } 413 414 // Try to take advantage of locality of references within a space, failing this find the space 415 // the hard way. 416 accounting::SpaceBitmap* object_bitmap = current_mark_bitmap_; 417 if (UNLIKELY(!object_bitmap->HasAddress(obj))) { 418 accounting::SpaceBitmap* new_bitmap = heap_->GetMarkBitmap()->GetContinuousSpaceBitmap(obj); 419 if (LIKELY(new_bitmap != NULL)) { 420 object_bitmap = new_bitmap; 421 } else { 422 MarkLargeObject(obj, false); 423 return; 424 } 425 } 426 427 DCHECK(object_bitmap->HasAddress(obj)); 428 object_bitmap->Clear(obj); 429} 430 431inline void MarkSweep::MarkObjectNonNull(const Object* obj) { 432 DCHECK(obj != NULL); 433 434 if (kUseBrooksPointer) { 435 // Verify all the objects have the correct Brooks pointer installed. 436 obj->AssertSelfBrooksPointer(); 437 } 438 439 if (immune_region_.ContainsObject(obj)) { 440 DCHECK(IsMarked(obj)); 441 return; 442 } 443 444 // Try to take advantage of locality of references within a space, failing this find the space 445 // the hard way. 446 accounting::SpaceBitmap* object_bitmap = current_mark_bitmap_; 447 if (UNLIKELY(!object_bitmap->HasAddress(obj))) { 448 accounting::SpaceBitmap* new_bitmap = heap_->GetMarkBitmap()->GetContinuousSpaceBitmap(obj); 449 if (LIKELY(new_bitmap != NULL)) { 450 object_bitmap = new_bitmap; 451 } else { 452 MarkLargeObject(obj, true); 453 return; 454 } 455 } 456 457 // This object was not previously marked. 458 if (!object_bitmap->Test(obj)) { 459 object_bitmap->Set(obj); 460 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 461 // Lock is not needed but is here anyways to please annotalysis. 462 MutexLock mu(Thread::Current(), mark_stack_lock_); 463 ExpandMarkStack(); 464 } 465 // The object must be pushed on to the mark stack. 466 mark_stack_->PushBack(const_cast<Object*>(obj)); 467 } 468} 469 470// Rare case, probably not worth inlining since it will increase instruction cache miss rate. 471bool MarkSweep::MarkLargeObject(const Object* obj, bool set) { 472 // TODO: support >1 discontinuous space. 473 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 474 accounting::ObjectSet* large_objects = large_object_space->GetMarkObjects(); 475 if (kProfileLargeObjects) { 476 ++large_object_test_; 477 } 478 if (UNLIKELY(!large_objects->Test(obj))) { 479 if (!large_object_space->Contains(obj)) { 480 LOG(ERROR) << "Tried to mark " << obj << " not contained by any spaces"; 481 LOG(ERROR) << "Attempting see if it's a bad root"; 482 VerifyRoots(); 483 LOG(FATAL) << "Can't mark bad root"; 484 } 485 if (kProfileLargeObjects) { 486 ++large_object_mark_; 487 } 488 if (set) { 489 large_objects->Set(obj); 490 } else { 491 large_objects->Clear(obj); 492 } 493 return true; 494 } 495 return false; 496} 497 498inline bool MarkSweep::MarkObjectParallel(const Object* obj) { 499 DCHECK(obj != NULL); 500 501 if (kUseBrooksPointer) { 502 // Verify all the objects have the correct Brooks pointer installed. 503 obj->AssertSelfBrooksPointer(); 504 } 505 506 if (immune_region_.ContainsObject(obj)) { 507 DCHECK(IsMarked(obj)); 508 return false; 509 } 510 511 // Try to take advantage of locality of references within a space, failing this find the space 512 // the hard way. 513 accounting::SpaceBitmap* object_bitmap = current_mark_bitmap_; 514 if (UNLIKELY(!object_bitmap->HasAddress(obj))) { 515 accounting::SpaceBitmap* new_bitmap = heap_->GetMarkBitmap()->GetContinuousSpaceBitmap(obj); 516 if (new_bitmap != NULL) { 517 object_bitmap = new_bitmap; 518 } else { 519 // TODO: Remove the Thread::Current here? 520 // TODO: Convert this to some kind of atomic marking? 521 MutexLock mu(Thread::Current(), large_object_lock_); 522 return MarkLargeObject(obj, true); 523 } 524 } 525 526 // Return true if the object was not previously marked. 527 return !object_bitmap->AtomicTestAndSet(obj); 528} 529 530// Used to mark objects when recursing. Recursion is done by moving 531// the finger across the bitmaps in address order and marking child 532// objects. Any newly-marked objects whose addresses are lower than 533// the finger won't be visited by the bitmap scan, so those objects 534// need to be added to the mark stack. 535inline void MarkSweep::MarkObject(const Object* obj) { 536 if (obj != NULL) { 537 MarkObjectNonNull(obj); 538 } 539} 540 541void MarkSweep::MarkRootParallelCallback(mirror::Object** root, void* arg, uint32_t /*thread_id*/, 542 RootType /*root_type*/) { 543 reinterpret_cast<MarkSweep*>(arg)->MarkObjectNonNullParallel(*root); 544} 545 546void MarkSweep::VerifyRootMarked(Object** root, void* arg, uint32_t /*thread_id*/, 547 RootType /*root_type*/) { 548 CHECK(reinterpret_cast<MarkSweep*>(arg)->IsMarked(*root)); 549} 550 551void MarkSweep::MarkRootCallback(Object** root, void* arg, uint32_t /*thread_id*/, 552 RootType /*root_type*/) { 553 reinterpret_cast<MarkSweep*>(arg)->MarkObjectNonNull(*root); 554} 555 556void MarkSweep::VerifyRootCallback(const Object* root, void* arg, size_t vreg, 557 const StackVisitor* visitor) { 558 reinterpret_cast<MarkSweep*>(arg)->VerifyRoot(root, vreg, visitor); 559} 560 561void MarkSweep::VerifyRoot(const Object* root, size_t vreg, const StackVisitor* visitor) { 562 // See if the root is on any space bitmap. 563 if (GetHeap()->GetLiveBitmap()->GetContinuousSpaceBitmap(root) == NULL) { 564 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 565 if (!large_object_space->Contains(root)) { 566 LOG(ERROR) << "Found invalid root: " << root; 567 if (visitor != NULL) { 568 LOG(ERROR) << visitor->DescribeLocation() << " in VReg: " << vreg; 569 } 570 } 571 } 572} 573 574void MarkSweep::VerifyRoots() { 575 Runtime::Current()->GetThreadList()->VerifyRoots(VerifyRootCallback, this); 576} 577 578void MarkSweep::MarkRoots(Thread* self) { 579 if (Locks::mutator_lock_->IsExclusiveHeld(self)) { 580 // If we exclusively hold the mutator lock, all threads must be suspended. 581 timings_.StartSplit("MarkRoots"); 582 Runtime::Current()->VisitRoots(MarkRootCallback, this); 583 timings_.EndSplit(); 584 RevokeAllThreadLocalAllocationStacks(self); 585 } else { 586 MarkThreadRoots(self); 587 // At this point the live stack should no longer have any mutators which push into it. 588 MarkNonThreadRoots(); 589 MarkConcurrentRoots( 590 static_cast<VisitRootFlags>(kVisitRootFlagAllRoots | kVisitRootFlagStartLoggingNewRoots)); 591 } 592} 593 594void MarkSweep::MarkNonThreadRoots() { 595 timings_.StartSplit("MarkNonThreadRoots"); 596 Runtime::Current()->VisitNonThreadRoots(MarkRootCallback, this); 597 timings_.EndSplit(); 598} 599 600void MarkSweep::MarkConcurrentRoots(VisitRootFlags flags) { 601 timings_.StartSplit("MarkConcurrentRoots"); 602 // Visit all runtime roots and clear dirty flags. 603 Runtime::Current()->VisitConcurrentRoots(MarkRootCallback, this, flags); 604 timings_.EndSplit(); 605} 606 607class ScanObjectVisitor { 608 public: 609 explicit ScanObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE 610 : mark_sweep_(mark_sweep) {} 611 612 // TODO: Fixme when anotatalysis works with visitors. 613 void operator()(Object* obj) const ALWAYS_INLINE NO_THREAD_SAFETY_ANALYSIS { 614 if (kCheckLocks) { 615 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 616 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 617 } 618 mark_sweep_->ScanObject(obj); 619 } 620 621 private: 622 MarkSweep* const mark_sweep_; 623}; 624 625template <bool kUseFinger = false> 626class MarkStackTask : public Task { 627 public: 628 MarkStackTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, size_t mark_stack_size, 629 const Object** mark_stack) 630 : mark_sweep_(mark_sweep), 631 thread_pool_(thread_pool), 632 mark_stack_pos_(mark_stack_size) { 633 // We may have to copy part of an existing mark stack when another mark stack overflows. 634 if (mark_stack_size != 0) { 635 DCHECK(mark_stack != NULL); 636 // TODO: Check performance? 637 std::copy(mark_stack, mark_stack + mark_stack_size, mark_stack_); 638 } 639 if (kCountTasks) { 640 ++mark_sweep_->work_chunks_created_; 641 } 642 } 643 644 static const size_t kMaxSize = 1 * KB; 645 646 protected: 647 class ScanObjectParallelVisitor { 648 public: 649 explicit ScanObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task) ALWAYS_INLINE 650 : chunk_task_(chunk_task) {} 651 652 void operator()(Object* obj) const { 653 MarkSweep* mark_sweep = chunk_task_->mark_sweep_; 654 mark_sweep->ScanObjectVisit(obj, 655 [mark_sweep, this](Object* /* obj */, Object* ref, const MemberOffset& /* offset */, 656 bool /* is_static */) ALWAYS_INLINE_LAMBDA { 657 if (ref != nullptr && mark_sweep->MarkObjectParallel(ref)) { 658 if (kUseFinger) { 659 android_memory_barrier(); 660 if (reinterpret_cast<uintptr_t>(ref) >= 661 static_cast<uintptr_t>(mark_sweep->atomic_finger_)) { 662 return; 663 } 664 } 665 chunk_task_->MarkStackPush(ref); 666 } 667 }); 668 } 669 670 private: 671 MarkStackTask<kUseFinger>* const chunk_task_; 672 }; 673 674 virtual ~MarkStackTask() { 675 // Make sure that we have cleared our mark stack. 676 DCHECK_EQ(mark_stack_pos_, 0U); 677 if (kCountTasks) { 678 ++mark_sweep_->work_chunks_deleted_; 679 } 680 } 681 682 MarkSweep* const mark_sweep_; 683 ThreadPool* const thread_pool_; 684 // Thread local mark stack for this task. 685 const Object* mark_stack_[kMaxSize]; 686 // Mark stack position. 687 size_t mark_stack_pos_; 688 689 void MarkStackPush(const Object* obj) ALWAYS_INLINE { 690 if (UNLIKELY(mark_stack_pos_ == kMaxSize)) { 691 // Mark stack overflow, give 1/2 the stack to the thread pool as a new work task. 692 mark_stack_pos_ /= 2; 693 auto* task = new MarkStackTask(thread_pool_, mark_sweep_, kMaxSize - mark_stack_pos_, 694 mark_stack_ + mark_stack_pos_); 695 thread_pool_->AddTask(Thread::Current(), task); 696 } 697 DCHECK(obj != nullptr); 698 DCHECK(mark_stack_pos_ < kMaxSize); 699 mark_stack_[mark_stack_pos_++] = obj; 700 } 701 702 virtual void Finalize() { 703 delete this; 704 } 705 706 // Scans all of the objects 707 virtual void Run(Thread* self) { 708 ScanObjectParallelVisitor visitor(this); 709 // TODO: Tune this. 710 static const size_t kFifoSize = 4; 711 BoundedFifoPowerOfTwo<const Object*, kFifoSize> prefetch_fifo; 712 for (;;) { 713 const Object* obj = nullptr; 714 if (kUseMarkStackPrefetch) { 715 while (mark_stack_pos_ != 0 && prefetch_fifo.size() < kFifoSize) { 716 const Object* obj = mark_stack_[--mark_stack_pos_]; 717 DCHECK(obj != nullptr); 718 __builtin_prefetch(obj); 719 prefetch_fifo.push_back(obj); 720 } 721 if (UNLIKELY(prefetch_fifo.empty())) { 722 break; 723 } 724 obj = prefetch_fifo.front(); 725 prefetch_fifo.pop_front(); 726 } else { 727 if (UNLIKELY(mark_stack_pos_ == 0)) { 728 break; 729 } 730 obj = mark_stack_[--mark_stack_pos_]; 731 } 732 DCHECK(obj != nullptr); 733 visitor(const_cast<mirror::Object*>(obj)); 734 } 735 } 736}; 737 738class CardScanTask : public MarkStackTask<false> { 739 public: 740 CardScanTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, accounting::SpaceBitmap* bitmap, 741 byte* begin, byte* end, byte minimum_age, size_t mark_stack_size, 742 const Object** mark_stack_obj) 743 : MarkStackTask<false>(thread_pool, mark_sweep, mark_stack_size, mark_stack_obj), 744 bitmap_(bitmap), 745 begin_(begin), 746 end_(end), 747 minimum_age_(minimum_age) { 748 } 749 750 protected: 751 accounting::SpaceBitmap* const bitmap_; 752 byte* const begin_; 753 byte* const end_; 754 const byte minimum_age_; 755 756 virtual void Finalize() { 757 delete this; 758 } 759 760 virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS { 761 ScanObjectParallelVisitor visitor(this); 762 accounting::CardTable* card_table = mark_sweep_->GetHeap()->GetCardTable(); 763 size_t cards_scanned = card_table->Scan(bitmap_, begin_, end_, visitor, minimum_age_); 764 VLOG(heap) << "Parallel scanning cards " << reinterpret_cast<void*>(begin_) << " - " 765 << reinterpret_cast<void*>(end_) << " = " << cards_scanned; 766 // Finish by emptying our local mark stack. 767 MarkStackTask::Run(self); 768 } 769}; 770 771size_t MarkSweep::GetThreadCount(bool paused) const { 772 if (heap_->GetThreadPool() == nullptr || !heap_->CareAboutPauseTimes()) { 773 return 0; 774 } 775 if (paused) { 776 return heap_->GetParallelGCThreadCount() + 1; 777 } else { 778 return heap_->GetConcGCThreadCount() + 1; 779 } 780} 781 782void MarkSweep::ScanGrayObjects(bool paused, byte minimum_age) { 783 accounting::CardTable* card_table = GetHeap()->GetCardTable(); 784 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 785 size_t thread_count = GetThreadCount(paused); 786 // The parallel version with only one thread is faster for card scanning, TODO: fix. 787 if (kParallelCardScan && thread_count > 0) { 788 Thread* self = Thread::Current(); 789 // Can't have a different split for each space since multiple spaces can have their cards being 790 // scanned at the same time. 791 timings_.StartSplit(paused ? "(Paused)ScanGrayObjects" : "ScanGrayObjects"); 792 // Try to take some of the mark stack since we can pass this off to the worker tasks. 793 const Object** mark_stack_begin = const_cast<const Object**>(mark_stack_->Begin()); 794 const Object** mark_stack_end = const_cast<const Object**>(mark_stack_->End()); 795 const size_t mark_stack_size = mark_stack_end - mark_stack_begin; 796 // Estimated number of work tasks we will create. 797 const size_t mark_stack_tasks = GetHeap()->GetContinuousSpaces().size() * thread_count; 798 DCHECK_NE(mark_stack_tasks, 0U); 799 const size_t mark_stack_delta = std::min(CardScanTask::kMaxSize / 2, 800 mark_stack_size / mark_stack_tasks + 1); 801 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 802 if (space->GetMarkBitmap() == nullptr) { 803 continue; 804 } 805 byte* card_begin = space->Begin(); 806 byte* card_end = space->End(); 807 // Align up the end address. For example, the image space's end 808 // may not be card-size-aligned. 809 card_end = AlignUp(card_end, accounting::CardTable::kCardSize); 810 DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_begin)); 811 DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_end)); 812 // Calculate how many bytes of heap we will scan, 813 const size_t address_range = card_end - card_begin; 814 // Calculate how much address range each task gets. 815 const size_t card_delta = RoundUp(address_range / thread_count + 1, 816 accounting::CardTable::kCardSize); 817 // Create the worker tasks for this space. 818 while (card_begin != card_end) { 819 // Add a range of cards. 820 size_t addr_remaining = card_end - card_begin; 821 size_t card_increment = std::min(card_delta, addr_remaining); 822 // Take from the back of the mark stack. 823 size_t mark_stack_remaining = mark_stack_end - mark_stack_begin; 824 size_t mark_stack_increment = std::min(mark_stack_delta, mark_stack_remaining); 825 mark_stack_end -= mark_stack_increment; 826 mark_stack_->PopBackCount(static_cast<int32_t>(mark_stack_increment)); 827 DCHECK_EQ(mark_stack_end, const_cast<const art::mirror::Object **>(mark_stack_->End())); 828 // Add the new task to the thread pool. 829 auto* task = new CardScanTask(thread_pool, this, space->GetMarkBitmap(), card_begin, 830 card_begin + card_increment, minimum_age, 831 mark_stack_increment, mark_stack_end); 832 thread_pool->AddTask(self, task); 833 card_begin += card_increment; 834 } 835 } 836 837 // Note: the card scan below may dirty new cards (and scan them) 838 // as a side effect when a Reference object is encountered and 839 // queued during the marking. See b/11465268. 840 thread_pool->SetMaxActiveWorkers(thread_count - 1); 841 thread_pool->StartWorkers(self); 842 thread_pool->Wait(self, true, true); 843 thread_pool->StopWorkers(self); 844 timings_.EndSplit(); 845 } else { 846 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 847 if (space->GetMarkBitmap() != nullptr) { 848 // Image spaces are handled properly since live == marked for them. 849 switch (space->GetGcRetentionPolicy()) { 850 case space::kGcRetentionPolicyNeverCollect: 851 timings_.StartSplit(paused ? "(Paused)ScanGrayImageSpaceObjects" : 852 "ScanGrayImageSpaceObjects"); 853 break; 854 case space::kGcRetentionPolicyFullCollect: 855 timings_.StartSplit(paused ? "(Paused)ScanGrayZygoteSpaceObjects" : 856 "ScanGrayZygoteSpaceObjects"); 857 break; 858 case space::kGcRetentionPolicyAlwaysCollect: 859 timings_.StartSplit(paused ? "(Paused)ScanGrayAllocSpaceObjects" : 860 "ScanGrayAllocSpaceObjects"); 861 break; 862 } 863 ScanObjectVisitor visitor(this); 864 card_table->Scan(space->GetMarkBitmap(), space->Begin(), space->End(), visitor, minimum_age); 865 timings_.EndSplit(); 866 } 867 } 868 } 869} 870 871class RecursiveMarkTask : public MarkStackTask<false> { 872 public: 873 RecursiveMarkTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, 874 accounting::SpaceBitmap* bitmap, uintptr_t begin, uintptr_t end) 875 : MarkStackTask<false>(thread_pool, mark_sweep, 0, NULL), 876 bitmap_(bitmap), 877 begin_(begin), 878 end_(end) { 879 } 880 881 protected: 882 accounting::SpaceBitmap* const bitmap_; 883 const uintptr_t begin_; 884 const uintptr_t end_; 885 886 virtual void Finalize() { 887 delete this; 888 } 889 890 // Scans all of the objects 891 virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS { 892 ScanObjectParallelVisitor visitor(this); 893 bitmap_->VisitMarkedRange(begin_, end_, visitor); 894 // Finish by emptying our local mark stack. 895 MarkStackTask::Run(self); 896 } 897}; 898 899// Populates the mark stack based on the set of marked objects and 900// recursively marks until the mark stack is emptied. 901void MarkSweep::RecursiveMark() { 902 TimingLogger::ScopedSplit split("RecursiveMark", &timings_); 903 // RecursiveMark will build the lists of known instances of the Reference classes. See 904 // DelayReferenceReferent for details. 905 if (kUseRecursiveMark) { 906 const bool partial = GetGcType() == kGcTypePartial; 907 ScanObjectVisitor scan_visitor(this); 908 auto* self = Thread::Current(); 909 ThreadPool* thread_pool = heap_->GetThreadPool(); 910 size_t thread_count = GetThreadCount(false); 911 const bool parallel = kParallelRecursiveMark && thread_count > 1; 912 mark_stack_->Reset(); 913 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 914 if ((space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) || 915 (!partial && space->GetGcRetentionPolicy() == space::kGcRetentionPolicyFullCollect)) { 916 current_mark_bitmap_ = space->GetMarkBitmap(); 917 if (current_mark_bitmap_ == nullptr) { 918 continue; 919 } 920 if (parallel) { 921 // We will use the mark stack the future. 922 // CHECK(mark_stack_->IsEmpty()); 923 // This function does not handle heap end increasing, so we must use the space end. 924 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 925 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 926 atomic_finger_ = static_cast<int32_t>(0xFFFFFFFF); 927 928 // Create a few worker tasks. 929 const size_t n = thread_count * 2; 930 while (begin != end) { 931 uintptr_t start = begin; 932 uintptr_t delta = (end - begin) / n; 933 delta = RoundUp(delta, KB); 934 if (delta < 16 * KB) delta = end - begin; 935 begin += delta; 936 auto* task = new RecursiveMarkTask(thread_pool, this, current_mark_bitmap_, start, 937 begin); 938 thread_pool->AddTask(self, task); 939 } 940 thread_pool->SetMaxActiveWorkers(thread_count - 1); 941 thread_pool->StartWorkers(self); 942 thread_pool->Wait(self, true, true); 943 thread_pool->StopWorkers(self); 944 } else { 945 // This function does not handle heap end increasing, so we must use the space end. 946 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 947 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 948 current_mark_bitmap_->VisitMarkedRange(begin, end, scan_visitor); 949 } 950 } 951 } 952 } 953 ProcessMarkStack(false); 954} 955 956mirror::Object* MarkSweep::IsMarkedCallback(mirror::Object* object, void* arg) { 957 if (reinterpret_cast<MarkSweep*>(arg)->IsMarked(object)) { 958 return object; 959 } 960 return nullptr; 961} 962 963void MarkSweep::RecursiveMarkDirtyObjects(bool paused, byte minimum_age) { 964 ScanGrayObjects(paused, minimum_age); 965 ProcessMarkStack(paused); 966} 967 968void MarkSweep::ReMarkRoots() { 969 Locks::mutator_lock_->AssertExclusiveHeld(Thread::Current()); 970 timings_.StartSplit("(Paused)ReMarkRoots"); 971 Runtime::Current()->VisitRoots( 972 MarkRootCallback, this, static_cast<VisitRootFlags>(kVisitRootFlagNewRoots | 973 kVisitRootFlagStopLoggingNewRoots | 974 kVisitRootFlagClearRootLog)); 975 timings_.EndSplit(); 976 if (kVerifyRoots) { 977 timings_.StartSplit("(Paused)VerifyRoots"); 978 Runtime::Current()->VisitRoots(VerifyRootMarked, this); 979 timings_.EndSplit(); 980 } 981} 982 983void MarkSweep::SweepSystemWeaks() { 984 Runtime* runtime = Runtime::Current(); 985 timings_.StartSplit("SweepSystemWeaks"); 986 runtime->SweepSystemWeaks(IsMarkedCallback, this); 987 timings_.EndSplit(); 988} 989 990mirror::Object* MarkSweep::VerifySystemWeakIsLiveCallback(Object* obj, void* arg) { 991 reinterpret_cast<MarkSweep*>(arg)->VerifyIsLive(obj); 992 // We don't actually want to sweep the object, so lets return "marked" 993 return obj; 994} 995 996void MarkSweep::VerifyIsLive(const Object* obj) { 997 Heap* heap = GetHeap(); 998 if (!heap->GetLiveBitmap()->Test(obj)) { 999 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 1000 if (!large_object_space->GetLiveObjects()->Test(obj)) { 1001 if (std::find(heap->allocation_stack_->Begin(), heap->allocation_stack_->End(), obj) == 1002 heap->allocation_stack_->End()) { 1003 // Object not found! 1004 heap->DumpSpaces(); 1005 LOG(FATAL) << "Found dead object " << obj; 1006 } 1007 } 1008 } 1009} 1010 1011void MarkSweep::VerifySystemWeaks() { 1012 // Verify system weaks, uses a special object visitor which returns the input object. 1013 Runtime::Current()->SweepSystemWeaks(VerifySystemWeakIsLiveCallback, this); 1014} 1015 1016class CheckpointMarkThreadRoots : public Closure { 1017 public: 1018 explicit CheckpointMarkThreadRoots(MarkSweep* mark_sweep) : mark_sweep_(mark_sweep) {} 1019 1020 virtual void Run(Thread* thread) NO_THREAD_SAFETY_ANALYSIS { 1021 ATRACE_BEGIN("Marking thread roots"); 1022 // Note: self is not necessarily equal to thread since thread may be suspended. 1023 Thread* self = Thread::Current(); 1024 CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc) 1025 << thread->GetState() << " thread " << thread << " self " << self; 1026 thread->VisitRoots(MarkSweep::MarkRootParallelCallback, mark_sweep_); 1027 ATRACE_END(); 1028 if (kUseThreadLocalAllocationStack) { 1029 thread->RevokeThreadLocalAllocationStack(); 1030 } 1031 mark_sweep_->GetBarrier().Pass(self); 1032 } 1033 1034 private: 1035 MarkSweep* mark_sweep_; 1036}; 1037 1038void MarkSweep::MarkRootsCheckpoint(Thread* self) { 1039 CheckpointMarkThreadRoots check_point(this); 1040 timings_.StartSplit("MarkRootsCheckpoint"); 1041 ThreadList* thread_list = Runtime::Current()->GetThreadList(); 1042 // Request the check point is run on all threads returning a count of the threads that must 1043 // run through the barrier including self. 1044 size_t barrier_count = thread_list->RunCheckpoint(&check_point); 1045 // Release locks then wait for all mutator threads to pass the barrier. 1046 // TODO: optimize to not release locks when there are no threads to wait for. 1047 Locks::heap_bitmap_lock_->ExclusiveUnlock(self); 1048 Locks::mutator_lock_->SharedUnlock(self); 1049 ThreadState old_state = self->SetState(kWaitingForCheckPointsToRun); 1050 CHECK_EQ(old_state, kWaitingPerformingGc); 1051 gc_barrier_->Increment(self, barrier_count); 1052 self->SetState(kWaitingPerformingGc); 1053 Locks::mutator_lock_->SharedLock(self); 1054 Locks::heap_bitmap_lock_->ExclusiveLock(self); 1055 timings_.EndSplit(); 1056} 1057 1058void MarkSweep::SweepArray(accounting::ObjectStack* allocations, bool swap_bitmaps) { 1059 timings_.StartSplit("SweepArray"); 1060 Thread* self = Thread::Current(); 1061 mirror::Object* chunk_free_buffer[kSweepArrayChunkFreeSize]; 1062 size_t chunk_free_pos = 0; 1063 size_t freed_bytes = 0; 1064 size_t freed_large_object_bytes = 0; 1065 size_t freed_objects = 0; 1066 size_t freed_large_objects = 0; 1067 // How many objects are left in the array, modified after each space is swept. 1068 Object** objects = const_cast<Object**>(allocations->Begin()); 1069 size_t count = allocations->Size(); 1070 // Change the order to ensure that the non-moving space last swept as an optimization. 1071 std::vector<space::ContinuousSpace*> sweep_spaces; 1072 space::ContinuousSpace* non_moving_space = nullptr; 1073 for (space::ContinuousSpace* space : heap_->GetContinuousSpaces()) { 1074 if (space->IsAllocSpace() && !immune_region_.ContainsSpace(space) && 1075 space->GetLiveBitmap() != nullptr) { 1076 if (space == heap_->GetNonMovingSpace()) { 1077 non_moving_space = space; 1078 } else { 1079 sweep_spaces.push_back(space); 1080 } 1081 } 1082 } 1083 // Unlikely to sweep a significant amount of non_movable objects, so we do these after the after 1084 // the other alloc spaces as an optimization. 1085 if (non_moving_space != nullptr) { 1086 sweep_spaces.push_back(non_moving_space); 1087 } 1088 // Start by sweeping the continuous spaces. 1089 for (space::ContinuousSpace* space : sweep_spaces) { 1090 space::AllocSpace* alloc_space = space->AsAllocSpace(); 1091 accounting::SpaceBitmap* live_bitmap = space->GetLiveBitmap(); 1092 accounting::SpaceBitmap* mark_bitmap = space->GetMarkBitmap(); 1093 if (swap_bitmaps) { 1094 std::swap(live_bitmap, mark_bitmap); 1095 } 1096 Object** out = objects; 1097 for (size_t i = 0; i < count; ++i) { 1098 Object* obj = objects[i]; 1099 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1100 continue; 1101 } 1102 if (space->HasAddress(obj)) { 1103 // This object is in the space, remove it from the array and add it to the sweep buffer 1104 // if needed. 1105 if (!mark_bitmap->Test(obj)) { 1106 if (chunk_free_pos >= kSweepArrayChunkFreeSize) { 1107 timings_.StartSplit("FreeList"); 1108 freed_objects += chunk_free_pos; 1109 freed_bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1110 timings_.EndSplit(); 1111 chunk_free_pos = 0; 1112 } 1113 chunk_free_buffer[chunk_free_pos++] = obj; 1114 } 1115 } else { 1116 *(out++) = obj; 1117 } 1118 } 1119 if (chunk_free_pos > 0) { 1120 timings_.StartSplit("FreeList"); 1121 freed_objects += chunk_free_pos; 1122 freed_bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1123 timings_.EndSplit(); 1124 chunk_free_pos = 0; 1125 } 1126 // All of the references which space contained are no longer in the allocation stack, update 1127 // the count. 1128 count = out - objects; 1129 } 1130 // Handle the large object space. 1131 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 1132 accounting::ObjectSet* large_live_objects = large_object_space->GetLiveObjects(); 1133 accounting::ObjectSet* large_mark_objects = large_object_space->GetMarkObjects(); 1134 if (swap_bitmaps) { 1135 std::swap(large_live_objects, large_mark_objects); 1136 } 1137 for (size_t i = 0; i < count; ++i) { 1138 Object* obj = objects[i]; 1139 // Handle large objects. 1140 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1141 continue; 1142 } 1143 if (!large_mark_objects->Test(obj)) { 1144 ++freed_large_objects; 1145 freed_large_object_bytes += large_object_space->Free(self, obj); 1146 } 1147 } 1148 timings_.EndSplit(); 1149 1150 timings_.StartSplit("RecordFree"); 1151 VLOG(heap) << "Freed " << freed_objects << "/" << count 1152 << " objects with size " << PrettySize(freed_bytes); 1153 heap_->RecordFree(freed_objects + freed_large_objects, freed_bytes + freed_large_object_bytes); 1154 freed_objects_.FetchAndAdd(freed_objects); 1155 freed_large_objects_.FetchAndAdd(freed_large_objects); 1156 freed_bytes_.FetchAndAdd(freed_bytes); 1157 freed_large_object_bytes_.FetchAndAdd(freed_large_object_bytes); 1158 timings_.EndSplit(); 1159 1160 timings_.StartSplit("ResetStack"); 1161 allocations->Reset(); 1162 timings_.EndSplit(); 1163} 1164 1165void MarkSweep::Sweep(bool swap_bitmaps) { 1166 DCHECK(mark_stack_->IsEmpty()); 1167 TimingLogger::ScopedSplit("Sweep", &timings_); 1168 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 1169 if (space->IsContinuousMemMapAllocSpace()) { 1170 space::ContinuousMemMapAllocSpace* alloc_space = space->AsContinuousMemMapAllocSpace(); 1171 TimingLogger::ScopedSplit split( 1172 alloc_space->IsZygoteSpace() ? "SweepZygoteSpace" : "SweepMallocSpace", &timings_); 1173 size_t freed_objects = 0; 1174 size_t freed_bytes = 0; 1175 alloc_space->Sweep(swap_bitmaps, &freed_objects, &freed_bytes); 1176 heap_->RecordFree(freed_objects, freed_bytes); 1177 freed_objects_.FetchAndAdd(freed_objects); 1178 freed_bytes_.FetchAndAdd(freed_bytes); 1179 } 1180 } 1181 SweepLargeObjects(swap_bitmaps); 1182} 1183 1184void MarkSweep::SweepLargeObjects(bool swap_bitmaps) { 1185 TimingLogger::ScopedSplit("SweepLargeObjects", &timings_); 1186 size_t freed_objects = 0; 1187 size_t freed_bytes = 0; 1188 GetHeap()->GetLargeObjectsSpace()->Sweep(swap_bitmaps, &freed_objects, &freed_bytes); 1189 freed_large_objects_.FetchAndAdd(freed_objects); 1190 freed_large_object_bytes_.FetchAndAdd(freed_bytes); 1191 GetHeap()->RecordFree(freed_objects, freed_bytes); 1192} 1193 1194// Process the "referent" field in a java.lang.ref.Reference. If the 1195// referent has not yet been marked, put it on the appropriate list in 1196// the heap for later processing. 1197void MarkSweep::DelayReferenceReferent(mirror::Class* klass, Object* obj) { 1198 DCHECK(klass != nullptr); 1199 heap_->DelayReferenceReferent(klass, obj->AsReference(), IsMarkedCallback, this); 1200} 1201 1202class MarkObjectVisitor { 1203 public: 1204 explicit MarkObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE : mark_sweep_(mark_sweep) {} 1205 1206 // TODO: Fixme when anotatalysis works with visitors. 1207 void operator()(const Object* /* obj */, const Object* ref, const MemberOffset& /* offset */, 1208 bool /* is_static */) const ALWAYS_INLINE 1209 NO_THREAD_SAFETY_ANALYSIS { 1210 if (kCheckLocks) { 1211 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 1212 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 1213 } 1214 mark_sweep_->MarkObject(ref); 1215 } 1216 1217 private: 1218 MarkSweep* const mark_sweep_; 1219}; 1220 1221// Scans an object reference. Determines the type of the reference 1222// and dispatches to a specialized scanning routine. 1223void MarkSweep::ScanObject(Object* obj) { 1224 MarkObjectVisitor visitor(this); 1225 ScanObjectVisit(obj, visitor); 1226} 1227 1228void MarkSweep::ProcessMarkStackPausedCallback(void* arg) { 1229 DCHECK(arg != nullptr); 1230 reinterpret_cast<MarkSweep*>(arg)->ProcessMarkStack(true); 1231} 1232 1233void MarkSweep::ProcessMarkStackParallel(size_t thread_count) { 1234 Thread* self = Thread::Current(); 1235 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 1236 const size_t chunk_size = std::min(mark_stack_->Size() / thread_count + 1, 1237 static_cast<size_t>(MarkStackTask<false>::kMaxSize)); 1238 CHECK_GT(chunk_size, 0U); 1239 // Split the current mark stack up into work tasks. 1240 for (mirror::Object **it = mark_stack_->Begin(), **end = mark_stack_->End(); it < end; ) { 1241 const size_t delta = std::min(static_cast<size_t>(end - it), chunk_size); 1242 thread_pool->AddTask(self, new MarkStackTask<false>(thread_pool, this, delta, 1243 const_cast<const mirror::Object**>(it))); 1244 it += delta; 1245 } 1246 thread_pool->SetMaxActiveWorkers(thread_count - 1); 1247 thread_pool->StartWorkers(self); 1248 thread_pool->Wait(self, true, true); 1249 thread_pool->StopWorkers(self); 1250 mark_stack_->Reset(); 1251 CHECK_EQ(work_chunks_created_, work_chunks_deleted_) << " some of the work chunks were leaked"; 1252} 1253 1254// Scan anything that's on the mark stack. 1255void MarkSweep::ProcessMarkStack(bool paused) { 1256 timings_.StartSplit(paused ? "(Paused)ProcessMarkStack" : "ProcessMarkStack"); 1257 size_t thread_count = GetThreadCount(paused); 1258 if (kParallelProcessMarkStack && thread_count > 1 && 1259 mark_stack_->Size() >= kMinimumParallelMarkStackSize) { 1260 ProcessMarkStackParallel(thread_count); 1261 } else { 1262 // TODO: Tune this. 1263 static const size_t kFifoSize = 4; 1264 BoundedFifoPowerOfTwo<Object*, kFifoSize> prefetch_fifo; 1265 for (;;) { 1266 Object* obj = NULL; 1267 if (kUseMarkStackPrefetch) { 1268 while (!mark_stack_->IsEmpty() && prefetch_fifo.size() < kFifoSize) { 1269 Object* obj = mark_stack_->PopBack(); 1270 DCHECK(obj != NULL); 1271 __builtin_prefetch(obj); 1272 prefetch_fifo.push_back(obj); 1273 } 1274 if (prefetch_fifo.empty()) { 1275 break; 1276 } 1277 obj = prefetch_fifo.front(); 1278 prefetch_fifo.pop_front(); 1279 } else { 1280 if (mark_stack_->IsEmpty()) { 1281 break; 1282 } 1283 obj = mark_stack_->PopBack(); 1284 } 1285 DCHECK(obj != NULL); 1286 ScanObject(obj); 1287 } 1288 } 1289 timings_.EndSplit(); 1290} 1291 1292inline bool MarkSweep::IsMarked(const Object* object) const 1293 SHARED_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 1294 if (immune_region_.ContainsObject(object)) { 1295 return true; 1296 } 1297 DCHECK(current_mark_bitmap_ != NULL); 1298 if (current_mark_bitmap_->HasAddress(object)) { 1299 return current_mark_bitmap_->Test(object); 1300 } 1301 return heap_->GetMarkBitmap()->Test(object); 1302} 1303 1304void MarkSweep::FinishPhase() { 1305 TimingLogger::ScopedSplit split("FinishPhase", &timings_); 1306 // Can't enqueue references if we hold the mutator lock. 1307 Heap* heap = GetHeap(); 1308 timings_.NewSplit("PostGcVerification"); 1309 heap->PostGcVerification(this); 1310 1311 // Update the cumulative statistics 1312 total_freed_objects_ += GetFreedObjects() + GetFreedLargeObjects(); 1313 total_freed_bytes_ += GetFreedBytes() + GetFreedLargeObjectBytes(); 1314 1315 // Ensure that the mark stack is empty. 1316 CHECK(mark_stack_->IsEmpty()); 1317 1318 if (kCountScannedTypes) { 1319 VLOG(gc) << "MarkSweep scanned classes=" << class_count_ << " arrays=" << array_count_ 1320 << " other=" << other_count_; 1321 } 1322 1323 if (kCountTasks) { 1324 VLOG(gc) << "Total number of work chunks allocated: " << work_chunks_created_; 1325 } 1326 1327 if (kMeasureOverhead) { 1328 VLOG(gc) << "Overhead time " << PrettyDuration(overhead_time_); 1329 } 1330 1331 if (kProfileLargeObjects) { 1332 VLOG(gc) << "Large objects tested " << large_object_test_ << " marked " << large_object_mark_; 1333 } 1334 1335 if (kCountClassesMarked) { 1336 VLOG(gc) << "Classes marked " << classes_marked_; 1337 } 1338 1339 if (kCountJavaLangRefs) { 1340 VLOG(gc) << "References scanned " << reference_count_; 1341 } 1342 1343 // Update the cumulative loggers. 1344 cumulative_timings_.Start(); 1345 cumulative_timings_.AddLogger(timings_); 1346 cumulative_timings_.End(); 1347 1348 // Clear all of the spaces' mark bitmaps. 1349 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 1350 accounting::SpaceBitmap* bitmap = space->GetMarkBitmap(); 1351 if (bitmap != nullptr && 1352 space->GetGcRetentionPolicy() != space::kGcRetentionPolicyNeverCollect) { 1353 bitmap->Clear(); 1354 } 1355 } 1356 mark_stack_->Reset(); 1357 1358 // Reset the marked large objects. 1359 space::LargeObjectSpace* large_objects = GetHeap()->GetLargeObjectsSpace(); 1360 large_objects->GetMarkObjects()->Clear(); 1361} 1362 1363} // namespace collector 1364} // namespace gc 1365} // namespace art 1366