mark_sweep.cc revision 0e54cd0d8fc635d3dc8bf88a465fdade151a098f
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 kProfileLargeObjects = false; 83static constexpr bool kMeasureOverhead = false; 84static constexpr bool kCountTasks = false; 85static constexpr bool kCountJavaLangRefs = false; 86static constexpr bool kCountMarkedObjects = 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 92// If true, revoke the rosalloc thread-local buffers at the 93// checkpoint, as opposed to during the pause. 94static constexpr bool kRevokeRosAllocThreadLocalBuffersAtCheckpoint = true; 95 96void MarkSweep::BindBitmaps() { 97 timings_.StartSplit("BindBitmaps"); 98 WriterMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 99 // Mark all of the spaces we never collect as immune. 100 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 101 if (space->GetGcRetentionPolicy() == space::kGcRetentionPolicyNeverCollect) { 102 CHECK(immune_region_.AddContinuousSpace(space)) << "Failed to add space " << *space; 103 } 104 } 105 timings_.EndSplit(); 106} 107 108MarkSweep::MarkSweep(Heap* heap, bool is_concurrent, const std::string& name_prefix) 109 : GarbageCollector(heap, 110 name_prefix + 111 (is_concurrent ? "concurrent mark sweep": "mark sweep")), 112 gc_barrier_(new Barrier(0)), 113 large_object_lock_("mark sweep large object lock", kMarkSweepLargeObjectLock), 114 mark_stack_lock_("mark sweep mark stack lock", kMarkSweepMarkStackLock), 115 is_concurrent_(is_concurrent) { 116} 117 118void MarkSweep::InitializePhase() { 119 timings_.Reset(); 120 TimingLogger::ScopedSplit split("InitializePhase", &timings_); 121 mark_stack_ = heap_->mark_stack_.get(); 122 DCHECK(mark_stack_ != nullptr); 123 immune_region_.Reset(); 124 class_count_ = 0; 125 array_count_ = 0; 126 other_count_ = 0; 127 large_object_test_ = 0; 128 large_object_mark_ = 0; 129 overhead_time_ = 0; 130 work_chunks_created_ = 0; 131 work_chunks_deleted_ = 0; 132 reference_count_ = 0; 133 mark_null_count_ = 0; 134 mark_immune_count_ = 0; 135 mark_fastpath_count_ = 0; 136 mark_slowpath_count_ = 0; 137 FindDefaultSpaceBitmap(); 138 { 139 // TODO: I don't think we should need heap bitmap lock to get the mark bitmap. 140 ReaderMutexLock mu(Thread::Current(), *Locks::heap_bitmap_lock_); 141 mark_bitmap_ = heap_->GetMarkBitmap(); 142 } 143 144 // Do any pre GC verification. 145 timings_.NewSplit("PreGcVerification"); 146 heap_->PreGcVerification(this); 147} 148 149void MarkSweep::ProcessReferences(Thread* self) { 150 TimingLogger::ScopedSplit split("ProcessReferences", &timings_); 151 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 152 GetHeap()->ProcessReferences(timings_, clear_soft_references_, &IsMarkedCallback, 153 &MarkObjectCallback, &ProcessMarkStackPausedCallback, this); 154} 155 156void MarkSweep::PreProcessReferences() { 157 if (IsConcurrent()) { 158 // No reason to do this for non-concurrent GC since pre processing soft references only helps 159 // pauses. 160 timings_.NewSplit("PreProcessReferences"); 161 GetHeap()->ProcessSoftReferences(timings_, clear_soft_references_, &IsMarkedCallback, 162 &MarkObjectCallback, &ProcessMarkStackPausedCallback, this); 163 } 164} 165 166void MarkSweep::HandleDirtyObjectsPhase() { 167 TimingLogger::ScopedSplit split("(Paused)HandleDirtyObjectsPhase", &timings_); 168 Thread* self = Thread::Current(); 169 Locks::mutator_lock_->AssertExclusiveHeld(self); 170 171 { 172 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 173 174 // Re-mark root set. 175 ReMarkRoots(); 176 177 // Scan dirty objects, this is only required if we are not doing concurrent GC. 178 RecursiveMarkDirtyObjects(true, accounting::CardTable::kCardDirty); 179 } 180 181 ProcessReferences(self); 182 183 // Only need to do this if we have the card mark verification on, and only during concurrent GC. 184 if (GetHeap()->verify_missing_card_marks_ || GetHeap()->verify_pre_gc_heap_|| 185 GetHeap()->verify_post_gc_heap_) { 186 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 187 // This second sweep makes sure that we don't have any objects in the live stack which point to 188 // freed objects. These cause problems since their references may be previously freed objects. 189 SweepArray(GetHeap()->allocation_stack_.get(), false); 190 // Since SweepArray() above resets the (active) allocation 191 // stack. Need to revoke the thread-local allocation stacks that 192 // point into it. 193 RevokeAllThreadLocalAllocationStacks(self); 194 } 195 196 timings_.StartSplit("PreSweepingGcVerification"); 197 heap_->PreSweepingGcVerification(this); 198 timings_.EndSplit(); 199 200 // Ensure that nobody inserted items in the live stack after we swapped the stacks. 201 ReaderMutexLock mu(self, *Locks::heap_bitmap_lock_); 202 CHECK_GE(live_stack_freeze_size_, GetHeap()->GetLiveStack()->Size()); 203 204 // Disallow new system weaks to prevent a race which occurs when someone adds a new system 205 // weak before we sweep them. Since this new system weak may not be marked, the GC may 206 // incorrectly sweep it. This also fixes a race where interning may attempt to return a strong 207 // reference to a string that is about to be swept. 208 Runtime::Current()->DisallowNewSystemWeaks(); 209} 210 211void MarkSweep::PreCleanCards() { 212 // Don't do this for non concurrent GCs since they don't have any dirty cards. 213 if (kPreCleanCards && IsConcurrent()) { 214 Thread* self = Thread::Current(); 215 CHECK(!Locks::mutator_lock_->IsExclusiveHeld(self)); 216 // Process dirty cards and add dirty cards to mod union tables, also ages cards. 217 heap_->ProcessCards(timings_, false); 218 // The checkpoint root marking is required to avoid a race condition which occurs if the 219 // following happens during a reference write: 220 // 1. mutator dirties the card (write barrier) 221 // 2. GC ages the card (the above ProcessCards call) 222 // 3. GC scans the object (the RecursiveMarkDirtyObjects call below) 223 // 4. mutator writes the value (corresponding to the write barrier in 1.) 224 // This causes the GC to age the card but not necessarily mark the reference which the mutator 225 // wrote into the object stored in the card. 226 // Having the checkpoint fixes this issue since it ensures that the card mark and the 227 // reference write are visible to the GC before the card is scanned (this is due to locks being 228 // acquired / released in the checkpoint code). 229 // The other roots are also marked to help reduce the pause. 230 MarkThreadRoots(self); 231 // TODO: Only mark the dirty roots. 232 MarkNonThreadRoots(); 233 MarkConcurrentRoots( 234 static_cast<VisitRootFlags>(kVisitRootFlagClearRootLog | kVisitRootFlagNewRoots)); 235 // Process the newly aged cards. 236 RecursiveMarkDirtyObjects(false, accounting::CardTable::kCardDirty - 1); 237 // TODO: Empty allocation stack to reduce the number of objects we need to test / mark as live 238 // in the next GC. 239 } 240} 241 242void MarkSweep::RevokeAllThreadLocalAllocationStacks(Thread* self) { 243 if (kUseThreadLocalAllocationStack) { 244 Locks::mutator_lock_->AssertExclusiveHeld(self); 245 heap_->RevokeAllThreadLocalAllocationStacks(self); 246 } 247} 248 249void MarkSweep::MarkingPhase() { 250 TimingLogger::ScopedSplit split("MarkingPhase", &timings_); 251 Thread* self = Thread::Current(); 252 253 BindBitmaps(); 254 FindDefaultSpaceBitmap(); 255 256 // Process dirty cards and add dirty cards to mod union tables. 257 heap_->ProcessCards(timings_, false); 258 259 // Need to do this before the checkpoint since we don't want any threads to add references to 260 // the live stack during the recursive mark. 261 timings_.NewSplit("SwapStacks"); 262 heap_->SwapStacks(self); 263 264 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 265 MarkRoots(self); 266 live_stack_freeze_size_ = heap_->GetLiveStack()->Size(); 267 UpdateAndMarkModUnion(); 268 MarkReachableObjects(); 269 // Pre-clean dirtied cards to reduce pauses. 270 PreCleanCards(); 271 PreProcessReferences(); 272} 273 274void MarkSweep::UpdateAndMarkModUnion() { 275 for (const auto& space : heap_->GetContinuousSpaces()) { 276 if (immune_region_.ContainsSpace(space)) { 277 const char* name = space->IsZygoteSpace() ? "UpdateAndMarkZygoteModUnionTable" : 278 "UpdateAndMarkImageModUnionTable"; 279 TimingLogger::ScopedSplit split(name, &timings_); 280 accounting::ModUnionTable* mod_union_table = heap_->FindModUnionTableFromSpace(space); 281 CHECK(mod_union_table != nullptr); 282 mod_union_table->UpdateAndMarkReferences(MarkObjectCallback, this); 283 } 284 } 285} 286 287void MarkSweep::MarkThreadRoots(Thread* self) { 288 MarkRootsCheckpoint(self); 289} 290 291void MarkSweep::MarkReachableObjects() { 292 // Mark everything allocated since the last as GC live so that we can sweep concurrently, 293 // knowing that new allocations won't be marked as live. 294 timings_.StartSplit("MarkStackAsLive"); 295 accounting::ObjectStack* live_stack = heap_->GetLiveStack(); 296 heap_->MarkAllocStackAsLive(live_stack); 297 live_stack->Reset(); 298 timings_.EndSplit(); 299 // Recursively mark all the non-image bits set in the mark bitmap. 300 RecursiveMark(); 301} 302 303void MarkSweep::ReclaimPhase() { 304 TimingLogger::ScopedSplit split("ReclaimPhase", &timings_); 305 Thread* self = Thread::Current(); 306 307 if (!IsConcurrent()) { 308 ProcessReferences(self); 309 } 310 311 { 312 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 313 SweepSystemWeaks(); 314 } 315 316 if (IsConcurrent()) { 317 Runtime::Current()->AllowNewSystemWeaks(); 318 319 TimingLogger::ScopedSplit split("UnMarkAllocStack", &timings_); 320 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 321 accounting::ObjectStack* allocation_stack = GetHeap()->allocation_stack_.get(); 322 if (!kPreCleanCards) { 323 // The allocation stack contains things allocated since the start of the GC. These may have 324 // been marked during this GC meaning they won't be eligible for reclaiming in the next 325 // sticky GC. Unmark these objects so that they are eligible for reclaiming in the next 326 // sticky GC. 327 // There is a race here which is safely handled. Another thread such as the hprof could 328 // have flushed the alloc stack after we resumed the threads. This is safe however, since 329 // reseting the allocation stack zeros it out with madvise. This means that we will either 330 // read NULLs or attempt to unmark a newly allocated object which will not be marked in the 331 // first place. 332 // We can't do this if we pre-clean cards since we will unmark objects which are no longer on 333 // a dirty card since we aged cards during the pre-cleaning process. 334 mirror::Object** end = allocation_stack->End(); 335 for (mirror::Object** it = allocation_stack->Begin(); it != end; ++it) { 336 const Object* obj = *it; 337 if (obj != nullptr) { 338 UnMarkObjectNonNull(obj); 339 } 340 } 341 } 342 } 343 344 { 345 WriterMutexLock mu(self, *Locks::heap_bitmap_lock_); 346 347 // Reclaim unmarked objects. 348 Sweep(false); 349 350 // Swap the live and mark bitmaps for each space which we modified space. This is an 351 // optimization that enables us to not clear live bits inside of the sweep. Only swaps unbound 352 // bitmaps. 353 timings_.StartSplit("SwapBitmaps"); 354 SwapBitmaps(); 355 timings_.EndSplit(); 356 357 // Unbind the live and mark bitmaps. 358 TimingLogger::ScopedSplit split("UnBindBitmaps", &timings_); 359 GetHeap()->UnBindBitmaps(); 360 } 361} 362 363void MarkSweep::FindDefaultSpaceBitmap() { 364 TimingLogger::ScopedSplit split("FindDefaultMarkBitmap", &timings_); 365 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 366 accounting::SpaceBitmap* bitmap = space->GetMarkBitmap(); 367 if (bitmap != nullptr && 368 space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) { 369 current_space_bitmap_ = bitmap; 370 return; 371 } 372 } 373 GetHeap()->DumpSpaces(); 374 LOG(FATAL) << "Could not find a default mark bitmap"; 375} 376 377void MarkSweep::ExpandMarkStack() { 378 ResizeMarkStack(mark_stack_->Capacity() * 2); 379} 380 381void MarkSweep::ResizeMarkStack(size_t new_size) { 382 // Rare case, no need to have Thread::Current be a parameter. 383 if (UNLIKELY(mark_stack_->Size() < mark_stack_->Capacity())) { 384 // Someone else acquired the lock and expanded the mark stack before us. 385 return; 386 } 387 std::vector<Object*> temp(mark_stack_->Begin(), mark_stack_->End()); 388 CHECK_LE(mark_stack_->Size(), new_size); 389 mark_stack_->Resize(new_size); 390 for (const auto& obj : temp) { 391 mark_stack_->PushBack(obj); 392 } 393} 394 395inline void MarkSweep::MarkObjectNonNullParallel(Object* obj) { 396 DCHECK(obj != NULL); 397 if (MarkObjectParallel(obj)) { 398 MutexLock mu(Thread::Current(), mark_stack_lock_); 399 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 400 ExpandMarkStack(); 401 } 402 // The object must be pushed on to the mark stack. 403 mark_stack_->PushBack(obj); 404 } 405} 406 407mirror::Object* MarkSweep::MarkObjectCallback(mirror::Object* obj, void* arg) { 408 MarkSweep* mark_sweep = reinterpret_cast<MarkSweep*>(arg); 409 mark_sweep->MarkObject(obj); 410 return obj; 411} 412 413inline void MarkSweep::UnMarkObjectNonNull(const Object* obj) { 414 DCHECK(!immune_region_.ContainsObject(obj)); 415 if (kUseBrooksPointer) { 416 // Verify all the objects have the correct Brooks pointer installed. 417 obj->AssertSelfBrooksPointer(); 418 } 419 // Try to take advantage of locality of references within a space, failing this find the space 420 // the hard way. 421 accounting::SpaceBitmap* object_bitmap = current_space_bitmap_; 422 if (UNLIKELY(!object_bitmap->HasAddress(obj))) { 423 accounting::SpaceBitmap* new_bitmap = mark_bitmap_->GetContinuousSpaceBitmap(obj); 424 if (LIKELY(new_bitmap != NULL)) { 425 object_bitmap = new_bitmap; 426 } else { 427 MarkLargeObject(obj, false); 428 return; 429 } 430 } 431 DCHECK(object_bitmap->HasAddress(obj)); 432 object_bitmap->Clear(obj); 433} 434 435inline void MarkSweep::MarkObjectNonNull(Object* obj) { 436 DCHECK(obj != nullptr); 437 if (kUseBrooksPointer) { 438 // Verify all the objects have the correct Brooks pointer installed. 439 obj->AssertSelfBrooksPointer(); 440 } 441 if (immune_region_.ContainsObject(obj)) { 442 if (kCountMarkedObjects) { 443 ++mark_immune_count_; 444 } 445 DCHECK(IsMarked(obj)); 446 return; 447 } 448 // Try to take advantage of locality of references within a space, failing this find the space 449 // the hard way. 450 accounting::SpaceBitmap* object_bitmap = current_space_bitmap_; 451 if (UNLIKELY(!object_bitmap->HasAddress(obj))) { 452 object_bitmap = mark_bitmap_->GetContinuousSpaceBitmap(obj); 453 if (kCountMarkedObjects) { 454 ++mark_slowpath_count_; 455 } 456 if (UNLIKELY(object_bitmap == nullptr)) { 457 MarkLargeObject(obj, true); 458 return; 459 } 460 } else if (kCountMarkedObjects) { 461 ++mark_fastpath_count_; 462 } 463 // This object was not previously marked. 464 if (!object_bitmap->Set(obj)) { 465 PushOnMarkStack(obj); 466 } 467} 468 469inline void MarkSweep::PushOnMarkStack(Object* obj) { 470 if (UNLIKELY(mark_stack_->Size() >= mark_stack_->Capacity())) { 471 // Lock is not needed but is here anyways to please annotalysis. 472 MutexLock mu(Thread::Current(), mark_stack_lock_); 473 ExpandMarkStack(); 474 } 475 // The object must be pushed on to the mark stack. 476 mark_stack_->PushBack(obj); 477} 478 479// Rare case, probably not worth inlining since it will increase instruction cache miss rate. 480bool MarkSweep::MarkLargeObject(const Object* obj, bool set) { 481 // TODO: support >1 discontinuous space. 482 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 483 accounting::ObjectSet* large_objects = large_object_space->GetMarkObjects(); 484 if (kProfileLargeObjects) { 485 ++large_object_test_; 486 } 487 if (UNLIKELY(!large_objects->Test(obj))) { 488 if (!large_object_space->Contains(obj)) { 489 LOG(ERROR) << "Tried to mark " << obj << " not contained by any spaces"; 490 LOG(ERROR) << "Attempting see if it's a bad root"; 491 VerifyRoots(); 492 LOG(FATAL) << "Can't mark bad root"; 493 } 494 if (kProfileLargeObjects) { 495 ++large_object_mark_; 496 } 497 if (set) { 498 large_objects->Set(obj); 499 } else { 500 large_objects->Clear(obj); 501 } 502 return true; 503 } 504 return false; 505} 506 507inline bool MarkSweep::MarkObjectParallel(const Object* obj) { 508 DCHECK(obj != nullptr); 509 if (kUseBrooksPointer) { 510 // Verify all the objects have the correct Brooks pointer installed. 511 obj->AssertSelfBrooksPointer(); 512 } 513 if (immune_region_.ContainsObject(obj)) { 514 DCHECK(IsMarked(obj)); 515 return false; 516 } 517 // Try to take advantage of locality of references within a space, failing this find the space 518 // the hard way. 519 accounting::SpaceBitmap* object_bitmap = current_space_bitmap_; 520 if (UNLIKELY(!object_bitmap->HasAddress(obj))) { 521 accounting::SpaceBitmap* new_bitmap = mark_bitmap_->GetContinuousSpaceBitmap(obj); 522 if (new_bitmap != NULL) { 523 object_bitmap = new_bitmap; 524 } else { 525 // TODO: Remove the Thread::Current here? 526 // TODO: Convert this to some kind of atomic marking? 527 MutexLock mu(Thread::Current(), large_object_lock_); 528 return MarkLargeObject(obj, true); 529 } 530 } 531 // Return true if the object was not previously marked. 532 return !object_bitmap->AtomicTestAndSet(obj); 533} 534 535// Used to mark objects when processing the mark stack. If an object is null, it is not marked. 536inline void MarkSweep::MarkObject(Object* obj) { 537 if (obj != nullptr) { 538 MarkObjectNonNull(obj); 539 } else if (kCountMarkedObjects) { 540 ++mark_null_count_; 541 } 542} 543 544void MarkSweep::MarkRootParallelCallback(Object** root, void* arg, uint32_t /*thread_id*/, 545 RootType /*root_type*/) { 546 reinterpret_cast<MarkSweep*>(arg)->MarkObjectNonNullParallel(*root); 547} 548 549void MarkSweep::VerifyRootMarked(Object** root, void* arg, uint32_t /*thread_id*/, 550 RootType /*root_type*/) { 551 CHECK(reinterpret_cast<MarkSweep*>(arg)->IsMarked(*root)); 552} 553 554void MarkSweep::MarkRootCallback(Object** root, void* arg, uint32_t /*thread_id*/, 555 RootType /*root_type*/) { 556 reinterpret_cast<MarkSweep*>(arg)->MarkObjectNonNull(*root); 557} 558 559void MarkSweep::VerifyRootCallback(const Object* root, void* arg, size_t vreg, 560 const StackVisitor* visitor) { 561 reinterpret_cast<MarkSweep*>(arg)->VerifyRoot(root, vreg, visitor); 562} 563 564void MarkSweep::VerifyRoot(const Object* root, size_t vreg, const StackVisitor* visitor) { 565 // See if the root is on any space bitmap. 566 if (GetHeap()->GetLiveBitmap()->GetContinuousSpaceBitmap(root) == NULL) { 567 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 568 if (!large_object_space->Contains(root)) { 569 LOG(ERROR) << "Found invalid root: " << root; 570 if (visitor != NULL) { 571 LOG(ERROR) << visitor->DescribeLocation() << " in VReg: " << vreg; 572 } 573 } 574 } 575} 576 577void MarkSweep::VerifyRoots() { 578 Runtime::Current()->GetThreadList()->VerifyRoots(VerifyRootCallback, this); 579} 580 581void MarkSweep::MarkRoots(Thread* self) { 582 if (Locks::mutator_lock_->IsExclusiveHeld(self)) { 583 // If we exclusively hold the mutator lock, all threads must be suspended. 584 timings_.StartSplit("MarkRoots"); 585 Runtime::Current()->VisitRoots(MarkRootCallback, this); 586 timings_.EndSplit(); 587 RevokeAllThreadLocalAllocationStacks(self); 588 } else { 589 MarkThreadRoots(self); 590 // At this point the live stack should no longer have any mutators which push into it. 591 MarkNonThreadRoots(); 592 MarkConcurrentRoots( 593 static_cast<VisitRootFlags>(kVisitRootFlagAllRoots | kVisitRootFlagStartLoggingNewRoots)); 594 } 595} 596 597void MarkSweep::MarkNonThreadRoots() { 598 timings_.StartSplit("MarkNonThreadRoots"); 599 Runtime::Current()->VisitNonThreadRoots(MarkRootCallback, this); 600 timings_.EndSplit(); 601} 602 603void MarkSweep::MarkConcurrentRoots(VisitRootFlags flags) { 604 timings_.StartSplit("MarkConcurrentRoots"); 605 // Visit all runtime roots and clear dirty flags. 606 Runtime::Current()->VisitConcurrentRoots(MarkRootCallback, this, flags); 607 timings_.EndSplit(); 608} 609 610class ScanObjectVisitor { 611 public: 612 explicit ScanObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE 613 : mark_sweep_(mark_sweep) {} 614 615 // TODO: Fixme when anotatalysis works with visitors. 616 void operator()(Object* obj) const ALWAYS_INLINE NO_THREAD_SAFETY_ANALYSIS { 617 if (kCheckLocks) { 618 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 619 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 620 } 621 mark_sweep_->ScanObject(obj); 622 } 623 624 private: 625 MarkSweep* const mark_sweep_; 626}; 627 628template <bool kUseFinger = false> 629class MarkStackTask : public Task { 630 public: 631 MarkStackTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, size_t mark_stack_size, 632 Object** mark_stack) 633 : mark_sweep_(mark_sweep), 634 thread_pool_(thread_pool), 635 mark_stack_pos_(mark_stack_size) { 636 // We may have to copy part of an existing mark stack when another mark stack overflows. 637 if (mark_stack_size != 0) { 638 DCHECK(mark_stack != NULL); 639 // TODO: Check performance? 640 std::copy(mark_stack, mark_stack + mark_stack_size, mark_stack_); 641 } 642 if (kCountTasks) { 643 ++mark_sweep_->work_chunks_created_; 644 } 645 } 646 647 static const size_t kMaxSize = 1 * KB; 648 649 protected: 650 class ScanObjectParallelVisitor { 651 public: 652 explicit ScanObjectParallelVisitor(MarkStackTask<kUseFinger>* chunk_task) ALWAYS_INLINE 653 : chunk_task_(chunk_task) {} 654 655 void operator()(Object* obj) const { 656 MarkSweep* mark_sweep = chunk_task_->mark_sweep_; 657 mark_sweep->ScanObjectVisit(obj, 658 [mark_sweep, this](Object* /* obj */, Object* ref, const MemberOffset& /* offset */, 659 bool /* is_static */) ALWAYS_INLINE_LAMBDA { 660 if (ref != nullptr && mark_sweep->MarkObjectParallel(ref)) { 661 if (kUseFinger) { 662 android_memory_barrier(); 663 if (reinterpret_cast<uintptr_t>(ref) >= 664 static_cast<uintptr_t>(mark_sweep->atomic_finger_)) { 665 return; 666 } 667 } 668 chunk_task_->MarkStackPush(ref); 669 } 670 }); 671 } 672 673 private: 674 MarkStackTask<kUseFinger>* const chunk_task_; 675 }; 676 677 virtual ~MarkStackTask() { 678 // Make sure that we have cleared our mark stack. 679 DCHECK_EQ(mark_stack_pos_, 0U); 680 if (kCountTasks) { 681 ++mark_sweep_->work_chunks_deleted_; 682 } 683 } 684 685 MarkSweep* const mark_sweep_; 686 ThreadPool* const thread_pool_; 687 // Thread local mark stack for this task. 688 Object* mark_stack_[kMaxSize]; 689 // Mark stack position. 690 size_t mark_stack_pos_; 691 692 void MarkStackPush(Object* obj) ALWAYS_INLINE { 693 if (UNLIKELY(mark_stack_pos_ == kMaxSize)) { 694 // Mark stack overflow, give 1/2 the stack to the thread pool as a new work task. 695 mark_stack_pos_ /= 2; 696 auto* task = new MarkStackTask(thread_pool_, mark_sweep_, kMaxSize - mark_stack_pos_, 697 mark_stack_ + mark_stack_pos_); 698 thread_pool_->AddTask(Thread::Current(), task); 699 } 700 DCHECK(obj != nullptr); 701 DCHECK_LT(mark_stack_pos_, kMaxSize); 702 mark_stack_[mark_stack_pos_++] = obj; 703 } 704 705 virtual void Finalize() { 706 delete this; 707 } 708 709 // Scans all of the objects 710 virtual void Run(Thread* self) { 711 ScanObjectParallelVisitor visitor(this); 712 // TODO: Tune this. 713 static const size_t kFifoSize = 4; 714 BoundedFifoPowerOfTwo<Object*, kFifoSize> prefetch_fifo; 715 for (;;) { 716 Object* obj = nullptr; 717 if (kUseMarkStackPrefetch) { 718 while (mark_stack_pos_ != 0 && prefetch_fifo.size() < kFifoSize) { 719 Object* obj = mark_stack_[--mark_stack_pos_]; 720 DCHECK(obj != nullptr); 721 __builtin_prefetch(obj); 722 prefetch_fifo.push_back(obj); 723 } 724 if (UNLIKELY(prefetch_fifo.empty())) { 725 break; 726 } 727 obj = prefetch_fifo.front(); 728 prefetch_fifo.pop_front(); 729 } else { 730 if (UNLIKELY(mark_stack_pos_ == 0)) { 731 break; 732 } 733 obj = mark_stack_[--mark_stack_pos_]; 734 } 735 DCHECK(obj != nullptr); 736 visitor(obj); 737 } 738 } 739}; 740 741class CardScanTask : public MarkStackTask<false> { 742 public: 743 CardScanTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, accounting::SpaceBitmap* bitmap, 744 byte* begin, byte* end, byte minimum_age, size_t mark_stack_size, 745 Object** mark_stack_obj) 746 : MarkStackTask<false>(thread_pool, mark_sweep, mark_stack_size, mark_stack_obj), 747 bitmap_(bitmap), 748 begin_(begin), 749 end_(end), 750 minimum_age_(minimum_age) { 751 } 752 753 protected: 754 accounting::SpaceBitmap* const bitmap_; 755 byte* const begin_; 756 byte* const end_; 757 const byte minimum_age_; 758 759 virtual void Finalize() { 760 delete this; 761 } 762 763 virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS { 764 ScanObjectParallelVisitor visitor(this); 765 accounting::CardTable* card_table = mark_sweep_->GetHeap()->GetCardTable(); 766 size_t cards_scanned = card_table->Scan(bitmap_, begin_, end_, visitor, minimum_age_); 767 VLOG(heap) << "Parallel scanning cards " << reinterpret_cast<void*>(begin_) << " - " 768 << reinterpret_cast<void*>(end_) << " = " << cards_scanned; 769 // Finish by emptying our local mark stack. 770 MarkStackTask::Run(self); 771 } 772}; 773 774size_t MarkSweep::GetThreadCount(bool paused) const { 775 if (heap_->GetThreadPool() == nullptr || !heap_->CareAboutPauseTimes()) { 776 return 0; 777 } 778 if (paused) { 779 return heap_->GetParallelGCThreadCount() + 1; 780 } else { 781 return heap_->GetConcGCThreadCount() + 1; 782 } 783} 784 785void MarkSweep::ScanGrayObjects(bool paused, byte minimum_age) { 786 accounting::CardTable* card_table = GetHeap()->GetCardTable(); 787 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 788 size_t thread_count = GetThreadCount(paused); 789 // The parallel version with only one thread is faster for card scanning, TODO: fix. 790 if (kParallelCardScan && thread_count > 0) { 791 Thread* self = Thread::Current(); 792 // Can't have a different split for each space since multiple spaces can have their cards being 793 // scanned at the same time. 794 timings_.StartSplit(paused ? "(Paused)ScanGrayObjects" : "ScanGrayObjects"); 795 // Try to take some of the mark stack since we can pass this off to the worker tasks. 796 Object** mark_stack_begin = mark_stack_->Begin(); 797 Object** mark_stack_end = mark_stack_->End(); 798 const size_t mark_stack_size = mark_stack_end - mark_stack_begin; 799 // Estimated number of work tasks we will create. 800 const size_t mark_stack_tasks = GetHeap()->GetContinuousSpaces().size() * thread_count; 801 DCHECK_NE(mark_stack_tasks, 0U); 802 const size_t mark_stack_delta = std::min(CardScanTask::kMaxSize / 2, 803 mark_stack_size / mark_stack_tasks + 1); 804 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 805 if (space->GetMarkBitmap() == nullptr) { 806 continue; 807 } 808 byte* card_begin = space->Begin(); 809 byte* card_end = space->End(); 810 // Align up the end address. For example, the image space's end 811 // may not be card-size-aligned. 812 card_end = AlignUp(card_end, accounting::CardTable::kCardSize); 813 DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_begin)); 814 DCHECK(IsAligned<accounting::CardTable::kCardSize>(card_end)); 815 // Calculate how many bytes of heap we will scan, 816 const size_t address_range = card_end - card_begin; 817 // Calculate how much address range each task gets. 818 const size_t card_delta = RoundUp(address_range / thread_count + 1, 819 accounting::CardTable::kCardSize); 820 // Create the worker tasks for this space. 821 while (card_begin != card_end) { 822 // Add a range of cards. 823 size_t addr_remaining = card_end - card_begin; 824 size_t card_increment = std::min(card_delta, addr_remaining); 825 // Take from the back of the mark stack. 826 size_t mark_stack_remaining = mark_stack_end - mark_stack_begin; 827 size_t mark_stack_increment = std::min(mark_stack_delta, mark_stack_remaining); 828 mark_stack_end -= mark_stack_increment; 829 mark_stack_->PopBackCount(static_cast<int32_t>(mark_stack_increment)); 830 DCHECK_EQ(mark_stack_end, mark_stack_->End()); 831 // Add the new task to the thread pool. 832 auto* task = new CardScanTask(thread_pool, this, space->GetMarkBitmap(), card_begin, 833 card_begin + card_increment, minimum_age, 834 mark_stack_increment, mark_stack_end); 835 thread_pool->AddTask(self, task); 836 card_begin += card_increment; 837 } 838 } 839 840 // Note: the card scan below may dirty new cards (and scan them) 841 // as a side effect when a Reference object is encountered and 842 // queued during the marking. See b/11465268. 843 thread_pool->SetMaxActiveWorkers(thread_count - 1); 844 thread_pool->StartWorkers(self); 845 thread_pool->Wait(self, true, true); 846 thread_pool->StopWorkers(self); 847 timings_.EndSplit(); 848 } else { 849 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 850 if (space->GetMarkBitmap() != nullptr) { 851 // Image spaces are handled properly since live == marked for them. 852 switch (space->GetGcRetentionPolicy()) { 853 case space::kGcRetentionPolicyNeverCollect: 854 timings_.StartSplit(paused ? "(Paused)ScanGrayImageSpaceObjects" : 855 "ScanGrayImageSpaceObjects"); 856 break; 857 case space::kGcRetentionPolicyFullCollect: 858 timings_.StartSplit(paused ? "(Paused)ScanGrayZygoteSpaceObjects" : 859 "ScanGrayZygoteSpaceObjects"); 860 break; 861 case space::kGcRetentionPolicyAlwaysCollect: 862 timings_.StartSplit(paused ? "(Paused)ScanGrayAllocSpaceObjects" : 863 "ScanGrayAllocSpaceObjects"); 864 break; 865 } 866 ScanObjectVisitor visitor(this); 867 card_table->Scan(space->GetMarkBitmap(), space->Begin(), space->End(), visitor, minimum_age); 868 timings_.EndSplit(); 869 } 870 } 871 } 872} 873 874class RecursiveMarkTask : public MarkStackTask<false> { 875 public: 876 RecursiveMarkTask(ThreadPool* thread_pool, MarkSweep* mark_sweep, 877 accounting::SpaceBitmap* bitmap, uintptr_t begin, uintptr_t end) 878 : MarkStackTask<false>(thread_pool, mark_sweep, 0, NULL), 879 bitmap_(bitmap), 880 begin_(begin), 881 end_(end) { 882 } 883 884 protected: 885 accounting::SpaceBitmap* const bitmap_; 886 const uintptr_t begin_; 887 const uintptr_t end_; 888 889 virtual void Finalize() { 890 delete this; 891 } 892 893 // Scans all of the objects 894 virtual void Run(Thread* self) NO_THREAD_SAFETY_ANALYSIS { 895 ScanObjectParallelVisitor visitor(this); 896 bitmap_->VisitMarkedRange(begin_, end_, visitor); 897 // Finish by emptying our local mark stack. 898 MarkStackTask::Run(self); 899 } 900}; 901 902// Populates the mark stack based on the set of marked objects and 903// recursively marks until the mark stack is emptied. 904void MarkSweep::RecursiveMark() { 905 TimingLogger::ScopedSplit split("RecursiveMark", &timings_); 906 // RecursiveMark will build the lists of known instances of the Reference classes. See 907 // DelayReferenceReferent for details. 908 if (kUseRecursiveMark) { 909 const bool partial = GetGcType() == kGcTypePartial; 910 ScanObjectVisitor scan_visitor(this); 911 auto* self = Thread::Current(); 912 ThreadPool* thread_pool = heap_->GetThreadPool(); 913 size_t thread_count = GetThreadCount(false); 914 const bool parallel = kParallelRecursiveMark && thread_count > 1; 915 mark_stack_->Reset(); 916 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 917 if ((space->GetGcRetentionPolicy() == space::kGcRetentionPolicyAlwaysCollect) || 918 (!partial && space->GetGcRetentionPolicy() == space::kGcRetentionPolicyFullCollect)) { 919 current_space_bitmap_ = space->GetMarkBitmap(); 920 if (current_space_bitmap_ == nullptr) { 921 continue; 922 } 923 if (parallel) { 924 // We will use the mark stack the future. 925 // CHECK(mark_stack_->IsEmpty()); 926 // This function does not handle heap end increasing, so we must use the space end. 927 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 928 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 929 atomic_finger_ = static_cast<int32_t>(0xFFFFFFFF); 930 931 // Create a few worker tasks. 932 const size_t n = thread_count * 2; 933 while (begin != end) { 934 uintptr_t start = begin; 935 uintptr_t delta = (end - begin) / n; 936 delta = RoundUp(delta, KB); 937 if (delta < 16 * KB) delta = end - begin; 938 begin += delta; 939 auto* task = new RecursiveMarkTask(thread_pool, this, current_space_bitmap_, start, 940 begin); 941 thread_pool->AddTask(self, task); 942 } 943 thread_pool->SetMaxActiveWorkers(thread_count - 1); 944 thread_pool->StartWorkers(self); 945 thread_pool->Wait(self, true, true); 946 thread_pool->StopWorkers(self); 947 } else { 948 // This function does not handle heap end increasing, so we must use the space end. 949 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 950 uintptr_t end = reinterpret_cast<uintptr_t>(space->End()); 951 current_space_bitmap_->VisitMarkedRange(begin, end, scan_visitor); 952 } 953 } 954 } 955 } 956 ProcessMarkStack(false); 957} 958 959mirror::Object* MarkSweep::IsMarkedCallback(mirror::Object* object, void* arg) { 960 if (reinterpret_cast<MarkSweep*>(arg)->IsMarked(object)) { 961 return object; 962 } 963 return nullptr; 964} 965 966void MarkSweep::RecursiveMarkDirtyObjects(bool paused, byte minimum_age) { 967 ScanGrayObjects(paused, minimum_age); 968 ProcessMarkStack(paused); 969} 970 971void MarkSweep::ReMarkRoots() { 972 Locks::mutator_lock_->AssertExclusiveHeld(Thread::Current()); 973 timings_.StartSplit("(Paused)ReMarkRoots"); 974 Runtime::Current()->VisitRoots( 975 MarkRootCallback, this, static_cast<VisitRootFlags>(kVisitRootFlagNewRoots | 976 kVisitRootFlagStopLoggingNewRoots | 977 kVisitRootFlagClearRootLog)); 978 timings_.EndSplit(); 979 if (kVerifyRoots) { 980 timings_.StartSplit("(Paused)VerifyRoots"); 981 Runtime::Current()->VisitRoots(VerifyRootMarked, this); 982 timings_.EndSplit(); 983 } 984} 985 986void MarkSweep::SweepSystemWeaks() { 987 Runtime* runtime = Runtime::Current(); 988 timings_.StartSplit("SweepSystemWeaks"); 989 runtime->SweepSystemWeaks(IsMarkedCallback, this); 990 timings_.EndSplit(); 991} 992 993mirror::Object* MarkSweep::VerifySystemWeakIsLiveCallback(Object* obj, void* arg) { 994 reinterpret_cast<MarkSweep*>(arg)->VerifyIsLive(obj); 995 // We don't actually want to sweep the object, so lets return "marked" 996 return obj; 997} 998 999void MarkSweep::VerifyIsLive(const Object* obj) { 1000 Heap* heap = GetHeap(); 1001 if (!heap->GetLiveBitmap()->Test(obj)) { 1002 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 1003 if (!large_object_space->GetLiveObjects()->Test(obj)) { 1004 if (std::find(heap->allocation_stack_->Begin(), heap->allocation_stack_->End(), obj) == 1005 heap->allocation_stack_->End()) { 1006 // Object not found! 1007 heap->DumpSpaces(); 1008 LOG(FATAL) << "Found dead object " << obj; 1009 } 1010 } 1011 } 1012} 1013 1014void MarkSweep::VerifySystemWeaks() { 1015 // Verify system weaks, uses a special object visitor which returns the input object. 1016 Runtime::Current()->SweepSystemWeaks(VerifySystemWeakIsLiveCallback, this); 1017} 1018 1019class CheckpointMarkThreadRoots : public Closure { 1020 public: 1021 explicit CheckpointMarkThreadRoots(MarkSweep* mark_sweep) : mark_sweep_(mark_sweep) {} 1022 1023 virtual void Run(Thread* thread) NO_THREAD_SAFETY_ANALYSIS { 1024 ATRACE_BEGIN("Marking thread roots"); 1025 // Note: self is not necessarily equal to thread since thread may be suspended. 1026 Thread* self = Thread::Current(); 1027 CHECK(thread == self || thread->IsSuspended() || thread->GetState() == kWaitingPerformingGc) 1028 << thread->GetState() << " thread " << thread << " self " << self; 1029 thread->VisitRoots(MarkSweep::MarkRootParallelCallback, mark_sweep_); 1030 ATRACE_END(); 1031 if (kUseThreadLocalAllocationStack) { 1032 thread->RevokeThreadLocalAllocationStack(); 1033 } 1034 if (kRevokeRosAllocThreadLocalBuffersAtCheckpoint) { 1035 mark_sweep_->GetHeap()->RevokeRosAllocThreadLocalBuffers(thread); 1036 } 1037 mark_sweep_->GetBarrier().Pass(self); 1038 } 1039 1040 private: 1041 MarkSweep* mark_sweep_; 1042}; 1043 1044void MarkSweep::MarkRootsCheckpoint(Thread* self) { 1045 CheckpointMarkThreadRoots check_point(this); 1046 timings_.StartSplit("MarkRootsCheckpoint"); 1047 ThreadList* thread_list = Runtime::Current()->GetThreadList(); 1048 // Request the check point is run on all threads returning a count of the threads that must 1049 // run through the barrier including self. 1050 size_t barrier_count = thread_list->RunCheckpoint(&check_point); 1051 // Release locks then wait for all mutator threads to pass the barrier. 1052 // TODO: optimize to not release locks when there are no threads to wait for. 1053 Locks::heap_bitmap_lock_->ExclusiveUnlock(self); 1054 Locks::mutator_lock_->SharedUnlock(self); 1055 ThreadState old_state = self->SetState(kWaitingForCheckPointsToRun); 1056 CHECK_EQ(old_state, kWaitingPerformingGc); 1057 gc_barrier_->Increment(self, barrier_count); 1058 self->SetState(kWaitingPerformingGc); 1059 Locks::mutator_lock_->SharedLock(self); 1060 Locks::heap_bitmap_lock_->ExclusiveLock(self); 1061 timings_.EndSplit(); 1062} 1063 1064void MarkSweep::SweepArray(accounting::ObjectStack* allocations, bool swap_bitmaps) { 1065 timings_.StartSplit("SweepArray"); 1066 Thread* self = Thread::Current(); 1067 mirror::Object* chunk_free_buffer[kSweepArrayChunkFreeSize]; 1068 size_t chunk_free_pos = 0; 1069 size_t freed_bytes = 0; 1070 size_t freed_large_object_bytes = 0; 1071 size_t freed_objects = 0; 1072 size_t freed_large_objects = 0; 1073 // How many objects are left in the array, modified after each space is swept. 1074 Object** objects = const_cast<Object**>(allocations->Begin()); 1075 size_t count = allocations->Size(); 1076 // Change the order to ensure that the non-moving space last swept as an optimization. 1077 std::vector<space::ContinuousSpace*> sweep_spaces; 1078 space::ContinuousSpace* non_moving_space = nullptr; 1079 for (space::ContinuousSpace* space : heap_->GetContinuousSpaces()) { 1080 if (space->IsAllocSpace() && !immune_region_.ContainsSpace(space) && 1081 space->GetLiveBitmap() != nullptr) { 1082 if (space == heap_->GetNonMovingSpace()) { 1083 non_moving_space = space; 1084 } else { 1085 sweep_spaces.push_back(space); 1086 } 1087 } 1088 } 1089 // Unlikely to sweep a significant amount of non_movable objects, so we do these after the after 1090 // the other alloc spaces as an optimization. 1091 if (non_moving_space != nullptr) { 1092 sweep_spaces.push_back(non_moving_space); 1093 } 1094 // Start by sweeping the continuous spaces. 1095 for (space::ContinuousSpace* space : sweep_spaces) { 1096 space::AllocSpace* alloc_space = space->AsAllocSpace(); 1097 accounting::SpaceBitmap* live_bitmap = space->GetLiveBitmap(); 1098 accounting::SpaceBitmap* mark_bitmap = space->GetMarkBitmap(); 1099 if (swap_bitmaps) { 1100 std::swap(live_bitmap, mark_bitmap); 1101 } 1102 Object** out = objects; 1103 for (size_t i = 0; i < count; ++i) { 1104 Object* obj = objects[i]; 1105 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1106 continue; 1107 } 1108 if (space->HasAddress(obj)) { 1109 // This object is in the space, remove it from the array and add it to the sweep buffer 1110 // if needed. 1111 if (!mark_bitmap->Test(obj)) { 1112 if (chunk_free_pos >= kSweepArrayChunkFreeSize) { 1113 timings_.StartSplit("FreeList"); 1114 freed_objects += chunk_free_pos; 1115 freed_bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1116 timings_.EndSplit(); 1117 chunk_free_pos = 0; 1118 } 1119 chunk_free_buffer[chunk_free_pos++] = obj; 1120 } 1121 } else { 1122 *(out++) = obj; 1123 } 1124 } 1125 if (chunk_free_pos > 0) { 1126 timings_.StartSplit("FreeList"); 1127 freed_objects += chunk_free_pos; 1128 freed_bytes += alloc_space->FreeList(self, chunk_free_pos, chunk_free_buffer); 1129 timings_.EndSplit(); 1130 chunk_free_pos = 0; 1131 } 1132 // All of the references which space contained are no longer in the allocation stack, update 1133 // the count. 1134 count = out - objects; 1135 } 1136 // Handle the large object space. 1137 space::LargeObjectSpace* large_object_space = GetHeap()->GetLargeObjectsSpace(); 1138 accounting::ObjectSet* large_live_objects = large_object_space->GetLiveObjects(); 1139 accounting::ObjectSet* large_mark_objects = large_object_space->GetMarkObjects(); 1140 if (swap_bitmaps) { 1141 std::swap(large_live_objects, large_mark_objects); 1142 } 1143 for (size_t i = 0; i < count; ++i) { 1144 Object* obj = objects[i]; 1145 // Handle large objects. 1146 if (kUseThreadLocalAllocationStack && obj == nullptr) { 1147 continue; 1148 } 1149 if (!large_mark_objects->Test(obj)) { 1150 ++freed_large_objects; 1151 freed_large_object_bytes += large_object_space->Free(self, obj); 1152 } 1153 } 1154 timings_.EndSplit(); 1155 1156 timings_.StartSplit("RecordFree"); 1157 VLOG(heap) << "Freed " << freed_objects << "/" << count 1158 << " objects with size " << PrettySize(freed_bytes); 1159 heap_->RecordFree(freed_objects + freed_large_objects, freed_bytes + freed_large_object_bytes); 1160 freed_objects_.FetchAndAdd(freed_objects); 1161 freed_large_objects_.FetchAndAdd(freed_large_objects); 1162 freed_bytes_.FetchAndAdd(freed_bytes); 1163 freed_large_object_bytes_.FetchAndAdd(freed_large_object_bytes); 1164 timings_.EndSplit(); 1165 1166 timings_.StartSplit("ResetStack"); 1167 allocations->Reset(); 1168 timings_.EndSplit(); 1169} 1170 1171void MarkSweep::Sweep(bool swap_bitmaps) { 1172 DCHECK(mark_stack_->IsEmpty()); 1173 TimingLogger::ScopedSplit("Sweep", &timings_); 1174 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 1175 if (space->IsContinuousMemMapAllocSpace()) { 1176 space::ContinuousMemMapAllocSpace* alloc_space = space->AsContinuousMemMapAllocSpace(); 1177 TimingLogger::ScopedSplit split( 1178 alloc_space->IsZygoteSpace() ? "SweepZygoteSpace" : "SweepMallocSpace", &timings_); 1179 size_t freed_objects = 0; 1180 size_t freed_bytes = 0; 1181 alloc_space->Sweep(swap_bitmaps, &freed_objects, &freed_bytes); 1182 heap_->RecordFree(freed_objects, freed_bytes); 1183 freed_objects_.FetchAndAdd(freed_objects); 1184 freed_bytes_.FetchAndAdd(freed_bytes); 1185 } 1186 } 1187 SweepLargeObjects(swap_bitmaps); 1188} 1189 1190void MarkSweep::SweepLargeObjects(bool swap_bitmaps) { 1191 TimingLogger::ScopedSplit("SweepLargeObjects", &timings_); 1192 size_t freed_objects = 0; 1193 size_t freed_bytes = 0; 1194 GetHeap()->GetLargeObjectsSpace()->Sweep(swap_bitmaps, &freed_objects, &freed_bytes); 1195 freed_large_objects_.FetchAndAdd(freed_objects); 1196 freed_large_object_bytes_.FetchAndAdd(freed_bytes); 1197 GetHeap()->RecordFree(freed_objects, freed_bytes); 1198} 1199 1200// Process the "referent" field in a java.lang.ref.Reference. If the 1201// referent has not yet been marked, put it on the appropriate list in 1202// the heap for later processing. 1203void MarkSweep::DelayReferenceReferent(mirror::Class* klass, Object* obj) { 1204 DCHECK(klass != nullptr); 1205 if (kCountJavaLangRefs) { 1206 ++reference_count_; 1207 } 1208 heap_->DelayReferenceReferent(klass, obj->AsReference(), IsMarkedCallback, this); 1209} 1210 1211class MarkObjectVisitor { 1212 public: 1213 explicit MarkObjectVisitor(MarkSweep* const mark_sweep) ALWAYS_INLINE : mark_sweep_(mark_sweep) {} 1214 1215 // TODO: Fixme when anotatalysis works with visitors. 1216 void operator()(Object* /* obj */, Object* ref, const MemberOffset& /* offset */, 1217 bool /* is_static */) const ALWAYS_INLINE 1218 NO_THREAD_SAFETY_ANALYSIS { 1219 if (kCheckLocks) { 1220 Locks::mutator_lock_->AssertSharedHeld(Thread::Current()); 1221 Locks::heap_bitmap_lock_->AssertExclusiveHeld(Thread::Current()); 1222 } 1223 mark_sweep_->MarkObject(ref); 1224 } 1225 1226 private: 1227 MarkSweep* const mark_sweep_; 1228}; 1229 1230// Scans an object reference. Determines the type of the reference 1231// and dispatches to a specialized scanning routine. 1232void MarkSweep::ScanObject(Object* obj) { 1233 MarkObjectVisitor visitor(this); 1234 ScanObjectVisit(obj, visitor); 1235} 1236 1237void MarkSweep::ProcessMarkStackPausedCallback(void* arg) { 1238 reinterpret_cast<MarkSweep*>(arg)->ProcessMarkStack(true); 1239} 1240 1241void MarkSweep::ProcessMarkStackParallel(size_t thread_count) { 1242 Thread* self = Thread::Current(); 1243 ThreadPool* thread_pool = GetHeap()->GetThreadPool(); 1244 const size_t chunk_size = std::min(mark_stack_->Size() / thread_count + 1, 1245 static_cast<size_t>(MarkStackTask<false>::kMaxSize)); 1246 CHECK_GT(chunk_size, 0U); 1247 // Split the current mark stack up into work tasks. 1248 for (mirror::Object **it = mark_stack_->Begin(), **end = mark_stack_->End(); it < end; ) { 1249 const size_t delta = std::min(static_cast<size_t>(end - it), chunk_size); 1250 thread_pool->AddTask(self, new MarkStackTask<false>(thread_pool, this, delta, it)); 1251 it += delta; 1252 } 1253 thread_pool->SetMaxActiveWorkers(thread_count - 1); 1254 thread_pool->StartWorkers(self); 1255 thread_pool->Wait(self, true, true); 1256 thread_pool->StopWorkers(self); 1257 mark_stack_->Reset(); 1258 CHECK_EQ(work_chunks_created_, work_chunks_deleted_) << " some of the work chunks were leaked"; 1259} 1260 1261// Scan anything that's on the mark stack. 1262void MarkSweep::ProcessMarkStack(bool paused) { 1263 timings_.StartSplit(paused ? "(Paused)ProcessMarkStack" : "ProcessMarkStack"); 1264 size_t thread_count = GetThreadCount(paused); 1265 if (kParallelProcessMarkStack && thread_count > 1 && 1266 mark_stack_->Size() >= kMinimumParallelMarkStackSize) { 1267 ProcessMarkStackParallel(thread_count); 1268 } else { 1269 // TODO: Tune this. 1270 static const size_t kFifoSize = 4; 1271 BoundedFifoPowerOfTwo<Object*, kFifoSize> prefetch_fifo; 1272 for (;;) { 1273 Object* obj = NULL; 1274 if (kUseMarkStackPrefetch) { 1275 while (!mark_stack_->IsEmpty() && prefetch_fifo.size() < kFifoSize) { 1276 Object* obj = mark_stack_->PopBack(); 1277 DCHECK(obj != NULL); 1278 __builtin_prefetch(obj); 1279 prefetch_fifo.push_back(obj); 1280 } 1281 if (prefetch_fifo.empty()) { 1282 break; 1283 } 1284 obj = prefetch_fifo.front(); 1285 prefetch_fifo.pop_front(); 1286 } else { 1287 if (mark_stack_->IsEmpty()) { 1288 break; 1289 } 1290 obj = mark_stack_->PopBack(); 1291 } 1292 DCHECK(obj != NULL); 1293 ScanObject(obj); 1294 } 1295 } 1296 timings_.EndSplit(); 1297} 1298 1299inline bool MarkSweep::IsMarked(const Object* object) const 1300 SHARED_LOCKS_REQUIRED(Locks::heap_bitmap_lock_) { 1301 if (immune_region_.ContainsObject(object)) { 1302 return true; 1303 } 1304 if (current_space_bitmap_->HasAddress(object)) { 1305 return current_space_bitmap_->Test(object); 1306 } 1307 return mark_bitmap_->Test(object); 1308} 1309 1310void MarkSweep::FinishPhase() { 1311 TimingLogger::ScopedSplit split("FinishPhase", &timings_); 1312 // Can't enqueue references if we hold the mutator lock. 1313 Heap* heap = GetHeap(); 1314 timings_.NewSplit("PostGcVerification"); 1315 heap->PostGcVerification(this); 1316 // Update the cumulative statistics. 1317 total_freed_objects_ += GetFreedObjects() + GetFreedLargeObjects(); 1318 total_freed_bytes_ += GetFreedBytes() + GetFreedLargeObjectBytes(); 1319 // Ensure that the mark stack is empty. 1320 CHECK(mark_stack_->IsEmpty()); 1321 if (kCountScannedTypes) { 1322 VLOG(gc) << "MarkSweep scanned classes=" << class_count_ << " arrays=" << array_count_ 1323 << " other=" << other_count_; 1324 } 1325 if (kCountTasks) { 1326 VLOG(gc) << "Total number of work chunks allocated: " << work_chunks_created_; 1327 } 1328 if (kMeasureOverhead) { 1329 VLOG(gc) << "Overhead time " << PrettyDuration(overhead_time_); 1330 } 1331 if (kProfileLargeObjects) { 1332 VLOG(gc) << "Large objects tested " << large_object_test_ << " marked " << large_object_mark_; 1333 } 1334 if (kCountJavaLangRefs) { 1335 VLOG(gc) << "References scanned " << reference_count_; 1336 } 1337 if (kCountMarkedObjects) { 1338 VLOG(gc) << "Marked: null=" << mark_null_count_ << " immune=" << mark_immune_count_ 1339 << " fastpath=" << mark_fastpath_count_ << " slowpath=" << mark_slowpath_count_; 1340 } 1341 // Update the cumulative loggers. 1342 cumulative_timings_.Start(); 1343 cumulative_timings_.AddLogger(timings_); 1344 cumulative_timings_.End(); 1345 // Clear all of the spaces' mark bitmaps. 1346 for (const auto& space : GetHeap()->GetContinuousSpaces()) { 1347 accounting::SpaceBitmap* bitmap = space->GetMarkBitmap(); 1348 if (bitmap != nullptr && 1349 space->GetGcRetentionPolicy() != space::kGcRetentionPolicyNeverCollect) { 1350 bitmap->Clear(); 1351 } 1352 } 1353 mark_stack_->Reset(); 1354 // Reset the marked large objects. 1355 space::LargeObjectSpace* large_objects = GetHeap()->GetLargeObjectsSpace(); 1356 large_objects->GetMarkObjects()->Clear(); 1357} 1358 1359void MarkSweep::RevokeAllThreadLocalBuffers() { 1360 if (kRevokeRosAllocThreadLocalBuffersAtCheckpoint && IsConcurrent()) { 1361 // If concurrent, rosalloc thread-local buffers are revoked at the 1362 // thread checkpoint. Bump pointer space thread-local buffers must 1363 // not be in use. 1364 GetHeap()->AssertAllBumpPointerSpaceThreadLocalBuffersAreRevoked(); 1365 } else { 1366 timings_.StartSplit("(Paused)RevokeAllThreadLocalBuffers"); 1367 GetHeap()->RevokeAllThreadLocalBuffers(); 1368 timings_.EndSplit(); 1369 } 1370} 1371 1372} // namespace collector 1373} // namespace gc 1374} // namespace art 1375