1 /*
2 * Copyright (C) 2013 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 #ifndef ART_RUNTIME_GC_HEAP_INL_H_
18 #define ART_RUNTIME_GC_HEAP_INL_H_
19
20 #include "heap.h"
21
22 #include "base/time_utils.h"
23 #include "debugger.h"
24 #include "gc/accounting/card_table-inl.h"
25 #include "gc/collector/semi_space.h"
26 #include "gc/space/bump_pointer_space-inl.h"
27 #include "gc/space/dlmalloc_space-inl.h"
28 #include "gc/space/large_object_space.h"
29 #include "gc/space/region_space-inl.h"
30 #include "gc/space/rosalloc_space-inl.h"
31 #include "runtime.h"
32 #include "handle_scope-inl.h"
33 #include "thread-inl.h"
34 #include "utils.h"
35 #include "verify_object-inl.h"
36
37 namespace art {
38 namespace gc {
39
40 template <bool kInstrumented, bool kCheckLargeObject, typename PreFenceVisitor>
AllocObjectWithAllocator(Thread * self,mirror::Class * klass,size_t byte_count,AllocatorType allocator,const PreFenceVisitor & pre_fence_visitor)41 inline mirror::Object* Heap::AllocObjectWithAllocator(Thread* self, mirror::Class* klass,
42 size_t byte_count, AllocatorType allocator,
43 const PreFenceVisitor& pre_fence_visitor) {
44 if (kIsDebugBuild) {
45 CheckPreconditionsForAllocObject(klass, byte_count);
46 // Since allocation can cause a GC which will need to SuspendAll, make sure all allocations are
47 // done in the runnable state where suspension is expected.
48 CHECK_EQ(self->GetState(), kRunnable);
49 self->AssertThreadSuspensionIsAllowable();
50 }
51 // Need to check that we arent the large object allocator since the large object allocation code
52 // path this function. If we didn't check we would have an infinite loop.
53 mirror::Object* obj;
54 if (kCheckLargeObject && UNLIKELY(ShouldAllocLargeObject(klass, byte_count))) {
55 obj = AllocLargeObject<kInstrumented, PreFenceVisitor>(self, &klass, byte_count,
56 pre_fence_visitor);
57 if (obj != nullptr) {
58 return obj;
59 } else {
60 // There should be an OOM exception, since we are retrying, clear it.
61 self->ClearException();
62 }
63 // If the large object allocation failed, try to use the normal spaces (main space,
64 // non moving space). This can happen if there is significant virtual address space
65 // fragmentation.
66 }
67 AllocationTimer alloc_timer(this, &obj);
68 // bytes allocated for the (individual) object.
69 size_t bytes_allocated;
70 size_t usable_size;
71 size_t new_num_bytes_allocated = 0;
72 if (allocator == kAllocatorTypeTLAB || allocator == kAllocatorTypeRegionTLAB) {
73 byte_count = RoundUp(byte_count, space::BumpPointerSpace::kAlignment);
74 }
75 // If we have a thread local allocation we don't need to update bytes allocated.
76 if ((allocator == kAllocatorTypeTLAB || allocator == kAllocatorTypeRegionTLAB) &&
77 byte_count <= self->TlabSize()) {
78 obj = self->AllocTlab(byte_count);
79 DCHECK(obj != nullptr) << "AllocTlab can't fail";
80 obj->SetClass(klass);
81 if (kUseBakerOrBrooksReadBarrier) {
82 if (kUseBrooksReadBarrier) {
83 obj->SetReadBarrierPointer(obj);
84 }
85 obj->AssertReadBarrierPointer();
86 }
87 bytes_allocated = byte_count;
88 usable_size = bytes_allocated;
89 pre_fence_visitor(obj, usable_size);
90 QuasiAtomic::ThreadFenceForConstructor();
91 } else if (!kInstrumented && allocator == kAllocatorTypeRosAlloc &&
92 (obj = rosalloc_space_->AllocThreadLocal(self, byte_count, &bytes_allocated)) &&
93 LIKELY(obj != nullptr)) {
94 DCHECK(!running_on_valgrind_);
95 obj->SetClass(klass);
96 if (kUseBakerOrBrooksReadBarrier) {
97 if (kUseBrooksReadBarrier) {
98 obj->SetReadBarrierPointer(obj);
99 }
100 obj->AssertReadBarrierPointer();
101 }
102 usable_size = bytes_allocated;
103 pre_fence_visitor(obj, usable_size);
104 QuasiAtomic::ThreadFenceForConstructor();
105 } else {
106 // bytes allocated that takes bulk thread-local buffer allocations into account.
107 size_t bytes_tl_bulk_allocated = 0;
108 obj = TryToAllocate<kInstrumented, false>(self, allocator, byte_count, &bytes_allocated,
109 &usable_size, &bytes_tl_bulk_allocated);
110 if (UNLIKELY(obj == nullptr)) {
111 bool is_current_allocator = allocator == GetCurrentAllocator();
112 obj = AllocateInternalWithGc(self, allocator, byte_count, &bytes_allocated, &usable_size,
113 &bytes_tl_bulk_allocated, &klass);
114 if (obj == nullptr) {
115 bool after_is_current_allocator = allocator == GetCurrentAllocator();
116 // If there is a pending exception, fail the allocation right away since the next one
117 // could cause OOM and abort the runtime.
118 if (!self->IsExceptionPending() && is_current_allocator && !after_is_current_allocator) {
119 // If the allocator changed, we need to restart the allocation.
120 return AllocObject<kInstrumented>(self, klass, byte_count, pre_fence_visitor);
121 }
122 return nullptr;
123 }
124 }
125 DCHECK_GT(bytes_allocated, 0u);
126 DCHECK_GT(usable_size, 0u);
127 obj->SetClass(klass);
128 if (kUseBakerOrBrooksReadBarrier) {
129 if (kUseBrooksReadBarrier) {
130 obj->SetReadBarrierPointer(obj);
131 }
132 obj->AssertReadBarrierPointer();
133 }
134 if (collector::SemiSpace::kUseRememberedSet && UNLIKELY(allocator == kAllocatorTypeNonMoving)) {
135 // (Note this if statement will be constant folded away for the
136 // fast-path quick entry points.) Because SetClass() has no write
137 // barrier, if a non-moving space allocation, we need a write
138 // barrier as the class pointer may point to the bump pointer
139 // space (where the class pointer is an "old-to-young" reference,
140 // though rare) under the GSS collector with the remembered set
141 // enabled. We don't need this for kAllocatorTypeRosAlloc/DlMalloc
142 // cases because we don't directly allocate into the main alloc
143 // space (besides promotions) under the SS/GSS collector.
144 WriteBarrierField(obj, mirror::Object::ClassOffset(), klass);
145 }
146 pre_fence_visitor(obj, usable_size);
147 new_num_bytes_allocated = static_cast<size_t>(
148 num_bytes_allocated_.FetchAndAddSequentiallyConsistent(bytes_tl_bulk_allocated))
149 + bytes_tl_bulk_allocated;
150 }
151 if (kIsDebugBuild && Runtime::Current()->IsStarted()) {
152 CHECK_LE(obj->SizeOf(), usable_size);
153 }
154 // TODO: Deprecate.
155 if (kInstrumented) {
156 if (Runtime::Current()->HasStatsEnabled()) {
157 RuntimeStats* thread_stats = self->GetStats();
158 ++thread_stats->allocated_objects;
159 thread_stats->allocated_bytes += bytes_allocated;
160 RuntimeStats* global_stats = Runtime::Current()->GetStats();
161 ++global_stats->allocated_objects;
162 global_stats->allocated_bytes += bytes_allocated;
163 }
164 } else {
165 DCHECK(!Runtime::Current()->HasStatsEnabled());
166 }
167 if (AllocatorHasAllocationStack(allocator)) {
168 PushOnAllocationStack(self, &obj);
169 }
170 if (kInstrumented) {
171 if (Dbg::IsAllocTrackingEnabled()) {
172 Dbg::RecordAllocation(self, klass, bytes_allocated);
173 }
174 } else {
175 DCHECK(!Dbg::IsAllocTrackingEnabled());
176 }
177 if (kInstrumented) {
178 if (gc_stress_mode_) {
179 CheckGcStressMode(self, &obj);
180 }
181 } else {
182 DCHECK(!gc_stress_mode_);
183 }
184 // IsConcurrentGc() isn't known at compile time so we can optimize by not checking it for
185 // the BumpPointer or TLAB allocators. This is nice since it allows the entire if statement to be
186 // optimized out. And for the other allocators, AllocatorMayHaveConcurrentGC is a constant since
187 // the allocator_type should be constant propagated.
188 if (AllocatorMayHaveConcurrentGC(allocator) && IsGcConcurrent()) {
189 CheckConcurrentGC(self, new_num_bytes_allocated, &obj);
190 }
191 VerifyObject(obj);
192 self->VerifyStack();
193 return obj;
194 }
195
196 // The size of a thread-local allocation stack in the number of references.
197 static constexpr size_t kThreadLocalAllocationStackSize = 128;
198
PushOnAllocationStack(Thread * self,mirror::Object ** obj)199 inline void Heap::PushOnAllocationStack(Thread* self, mirror::Object** obj) {
200 if (kUseThreadLocalAllocationStack) {
201 if (UNLIKELY(!self->PushOnThreadLocalAllocationStack(*obj))) {
202 PushOnThreadLocalAllocationStackWithInternalGC(self, obj);
203 }
204 } else if (UNLIKELY(!allocation_stack_->AtomicPushBack(*obj))) {
205 PushOnAllocationStackWithInternalGC(self, obj);
206 }
207 }
208
209 template <bool kInstrumented, typename PreFenceVisitor>
AllocLargeObject(Thread * self,mirror::Class ** klass,size_t byte_count,const PreFenceVisitor & pre_fence_visitor)210 inline mirror::Object* Heap::AllocLargeObject(Thread* self, mirror::Class** klass,
211 size_t byte_count,
212 const PreFenceVisitor& pre_fence_visitor) {
213 // Save and restore the class in case it moves.
214 StackHandleScope<1> hs(self);
215 auto klass_wrapper = hs.NewHandleWrapper(klass);
216 return AllocObjectWithAllocator<kInstrumented, false, PreFenceVisitor>(self, *klass, byte_count,
217 kAllocatorTypeLOS,
218 pre_fence_visitor);
219 }
220
221 template <const bool kInstrumented, const bool kGrow>
TryToAllocate(Thread * self,AllocatorType allocator_type,size_t alloc_size,size_t * bytes_allocated,size_t * usable_size,size_t * bytes_tl_bulk_allocated)222 inline mirror::Object* Heap::TryToAllocate(Thread* self, AllocatorType allocator_type,
223 size_t alloc_size, size_t* bytes_allocated,
224 size_t* usable_size,
225 size_t* bytes_tl_bulk_allocated) {
226 if (allocator_type != kAllocatorTypeTLAB && allocator_type != kAllocatorTypeRegionTLAB &&
227 allocator_type != kAllocatorTypeRosAlloc &&
228 UNLIKELY(IsOutOfMemoryOnAllocation<kGrow>(allocator_type, alloc_size))) {
229 return nullptr;
230 }
231 mirror::Object* ret;
232 switch (allocator_type) {
233 case kAllocatorTypeBumpPointer: {
234 DCHECK(bump_pointer_space_ != nullptr);
235 alloc_size = RoundUp(alloc_size, space::BumpPointerSpace::kAlignment);
236 ret = bump_pointer_space_->AllocNonvirtual(alloc_size);
237 if (LIKELY(ret != nullptr)) {
238 *bytes_allocated = alloc_size;
239 *usable_size = alloc_size;
240 *bytes_tl_bulk_allocated = alloc_size;
241 }
242 break;
243 }
244 case kAllocatorTypeRosAlloc: {
245 if (kInstrumented && UNLIKELY(running_on_valgrind_)) {
246 // If running on valgrind, we should be using the instrumented path.
247 size_t max_bytes_tl_bulk_allocated = rosalloc_space_->MaxBytesBulkAllocatedFor(alloc_size);
248 if (UNLIKELY(IsOutOfMemoryOnAllocation<kGrow>(allocator_type,
249 max_bytes_tl_bulk_allocated))) {
250 return nullptr;
251 }
252 ret = rosalloc_space_->Alloc(self, alloc_size, bytes_allocated, usable_size,
253 bytes_tl_bulk_allocated);
254 } else {
255 DCHECK(!running_on_valgrind_);
256 size_t max_bytes_tl_bulk_allocated =
257 rosalloc_space_->MaxBytesBulkAllocatedForNonvirtual(alloc_size);
258 if (UNLIKELY(IsOutOfMemoryOnAllocation<kGrow>(allocator_type,
259 max_bytes_tl_bulk_allocated))) {
260 return nullptr;
261 }
262 if (!kInstrumented) {
263 DCHECK(!rosalloc_space_->CanAllocThreadLocal(self, alloc_size));
264 }
265 ret = rosalloc_space_->AllocNonvirtual(self, alloc_size, bytes_allocated, usable_size,
266 bytes_tl_bulk_allocated);
267 }
268 break;
269 }
270 case kAllocatorTypeDlMalloc: {
271 if (kInstrumented && UNLIKELY(running_on_valgrind_)) {
272 // If running on valgrind, we should be using the instrumented path.
273 ret = dlmalloc_space_->Alloc(self, alloc_size, bytes_allocated, usable_size,
274 bytes_tl_bulk_allocated);
275 } else {
276 DCHECK(!running_on_valgrind_);
277 ret = dlmalloc_space_->AllocNonvirtual(self, alloc_size, bytes_allocated, usable_size,
278 bytes_tl_bulk_allocated);
279 }
280 break;
281 }
282 case kAllocatorTypeNonMoving: {
283 ret = non_moving_space_->Alloc(self, alloc_size, bytes_allocated, usable_size,
284 bytes_tl_bulk_allocated);
285 break;
286 }
287 case kAllocatorTypeLOS: {
288 ret = large_object_space_->Alloc(self, alloc_size, bytes_allocated, usable_size,
289 bytes_tl_bulk_allocated);
290 // Note that the bump pointer spaces aren't necessarily next to
291 // the other continuous spaces like the non-moving alloc space or
292 // the zygote space.
293 DCHECK(ret == nullptr || large_object_space_->Contains(ret));
294 break;
295 }
296 case kAllocatorTypeTLAB: {
297 DCHECK_ALIGNED(alloc_size, space::BumpPointerSpace::kAlignment);
298 if (UNLIKELY(self->TlabSize() < alloc_size)) {
299 const size_t new_tlab_size = alloc_size + kDefaultTLABSize;
300 if (UNLIKELY(IsOutOfMemoryOnAllocation<kGrow>(allocator_type, new_tlab_size))) {
301 return nullptr;
302 }
303 // Try allocating a new thread local buffer, if the allocaiton fails the space must be
304 // full so return null.
305 if (!bump_pointer_space_->AllocNewTlab(self, new_tlab_size)) {
306 return nullptr;
307 }
308 *bytes_tl_bulk_allocated = new_tlab_size;
309 } else {
310 *bytes_tl_bulk_allocated = 0;
311 }
312 // The allocation can't fail.
313 ret = self->AllocTlab(alloc_size);
314 DCHECK(ret != nullptr);
315 *bytes_allocated = alloc_size;
316 *usable_size = alloc_size;
317 break;
318 }
319 case kAllocatorTypeRegion: {
320 DCHECK(region_space_ != nullptr);
321 alloc_size = RoundUp(alloc_size, space::RegionSpace::kAlignment);
322 ret = region_space_->AllocNonvirtual<false>(alloc_size, bytes_allocated, usable_size,
323 bytes_tl_bulk_allocated);
324 break;
325 }
326 case kAllocatorTypeRegionTLAB: {
327 DCHECK(region_space_ != nullptr);
328 DCHECK_ALIGNED(alloc_size, space::RegionSpace::kAlignment);
329 if (UNLIKELY(self->TlabSize() < alloc_size)) {
330 if (space::RegionSpace::kRegionSize >= alloc_size) {
331 // Non-large. Check OOME for a tlab.
332 if (LIKELY(!IsOutOfMemoryOnAllocation<kGrow>(allocator_type, space::RegionSpace::kRegionSize))) {
333 // Try to allocate a tlab.
334 if (!region_space_->AllocNewTlab(self)) {
335 // Failed to allocate a tlab. Try non-tlab.
336 ret = region_space_->AllocNonvirtual<false>(alloc_size, bytes_allocated, usable_size,
337 bytes_tl_bulk_allocated);
338 return ret;
339 }
340 *bytes_tl_bulk_allocated = space::RegionSpace::kRegionSize;
341 // Fall-through.
342 } else {
343 // Check OOME for a non-tlab allocation.
344 if (!IsOutOfMemoryOnAllocation<kGrow>(allocator_type, alloc_size)) {
345 ret = region_space_->AllocNonvirtual<false>(alloc_size, bytes_allocated, usable_size,
346 bytes_tl_bulk_allocated);
347 return ret;
348 } else {
349 // Neither tlab or non-tlab works. Give up.
350 return nullptr;
351 }
352 }
353 } else {
354 // Large. Check OOME.
355 if (LIKELY(!IsOutOfMemoryOnAllocation<kGrow>(allocator_type, alloc_size))) {
356 ret = region_space_->AllocNonvirtual<false>(alloc_size, bytes_allocated, usable_size,
357 bytes_tl_bulk_allocated);
358 return ret;
359 } else {
360 return nullptr;
361 }
362 }
363 } else {
364 *bytes_tl_bulk_allocated = 0; // Allocated in an existing buffer.
365 }
366 // The allocation can't fail.
367 ret = self->AllocTlab(alloc_size);
368 DCHECK(ret != nullptr);
369 *bytes_allocated = alloc_size;
370 *usable_size = alloc_size;
371 break;
372 }
373 default: {
374 LOG(FATAL) << "Invalid allocator type";
375 ret = nullptr;
376 }
377 }
378 return ret;
379 }
380
AllocationTimer(Heap * heap,mirror::Object ** allocated_obj_ptr)381 inline Heap::AllocationTimer::AllocationTimer(Heap* heap, mirror::Object** allocated_obj_ptr)
382 : heap_(heap), allocated_obj_ptr_(allocated_obj_ptr),
383 allocation_start_time_(kMeasureAllocationTime ? NanoTime() / kTimeAdjust : 0u) { }
384
~AllocationTimer()385 inline Heap::AllocationTimer::~AllocationTimer() {
386 if (kMeasureAllocationTime) {
387 mirror::Object* allocated_obj = *allocated_obj_ptr_;
388 // Only if the allocation succeeded, record the time.
389 if (allocated_obj != nullptr) {
390 uint64_t allocation_end_time = NanoTime() / kTimeAdjust;
391 heap_->total_allocation_time_.FetchAndAddSequentiallyConsistent(allocation_end_time - allocation_start_time_);
392 }
393 }
394 }
395
ShouldAllocLargeObject(mirror::Class * c,size_t byte_count)396 inline bool Heap::ShouldAllocLargeObject(mirror::Class* c, size_t byte_count) const {
397 // We need to have a zygote space or else our newly allocated large object can end up in the
398 // Zygote resulting in it being prematurely freed.
399 // We can only do this for primitive objects since large objects will not be within the card table
400 // range. This also means that we rely on SetClass not dirtying the object's card.
401 return byte_count >= large_object_threshold_ && (c->IsPrimitiveArray() || c->IsStringClass());
402 }
403
404 template <bool kGrow>
IsOutOfMemoryOnAllocation(AllocatorType allocator_type,size_t alloc_size)405 inline bool Heap::IsOutOfMemoryOnAllocation(AllocatorType allocator_type, size_t alloc_size) {
406 size_t new_footprint = num_bytes_allocated_.LoadSequentiallyConsistent() + alloc_size;
407 if (UNLIKELY(new_footprint > max_allowed_footprint_)) {
408 if (UNLIKELY(new_footprint > growth_limit_)) {
409 return true;
410 }
411 if (!AllocatorMayHaveConcurrentGC(allocator_type) || !IsGcConcurrent()) {
412 if (!kGrow) {
413 return true;
414 }
415 // TODO: Grow for allocation is racy, fix it.
416 VLOG(heap) << "Growing heap from " << PrettySize(max_allowed_footprint_) << " to "
417 << PrettySize(new_footprint) << " for a " << PrettySize(alloc_size) << " allocation";
418 max_allowed_footprint_ = new_footprint;
419 }
420 }
421 return false;
422 }
423
CheckConcurrentGC(Thread * self,size_t new_num_bytes_allocated,mirror::Object ** obj)424 inline void Heap::CheckConcurrentGC(Thread* self, size_t new_num_bytes_allocated,
425 mirror::Object** obj) {
426 if (UNLIKELY(new_num_bytes_allocated >= concurrent_start_bytes_)) {
427 RequestConcurrentGCAndSaveObject(self, false, obj);
428 }
429 }
430
431 } // namespace gc
432 } // namespace art
433
434 #endif // ART_RUNTIME_GC_HEAP_INL_H_
435