1 /*
2 * Copyright 2014 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "jit_code_cache.h"
18
19 #include <sstream>
20
21 #include "arch/context.h"
22 #include "art_method-inl.h"
23 #include "base/enums.h"
24 #include "base/stl_util.h"
25 #include "base/systrace.h"
26 #include "base/time_utils.h"
27 #include "cha.h"
28 #include "debugger_interface.h"
29 #include "entrypoints/runtime_asm_entrypoints.h"
30 #include "gc/accounting/bitmap-inl.h"
31 #include "gc/scoped_gc_critical_section.h"
32 #include "intern_table.h"
33 #include "jit/jit.h"
34 #include "jit/profiling_info.h"
35 #include "linear_alloc.h"
36 #include "mem_map.h"
37 #include "oat_file-inl.h"
38 #include "oat_quick_method_header.h"
39 #include "object_callbacks.h"
40 #include "scoped_thread_state_change-inl.h"
41 #include "stack.h"
42 #include "thread_list.h"
43
44 namespace art {
45 namespace jit {
46
47 static constexpr int kProtAll = PROT_READ | PROT_WRITE | PROT_EXEC;
48 static constexpr int kProtData = PROT_READ | PROT_WRITE;
49 static constexpr int kProtCode = PROT_READ | PROT_EXEC;
50
51 static constexpr size_t kCodeSizeLogThreshold = 50 * KB;
52 static constexpr size_t kStackMapSizeLogThreshold = 50 * KB;
53
54 #define CHECKED_MPROTECT(memory, size, prot) \
55 do { \
56 int rc = mprotect(memory, size, prot); \
57 if (UNLIKELY(rc != 0)) { \
58 errno = rc; \
59 PLOG(FATAL) << "Failed to mprotect jit code cache"; \
60 } \
61 } while (false) \
62
Create(size_t initial_capacity,size_t max_capacity,bool generate_debug_info,std::string * error_msg)63 JitCodeCache* JitCodeCache::Create(size_t initial_capacity,
64 size_t max_capacity,
65 bool generate_debug_info,
66 std::string* error_msg) {
67 ScopedTrace trace(__PRETTY_FUNCTION__);
68 CHECK_GE(max_capacity, initial_capacity);
69
70 // Generating debug information is for using the Linux perf tool on
71 // host which does not work with ashmem.
72 // Also, target linux does not support ashmem.
73 bool use_ashmem = !generate_debug_info && !kIsTargetLinux;
74
75 // With 'perf', we want a 1-1 mapping between an address and a method.
76 bool garbage_collect_code = !generate_debug_info;
77
78 // We need to have 32 bit offsets from method headers in code cache which point to things
79 // in the data cache. If the maps are more than 4G apart, having multiple maps wouldn't work.
80 // Ensure we're below 1 GB to be safe.
81 if (max_capacity > 1 * GB) {
82 std::ostringstream oss;
83 oss << "Maxium code cache capacity is limited to 1 GB, "
84 << PrettySize(max_capacity) << " is too big";
85 *error_msg = oss.str();
86 return nullptr;
87 }
88
89 std::string error_str;
90 // Map name specific for android_os_Debug.cpp accounting.
91 // Map in low 4gb to simplify accessing root tables for x86_64.
92 // We could do PC-relative addressing to avoid this problem, but that
93 // would require reserving code and data area before submitting, which
94 // means more windows for the code memory to be RWX.
95 std::unique_ptr<MemMap> data_map(MemMap::MapAnonymous(
96 "data-code-cache", nullptr,
97 max_capacity,
98 kProtData,
99 /* low_4gb */ true,
100 /* reuse */ false,
101 &error_str,
102 use_ashmem));
103 if (data_map == nullptr) {
104 std::ostringstream oss;
105 oss << "Failed to create read write cache: " << error_str << " size=" << max_capacity;
106 *error_msg = oss.str();
107 return nullptr;
108 }
109
110 // Align both capacities to page size, as that's the unit mspaces use.
111 initial_capacity = RoundDown(initial_capacity, 2 * kPageSize);
112 max_capacity = RoundDown(max_capacity, 2 * kPageSize);
113
114 // Data cache is 1 / 2 of the map.
115 // TODO: Make this variable?
116 size_t data_size = max_capacity / 2;
117 size_t code_size = max_capacity - data_size;
118 DCHECK_EQ(code_size + data_size, max_capacity);
119 uint8_t* divider = data_map->Begin() + data_size;
120
121 MemMap* code_map =
122 data_map->RemapAtEnd(divider, "jit-code-cache", kProtAll, &error_str, use_ashmem);
123 if (code_map == nullptr) {
124 std::ostringstream oss;
125 oss << "Failed to create read write execute cache: " << error_str << " size=" << max_capacity;
126 *error_msg = oss.str();
127 return nullptr;
128 }
129 DCHECK_EQ(code_map->Begin(), divider);
130 data_size = initial_capacity / 2;
131 code_size = initial_capacity - data_size;
132 DCHECK_EQ(code_size + data_size, initial_capacity);
133 return new JitCodeCache(
134 code_map, data_map.release(), code_size, data_size, max_capacity, garbage_collect_code);
135 }
136
JitCodeCache(MemMap * code_map,MemMap * data_map,size_t initial_code_capacity,size_t initial_data_capacity,size_t max_capacity,bool garbage_collect_code)137 JitCodeCache::JitCodeCache(MemMap* code_map,
138 MemMap* data_map,
139 size_t initial_code_capacity,
140 size_t initial_data_capacity,
141 size_t max_capacity,
142 bool garbage_collect_code)
143 : lock_("Jit code cache", kJitCodeCacheLock),
144 lock_cond_("Jit code cache condition variable", lock_),
145 collection_in_progress_(false),
146 code_map_(code_map),
147 data_map_(data_map),
148 max_capacity_(max_capacity),
149 current_capacity_(initial_code_capacity + initial_data_capacity),
150 code_end_(initial_code_capacity),
151 data_end_(initial_data_capacity),
152 last_collection_increased_code_cache_(false),
153 last_update_time_ns_(0),
154 garbage_collect_code_(garbage_collect_code),
155 used_memory_for_data_(0),
156 used_memory_for_code_(0),
157 number_of_compilations_(0),
158 number_of_osr_compilations_(0),
159 number_of_collections_(0),
160 histogram_stack_map_memory_use_("Memory used for stack maps", 16),
161 histogram_code_memory_use_("Memory used for compiled code", 16),
162 histogram_profiling_info_memory_use_("Memory used for profiling info", 16),
163 is_weak_access_enabled_(true),
164 inline_cache_cond_("Jit inline cache condition variable", lock_) {
165
166 DCHECK_GE(max_capacity, initial_code_capacity + initial_data_capacity);
167 code_mspace_ = create_mspace_with_base(code_map_->Begin(), code_end_, false /*locked*/);
168 data_mspace_ = create_mspace_with_base(data_map_->Begin(), data_end_, false /*locked*/);
169
170 if (code_mspace_ == nullptr || data_mspace_ == nullptr) {
171 PLOG(FATAL) << "create_mspace_with_base failed";
172 }
173
174 SetFootprintLimit(current_capacity_);
175
176 CHECKED_MPROTECT(code_map_->Begin(), code_map_->Size(), kProtCode);
177 CHECKED_MPROTECT(data_map_->Begin(), data_map_->Size(), kProtData);
178
179 VLOG(jit) << "Created jit code cache: initial data size="
180 << PrettySize(initial_data_capacity)
181 << ", initial code size="
182 << PrettySize(initial_code_capacity);
183 }
184
ContainsPc(const void * ptr) const185 bool JitCodeCache::ContainsPc(const void* ptr) const {
186 return code_map_->Begin() <= ptr && ptr < code_map_->End();
187 }
188
ContainsMethod(ArtMethod * method)189 bool JitCodeCache::ContainsMethod(ArtMethod* method) {
190 MutexLock mu(Thread::Current(), lock_);
191 for (auto& it : method_code_map_) {
192 if (it.second == method) {
193 return true;
194 }
195 }
196 return false;
197 }
198
199 class ScopedCodeCacheWrite : ScopedTrace {
200 public:
ScopedCodeCacheWrite(MemMap * code_map,bool only_for_tlb_shootdown=false)201 explicit ScopedCodeCacheWrite(MemMap* code_map, bool only_for_tlb_shootdown = false)
202 : ScopedTrace("ScopedCodeCacheWrite"),
203 code_map_(code_map),
204 only_for_tlb_shootdown_(only_for_tlb_shootdown) {
205 ScopedTrace trace("mprotect all");
206 CHECKED_MPROTECT(
207 code_map_->Begin(), only_for_tlb_shootdown_ ? kPageSize : code_map_->Size(), kProtAll);
208 }
~ScopedCodeCacheWrite()209 ~ScopedCodeCacheWrite() {
210 ScopedTrace trace("mprotect code");
211 CHECKED_MPROTECT(
212 code_map_->Begin(), only_for_tlb_shootdown_ ? kPageSize : code_map_->Size(), kProtCode);
213 }
214 private:
215 MemMap* const code_map_;
216
217 // If we're using ScopedCacheWrite only for TLB shootdown, we limit the scope of mprotect to
218 // one page.
219 const bool only_for_tlb_shootdown_;
220
221 DISALLOW_COPY_AND_ASSIGN(ScopedCodeCacheWrite);
222 };
223
CommitCode(Thread * self,ArtMethod * method,uint8_t * stack_map,uint8_t * method_info,uint8_t * roots_data,size_t frame_size_in_bytes,size_t core_spill_mask,size_t fp_spill_mask,const uint8_t * code,size_t code_size,size_t data_size,bool osr,Handle<mirror::ObjectArray<mirror::Object>> roots,bool has_should_deoptimize_flag,const ArenaSet<ArtMethod * > & cha_single_implementation_list)224 uint8_t* JitCodeCache::CommitCode(Thread* self,
225 ArtMethod* method,
226 uint8_t* stack_map,
227 uint8_t* method_info,
228 uint8_t* roots_data,
229 size_t frame_size_in_bytes,
230 size_t core_spill_mask,
231 size_t fp_spill_mask,
232 const uint8_t* code,
233 size_t code_size,
234 size_t data_size,
235 bool osr,
236 Handle<mirror::ObjectArray<mirror::Object>> roots,
237 bool has_should_deoptimize_flag,
238 const ArenaSet<ArtMethod*>& cha_single_implementation_list) {
239 uint8_t* result = CommitCodeInternal(self,
240 method,
241 stack_map,
242 method_info,
243 roots_data,
244 frame_size_in_bytes,
245 core_spill_mask,
246 fp_spill_mask,
247 code,
248 code_size,
249 data_size,
250 osr,
251 roots,
252 has_should_deoptimize_flag,
253 cha_single_implementation_list);
254 if (result == nullptr) {
255 // Retry.
256 GarbageCollectCache(self);
257 result = CommitCodeInternal(self,
258 method,
259 stack_map,
260 method_info,
261 roots_data,
262 frame_size_in_bytes,
263 core_spill_mask,
264 fp_spill_mask,
265 code,
266 code_size,
267 data_size,
268 osr,
269 roots,
270 has_should_deoptimize_flag,
271 cha_single_implementation_list);
272 }
273 return result;
274 }
275
WaitForPotentialCollectionToComplete(Thread * self)276 bool JitCodeCache::WaitForPotentialCollectionToComplete(Thread* self) {
277 bool in_collection = false;
278 while (collection_in_progress_) {
279 in_collection = true;
280 lock_cond_.Wait(self);
281 }
282 return in_collection;
283 }
284
FromCodeToAllocation(const void * code)285 static uintptr_t FromCodeToAllocation(const void* code) {
286 size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
287 return reinterpret_cast<uintptr_t>(code) - RoundUp(sizeof(OatQuickMethodHeader), alignment);
288 }
289
ComputeRootTableSize(uint32_t number_of_roots)290 static uint32_t ComputeRootTableSize(uint32_t number_of_roots) {
291 return sizeof(uint32_t) + number_of_roots * sizeof(GcRoot<mirror::Object>);
292 }
293
GetNumberOfRoots(const uint8_t * stack_map)294 static uint32_t GetNumberOfRoots(const uint8_t* stack_map) {
295 // The length of the table is stored just before the stack map (and therefore at the end of
296 // the table itself), in order to be able to fetch it from a `stack_map` pointer.
297 return reinterpret_cast<const uint32_t*>(stack_map)[-1];
298 }
299
FillRootTableLength(uint8_t * roots_data,uint32_t length)300 static void FillRootTableLength(uint8_t* roots_data, uint32_t length) {
301 // Store the length of the table at the end. This will allow fetching it from a `stack_map`
302 // pointer.
303 reinterpret_cast<uint32_t*>(roots_data)[length] = length;
304 }
305
FromStackMapToRoots(const uint8_t * stack_map_data)306 static const uint8_t* FromStackMapToRoots(const uint8_t* stack_map_data) {
307 return stack_map_data - ComputeRootTableSize(GetNumberOfRoots(stack_map_data));
308 }
309
FillRootTable(uint8_t * roots_data,Handle<mirror::ObjectArray<mirror::Object>> roots)310 static void FillRootTable(uint8_t* roots_data, Handle<mirror::ObjectArray<mirror::Object>> roots)
311 REQUIRES_SHARED(Locks::mutator_lock_) {
312 GcRoot<mirror::Object>* gc_roots = reinterpret_cast<GcRoot<mirror::Object>*>(roots_data);
313 const uint32_t length = roots->GetLength();
314 // Put all roots in `roots_data`.
315 for (uint32_t i = 0; i < length; ++i) {
316 ObjPtr<mirror::Object> object = roots->Get(i);
317 if (kIsDebugBuild) {
318 // Ensure the string is strongly interned. b/32995596
319 if (object->IsString()) {
320 ObjPtr<mirror::String> str = reinterpret_cast<mirror::String*>(object.Ptr());
321 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
322 CHECK(class_linker->GetInternTable()->LookupStrong(Thread::Current(), str) != nullptr);
323 }
324 }
325 gc_roots[i] = GcRoot<mirror::Object>(object);
326 }
327 }
328
GetRootTable(const void * code_ptr,uint32_t * number_of_roots=nullptr)329 static uint8_t* GetRootTable(const void* code_ptr, uint32_t* number_of_roots = nullptr) {
330 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
331 uint8_t* data = method_header->GetOptimizedCodeInfoPtr();
332 uint32_t roots = GetNumberOfRoots(data);
333 if (number_of_roots != nullptr) {
334 *number_of_roots = roots;
335 }
336 return data - ComputeRootTableSize(roots);
337 }
338
339 // Use a sentinel for marking entries in the JIT table that have been cleared.
340 // This helps diagnosing in case the compiled code tries to wrongly access such
341 // entries.
342 static mirror::Class* const weak_sentinel =
343 reinterpret_cast<mirror::Class*>(Context::kBadGprBase + 0xff);
344
345 // Helper for the GC to process a weak class in a JIT root table.
ProcessWeakClass(GcRoot<mirror::Class> * root_ptr,IsMarkedVisitor * visitor,mirror::Class * update)346 static inline void ProcessWeakClass(GcRoot<mirror::Class>* root_ptr,
347 IsMarkedVisitor* visitor,
348 mirror::Class* update)
349 REQUIRES_SHARED(Locks::mutator_lock_) {
350 // This does not need a read barrier because this is called by GC.
351 mirror::Class* cls = root_ptr->Read<kWithoutReadBarrier>();
352 if (cls != nullptr && cls != weak_sentinel) {
353 DCHECK((cls->IsClass<kDefaultVerifyFlags, kWithoutReadBarrier>()));
354 // Look at the classloader of the class to know if it has been unloaded.
355 // This does not need a read barrier because this is called by GC.
356 mirror::Object* class_loader =
357 cls->GetClassLoader<kDefaultVerifyFlags, kWithoutReadBarrier>();
358 if (class_loader == nullptr || visitor->IsMarked(class_loader) != nullptr) {
359 // The class loader is live, update the entry if the class has moved.
360 mirror::Class* new_cls = down_cast<mirror::Class*>(visitor->IsMarked(cls));
361 // Note that new_object can be null for CMS and newly allocated objects.
362 if (new_cls != nullptr && new_cls != cls) {
363 *root_ptr = GcRoot<mirror::Class>(new_cls);
364 }
365 } else {
366 // The class loader is not live, clear the entry.
367 *root_ptr = GcRoot<mirror::Class>(update);
368 }
369 }
370 }
371
SweepRootTables(IsMarkedVisitor * visitor)372 void JitCodeCache::SweepRootTables(IsMarkedVisitor* visitor) {
373 MutexLock mu(Thread::Current(), lock_);
374 for (const auto& entry : method_code_map_) {
375 uint32_t number_of_roots = 0;
376 uint8_t* roots_data = GetRootTable(entry.first, &number_of_roots);
377 GcRoot<mirror::Object>* roots = reinterpret_cast<GcRoot<mirror::Object>*>(roots_data);
378 for (uint32_t i = 0; i < number_of_roots; ++i) {
379 // This does not need a read barrier because this is called by GC.
380 mirror::Object* object = roots[i].Read<kWithoutReadBarrier>();
381 if (object == nullptr || object == weak_sentinel) {
382 // entry got deleted in a previous sweep.
383 } else if (object->IsString<kDefaultVerifyFlags, kWithoutReadBarrier>()) {
384 mirror::Object* new_object = visitor->IsMarked(object);
385 // We know the string is marked because it's a strongly-interned string that
386 // is always alive. The IsMarked implementation of the CMS collector returns
387 // null for newly allocated objects, but we know those haven't moved. Therefore,
388 // only update the entry if we get a different non-null string.
389 // TODO: Do not use IsMarked for j.l.Class, and adjust once we move this method
390 // out of the weak access/creation pause. b/32167580
391 if (new_object != nullptr && new_object != object) {
392 DCHECK(new_object->IsString());
393 roots[i] = GcRoot<mirror::Object>(new_object);
394 }
395 } else {
396 ProcessWeakClass(
397 reinterpret_cast<GcRoot<mirror::Class>*>(&roots[i]), visitor, weak_sentinel);
398 }
399 }
400 }
401 // Walk over inline caches to clear entries containing unloaded classes.
402 for (ProfilingInfo* info : profiling_infos_) {
403 for (size_t i = 0; i < info->number_of_inline_caches_; ++i) {
404 InlineCache* cache = &info->cache_[i];
405 for (size_t j = 0; j < InlineCache::kIndividualCacheSize; ++j) {
406 ProcessWeakClass(&cache->classes_[j], visitor, nullptr);
407 }
408 }
409 }
410 }
411
FreeCode(const void * code_ptr)412 void JitCodeCache::FreeCode(const void* code_ptr) {
413 uintptr_t allocation = FromCodeToAllocation(code_ptr);
414 // Notify native debugger that we are about to remove the code.
415 // It does nothing if we are not using native debugger.
416 DeleteJITCodeEntryForAddress(reinterpret_cast<uintptr_t>(code_ptr));
417 FreeData(GetRootTable(code_ptr));
418 FreeCode(reinterpret_cast<uint8_t*>(allocation));
419 }
420
FreeAllMethodHeaders(const std::unordered_set<OatQuickMethodHeader * > & method_headers)421 void JitCodeCache::FreeAllMethodHeaders(
422 const std::unordered_set<OatQuickMethodHeader*>& method_headers) {
423 {
424 MutexLock mu(Thread::Current(), *Locks::cha_lock_);
425 Runtime::Current()->GetClassLinker()->GetClassHierarchyAnalysis()
426 ->RemoveDependentsWithMethodHeaders(method_headers);
427 }
428
429 // We need to remove entries in method_headers from CHA dependencies
430 // first since once we do FreeCode() below, the memory can be reused
431 // so it's possible for the same method_header to start representing
432 // different compile code.
433 MutexLock mu(Thread::Current(), lock_);
434 ScopedCodeCacheWrite scc(code_map_.get());
435 for (const OatQuickMethodHeader* method_header : method_headers) {
436 FreeCode(method_header->GetCode());
437 }
438 }
439
RemoveMethodsIn(Thread * self,const LinearAlloc & alloc)440 void JitCodeCache::RemoveMethodsIn(Thread* self, const LinearAlloc& alloc) {
441 ScopedTrace trace(__PRETTY_FUNCTION__);
442 // We use a set to first collect all method_headers whose code need to be
443 // removed. We need to free the underlying code after we remove CHA dependencies
444 // for entries in this set. And it's more efficient to iterate through
445 // the CHA dependency map just once with an unordered_set.
446 std::unordered_set<OatQuickMethodHeader*> method_headers;
447 {
448 MutexLock mu(self, lock_);
449 // We do not check if a code cache GC is in progress, as this method comes
450 // with the classlinker_classes_lock_ held, and suspending ourselves could
451 // lead to a deadlock.
452 {
453 ScopedCodeCacheWrite scc(code_map_.get());
454 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
455 if (alloc.ContainsUnsafe(it->second)) {
456 method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->first));
457 it = method_code_map_.erase(it);
458 } else {
459 ++it;
460 }
461 }
462 }
463 for (auto it = osr_code_map_.begin(); it != osr_code_map_.end();) {
464 if (alloc.ContainsUnsafe(it->first)) {
465 // Note that the code has already been pushed to method_headers in the loop
466 // above and is going to be removed in FreeCode() below.
467 it = osr_code_map_.erase(it);
468 } else {
469 ++it;
470 }
471 }
472 for (auto it = profiling_infos_.begin(); it != profiling_infos_.end();) {
473 ProfilingInfo* info = *it;
474 if (alloc.ContainsUnsafe(info->GetMethod())) {
475 info->GetMethod()->SetProfilingInfo(nullptr);
476 FreeData(reinterpret_cast<uint8_t*>(info));
477 it = profiling_infos_.erase(it);
478 } else {
479 ++it;
480 }
481 }
482 }
483 FreeAllMethodHeaders(method_headers);
484 }
485
IsWeakAccessEnabled(Thread * self) const486 bool JitCodeCache::IsWeakAccessEnabled(Thread* self) const {
487 return kUseReadBarrier
488 ? self->GetWeakRefAccessEnabled()
489 : is_weak_access_enabled_.LoadSequentiallyConsistent();
490 }
491
WaitUntilInlineCacheAccessible(Thread * self)492 void JitCodeCache::WaitUntilInlineCacheAccessible(Thread* self) {
493 if (IsWeakAccessEnabled(self)) {
494 return;
495 }
496 ScopedThreadSuspension sts(self, kWaitingWeakGcRootRead);
497 MutexLock mu(self, lock_);
498 while (!IsWeakAccessEnabled(self)) {
499 inline_cache_cond_.Wait(self);
500 }
501 }
502
BroadcastForInlineCacheAccess()503 void JitCodeCache::BroadcastForInlineCacheAccess() {
504 Thread* self = Thread::Current();
505 MutexLock mu(self, lock_);
506 inline_cache_cond_.Broadcast(self);
507 }
508
AllowInlineCacheAccess()509 void JitCodeCache::AllowInlineCacheAccess() {
510 DCHECK(!kUseReadBarrier);
511 is_weak_access_enabled_.StoreSequentiallyConsistent(true);
512 BroadcastForInlineCacheAccess();
513 }
514
DisallowInlineCacheAccess()515 void JitCodeCache::DisallowInlineCacheAccess() {
516 DCHECK(!kUseReadBarrier);
517 is_weak_access_enabled_.StoreSequentiallyConsistent(false);
518 }
519
CopyInlineCacheInto(const InlineCache & ic,Handle<mirror::ObjectArray<mirror::Class>> array)520 void JitCodeCache::CopyInlineCacheInto(const InlineCache& ic,
521 Handle<mirror::ObjectArray<mirror::Class>> array) {
522 WaitUntilInlineCacheAccessible(Thread::Current());
523 // Note that we don't need to lock `lock_` here, the compiler calling
524 // this method has already ensured the inline cache will not be deleted.
525 for (size_t in_cache = 0, in_array = 0;
526 in_cache < InlineCache::kIndividualCacheSize;
527 ++in_cache) {
528 mirror::Class* object = ic.classes_[in_cache].Read();
529 if (object != nullptr) {
530 array->Set(in_array++, object);
531 }
532 }
533 }
534
ClearMethodCounter(ArtMethod * method,bool was_warm)535 static void ClearMethodCounter(ArtMethod* method, bool was_warm) {
536 if (was_warm) {
537 method->AddAccessFlags(kAccPreviouslyWarm);
538 }
539 // We reset the counter to 1 so that the profile knows that the method was executed at least once.
540 // This is required for layout purposes.
541 // We also need to make sure we'll pass the warmup threshold again, so we set to 0 if
542 // the warmup threshold is 1.
543 uint16_t jit_warmup_threshold = Runtime::Current()->GetJITOptions()->GetWarmupThreshold();
544 method->SetCounter(std::min(jit_warmup_threshold - 1, 1));
545 }
546
CommitCodeInternal(Thread * self,ArtMethod * method,uint8_t * stack_map,uint8_t * method_info,uint8_t * roots_data,size_t frame_size_in_bytes,size_t core_spill_mask,size_t fp_spill_mask,const uint8_t * code,size_t code_size,size_t data_size,bool osr,Handle<mirror::ObjectArray<mirror::Object>> roots,bool has_should_deoptimize_flag,const ArenaSet<ArtMethod * > & cha_single_implementation_list)547 uint8_t* JitCodeCache::CommitCodeInternal(Thread* self,
548 ArtMethod* method,
549 uint8_t* stack_map,
550 uint8_t* method_info,
551 uint8_t* roots_data,
552 size_t frame_size_in_bytes,
553 size_t core_spill_mask,
554 size_t fp_spill_mask,
555 const uint8_t* code,
556 size_t code_size,
557 size_t data_size,
558 bool osr,
559 Handle<mirror::ObjectArray<mirror::Object>> roots,
560 bool has_should_deoptimize_flag,
561 const ArenaSet<ArtMethod*>&
562 cha_single_implementation_list) {
563 DCHECK(stack_map != nullptr);
564 size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
565 // Ensure the header ends up at expected instruction alignment.
566 size_t header_size = RoundUp(sizeof(OatQuickMethodHeader), alignment);
567 size_t total_size = header_size + code_size;
568
569 OatQuickMethodHeader* method_header = nullptr;
570 uint8_t* code_ptr = nullptr;
571 uint8_t* memory = nullptr;
572 {
573 ScopedThreadSuspension sts(self, kSuspended);
574 MutexLock mu(self, lock_);
575 WaitForPotentialCollectionToComplete(self);
576 {
577 ScopedCodeCacheWrite scc(code_map_.get());
578 memory = AllocateCode(total_size);
579 if (memory == nullptr) {
580 return nullptr;
581 }
582 code_ptr = memory + header_size;
583
584 std::copy(code, code + code_size, code_ptr);
585 method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
586 new (method_header) OatQuickMethodHeader(
587 code_ptr - stack_map,
588 code_ptr - method_info,
589 frame_size_in_bytes,
590 core_spill_mask,
591 fp_spill_mask,
592 code_size);
593 // Flush caches before we remove write permission because some ARMv8 Qualcomm kernels may
594 // trigger a segfault if a page fault occurs when requesting a cache maintenance operation.
595 // This is a kernel bug that we need to work around until affected devices (e.g. Nexus 5X and
596 // 6P) stop being supported or their kernels are fixed.
597 //
598 // For reference, this behavior is caused by this commit:
599 // https://android.googlesource.com/kernel/msm/+/3fbe6bc28a6b9939d0650f2f17eb5216c719950c
600 FlushInstructionCache(reinterpret_cast<char*>(code_ptr),
601 reinterpret_cast<char*>(code_ptr + code_size));
602 DCHECK(!Runtime::Current()->IsAotCompiler());
603 if (has_should_deoptimize_flag) {
604 method_header->SetHasShouldDeoptimizeFlag();
605 }
606 }
607
608 number_of_compilations_++;
609 }
610 // We need to update the entry point in the runnable state for the instrumentation.
611 {
612 // Need cha_lock_ for checking all single-implementation flags and register
613 // dependencies.
614 MutexLock cha_mu(self, *Locks::cha_lock_);
615 bool single_impl_still_valid = true;
616 for (ArtMethod* single_impl : cha_single_implementation_list) {
617 if (!single_impl->HasSingleImplementation()) {
618 // Simply discard the compiled code. Clear the counter so that it may be recompiled later.
619 // Hopefully the class hierarchy will be more stable when compilation is retried.
620 single_impl_still_valid = false;
621 ClearMethodCounter(method, /*was_warm*/ false);
622 break;
623 }
624 }
625
626 // Discard the code if any single-implementation assumptions are now invalid.
627 if (!single_impl_still_valid) {
628 VLOG(jit) << "JIT discarded jitted code due to invalid single-implementation assumptions.";
629 return nullptr;
630 }
631 DCHECK(cha_single_implementation_list.empty() || !Runtime::Current()->IsJavaDebuggable())
632 << "Should not be using cha on debuggable apps/runs!";
633
634 for (ArtMethod* single_impl : cha_single_implementation_list) {
635 Runtime::Current()->GetClassLinker()->GetClassHierarchyAnalysis()->AddDependency(
636 single_impl, method, method_header);
637 }
638
639 // The following needs to be guarded by cha_lock_ also. Otherwise it's
640 // possible that the compiled code is considered invalidated by some class linking,
641 // but below we still make the compiled code valid for the method.
642 MutexLock mu(self, lock_);
643 // Fill the root table before updating the entry point.
644 DCHECK_EQ(FromStackMapToRoots(stack_map), roots_data);
645 DCHECK_LE(roots_data, stack_map);
646 FillRootTable(roots_data, roots);
647 {
648 // Flush data cache, as compiled code references literals in it.
649 // We also need a TLB shootdown to act as memory barrier across cores.
650 ScopedCodeCacheWrite ccw(code_map_.get(), /* only_for_tlb_shootdown */ true);
651 FlushDataCache(reinterpret_cast<char*>(roots_data),
652 reinterpret_cast<char*>(roots_data + data_size));
653 }
654 method_code_map_.Put(code_ptr, method);
655 if (osr) {
656 number_of_osr_compilations_++;
657 osr_code_map_.Put(method, code_ptr);
658 } else {
659 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
660 method, method_header->GetEntryPoint());
661 }
662 if (collection_in_progress_) {
663 // We need to update the live bitmap if there is a GC to ensure it sees this new
664 // code.
665 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
666 }
667 last_update_time_ns_.StoreRelease(NanoTime());
668 VLOG(jit)
669 << "JIT added (osr=" << std::boolalpha << osr << std::noboolalpha << ") "
670 << ArtMethod::PrettyMethod(method) << "@" << method
671 << " ccache_size=" << PrettySize(CodeCacheSizeLocked()) << ": "
672 << " dcache_size=" << PrettySize(DataCacheSizeLocked()) << ": "
673 << reinterpret_cast<const void*>(method_header->GetEntryPoint()) << ","
674 << reinterpret_cast<const void*>(method_header->GetEntryPoint() +
675 method_header->GetCodeSize());
676 histogram_code_memory_use_.AddValue(code_size);
677 if (code_size > kCodeSizeLogThreshold) {
678 LOG(INFO) << "JIT allocated "
679 << PrettySize(code_size)
680 << " for compiled code of "
681 << ArtMethod::PrettyMethod(method);
682 }
683 }
684
685 return reinterpret_cast<uint8_t*>(method_header);
686 }
687
CodeCacheSize()688 size_t JitCodeCache::CodeCacheSize() {
689 MutexLock mu(Thread::Current(), lock_);
690 return CodeCacheSizeLocked();
691 }
692
RemoveMethod(ArtMethod * method,bool release_memory)693 bool JitCodeCache::RemoveMethod(ArtMethod* method, bool release_memory) {
694 MutexLock mu(Thread::Current(), lock_);
695 if (method->IsNative()) {
696 return false;
697 }
698
699 bool in_cache = false;
700 {
701 ScopedCodeCacheWrite ccw(code_map_.get());
702 for (auto code_iter = method_code_map_.begin(); code_iter != method_code_map_.end();) {
703 if (code_iter->second == method) {
704 if (release_memory) {
705 FreeCode(code_iter->first);
706 }
707 code_iter = method_code_map_.erase(code_iter);
708 in_cache = true;
709 continue;
710 }
711 ++code_iter;
712 }
713 }
714
715 bool osr = false;
716 auto code_map = osr_code_map_.find(method);
717 if (code_map != osr_code_map_.end()) {
718 osr_code_map_.erase(code_map);
719 osr = true;
720 }
721
722 if (!in_cache) {
723 return false;
724 }
725
726 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
727 if (info != nullptr) {
728 auto profile = std::find(profiling_infos_.begin(), profiling_infos_.end(), info);
729 DCHECK(profile != profiling_infos_.end());
730 profiling_infos_.erase(profile);
731 }
732 method->SetProfilingInfo(nullptr);
733 method->ClearCounter();
734 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
735 method, GetQuickToInterpreterBridge());
736 VLOG(jit)
737 << "JIT removed (osr=" << std::boolalpha << osr << std::noboolalpha << ") "
738 << ArtMethod::PrettyMethod(method) << "@" << method
739 << " ccache_size=" << PrettySize(CodeCacheSizeLocked()) << ": "
740 << " dcache_size=" << PrettySize(DataCacheSizeLocked());
741 return true;
742 }
743
744 // This notifies the code cache that the given method has been redefined and that it should remove
745 // any cached information it has on the method. All threads must be suspended before calling this
746 // method. The compiled code for the method (if there is any) must not be in any threads call stack.
NotifyMethodRedefined(ArtMethod * method)747 void JitCodeCache::NotifyMethodRedefined(ArtMethod* method) {
748 MutexLock mu(Thread::Current(), lock_);
749 if (method->IsNative()) {
750 return;
751 }
752 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
753 if (info != nullptr) {
754 auto profile = std::find(profiling_infos_.begin(), profiling_infos_.end(), info);
755 DCHECK(profile != profiling_infos_.end());
756 profiling_infos_.erase(profile);
757 }
758 method->SetProfilingInfo(nullptr);
759 ScopedCodeCacheWrite ccw(code_map_.get());
760 for (auto code_iter = method_code_map_.begin(); code_iter != method_code_map_.end();) {
761 if (code_iter->second == method) {
762 FreeCode(code_iter->first);
763 code_iter = method_code_map_.erase(code_iter);
764 continue;
765 }
766 ++code_iter;
767 }
768 auto code_map = osr_code_map_.find(method);
769 if (code_map != osr_code_map_.end()) {
770 osr_code_map_.erase(code_map);
771 }
772 }
773
774 // This invalidates old_method. Once this function returns one can no longer use old_method to
775 // execute code unless it is fixed up. This fixup will happen later in the process of installing a
776 // class redefinition.
777 // TODO We should add some info to ArtMethod to note that 'old_method' has been invalidated and
778 // shouldn't be used since it is no longer logically in the jit code cache.
779 // TODO We should add DCHECKS that validate that the JIT is paused when this method is entered.
MoveObsoleteMethod(ArtMethod * old_method,ArtMethod * new_method)780 void JitCodeCache::MoveObsoleteMethod(ArtMethod* old_method, ArtMethod* new_method) {
781 // Native methods have no profiling info and need no special handling from the JIT code cache.
782 if (old_method->IsNative()) {
783 return;
784 }
785 MutexLock mu(Thread::Current(), lock_);
786 // Update ProfilingInfo to the new one and remove it from the old_method.
787 if (old_method->GetProfilingInfo(kRuntimePointerSize) != nullptr) {
788 DCHECK_EQ(old_method->GetProfilingInfo(kRuntimePointerSize)->GetMethod(), old_method);
789 ProfilingInfo* info = old_method->GetProfilingInfo(kRuntimePointerSize);
790 old_method->SetProfilingInfo(nullptr);
791 // Since the JIT should be paused and all threads suspended by the time this is called these
792 // checks should always pass.
793 DCHECK(!info->IsInUseByCompiler());
794 new_method->SetProfilingInfo(info);
795 info->method_ = new_method;
796 }
797 // Update method_code_map_ to point to the new method.
798 for (auto& it : method_code_map_) {
799 if (it.second == old_method) {
800 it.second = new_method;
801 }
802 }
803 // Update osr_code_map_ to point to the new method.
804 auto code_map = osr_code_map_.find(old_method);
805 if (code_map != osr_code_map_.end()) {
806 osr_code_map_.Put(new_method, code_map->second);
807 osr_code_map_.erase(old_method);
808 }
809 }
810
CodeCacheSizeLocked()811 size_t JitCodeCache::CodeCacheSizeLocked() {
812 return used_memory_for_code_;
813 }
814
DataCacheSize()815 size_t JitCodeCache::DataCacheSize() {
816 MutexLock mu(Thread::Current(), lock_);
817 return DataCacheSizeLocked();
818 }
819
DataCacheSizeLocked()820 size_t JitCodeCache::DataCacheSizeLocked() {
821 return used_memory_for_data_;
822 }
823
ClearData(Thread * self,uint8_t * stack_map_data,uint8_t * roots_data)824 void JitCodeCache::ClearData(Thread* self,
825 uint8_t* stack_map_data,
826 uint8_t* roots_data) {
827 DCHECK_EQ(FromStackMapToRoots(stack_map_data), roots_data);
828 MutexLock mu(self, lock_);
829 FreeData(reinterpret_cast<uint8_t*>(roots_data));
830 }
831
ReserveData(Thread * self,size_t stack_map_size,size_t method_info_size,size_t number_of_roots,ArtMethod * method,uint8_t ** stack_map_data,uint8_t ** method_info_data,uint8_t ** roots_data)832 size_t JitCodeCache::ReserveData(Thread* self,
833 size_t stack_map_size,
834 size_t method_info_size,
835 size_t number_of_roots,
836 ArtMethod* method,
837 uint8_t** stack_map_data,
838 uint8_t** method_info_data,
839 uint8_t** roots_data) {
840 size_t table_size = ComputeRootTableSize(number_of_roots);
841 size_t size = RoundUp(stack_map_size + method_info_size + table_size, sizeof(void*));
842 uint8_t* result = nullptr;
843
844 {
845 ScopedThreadSuspension sts(self, kSuspended);
846 MutexLock mu(self, lock_);
847 WaitForPotentialCollectionToComplete(self);
848 result = AllocateData(size);
849 }
850
851 if (result == nullptr) {
852 // Retry.
853 GarbageCollectCache(self);
854 ScopedThreadSuspension sts(self, kSuspended);
855 MutexLock mu(self, lock_);
856 WaitForPotentialCollectionToComplete(self);
857 result = AllocateData(size);
858 }
859
860 MutexLock mu(self, lock_);
861 histogram_stack_map_memory_use_.AddValue(size);
862 if (size > kStackMapSizeLogThreshold) {
863 LOG(INFO) << "JIT allocated "
864 << PrettySize(size)
865 << " for stack maps of "
866 << ArtMethod::PrettyMethod(method);
867 }
868 if (result != nullptr) {
869 *roots_data = result;
870 *stack_map_data = result + table_size;
871 *method_info_data = *stack_map_data + stack_map_size;
872 FillRootTableLength(*roots_data, number_of_roots);
873 return size;
874 } else {
875 *roots_data = nullptr;
876 *stack_map_data = nullptr;
877 *method_info_data = nullptr;
878 return 0;
879 }
880 }
881
882 class MarkCodeVisitor FINAL : public StackVisitor {
883 public:
MarkCodeVisitor(Thread * thread_in,JitCodeCache * code_cache_in)884 MarkCodeVisitor(Thread* thread_in, JitCodeCache* code_cache_in)
885 : StackVisitor(thread_in, nullptr, StackVisitor::StackWalkKind::kSkipInlinedFrames),
886 code_cache_(code_cache_in),
887 bitmap_(code_cache_->GetLiveBitmap()) {}
888
VisitFrame()889 bool VisitFrame() OVERRIDE REQUIRES_SHARED(Locks::mutator_lock_) {
890 const OatQuickMethodHeader* method_header = GetCurrentOatQuickMethodHeader();
891 if (method_header == nullptr) {
892 return true;
893 }
894 const void* code = method_header->GetCode();
895 if (code_cache_->ContainsPc(code)) {
896 // Use the atomic set version, as multiple threads are executing this code.
897 bitmap_->AtomicTestAndSet(FromCodeToAllocation(code));
898 }
899 return true;
900 }
901
902 private:
903 JitCodeCache* const code_cache_;
904 CodeCacheBitmap* const bitmap_;
905 };
906
907 class MarkCodeClosure FINAL : public Closure {
908 public:
MarkCodeClosure(JitCodeCache * code_cache,Barrier * barrier)909 MarkCodeClosure(JitCodeCache* code_cache, Barrier* barrier)
910 : code_cache_(code_cache), barrier_(barrier) {}
911
Run(Thread * thread)912 void Run(Thread* thread) OVERRIDE REQUIRES_SHARED(Locks::mutator_lock_) {
913 ScopedTrace trace(__PRETTY_FUNCTION__);
914 DCHECK(thread == Thread::Current() || thread->IsSuspended());
915 MarkCodeVisitor visitor(thread, code_cache_);
916 visitor.WalkStack();
917 if (kIsDebugBuild) {
918 // The stack walking code queries the side instrumentation stack if it
919 // sees an instrumentation exit pc, so the JIT code of methods in that stack
920 // must have been seen. We sanity check this below.
921 for (const instrumentation::InstrumentationStackFrame& frame
922 : *thread->GetInstrumentationStack()) {
923 // The 'method_' in InstrumentationStackFrame is the one that has return_pc_ in
924 // its stack frame, it is not the method owning return_pc_. We just pass null to
925 // LookupMethodHeader: the method is only checked against in debug builds.
926 OatQuickMethodHeader* method_header =
927 code_cache_->LookupMethodHeader(frame.return_pc_, nullptr);
928 if (method_header != nullptr) {
929 const void* code = method_header->GetCode();
930 CHECK(code_cache_->GetLiveBitmap()->Test(FromCodeToAllocation(code)));
931 }
932 }
933 }
934 barrier_->Pass(Thread::Current());
935 }
936
937 private:
938 JitCodeCache* const code_cache_;
939 Barrier* const barrier_;
940 };
941
NotifyCollectionDone(Thread * self)942 void JitCodeCache::NotifyCollectionDone(Thread* self) {
943 collection_in_progress_ = false;
944 lock_cond_.Broadcast(self);
945 }
946
SetFootprintLimit(size_t new_footprint)947 void JitCodeCache::SetFootprintLimit(size_t new_footprint) {
948 size_t per_space_footprint = new_footprint / 2;
949 DCHECK(IsAlignedParam(per_space_footprint, kPageSize));
950 DCHECK_EQ(per_space_footprint * 2, new_footprint);
951 mspace_set_footprint_limit(data_mspace_, per_space_footprint);
952 {
953 ScopedCodeCacheWrite scc(code_map_.get());
954 mspace_set_footprint_limit(code_mspace_, per_space_footprint);
955 }
956 }
957
IncreaseCodeCacheCapacity()958 bool JitCodeCache::IncreaseCodeCacheCapacity() {
959 if (current_capacity_ == max_capacity_) {
960 return false;
961 }
962
963 // Double the capacity if we're below 1MB, or increase it by 1MB if
964 // we're above.
965 if (current_capacity_ < 1 * MB) {
966 current_capacity_ *= 2;
967 } else {
968 current_capacity_ += 1 * MB;
969 }
970 if (current_capacity_ > max_capacity_) {
971 current_capacity_ = max_capacity_;
972 }
973
974 if (!kIsDebugBuild || VLOG_IS_ON(jit)) {
975 LOG(INFO) << "Increasing code cache capacity to " << PrettySize(current_capacity_);
976 }
977
978 SetFootprintLimit(current_capacity_);
979
980 return true;
981 }
982
MarkCompiledCodeOnThreadStacks(Thread * self)983 void JitCodeCache::MarkCompiledCodeOnThreadStacks(Thread* self) {
984 Barrier barrier(0);
985 size_t threads_running_checkpoint = 0;
986 MarkCodeClosure closure(this, &barrier);
987 threads_running_checkpoint = Runtime::Current()->GetThreadList()->RunCheckpoint(&closure);
988 // Now that we have run our checkpoint, move to a suspended state and wait
989 // for other threads to run the checkpoint.
990 ScopedThreadSuspension sts(self, kSuspended);
991 if (threads_running_checkpoint != 0) {
992 barrier.Increment(self, threads_running_checkpoint);
993 }
994 }
995
ShouldDoFullCollection()996 bool JitCodeCache::ShouldDoFullCollection() {
997 if (current_capacity_ == max_capacity_) {
998 // Always do a full collection when the code cache is full.
999 return true;
1000 } else if (current_capacity_ < kReservedCapacity) {
1001 // Always do partial collection when the code cache size is below the reserved
1002 // capacity.
1003 return false;
1004 } else if (last_collection_increased_code_cache_) {
1005 // This time do a full collection.
1006 return true;
1007 } else {
1008 // This time do a partial collection.
1009 return false;
1010 }
1011 }
1012
GarbageCollectCache(Thread * self)1013 void JitCodeCache::GarbageCollectCache(Thread* self) {
1014 ScopedTrace trace(__FUNCTION__);
1015 if (!garbage_collect_code_) {
1016 MutexLock mu(self, lock_);
1017 IncreaseCodeCacheCapacity();
1018 return;
1019 }
1020
1021 // Wait for an existing collection, or let everyone know we are starting one.
1022 {
1023 ScopedThreadSuspension sts(self, kSuspended);
1024 MutexLock mu(self, lock_);
1025 if (WaitForPotentialCollectionToComplete(self)) {
1026 return;
1027 } else {
1028 number_of_collections_++;
1029 live_bitmap_.reset(CodeCacheBitmap::Create(
1030 "code-cache-bitmap",
1031 reinterpret_cast<uintptr_t>(code_map_->Begin()),
1032 reinterpret_cast<uintptr_t>(code_map_->Begin() + current_capacity_ / 2)));
1033 collection_in_progress_ = true;
1034 }
1035 }
1036
1037 TimingLogger logger("JIT code cache timing logger", true, VLOG_IS_ON(jit));
1038 {
1039 TimingLogger::ScopedTiming st("Code cache collection", &logger);
1040
1041 bool do_full_collection = false;
1042 {
1043 MutexLock mu(self, lock_);
1044 do_full_collection = ShouldDoFullCollection();
1045 }
1046
1047 if (!kIsDebugBuild || VLOG_IS_ON(jit)) {
1048 LOG(INFO) << "Do "
1049 << (do_full_collection ? "full" : "partial")
1050 << " code cache collection, code="
1051 << PrettySize(CodeCacheSize())
1052 << ", data=" << PrettySize(DataCacheSize());
1053 }
1054
1055 DoCollection(self, /* collect_profiling_info */ do_full_collection);
1056
1057 if (!kIsDebugBuild || VLOG_IS_ON(jit)) {
1058 LOG(INFO) << "After code cache collection, code="
1059 << PrettySize(CodeCacheSize())
1060 << ", data=" << PrettySize(DataCacheSize());
1061 }
1062
1063 {
1064 MutexLock mu(self, lock_);
1065
1066 // Increase the code cache only when we do partial collections.
1067 // TODO: base this strategy on how full the code cache is?
1068 if (do_full_collection) {
1069 last_collection_increased_code_cache_ = false;
1070 } else {
1071 last_collection_increased_code_cache_ = true;
1072 IncreaseCodeCacheCapacity();
1073 }
1074
1075 bool next_collection_will_be_full = ShouldDoFullCollection();
1076
1077 // Start polling the liveness of compiled code to prepare for the next full collection.
1078 if (next_collection_will_be_full) {
1079 // Save the entry point of methods we have compiled, and update the entry
1080 // point of those methods to the interpreter. If the method is invoked, the
1081 // interpreter will update its entry point to the compiled code and call it.
1082 for (ProfilingInfo* info : profiling_infos_) {
1083 const void* entry_point = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1084 if (ContainsPc(entry_point)) {
1085 info->SetSavedEntryPoint(entry_point);
1086 // Don't call Instrumentation::UpdateMethods, as it can check the declaring
1087 // class of the method. We may be concurrently running a GC which makes accessing
1088 // the class unsafe. We know it is OK to bypass the instrumentation as we've just
1089 // checked that the current entry point is JIT compiled code.
1090 info->GetMethod()->SetEntryPointFromQuickCompiledCode(GetQuickToInterpreterBridge());
1091 }
1092 }
1093
1094 DCHECK(CheckLiveCompiledCodeHasProfilingInfo());
1095 }
1096 live_bitmap_.reset(nullptr);
1097 NotifyCollectionDone(self);
1098 }
1099 }
1100 Runtime::Current()->GetJit()->AddTimingLogger(logger);
1101 }
1102
RemoveUnmarkedCode(Thread * self)1103 void JitCodeCache::RemoveUnmarkedCode(Thread* self) {
1104 ScopedTrace trace(__FUNCTION__);
1105 std::unordered_set<OatQuickMethodHeader*> method_headers;
1106 {
1107 MutexLock mu(self, lock_);
1108 ScopedCodeCacheWrite scc(code_map_.get());
1109 // Iterate over all compiled code and remove entries that are not marked.
1110 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
1111 const void* code_ptr = it->first;
1112 uintptr_t allocation = FromCodeToAllocation(code_ptr);
1113 if (GetLiveBitmap()->Test(allocation)) {
1114 ++it;
1115 } else {
1116 method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->first));
1117 it = method_code_map_.erase(it);
1118 }
1119 }
1120 }
1121 FreeAllMethodHeaders(method_headers);
1122 }
1123
DoCollection(Thread * self,bool collect_profiling_info)1124 void JitCodeCache::DoCollection(Thread* self, bool collect_profiling_info) {
1125 ScopedTrace trace(__FUNCTION__);
1126 {
1127 MutexLock mu(self, lock_);
1128 if (collect_profiling_info) {
1129 // Clear the profiling info of methods that do not have compiled code as entrypoint.
1130 // Also remove the saved entry point from the ProfilingInfo objects.
1131 for (ProfilingInfo* info : profiling_infos_) {
1132 const void* ptr = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1133 if (!ContainsPc(ptr) && !info->IsInUseByCompiler()) {
1134 info->GetMethod()->SetProfilingInfo(nullptr);
1135 }
1136
1137 if (info->GetSavedEntryPoint() != nullptr) {
1138 info->SetSavedEntryPoint(nullptr);
1139 // We are going to move this method back to interpreter. Clear the counter now to
1140 // give it a chance to be hot again.
1141 ClearMethodCounter(info->GetMethod(), /*was_warm*/ true);
1142 }
1143 }
1144 } else if (kIsDebugBuild) {
1145 // Sanity check that the profiling infos do not have a dangling entry point.
1146 for (ProfilingInfo* info : profiling_infos_) {
1147 DCHECK(info->GetSavedEntryPoint() == nullptr);
1148 }
1149 }
1150
1151 // Mark compiled code that are entrypoints of ArtMethods. Compiled code that is not
1152 // an entry point is either:
1153 // - an osr compiled code, that will be removed if not in a thread call stack.
1154 // - discarded compiled code, that will be removed if not in a thread call stack.
1155 for (const auto& it : method_code_map_) {
1156 ArtMethod* method = it.second;
1157 const void* code_ptr = it.first;
1158 const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1159 if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) {
1160 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
1161 }
1162 }
1163
1164 // Empty osr method map, as osr compiled code will be deleted (except the ones
1165 // on thread stacks).
1166 osr_code_map_.clear();
1167 }
1168
1169 // Run a checkpoint on all threads to mark the JIT compiled code they are running.
1170 MarkCompiledCodeOnThreadStacks(self);
1171
1172 // At this point, mutator threads are still running, and entrypoints of methods can
1173 // change. We do know they cannot change to a code cache entry that is not marked,
1174 // therefore we can safely remove those entries.
1175 RemoveUnmarkedCode(self);
1176
1177 if (collect_profiling_info) {
1178 ScopedThreadSuspension sts(self, kSuspended);
1179 MutexLock mu(self, lock_);
1180 // Free all profiling infos of methods not compiled nor being compiled.
1181 auto profiling_kept_end = std::remove_if(profiling_infos_.begin(), profiling_infos_.end(),
1182 [this] (ProfilingInfo* info) NO_THREAD_SAFETY_ANALYSIS {
1183 const void* ptr = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1184 // We have previously cleared the ProfilingInfo pointer in the ArtMethod in the hope
1185 // that the compiled code would not get revived. As mutator threads run concurrently,
1186 // they may have revived the compiled code, and now we are in the situation where
1187 // a method has compiled code but no ProfilingInfo.
1188 // We make sure compiled methods have a ProfilingInfo object. It is needed for
1189 // code cache collection.
1190 if (ContainsPc(ptr) &&
1191 info->GetMethod()->GetProfilingInfo(kRuntimePointerSize) == nullptr) {
1192 info->GetMethod()->SetProfilingInfo(info);
1193 } else if (info->GetMethod()->GetProfilingInfo(kRuntimePointerSize) != info) {
1194 // No need for this ProfilingInfo object anymore.
1195 FreeData(reinterpret_cast<uint8_t*>(info));
1196 return true;
1197 }
1198 return false;
1199 });
1200 profiling_infos_.erase(profiling_kept_end, profiling_infos_.end());
1201 DCHECK(CheckLiveCompiledCodeHasProfilingInfo());
1202 }
1203 }
1204
CheckLiveCompiledCodeHasProfilingInfo()1205 bool JitCodeCache::CheckLiveCompiledCodeHasProfilingInfo() {
1206 ScopedTrace trace(__FUNCTION__);
1207 // Check that methods we have compiled do have a ProfilingInfo object. We would
1208 // have memory leaks of compiled code otherwise.
1209 for (const auto& it : method_code_map_) {
1210 ArtMethod* method = it.second;
1211 if (method->GetProfilingInfo(kRuntimePointerSize) == nullptr) {
1212 const void* code_ptr = it.first;
1213 const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1214 if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) {
1215 // If the code is not dead, then we have a problem. Note that this can even
1216 // happen just after a collection, as mutator threads are running in parallel
1217 // and could deoptimize an existing compiled code.
1218 return false;
1219 }
1220 }
1221 }
1222 return true;
1223 }
1224
LookupMethodHeader(uintptr_t pc,ArtMethod * method)1225 OatQuickMethodHeader* JitCodeCache::LookupMethodHeader(uintptr_t pc, ArtMethod* method) {
1226 static_assert(kRuntimeISA != kThumb2, "kThumb2 cannot be a runtime ISA");
1227 if (kRuntimeISA == kArm) {
1228 // On Thumb-2, the pc is offset by one.
1229 --pc;
1230 }
1231 if (!ContainsPc(reinterpret_cast<const void*>(pc))) {
1232 return nullptr;
1233 }
1234
1235 MutexLock mu(Thread::Current(), lock_);
1236 if (method_code_map_.empty()) {
1237 return nullptr;
1238 }
1239 auto it = method_code_map_.lower_bound(reinterpret_cast<const void*>(pc));
1240 --it;
1241
1242 const void* code_ptr = it->first;
1243 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1244 if (!method_header->Contains(pc)) {
1245 return nullptr;
1246 }
1247 if (kIsDebugBuild && method != nullptr) {
1248 // When we are walking the stack to redefine classes and creating obsolete methods it is
1249 // possible that we might have updated the method_code_map by making this method obsolete in a
1250 // previous frame. Therefore we should just check that the non-obsolete version of this method
1251 // is the one we expect. We change to the non-obsolete versions in the error message since the
1252 // obsolete version of the method might not be fully initialized yet. This situation can only
1253 // occur when we are in the process of allocating and setting up obsolete methods. Otherwise
1254 // method and it->second should be identical. (See runtime/openjdkjvmti/ti_redefine.cc for more
1255 // information.)
1256 DCHECK_EQ(it->second->GetNonObsoleteMethod(), method->GetNonObsoleteMethod())
1257 << ArtMethod::PrettyMethod(method->GetNonObsoleteMethod()) << " "
1258 << ArtMethod::PrettyMethod(it->second->GetNonObsoleteMethod()) << " "
1259 << std::hex << pc;
1260 }
1261 return method_header;
1262 }
1263
LookupOsrMethodHeader(ArtMethod * method)1264 OatQuickMethodHeader* JitCodeCache::LookupOsrMethodHeader(ArtMethod* method) {
1265 MutexLock mu(Thread::Current(), lock_);
1266 auto it = osr_code_map_.find(method);
1267 if (it == osr_code_map_.end()) {
1268 return nullptr;
1269 }
1270 return OatQuickMethodHeader::FromCodePointer(it->second);
1271 }
1272
AddProfilingInfo(Thread * self,ArtMethod * method,const std::vector<uint32_t> & entries,bool retry_allocation)1273 ProfilingInfo* JitCodeCache::AddProfilingInfo(Thread* self,
1274 ArtMethod* method,
1275 const std::vector<uint32_t>& entries,
1276 bool retry_allocation)
1277 // No thread safety analysis as we are using TryLock/Unlock explicitly.
1278 NO_THREAD_SAFETY_ANALYSIS {
1279 ProfilingInfo* info = nullptr;
1280 if (!retry_allocation) {
1281 // If we are allocating for the interpreter, just try to lock, to avoid
1282 // lock contention with the JIT.
1283 if (lock_.ExclusiveTryLock(self)) {
1284 info = AddProfilingInfoInternal(self, method, entries);
1285 lock_.ExclusiveUnlock(self);
1286 }
1287 } else {
1288 {
1289 MutexLock mu(self, lock_);
1290 info = AddProfilingInfoInternal(self, method, entries);
1291 }
1292
1293 if (info == nullptr) {
1294 GarbageCollectCache(self);
1295 MutexLock mu(self, lock_);
1296 info = AddProfilingInfoInternal(self, method, entries);
1297 }
1298 }
1299 return info;
1300 }
1301
AddProfilingInfoInternal(Thread * self ATTRIBUTE_UNUSED,ArtMethod * method,const std::vector<uint32_t> & entries)1302 ProfilingInfo* JitCodeCache::AddProfilingInfoInternal(Thread* self ATTRIBUTE_UNUSED,
1303 ArtMethod* method,
1304 const std::vector<uint32_t>& entries) {
1305 size_t profile_info_size = RoundUp(
1306 sizeof(ProfilingInfo) + sizeof(InlineCache) * entries.size(),
1307 sizeof(void*));
1308
1309 // Check whether some other thread has concurrently created it.
1310 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1311 if (info != nullptr) {
1312 return info;
1313 }
1314
1315 uint8_t* data = AllocateData(profile_info_size);
1316 if (data == nullptr) {
1317 return nullptr;
1318 }
1319 info = new (data) ProfilingInfo(method, entries);
1320
1321 // Make sure other threads see the data in the profiling info object before the
1322 // store in the ArtMethod's ProfilingInfo pointer.
1323 QuasiAtomic::ThreadFenceRelease();
1324
1325 method->SetProfilingInfo(info);
1326 profiling_infos_.push_back(info);
1327 histogram_profiling_info_memory_use_.AddValue(profile_info_size);
1328 return info;
1329 }
1330
1331 // NO_THREAD_SAFETY_ANALYSIS as this is called from mspace code, at which point the lock
1332 // is already held.
MoreCore(const void * mspace,intptr_t increment)1333 void* JitCodeCache::MoreCore(const void* mspace, intptr_t increment) NO_THREAD_SAFETY_ANALYSIS {
1334 if (code_mspace_ == mspace) {
1335 size_t result = code_end_;
1336 code_end_ += increment;
1337 return reinterpret_cast<void*>(result + code_map_->Begin());
1338 } else {
1339 DCHECK_EQ(data_mspace_, mspace);
1340 size_t result = data_end_;
1341 data_end_ += increment;
1342 return reinterpret_cast<void*>(result + data_map_->Begin());
1343 }
1344 }
1345
GetProfiledMethods(const std::set<std::string> & dex_base_locations,std::vector<ProfileMethodInfo> & methods)1346 void JitCodeCache::GetProfiledMethods(const std::set<std::string>& dex_base_locations,
1347 std::vector<ProfileMethodInfo>& methods) {
1348 ScopedTrace trace(__FUNCTION__);
1349 MutexLock mu(Thread::Current(), lock_);
1350 uint16_t jit_compile_threshold = Runtime::Current()->GetJITOptions()->GetCompileThreshold();
1351 for (const ProfilingInfo* info : profiling_infos_) {
1352 ArtMethod* method = info->GetMethod();
1353 const DexFile* dex_file = method->GetDexFile();
1354 if (!ContainsElement(dex_base_locations, dex_file->GetBaseLocation())) {
1355 // Skip dex files which are not profiled.
1356 continue;
1357 }
1358 std::vector<ProfileMethodInfo::ProfileInlineCache> inline_caches;
1359
1360 // If the method didn't reach the compilation threshold don't save the inline caches.
1361 // They might be incomplete and cause unnecessary deoptimizations.
1362 // If the inline cache is empty the compiler will generate a regular invoke virtual/interface.
1363 if (method->GetCounter() < jit_compile_threshold) {
1364 methods.emplace_back(/*ProfileMethodInfo*/
1365 MethodReference(dex_file, method->GetDexMethodIndex()), inline_caches);
1366 continue;
1367 }
1368
1369 for (size_t i = 0; i < info->number_of_inline_caches_; ++i) {
1370 std::vector<TypeReference> profile_classes;
1371 const InlineCache& cache = info->cache_[i];
1372 ArtMethod* caller = info->GetMethod();
1373 bool is_missing_types = false;
1374 for (size_t k = 0; k < InlineCache::kIndividualCacheSize; k++) {
1375 mirror::Class* cls = cache.classes_[k].Read();
1376 if (cls == nullptr) {
1377 break;
1378 }
1379
1380 // Check if the receiver is in the boot class path or if it's in the
1381 // same class loader as the caller. If not, skip it, as there is not
1382 // much we can do during AOT.
1383 if (!cls->IsBootStrapClassLoaded() &&
1384 caller->GetClassLoader() != cls->GetClassLoader()) {
1385 is_missing_types = true;
1386 continue;
1387 }
1388
1389 const DexFile* class_dex_file = nullptr;
1390 dex::TypeIndex type_index;
1391
1392 if (cls->GetDexCache() == nullptr) {
1393 DCHECK(cls->IsArrayClass()) << cls->PrettyClass();
1394 // Make a best effort to find the type index in the method's dex file.
1395 // We could search all open dex files but that might turn expensive
1396 // and probably not worth it.
1397 class_dex_file = dex_file;
1398 type_index = cls->FindTypeIndexInOtherDexFile(*dex_file);
1399 } else {
1400 class_dex_file = &(cls->GetDexFile());
1401 type_index = cls->GetDexTypeIndex();
1402 }
1403 if (!type_index.IsValid()) {
1404 // Could be a proxy class or an array for which we couldn't find the type index.
1405 is_missing_types = true;
1406 continue;
1407 }
1408 if (ContainsElement(dex_base_locations, class_dex_file->GetBaseLocation())) {
1409 // Only consider classes from the same apk (including multidex).
1410 profile_classes.emplace_back(/*ProfileMethodInfo::ProfileClassReference*/
1411 class_dex_file, type_index);
1412 } else {
1413 is_missing_types = true;
1414 }
1415 }
1416 if (!profile_classes.empty()) {
1417 inline_caches.emplace_back(/*ProfileMethodInfo::ProfileInlineCache*/
1418 cache.dex_pc_, is_missing_types, profile_classes);
1419 }
1420 }
1421 methods.emplace_back(/*ProfileMethodInfo*/
1422 MethodReference(dex_file, method->GetDexMethodIndex()), inline_caches);
1423 }
1424 }
1425
GetLastUpdateTimeNs() const1426 uint64_t JitCodeCache::GetLastUpdateTimeNs() const {
1427 return last_update_time_ns_.LoadAcquire();
1428 }
1429
IsOsrCompiled(ArtMethod * method)1430 bool JitCodeCache::IsOsrCompiled(ArtMethod* method) {
1431 MutexLock mu(Thread::Current(), lock_);
1432 return osr_code_map_.find(method) != osr_code_map_.end();
1433 }
1434
NotifyCompilationOf(ArtMethod * method,Thread * self,bool osr)1435 bool JitCodeCache::NotifyCompilationOf(ArtMethod* method, Thread* self, bool osr) {
1436 if (!osr && ContainsPc(method->GetEntryPointFromQuickCompiledCode())) {
1437 return false;
1438 }
1439
1440 MutexLock mu(self, lock_);
1441 if (osr && (osr_code_map_.find(method) != osr_code_map_.end())) {
1442 return false;
1443 }
1444
1445 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1446 if (info == nullptr) {
1447 VLOG(jit) << method->PrettyMethod() << " needs a ProfilingInfo to be compiled";
1448 // Because the counter is not atomic, there are some rare cases where we may not hit the
1449 // threshold for creating the ProfilingInfo. Reset the counter now to "correct" this.
1450 ClearMethodCounter(method, /*was_warm*/ false);
1451 return false;
1452 }
1453
1454 if (info->IsMethodBeingCompiled(osr)) {
1455 return false;
1456 }
1457
1458 info->SetIsMethodBeingCompiled(true, osr);
1459 return true;
1460 }
1461
NotifyCompilerUse(ArtMethod * method,Thread * self)1462 ProfilingInfo* JitCodeCache::NotifyCompilerUse(ArtMethod* method, Thread* self) {
1463 MutexLock mu(self, lock_);
1464 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1465 if (info != nullptr) {
1466 if (!info->IncrementInlineUse()) {
1467 // Overflow of inlining uses, just bail.
1468 return nullptr;
1469 }
1470 }
1471 return info;
1472 }
1473
DoneCompilerUse(ArtMethod * method,Thread * self)1474 void JitCodeCache::DoneCompilerUse(ArtMethod* method, Thread* self) {
1475 MutexLock mu(self, lock_);
1476 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1477 DCHECK(info != nullptr);
1478 info->DecrementInlineUse();
1479 }
1480
DoneCompiling(ArtMethod * method,Thread * self ATTRIBUTE_UNUSED,bool osr)1481 void JitCodeCache::DoneCompiling(ArtMethod* method, Thread* self ATTRIBUTE_UNUSED, bool osr) {
1482 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1483 DCHECK(info->IsMethodBeingCompiled(osr));
1484 info->SetIsMethodBeingCompiled(false, osr);
1485 }
1486
GetMemorySizeOfCodePointer(const void * ptr)1487 size_t JitCodeCache::GetMemorySizeOfCodePointer(const void* ptr) {
1488 MutexLock mu(Thread::Current(), lock_);
1489 return mspace_usable_size(reinterpret_cast<const void*>(FromCodeToAllocation(ptr)));
1490 }
1491
InvalidateCompiledCodeFor(ArtMethod * method,const OatQuickMethodHeader * header)1492 void JitCodeCache::InvalidateCompiledCodeFor(ArtMethod* method,
1493 const OatQuickMethodHeader* header) {
1494 ProfilingInfo* profiling_info = method->GetProfilingInfo(kRuntimePointerSize);
1495 if ((profiling_info != nullptr) &&
1496 (profiling_info->GetSavedEntryPoint() == header->GetEntryPoint())) {
1497 // Prevent future uses of the compiled code.
1498 profiling_info->SetSavedEntryPoint(nullptr);
1499 }
1500
1501 if (method->GetEntryPointFromQuickCompiledCode() == header->GetEntryPoint()) {
1502 // The entrypoint is the one to invalidate, so we just update it to the interpreter entry point
1503 // and clear the counter to get the method Jitted again.
1504 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
1505 method, GetQuickToInterpreterBridge());
1506 ClearMethodCounter(method, /*was_warm*/ profiling_info != nullptr);
1507 } else {
1508 MutexLock mu(Thread::Current(), lock_);
1509 auto it = osr_code_map_.find(method);
1510 if (it != osr_code_map_.end() && OatQuickMethodHeader::FromCodePointer(it->second) == header) {
1511 // Remove the OSR method, to avoid using it again.
1512 osr_code_map_.erase(it);
1513 }
1514 }
1515 }
1516
AllocateCode(size_t code_size)1517 uint8_t* JitCodeCache::AllocateCode(size_t code_size) {
1518 size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
1519 uint8_t* result = reinterpret_cast<uint8_t*>(
1520 mspace_memalign(code_mspace_, alignment, code_size));
1521 size_t header_size = RoundUp(sizeof(OatQuickMethodHeader), alignment);
1522 // Ensure the header ends up at expected instruction alignment.
1523 DCHECK_ALIGNED_PARAM(reinterpret_cast<uintptr_t>(result + header_size), alignment);
1524 used_memory_for_code_ += mspace_usable_size(result);
1525 return result;
1526 }
1527
FreeCode(uint8_t * code)1528 void JitCodeCache::FreeCode(uint8_t* code) {
1529 used_memory_for_code_ -= mspace_usable_size(code);
1530 mspace_free(code_mspace_, code);
1531 }
1532
AllocateData(size_t data_size)1533 uint8_t* JitCodeCache::AllocateData(size_t data_size) {
1534 void* result = mspace_malloc(data_mspace_, data_size);
1535 used_memory_for_data_ += mspace_usable_size(result);
1536 return reinterpret_cast<uint8_t*>(result);
1537 }
1538
FreeData(uint8_t * data)1539 void JitCodeCache::FreeData(uint8_t* data) {
1540 used_memory_for_data_ -= mspace_usable_size(data);
1541 mspace_free(data_mspace_, data);
1542 }
1543
Dump(std::ostream & os)1544 void JitCodeCache::Dump(std::ostream& os) {
1545 MutexLock mu(Thread::Current(), lock_);
1546 os << "Current JIT code cache size: " << PrettySize(used_memory_for_code_) << "\n"
1547 << "Current JIT data cache size: " << PrettySize(used_memory_for_data_) << "\n"
1548 << "Current JIT capacity: " << PrettySize(current_capacity_) << "\n"
1549 << "Current number of JIT code cache entries: " << method_code_map_.size() << "\n"
1550 << "Total number of JIT compilations: " << number_of_compilations_ << "\n"
1551 << "Total number of JIT compilations for on stack replacement: "
1552 << number_of_osr_compilations_ << "\n"
1553 << "Total number of JIT code cache collections: " << number_of_collections_ << std::endl;
1554 histogram_stack_map_memory_use_.PrintMemoryUse(os);
1555 histogram_code_memory_use_.PrintMemoryUse(os);
1556 histogram_profiling_info_memory_use_.PrintMemoryUse(os);
1557 }
1558
1559 } // namespace jit
1560 } // namespace art
1561