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 "art_method-inl.h"
22 #include "base/enums.h"
23 #include "base/stl_util.h"
24 #include "base/systrace.h"
25 #include "base/time_utils.h"
26 #include "cha.h"
27 #include "debugger_interface.h"
28 #include "entrypoints/runtime_asm_entrypoints.h"
29 #include "gc/accounting/bitmap-inl.h"
30 #include "gc/scoped_gc_critical_section.h"
31 #include "jit/jit.h"
32 #include "jit/profiling_info.h"
33 #include "linear_alloc.h"
34 #include "mem_map.h"
35 #include "oat_file-inl.h"
36 #include "scoped_thread_state_change-inl.h"
37 #include "thread_list.h"
38
39 namespace art {
40 namespace jit {
41
42 static constexpr int kProtAll = PROT_READ | PROT_WRITE | PROT_EXEC;
43 static constexpr int kProtData = PROT_READ | PROT_WRITE;
44 static constexpr int kProtCode = PROT_READ | PROT_EXEC;
45
46 static constexpr size_t kCodeSizeLogThreshold = 50 * KB;
47 static constexpr size_t kStackMapSizeLogThreshold = 50 * KB;
48
49 #define CHECKED_MPROTECT(memory, size, prot) \
50 do { \
51 int rc = mprotect(memory, size, prot); \
52 if (UNLIKELY(rc != 0)) { \
53 errno = rc; \
54 PLOG(FATAL) << "Failed to mprotect jit code cache"; \
55 } \
56 } while (false) \
57
Create(size_t initial_capacity,size_t max_capacity,bool generate_debug_info,std::string * error_msg)58 JitCodeCache* JitCodeCache::Create(size_t initial_capacity,
59 size_t max_capacity,
60 bool generate_debug_info,
61 std::string* error_msg) {
62 ScopedTrace trace(__PRETTY_FUNCTION__);
63 CHECK_GE(max_capacity, initial_capacity);
64
65 // Generating debug information is mostly for using the 'perf' tool, which does
66 // not work with ashmem.
67 bool use_ashmem = !generate_debug_info;
68 // With 'perf', we want a 1-1 mapping between an address and a method.
69 bool garbage_collect_code = !generate_debug_info;
70
71 // We need to have 32 bit offsets from method headers in code cache which point to things
72 // in the data cache. If the maps are more than 4G apart, having multiple maps wouldn't work.
73 // Ensure we're below 1 GB to be safe.
74 if (max_capacity > 1 * GB) {
75 std::ostringstream oss;
76 oss << "Maxium code cache capacity is limited to 1 GB, "
77 << PrettySize(max_capacity) << " is too big";
78 *error_msg = oss.str();
79 return nullptr;
80 }
81
82 std::string error_str;
83 // Map name specific for android_os_Debug.cpp accounting.
84 // Map in low 4gb to simplify accessing root tables for x86_64.
85 // We could do PC-relative addressing to avoid this problem, but that
86 // would require reserving code and data area before submitting, which
87 // means more windows for the code memory to be RWX.
88 MemMap* data_map = MemMap::MapAnonymous(
89 "data-code-cache", nullptr,
90 max_capacity,
91 kProtAll,
92 /* low_4gb */ true,
93 /* reuse */ false,
94 &error_str,
95 use_ashmem);
96 if (data_map == nullptr) {
97 std::ostringstream oss;
98 oss << "Failed to create read write execute cache: " << error_str << " size=" << max_capacity;
99 *error_msg = oss.str();
100 return nullptr;
101 }
102
103 // Align both capacities to page size, as that's the unit mspaces use.
104 initial_capacity = RoundDown(initial_capacity, 2 * kPageSize);
105 max_capacity = RoundDown(max_capacity, 2 * kPageSize);
106
107 // Data cache is 1 / 2 of the map.
108 // TODO: Make this variable?
109 size_t data_size = max_capacity / 2;
110 size_t code_size = max_capacity - data_size;
111 DCHECK_EQ(code_size + data_size, max_capacity);
112 uint8_t* divider = data_map->Begin() + data_size;
113
114 MemMap* code_map =
115 data_map->RemapAtEnd(divider, "jit-code-cache", kProtAll, &error_str, use_ashmem);
116 if (code_map == nullptr) {
117 std::ostringstream oss;
118 oss << "Failed to create read write execute cache: " << error_str << " size=" << max_capacity;
119 *error_msg = oss.str();
120 return nullptr;
121 }
122 DCHECK_EQ(code_map->Begin(), divider);
123 data_size = initial_capacity / 2;
124 code_size = initial_capacity - data_size;
125 DCHECK_EQ(code_size + data_size, initial_capacity);
126 return new JitCodeCache(
127 code_map, data_map, code_size, data_size, max_capacity, garbage_collect_code);
128 }
129
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)130 JitCodeCache::JitCodeCache(MemMap* code_map,
131 MemMap* data_map,
132 size_t initial_code_capacity,
133 size_t initial_data_capacity,
134 size_t max_capacity,
135 bool garbage_collect_code)
136 : lock_("Jit code cache", kJitCodeCacheLock),
137 lock_cond_("Jit code cache condition variable", lock_),
138 collection_in_progress_(false),
139 code_map_(code_map),
140 data_map_(data_map),
141 max_capacity_(max_capacity),
142 current_capacity_(initial_code_capacity + initial_data_capacity),
143 code_end_(initial_code_capacity),
144 data_end_(initial_data_capacity),
145 last_collection_increased_code_cache_(false),
146 last_update_time_ns_(0),
147 garbage_collect_code_(garbage_collect_code),
148 used_memory_for_data_(0),
149 used_memory_for_code_(0),
150 number_of_compilations_(0),
151 number_of_osr_compilations_(0),
152 number_of_collections_(0),
153 histogram_stack_map_memory_use_("Memory used for stack maps", 16),
154 histogram_code_memory_use_("Memory used for compiled code", 16),
155 histogram_profiling_info_memory_use_("Memory used for profiling info", 16),
156 is_weak_access_enabled_(true),
157 inline_cache_cond_("Jit inline cache condition variable", lock_) {
158
159 DCHECK_GE(max_capacity, initial_code_capacity + initial_data_capacity);
160 code_mspace_ = create_mspace_with_base(code_map_->Begin(), code_end_, false /*locked*/);
161 data_mspace_ = create_mspace_with_base(data_map_->Begin(), data_end_, false /*locked*/);
162
163 if (code_mspace_ == nullptr || data_mspace_ == nullptr) {
164 PLOG(FATAL) << "create_mspace_with_base failed";
165 }
166
167 SetFootprintLimit(current_capacity_);
168
169 CHECKED_MPROTECT(code_map_->Begin(), code_map_->Size(), kProtCode);
170 CHECKED_MPROTECT(data_map_->Begin(), data_map_->Size(), kProtData);
171
172 VLOG(jit) << "Created jit code cache: initial data size="
173 << PrettySize(initial_data_capacity)
174 << ", initial code size="
175 << PrettySize(initial_code_capacity);
176 }
177
ContainsPc(const void * ptr) const178 bool JitCodeCache::ContainsPc(const void* ptr) const {
179 return code_map_->Begin() <= ptr && ptr < code_map_->End();
180 }
181
ContainsMethod(ArtMethod * method)182 bool JitCodeCache::ContainsMethod(ArtMethod* method) {
183 MutexLock mu(Thread::Current(), lock_);
184 for (auto& it : method_code_map_) {
185 if (it.second == method) {
186 return true;
187 }
188 }
189 return false;
190 }
191
192 class ScopedCodeCacheWrite : ScopedTrace {
193 public:
ScopedCodeCacheWrite(MemMap * code_map,bool only_for_tlb_shootdown=false)194 explicit ScopedCodeCacheWrite(MemMap* code_map, bool only_for_tlb_shootdown = false)
195 : ScopedTrace("ScopedCodeCacheWrite"),
196 code_map_(code_map),
197 only_for_tlb_shootdown_(only_for_tlb_shootdown) {
198 ScopedTrace trace("mprotect all");
199 CHECKED_MPROTECT(
200 code_map_->Begin(), only_for_tlb_shootdown_ ? kPageSize : code_map_->Size(), kProtAll);
201 }
~ScopedCodeCacheWrite()202 ~ScopedCodeCacheWrite() {
203 ScopedTrace trace("mprotect code");
204 CHECKED_MPROTECT(
205 code_map_->Begin(), only_for_tlb_shootdown_ ? kPageSize : code_map_->Size(), kProtCode);
206 }
207 private:
208 MemMap* const code_map_;
209
210 // If we're using ScopedCacheWrite only for TLB shootdown, we limit the scope of mprotect to
211 // one page.
212 const bool only_for_tlb_shootdown_;
213
214 DISALLOW_COPY_AND_ASSIGN(ScopedCodeCacheWrite);
215 };
216
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)217 uint8_t* JitCodeCache::CommitCode(Thread* self,
218 ArtMethod* method,
219 uint8_t* stack_map,
220 uint8_t* method_info,
221 uint8_t* roots_data,
222 size_t frame_size_in_bytes,
223 size_t core_spill_mask,
224 size_t fp_spill_mask,
225 const uint8_t* code,
226 size_t code_size,
227 size_t data_size,
228 bool osr,
229 Handle<mirror::ObjectArray<mirror::Object>> roots,
230 bool has_should_deoptimize_flag,
231 const ArenaSet<ArtMethod*>& cha_single_implementation_list) {
232 uint8_t* result = CommitCodeInternal(self,
233 method,
234 stack_map,
235 method_info,
236 roots_data,
237 frame_size_in_bytes,
238 core_spill_mask,
239 fp_spill_mask,
240 code,
241 code_size,
242 data_size,
243 osr,
244 roots,
245 has_should_deoptimize_flag,
246 cha_single_implementation_list);
247 if (result == nullptr) {
248 // Retry.
249 GarbageCollectCache(self);
250 result = CommitCodeInternal(self,
251 method,
252 stack_map,
253 method_info,
254 roots_data,
255 frame_size_in_bytes,
256 core_spill_mask,
257 fp_spill_mask,
258 code,
259 code_size,
260 data_size,
261 osr,
262 roots,
263 has_should_deoptimize_flag,
264 cha_single_implementation_list);
265 }
266 return result;
267 }
268
WaitForPotentialCollectionToComplete(Thread * self)269 bool JitCodeCache::WaitForPotentialCollectionToComplete(Thread* self) {
270 bool in_collection = false;
271 while (collection_in_progress_) {
272 in_collection = true;
273 lock_cond_.Wait(self);
274 }
275 return in_collection;
276 }
277
FromCodeToAllocation(const void * code)278 static uintptr_t FromCodeToAllocation(const void* code) {
279 size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
280 return reinterpret_cast<uintptr_t>(code) - RoundUp(sizeof(OatQuickMethodHeader), alignment);
281 }
282
ComputeRootTableSize(uint32_t number_of_roots)283 static uint32_t ComputeRootTableSize(uint32_t number_of_roots) {
284 return sizeof(uint32_t) + number_of_roots * sizeof(GcRoot<mirror::Object>);
285 }
286
GetNumberOfRoots(const uint8_t * stack_map)287 static uint32_t GetNumberOfRoots(const uint8_t* stack_map) {
288 // The length of the table is stored just before the stack map (and therefore at the end of
289 // the table itself), in order to be able to fetch it from a `stack_map` pointer.
290 return reinterpret_cast<const uint32_t*>(stack_map)[-1];
291 }
292
FillRootTableLength(uint8_t * roots_data,uint32_t length)293 static void FillRootTableLength(uint8_t* roots_data, uint32_t length) {
294 // Store the length of the table at the end. This will allow fetching it from a `stack_map`
295 // pointer.
296 reinterpret_cast<uint32_t*>(roots_data)[length] = length;
297 }
298
FromStackMapToRoots(const uint8_t * stack_map_data)299 static const uint8_t* FromStackMapToRoots(const uint8_t* stack_map_data) {
300 return stack_map_data - ComputeRootTableSize(GetNumberOfRoots(stack_map_data));
301 }
302
FillRootTable(uint8_t * roots_data,Handle<mirror::ObjectArray<mirror::Object>> roots)303 static void FillRootTable(uint8_t* roots_data, Handle<mirror::ObjectArray<mirror::Object>> roots)
304 REQUIRES_SHARED(Locks::mutator_lock_) {
305 GcRoot<mirror::Object>* gc_roots = reinterpret_cast<GcRoot<mirror::Object>*>(roots_data);
306 const uint32_t length = roots->GetLength();
307 // Put all roots in `roots_data`.
308 for (uint32_t i = 0; i < length; ++i) {
309 ObjPtr<mirror::Object> object = roots->Get(i);
310 if (kIsDebugBuild) {
311 // Ensure the string is strongly interned. b/32995596
312 if (object->IsString()) {
313 ObjPtr<mirror::String> str = reinterpret_cast<mirror::String*>(object.Ptr());
314 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
315 CHECK(class_linker->GetInternTable()->LookupStrong(Thread::Current(), str) != nullptr);
316 }
317 }
318 gc_roots[i] = GcRoot<mirror::Object>(object);
319 }
320 }
321
GetRootTable(const void * code_ptr,uint32_t * number_of_roots=nullptr)322 static uint8_t* GetRootTable(const void* code_ptr, uint32_t* number_of_roots = nullptr) {
323 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
324 uint8_t* data = method_header->GetOptimizedCodeInfoPtr();
325 uint32_t roots = GetNumberOfRoots(data);
326 if (number_of_roots != nullptr) {
327 *number_of_roots = roots;
328 }
329 return data - ComputeRootTableSize(roots);
330 }
331
332 // Use a sentinel for marking entries in the JIT table that have been cleared.
333 // This helps diagnosing in case the compiled code tries to wrongly access such
334 // entries.
335 static mirror::Class* const weak_sentinel = reinterpret_cast<mirror::Class*>(0x1);
336
337 // 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)338 static inline void ProcessWeakClass(GcRoot<mirror::Class>* root_ptr,
339 IsMarkedVisitor* visitor,
340 mirror::Class* update)
341 REQUIRES_SHARED(Locks::mutator_lock_) {
342 // This does not need a read barrier because this is called by GC.
343 mirror::Class* cls = root_ptr->Read<kWithoutReadBarrier>();
344 if (cls != nullptr && cls != weak_sentinel) {
345 DCHECK((cls->IsClass<kDefaultVerifyFlags, kWithoutReadBarrier>()));
346 // Look at the classloader of the class to know if it has been unloaded.
347 // This does not need a read barrier because this is called by GC.
348 mirror::Object* class_loader =
349 cls->GetClassLoader<kDefaultVerifyFlags, kWithoutReadBarrier>();
350 if (class_loader == nullptr || visitor->IsMarked(class_loader) != nullptr) {
351 // The class loader is live, update the entry if the class has moved.
352 mirror::Class* new_cls = down_cast<mirror::Class*>(visitor->IsMarked(cls));
353 // Note that new_object can be null for CMS and newly allocated objects.
354 if (new_cls != nullptr && new_cls != cls) {
355 *root_ptr = GcRoot<mirror::Class>(new_cls);
356 }
357 } else {
358 // The class loader is not live, clear the entry.
359 *root_ptr = GcRoot<mirror::Class>(update);
360 }
361 }
362 }
363
SweepRootTables(IsMarkedVisitor * visitor)364 void JitCodeCache::SweepRootTables(IsMarkedVisitor* visitor) {
365 MutexLock mu(Thread::Current(), lock_);
366 for (const auto& entry : method_code_map_) {
367 uint32_t number_of_roots = 0;
368 uint8_t* roots_data = GetRootTable(entry.first, &number_of_roots);
369 GcRoot<mirror::Object>* roots = reinterpret_cast<GcRoot<mirror::Object>*>(roots_data);
370 for (uint32_t i = 0; i < number_of_roots; ++i) {
371 // This does not need a read barrier because this is called by GC.
372 mirror::Object* object = roots[i].Read<kWithoutReadBarrier>();
373 if (object == nullptr || object == weak_sentinel) {
374 // entry got deleted in a previous sweep.
375 } else if (object->IsString<kDefaultVerifyFlags, kWithoutReadBarrier>()) {
376 mirror::Object* new_object = visitor->IsMarked(object);
377 // We know the string is marked because it's a strongly-interned string that
378 // is always alive. The IsMarked implementation of the CMS collector returns
379 // null for newly allocated objects, but we know those haven't moved. Therefore,
380 // only update the entry if we get a different non-null string.
381 // TODO: Do not use IsMarked for j.l.Class, and adjust once we move this method
382 // out of the weak access/creation pause. b/32167580
383 if (new_object != nullptr && new_object != object) {
384 DCHECK(new_object->IsString());
385 roots[i] = GcRoot<mirror::Object>(new_object);
386 }
387 } else {
388 ProcessWeakClass(
389 reinterpret_cast<GcRoot<mirror::Class>*>(&roots[i]), visitor, weak_sentinel);
390 }
391 }
392 }
393 // Walk over inline caches to clear entries containing unloaded classes.
394 for (ProfilingInfo* info : profiling_infos_) {
395 for (size_t i = 0; i < info->number_of_inline_caches_; ++i) {
396 InlineCache* cache = &info->cache_[i];
397 for (size_t j = 0; j < InlineCache::kIndividualCacheSize; ++j) {
398 ProcessWeakClass(&cache->classes_[j], visitor, nullptr);
399 }
400 }
401 }
402 }
403
FreeCode(const void * code_ptr)404 void JitCodeCache::FreeCode(const void* code_ptr) {
405 uintptr_t allocation = FromCodeToAllocation(code_ptr);
406 // Notify native debugger that we are about to remove the code.
407 // It does nothing if we are not using native debugger.
408 DeleteJITCodeEntryForAddress(reinterpret_cast<uintptr_t>(code_ptr));
409 FreeData(GetRootTable(code_ptr));
410 FreeCode(reinterpret_cast<uint8_t*>(allocation));
411 }
412
FreeAllMethodHeaders(const std::unordered_set<OatQuickMethodHeader * > & method_headers)413 void JitCodeCache::FreeAllMethodHeaders(
414 const std::unordered_set<OatQuickMethodHeader*>& method_headers) {
415 {
416 MutexLock mu(Thread::Current(), *Locks::cha_lock_);
417 Runtime::Current()->GetClassHierarchyAnalysis()
418 ->RemoveDependentsWithMethodHeaders(method_headers);
419 }
420
421 // We need to remove entries in method_headers from CHA dependencies
422 // first since once we do FreeCode() below, the memory can be reused
423 // so it's possible for the same method_header to start representing
424 // different compile code.
425 MutexLock mu(Thread::Current(), lock_);
426 ScopedCodeCacheWrite scc(code_map_.get());
427 for (const OatQuickMethodHeader* method_header : method_headers) {
428 FreeCode(method_header->GetCode());
429 }
430 }
431
RemoveMethodsIn(Thread * self,const LinearAlloc & alloc)432 void JitCodeCache::RemoveMethodsIn(Thread* self, const LinearAlloc& alloc) {
433 ScopedTrace trace(__PRETTY_FUNCTION__);
434 // We use a set to first collect all method_headers whose code need to be
435 // removed. We need to free the underlying code after we remove CHA dependencies
436 // for entries in this set. And it's more efficient to iterate through
437 // the CHA dependency map just once with an unordered_set.
438 std::unordered_set<OatQuickMethodHeader*> method_headers;
439 {
440 MutexLock mu(self, lock_);
441 // We do not check if a code cache GC is in progress, as this method comes
442 // with the classlinker_classes_lock_ held, and suspending ourselves could
443 // lead to a deadlock.
444 {
445 ScopedCodeCacheWrite scc(code_map_.get());
446 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
447 if (alloc.ContainsUnsafe(it->second)) {
448 method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->first));
449 it = method_code_map_.erase(it);
450 } else {
451 ++it;
452 }
453 }
454 }
455 for (auto it = osr_code_map_.begin(); it != osr_code_map_.end();) {
456 if (alloc.ContainsUnsafe(it->first)) {
457 // Note that the code has already been pushed to method_headers in the loop
458 // above and is going to be removed in FreeCode() below.
459 it = osr_code_map_.erase(it);
460 } else {
461 ++it;
462 }
463 }
464 for (auto it = profiling_infos_.begin(); it != profiling_infos_.end();) {
465 ProfilingInfo* info = *it;
466 if (alloc.ContainsUnsafe(info->GetMethod())) {
467 info->GetMethod()->SetProfilingInfo(nullptr);
468 FreeData(reinterpret_cast<uint8_t*>(info));
469 it = profiling_infos_.erase(it);
470 } else {
471 ++it;
472 }
473 }
474 }
475 FreeAllMethodHeaders(method_headers);
476 }
477
IsWeakAccessEnabled(Thread * self) const478 bool JitCodeCache::IsWeakAccessEnabled(Thread* self) const {
479 return kUseReadBarrier
480 ? self->GetWeakRefAccessEnabled()
481 : is_weak_access_enabled_.LoadSequentiallyConsistent();
482 }
483
WaitUntilInlineCacheAccessible(Thread * self)484 void JitCodeCache::WaitUntilInlineCacheAccessible(Thread* self) {
485 if (IsWeakAccessEnabled(self)) {
486 return;
487 }
488 ScopedThreadSuspension sts(self, kWaitingWeakGcRootRead);
489 MutexLock mu(self, lock_);
490 while (!IsWeakAccessEnabled(self)) {
491 inline_cache_cond_.Wait(self);
492 }
493 }
494
BroadcastForInlineCacheAccess()495 void JitCodeCache::BroadcastForInlineCacheAccess() {
496 Thread* self = Thread::Current();
497 MutexLock mu(self, lock_);
498 inline_cache_cond_.Broadcast(self);
499 }
500
AllowInlineCacheAccess()501 void JitCodeCache::AllowInlineCacheAccess() {
502 DCHECK(!kUseReadBarrier);
503 is_weak_access_enabled_.StoreSequentiallyConsistent(true);
504 BroadcastForInlineCacheAccess();
505 }
506
DisallowInlineCacheAccess()507 void JitCodeCache::DisallowInlineCacheAccess() {
508 DCHECK(!kUseReadBarrier);
509 is_weak_access_enabled_.StoreSequentiallyConsistent(false);
510 }
511
CopyInlineCacheInto(const InlineCache & ic,Handle<mirror::ObjectArray<mirror::Class>> array)512 void JitCodeCache::CopyInlineCacheInto(const InlineCache& ic,
513 Handle<mirror::ObjectArray<mirror::Class>> array) {
514 WaitUntilInlineCacheAccessible(Thread::Current());
515 // Note that we don't need to lock `lock_` here, the compiler calling
516 // this method has already ensured the inline cache will not be deleted.
517 for (size_t in_cache = 0, in_array = 0;
518 in_cache < InlineCache::kIndividualCacheSize;
519 ++in_cache) {
520 mirror::Class* object = ic.classes_[in_cache].Read();
521 if (object != nullptr) {
522 array->Set(in_array++, object);
523 }
524 }
525 }
526
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)527 uint8_t* JitCodeCache::CommitCodeInternal(Thread* self,
528 ArtMethod* method,
529 uint8_t* stack_map,
530 uint8_t* method_info,
531 uint8_t* roots_data,
532 size_t frame_size_in_bytes,
533 size_t core_spill_mask,
534 size_t fp_spill_mask,
535 const uint8_t* code,
536 size_t code_size,
537 size_t data_size,
538 bool osr,
539 Handle<mirror::ObjectArray<mirror::Object>> roots,
540 bool has_should_deoptimize_flag,
541 const ArenaSet<ArtMethod*>&
542 cha_single_implementation_list) {
543 DCHECK(stack_map != nullptr);
544 size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
545 // Ensure the header ends up at expected instruction alignment.
546 size_t header_size = RoundUp(sizeof(OatQuickMethodHeader), alignment);
547 size_t total_size = header_size + code_size;
548
549 OatQuickMethodHeader* method_header = nullptr;
550 uint8_t* code_ptr = nullptr;
551 uint8_t* memory = nullptr;
552 {
553 ScopedThreadSuspension sts(self, kSuspended);
554 MutexLock mu(self, lock_);
555 WaitForPotentialCollectionToComplete(self);
556 {
557 ScopedCodeCacheWrite scc(code_map_.get());
558 memory = AllocateCode(total_size);
559 if (memory == nullptr) {
560 return nullptr;
561 }
562 code_ptr = memory + header_size;
563
564 std::copy(code, code + code_size, code_ptr);
565 method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
566 new (method_header) OatQuickMethodHeader(
567 code_ptr - stack_map,
568 code_ptr - method_info,
569 frame_size_in_bytes,
570 core_spill_mask,
571 fp_spill_mask,
572 code_size);
573 // Flush caches before we remove write permission because some ARMv8 Qualcomm kernels may
574 // trigger a segfault if a page fault occurs when requesting a cache maintenance operation.
575 // This is a kernel bug that we need to work around until affected devices (e.g. Nexus 5X and
576 // 6P) stop being supported or their kernels are fixed.
577 //
578 // For reference, this behavior is caused by this commit:
579 // https://android.googlesource.com/kernel/msm/+/3fbe6bc28a6b9939d0650f2f17eb5216c719950c
580 FlushInstructionCache(reinterpret_cast<char*>(code_ptr),
581 reinterpret_cast<char*>(code_ptr + code_size));
582 DCHECK(!Runtime::Current()->IsAotCompiler());
583 if (has_should_deoptimize_flag) {
584 method_header->SetHasShouldDeoptimizeFlag();
585 }
586 }
587
588 number_of_compilations_++;
589 }
590 // We need to update the entry point in the runnable state for the instrumentation.
591 {
592 // Need cha_lock_ for checking all single-implementation flags and register
593 // dependencies.
594 MutexLock cha_mu(self, *Locks::cha_lock_);
595 bool single_impl_still_valid = true;
596 for (ArtMethod* single_impl : cha_single_implementation_list) {
597 if (!single_impl->HasSingleImplementation()) {
598 // We simply discard the compiled code. Clear the
599 // counter so that it may be recompiled later. Hopefully the
600 // class hierarchy will be more stable when compilation is retried.
601 single_impl_still_valid = false;
602 method->ClearCounter();
603 break;
604 }
605 }
606
607 // Discard the code if any single-implementation assumptions are now invalid.
608 if (!single_impl_still_valid) {
609 VLOG(jit) << "JIT discarded jitted code due to invalid single-implementation assumptions.";
610 return nullptr;
611 }
612 DCHECK(cha_single_implementation_list.empty() || !Runtime::Current()->IsJavaDebuggable())
613 << "Should not be using cha on debuggable apps/runs!";
614
615 for (ArtMethod* single_impl : cha_single_implementation_list) {
616 Runtime::Current()->GetClassHierarchyAnalysis()->AddDependency(
617 single_impl, method, method_header);
618 }
619
620 // The following needs to be guarded by cha_lock_ also. Otherwise it's
621 // possible that the compiled code is considered invalidated by some class linking,
622 // but below we still make the compiled code valid for the method.
623 MutexLock mu(self, lock_);
624 // Fill the root table before updating the entry point.
625 DCHECK_EQ(FromStackMapToRoots(stack_map), roots_data);
626 DCHECK_LE(roots_data, stack_map);
627 FillRootTable(roots_data, roots);
628 {
629 // Flush data cache, as compiled code references literals in it.
630 // We also need a TLB shootdown to act as memory barrier across cores.
631 ScopedCodeCacheWrite ccw(code_map_.get(), /* only_for_tlb_shootdown */ true);
632 FlushDataCache(reinterpret_cast<char*>(roots_data),
633 reinterpret_cast<char*>(roots_data + data_size));
634 }
635 method_code_map_.Put(code_ptr, method);
636 if (osr) {
637 number_of_osr_compilations_++;
638 osr_code_map_.Put(method, code_ptr);
639 } else {
640 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
641 method, method_header->GetEntryPoint());
642 }
643 if (collection_in_progress_) {
644 // We need to update the live bitmap if there is a GC to ensure it sees this new
645 // code.
646 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
647 }
648 last_update_time_ns_.StoreRelease(NanoTime());
649 VLOG(jit)
650 << "JIT added (osr=" << std::boolalpha << osr << std::noboolalpha << ") "
651 << ArtMethod::PrettyMethod(method) << "@" << method
652 << " ccache_size=" << PrettySize(CodeCacheSizeLocked()) << ": "
653 << " dcache_size=" << PrettySize(DataCacheSizeLocked()) << ": "
654 << reinterpret_cast<const void*>(method_header->GetEntryPoint()) << ","
655 << reinterpret_cast<const void*>(method_header->GetEntryPoint() +
656 method_header->GetCodeSize());
657 histogram_code_memory_use_.AddValue(code_size);
658 if (code_size > kCodeSizeLogThreshold) {
659 LOG(INFO) << "JIT allocated "
660 << PrettySize(code_size)
661 << " for compiled code of "
662 << ArtMethod::PrettyMethod(method);
663 }
664 }
665
666 return reinterpret_cast<uint8_t*>(method_header);
667 }
668
CodeCacheSize()669 size_t JitCodeCache::CodeCacheSize() {
670 MutexLock mu(Thread::Current(), lock_);
671 return CodeCacheSizeLocked();
672 }
673
674 // This notifies the code cache that the given method has been redefined and that it should remove
675 // any cached information it has on the method. All threads must be suspended before calling this
676 // method. The compiled code for the method (if there is any) must not be in any threads call stack.
NotifyMethodRedefined(ArtMethod * method)677 void JitCodeCache::NotifyMethodRedefined(ArtMethod* method) {
678 MutexLock mu(Thread::Current(), lock_);
679 if (method->IsNative()) {
680 return;
681 }
682 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
683 if (info != nullptr) {
684 auto profile = std::find(profiling_infos_.begin(), profiling_infos_.end(), info);
685 DCHECK(profile != profiling_infos_.end());
686 profiling_infos_.erase(profile);
687 }
688 method->SetProfilingInfo(nullptr);
689 ScopedCodeCacheWrite ccw(code_map_.get());
690 for (auto code_iter = method_code_map_.begin(); code_iter != method_code_map_.end();) {
691 if (code_iter->second == method) {
692 FreeCode(code_iter->first);
693 code_iter = method_code_map_.erase(code_iter);
694 continue;
695 }
696 ++code_iter;
697 }
698 auto code_map = osr_code_map_.find(method);
699 if (code_map != osr_code_map_.end()) {
700 osr_code_map_.erase(code_map);
701 }
702 }
703
704 // This invalidates old_method. Once this function returns one can no longer use old_method to
705 // execute code unless it is fixed up. This fixup will happen later in the process of installing a
706 // class redefinition.
707 // TODO We should add some info to ArtMethod to note that 'old_method' has been invalidated and
708 // shouldn't be used since it is no longer logically in the jit code cache.
709 // TODO We should add DCHECKS that validate that the JIT is paused when this method is entered.
MoveObsoleteMethod(ArtMethod * old_method,ArtMethod * new_method)710 void JitCodeCache::MoveObsoleteMethod(ArtMethod* old_method, ArtMethod* new_method) {
711 // Native methods have no profiling info and need no special handling from the JIT code cache.
712 if (old_method->IsNative()) {
713 return;
714 }
715 MutexLock mu(Thread::Current(), lock_);
716 // Update ProfilingInfo to the new one and remove it from the old_method.
717 if (old_method->GetProfilingInfo(kRuntimePointerSize) != nullptr) {
718 DCHECK_EQ(old_method->GetProfilingInfo(kRuntimePointerSize)->GetMethod(), old_method);
719 ProfilingInfo* info = old_method->GetProfilingInfo(kRuntimePointerSize);
720 old_method->SetProfilingInfo(nullptr);
721 // Since the JIT should be paused and all threads suspended by the time this is called these
722 // checks should always pass.
723 DCHECK(!info->IsInUseByCompiler());
724 new_method->SetProfilingInfo(info);
725 info->method_ = new_method;
726 }
727 // Update method_code_map_ to point to the new method.
728 for (auto& it : method_code_map_) {
729 if (it.second == old_method) {
730 it.second = new_method;
731 }
732 }
733 // Update osr_code_map_ to point to the new method.
734 auto code_map = osr_code_map_.find(old_method);
735 if (code_map != osr_code_map_.end()) {
736 osr_code_map_.Put(new_method, code_map->second);
737 osr_code_map_.erase(old_method);
738 }
739 }
740
CodeCacheSizeLocked()741 size_t JitCodeCache::CodeCacheSizeLocked() {
742 return used_memory_for_code_;
743 }
744
DataCacheSize()745 size_t JitCodeCache::DataCacheSize() {
746 MutexLock mu(Thread::Current(), lock_);
747 return DataCacheSizeLocked();
748 }
749
DataCacheSizeLocked()750 size_t JitCodeCache::DataCacheSizeLocked() {
751 return used_memory_for_data_;
752 }
753
ClearData(Thread * self,uint8_t * stack_map_data,uint8_t * roots_data)754 void JitCodeCache::ClearData(Thread* self,
755 uint8_t* stack_map_data,
756 uint8_t* roots_data) {
757 DCHECK_EQ(FromStackMapToRoots(stack_map_data), roots_data);
758 MutexLock mu(self, lock_);
759 FreeData(reinterpret_cast<uint8_t*>(roots_data));
760 }
761
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)762 size_t JitCodeCache::ReserveData(Thread* self,
763 size_t stack_map_size,
764 size_t method_info_size,
765 size_t number_of_roots,
766 ArtMethod* method,
767 uint8_t** stack_map_data,
768 uint8_t** method_info_data,
769 uint8_t** roots_data) {
770 size_t table_size = ComputeRootTableSize(number_of_roots);
771 size_t size = RoundUp(stack_map_size + method_info_size + table_size, sizeof(void*));
772 uint8_t* result = nullptr;
773
774 {
775 ScopedThreadSuspension sts(self, kSuspended);
776 MutexLock mu(self, lock_);
777 WaitForPotentialCollectionToComplete(self);
778 result = AllocateData(size);
779 }
780
781 if (result == nullptr) {
782 // Retry.
783 GarbageCollectCache(self);
784 ScopedThreadSuspension sts(self, kSuspended);
785 MutexLock mu(self, lock_);
786 WaitForPotentialCollectionToComplete(self);
787 result = AllocateData(size);
788 }
789
790 MutexLock mu(self, lock_);
791 histogram_stack_map_memory_use_.AddValue(size);
792 if (size > kStackMapSizeLogThreshold) {
793 LOG(INFO) << "JIT allocated "
794 << PrettySize(size)
795 << " for stack maps of "
796 << ArtMethod::PrettyMethod(method);
797 }
798 if (result != nullptr) {
799 *roots_data = result;
800 *stack_map_data = result + table_size;
801 *method_info_data = *stack_map_data + stack_map_size;
802 FillRootTableLength(*roots_data, number_of_roots);
803 return size;
804 } else {
805 *roots_data = nullptr;
806 *stack_map_data = nullptr;
807 *method_info_data = nullptr;
808 return 0;
809 }
810 }
811
812 class MarkCodeVisitor FINAL : public StackVisitor {
813 public:
MarkCodeVisitor(Thread * thread_in,JitCodeCache * code_cache_in)814 MarkCodeVisitor(Thread* thread_in, JitCodeCache* code_cache_in)
815 : StackVisitor(thread_in, nullptr, StackVisitor::StackWalkKind::kSkipInlinedFrames),
816 code_cache_(code_cache_in),
817 bitmap_(code_cache_->GetLiveBitmap()) {}
818
VisitFrame()819 bool VisitFrame() OVERRIDE REQUIRES_SHARED(Locks::mutator_lock_) {
820 const OatQuickMethodHeader* method_header = GetCurrentOatQuickMethodHeader();
821 if (method_header == nullptr) {
822 return true;
823 }
824 const void* code = method_header->GetCode();
825 if (code_cache_->ContainsPc(code)) {
826 // Use the atomic set version, as multiple threads are executing this code.
827 bitmap_->AtomicTestAndSet(FromCodeToAllocation(code));
828 }
829 return true;
830 }
831
832 private:
833 JitCodeCache* const code_cache_;
834 CodeCacheBitmap* const bitmap_;
835 };
836
837 class MarkCodeClosure FINAL : public Closure {
838 public:
MarkCodeClosure(JitCodeCache * code_cache,Barrier * barrier)839 MarkCodeClosure(JitCodeCache* code_cache, Barrier* barrier)
840 : code_cache_(code_cache), barrier_(barrier) {}
841
Run(Thread * thread)842 void Run(Thread* thread) OVERRIDE REQUIRES_SHARED(Locks::mutator_lock_) {
843 ScopedTrace trace(__PRETTY_FUNCTION__);
844 DCHECK(thread == Thread::Current() || thread->IsSuspended());
845 MarkCodeVisitor visitor(thread, code_cache_);
846 visitor.WalkStack();
847 if (kIsDebugBuild) {
848 // The stack walking code queries the side instrumentation stack if it
849 // sees an instrumentation exit pc, so the JIT code of methods in that stack
850 // must have been seen. We sanity check this below.
851 for (const instrumentation::InstrumentationStackFrame& frame
852 : *thread->GetInstrumentationStack()) {
853 // The 'method_' in InstrumentationStackFrame is the one that has return_pc_ in
854 // its stack frame, it is not the method owning return_pc_. We just pass null to
855 // LookupMethodHeader: the method is only checked against in debug builds.
856 OatQuickMethodHeader* method_header =
857 code_cache_->LookupMethodHeader(frame.return_pc_, nullptr);
858 if (method_header != nullptr) {
859 const void* code = method_header->GetCode();
860 CHECK(code_cache_->GetLiveBitmap()->Test(FromCodeToAllocation(code)));
861 }
862 }
863 }
864 barrier_->Pass(Thread::Current());
865 }
866
867 private:
868 JitCodeCache* const code_cache_;
869 Barrier* const barrier_;
870 };
871
NotifyCollectionDone(Thread * self)872 void JitCodeCache::NotifyCollectionDone(Thread* self) {
873 collection_in_progress_ = false;
874 lock_cond_.Broadcast(self);
875 }
876
SetFootprintLimit(size_t new_footprint)877 void JitCodeCache::SetFootprintLimit(size_t new_footprint) {
878 size_t per_space_footprint = new_footprint / 2;
879 DCHECK(IsAlignedParam(per_space_footprint, kPageSize));
880 DCHECK_EQ(per_space_footprint * 2, new_footprint);
881 mspace_set_footprint_limit(data_mspace_, per_space_footprint);
882 {
883 ScopedCodeCacheWrite scc(code_map_.get());
884 mspace_set_footprint_limit(code_mspace_, per_space_footprint);
885 }
886 }
887
IncreaseCodeCacheCapacity()888 bool JitCodeCache::IncreaseCodeCacheCapacity() {
889 if (current_capacity_ == max_capacity_) {
890 return false;
891 }
892
893 // Double the capacity if we're below 1MB, or increase it by 1MB if
894 // we're above.
895 if (current_capacity_ < 1 * MB) {
896 current_capacity_ *= 2;
897 } else {
898 current_capacity_ += 1 * MB;
899 }
900 if (current_capacity_ > max_capacity_) {
901 current_capacity_ = max_capacity_;
902 }
903
904 if (!kIsDebugBuild || VLOG_IS_ON(jit)) {
905 LOG(INFO) << "Increasing code cache capacity to " << PrettySize(current_capacity_);
906 }
907
908 SetFootprintLimit(current_capacity_);
909
910 return true;
911 }
912
MarkCompiledCodeOnThreadStacks(Thread * self)913 void JitCodeCache::MarkCompiledCodeOnThreadStacks(Thread* self) {
914 Barrier barrier(0);
915 size_t threads_running_checkpoint = 0;
916 MarkCodeClosure closure(this, &barrier);
917 threads_running_checkpoint = Runtime::Current()->GetThreadList()->RunCheckpoint(&closure);
918 // Now that we have run our checkpoint, move to a suspended state and wait
919 // for other threads to run the checkpoint.
920 ScopedThreadSuspension sts(self, kSuspended);
921 if (threads_running_checkpoint != 0) {
922 barrier.Increment(self, threads_running_checkpoint);
923 }
924 }
925
ShouldDoFullCollection()926 bool JitCodeCache::ShouldDoFullCollection() {
927 if (current_capacity_ == max_capacity_) {
928 // Always do a full collection when the code cache is full.
929 return true;
930 } else if (current_capacity_ < kReservedCapacity) {
931 // Always do partial collection when the code cache size is below the reserved
932 // capacity.
933 return false;
934 } else if (last_collection_increased_code_cache_) {
935 // This time do a full collection.
936 return true;
937 } else {
938 // This time do a partial collection.
939 return false;
940 }
941 }
942
GarbageCollectCache(Thread * self)943 void JitCodeCache::GarbageCollectCache(Thread* self) {
944 ScopedTrace trace(__FUNCTION__);
945 if (!garbage_collect_code_) {
946 MutexLock mu(self, lock_);
947 IncreaseCodeCacheCapacity();
948 return;
949 }
950
951 // Wait for an existing collection, or let everyone know we are starting one.
952 {
953 ScopedThreadSuspension sts(self, kSuspended);
954 MutexLock mu(self, lock_);
955 if (WaitForPotentialCollectionToComplete(self)) {
956 return;
957 } else {
958 number_of_collections_++;
959 live_bitmap_.reset(CodeCacheBitmap::Create(
960 "code-cache-bitmap",
961 reinterpret_cast<uintptr_t>(code_map_->Begin()),
962 reinterpret_cast<uintptr_t>(code_map_->Begin() + current_capacity_ / 2)));
963 collection_in_progress_ = true;
964 }
965 }
966
967 TimingLogger logger("JIT code cache timing logger", true, VLOG_IS_ON(jit));
968 {
969 TimingLogger::ScopedTiming st("Code cache collection", &logger);
970
971 bool do_full_collection = false;
972 {
973 MutexLock mu(self, lock_);
974 do_full_collection = ShouldDoFullCollection();
975 }
976
977 if (!kIsDebugBuild || VLOG_IS_ON(jit)) {
978 LOG(INFO) << "Do "
979 << (do_full_collection ? "full" : "partial")
980 << " code cache collection, code="
981 << PrettySize(CodeCacheSize())
982 << ", data=" << PrettySize(DataCacheSize());
983 }
984
985 DoCollection(self, /* collect_profiling_info */ do_full_collection);
986
987 if (!kIsDebugBuild || VLOG_IS_ON(jit)) {
988 LOG(INFO) << "After code cache collection, code="
989 << PrettySize(CodeCacheSize())
990 << ", data=" << PrettySize(DataCacheSize());
991 }
992
993 {
994 MutexLock mu(self, lock_);
995
996 // Increase the code cache only when we do partial collections.
997 // TODO: base this strategy on how full the code cache is?
998 if (do_full_collection) {
999 last_collection_increased_code_cache_ = false;
1000 } else {
1001 last_collection_increased_code_cache_ = true;
1002 IncreaseCodeCacheCapacity();
1003 }
1004
1005 bool next_collection_will_be_full = ShouldDoFullCollection();
1006
1007 // Start polling the liveness of compiled code to prepare for the next full collection.
1008 if (next_collection_will_be_full) {
1009 // Save the entry point of methods we have compiled, and update the entry
1010 // point of those methods to the interpreter. If the method is invoked, the
1011 // interpreter will update its entry point to the compiled code and call it.
1012 for (ProfilingInfo* info : profiling_infos_) {
1013 const void* entry_point = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1014 if (ContainsPc(entry_point)) {
1015 info->SetSavedEntryPoint(entry_point);
1016 // Don't call Instrumentation::UpdateMethods, as it can check the declaring
1017 // class of the method. We may be concurrently running a GC which makes accessing
1018 // the class unsafe. We know it is OK to bypass the instrumentation as we've just
1019 // checked that the current entry point is JIT compiled code.
1020 info->GetMethod()->SetEntryPointFromQuickCompiledCode(GetQuickToInterpreterBridge());
1021 }
1022 }
1023
1024 DCHECK(CheckLiveCompiledCodeHasProfilingInfo());
1025 }
1026 live_bitmap_.reset(nullptr);
1027 NotifyCollectionDone(self);
1028 }
1029 }
1030 Runtime::Current()->GetJit()->AddTimingLogger(logger);
1031 }
1032
RemoveUnmarkedCode(Thread * self)1033 void JitCodeCache::RemoveUnmarkedCode(Thread* self) {
1034 ScopedTrace trace(__FUNCTION__);
1035 std::unordered_set<OatQuickMethodHeader*> method_headers;
1036 {
1037 MutexLock mu(self, lock_);
1038 ScopedCodeCacheWrite scc(code_map_.get());
1039 // Iterate over all compiled code and remove entries that are not marked.
1040 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
1041 const void* code_ptr = it->first;
1042 uintptr_t allocation = FromCodeToAllocation(code_ptr);
1043 if (GetLiveBitmap()->Test(allocation)) {
1044 ++it;
1045 } else {
1046 method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->first));
1047 it = method_code_map_.erase(it);
1048 }
1049 }
1050 }
1051 FreeAllMethodHeaders(method_headers);
1052 }
1053
DoCollection(Thread * self,bool collect_profiling_info)1054 void JitCodeCache::DoCollection(Thread* self, bool collect_profiling_info) {
1055 ScopedTrace trace(__FUNCTION__);
1056 {
1057 MutexLock mu(self, lock_);
1058 if (collect_profiling_info) {
1059 // Clear the profiling info of methods that do not have compiled code as entrypoint.
1060 // Also remove the saved entry point from the ProfilingInfo objects.
1061 for (ProfilingInfo* info : profiling_infos_) {
1062 const void* ptr = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1063 if (!ContainsPc(ptr) && !info->IsInUseByCompiler()) {
1064 info->GetMethod()->SetProfilingInfo(nullptr);
1065 }
1066
1067 if (info->GetSavedEntryPoint() != nullptr) {
1068 info->SetSavedEntryPoint(nullptr);
1069 // We are going to move this method back to interpreter. Clear the counter now to
1070 // give it a chance to be hot again.
1071 info->GetMethod()->ClearCounter();
1072 }
1073 }
1074 } else if (kIsDebugBuild) {
1075 // Sanity check that the profiling infos do not have a dangling entry point.
1076 for (ProfilingInfo* info : profiling_infos_) {
1077 DCHECK(info->GetSavedEntryPoint() == nullptr);
1078 }
1079 }
1080
1081 // Mark compiled code that are entrypoints of ArtMethods. Compiled code that is not
1082 // an entry point is either:
1083 // - an osr compiled code, that will be removed if not in a thread call stack.
1084 // - discarded compiled code, that will be removed if not in a thread call stack.
1085 for (const auto& it : method_code_map_) {
1086 ArtMethod* method = it.second;
1087 const void* code_ptr = it.first;
1088 const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1089 if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) {
1090 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
1091 }
1092 }
1093
1094 // Empty osr method map, as osr compiled code will be deleted (except the ones
1095 // on thread stacks).
1096 osr_code_map_.clear();
1097 }
1098
1099 // Run a checkpoint on all threads to mark the JIT compiled code they are running.
1100 MarkCompiledCodeOnThreadStacks(self);
1101
1102 // At this point, mutator threads are still running, and entrypoints of methods can
1103 // change. We do know they cannot change to a code cache entry that is not marked,
1104 // therefore we can safely remove those entries.
1105 RemoveUnmarkedCode(self);
1106
1107 if (collect_profiling_info) {
1108 ScopedThreadSuspension sts(self, kSuspended);
1109 MutexLock mu(self, lock_);
1110 // Free all profiling infos of methods not compiled nor being compiled.
1111 auto profiling_kept_end = std::remove_if(profiling_infos_.begin(), profiling_infos_.end(),
1112 [this] (ProfilingInfo* info) NO_THREAD_SAFETY_ANALYSIS {
1113 const void* ptr = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1114 // We have previously cleared the ProfilingInfo pointer in the ArtMethod in the hope
1115 // that the compiled code would not get revived. As mutator threads run concurrently,
1116 // they may have revived the compiled code, and now we are in the situation where
1117 // a method has compiled code but no ProfilingInfo.
1118 // We make sure compiled methods have a ProfilingInfo object. It is needed for
1119 // code cache collection.
1120 if (ContainsPc(ptr) &&
1121 info->GetMethod()->GetProfilingInfo(kRuntimePointerSize) == nullptr) {
1122 info->GetMethod()->SetProfilingInfo(info);
1123 } else if (info->GetMethod()->GetProfilingInfo(kRuntimePointerSize) != info) {
1124 // No need for this ProfilingInfo object anymore.
1125 FreeData(reinterpret_cast<uint8_t*>(info));
1126 return true;
1127 }
1128 return false;
1129 });
1130 profiling_infos_.erase(profiling_kept_end, profiling_infos_.end());
1131 DCHECK(CheckLiveCompiledCodeHasProfilingInfo());
1132 }
1133 }
1134
CheckLiveCompiledCodeHasProfilingInfo()1135 bool JitCodeCache::CheckLiveCompiledCodeHasProfilingInfo() {
1136 ScopedTrace trace(__FUNCTION__);
1137 // Check that methods we have compiled do have a ProfilingInfo object. We would
1138 // have memory leaks of compiled code otherwise.
1139 for (const auto& it : method_code_map_) {
1140 ArtMethod* method = it.second;
1141 if (method->GetProfilingInfo(kRuntimePointerSize) == nullptr) {
1142 const void* code_ptr = it.first;
1143 const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1144 if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) {
1145 // If the code is not dead, then we have a problem. Note that this can even
1146 // happen just after a collection, as mutator threads are running in parallel
1147 // and could deoptimize an existing compiled code.
1148 return false;
1149 }
1150 }
1151 }
1152 return true;
1153 }
1154
LookupMethodHeader(uintptr_t pc,ArtMethod * method)1155 OatQuickMethodHeader* JitCodeCache::LookupMethodHeader(uintptr_t pc, ArtMethod* method) {
1156 static_assert(kRuntimeISA != kThumb2, "kThumb2 cannot be a runtime ISA");
1157 if (kRuntimeISA == kArm) {
1158 // On Thumb-2, the pc is offset by one.
1159 --pc;
1160 }
1161 if (!ContainsPc(reinterpret_cast<const void*>(pc))) {
1162 return nullptr;
1163 }
1164
1165 MutexLock mu(Thread::Current(), lock_);
1166 if (method_code_map_.empty()) {
1167 return nullptr;
1168 }
1169 auto it = method_code_map_.lower_bound(reinterpret_cast<const void*>(pc));
1170 --it;
1171
1172 const void* code_ptr = it->first;
1173 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1174 if (!method_header->Contains(pc)) {
1175 return nullptr;
1176 }
1177 if (kIsDebugBuild && method != nullptr) {
1178 // When we are walking the stack to redefine classes and creating obsolete methods it is
1179 // possible that we might have updated the method_code_map by making this method obsolete in a
1180 // previous frame. Therefore we should just check that the non-obsolete version of this method
1181 // is the one we expect. We change to the non-obsolete versions in the error message since the
1182 // obsolete version of the method might not be fully initialized yet. This situation can only
1183 // occur when we are in the process of allocating and setting up obsolete methods. Otherwise
1184 // method and it->second should be identical. (See runtime/openjdkjvmti/ti_redefine.cc for more
1185 // information.)
1186 DCHECK_EQ(it->second->GetNonObsoleteMethod(), method->GetNonObsoleteMethod())
1187 << ArtMethod::PrettyMethod(method->GetNonObsoleteMethod()) << " "
1188 << ArtMethod::PrettyMethod(it->second->GetNonObsoleteMethod()) << " "
1189 << std::hex << pc;
1190 }
1191 return method_header;
1192 }
1193
LookupOsrMethodHeader(ArtMethod * method)1194 OatQuickMethodHeader* JitCodeCache::LookupOsrMethodHeader(ArtMethod* method) {
1195 MutexLock mu(Thread::Current(), lock_);
1196 auto it = osr_code_map_.find(method);
1197 if (it == osr_code_map_.end()) {
1198 return nullptr;
1199 }
1200 return OatQuickMethodHeader::FromCodePointer(it->second);
1201 }
1202
AddProfilingInfo(Thread * self,ArtMethod * method,const std::vector<uint32_t> & entries,bool retry_allocation)1203 ProfilingInfo* JitCodeCache::AddProfilingInfo(Thread* self,
1204 ArtMethod* method,
1205 const std::vector<uint32_t>& entries,
1206 bool retry_allocation)
1207 // No thread safety analysis as we are using TryLock/Unlock explicitly.
1208 NO_THREAD_SAFETY_ANALYSIS {
1209 ProfilingInfo* info = nullptr;
1210 if (!retry_allocation) {
1211 // If we are allocating for the interpreter, just try to lock, to avoid
1212 // lock contention with the JIT.
1213 if (lock_.ExclusiveTryLock(self)) {
1214 info = AddProfilingInfoInternal(self, method, entries);
1215 lock_.ExclusiveUnlock(self);
1216 }
1217 } else {
1218 {
1219 MutexLock mu(self, lock_);
1220 info = AddProfilingInfoInternal(self, method, entries);
1221 }
1222
1223 if (info == nullptr) {
1224 GarbageCollectCache(self);
1225 MutexLock mu(self, lock_);
1226 info = AddProfilingInfoInternal(self, method, entries);
1227 }
1228 }
1229 return info;
1230 }
1231
AddProfilingInfoInternal(Thread * self ATTRIBUTE_UNUSED,ArtMethod * method,const std::vector<uint32_t> & entries)1232 ProfilingInfo* JitCodeCache::AddProfilingInfoInternal(Thread* self ATTRIBUTE_UNUSED,
1233 ArtMethod* method,
1234 const std::vector<uint32_t>& entries) {
1235 size_t profile_info_size = RoundUp(
1236 sizeof(ProfilingInfo) + sizeof(InlineCache) * entries.size(),
1237 sizeof(void*));
1238
1239 // Check whether some other thread has concurrently created it.
1240 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1241 if (info != nullptr) {
1242 return info;
1243 }
1244
1245 uint8_t* data = AllocateData(profile_info_size);
1246 if (data == nullptr) {
1247 return nullptr;
1248 }
1249 info = new (data) ProfilingInfo(method, entries);
1250
1251 // Make sure other threads see the data in the profiling info object before the
1252 // store in the ArtMethod's ProfilingInfo pointer.
1253 QuasiAtomic::ThreadFenceRelease();
1254
1255 method->SetProfilingInfo(info);
1256 profiling_infos_.push_back(info);
1257 histogram_profiling_info_memory_use_.AddValue(profile_info_size);
1258 return info;
1259 }
1260
1261 // NO_THREAD_SAFETY_ANALYSIS as this is called from mspace code, at which point the lock
1262 // is already held.
MoreCore(const void * mspace,intptr_t increment)1263 void* JitCodeCache::MoreCore(const void* mspace, intptr_t increment) NO_THREAD_SAFETY_ANALYSIS {
1264 if (code_mspace_ == mspace) {
1265 size_t result = code_end_;
1266 code_end_ += increment;
1267 return reinterpret_cast<void*>(result + code_map_->Begin());
1268 } else {
1269 DCHECK_EQ(data_mspace_, mspace);
1270 size_t result = data_end_;
1271 data_end_ += increment;
1272 return reinterpret_cast<void*>(result + data_map_->Begin());
1273 }
1274 }
1275
GetProfiledMethods(const std::set<std::string> & dex_base_locations,std::vector<ProfileMethodInfo> & methods)1276 void JitCodeCache::GetProfiledMethods(const std::set<std::string>& dex_base_locations,
1277 std::vector<ProfileMethodInfo>& methods) {
1278 ScopedTrace trace(__FUNCTION__);
1279 MutexLock mu(Thread::Current(), lock_);
1280 uint16_t jit_compile_threshold = Runtime::Current()->GetJITOptions()->GetCompileThreshold();
1281 for (const ProfilingInfo* info : profiling_infos_) {
1282 ArtMethod* method = info->GetMethod();
1283 const DexFile* dex_file = method->GetDexFile();
1284 if (!ContainsElement(dex_base_locations, dex_file->GetBaseLocation())) {
1285 // Skip dex files which are not profiled.
1286 continue;
1287 }
1288 std::vector<ProfileMethodInfo::ProfileInlineCache> inline_caches;
1289
1290 // If the method didn't reach the compilation threshold don't save the inline caches.
1291 // They might be incomplete and cause unnecessary deoptimizations.
1292 // If the inline cache is empty the compiler will generate a regular invoke virtual/interface.
1293 if (method->GetCounter() < jit_compile_threshold) {
1294 methods.emplace_back(/*ProfileMethodInfo*/
1295 dex_file, method->GetDexMethodIndex(), inline_caches);
1296 continue;
1297 }
1298
1299 for (size_t i = 0; i < info->number_of_inline_caches_; ++i) {
1300 std::vector<ProfileMethodInfo::ProfileClassReference> profile_classes;
1301 const InlineCache& cache = info->cache_[i];
1302 ArtMethod* caller = info->GetMethod();
1303 bool is_missing_types = false;
1304 for (size_t k = 0; k < InlineCache::kIndividualCacheSize; k++) {
1305 mirror::Class* cls = cache.classes_[k].Read();
1306 if (cls == nullptr) {
1307 break;
1308 }
1309
1310 // Check if the receiver is in the boot class path or if it's in the
1311 // same class loader as the caller. If not, skip it, as there is not
1312 // much we can do during AOT.
1313 if (!cls->IsBootStrapClassLoaded() &&
1314 caller->GetClassLoader() != cls->GetClassLoader()) {
1315 is_missing_types = true;
1316 continue;
1317 }
1318
1319 const DexFile* class_dex_file = nullptr;
1320 dex::TypeIndex type_index;
1321
1322 if (cls->GetDexCache() == nullptr) {
1323 DCHECK(cls->IsArrayClass()) << cls->PrettyClass();
1324 // Make a best effort to find the type index in the method's dex file.
1325 // We could search all open dex files but that might turn expensive
1326 // and probably not worth it.
1327 class_dex_file = dex_file;
1328 type_index = cls->FindTypeIndexInOtherDexFile(*dex_file);
1329 } else {
1330 class_dex_file = &(cls->GetDexFile());
1331 type_index = cls->GetDexTypeIndex();
1332 }
1333 if (!type_index.IsValid()) {
1334 // Could be a proxy class or an array for which we couldn't find the type index.
1335 is_missing_types = true;
1336 continue;
1337 }
1338 if (ContainsElement(dex_base_locations, class_dex_file->GetBaseLocation())) {
1339 // Only consider classes from the same apk (including multidex).
1340 profile_classes.emplace_back(/*ProfileMethodInfo::ProfileClassReference*/
1341 class_dex_file, type_index);
1342 } else {
1343 is_missing_types = true;
1344 }
1345 }
1346 if (!profile_classes.empty()) {
1347 inline_caches.emplace_back(/*ProfileMethodInfo::ProfileInlineCache*/
1348 cache.dex_pc_, is_missing_types, profile_classes);
1349 }
1350 }
1351 methods.emplace_back(/*ProfileMethodInfo*/
1352 dex_file, method->GetDexMethodIndex(), inline_caches);
1353 }
1354 }
1355
GetLastUpdateTimeNs() const1356 uint64_t JitCodeCache::GetLastUpdateTimeNs() const {
1357 return last_update_time_ns_.LoadAcquire();
1358 }
1359
IsOsrCompiled(ArtMethod * method)1360 bool JitCodeCache::IsOsrCompiled(ArtMethod* method) {
1361 MutexLock mu(Thread::Current(), lock_);
1362 return osr_code_map_.find(method) != osr_code_map_.end();
1363 }
1364
NotifyCompilationOf(ArtMethod * method,Thread * self,bool osr)1365 bool JitCodeCache::NotifyCompilationOf(ArtMethod* method, Thread* self, bool osr) {
1366 if (!osr && ContainsPc(method->GetEntryPointFromQuickCompiledCode())) {
1367 return false;
1368 }
1369
1370 MutexLock mu(self, lock_);
1371 if (osr && (osr_code_map_.find(method) != osr_code_map_.end())) {
1372 return false;
1373 }
1374
1375 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1376 if (info == nullptr) {
1377 VLOG(jit) << method->PrettyMethod() << " needs a ProfilingInfo to be compiled";
1378 // Because the counter is not atomic, there are some rare cases where we may not
1379 // hit the threshold for creating the ProfilingInfo. Reset the counter now to
1380 // "correct" this.
1381 method->ClearCounter();
1382 return false;
1383 }
1384
1385 if (info->IsMethodBeingCompiled(osr)) {
1386 return false;
1387 }
1388
1389 info->SetIsMethodBeingCompiled(true, osr);
1390 return true;
1391 }
1392
NotifyCompilerUse(ArtMethod * method,Thread * self)1393 ProfilingInfo* JitCodeCache::NotifyCompilerUse(ArtMethod* method, Thread* self) {
1394 MutexLock mu(self, lock_);
1395 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1396 if (info != nullptr) {
1397 if (!info->IncrementInlineUse()) {
1398 // Overflow of inlining uses, just bail.
1399 return nullptr;
1400 }
1401 }
1402 return info;
1403 }
1404
DoneCompilerUse(ArtMethod * method,Thread * self)1405 void JitCodeCache::DoneCompilerUse(ArtMethod* method, Thread* self) {
1406 MutexLock mu(self, lock_);
1407 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1408 DCHECK(info != nullptr);
1409 info->DecrementInlineUse();
1410 }
1411
DoneCompiling(ArtMethod * method,Thread * self ATTRIBUTE_UNUSED,bool osr)1412 void JitCodeCache::DoneCompiling(ArtMethod* method, Thread* self ATTRIBUTE_UNUSED, bool osr) {
1413 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1414 DCHECK(info->IsMethodBeingCompiled(osr));
1415 info->SetIsMethodBeingCompiled(false, osr);
1416 }
1417
GetMemorySizeOfCodePointer(const void * ptr)1418 size_t JitCodeCache::GetMemorySizeOfCodePointer(const void* ptr) {
1419 MutexLock mu(Thread::Current(), lock_);
1420 return mspace_usable_size(reinterpret_cast<const void*>(FromCodeToAllocation(ptr)));
1421 }
1422
InvalidateCompiledCodeFor(ArtMethod * method,const OatQuickMethodHeader * header)1423 void JitCodeCache::InvalidateCompiledCodeFor(ArtMethod* method,
1424 const OatQuickMethodHeader* header) {
1425 ProfilingInfo* profiling_info = method->GetProfilingInfo(kRuntimePointerSize);
1426 if ((profiling_info != nullptr) &&
1427 (profiling_info->GetSavedEntryPoint() == header->GetEntryPoint())) {
1428 // Prevent future uses of the compiled code.
1429 profiling_info->SetSavedEntryPoint(nullptr);
1430 }
1431
1432 if (method->GetEntryPointFromQuickCompiledCode() == header->GetEntryPoint()) {
1433 // The entrypoint is the one to invalidate, so we just update
1434 // it to the interpreter entry point and clear the counter to get the method
1435 // Jitted again.
1436 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
1437 method, GetQuickToInterpreterBridge());
1438 method->ClearCounter();
1439 } else {
1440 MutexLock mu(Thread::Current(), lock_);
1441 auto it = osr_code_map_.find(method);
1442 if (it != osr_code_map_.end() && OatQuickMethodHeader::FromCodePointer(it->second) == header) {
1443 // Remove the OSR method, to avoid using it again.
1444 osr_code_map_.erase(it);
1445 }
1446 }
1447 }
1448
AllocateCode(size_t code_size)1449 uint8_t* JitCodeCache::AllocateCode(size_t code_size) {
1450 size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
1451 uint8_t* result = reinterpret_cast<uint8_t*>(
1452 mspace_memalign(code_mspace_, alignment, code_size));
1453 size_t header_size = RoundUp(sizeof(OatQuickMethodHeader), alignment);
1454 // Ensure the header ends up at expected instruction alignment.
1455 DCHECK_ALIGNED_PARAM(reinterpret_cast<uintptr_t>(result + header_size), alignment);
1456 used_memory_for_code_ += mspace_usable_size(result);
1457 return result;
1458 }
1459
FreeCode(uint8_t * code)1460 void JitCodeCache::FreeCode(uint8_t* code) {
1461 used_memory_for_code_ -= mspace_usable_size(code);
1462 mspace_free(code_mspace_, code);
1463 }
1464
AllocateData(size_t data_size)1465 uint8_t* JitCodeCache::AllocateData(size_t data_size) {
1466 void* result = mspace_malloc(data_mspace_, data_size);
1467 used_memory_for_data_ += mspace_usable_size(result);
1468 return reinterpret_cast<uint8_t*>(result);
1469 }
1470
FreeData(uint8_t * data)1471 void JitCodeCache::FreeData(uint8_t* data) {
1472 used_memory_for_data_ -= mspace_usable_size(data);
1473 mspace_free(data_mspace_, data);
1474 }
1475
Dump(std::ostream & os)1476 void JitCodeCache::Dump(std::ostream& os) {
1477 MutexLock mu(Thread::Current(), lock_);
1478 os << "Current JIT code cache size: " << PrettySize(used_memory_for_code_) << "\n"
1479 << "Current JIT data cache size: " << PrettySize(used_memory_for_data_) << "\n"
1480 << "Current JIT capacity: " << PrettySize(current_capacity_) << "\n"
1481 << "Current number of JIT code cache entries: " << method_code_map_.size() << "\n"
1482 << "Total number of JIT compilations: " << number_of_compilations_ << "\n"
1483 << "Total number of JIT compilations for on stack replacement: "
1484 << number_of_osr_compilations_ << "\n"
1485 << "Total number of JIT code cache collections: " << number_of_collections_ << std::endl;
1486 histogram_stack_map_memory_use_.PrintMemoryUse(os);
1487 histogram_code_memory_use_.PrintMemoryUse(os);
1488 histogram_profiling_info_memory_use_.PrintMemoryUse(os);
1489 }
1490
1491 } // namespace jit
1492 } // namespace art
1493