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 <android-base/logging.h>
22
23 #include "arch/context.h"
24 #include "art_method-inl.h"
25 #include "base/enums.h"
26 #include "base/histogram-inl.h"
27 #include "base/logging.h" // For VLOG.
28 #include "base/membarrier.h"
29 #include "base/memfd.h"
30 #include "base/mem_map.h"
31 #include "base/quasi_atomic.h"
32 #include "base/stl_util.h"
33 #include "base/systrace.h"
34 #include "base/time_utils.h"
35 #include "base/utils.h"
36 #include "cha.h"
37 #include "debugger_interface.h"
38 #include "dex/dex_file_loader.h"
39 #include "dex/method_reference.h"
40 #include "entrypoints/entrypoint_utils-inl.h"
41 #include "entrypoints/runtime_asm_entrypoints.h"
42 #include "gc/accounting/bitmap-inl.h"
43 #include "gc/allocator/dlmalloc.h"
44 #include "gc/scoped_gc_critical_section.h"
45 #include "handle.h"
46 #include "instrumentation.h"
47 #include "intern_table.h"
48 #include "jit/jit.h"
49 #include "jit/profiling_info.h"
50 #include "jit/jit_scoped_code_cache_write.h"
51 #include "linear_alloc.h"
52 #include "oat_file-inl.h"
53 #include "oat_quick_method_header.h"
54 #include "object_callbacks.h"
55 #include "profile/profile_compilation_info.h"
56 #include "scoped_thread_state_change-inl.h"
57 #include "stack.h"
58 #include "thread-current-inl.h"
59 #include "thread_list.h"
60
61 namespace art {
62 namespace jit {
63
64 static constexpr size_t kCodeSizeLogThreshold = 50 * KB;
65 static constexpr size_t kStackMapSizeLogThreshold = 50 * KB;
66
67 class JitCodeCache::JniStubKey {
68 public:
REQUIRES_SHARED(Locks::mutator_lock_)69 explicit JniStubKey(ArtMethod* method) REQUIRES_SHARED(Locks::mutator_lock_)
70 : shorty_(method->GetShorty()),
71 is_static_(method->IsStatic()),
72 is_fast_native_(method->IsFastNative()),
73 is_critical_native_(method->IsCriticalNative()),
74 is_synchronized_(method->IsSynchronized()) {
75 DCHECK(!(is_fast_native_ && is_critical_native_));
76 }
77
operator <(const JniStubKey & rhs) const78 bool operator<(const JniStubKey& rhs) const {
79 if (is_static_ != rhs.is_static_) {
80 return rhs.is_static_;
81 }
82 if (is_synchronized_ != rhs.is_synchronized_) {
83 return rhs.is_synchronized_;
84 }
85 if (is_fast_native_ != rhs.is_fast_native_) {
86 return rhs.is_fast_native_;
87 }
88 if (is_critical_native_ != rhs.is_critical_native_) {
89 return rhs.is_critical_native_;
90 }
91 return strcmp(shorty_, rhs.shorty_) < 0;
92 }
93
94 // Update the shorty to point to another method's shorty. Call this function when removing
95 // the method that references the old shorty from JniCodeData and not removing the entire
96 // JniCodeData; the old shorty may become a dangling pointer when that method is unloaded.
UpdateShorty(ArtMethod * method) const97 void UpdateShorty(ArtMethod* method) const REQUIRES_SHARED(Locks::mutator_lock_) {
98 const char* shorty = method->GetShorty();
99 DCHECK_STREQ(shorty_, shorty);
100 shorty_ = shorty;
101 }
102
103 private:
104 // The shorty points to a DexFile data and may need to change
105 // to point to the same shorty in a different DexFile.
106 mutable const char* shorty_;
107
108 const bool is_static_;
109 const bool is_fast_native_;
110 const bool is_critical_native_;
111 const bool is_synchronized_;
112 };
113
114 class JitCodeCache::JniStubData {
115 public:
JniStubData()116 JniStubData() : code_(nullptr), methods_() {}
117
SetCode(const void * code)118 void SetCode(const void* code) {
119 DCHECK(code != nullptr);
120 code_ = code;
121 }
122
UpdateEntryPoints(const void * entrypoint)123 void UpdateEntryPoints(const void* entrypoint) REQUIRES_SHARED(Locks::mutator_lock_) {
124 DCHECK(IsCompiled());
125 DCHECK(entrypoint == OatQuickMethodHeader::FromCodePointer(GetCode())->GetEntryPoint());
126 instrumentation::Instrumentation* instrum = Runtime::Current()->GetInstrumentation();
127 for (ArtMethod* m : GetMethods()) {
128 // Because `m` might be in the process of being deleted:
129 // - Call the dedicated method instead of the more generic UpdateMethodsCode
130 // - Check the class status without a full read barrier; use ReadBarrier::IsMarked().
131 bool can_set_entrypoint = true;
132 if (NeedsClinitCheckBeforeCall(m)) {
133 // To avoid resurrecting an unreachable object, we must not use a full read
134 // barrier but we do not want to miss updating an entrypoint under common
135 // circumstances, i.e. during a GC the class becomes visibly initialized,
136 // the method becomes hot, we compile the thunk and want to update the
137 // entrypoint while the method's declaring class field still points to the
138 // from-space class object with the old status. Therefore we read the
139 // declaring class without a read barrier and check if it's already marked.
140 // If yes, we check the status of the to-space class object as intended.
141 // Otherwise, there is no to-space object and the from-space class object
142 // contains the most recent value of the status field; even if this races
143 // with another thread doing a read barrier and updating the status, that's
144 // no different from a race with a thread that just updates the status.
145 // Such race can happen only for the zygote method pre-compilation, as we
146 // otherwise compile only thunks for methods of visibly initialized classes.
147 ObjPtr<mirror::Class> klass = m->GetDeclaringClass<kWithoutReadBarrier>();
148 ObjPtr<mirror::Class> marked = ReadBarrier::IsMarked(klass.Ptr());
149 ObjPtr<mirror::Class> checked_klass = (marked != nullptr) ? marked : klass;
150 can_set_entrypoint = checked_klass->IsVisiblyInitialized();
151 }
152 if (can_set_entrypoint) {
153 instrum->UpdateNativeMethodsCodeToJitCode(m, entrypoint);
154 }
155 }
156 }
157
GetCode() const158 const void* GetCode() const {
159 return code_;
160 }
161
IsCompiled() const162 bool IsCompiled() const {
163 return GetCode() != nullptr;
164 }
165
AddMethod(ArtMethod * method)166 void AddMethod(ArtMethod* method) {
167 if (!ContainsElement(methods_, method)) {
168 methods_.push_back(method);
169 }
170 }
171
GetMethods() const172 const std::vector<ArtMethod*>& GetMethods() const {
173 return methods_;
174 }
175
RemoveMethodsIn(const LinearAlloc & alloc)176 void RemoveMethodsIn(const LinearAlloc& alloc) {
177 auto kept_end = std::remove_if(
178 methods_.begin(),
179 methods_.end(),
180 [&alloc](ArtMethod* method) { return alloc.ContainsUnsafe(method); });
181 methods_.erase(kept_end, methods_.end());
182 }
183
RemoveMethod(ArtMethod * method)184 bool RemoveMethod(ArtMethod* method) {
185 auto it = std::find(methods_.begin(), methods_.end(), method);
186 if (it != methods_.end()) {
187 methods_.erase(it);
188 return true;
189 } else {
190 return false;
191 }
192 }
193
MoveObsoleteMethod(ArtMethod * old_method,ArtMethod * new_method)194 void MoveObsoleteMethod(ArtMethod* old_method, ArtMethod* new_method) {
195 std::replace(methods_.begin(), methods_.end(), old_method, new_method);
196 }
197
198 private:
199 const void* code_;
200 std::vector<ArtMethod*> methods_;
201 };
202
Create(bool used_only_for_profile_data,bool rwx_memory_allowed,bool is_zygote,std::string * error_msg)203 JitCodeCache* JitCodeCache::Create(bool used_only_for_profile_data,
204 bool rwx_memory_allowed,
205 bool is_zygote,
206 std::string* error_msg) {
207 // Register for membarrier expedited sync core if JIT will be generating code.
208 if (!used_only_for_profile_data) {
209 if (art::membarrier(art::MembarrierCommand::kRegisterPrivateExpeditedSyncCore) != 0) {
210 // MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE ensures that CPU instruction pipelines are
211 // flushed and it's used when adding code to the JIT. The memory used by the new code may
212 // have just been released and, in theory, the old code could still be in a pipeline.
213 VLOG(jit) << "Kernel does not support membarrier sync-core";
214 }
215 }
216
217 size_t initial_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheInitialCapacity();
218 // Check whether the provided max capacity in options is below 1GB.
219 size_t max_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheMaxCapacity();
220 // We need to have 32 bit offsets from method headers in code cache which point to things
221 // in the data cache. If the maps are more than 4G apart, having multiple maps wouldn't work.
222 // Ensure we're below 1 GB to be safe.
223 if (max_capacity > 1 * GB) {
224 std::ostringstream oss;
225 oss << "Maxium code cache capacity is limited to 1 GB, "
226 << PrettySize(max_capacity) << " is too big";
227 *error_msg = oss.str();
228 return nullptr;
229 }
230
231 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
232 JitMemoryRegion region;
233 if (!region.Initialize(initial_capacity,
234 max_capacity,
235 rwx_memory_allowed,
236 is_zygote,
237 error_msg)) {
238 return nullptr;
239 }
240
241 std::unique_ptr<JitCodeCache> jit_code_cache(new JitCodeCache());
242 if (is_zygote) {
243 // Zygote should never collect code to share the memory with the children.
244 jit_code_cache->garbage_collect_code_ = false;
245 jit_code_cache->shared_region_ = std::move(region);
246 } else {
247 jit_code_cache->private_region_ = std::move(region);
248 }
249
250 VLOG(jit) << "Created jit code cache: initial capacity="
251 << PrettySize(initial_capacity)
252 << ", maximum capacity="
253 << PrettySize(max_capacity);
254
255 return jit_code_cache.release();
256 }
257
JitCodeCache()258 JitCodeCache::JitCodeCache()
259 : is_weak_access_enabled_(true),
260 inline_cache_cond_("Jit inline cache condition variable", *Locks::jit_lock_),
261 zygote_map_(&shared_region_),
262 lock_cond_("Jit code cache condition variable", *Locks::jit_lock_),
263 collection_in_progress_(false),
264 last_collection_increased_code_cache_(false),
265 garbage_collect_code_(true),
266 number_of_compilations_(0),
267 number_of_osr_compilations_(0),
268 number_of_collections_(0),
269 histogram_stack_map_memory_use_("Memory used for stack maps", 16),
270 histogram_code_memory_use_("Memory used for compiled code", 16),
271 histogram_profiling_info_memory_use_("Memory used for profiling info", 16) {
272 }
273
~JitCodeCache()274 JitCodeCache::~JitCodeCache() {}
275
PrivateRegionContainsPc(const void * ptr) const276 bool JitCodeCache::PrivateRegionContainsPc(const void* ptr) const {
277 return private_region_.IsInExecSpace(ptr);
278 }
279
ContainsPc(const void * ptr) const280 bool JitCodeCache::ContainsPc(const void* ptr) const {
281 return PrivateRegionContainsPc(ptr) || shared_region_.IsInExecSpace(ptr);
282 }
283
WillExecuteJitCode(ArtMethod * method)284 bool JitCodeCache::WillExecuteJitCode(ArtMethod* method) {
285 ScopedObjectAccess soa(art::Thread::Current());
286 ScopedAssertNoThreadSuspension sants(__FUNCTION__);
287 if (ContainsPc(method->GetEntryPointFromQuickCompiledCode())) {
288 return true;
289 } else if (method->GetEntryPointFromQuickCompiledCode() == GetQuickInstrumentationEntryPoint()) {
290 return FindCompiledCodeForInstrumentation(method) != nullptr;
291 }
292 return false;
293 }
294
ContainsMethod(ArtMethod * method)295 bool JitCodeCache::ContainsMethod(ArtMethod* method) {
296 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
297 if (UNLIKELY(method->IsNative())) {
298 auto it = jni_stubs_map_.find(JniStubKey(method));
299 if (it != jni_stubs_map_.end() &&
300 it->second.IsCompiled() &&
301 ContainsElement(it->second.GetMethods(), method)) {
302 return true;
303 }
304 } else {
305 for (const auto& it : method_code_map_) {
306 if (it.second == method) {
307 return true;
308 }
309 }
310 if (zygote_map_.ContainsMethod(method)) {
311 return true;
312 }
313 }
314 return false;
315 }
316
GetJniStubCode(ArtMethod * method)317 const void* JitCodeCache::GetJniStubCode(ArtMethod* method) {
318 DCHECK(method->IsNative());
319 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
320 auto it = jni_stubs_map_.find(JniStubKey(method));
321 if (it != jni_stubs_map_.end()) {
322 JniStubData& data = it->second;
323 if (data.IsCompiled() && ContainsElement(data.GetMethods(), method)) {
324 return data.GetCode();
325 }
326 }
327 return nullptr;
328 }
329
FindCompiledCodeForInstrumentation(ArtMethod * method)330 const void* JitCodeCache::FindCompiledCodeForInstrumentation(ArtMethod* method) {
331 // If jit-gc is still on we use the SavedEntryPoint field for doing that and so cannot use it to
332 // find the instrumentation entrypoint.
333 if (LIKELY(GetGarbageCollectCode())) {
334 return nullptr;
335 }
336 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
337 if (info == nullptr) {
338 return nullptr;
339 }
340 // When GC is disabled for trampoline tracing we will use SavedEntrypoint to hold the actual
341 // jit-compiled version of the method. If jit-gc is disabled for other reasons this will just be
342 // nullptr.
343 return info->GetSavedEntryPoint();
344 }
345
GetSavedEntryPointOfPreCompiledMethod(ArtMethod * method)346 const void* JitCodeCache::GetSavedEntryPointOfPreCompiledMethod(ArtMethod* method) {
347 if (method->IsPreCompiled()) {
348 const void* code_ptr = nullptr;
349 if (method->GetDeclaringClass()->GetClassLoader() == nullptr) {
350 code_ptr = zygote_map_.GetCodeFor(method);
351 } else {
352 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
353 auto it = saved_compiled_methods_map_.find(method);
354 if (it != saved_compiled_methods_map_.end()) {
355 code_ptr = it->second;
356 }
357 }
358 if (code_ptr != nullptr) {
359 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
360 return method_header->GetEntryPoint();
361 }
362 }
363 return nullptr;
364 }
365
WaitForPotentialCollectionToComplete(Thread * self)366 bool JitCodeCache::WaitForPotentialCollectionToComplete(Thread* self) {
367 bool in_collection = false;
368 while (collection_in_progress_) {
369 in_collection = true;
370 lock_cond_.Wait(self);
371 }
372 return in_collection;
373 }
374
FromCodeToAllocation(const void * code)375 static uintptr_t FromCodeToAllocation(const void* code) {
376 size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
377 return reinterpret_cast<uintptr_t>(code) - RoundUp(sizeof(OatQuickMethodHeader), alignment);
378 }
379
GetNumberOfRoots(const uint8_t * stack_map)380 static uint32_t GetNumberOfRoots(const uint8_t* stack_map) {
381 // The length of the table is stored just before the stack map (and therefore at the end of
382 // the table itself), in order to be able to fetch it from a `stack_map` pointer.
383 return reinterpret_cast<const uint32_t*>(stack_map)[-1];
384 }
385
DCheckRootsAreValid(const std::vector<Handle<mirror::Object>> & roots,bool is_shared_region)386 static void DCheckRootsAreValid(const std::vector<Handle<mirror::Object>>& roots,
387 bool is_shared_region)
388 REQUIRES(!Locks::intern_table_lock_) REQUIRES_SHARED(Locks::mutator_lock_) {
389 if (!kIsDebugBuild) {
390 return;
391 }
392 // Put all roots in `roots_data`.
393 for (Handle<mirror::Object> object : roots) {
394 // Ensure the string is strongly interned. b/32995596
395 if (object->IsString()) {
396 ObjPtr<mirror::String> str = object->AsString();
397 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
398 CHECK(class_linker->GetInternTable()->LookupStrong(Thread::Current(), str) != nullptr);
399 }
400 // Ensure that we don't put movable objects in the shared region.
401 if (is_shared_region) {
402 CHECK(!Runtime::Current()->GetHeap()->IsMovableObject(object.Get()));
403 }
404 }
405 }
406
GetRootTable(const void * code_ptr,uint32_t * number_of_roots=nullptr)407 static const uint8_t* GetRootTable(const void* code_ptr, uint32_t* number_of_roots = nullptr) {
408 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
409 uint8_t* data = method_header->GetOptimizedCodeInfoPtr();
410 uint32_t roots = GetNumberOfRoots(data);
411 if (number_of_roots != nullptr) {
412 *number_of_roots = roots;
413 }
414 return data - ComputeRootTableSize(roots);
415 }
416
SweepRootTables(IsMarkedVisitor * visitor)417 void JitCodeCache::SweepRootTables(IsMarkedVisitor* visitor) {
418 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
419 for (const auto& entry : method_code_map_) {
420 uint32_t number_of_roots = 0;
421 const uint8_t* root_table = GetRootTable(entry.first, &number_of_roots);
422 uint8_t* roots_data = private_region_.IsInDataSpace(root_table)
423 ? private_region_.GetWritableDataAddress(root_table)
424 : shared_region_.GetWritableDataAddress(root_table);
425 GcRoot<mirror::Object>* roots = reinterpret_cast<GcRoot<mirror::Object>*>(roots_data);
426 for (uint32_t i = 0; i < number_of_roots; ++i) {
427 // This does not need a read barrier because this is called by GC.
428 mirror::Object* object = roots[i].Read<kWithoutReadBarrier>();
429 if (object == nullptr || object == Runtime::GetWeakClassSentinel()) {
430 // entry got deleted in a previous sweep.
431 } else if (object->IsString<kDefaultVerifyFlags>()) {
432 mirror::Object* new_object = visitor->IsMarked(object);
433 // We know the string is marked because it's a strongly-interned string that
434 // is always alive. The IsMarked implementation of the CMS collector returns
435 // null for newly allocated objects, but we know those haven't moved. Therefore,
436 // only update the entry if we get a different non-null string.
437 // TODO: Do not use IsMarked for j.l.Class, and adjust once we move this method
438 // out of the weak access/creation pause. b/32167580
439 if (new_object != nullptr && new_object != object) {
440 DCHECK(new_object->IsString());
441 roots[i] = GcRoot<mirror::Object>(new_object);
442 }
443 } else {
444 Runtime::ProcessWeakClass(
445 reinterpret_cast<GcRoot<mirror::Class>*>(&roots[i]),
446 visitor,
447 Runtime::GetWeakClassSentinel());
448 }
449 }
450 }
451 // Walk over inline caches to clear entries containing unloaded classes.
452 for (ProfilingInfo* info : profiling_infos_) {
453 for (size_t i = 0; i < info->number_of_inline_caches_; ++i) {
454 InlineCache* cache = &info->cache_[i];
455 for (size_t j = 0; j < InlineCache::kIndividualCacheSize; ++j) {
456 Runtime::ProcessWeakClass(&cache->classes_[j], visitor, nullptr);
457 }
458 }
459 }
460 }
461
FreeCodeAndData(const void * code_ptr,bool free_debug_info)462 void JitCodeCache::FreeCodeAndData(const void* code_ptr, bool free_debug_info) {
463 if (IsInZygoteExecSpace(code_ptr)) {
464 // No need to free, this is shared memory.
465 return;
466 }
467 uintptr_t allocation = FromCodeToAllocation(code_ptr);
468 if (free_debug_info) {
469 // Remove compressed mini-debug info for the method.
470 // TODO: This is expensive, so we should always do it in the caller in bulk.
471 RemoveNativeDebugInfoForJit(ArrayRef<const void*>(&code_ptr, 1));
472 }
473 if (OatQuickMethodHeader::FromCodePointer(code_ptr)->IsOptimized()) {
474 private_region_.FreeData(GetRootTable(code_ptr));
475 } // else this is a JNI stub without any data.
476
477 private_region_.FreeCode(reinterpret_cast<uint8_t*>(allocation));
478 }
479
FreeAllMethodHeaders(const std::unordered_set<OatQuickMethodHeader * > & method_headers)480 void JitCodeCache::FreeAllMethodHeaders(
481 const std::unordered_set<OatQuickMethodHeader*>& method_headers) {
482 // We need to remove entries in method_headers from CHA dependencies
483 // first since once we do FreeCode() below, the memory can be reused
484 // so it's possible for the same method_header to start representing
485 // different compile code.
486 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
487 {
488 MutexLock mu2(Thread::Current(), *Locks::cha_lock_);
489 Runtime::Current()->GetClassLinker()->GetClassHierarchyAnalysis()
490 ->RemoveDependentsWithMethodHeaders(method_headers);
491 }
492
493 // Remove compressed mini-debug info for the methods.
494 std::vector<const void*> removed_symbols;
495 removed_symbols.reserve(method_headers.size());
496 for (const OatQuickMethodHeader* method_header : method_headers) {
497 removed_symbols.push_back(method_header->GetCode());
498 }
499 std::sort(removed_symbols.begin(), removed_symbols.end());
500 RemoveNativeDebugInfoForJit(ArrayRef<const void*>(removed_symbols));
501
502 ScopedCodeCacheWrite scc(private_region_);
503 for (const OatQuickMethodHeader* method_header : method_headers) {
504 FreeCodeAndData(method_header->GetCode(), /*free_debug_info=*/ false);
505 }
506 }
507
RemoveMethodsIn(Thread * self,const LinearAlloc & alloc)508 void JitCodeCache::RemoveMethodsIn(Thread* self, const LinearAlloc& alloc) {
509 ScopedTrace trace(__PRETTY_FUNCTION__);
510 // We use a set to first collect all method_headers whose code need to be
511 // removed. We need to free the underlying code after we remove CHA dependencies
512 // for entries in this set. And it's more efficient to iterate through
513 // the CHA dependency map just once with an unordered_set.
514 std::unordered_set<OatQuickMethodHeader*> method_headers;
515 {
516 MutexLock mu(self, *Locks::jit_lock_);
517 // We do not check if a code cache GC is in progress, as this method comes
518 // with the classlinker_classes_lock_ held, and suspending ourselves could
519 // lead to a deadlock.
520 {
521 for (auto it = jni_stubs_map_.begin(); it != jni_stubs_map_.end();) {
522 it->second.RemoveMethodsIn(alloc);
523 if (it->second.GetMethods().empty()) {
524 method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->second.GetCode()));
525 it = jni_stubs_map_.erase(it);
526 } else {
527 it->first.UpdateShorty(it->second.GetMethods().front());
528 ++it;
529 }
530 }
531 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
532 if (alloc.ContainsUnsafe(it->second)) {
533 method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->first));
534 it = method_code_map_.erase(it);
535 } else {
536 ++it;
537 }
538 }
539 }
540 for (auto it = osr_code_map_.begin(); it != osr_code_map_.end();) {
541 if (alloc.ContainsUnsafe(it->first)) {
542 // Note that the code has already been pushed to method_headers in the loop
543 // above and is going to be removed in FreeCode() below.
544 it = osr_code_map_.erase(it);
545 } else {
546 ++it;
547 }
548 }
549 for (auto it = profiling_infos_.begin(); it != profiling_infos_.end();) {
550 ProfilingInfo* info = *it;
551 if (alloc.ContainsUnsafe(info->GetMethod())) {
552 info->GetMethod()->SetProfilingInfo(nullptr);
553 private_region_.FreeWritableData(reinterpret_cast<uint8_t*>(info));
554 it = profiling_infos_.erase(it);
555 } else {
556 ++it;
557 }
558 }
559 }
560 FreeAllMethodHeaders(method_headers);
561 }
562
IsWeakAccessEnabled(Thread * self) const563 bool JitCodeCache::IsWeakAccessEnabled(Thread* self) const {
564 return kUseReadBarrier
565 ? self->GetWeakRefAccessEnabled()
566 : is_weak_access_enabled_.load(std::memory_order_seq_cst);
567 }
568
WaitUntilInlineCacheAccessible(Thread * self)569 void JitCodeCache::WaitUntilInlineCacheAccessible(Thread* self) {
570 if (IsWeakAccessEnabled(self)) {
571 return;
572 }
573 ScopedThreadSuspension sts(self, kWaitingWeakGcRootRead);
574 MutexLock mu(self, *Locks::jit_lock_);
575 while (!IsWeakAccessEnabled(self)) {
576 inline_cache_cond_.Wait(self);
577 }
578 }
579
BroadcastForInlineCacheAccess()580 void JitCodeCache::BroadcastForInlineCacheAccess() {
581 Thread* self = Thread::Current();
582 MutexLock mu(self, *Locks::jit_lock_);
583 inline_cache_cond_.Broadcast(self);
584 }
585
AllowInlineCacheAccess()586 void JitCodeCache::AllowInlineCacheAccess() {
587 DCHECK(!kUseReadBarrier);
588 is_weak_access_enabled_.store(true, std::memory_order_seq_cst);
589 BroadcastForInlineCacheAccess();
590 }
591
DisallowInlineCacheAccess()592 void JitCodeCache::DisallowInlineCacheAccess() {
593 DCHECK(!kUseReadBarrier);
594 is_weak_access_enabled_.store(false, std::memory_order_seq_cst);
595 }
596
CopyInlineCacheInto(const InlineCache & ic,Handle<mirror::ObjectArray<mirror::Class>> array)597 void JitCodeCache::CopyInlineCacheInto(const InlineCache& ic,
598 Handle<mirror::ObjectArray<mirror::Class>> array) {
599 WaitUntilInlineCacheAccessible(Thread::Current());
600 // Note that we don't need to lock `lock_` here, the compiler calling
601 // this method has already ensured the inline cache will not be deleted.
602 for (size_t in_cache = 0, in_array = 0;
603 in_cache < InlineCache::kIndividualCacheSize;
604 ++in_cache) {
605 mirror::Class* object = ic.classes_[in_cache].Read();
606 if (object != nullptr) {
607 array->Set(in_array++, object);
608 }
609 }
610 }
611
ClearMethodCounter(ArtMethod * method,bool was_warm)612 static void ClearMethodCounter(ArtMethod* method, bool was_warm)
613 REQUIRES_SHARED(Locks::mutator_lock_) {
614 if (was_warm) {
615 method->SetPreviouslyWarm();
616 }
617 // We reset the counter to 1 so that the profile knows that the method was executed at least once.
618 // This is required for layout purposes.
619 // We also need to make sure we'll pass the warmup threshold again, so we set to 0 if
620 // the warmup threshold is 1.
621 uint16_t jit_warmup_threshold = Runtime::Current()->GetJITOptions()->GetWarmupThreshold();
622 method->SetCounter(std::min(jit_warmup_threshold - 1, 1));
623 }
624
WaitForPotentialCollectionToCompleteRunnable(Thread * self)625 void JitCodeCache::WaitForPotentialCollectionToCompleteRunnable(Thread* self) {
626 while (collection_in_progress_) {
627 Locks::jit_lock_->Unlock(self);
628 {
629 ScopedThreadSuspension sts(self, kSuspended);
630 MutexLock mu(self, *Locks::jit_lock_);
631 WaitForPotentialCollectionToComplete(self);
632 }
633 Locks::jit_lock_->Lock(self);
634 }
635 }
636
Commit(Thread * self,JitMemoryRegion * region,ArtMethod * method,ArrayRef<const uint8_t> reserved_code,ArrayRef<const uint8_t> code,ArrayRef<const uint8_t> reserved_data,const std::vector<Handle<mirror::Object>> & roots,ArrayRef<const uint8_t> stack_map,bool osr,bool has_should_deoptimize_flag,const ArenaSet<ArtMethod * > & cha_single_implementation_list)637 bool JitCodeCache::Commit(Thread* self,
638 JitMemoryRegion* region,
639 ArtMethod* method,
640 ArrayRef<const uint8_t> reserved_code,
641 ArrayRef<const uint8_t> code,
642 ArrayRef<const uint8_t> reserved_data,
643 const std::vector<Handle<mirror::Object>>& roots,
644 ArrayRef<const uint8_t> stack_map,
645 bool osr,
646 bool has_should_deoptimize_flag,
647 const ArenaSet<ArtMethod*>& cha_single_implementation_list) {
648 DCHECK(!method->IsNative() || !osr);
649
650 if (!method->IsNative()) {
651 // We need to do this before grabbing the lock_ because it needs to be able to see the string
652 // InternTable. Native methods do not have roots.
653 DCheckRootsAreValid(roots, IsSharedRegion(*region));
654 }
655
656 const uint8_t* roots_data = reserved_data.data();
657 size_t root_table_size = ComputeRootTableSize(roots.size());
658 const uint8_t* stack_map_data = roots_data + root_table_size;
659
660 MutexLock mu(self, *Locks::jit_lock_);
661 // We need to make sure that there will be no jit-gcs going on and wait for any ongoing one to
662 // finish.
663 WaitForPotentialCollectionToCompleteRunnable(self);
664 const uint8_t* code_ptr = region->CommitCode(
665 reserved_code, code, stack_map_data, has_should_deoptimize_flag);
666 if (code_ptr == nullptr) {
667 return false;
668 }
669 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
670
671 // Commit roots and stack maps before updating the entry point.
672 if (!region->CommitData(reserved_data, roots, stack_map)) {
673 return false;
674 }
675
676 number_of_compilations_++;
677
678 // We need to update the entry point in the runnable state for the instrumentation.
679 {
680 // The following needs to be guarded by cha_lock_ also. Otherwise it's possible that the
681 // compiled code is considered invalidated by some class linking, but below we still make the
682 // compiled code valid for the method. Need cha_lock_ for checking all single-implementation
683 // flags and register dependencies.
684 MutexLock cha_mu(self, *Locks::cha_lock_);
685 bool single_impl_still_valid = true;
686 for (ArtMethod* single_impl : cha_single_implementation_list) {
687 if (!single_impl->HasSingleImplementation()) {
688 // Simply discard the compiled code. Clear the counter so that it may be recompiled later.
689 // Hopefully the class hierarchy will be more stable when compilation is retried.
690 single_impl_still_valid = false;
691 ClearMethodCounter(method, /*was_warm=*/ false);
692 break;
693 }
694 }
695
696 // Discard the code if any single-implementation assumptions are now invalid.
697 if (UNLIKELY(!single_impl_still_valid)) {
698 VLOG(jit) << "JIT discarded jitted code due to invalid single-implementation assumptions.";
699 return false;
700 }
701 DCHECK(cha_single_implementation_list.empty() || !Runtime::Current()->IsJavaDebuggable())
702 << "Should not be using cha on debuggable apps/runs!";
703
704 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
705 for (ArtMethod* single_impl : cha_single_implementation_list) {
706 class_linker->GetClassHierarchyAnalysis()->AddDependency(single_impl, method, method_header);
707 }
708
709 if (UNLIKELY(method->IsNative())) {
710 auto it = jni_stubs_map_.find(JniStubKey(method));
711 DCHECK(it != jni_stubs_map_.end())
712 << "Entry inserted in NotifyCompilationOf() should be alive.";
713 JniStubData* data = &it->second;
714 DCHECK(ContainsElement(data->GetMethods(), method))
715 << "Entry inserted in NotifyCompilationOf() should contain this method.";
716 data->SetCode(code_ptr);
717 data->UpdateEntryPoints(method_header->GetEntryPoint());
718 } else {
719 if (method->IsPreCompiled() && IsSharedRegion(*region)) {
720 zygote_map_.Put(code_ptr, method);
721 } else {
722 method_code_map_.Put(code_ptr, method);
723 }
724 if (osr) {
725 number_of_osr_compilations_++;
726 osr_code_map_.Put(method, code_ptr);
727 } else if (NeedsClinitCheckBeforeCall(method) &&
728 !method->GetDeclaringClass()->IsVisiblyInitialized()) {
729 // This situation currently only occurs in the jit-zygote mode.
730 DCHECK(!garbage_collect_code_);
731 DCHECK(method->IsPreCompiled());
732 // The shared region can easily be queried. For the private region, we
733 // use a side map.
734 if (!IsSharedRegion(*region)) {
735 saved_compiled_methods_map_.Put(method, code_ptr);
736 }
737 } else {
738 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
739 method, method_header->GetEntryPoint());
740 }
741 }
742 if (collection_in_progress_) {
743 // We need to update the live bitmap if there is a GC to ensure it sees this new
744 // code.
745 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
746 }
747 VLOG(jit)
748 << "JIT added (osr=" << std::boolalpha << osr << std::noboolalpha << ") "
749 << ArtMethod::PrettyMethod(method) << "@" << method
750 << " ccache_size=" << PrettySize(CodeCacheSizeLocked()) << ": "
751 << " dcache_size=" << PrettySize(DataCacheSizeLocked()) << ": "
752 << reinterpret_cast<const void*>(method_header->GetEntryPoint()) << ","
753 << reinterpret_cast<const void*>(method_header->GetEntryPoint() +
754 method_header->GetCodeSize());
755 }
756
757 return true;
758 }
759
CodeCacheSize()760 size_t JitCodeCache::CodeCacheSize() {
761 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
762 return CodeCacheSizeLocked();
763 }
764
RemoveMethod(ArtMethod * method,bool release_memory)765 bool JitCodeCache::RemoveMethod(ArtMethod* method, bool release_memory) {
766 // This function is used only for testing and only with non-native methods.
767 CHECK(!method->IsNative());
768
769 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
770
771 bool osr = osr_code_map_.find(method) != osr_code_map_.end();
772 bool in_cache = RemoveMethodLocked(method, release_memory);
773
774 if (!in_cache) {
775 return false;
776 }
777
778 method->SetCounter(0);
779 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
780 method, GetQuickToInterpreterBridge());
781 VLOG(jit)
782 << "JIT removed (osr=" << std::boolalpha << osr << std::noboolalpha << ") "
783 << ArtMethod::PrettyMethod(method) << "@" << method
784 << " ccache_size=" << PrettySize(CodeCacheSizeLocked()) << ": "
785 << " dcache_size=" << PrettySize(DataCacheSizeLocked());
786 return true;
787 }
788
RemoveMethodLocked(ArtMethod * method,bool release_memory)789 bool JitCodeCache::RemoveMethodLocked(ArtMethod* method, bool release_memory) {
790 if (LIKELY(!method->IsNative())) {
791 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
792 if (info != nullptr) {
793 RemoveElement(profiling_infos_, info);
794 }
795 method->SetProfilingInfo(nullptr);
796 }
797
798 bool in_cache = false;
799 ScopedCodeCacheWrite ccw(private_region_);
800 if (UNLIKELY(method->IsNative())) {
801 auto it = jni_stubs_map_.find(JniStubKey(method));
802 if (it != jni_stubs_map_.end() && it->second.RemoveMethod(method)) {
803 in_cache = true;
804 if (it->second.GetMethods().empty()) {
805 if (release_memory) {
806 FreeCodeAndData(it->second.GetCode());
807 }
808 jni_stubs_map_.erase(it);
809 } else {
810 it->first.UpdateShorty(it->second.GetMethods().front());
811 }
812 }
813 } else {
814 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
815 if (it->second == method) {
816 in_cache = true;
817 if (release_memory) {
818 FreeCodeAndData(it->first);
819 }
820 it = method_code_map_.erase(it);
821 } else {
822 ++it;
823 }
824 }
825
826 auto osr_it = osr_code_map_.find(method);
827 if (osr_it != osr_code_map_.end()) {
828 osr_code_map_.erase(osr_it);
829 }
830 }
831
832 return in_cache;
833 }
834
835 // This notifies the code cache that the given method has been redefined and that it should remove
836 // any cached information it has on the method. All threads must be suspended before calling this
837 // method. The compiled code for the method (if there is any) must not be in any threads call stack.
NotifyMethodRedefined(ArtMethod * method)838 void JitCodeCache::NotifyMethodRedefined(ArtMethod* method) {
839 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
840 RemoveMethodLocked(method, /* release_memory= */ true);
841 }
842
843 // This invalidates old_method. Once this function returns one can no longer use old_method to
844 // execute code unless it is fixed up. This fixup will happen later in the process of installing a
845 // class redefinition.
846 // TODO We should add some info to ArtMethod to note that 'old_method' has been invalidated and
847 // shouldn't be used since it is no longer logically in the jit code cache.
848 // TODO We should add DCHECKS that validate that the JIT is paused when this method is entered.
MoveObsoleteMethod(ArtMethod * old_method,ArtMethod * new_method)849 void JitCodeCache::MoveObsoleteMethod(ArtMethod* old_method, ArtMethod* new_method) {
850 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
851 if (old_method->IsNative()) {
852 // Update methods in jni_stubs_map_.
853 for (auto& entry : jni_stubs_map_) {
854 JniStubData& data = entry.second;
855 data.MoveObsoleteMethod(old_method, new_method);
856 }
857 return;
858 }
859 // Update ProfilingInfo to the new one and remove it from the old_method.
860 if (old_method->GetProfilingInfo(kRuntimePointerSize) != nullptr) {
861 DCHECK_EQ(old_method->GetProfilingInfo(kRuntimePointerSize)->GetMethod(), old_method);
862 ProfilingInfo* info = old_method->GetProfilingInfo(kRuntimePointerSize);
863 old_method->SetProfilingInfo(nullptr);
864 // Since the JIT should be paused and all threads suspended by the time this is called these
865 // checks should always pass.
866 DCHECK(!info->IsInUseByCompiler());
867 new_method->SetProfilingInfo(info);
868 // Get rid of the old saved entrypoint if it is there.
869 info->SetSavedEntryPoint(nullptr);
870 info->method_ = new_method;
871 }
872 // Update method_code_map_ to point to the new method.
873 for (auto& it : method_code_map_) {
874 if (it.second == old_method) {
875 it.second = new_method;
876 }
877 }
878 // Update osr_code_map_ to point to the new method.
879 auto code_map = osr_code_map_.find(old_method);
880 if (code_map != osr_code_map_.end()) {
881 osr_code_map_.Put(new_method, code_map->second);
882 osr_code_map_.erase(old_method);
883 }
884 }
885
TransitionToDebuggable()886 void JitCodeCache::TransitionToDebuggable() {
887 // Check that none of our methods have an entrypoint in the zygote exec
888 // space (this should be taken care of by
889 // ClassLinker::UpdateEntryPointsClassVisitor.
890 {
891 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
892 if (kIsDebugBuild) {
893 for (const auto& it : method_code_map_) {
894 ArtMethod* method = it.second;
895 DCHECK(!method->IsPreCompiled());
896 DCHECK(!IsInZygoteExecSpace(method->GetEntryPointFromQuickCompiledCode()));
897 }
898 }
899 // Not strictly necessary, but this map is useless now.
900 saved_compiled_methods_map_.clear();
901 }
902 if (kIsDebugBuild) {
903 for (const auto& entry : zygote_map_) {
904 ArtMethod* method = entry.method;
905 if (method != nullptr) {
906 DCHECK(!method->IsPreCompiled());
907 DCHECK(!IsInZygoteExecSpace(method->GetEntryPointFromQuickCompiledCode()));
908 }
909 }
910 }
911 }
912
CodeCacheSizeLocked()913 size_t JitCodeCache::CodeCacheSizeLocked() {
914 return GetCurrentRegion()->GetUsedMemoryForCode();
915 }
916
DataCacheSize()917 size_t JitCodeCache::DataCacheSize() {
918 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
919 return DataCacheSizeLocked();
920 }
921
DataCacheSizeLocked()922 size_t JitCodeCache::DataCacheSizeLocked() {
923 return GetCurrentRegion()->GetUsedMemoryForData();
924 }
925
Reserve(Thread * self,JitMemoryRegion * region,size_t code_size,size_t stack_map_size,size_t number_of_roots,ArtMethod * method,ArrayRef<const uint8_t> * reserved_code,ArrayRef<const uint8_t> * reserved_data)926 bool JitCodeCache::Reserve(Thread* self,
927 JitMemoryRegion* region,
928 size_t code_size,
929 size_t stack_map_size,
930 size_t number_of_roots,
931 ArtMethod* method,
932 /*out*/ArrayRef<const uint8_t>* reserved_code,
933 /*out*/ArrayRef<const uint8_t>* reserved_data) {
934 code_size = OatQuickMethodHeader::InstructionAlignedSize() + code_size;
935 size_t data_size = RoundUp(ComputeRootTableSize(number_of_roots) + stack_map_size, sizeof(void*));
936
937 const uint8_t* code;
938 const uint8_t* data;
939 // We might need to try the allocation twice (with GC in between to free up memory).
940 for (int i = 0; i < 2; i++) {
941 {
942 ScopedThreadSuspension sts(self, kSuspended);
943 MutexLock mu(self, *Locks::jit_lock_);
944 WaitForPotentialCollectionToComplete(self);
945 ScopedCodeCacheWrite ccw(*region);
946 code = region->AllocateCode(code_size);
947 data = region->AllocateData(data_size);
948 }
949 if (code == nullptr || data == nullptr) {
950 Free(self, region, code, data);
951 if (i == 0) {
952 GarbageCollectCache(self);
953 continue; // Retry after GC.
954 } else {
955 return false; // Fail.
956 }
957 }
958 break; // Success.
959 }
960 *reserved_code = ArrayRef<const uint8_t>(code, code_size);
961 *reserved_data = ArrayRef<const uint8_t>(data, data_size);
962
963 MutexLock mu(self, *Locks::jit_lock_);
964 histogram_code_memory_use_.AddValue(code_size);
965 if (code_size > kCodeSizeLogThreshold) {
966 LOG(INFO) << "JIT allocated "
967 << PrettySize(code_size)
968 << " for compiled code of "
969 << ArtMethod::PrettyMethod(method);
970 }
971 histogram_stack_map_memory_use_.AddValue(data_size);
972 if (data_size > kStackMapSizeLogThreshold) {
973 LOG(INFO) << "JIT allocated "
974 << PrettySize(data_size)
975 << " for stack maps of "
976 << ArtMethod::PrettyMethod(method);
977 }
978 return true;
979 }
980
Free(Thread * self,JitMemoryRegion * region,const uint8_t * code,const uint8_t * data)981 void JitCodeCache::Free(Thread* self,
982 JitMemoryRegion* region,
983 const uint8_t* code,
984 const uint8_t* data) {
985 MutexLock mu(self, *Locks::jit_lock_);
986 ScopedCodeCacheWrite ccw(*region);
987 if (code != nullptr) {
988 region->FreeCode(code);
989 }
990 if (data != nullptr) {
991 region->FreeData(data);
992 }
993 }
994
995 class MarkCodeClosure final : public Closure {
996 public:
MarkCodeClosure(JitCodeCache * code_cache,CodeCacheBitmap * bitmap,Barrier * barrier)997 MarkCodeClosure(JitCodeCache* code_cache, CodeCacheBitmap* bitmap, Barrier* barrier)
998 : code_cache_(code_cache), bitmap_(bitmap), barrier_(barrier) {}
999
Run(Thread * thread)1000 void Run(Thread* thread) override REQUIRES_SHARED(Locks::mutator_lock_) {
1001 ScopedTrace trace(__PRETTY_FUNCTION__);
1002 DCHECK(thread == Thread::Current() || thread->IsSuspended());
1003 StackVisitor::WalkStack(
1004 [&](const art::StackVisitor* stack_visitor) {
1005 const OatQuickMethodHeader* method_header =
1006 stack_visitor->GetCurrentOatQuickMethodHeader();
1007 if (method_header == nullptr) {
1008 return true;
1009 }
1010 const void* code = method_header->GetCode();
1011 if (code_cache_->ContainsPc(code) && !code_cache_->IsInZygoteExecSpace(code)) {
1012 // Use the atomic set version, as multiple threads are executing this code.
1013 bitmap_->AtomicTestAndSet(FromCodeToAllocation(code));
1014 }
1015 return true;
1016 },
1017 thread,
1018 /* context= */ nullptr,
1019 art::StackVisitor::StackWalkKind::kSkipInlinedFrames);
1020
1021 if (kIsDebugBuild) {
1022 // The stack walking code queries the side instrumentation stack if it
1023 // sees an instrumentation exit pc, so the JIT code of methods in that stack
1024 // must have been seen. We sanity check this below.
1025 for (const auto& it : *thread->GetInstrumentationStack()) {
1026 // The 'method_' in InstrumentationStackFrame is the one that has return_pc_ in
1027 // its stack frame, it is not the method owning return_pc_. We just pass null to
1028 // LookupMethodHeader: the method is only checked against in debug builds.
1029 OatQuickMethodHeader* method_header =
1030 code_cache_->LookupMethodHeader(it.second.return_pc_, /* method= */ nullptr);
1031 if (method_header != nullptr) {
1032 const void* code = method_header->GetCode();
1033 CHECK(bitmap_->Test(FromCodeToAllocation(code)));
1034 }
1035 }
1036 }
1037 barrier_->Pass(Thread::Current());
1038 }
1039
1040 private:
1041 JitCodeCache* const code_cache_;
1042 CodeCacheBitmap* const bitmap_;
1043 Barrier* const barrier_;
1044 };
1045
NotifyCollectionDone(Thread * self)1046 void JitCodeCache::NotifyCollectionDone(Thread* self) {
1047 collection_in_progress_ = false;
1048 lock_cond_.Broadcast(self);
1049 }
1050
MarkCompiledCodeOnThreadStacks(Thread * self)1051 void JitCodeCache::MarkCompiledCodeOnThreadStacks(Thread* self) {
1052 Barrier barrier(0);
1053 size_t threads_running_checkpoint = 0;
1054 MarkCodeClosure closure(this, GetLiveBitmap(), &barrier);
1055 threads_running_checkpoint = Runtime::Current()->GetThreadList()->RunCheckpoint(&closure);
1056 // Now that we have run our checkpoint, move to a suspended state and wait
1057 // for other threads to run the checkpoint.
1058 ScopedThreadSuspension sts(self, kSuspended);
1059 if (threads_running_checkpoint != 0) {
1060 barrier.Increment(self, threads_running_checkpoint);
1061 }
1062 }
1063
ShouldDoFullCollection()1064 bool JitCodeCache::ShouldDoFullCollection() {
1065 if (private_region_.GetCurrentCapacity() == private_region_.GetMaxCapacity()) {
1066 // Always do a full collection when the code cache is full.
1067 return true;
1068 } else if (private_region_.GetCurrentCapacity() < kReservedCapacity) {
1069 // Always do partial collection when the code cache size is below the reserved
1070 // capacity.
1071 return false;
1072 } else if (last_collection_increased_code_cache_) {
1073 // This time do a full collection.
1074 return true;
1075 } else {
1076 // This time do a partial collection.
1077 return false;
1078 }
1079 }
1080
GarbageCollectCache(Thread * self)1081 void JitCodeCache::GarbageCollectCache(Thread* self) {
1082 ScopedTrace trace(__FUNCTION__);
1083 // Wait for an existing collection, or let everyone know we are starting one.
1084 {
1085 ScopedThreadSuspension sts(self, kSuspended);
1086 MutexLock mu(self, *Locks::jit_lock_);
1087 if (!garbage_collect_code_) {
1088 private_region_.IncreaseCodeCacheCapacity();
1089 return;
1090 } else if (WaitForPotentialCollectionToComplete(self)) {
1091 return;
1092 } else {
1093 number_of_collections_++;
1094 live_bitmap_.reset(CodeCacheBitmap::Create(
1095 "code-cache-bitmap",
1096 reinterpret_cast<uintptr_t>(private_region_.GetExecPages()->Begin()),
1097 reinterpret_cast<uintptr_t>(
1098 private_region_.GetExecPages()->Begin() + private_region_.GetCurrentCapacity() / 2)));
1099 collection_in_progress_ = true;
1100 }
1101 }
1102
1103 TimingLogger logger("JIT code cache timing logger", true, VLOG_IS_ON(jit));
1104 {
1105 TimingLogger::ScopedTiming st("Code cache collection", &logger);
1106
1107 bool do_full_collection = false;
1108 {
1109 MutexLock mu(self, *Locks::jit_lock_);
1110 do_full_collection = ShouldDoFullCollection();
1111 }
1112
1113 VLOG(jit) << "Do "
1114 << (do_full_collection ? "full" : "partial")
1115 << " code cache collection, code="
1116 << PrettySize(CodeCacheSize())
1117 << ", data=" << PrettySize(DataCacheSize());
1118
1119 DoCollection(self, /* collect_profiling_info= */ do_full_collection);
1120
1121 VLOG(jit) << "After code cache collection, code="
1122 << PrettySize(CodeCacheSize())
1123 << ", data=" << PrettySize(DataCacheSize());
1124
1125 {
1126 MutexLock mu(self, *Locks::jit_lock_);
1127
1128 // Increase the code cache only when we do partial collections.
1129 // TODO: base this strategy on how full the code cache is?
1130 if (do_full_collection) {
1131 last_collection_increased_code_cache_ = false;
1132 } else {
1133 last_collection_increased_code_cache_ = true;
1134 private_region_.IncreaseCodeCacheCapacity();
1135 }
1136
1137 bool next_collection_will_be_full = ShouldDoFullCollection();
1138
1139 // Start polling the liveness of compiled code to prepare for the next full collection.
1140 if (next_collection_will_be_full) {
1141 if (Runtime::Current()->GetJITOptions()->CanCompileBaseline()) {
1142 for (ProfilingInfo* info : profiling_infos_) {
1143 info->SetBaselineHotnessCount(0);
1144 }
1145 } else {
1146 // Save the entry point of methods we have compiled, and update the entry
1147 // point of those methods to the interpreter. If the method is invoked, the
1148 // interpreter will update its entry point to the compiled code and call it.
1149 for (ProfilingInfo* info : profiling_infos_) {
1150 const void* entry_point = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1151 if (!IsInZygoteDataSpace(info) && ContainsPc(entry_point)) {
1152 info->SetSavedEntryPoint(entry_point);
1153 // Don't call Instrumentation::UpdateMethodsCode(), as it can check the declaring
1154 // class of the method. We may be concurrently running a GC which makes accessing
1155 // the class unsafe. We know it is OK to bypass the instrumentation as we've just
1156 // checked that the current entry point is JIT compiled code.
1157 info->GetMethod()->SetEntryPointFromQuickCompiledCode(GetQuickToInterpreterBridge());
1158 }
1159 }
1160 }
1161
1162 // Change entry points of native methods back to the GenericJNI entrypoint.
1163 for (const auto& entry : jni_stubs_map_) {
1164 const JniStubData& data = entry.second;
1165 if (!data.IsCompiled() || IsInZygoteExecSpace(data.GetCode())) {
1166 continue;
1167 }
1168 // Make sure a single invocation of the GenericJNI trampoline tries to recompile.
1169 uint16_t new_counter = Runtime::Current()->GetJit()->HotMethodThreshold() - 1u;
1170 const OatQuickMethodHeader* method_header =
1171 OatQuickMethodHeader::FromCodePointer(data.GetCode());
1172 for (ArtMethod* method : data.GetMethods()) {
1173 if (method->GetEntryPointFromQuickCompiledCode() == method_header->GetEntryPoint()) {
1174 // Don't call Instrumentation::UpdateMethodsCode(), same as for normal methods above.
1175 method->SetCounter(new_counter);
1176 method->SetEntryPointFromQuickCompiledCode(GetQuickGenericJniStub());
1177 }
1178 }
1179 }
1180 }
1181 live_bitmap_.reset(nullptr);
1182 NotifyCollectionDone(self);
1183 }
1184 }
1185 Runtime::Current()->GetJit()->AddTimingLogger(logger);
1186 }
1187
RemoveUnmarkedCode(Thread * self)1188 void JitCodeCache::RemoveUnmarkedCode(Thread* self) {
1189 ScopedTrace trace(__FUNCTION__);
1190 std::unordered_set<OatQuickMethodHeader*> method_headers;
1191 {
1192 MutexLock mu(self, *Locks::jit_lock_);
1193 // Iterate over all compiled code and remove entries that are not marked.
1194 for (auto it = jni_stubs_map_.begin(); it != jni_stubs_map_.end();) {
1195 JniStubData* data = &it->second;
1196 if (IsInZygoteExecSpace(data->GetCode()) ||
1197 !data->IsCompiled() ||
1198 GetLiveBitmap()->Test(FromCodeToAllocation(data->GetCode()))) {
1199 ++it;
1200 } else {
1201 method_headers.insert(OatQuickMethodHeader::FromCodePointer(data->GetCode()));
1202 it = jni_stubs_map_.erase(it);
1203 }
1204 }
1205 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
1206 const void* code_ptr = it->first;
1207 uintptr_t allocation = FromCodeToAllocation(code_ptr);
1208 if (IsInZygoteExecSpace(code_ptr) || GetLiveBitmap()->Test(allocation)) {
1209 ++it;
1210 } else {
1211 OatQuickMethodHeader* header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1212 method_headers.insert(header);
1213 it = method_code_map_.erase(it);
1214 }
1215 }
1216 }
1217 FreeAllMethodHeaders(method_headers);
1218 }
1219
GetGarbageCollectCode()1220 bool JitCodeCache::GetGarbageCollectCode() {
1221 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1222 return garbage_collect_code_;
1223 }
1224
SetGarbageCollectCode(bool value)1225 void JitCodeCache::SetGarbageCollectCode(bool value) {
1226 Thread* self = Thread::Current();
1227 MutexLock mu(self, *Locks::jit_lock_);
1228 if (garbage_collect_code_ != value) {
1229 if (garbage_collect_code_) {
1230 // When dynamically disabling the garbage collection, we neee
1231 // to make sure that a potential current collection is finished, and also
1232 // clear the saved entry point in profiling infos to avoid dangling pointers.
1233 WaitForPotentialCollectionToComplete(self);
1234 for (ProfilingInfo* info : profiling_infos_) {
1235 info->SetSavedEntryPoint(nullptr);
1236 }
1237 }
1238 // Update the flag while holding the lock to ensure no thread will try to GC.
1239 garbage_collect_code_ = value;
1240 }
1241 }
1242
DoCollection(Thread * self,bool collect_profiling_info)1243 void JitCodeCache::DoCollection(Thread* self, bool collect_profiling_info) {
1244 ScopedTrace trace(__FUNCTION__);
1245 {
1246 MutexLock mu(self, *Locks::jit_lock_);
1247
1248 if (Runtime::Current()->GetJITOptions()->CanCompileBaseline()) {
1249 // Update to interpreter the methods that have baseline entrypoints and whose baseline
1250 // hotness count is zero.
1251 // Note that these methods may be in thread stack or concurrently revived
1252 // between. That's OK, as the thread executing it will mark it.
1253 for (ProfilingInfo* info : profiling_infos_) {
1254 if (info->GetBaselineHotnessCount() == 0) {
1255 const void* entry_point = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1256 if (ContainsPc(entry_point)) {
1257 OatQuickMethodHeader* method_header =
1258 OatQuickMethodHeader::FromEntryPoint(entry_point);
1259 if (CodeInfo::IsBaseline(method_header->GetOptimizedCodeInfoPtr())) {
1260 info->GetMethod()->SetEntryPointFromQuickCompiledCode(GetQuickToInterpreterBridge());
1261 }
1262 }
1263 }
1264 }
1265 // TODO: collect profiling info
1266 // TODO: collect optimized code?
1267 } else {
1268 if (collect_profiling_info) {
1269 // Clear the profiling info of methods that do not have compiled code as entrypoint.
1270 // Also remove the saved entry point from the ProfilingInfo objects.
1271 for (ProfilingInfo* info : profiling_infos_) {
1272 const void* ptr = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1273 if (!ContainsPc(ptr) && !info->IsInUseByCompiler() && !IsInZygoteDataSpace(info)) {
1274 info->GetMethod()->SetProfilingInfo(nullptr);
1275 }
1276
1277 if (info->GetSavedEntryPoint() != nullptr) {
1278 info->SetSavedEntryPoint(nullptr);
1279 // We are going to move this method back to interpreter. Clear the counter now to
1280 // give it a chance to be hot again.
1281 ClearMethodCounter(info->GetMethod(), /*was_warm=*/ true);
1282 }
1283 }
1284 } else if (kIsDebugBuild) {
1285 // Sanity check that the profiling infos do not have a dangling entry point.
1286 for (ProfilingInfo* info : profiling_infos_) {
1287 DCHECK(!Runtime::Current()->IsZygote());
1288 const void* entry_point = info->GetSavedEntryPoint();
1289 DCHECK(entry_point == nullptr || IsInZygoteExecSpace(entry_point));
1290 }
1291 }
1292 }
1293
1294 // Mark compiled code that are entrypoints of ArtMethods. Compiled code that is not
1295 // an entry point is either:
1296 // - an osr compiled code, that will be removed if not in a thread call stack.
1297 // - discarded compiled code, that will be removed if not in a thread call stack.
1298 for (const auto& entry : jni_stubs_map_) {
1299 const JniStubData& data = entry.second;
1300 const void* code_ptr = data.GetCode();
1301 if (IsInZygoteExecSpace(code_ptr)) {
1302 continue;
1303 }
1304 const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1305 for (ArtMethod* method : data.GetMethods()) {
1306 if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) {
1307 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
1308 break;
1309 }
1310 }
1311 }
1312 for (const auto& it : method_code_map_) {
1313 ArtMethod* method = it.second;
1314 const void* code_ptr = it.first;
1315 if (IsInZygoteExecSpace(code_ptr)) {
1316 continue;
1317 }
1318 const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1319 if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) {
1320 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
1321 }
1322 }
1323
1324 // Empty osr method map, as osr compiled code will be deleted (except the ones
1325 // on thread stacks).
1326 osr_code_map_.clear();
1327 }
1328
1329 // Run a checkpoint on all threads to mark the JIT compiled code they are running.
1330 MarkCompiledCodeOnThreadStacks(self);
1331
1332 // At this point, mutator threads are still running, and entrypoints of methods can
1333 // change. We do know they cannot change to a code cache entry that is not marked,
1334 // therefore we can safely remove those entries.
1335 RemoveUnmarkedCode(self);
1336
1337 if (collect_profiling_info) {
1338 MutexLock mu(self, *Locks::jit_lock_);
1339 // Free all profiling infos of methods not compiled nor being compiled.
1340 auto profiling_kept_end = std::remove_if(profiling_infos_.begin(), profiling_infos_.end(),
1341 [this] (ProfilingInfo* info) NO_THREAD_SAFETY_ANALYSIS {
1342 const void* ptr = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1343 // We have previously cleared the ProfilingInfo pointer in the ArtMethod in the hope
1344 // that the compiled code would not get revived. As mutator threads run concurrently,
1345 // they may have revived the compiled code, and now we are in the situation where
1346 // a method has compiled code but no ProfilingInfo.
1347 // We make sure compiled methods have a ProfilingInfo object. It is needed for
1348 // code cache collection.
1349 if (ContainsPc(ptr) &&
1350 info->GetMethod()->GetProfilingInfo(kRuntimePointerSize) == nullptr) {
1351 info->GetMethod()->SetProfilingInfo(info);
1352 } else if (info->GetMethod()->GetProfilingInfo(kRuntimePointerSize) != info) {
1353 // No need for this ProfilingInfo object anymore.
1354 private_region_.FreeWritableData(reinterpret_cast<uint8_t*>(info));
1355 return true;
1356 }
1357 return false;
1358 });
1359 profiling_infos_.erase(profiling_kept_end, profiling_infos_.end());
1360 }
1361 }
1362
LookupMethodHeader(uintptr_t pc,ArtMethod * method)1363 OatQuickMethodHeader* JitCodeCache::LookupMethodHeader(uintptr_t pc, ArtMethod* method) {
1364 static_assert(kRuntimeISA != InstructionSet::kThumb2, "kThumb2 cannot be a runtime ISA");
1365 if (kRuntimeISA == InstructionSet::kArm) {
1366 // On Thumb-2, the pc is offset by one.
1367 --pc;
1368 }
1369 if (!ContainsPc(reinterpret_cast<const void*>(pc))) {
1370 return nullptr;
1371 }
1372
1373 if (!kIsDebugBuild) {
1374 // Called with null `method` only from MarkCodeClosure::Run() in debug build.
1375 CHECK(method != nullptr);
1376 }
1377
1378 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1379 OatQuickMethodHeader* method_header = nullptr;
1380 ArtMethod* found_method = nullptr; // Only for DCHECK(), not for JNI stubs.
1381 if (method != nullptr && UNLIKELY(method->IsNative())) {
1382 auto it = jni_stubs_map_.find(JniStubKey(method));
1383 if (it == jni_stubs_map_.end() || !ContainsElement(it->second.GetMethods(), method)) {
1384 return nullptr;
1385 }
1386 const void* code_ptr = it->second.GetCode();
1387 method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1388 if (!method_header->Contains(pc)) {
1389 return nullptr;
1390 }
1391 } else {
1392 if (shared_region_.IsInExecSpace(reinterpret_cast<const void*>(pc))) {
1393 const void* code_ptr = zygote_map_.GetCodeFor(method, pc);
1394 if (code_ptr != nullptr) {
1395 return OatQuickMethodHeader::FromCodePointer(code_ptr);
1396 }
1397 }
1398 auto it = method_code_map_.lower_bound(reinterpret_cast<const void*>(pc));
1399 if (it != method_code_map_.begin()) {
1400 --it;
1401 const void* code_ptr = it->first;
1402 if (OatQuickMethodHeader::FromCodePointer(code_ptr)->Contains(pc)) {
1403 method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1404 found_method = it->second;
1405 }
1406 }
1407 if (method_header == nullptr && method == nullptr) {
1408 // Scan all compiled JNI stubs as well. This slow search is used only
1409 // for checks in debug build, for release builds the `method` is not null.
1410 for (auto&& entry : jni_stubs_map_) {
1411 const JniStubData& data = entry.second;
1412 if (data.IsCompiled() &&
1413 OatQuickMethodHeader::FromCodePointer(data.GetCode())->Contains(pc)) {
1414 method_header = OatQuickMethodHeader::FromCodePointer(data.GetCode());
1415 }
1416 }
1417 }
1418 if (method_header == nullptr) {
1419 return nullptr;
1420 }
1421 }
1422
1423 if (kIsDebugBuild && method != nullptr && !method->IsNative()) {
1424 DCHECK_EQ(found_method, method)
1425 << ArtMethod::PrettyMethod(method) << " "
1426 << ArtMethod::PrettyMethod(found_method) << " "
1427 << std::hex << pc;
1428 }
1429 return method_header;
1430 }
1431
LookupOsrMethodHeader(ArtMethod * method)1432 OatQuickMethodHeader* JitCodeCache::LookupOsrMethodHeader(ArtMethod* method) {
1433 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1434 auto it = osr_code_map_.find(method);
1435 if (it == osr_code_map_.end()) {
1436 return nullptr;
1437 }
1438 return OatQuickMethodHeader::FromCodePointer(it->second);
1439 }
1440
AddProfilingInfo(Thread * self,ArtMethod * method,const std::vector<uint32_t> & entries,bool retry_allocation)1441 ProfilingInfo* JitCodeCache::AddProfilingInfo(Thread* self,
1442 ArtMethod* method,
1443 const std::vector<uint32_t>& entries,
1444 bool retry_allocation)
1445 // No thread safety analysis as we are using TryLock/Unlock explicitly.
1446 NO_THREAD_SAFETY_ANALYSIS {
1447 DCHECK(CanAllocateProfilingInfo());
1448 ProfilingInfo* info = nullptr;
1449 if (!retry_allocation) {
1450 // If we are allocating for the interpreter, just try to lock, to avoid
1451 // lock contention with the JIT.
1452 if (Locks::jit_lock_->ExclusiveTryLock(self)) {
1453 info = AddProfilingInfoInternal(self, method, entries);
1454 Locks::jit_lock_->ExclusiveUnlock(self);
1455 }
1456 } else {
1457 {
1458 MutexLock mu(self, *Locks::jit_lock_);
1459 info = AddProfilingInfoInternal(self, method, entries);
1460 }
1461
1462 if (info == nullptr) {
1463 GarbageCollectCache(self);
1464 MutexLock mu(self, *Locks::jit_lock_);
1465 info = AddProfilingInfoInternal(self, method, entries);
1466 }
1467 }
1468 return info;
1469 }
1470
AddProfilingInfoInternal(Thread * self ATTRIBUTE_UNUSED,ArtMethod * method,const std::vector<uint32_t> & entries)1471 ProfilingInfo* JitCodeCache::AddProfilingInfoInternal(Thread* self ATTRIBUTE_UNUSED,
1472 ArtMethod* method,
1473 const std::vector<uint32_t>& entries) {
1474 size_t profile_info_size = RoundUp(
1475 sizeof(ProfilingInfo) + sizeof(InlineCache) * entries.size(),
1476 sizeof(void*));
1477
1478 // Check whether some other thread has concurrently created it.
1479 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1480 if (info != nullptr) {
1481 return info;
1482 }
1483
1484 const uint8_t* data = private_region_.AllocateData(profile_info_size);
1485 if (data == nullptr) {
1486 return nullptr;
1487 }
1488 uint8_t* writable_data = private_region_.GetWritableDataAddress(data);
1489 info = new (writable_data) ProfilingInfo(method, entries);
1490
1491 // Make sure other threads see the data in the profiling info object before the
1492 // store in the ArtMethod's ProfilingInfo pointer.
1493 std::atomic_thread_fence(std::memory_order_release);
1494
1495 method->SetProfilingInfo(info);
1496 profiling_infos_.push_back(info);
1497 histogram_profiling_info_memory_use_.AddValue(profile_info_size);
1498 return info;
1499 }
1500
MoreCore(const void * mspace,intptr_t increment)1501 void* JitCodeCache::MoreCore(const void* mspace, intptr_t increment) {
1502 return shared_region_.OwnsSpace(mspace)
1503 ? shared_region_.MoreCore(mspace, increment)
1504 : private_region_.MoreCore(mspace, increment);
1505 }
1506
GetProfiledMethods(const std::set<std::string> & dex_base_locations,std::vector<ProfileMethodInfo> & methods)1507 void JitCodeCache::GetProfiledMethods(const std::set<std::string>& dex_base_locations,
1508 std::vector<ProfileMethodInfo>& methods) {
1509 Thread* self = Thread::Current();
1510 WaitUntilInlineCacheAccessible(self);
1511 MutexLock mu(self, *Locks::jit_lock_);
1512 ScopedTrace trace(__FUNCTION__);
1513 uint16_t jit_compile_threshold = Runtime::Current()->GetJITOptions()->GetCompileThreshold();
1514 for (const ProfilingInfo* info : profiling_infos_) {
1515 ArtMethod* method = info->GetMethod();
1516 const DexFile* dex_file = method->GetDexFile();
1517 const std::string base_location = DexFileLoader::GetBaseLocation(dex_file->GetLocation());
1518 if (!ContainsElement(dex_base_locations, base_location)) {
1519 // Skip dex files which are not profiled.
1520 continue;
1521 }
1522 std::vector<ProfileMethodInfo::ProfileInlineCache> inline_caches;
1523
1524 // If the method didn't reach the compilation threshold don't save the inline caches.
1525 // They might be incomplete and cause unnecessary deoptimizations.
1526 // If the inline cache is empty the compiler will generate a regular invoke virtual/interface.
1527 if (method->GetCounter() < jit_compile_threshold) {
1528 methods.emplace_back(/*ProfileMethodInfo*/
1529 MethodReference(dex_file, method->GetDexMethodIndex()), inline_caches);
1530 continue;
1531 }
1532
1533 for (size_t i = 0; i < info->number_of_inline_caches_; ++i) {
1534 std::vector<TypeReference> profile_classes;
1535 const InlineCache& cache = info->cache_[i];
1536 ArtMethod* caller = info->GetMethod();
1537 bool is_missing_types = false;
1538 for (size_t k = 0; k < InlineCache::kIndividualCacheSize; k++) {
1539 mirror::Class* cls = cache.classes_[k].Read();
1540 if (cls == nullptr) {
1541 break;
1542 }
1543
1544 // Check if the receiver is in the boot class path or if it's in the
1545 // same class loader as the caller. If not, skip it, as there is not
1546 // much we can do during AOT.
1547 if (!cls->IsBootStrapClassLoaded() &&
1548 caller->GetClassLoader() != cls->GetClassLoader()) {
1549 is_missing_types = true;
1550 continue;
1551 }
1552
1553 const DexFile* class_dex_file = nullptr;
1554 dex::TypeIndex type_index;
1555
1556 if (cls->GetDexCache() == nullptr) {
1557 DCHECK(cls->IsArrayClass()) << cls->PrettyClass();
1558 // Make a best effort to find the type index in the method's dex file.
1559 // We could search all open dex files but that might turn expensive
1560 // and probably not worth it.
1561 class_dex_file = dex_file;
1562 type_index = cls->FindTypeIndexInOtherDexFile(*dex_file);
1563 } else {
1564 class_dex_file = &(cls->GetDexFile());
1565 type_index = cls->GetDexTypeIndex();
1566 }
1567 if (!type_index.IsValid()) {
1568 // Could be a proxy class or an array for which we couldn't find the type index.
1569 is_missing_types = true;
1570 continue;
1571 }
1572 if (ContainsElement(dex_base_locations,
1573 DexFileLoader::GetBaseLocation(class_dex_file->GetLocation()))) {
1574 // Only consider classes from the same apk (including multidex).
1575 profile_classes.emplace_back(/*ProfileMethodInfo::ProfileClassReference*/
1576 class_dex_file, type_index);
1577 } else {
1578 is_missing_types = true;
1579 }
1580 }
1581 if (!profile_classes.empty()) {
1582 inline_caches.emplace_back(/*ProfileMethodInfo::ProfileInlineCache*/
1583 cache.dex_pc_, is_missing_types, profile_classes);
1584 }
1585 }
1586 methods.emplace_back(/*ProfileMethodInfo*/
1587 MethodReference(dex_file, method->GetDexMethodIndex()), inline_caches);
1588 }
1589 }
1590
IsOsrCompiled(ArtMethod * method)1591 bool JitCodeCache::IsOsrCompiled(ArtMethod* method) {
1592 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1593 return osr_code_map_.find(method) != osr_code_map_.end();
1594 }
1595
NotifyCompilationOf(ArtMethod * method,Thread * self,bool osr,bool prejit,bool baseline,JitMemoryRegion * region)1596 bool JitCodeCache::NotifyCompilationOf(ArtMethod* method,
1597 Thread* self,
1598 bool osr,
1599 bool prejit,
1600 bool baseline,
1601 JitMemoryRegion* region) {
1602 const void* existing_entry_point = method->GetEntryPointFromQuickCompiledCode();
1603 if (!osr && ContainsPc(existing_entry_point)) {
1604 OatQuickMethodHeader* method_header =
1605 OatQuickMethodHeader::FromEntryPoint(existing_entry_point);
1606 if (CodeInfo::IsBaseline(method_header->GetOptimizedCodeInfoPtr()) == baseline) {
1607 VLOG(jit) << "Not compiling "
1608 << method->PrettyMethod()
1609 << " because it has already been compiled"
1610 << " baseline=" << std::boolalpha << baseline;
1611 return false;
1612 }
1613 }
1614
1615 if (NeedsClinitCheckBeforeCall(method) && !prejit) {
1616 // We do not need a synchronization barrier for checking the visibly initialized status
1617 // or checking the initialized status just for requesting visible initialization.
1618 ClassStatus status = method->GetDeclaringClass()
1619 ->GetStatus<kDefaultVerifyFlags, /*kWithSynchronizationBarrier=*/ false>();
1620 if (status != ClassStatus::kVisiblyInitialized) {
1621 // Unless we're pre-jitting, we currently don't save the JIT compiled code if we cannot
1622 // update the entrypoint due to needing an initialization check.
1623 if (status == ClassStatus::kInitialized) {
1624 // Request visible initialization but do not block to allow compiling other methods.
1625 // Hopefully, this will complete by the time the method becomes hot again.
1626 Runtime::Current()->GetClassLinker()->MakeInitializedClassesVisiblyInitialized(
1627 self, /*wait=*/ false);
1628 }
1629 VLOG(jit) << "Not compiling "
1630 << method->PrettyMethod()
1631 << " because it has the resolution stub";
1632 // Give it a new chance to be hot.
1633 ClearMethodCounter(method, /*was_warm=*/ false);
1634 return false;
1635 }
1636 }
1637
1638 if (osr) {
1639 MutexLock mu(self, *Locks::jit_lock_);
1640 if (osr_code_map_.find(method) != osr_code_map_.end()) {
1641 return false;
1642 }
1643 }
1644
1645 if (UNLIKELY(method->IsNative())) {
1646 MutexLock mu(self, *Locks::jit_lock_);
1647 JniStubKey key(method);
1648 auto it = jni_stubs_map_.find(key);
1649 bool new_compilation = false;
1650 if (it == jni_stubs_map_.end()) {
1651 // Create a new entry to mark the stub as being compiled.
1652 it = jni_stubs_map_.Put(key, JniStubData{});
1653 new_compilation = true;
1654 }
1655 JniStubData* data = &it->second;
1656 data->AddMethod(method);
1657 if (data->IsCompiled()) {
1658 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(data->GetCode());
1659 const void* entrypoint = method_header->GetEntryPoint();
1660 // Update also entrypoints of other methods held by the JniStubData.
1661 // We could simply update the entrypoint of `method` but if the last JIT GC has
1662 // changed these entrypoints to GenericJNI in preparation for a full GC, we may
1663 // as well change them back as this stub shall not be collected anyway and this
1664 // can avoid a few expensive GenericJNI calls.
1665 data->UpdateEntryPoints(entrypoint);
1666 if (collection_in_progress_) {
1667 if (!IsInZygoteExecSpace(data->GetCode())) {
1668 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(data->GetCode()));
1669 }
1670 }
1671 }
1672 return new_compilation;
1673 } else {
1674 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1675 if (CanAllocateProfilingInfo() && baseline && info == nullptr) {
1676 // We can retry allocation here as we're the JIT thread.
1677 if (ProfilingInfo::Create(self, method, /* retry_allocation= */ true)) {
1678 info = method->GetProfilingInfo(kRuntimePointerSize);
1679 }
1680 }
1681 if (info == nullptr) {
1682 // When prejitting, we don't allocate a profiling info.
1683 if (!prejit && !IsSharedRegion(*region)) {
1684 VLOG(jit) << method->PrettyMethod() << " needs a ProfilingInfo to be compiled";
1685 // Because the counter is not atomic, there are some rare cases where we may not hit the
1686 // threshold for creating the ProfilingInfo. Reset the counter now to "correct" this.
1687 ClearMethodCounter(method, /*was_warm=*/ false);
1688 return false;
1689 }
1690 } else {
1691 MutexLock mu(self, *Locks::jit_lock_);
1692 if (info->IsMethodBeingCompiled(osr)) {
1693 return false;
1694 }
1695 info->SetIsMethodBeingCompiled(true, osr);
1696 }
1697 return true;
1698 }
1699 }
1700
NotifyCompilerUse(ArtMethod * method,Thread * self)1701 ProfilingInfo* JitCodeCache::NotifyCompilerUse(ArtMethod* method, Thread* self) {
1702 MutexLock mu(self, *Locks::jit_lock_);
1703 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1704 if (info != nullptr) {
1705 if (!info->IncrementInlineUse()) {
1706 // Overflow of inlining uses, just bail.
1707 return nullptr;
1708 }
1709 }
1710 return info;
1711 }
1712
DoneCompilerUse(ArtMethod * method,Thread * self)1713 void JitCodeCache::DoneCompilerUse(ArtMethod* method, Thread* self) {
1714 MutexLock mu(self, *Locks::jit_lock_);
1715 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1716 DCHECK(info != nullptr);
1717 info->DecrementInlineUse();
1718 }
1719
DoneCompiling(ArtMethod * method,Thread * self,bool osr)1720 void JitCodeCache::DoneCompiling(ArtMethod* method, Thread* self, bool osr) {
1721 DCHECK_EQ(Thread::Current(), self);
1722 MutexLock mu(self, *Locks::jit_lock_);
1723 if (UNLIKELY(method->IsNative())) {
1724 auto it = jni_stubs_map_.find(JniStubKey(method));
1725 DCHECK(it != jni_stubs_map_.end());
1726 JniStubData* data = &it->second;
1727 DCHECK(ContainsElement(data->GetMethods(), method));
1728 if (UNLIKELY(!data->IsCompiled())) {
1729 // Failed to compile; the JNI compiler never fails, but the cache may be full.
1730 jni_stubs_map_.erase(it); // Remove the entry added in NotifyCompilationOf().
1731 } // else Commit() updated entrypoints of all methods in the JniStubData.
1732 } else {
1733 ProfilingInfo* info = method->GetProfilingInfo(kRuntimePointerSize);
1734 if (info != nullptr) {
1735 DCHECK(info->IsMethodBeingCompiled(osr));
1736 info->SetIsMethodBeingCompiled(false, osr);
1737 }
1738 }
1739 }
1740
InvalidateAllCompiledCode()1741 void JitCodeCache::InvalidateAllCompiledCode() {
1742 art::MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1743 size_t cnt = profiling_infos_.size();
1744 size_t osr_size = osr_code_map_.size();
1745 for (ProfilingInfo* pi : profiling_infos_) {
1746 // NB Due to OSR we might run this on some methods multiple times but this should be fine.
1747 ArtMethod* meth = pi->GetMethod();
1748 pi->SetSavedEntryPoint(nullptr);
1749 // We had a ProfilingInfo so we must be warm.
1750 ClearMethodCounter(meth, /*was_warm=*/true);
1751 ClassLinker* linker = Runtime::Current()->GetClassLinker();
1752 if (meth->IsObsolete()) {
1753 linker->SetEntryPointsForObsoleteMethod(meth);
1754 } else {
1755 linker->SetEntryPointsToInterpreter(meth);
1756 }
1757 }
1758 osr_code_map_.clear();
1759 VLOG(jit) << "Invalidated the compiled code of " << (cnt - osr_size) << " methods and "
1760 << osr_size << " OSRs.";
1761 }
1762
InvalidateCompiledCodeFor(ArtMethod * method,const OatQuickMethodHeader * header)1763 void JitCodeCache::InvalidateCompiledCodeFor(ArtMethod* method,
1764 const OatQuickMethodHeader* header) {
1765 DCHECK(!method->IsNative());
1766 ProfilingInfo* profiling_info = method->GetProfilingInfo(kRuntimePointerSize);
1767 const void* method_entrypoint = method->GetEntryPointFromQuickCompiledCode();
1768 if ((profiling_info != nullptr) &&
1769 (profiling_info->GetSavedEntryPoint() == header->GetEntryPoint())) {
1770 // When instrumentation is set, the actual entrypoint is the one in the profiling info.
1771 method_entrypoint = profiling_info->GetSavedEntryPoint();
1772 // Prevent future uses of the compiled code.
1773 profiling_info->SetSavedEntryPoint(nullptr);
1774 }
1775
1776 // Clear the method counter if we are running jitted code since we might want to jit this again in
1777 // the future.
1778 if (method_entrypoint == header->GetEntryPoint()) {
1779 // The entrypoint is the one to invalidate, so we just update it to the interpreter entry point
1780 // and clear the counter to get the method Jitted again.
1781 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
1782 method, GetQuickToInterpreterBridge());
1783 ClearMethodCounter(method, /*was_warm=*/ profiling_info != nullptr);
1784 } else {
1785 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1786 auto it = osr_code_map_.find(method);
1787 if (it != osr_code_map_.end() && OatQuickMethodHeader::FromCodePointer(it->second) == header) {
1788 // Remove the OSR method, to avoid using it again.
1789 osr_code_map_.erase(it);
1790 }
1791 }
1792
1793 // In case the method was pre-compiled, clear that information so we
1794 // can recompile it ourselves.
1795 if (method->IsPreCompiled()) {
1796 method->ClearPreCompiled();
1797 }
1798 }
1799
Dump(std::ostream & os)1800 void JitCodeCache::Dump(std::ostream& os) {
1801 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1802 os << "Current JIT code cache size (used / resident): "
1803 << GetCurrentRegion()->GetUsedMemoryForCode() / KB << "KB / "
1804 << GetCurrentRegion()->GetResidentMemoryForCode() / KB << "KB\n"
1805 << "Current JIT data cache size (used / resident): "
1806 << GetCurrentRegion()->GetUsedMemoryForData() / KB << "KB / "
1807 << GetCurrentRegion()->GetResidentMemoryForData() / KB << "KB\n";
1808 if (!Runtime::Current()->IsZygote()) {
1809 os << "Zygote JIT code cache size (at point of fork): "
1810 << shared_region_.GetUsedMemoryForCode() / KB << "KB / "
1811 << shared_region_.GetResidentMemoryForCode() / KB << "KB\n"
1812 << "Zygote JIT data cache size (at point of fork): "
1813 << shared_region_.GetUsedMemoryForData() / KB << "KB / "
1814 << shared_region_.GetResidentMemoryForData() / KB << "KB\n";
1815 }
1816 os << "Current JIT mini-debug-info size: " << PrettySize(GetJitMiniDebugInfoMemUsage()) << "\n"
1817 << "Current JIT capacity: " << PrettySize(GetCurrentRegion()->GetCurrentCapacity()) << "\n"
1818 << "Current number of JIT JNI stub entries: " << jni_stubs_map_.size() << "\n"
1819 << "Current number of JIT code cache entries: " << method_code_map_.size() << "\n"
1820 << "Total number of JIT compilations: " << number_of_compilations_ << "\n"
1821 << "Total number of JIT compilations for on stack replacement: "
1822 << number_of_osr_compilations_ << "\n"
1823 << "Total number of JIT code cache collections: " << number_of_collections_ << std::endl;
1824 histogram_stack_map_memory_use_.PrintMemoryUse(os);
1825 histogram_code_memory_use_.PrintMemoryUse(os);
1826 histogram_profiling_info_memory_use_.PrintMemoryUse(os);
1827 }
1828
PostForkChildAction(bool is_system_server,bool is_zygote)1829 void JitCodeCache::PostForkChildAction(bool is_system_server, bool is_zygote) {
1830 Thread* self = Thread::Current();
1831
1832 // Remove potential tasks that have been inherited from the zygote.
1833 // We do this now and not in Jit::PostForkChildAction, as system server calls
1834 // JitCodeCache::PostForkChildAction first, and then does some code loading
1835 // that may result in new JIT tasks that we want to keep.
1836 ThreadPool* pool = Runtime::Current()->GetJit()->GetThreadPool();
1837 if (pool != nullptr) {
1838 pool->RemoveAllTasks(self);
1839 }
1840
1841 MutexLock mu(self, *Locks::jit_lock_);
1842
1843 // Reset potential writable MemMaps inherited from the zygote. We never want
1844 // to write to them.
1845 shared_region_.ResetWritableMappings();
1846
1847 if (is_zygote || Runtime::Current()->IsSafeMode()) {
1848 // Don't create a private region for a child zygote. Regions are usually map shared
1849 // (to satisfy dual-view), and we don't want children of a child zygote to inherit it.
1850 return;
1851 }
1852
1853 // Reset all statistics to be specific to this process.
1854 number_of_compilations_ = 0;
1855 number_of_osr_compilations_ = 0;
1856 number_of_collections_ = 0;
1857 histogram_stack_map_memory_use_.Reset();
1858 histogram_code_memory_use_.Reset();
1859 histogram_profiling_info_memory_use_.Reset();
1860
1861 size_t initial_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheInitialCapacity();
1862 size_t max_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheMaxCapacity();
1863 std::string error_msg;
1864 if (!private_region_.Initialize(initial_capacity,
1865 max_capacity,
1866 /* rwx_memory_allowed= */ !is_system_server,
1867 is_zygote,
1868 &error_msg)) {
1869 LOG(WARNING) << "Could not create private region after zygote fork: " << error_msg;
1870 }
1871 }
1872
GetCurrentRegion()1873 JitMemoryRegion* JitCodeCache::GetCurrentRegion() {
1874 return Runtime::Current()->IsZygote() ? &shared_region_ : &private_region_;
1875 }
1876
Initialize(uint32_t number_of_methods)1877 void ZygoteMap::Initialize(uint32_t number_of_methods) {
1878 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1879 // Allocate for 40-80% capacity. This will offer OK lookup times, and termination
1880 // cases.
1881 size_t capacity = RoundUpToPowerOfTwo(number_of_methods * 100 / 80);
1882 const uint8_t* memory = region_->AllocateData(
1883 capacity * sizeof(Entry) + sizeof(ZygoteCompilationState));
1884 if (memory == nullptr) {
1885 LOG(WARNING) << "Could not allocate data for the zygote map";
1886 return;
1887 }
1888 const Entry* data = reinterpret_cast<const Entry*>(memory);
1889 region_->FillData(data, capacity, Entry { nullptr, nullptr });
1890 map_ = ArrayRef(data, capacity);
1891 compilation_state_ = reinterpret_cast<const ZygoteCompilationState*>(
1892 memory + capacity * sizeof(Entry));
1893 region_->WriteData(compilation_state_, ZygoteCompilationState::kInProgress);
1894 }
1895
GetCodeFor(ArtMethod * method,uintptr_t pc) const1896 const void* ZygoteMap::GetCodeFor(ArtMethod* method, uintptr_t pc) const {
1897 if (map_.empty()) {
1898 return nullptr;
1899 }
1900
1901 if (method == nullptr) {
1902 // Do a linear search. This should only be used in debug builds.
1903 CHECK(kIsDebugBuild);
1904 for (const Entry& entry : map_) {
1905 const void* code_ptr = entry.code_ptr;
1906 if (code_ptr != nullptr) {
1907 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1908 if (method_header->Contains(pc)) {
1909 return code_ptr;
1910 }
1911 }
1912 }
1913 return nullptr;
1914 }
1915
1916 std::hash<ArtMethod*> hf;
1917 size_t index = hf(method) & (map_.size() - 1u);
1918 size_t original_index = index;
1919 // Loop over the array: we know this loop terminates as we will either
1920 // encounter the given method, or a null entry. Both terminate the loop.
1921 // Note that the zygote may concurrently write new entries to the map. That's OK as the
1922 // map is never resized.
1923 while (true) {
1924 const Entry& entry = map_[index];
1925 if (entry.method == nullptr) {
1926 // Not compiled yet.
1927 return nullptr;
1928 }
1929 if (entry.method == method) {
1930 if (entry.code_ptr == nullptr) {
1931 // This is a race with the zygote which wrote the method, but hasn't written the
1932 // code. Just bail and wait for the next time we need the method.
1933 return nullptr;
1934 }
1935 if (pc != 0 && !OatQuickMethodHeader::FromCodePointer(entry.code_ptr)->Contains(pc)) {
1936 return nullptr;
1937 }
1938 return entry.code_ptr;
1939 }
1940 index = (index + 1) & (map_.size() - 1);
1941 DCHECK_NE(original_index, index);
1942 }
1943 }
1944
Put(const void * code,ArtMethod * method)1945 void ZygoteMap::Put(const void* code, ArtMethod* method) {
1946 if (map_.empty()) {
1947 return;
1948 }
1949 CHECK(Runtime::Current()->IsZygote());
1950 std::hash<ArtMethod*> hf;
1951 size_t index = hf(method) & (map_.size() - 1);
1952 size_t original_index = index;
1953 // Because the size of the map is bigger than the number of methods that will
1954 // be added, we are guaranteed to find a free slot in the array, and
1955 // therefore for this loop to terminate.
1956 while (true) {
1957 const Entry* entry = &map_[index];
1958 if (entry->method == nullptr) {
1959 // Note that readers can read this memory concurrently, but that's OK as
1960 // we are writing pointers.
1961 region_->WriteData(entry, Entry { method, code });
1962 break;
1963 }
1964 index = (index + 1) & (map_.size() - 1);
1965 DCHECK_NE(original_index, index);
1966 }
1967 DCHECK_EQ(GetCodeFor(method), code);
1968 }
1969
1970 } // namespace jit
1971 } // namespace art
1972