indirect_reference_table.h revision 3f5881fda3606b27e30bf903052c73b03910f90b
1/* 2 * Copyright (C) 2009 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17#ifndef ART_RUNTIME_INDIRECT_REFERENCE_TABLE_H_ 18#define ART_RUNTIME_INDIRECT_REFERENCE_TABLE_H_ 19 20#include <stdint.h> 21 22#include <iosfwd> 23#include <string> 24 25#include "base/logging.h" 26#include "base/mutex.h" 27#include "gc_root.h" 28#include "object_callbacks.h" 29#include "offsets.h" 30#include "read_barrier_option.h" 31 32namespace art { 33 34class RootInfo; 35 36namespace mirror { 37class Object; 38} // namespace mirror 39 40class MemMap; 41 42/* 43 * Maintain a table of indirect references. Used for local/global JNI 44 * references. 45 * 46 * The table contains object references that are part of the GC root set. 47 * When an object is added we return an IndirectRef that is not a valid 48 * pointer but can be used to find the original value in O(1) time. 49 * Conversions to and from indirect references are performed on upcalls 50 * and downcalls, so they need to be very fast. 51 * 52 * To be efficient for JNI local variable storage, we need to provide 53 * operations that allow us to operate on segments of the table, where 54 * segments are pushed and popped as if on a stack. For example, deletion 55 * of an entry should only succeed if it appears in the current segment, 56 * and we want to be able to strip off the current segment quickly when 57 * a method returns. Additions to the table must be made in the current 58 * segment even if space is available in an earlier area. 59 * 60 * A new segment is created when we call into native code from interpreted 61 * code, or when we handle the JNI PushLocalFrame function. 62 * 63 * The GC must be able to scan the entire table quickly. 64 * 65 * In summary, these must be very fast: 66 * - adding or removing a segment 67 * - adding references to a new segment 68 * - converting an indirect reference back to an Object 69 * These can be a little slower, but must still be pretty quick: 70 * - adding references to a "mature" segment 71 * - removing individual references 72 * - scanning the entire table straight through 73 * 74 * If there's more than one segment, we don't guarantee that the table 75 * will fill completely before we fail due to lack of space. We do ensure 76 * that the current segment will pack tightly, which should satisfy JNI 77 * requirements (e.g. EnsureLocalCapacity). 78 * 79 * To make everything fit nicely in 32-bit integers, the maximum size of 80 * the table is capped at 64K. 81 * 82 * Only SynchronizedGet is synchronized. 83 */ 84 85/* 86 * Indirect reference definition. This must be interchangeable with JNI's 87 * jobject, and it's convenient to let null be null, so we use void*. 88 * 89 * We need a 16-bit table index and a 2-bit reference type (global, local, 90 * weak global). Real object pointers will have zeroes in the low 2 or 3 91 * bits (4- or 8-byte alignment), so it's useful to put the ref type 92 * in the low bits and reserve zero as an invalid value. 93 * 94 * The remaining 14 bits can be used to detect stale indirect references. 95 * For example, if objects don't move, we can use a hash of the original 96 * Object* to make sure the entry hasn't been re-used. (If the Object* 97 * we find there doesn't match because of heap movement, we could do a 98 * secondary check on the preserved hash value; this implies that creating 99 * a global/local ref queries the hash value and forces it to be saved.) 100 * 101 * A more rigorous approach would be to put a serial number in the extra 102 * bits, and keep a copy of the serial number in a parallel table. This is 103 * easier when objects can move, but requires 2x the memory and additional 104 * memory accesses on add/get. It will catch additional problems, e.g.: 105 * create iref1 for obj, delete iref1, create iref2 for same obj, lookup 106 * iref1. A pattern based on object bits will miss this. 107 */ 108typedef void* IndirectRef; 109 110/* 111 * Indirect reference kind, used as the two low bits of IndirectRef. 112 * 113 * For convenience these match up with enum jobjectRefType from jni.h. 114 */ 115enum IndirectRefKind { 116 kHandleScopeOrInvalid = 0, // <<stack indirect reference table or invalid reference>> 117 kLocal = 1, // <<local reference>> 118 kGlobal = 2, // <<global reference>> 119 kWeakGlobal = 3 // <<weak global reference>> 120}; 121std::ostream& operator<<(std::ostream& os, const IndirectRefKind& rhs); 122 123/* 124 * Determine what kind of indirect reference this is. 125 */ 126static inline IndirectRefKind GetIndirectRefKind(IndirectRef iref) { 127 return static_cast<IndirectRefKind>(reinterpret_cast<uintptr_t>(iref) & 0x03); 128} 129 130/* use as initial value for "cookie", and when table has only one segment */ 131static const uint32_t IRT_FIRST_SEGMENT = 0; 132 133/* 134 * Table definition. 135 * 136 * For the global reference table, the expected common operations are 137 * adding a new entry and removing a recently-added entry (usually the 138 * most-recently-added entry). For JNI local references, the common 139 * operations are adding a new entry and removing an entire table segment. 140 * 141 * If "alloc_entries_" is not equal to "max_entries_", the table may expand 142 * when entries are added, which means the memory may move. If you want 143 * to keep pointers into "table" rather than offsets, you must use a 144 * fixed-size table. 145 * 146 * If we delete entries from the middle of the list, we will be left with 147 * "holes". We track the number of holes so that, when adding new elements, 148 * we can quickly decide to do a trivial append or go slot-hunting. 149 * 150 * When the top-most entry is removed, any holes immediately below it are 151 * also removed. Thus, deletion of an entry may reduce "topIndex" by more 152 * than one. 153 * 154 * To get the desired behavior for JNI locals, we need to know the bottom 155 * and top of the current "segment". The top is managed internally, and 156 * the bottom is passed in as a function argument. When we call a native method or 157 * push a local frame, the current top index gets pushed on, and serves 158 * as the new bottom. When we pop a frame off, the value from the stack 159 * becomes the new top index, and the value stored in the previous frame 160 * becomes the new bottom. 161 * 162 * To avoid having to re-scan the table after a pop, we want to push the 163 * number of holes in the table onto the stack. Because of our 64K-entry 164 * cap, we can combine the two into a single unsigned 32-bit value. 165 * Instead of a "bottom" argument we take a "cookie", which includes the 166 * bottom index and the count of holes below the bottom. 167 * 168 * Common alternative implementation: make IndirectRef a pointer to the 169 * actual reference slot. Instead of getting a table and doing a lookup, 170 * the lookup can be done instantly. Operations like determining the 171 * type and deleting the reference are more expensive because the table 172 * must be hunted for (i.e. you have to do a pointer comparison to see 173 * which table it's in), you can't move the table when expanding it (so 174 * realloc() is out), and tricks like serial number checking to detect 175 * stale references aren't possible (though we may be able to get similar 176 * benefits with other approaches). 177 * 178 * TODO: consider a "lastDeleteIndex" for quick hole-filling when an 179 * add immediately follows a delete; must invalidate after segment pop 180 * (which could increase the cost/complexity of method call/return). 181 * Might be worth only using it for JNI globals. 182 * 183 * TODO: may want completely different add/remove algorithms for global 184 * and local refs to improve performance. A large circular buffer might 185 * reduce the amortized cost of adding global references. 186 * 187 */ 188union IRTSegmentState { 189 uint32_t all; 190 struct { 191 uint32_t topIndex:16; /* index of first unused entry */ 192 uint32_t numHoles:16; /* #of holes in entire table */ 193 } parts; 194}; 195 196// Try to choose kIRTPrevCount so that sizeof(IrtEntry) is a power of 2. 197// Contains multiple entries but only one active one, this helps us detect use after free errors 198// since the serial stored in the indirect ref wont match. 199static const size_t kIRTPrevCount = kIsDebugBuild ? 7 : 3; 200class IrtEntry { 201 public: 202 void Add(mirror::Object* obj) SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 203 ++serial_; 204 if (serial_ == kIRTPrevCount) { 205 serial_ = 0; 206 } 207 references_[serial_] = GcRoot<mirror::Object>(obj); 208 } 209 GcRoot<mirror::Object>* GetReference() { 210 DCHECK_LT(serial_, kIRTPrevCount); 211 return &references_[serial_]; 212 } 213 uint32_t GetSerial() const { 214 return serial_; 215 } 216 217 private: 218 uint32_t serial_; 219 GcRoot<mirror::Object> references_[kIRTPrevCount]; 220}; 221static_assert(sizeof(IrtEntry) == (1 + kIRTPrevCount) * sizeof(uint32_t), 222 "Unexpected sizeof(IrtEntry)"); 223 224class IrtIterator { 225 public: 226 explicit IrtIterator(IrtEntry* table, size_t i, size_t capacity) 227 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 228 : table_(table), i_(i), capacity_(capacity) { 229 } 230 231 IrtIterator& operator++() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 232 ++i_; 233 return *this; 234 } 235 236 GcRoot<mirror::Object>* operator*() { 237 // This does not have a read barrier as this is used to visit roots. 238 return table_[i_].GetReference(); 239 } 240 241 bool equals(const IrtIterator& rhs) const { 242 return (i_ == rhs.i_ && table_ == rhs.table_); 243 } 244 245 private: 246 IrtEntry* const table_; 247 size_t i_; 248 const size_t capacity_; 249}; 250 251bool inline operator==(const IrtIterator& lhs, const IrtIterator& rhs) { 252 return lhs.equals(rhs); 253} 254 255bool inline operator!=(const IrtIterator& lhs, const IrtIterator& rhs) { 256 return !lhs.equals(rhs); 257} 258 259class IndirectReferenceTable { 260 public: 261 // WARNING: When using with abort_on_error = false, the object may be in a partially 262 // initialized state. Use IsValid() to check. 263 IndirectReferenceTable(size_t initialCount, size_t maxCount, IndirectRefKind kind, 264 bool abort_on_error = true); 265 266 ~IndirectReferenceTable(); 267 268 bool IsValid() const; 269 270 /* 271 * Add a new entry. "obj" must be a valid non-NULL object reference. 272 * 273 * Returns NULL if the table is full (max entries reached, or alloc 274 * failed during expansion). 275 */ 276 IndirectRef Add(uint32_t cookie, mirror::Object* obj) 277 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_); 278 279 /* 280 * Given an IndirectRef in the table, return the Object it refers to. 281 * 282 * Returns kInvalidIndirectRefObject if iref is invalid. 283 */ 284 template<ReadBarrierOption kReadBarrierOption = kWithReadBarrier> 285 mirror::Object* Get(IndirectRef iref) const SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) 286 ALWAYS_INLINE; 287 288 // Synchronized get which reads a reference, acquiring a lock if necessary. 289 template<ReadBarrierOption kReadBarrierOption = kWithReadBarrier> 290 mirror::Object* SynchronizedGet(Thread* /*self*/, ReaderWriterMutex* /*mutex*/, 291 IndirectRef iref) const 292 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) { 293 return Get<kReadBarrierOption>(iref); 294 } 295 296 /* 297 * Remove an existing entry. 298 * 299 * If the entry is not between the current top index and the bottom index 300 * specified by the cookie, we don't remove anything. This is the behavior 301 * required by JNI's DeleteLocalRef function. 302 * 303 * Returns "false" if nothing was removed. 304 */ 305 bool Remove(uint32_t cookie, IndirectRef iref); 306 307 void AssertEmpty(); 308 309 void Dump(std::ostream& os) const SHARED_LOCKS_REQUIRED(Locks::mutator_lock_); 310 311 /* 312 * Return the #of entries in the entire table. This includes holes, and 313 * so may be larger than the actual number of "live" entries. 314 */ 315 size_t Capacity() const { 316 return segment_state_.parts.topIndex; 317 } 318 319 // Note IrtIterator does not have a read barrier as it's used to visit roots. 320 IrtIterator begin() { 321 return IrtIterator(table_, 0, Capacity()); 322 } 323 324 IrtIterator end() { 325 return IrtIterator(table_, Capacity(), Capacity()); 326 } 327 328 void VisitRoots(RootVisitor* visitor, const RootInfo& root_info) 329 SHARED_LOCKS_REQUIRED(Locks::mutator_lock_); 330 331 uint32_t GetSegmentState() const { 332 return segment_state_.all; 333 } 334 335 void SetSegmentState(uint32_t new_state) { 336 segment_state_.all = new_state; 337 } 338 339 static Offset SegmentStateOffset() { 340 return Offset(OFFSETOF_MEMBER(IndirectReferenceTable, segment_state_)); 341 } 342 343 // Release pages past the end of the table that may have previously held references. 344 void Trim() SHARED_LOCKS_REQUIRED(Locks::mutator_lock_); 345 346 private: 347 // Extract the table index from an indirect reference. 348 static uint32_t ExtractIndex(IndirectRef iref) { 349 uintptr_t uref = reinterpret_cast<uintptr_t>(iref); 350 return (uref >> 2) & 0xffff; 351 } 352 353 /* 354 * The object pointer itself is subject to relocation in some GC 355 * implementations, so we shouldn't really be using it here. 356 */ 357 IndirectRef ToIndirectRef(uint32_t tableIndex) const { 358 DCHECK_LT(tableIndex, 65536U); 359 uint32_t serialChunk = table_[tableIndex].GetSerial(); 360 uintptr_t uref = (serialChunk << 20) | (tableIndex << 2) | kind_; 361 return reinterpret_cast<IndirectRef>(uref); 362 } 363 364 // Abort if check_jni is not enabled. 365 static void AbortIfNoCheckJNI(); 366 367 /* extra debugging checks */ 368 bool GetChecked(IndirectRef) const; 369 bool CheckEntry(const char*, IndirectRef, int) const; 370 371 /* semi-public - read/write by jni down calls */ 372 IRTSegmentState segment_state_; 373 374 // Mem map where we store the indirect refs. 375 std::unique_ptr<MemMap> table_mem_map_; 376 // bottom of the stack. Do not directly access the object references 377 // in this as they are roots. Use Get() that has a read barrier. 378 IrtEntry* table_; 379 /* bit mask, ORed into all irefs */ 380 const IndirectRefKind kind_; 381 /* max #of entries allowed */ 382 const size_t max_entries_; 383}; 384 385} // namespace art 386 387#endif // ART_RUNTIME_INDIRECT_REFERENCE_TABLE_H_ 388