1 //===- HWAddressSanitizer.cpp - detector of uninitialized reads -------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 /// \file 10 /// This file is a part of HWAddressSanitizer, an address basic correctness 11 /// checker based on tagged addressing. 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Transforms/Instrumentation/HWAddressSanitizer.h" 15 #include "llvm/ADT/MapVector.h" 16 #include "llvm/ADT/STLExtras.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/ADT/StringExtras.h" 19 #include "llvm/ADT/StringRef.h" 20 #include "llvm/ADT/Triple.h" 21 #include "llvm/Analysis/PostDominators.h" 22 #include "llvm/Analysis/StackSafetyAnalysis.h" 23 #include "llvm/Analysis/ValueTracking.h" 24 #include "llvm/BinaryFormat/Dwarf.h" 25 #include "llvm/BinaryFormat/ELF.h" 26 #include "llvm/IR/Attributes.h" 27 #include "llvm/IR/BasicBlock.h" 28 #include "llvm/IR/Constant.h" 29 #include "llvm/IR/Constants.h" 30 #include "llvm/IR/DataLayout.h" 31 #include "llvm/IR/DebugInfoMetadata.h" 32 #include "llvm/IR/DerivedTypes.h" 33 #include "llvm/IR/Dominators.h" 34 #include "llvm/IR/Function.h" 35 #include "llvm/IR/IRBuilder.h" 36 #include "llvm/IR/InlineAsm.h" 37 #include "llvm/IR/InstIterator.h" 38 #include "llvm/IR/Instruction.h" 39 #include "llvm/IR/Instructions.h" 40 #include "llvm/IR/IntrinsicInst.h" 41 #include "llvm/IR/Intrinsics.h" 42 #include "llvm/IR/LLVMContext.h" 43 #include "llvm/IR/MDBuilder.h" 44 #include "llvm/IR/Module.h" 45 #include "llvm/IR/Type.h" 46 #include "llvm/IR/Value.h" 47 #include "llvm/Support/Casting.h" 48 #include "llvm/Support/CommandLine.h" 49 #include "llvm/Support/Debug.h" 50 #include "llvm/Support/raw_ostream.h" 51 #include "llvm/Transforms/Instrumentation/AddressSanitizerCommon.h" 52 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 53 #include "llvm/Transforms/Utils/MemoryTaggingSupport.h" 54 #include "llvm/Transforms/Utils/ModuleUtils.h" 55 #include "llvm/Transforms/Utils/PromoteMemToReg.h" 56 57 using namespace llvm; 58 59 #define DEBUG_TYPE "hwasan" 60 61 const char kHwasanModuleCtorName[] = "hwasan.module_ctor"; 62 const char kHwasanNoteName[] = "hwasan.note"; 63 const char kHwasanInitName[] = "__hwasan_init"; 64 const char kHwasanPersonalityThunkName[] = "__hwasan_personality_thunk"; 65 66 const char kHwasanShadowMemoryDynamicAddress[] = 67 "__hwasan_shadow_memory_dynamic_address"; 68 69 // Accesses sizes are powers of two: 1, 2, 4, 8, 16. 70 static const size_t kNumberOfAccessSizes = 5; 71 72 static const size_t kDefaultShadowScale = 4; 73 static const uint64_t kDynamicShadowSentinel = 74 std::numeric_limits<uint64_t>::max(); 75 76 static const unsigned kShadowBaseAlignment = 32; 77 78 static cl::opt<std::string> 79 ClMemoryAccessCallbackPrefix("hwasan-memory-access-callback-prefix", 80 cl::desc("Prefix for memory access callbacks"), 81 cl::Hidden, cl::init("__hwasan_")); 82 83 static cl::opt<bool> ClKasanMemIntrinCallbackPrefix( 84 "hwasan-kernel-mem-intrinsic-prefix", 85 cl::desc("Use prefix for memory intrinsics in KASAN mode"), cl::Hidden, 86 cl::init(false)); 87 88 static cl::opt<bool> ClInstrumentWithCalls( 89 "hwasan-instrument-with-calls", 90 cl::desc("instrument reads and writes with callbacks"), cl::Hidden, 91 cl::init(false)); 92 93 static cl::opt<bool> ClInstrumentReads("hwasan-instrument-reads", 94 cl::desc("instrument read instructions"), 95 cl::Hidden, cl::init(true)); 96 97 static cl::opt<bool> 98 ClInstrumentWrites("hwasan-instrument-writes", 99 cl::desc("instrument write instructions"), cl::Hidden, 100 cl::init(true)); 101 102 static cl::opt<bool> ClInstrumentAtomics( 103 "hwasan-instrument-atomics", 104 cl::desc("instrument atomic instructions (rmw, cmpxchg)"), cl::Hidden, 105 cl::init(true)); 106 107 static cl::opt<bool> ClInstrumentByval("hwasan-instrument-byval", 108 cl::desc("instrument byval arguments"), 109 cl::Hidden, cl::init(true)); 110 111 static cl::opt<bool> 112 ClRecover("hwasan-recover", 113 cl::desc("Enable recovery mode (continue-after-error)."), 114 cl::Hidden, cl::init(false)); 115 116 static cl::opt<bool> ClInstrumentStack("hwasan-instrument-stack", 117 cl::desc("instrument stack (allocas)"), 118 cl::Hidden, cl::init(true)); 119 120 static cl::opt<bool> 121 ClUseStackSafety("hwasan-use-stack-safety", cl::Hidden, cl::init(true), 122 cl::Hidden, cl::desc("Use Stack Safety analysis results"), 123 cl::Optional); 124 125 static cl::opt<size_t> ClMaxLifetimes( 126 "hwasan-max-lifetimes-for-alloca", cl::Hidden, cl::init(3), 127 cl::ReallyHidden, 128 cl::desc("How many lifetime ends to handle for a single alloca."), 129 cl::Optional); 130 131 static cl::opt<bool> 132 ClUseAfterScope("hwasan-use-after-scope", 133 cl::desc("detect use after scope within function"), 134 cl::Hidden, cl::init(false)); 135 136 static cl::opt<bool> ClUARRetagToZero( 137 "hwasan-uar-retag-to-zero", 138 cl::desc("Clear alloca tags before returning from the function to allow " 139 "non-instrumented and instrumented function calls mix. When set " 140 "to false, allocas are retagged before returning from the " 141 "function to detect use after return."), 142 cl::Hidden, cl::init(true)); 143 144 static cl::opt<bool> ClGenerateTagsWithCalls( 145 "hwasan-generate-tags-with-calls", 146 cl::desc("generate new tags with runtime library calls"), cl::Hidden, 147 cl::init(false)); 148 149 static cl::opt<bool> ClGlobals("hwasan-globals", cl::desc("Instrument globals"), 150 cl::Hidden, cl::init(false)); 151 152 static cl::opt<int> ClMatchAllTag( 153 "hwasan-match-all-tag", 154 cl::desc("don't report bad accesses via pointers with this tag"), 155 cl::Hidden, cl::init(-1)); 156 157 static cl::opt<bool> 158 ClEnableKhwasan("hwasan-kernel", 159 cl::desc("Enable KernelHWAddressSanitizer instrumentation"), 160 cl::Hidden, cl::init(false)); 161 162 // These flags allow to change the shadow mapping and control how shadow memory 163 // is accessed. The shadow mapping looks like: 164 // Shadow = (Mem >> scale) + offset 165 166 static cl::opt<uint64_t> 167 ClMappingOffset("hwasan-mapping-offset", 168 cl::desc("HWASan shadow mapping offset [EXPERIMENTAL]"), 169 cl::Hidden, cl::init(0)); 170 171 static cl::opt<bool> 172 ClWithIfunc("hwasan-with-ifunc", 173 cl::desc("Access dynamic shadow through an ifunc global on " 174 "platforms that support this"), 175 cl::Hidden, cl::init(false)); 176 177 static cl::opt<bool> ClWithTls( 178 "hwasan-with-tls", 179 cl::desc("Access dynamic shadow through an thread-local pointer on " 180 "platforms that support this"), 181 cl::Hidden, cl::init(true)); 182 183 static cl::opt<bool> 184 ClRecordStackHistory("hwasan-record-stack-history", 185 cl::desc("Record stack frames with tagged allocations " 186 "in a thread-local ring buffer"), 187 cl::Hidden, cl::init(true)); 188 static cl::opt<bool> 189 ClInstrumentMemIntrinsics("hwasan-instrument-mem-intrinsics", 190 cl::desc("instrument memory intrinsics"), 191 cl::Hidden, cl::init(true)); 192 193 static cl::opt<bool> 194 ClInstrumentLandingPads("hwasan-instrument-landing-pads", 195 cl::desc("instrument landing pads"), cl::Hidden, 196 cl::init(false)); 197 198 static cl::opt<bool> ClUseShortGranules( 199 "hwasan-use-short-granules", 200 cl::desc("use short granules in allocas and outlined checks"), cl::Hidden, 201 cl::init(false)); 202 203 static cl::opt<bool> ClInstrumentPersonalityFunctions( 204 "hwasan-instrument-personality-functions", 205 cl::desc("instrument personality functions"), cl::Hidden); 206 207 static cl::opt<bool> ClInlineAllChecks("hwasan-inline-all-checks", 208 cl::desc("inline all checks"), 209 cl::Hidden, cl::init(false)); 210 211 // Enabled from clang by "-fsanitize-hwaddress-experimental-aliasing". 212 static cl::opt<bool> ClUsePageAliases("hwasan-experimental-use-page-aliases", 213 cl::desc("Use page aliasing in HWASan"), 214 cl::Hidden, cl::init(false)); 215 216 namespace { 217 218 bool shouldUsePageAliases(const Triple &TargetTriple) { 219 return ClUsePageAliases && TargetTriple.getArch() == Triple::x86_64; 220 } 221 222 bool shouldInstrumentStack(const Triple &TargetTriple) { 223 return !shouldUsePageAliases(TargetTriple) && ClInstrumentStack; 224 } 225 226 bool shouldInstrumentWithCalls(const Triple &TargetTriple) { 227 return ClInstrumentWithCalls || TargetTriple.getArch() == Triple::x86_64; 228 } 229 230 bool mightUseStackSafetyAnalysis(bool DisableOptimization) { 231 return ClUseStackSafety.getNumOccurrences() ? ClUseStackSafety 232 : !DisableOptimization; 233 } 234 235 bool shouldUseStackSafetyAnalysis(const Triple &TargetTriple, 236 bool DisableOptimization) { 237 return shouldInstrumentStack(TargetTriple) && 238 mightUseStackSafetyAnalysis(DisableOptimization); 239 } 240 241 bool shouldDetectUseAfterScope(const Triple &TargetTriple) { 242 return ClUseAfterScope && shouldInstrumentStack(TargetTriple); 243 } 244 245 /// An instrumentation pass implementing detection of addressability bugs 246 /// using tagged pointers. 247 class HWAddressSanitizer { 248 public: 249 HWAddressSanitizer(Module &M, bool CompileKernel, bool Recover, 250 const StackSafetyGlobalInfo *SSI) 251 : M(M), SSI(SSI) { 252 this->Recover = ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover; 253 this->CompileKernel = ClEnableKhwasan.getNumOccurrences() > 0 254 ? ClEnableKhwasan 255 : CompileKernel; 256 257 initializeModule(); 258 } 259 260 void setSSI(const StackSafetyGlobalInfo *S) { SSI = S; } 261 262 bool sanitizeFunction(Function &F, FunctionAnalysisManager &FAM); 263 void initializeModule(); 264 void createHwasanCtorComdat(); 265 266 void initializeCallbacks(Module &M); 267 268 Value *getOpaqueNoopCast(IRBuilder<> &IRB, Value *Val); 269 270 Value *getDynamicShadowIfunc(IRBuilder<> &IRB); 271 Value *getShadowNonTls(IRBuilder<> &IRB); 272 273 void untagPointerOperand(Instruction *I, Value *Addr); 274 Value *memToShadow(Value *Shadow, IRBuilder<> &IRB); 275 276 int64_t getAccessInfo(bool IsWrite, unsigned AccessSizeIndex); 277 void instrumentMemAccessOutline(Value *Ptr, bool IsWrite, 278 unsigned AccessSizeIndex, 279 Instruction *InsertBefore); 280 void instrumentMemAccessInline(Value *Ptr, bool IsWrite, 281 unsigned AccessSizeIndex, 282 Instruction *InsertBefore); 283 bool ignoreMemIntrinsic(MemIntrinsic *MI); 284 void instrumentMemIntrinsic(MemIntrinsic *MI); 285 bool instrumentMemAccess(InterestingMemoryOperand &O); 286 bool ignoreAccess(Instruction *Inst, Value *Ptr); 287 void getInterestingMemoryOperands( 288 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting); 289 290 bool isInterestingAlloca(const AllocaInst &AI); 291 void tagAlloca(IRBuilder<> &IRB, AllocaInst *AI, Value *Tag, size_t Size); 292 Value *tagPointer(IRBuilder<> &IRB, Type *Ty, Value *PtrLong, Value *Tag); 293 Value *untagPointer(IRBuilder<> &IRB, Value *PtrLong); 294 bool instrumentStack(memtag::StackInfo &Info, Value *StackTag, 295 const DominatorTree &DT, const PostDominatorTree &PDT, 296 const LoopInfo &LI); 297 Value *readRegister(IRBuilder<> &IRB, StringRef Name); 298 bool instrumentLandingPads(SmallVectorImpl<Instruction *> &RetVec); 299 Value *getNextTagWithCall(IRBuilder<> &IRB); 300 Value *getStackBaseTag(IRBuilder<> &IRB); 301 Value *getAllocaTag(IRBuilder<> &IRB, Value *StackTag, AllocaInst *AI, 302 unsigned AllocaNo); 303 Value *getUARTag(IRBuilder<> &IRB, Value *StackTag); 304 305 Value *getHwasanThreadSlotPtr(IRBuilder<> &IRB, Type *Ty); 306 Value *applyTagMask(IRBuilder<> &IRB, Value *OldTag); 307 unsigned retagMask(unsigned AllocaNo); 308 309 void emitPrologue(IRBuilder<> &IRB, bool WithFrameRecord); 310 311 void instrumentGlobal(GlobalVariable *GV, uint8_t Tag); 312 void instrumentGlobals(); 313 314 Value *getPC(IRBuilder<> &IRB); 315 Value *getSP(IRBuilder<> &IRB); 316 317 void instrumentPersonalityFunctions(); 318 319 private: 320 LLVMContext *C; 321 Module &M; 322 const StackSafetyGlobalInfo *SSI; 323 Triple TargetTriple; 324 FunctionCallee HWAsanMemmove, HWAsanMemcpy, HWAsanMemset; 325 FunctionCallee HWAsanHandleVfork; 326 327 /// This struct defines the shadow mapping using the rule: 328 /// shadow = (mem >> Scale) + Offset. 329 /// If InGlobal is true, then 330 /// extern char __hwasan_shadow[]; 331 /// shadow = (mem >> Scale) + &__hwasan_shadow 332 /// If InTls is true, then 333 /// extern char *__hwasan_tls; 334 /// shadow = (mem>>Scale) + align_up(__hwasan_shadow, kShadowBaseAlignment) 335 /// 336 /// If WithFrameRecord is true, then __hwasan_tls will be used to access the 337 /// ring buffer for storing stack allocations on targets that support it. 338 struct ShadowMapping { 339 int Scale; 340 uint64_t Offset; 341 bool InGlobal; 342 bool InTls; 343 bool WithFrameRecord; 344 345 void init(Triple &TargetTriple, bool InstrumentWithCalls); 346 uint64_t getObjectAlignment() const { return 1ULL << Scale; } 347 }; 348 349 ShadowMapping Mapping; 350 351 Type *VoidTy = Type::getVoidTy(M.getContext()); 352 Type *IntptrTy; 353 Type *Int8PtrTy; 354 Type *Int8Ty; 355 Type *Int32Ty; 356 Type *Int64Ty = Type::getInt64Ty(M.getContext()); 357 358 bool CompileKernel; 359 bool Recover; 360 bool OutlinedChecks; 361 bool UseShortGranules; 362 bool InstrumentLandingPads; 363 bool InstrumentWithCalls; 364 bool InstrumentStack; 365 bool DetectUseAfterScope; 366 bool UsePageAliases; 367 368 bool HasMatchAllTag = false; 369 uint8_t MatchAllTag = 0; 370 371 unsigned PointerTagShift; 372 uint64_t TagMaskByte; 373 374 Function *HwasanCtorFunction; 375 376 FunctionCallee HwasanMemoryAccessCallback[2][kNumberOfAccessSizes]; 377 FunctionCallee HwasanMemoryAccessCallbackSized[2]; 378 379 FunctionCallee HwasanTagMemoryFunc; 380 FunctionCallee HwasanGenerateTagFunc; 381 382 Constant *ShadowGlobal; 383 384 Value *ShadowBase = nullptr; 385 Value *StackBaseTag = nullptr; 386 Value *CachedSP = nullptr; 387 GlobalValue *ThreadPtrGlobal = nullptr; 388 }; 389 390 } // end anonymous namespace 391 392 PreservedAnalyses HWAddressSanitizerPass::run(Module &M, 393 ModuleAnalysisManager &MAM) { 394 const StackSafetyGlobalInfo *SSI = nullptr; 395 auto TargetTriple = llvm::Triple(M.getTargetTriple()); 396 if (shouldUseStackSafetyAnalysis(TargetTriple, Options.DisableOptimization)) 397 SSI = &MAM.getResult<StackSafetyGlobalAnalysis>(M); 398 399 HWAddressSanitizer HWASan(M, Options.CompileKernel, Options.Recover, SSI); 400 bool Modified = false; 401 auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 402 for (Function &F : M) 403 Modified |= HWASan.sanitizeFunction(F, FAM); 404 if (Modified) 405 return PreservedAnalyses::none(); 406 return PreservedAnalyses::all(); 407 } 408 void HWAddressSanitizerPass::printPipeline( 409 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { 410 static_cast<PassInfoMixin<HWAddressSanitizerPass> *>(this)->printPipeline( 411 OS, MapClassName2PassName); 412 OS << "<"; 413 if (Options.CompileKernel) 414 OS << "kernel;"; 415 if (Options.Recover) 416 OS << "recover"; 417 OS << ">"; 418 } 419 420 void HWAddressSanitizer::createHwasanCtorComdat() { 421 std::tie(HwasanCtorFunction, std::ignore) = 422 getOrCreateSanitizerCtorAndInitFunctions( 423 M, kHwasanModuleCtorName, kHwasanInitName, 424 /*InitArgTypes=*/{}, 425 /*InitArgs=*/{}, 426 // This callback is invoked when the functions are created the first 427 // time. Hook them into the global ctors list in that case: 428 [&](Function *Ctor, FunctionCallee) { 429 Comdat *CtorComdat = M.getOrInsertComdat(kHwasanModuleCtorName); 430 Ctor->setComdat(CtorComdat); 431 appendToGlobalCtors(M, Ctor, 0, Ctor); 432 }); 433 434 // Create a note that contains pointers to the list of global 435 // descriptors. Adding a note to the output file will cause the linker to 436 // create a PT_NOTE program header pointing to the note that we can use to 437 // find the descriptor list starting from the program headers. A function 438 // provided by the runtime initializes the shadow memory for the globals by 439 // accessing the descriptor list via the note. The dynamic loader needs to 440 // call this function whenever a library is loaded. 441 // 442 // The reason why we use a note for this instead of a more conventional 443 // approach of having a global constructor pass a descriptor list pointer to 444 // the runtime is because of an order of initialization problem. With 445 // constructors we can encounter the following problematic scenario: 446 // 447 // 1) library A depends on library B and also interposes one of B's symbols 448 // 2) B's constructors are called before A's (as required for correctness) 449 // 3) during construction, B accesses one of its "own" globals (actually 450 // interposed by A) and triggers a HWASAN failure due to the initialization 451 // for A not having happened yet 452 // 453 // Even without interposition it is possible to run into similar situations in 454 // cases where two libraries mutually depend on each other. 455 // 456 // We only need one note per binary, so put everything for the note in a 457 // comdat. This needs to be a comdat with an .init_array section to prevent 458 // newer versions of lld from discarding the note. 459 // 460 // Create the note even if we aren't instrumenting globals. This ensures that 461 // binaries linked from object files with both instrumented and 462 // non-instrumented globals will end up with a note, even if a comdat from an 463 // object file with non-instrumented globals is selected. The note is harmless 464 // if the runtime doesn't support it, since it will just be ignored. 465 Comdat *NoteComdat = M.getOrInsertComdat(kHwasanModuleCtorName); 466 467 Type *Int8Arr0Ty = ArrayType::get(Int8Ty, 0); 468 auto Start = 469 new GlobalVariable(M, Int8Arr0Ty, true, GlobalVariable::ExternalLinkage, 470 nullptr, "__start_hwasan_globals"); 471 Start->setVisibility(GlobalValue::HiddenVisibility); 472 Start->setDSOLocal(true); 473 auto Stop = 474 new GlobalVariable(M, Int8Arr0Ty, true, GlobalVariable::ExternalLinkage, 475 nullptr, "__stop_hwasan_globals"); 476 Stop->setVisibility(GlobalValue::HiddenVisibility); 477 Stop->setDSOLocal(true); 478 479 // Null-terminated so actually 8 bytes, which are required in order to align 480 // the note properly. 481 auto *Name = ConstantDataArray::get(*C, "LLVM\0\0\0"); 482 483 auto *NoteTy = StructType::get(Int32Ty, Int32Ty, Int32Ty, Name->getType(), 484 Int32Ty, Int32Ty); 485 auto *Note = 486 new GlobalVariable(M, NoteTy, /*isConstant=*/true, 487 GlobalValue::PrivateLinkage, nullptr, kHwasanNoteName); 488 Note->setSection(".note.hwasan.globals"); 489 Note->setComdat(NoteComdat); 490 Note->setAlignment(Align(4)); 491 Note->setDSOLocal(true); 492 493 // The pointers in the note need to be relative so that the note ends up being 494 // placed in rodata, which is the standard location for notes. 495 auto CreateRelPtr = [&](Constant *Ptr) { 496 return ConstantExpr::getTrunc( 497 ConstantExpr::getSub(ConstantExpr::getPtrToInt(Ptr, Int64Ty), 498 ConstantExpr::getPtrToInt(Note, Int64Ty)), 499 Int32Ty); 500 }; 501 Note->setInitializer(ConstantStruct::getAnon( 502 {ConstantInt::get(Int32Ty, 8), // n_namesz 503 ConstantInt::get(Int32Ty, 8), // n_descsz 504 ConstantInt::get(Int32Ty, ELF::NT_LLVM_HWASAN_GLOBALS), // n_type 505 Name, CreateRelPtr(Start), CreateRelPtr(Stop)})); 506 appendToCompilerUsed(M, Note); 507 508 // Create a zero-length global in hwasan_globals so that the linker will 509 // always create start and stop symbols. 510 auto Dummy = new GlobalVariable( 511 M, Int8Arr0Ty, /*isConstantGlobal*/ true, GlobalVariable::PrivateLinkage, 512 Constant::getNullValue(Int8Arr0Ty), "hwasan.dummy.global"); 513 Dummy->setSection("hwasan_globals"); 514 Dummy->setComdat(NoteComdat); 515 Dummy->setMetadata(LLVMContext::MD_associated, 516 MDNode::get(*C, ValueAsMetadata::get(Note))); 517 appendToCompilerUsed(M, Dummy); 518 } 519 520 /// Module-level initialization. 521 /// 522 /// inserts a call to __hwasan_init to the module's constructor list. 523 void HWAddressSanitizer::initializeModule() { 524 LLVM_DEBUG(dbgs() << "Init " << M.getName() << "\n"); 525 auto &DL = M.getDataLayout(); 526 527 TargetTriple = Triple(M.getTargetTriple()); 528 529 // x86_64 currently has two modes: 530 // - Intel LAM (default) 531 // - pointer aliasing (heap only) 532 bool IsX86_64 = TargetTriple.getArch() == Triple::x86_64; 533 UsePageAliases = shouldUsePageAliases(TargetTriple); 534 InstrumentWithCalls = shouldInstrumentWithCalls(TargetTriple); 535 InstrumentStack = shouldInstrumentStack(TargetTriple); 536 DetectUseAfterScope = shouldDetectUseAfterScope(TargetTriple); 537 PointerTagShift = IsX86_64 ? 57 : 56; 538 TagMaskByte = IsX86_64 ? 0x3F : 0xFF; 539 540 Mapping.init(TargetTriple, InstrumentWithCalls); 541 542 C = &(M.getContext()); 543 IRBuilder<> IRB(*C); 544 IntptrTy = IRB.getIntPtrTy(DL); 545 Int8PtrTy = IRB.getInt8PtrTy(); 546 Int8Ty = IRB.getInt8Ty(); 547 Int32Ty = IRB.getInt32Ty(); 548 549 HwasanCtorFunction = nullptr; 550 551 // Older versions of Android do not have the required runtime support for 552 // short granules, global or personality function instrumentation. On other 553 // platforms we currently require using the latest version of the runtime. 554 bool NewRuntime = 555 !TargetTriple.isAndroid() || !TargetTriple.isAndroidVersionLT(30); 556 557 UseShortGranules = 558 ClUseShortGranules.getNumOccurrences() ? ClUseShortGranules : NewRuntime; 559 OutlinedChecks = 560 TargetTriple.isAArch64() && TargetTriple.isOSBinFormatELF() && 561 (ClInlineAllChecks.getNumOccurrences() ? !ClInlineAllChecks : !Recover); 562 563 if (ClMatchAllTag.getNumOccurrences()) { 564 if (ClMatchAllTag != -1) { 565 HasMatchAllTag = true; 566 MatchAllTag = ClMatchAllTag & 0xFF; 567 } 568 } else if (CompileKernel) { 569 HasMatchAllTag = true; 570 MatchAllTag = 0xFF; 571 } 572 573 // If we don't have personality function support, fall back to landing pads. 574 InstrumentLandingPads = ClInstrumentLandingPads.getNumOccurrences() 575 ? ClInstrumentLandingPads 576 : !NewRuntime; 577 578 if (!CompileKernel) { 579 createHwasanCtorComdat(); 580 bool InstrumentGlobals = 581 ClGlobals.getNumOccurrences() ? ClGlobals : NewRuntime; 582 583 if (InstrumentGlobals && !UsePageAliases) 584 instrumentGlobals(); 585 586 bool InstrumentPersonalityFunctions = 587 ClInstrumentPersonalityFunctions.getNumOccurrences() 588 ? ClInstrumentPersonalityFunctions 589 : NewRuntime; 590 if (InstrumentPersonalityFunctions) 591 instrumentPersonalityFunctions(); 592 } 593 594 if (!TargetTriple.isAndroid()) { 595 Constant *C = M.getOrInsertGlobal("__hwasan_tls", IntptrTy, [&] { 596 auto *GV = new GlobalVariable(M, IntptrTy, /*isConstant=*/false, 597 GlobalValue::ExternalLinkage, nullptr, 598 "__hwasan_tls", nullptr, 599 GlobalVariable::InitialExecTLSModel); 600 appendToCompilerUsed(M, GV); 601 return GV; 602 }); 603 ThreadPtrGlobal = cast<GlobalVariable>(C); 604 } 605 } 606 607 void HWAddressSanitizer::initializeCallbacks(Module &M) { 608 IRBuilder<> IRB(*C); 609 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) { 610 const std::string TypeStr = AccessIsWrite ? "store" : "load"; 611 const std::string EndingStr = Recover ? "_noabort" : ""; 612 613 HwasanMemoryAccessCallbackSized[AccessIsWrite] = M.getOrInsertFunction( 614 ClMemoryAccessCallbackPrefix + TypeStr + "N" + EndingStr, 615 FunctionType::get(IRB.getVoidTy(), {IntptrTy, IntptrTy}, false)); 616 617 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes; 618 AccessSizeIndex++) { 619 HwasanMemoryAccessCallback[AccessIsWrite][AccessSizeIndex] = 620 M.getOrInsertFunction( 621 ClMemoryAccessCallbackPrefix + TypeStr + 622 itostr(1ULL << AccessSizeIndex) + EndingStr, 623 FunctionType::get(IRB.getVoidTy(), {IntptrTy}, false)); 624 } 625 } 626 627 HwasanTagMemoryFunc = M.getOrInsertFunction( 628 "__hwasan_tag_memory", IRB.getVoidTy(), Int8PtrTy, Int8Ty, IntptrTy); 629 HwasanGenerateTagFunc = 630 M.getOrInsertFunction("__hwasan_generate_tag", Int8Ty); 631 632 ShadowGlobal = M.getOrInsertGlobal("__hwasan_shadow", 633 ArrayType::get(IRB.getInt8Ty(), 0)); 634 635 const std::string MemIntrinCallbackPrefix = 636 (CompileKernel && !ClKasanMemIntrinCallbackPrefix) 637 ? std::string("") 638 : ClMemoryAccessCallbackPrefix; 639 HWAsanMemmove = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memmove", 640 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 641 IRB.getInt8PtrTy(), IntptrTy); 642 HWAsanMemcpy = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memcpy", 643 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 644 IRB.getInt8PtrTy(), IntptrTy); 645 HWAsanMemset = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memset", 646 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 647 IRB.getInt32Ty(), IntptrTy); 648 649 HWAsanHandleVfork = 650 M.getOrInsertFunction("__hwasan_handle_vfork", IRB.getVoidTy(), IntptrTy); 651 } 652 653 Value *HWAddressSanitizer::getOpaqueNoopCast(IRBuilder<> &IRB, Value *Val) { 654 // An empty inline asm with input reg == output reg. 655 // An opaque no-op cast, basically. 656 // This prevents code bloat as a result of rematerializing trivial definitions 657 // such as constants or global addresses at every load and store. 658 InlineAsm *Asm = 659 InlineAsm::get(FunctionType::get(Int8PtrTy, {Val->getType()}, false), 660 StringRef(""), StringRef("=r,0"), 661 /*hasSideEffects=*/false); 662 return IRB.CreateCall(Asm, {Val}, ".hwasan.shadow"); 663 } 664 665 Value *HWAddressSanitizer::getDynamicShadowIfunc(IRBuilder<> &IRB) { 666 return getOpaqueNoopCast(IRB, ShadowGlobal); 667 } 668 669 Value *HWAddressSanitizer::getShadowNonTls(IRBuilder<> &IRB) { 670 if (Mapping.Offset != kDynamicShadowSentinel) 671 return getOpaqueNoopCast( 672 IRB, ConstantExpr::getIntToPtr( 673 ConstantInt::get(IntptrTy, Mapping.Offset), Int8PtrTy)); 674 675 if (Mapping.InGlobal) { 676 return getDynamicShadowIfunc(IRB); 677 } else { 678 Value *GlobalDynamicAddress = 679 IRB.GetInsertBlock()->getParent()->getParent()->getOrInsertGlobal( 680 kHwasanShadowMemoryDynamicAddress, Int8PtrTy); 681 return IRB.CreateLoad(Int8PtrTy, GlobalDynamicAddress); 682 } 683 } 684 685 bool HWAddressSanitizer::ignoreAccess(Instruction *Inst, Value *Ptr) { 686 // Do not instrument acesses from different address spaces; we cannot deal 687 // with them. 688 Type *PtrTy = cast<PointerType>(Ptr->getType()->getScalarType()); 689 if (PtrTy->getPointerAddressSpace() != 0) 690 return true; 691 692 // Ignore swifterror addresses. 693 // swifterror memory addresses are mem2reg promoted by instruction 694 // selection. As such they cannot have regular uses like an instrumentation 695 // function and it makes no sense to track them as memory. 696 if (Ptr->isSwiftError()) 697 return true; 698 699 if (findAllocaForValue(Ptr)) { 700 if (!InstrumentStack) 701 return true; 702 if (SSI && SSI->stackAccessIsSafe(*Inst)) 703 return true; 704 } 705 return false; 706 } 707 708 void HWAddressSanitizer::getInterestingMemoryOperands( 709 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting) { 710 // Skip memory accesses inserted by another instrumentation. 711 if (I->hasMetadata(LLVMContext::MD_nosanitize)) 712 return; 713 714 // Do not instrument the load fetching the dynamic shadow address. 715 if (ShadowBase == I) 716 return; 717 718 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 719 if (!ClInstrumentReads || ignoreAccess(I, LI->getPointerOperand())) 720 return; 721 Interesting.emplace_back(I, LI->getPointerOperandIndex(), false, 722 LI->getType(), LI->getAlign()); 723 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 724 if (!ClInstrumentWrites || ignoreAccess(I, SI->getPointerOperand())) 725 return; 726 Interesting.emplace_back(I, SI->getPointerOperandIndex(), true, 727 SI->getValueOperand()->getType(), SI->getAlign()); 728 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) { 729 if (!ClInstrumentAtomics || ignoreAccess(I, RMW->getPointerOperand())) 730 return; 731 Interesting.emplace_back(I, RMW->getPointerOperandIndex(), true, 732 RMW->getValOperand()->getType(), None); 733 } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) { 734 if (!ClInstrumentAtomics || ignoreAccess(I, XCHG->getPointerOperand())) 735 return; 736 Interesting.emplace_back(I, XCHG->getPointerOperandIndex(), true, 737 XCHG->getCompareOperand()->getType(), None); 738 } else if (auto CI = dyn_cast<CallInst>(I)) { 739 for (unsigned ArgNo = 0; ArgNo < CI->arg_size(); ArgNo++) { 740 if (!ClInstrumentByval || !CI->isByValArgument(ArgNo) || 741 ignoreAccess(I, CI->getArgOperand(ArgNo))) 742 continue; 743 Type *Ty = CI->getParamByValType(ArgNo); 744 Interesting.emplace_back(I, ArgNo, false, Ty, Align(1)); 745 } 746 } 747 } 748 749 static unsigned getPointerOperandIndex(Instruction *I) { 750 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 751 return LI->getPointerOperandIndex(); 752 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 753 return SI->getPointerOperandIndex(); 754 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) 755 return RMW->getPointerOperandIndex(); 756 if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) 757 return XCHG->getPointerOperandIndex(); 758 report_fatal_error("Unexpected instruction"); 759 return -1; 760 } 761 762 static size_t TypeSizeToSizeIndex(uint32_t TypeSize) { 763 size_t Res = countTrailingZeros(TypeSize / 8); 764 assert(Res < kNumberOfAccessSizes); 765 return Res; 766 } 767 768 void HWAddressSanitizer::untagPointerOperand(Instruction *I, Value *Addr) { 769 if (TargetTriple.isAArch64() || TargetTriple.getArch() == Triple::x86_64) 770 return; 771 772 IRBuilder<> IRB(I); 773 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy); 774 Value *UntaggedPtr = 775 IRB.CreateIntToPtr(untagPointer(IRB, AddrLong), Addr->getType()); 776 I->setOperand(getPointerOperandIndex(I), UntaggedPtr); 777 } 778 779 Value *HWAddressSanitizer::memToShadow(Value *Mem, IRBuilder<> &IRB) { 780 // Mem >> Scale 781 Value *Shadow = IRB.CreateLShr(Mem, Mapping.Scale); 782 if (Mapping.Offset == 0) 783 return IRB.CreateIntToPtr(Shadow, Int8PtrTy); 784 // (Mem >> Scale) + Offset 785 return IRB.CreateGEP(Int8Ty, ShadowBase, Shadow); 786 } 787 788 int64_t HWAddressSanitizer::getAccessInfo(bool IsWrite, 789 unsigned AccessSizeIndex) { 790 return (CompileKernel << HWASanAccessInfo::CompileKernelShift) + 791 (HasMatchAllTag << HWASanAccessInfo::HasMatchAllShift) + 792 (MatchAllTag << HWASanAccessInfo::MatchAllShift) + 793 (Recover << HWASanAccessInfo::RecoverShift) + 794 (IsWrite << HWASanAccessInfo::IsWriteShift) + 795 (AccessSizeIndex << HWASanAccessInfo::AccessSizeShift); 796 } 797 798 void HWAddressSanitizer::instrumentMemAccessOutline(Value *Ptr, bool IsWrite, 799 unsigned AccessSizeIndex, 800 Instruction *InsertBefore) { 801 assert(!UsePageAliases); 802 const int64_t AccessInfo = getAccessInfo(IsWrite, AccessSizeIndex); 803 IRBuilder<> IRB(InsertBefore); 804 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 805 Ptr = IRB.CreateBitCast(Ptr, Int8PtrTy); 806 IRB.CreateCall(Intrinsic::getDeclaration( 807 M, UseShortGranules 808 ? Intrinsic::hwasan_check_memaccess_shortgranules 809 : Intrinsic::hwasan_check_memaccess), 810 {ShadowBase, Ptr, ConstantInt::get(Int32Ty, AccessInfo)}); 811 } 812 813 void HWAddressSanitizer::instrumentMemAccessInline(Value *Ptr, bool IsWrite, 814 unsigned AccessSizeIndex, 815 Instruction *InsertBefore) { 816 assert(!UsePageAliases); 817 const int64_t AccessInfo = getAccessInfo(IsWrite, AccessSizeIndex); 818 IRBuilder<> IRB(InsertBefore); 819 820 Value *PtrLong = IRB.CreatePointerCast(Ptr, IntptrTy); 821 Value *PtrTag = IRB.CreateTrunc(IRB.CreateLShr(PtrLong, PointerTagShift), 822 IRB.getInt8Ty()); 823 Value *AddrLong = untagPointer(IRB, PtrLong); 824 Value *Shadow = memToShadow(AddrLong, IRB); 825 Value *MemTag = IRB.CreateLoad(Int8Ty, Shadow); 826 Value *TagMismatch = IRB.CreateICmpNE(PtrTag, MemTag); 827 828 if (HasMatchAllTag) { 829 Value *TagNotIgnored = IRB.CreateICmpNE( 830 PtrTag, ConstantInt::get(PtrTag->getType(), MatchAllTag)); 831 TagMismatch = IRB.CreateAnd(TagMismatch, TagNotIgnored); 832 } 833 834 Instruction *CheckTerm = 835 SplitBlockAndInsertIfThen(TagMismatch, InsertBefore, false, 836 MDBuilder(*C).createBranchWeights(1, 100000)); 837 838 IRB.SetInsertPoint(CheckTerm); 839 Value *OutOfShortGranuleTagRange = 840 IRB.CreateICmpUGT(MemTag, ConstantInt::get(Int8Ty, 15)); 841 Instruction *CheckFailTerm = 842 SplitBlockAndInsertIfThen(OutOfShortGranuleTagRange, CheckTerm, !Recover, 843 MDBuilder(*C).createBranchWeights(1, 100000)); 844 845 IRB.SetInsertPoint(CheckTerm); 846 Value *PtrLowBits = IRB.CreateTrunc(IRB.CreateAnd(PtrLong, 15), Int8Ty); 847 PtrLowBits = IRB.CreateAdd( 848 PtrLowBits, ConstantInt::get(Int8Ty, (1 << AccessSizeIndex) - 1)); 849 Value *PtrLowBitsOOB = IRB.CreateICmpUGE(PtrLowBits, MemTag); 850 SplitBlockAndInsertIfThen(PtrLowBitsOOB, CheckTerm, false, 851 MDBuilder(*C).createBranchWeights(1, 100000), 852 (DomTreeUpdater *)nullptr, nullptr, 853 CheckFailTerm->getParent()); 854 855 IRB.SetInsertPoint(CheckTerm); 856 Value *InlineTagAddr = IRB.CreateOr(AddrLong, 15); 857 InlineTagAddr = IRB.CreateIntToPtr(InlineTagAddr, Int8PtrTy); 858 Value *InlineTag = IRB.CreateLoad(Int8Ty, InlineTagAddr); 859 Value *InlineTagMismatch = IRB.CreateICmpNE(PtrTag, InlineTag); 860 SplitBlockAndInsertIfThen(InlineTagMismatch, CheckTerm, false, 861 MDBuilder(*C).createBranchWeights(1, 100000), 862 (DomTreeUpdater *)nullptr, nullptr, 863 CheckFailTerm->getParent()); 864 865 IRB.SetInsertPoint(CheckFailTerm); 866 InlineAsm *Asm; 867 switch (TargetTriple.getArch()) { 868 case Triple::x86_64: 869 // The signal handler will find the data address in rdi. 870 Asm = InlineAsm::get( 871 FunctionType::get(IRB.getVoidTy(), {PtrLong->getType()}, false), 872 "int3\nnopl " + 873 itostr(0x40 + (AccessInfo & HWASanAccessInfo::RuntimeMask)) + 874 "(%rax)", 875 "{rdi}", 876 /*hasSideEffects=*/true); 877 break; 878 case Triple::aarch64: 879 case Triple::aarch64_be: 880 // The signal handler will find the data address in x0. 881 Asm = InlineAsm::get( 882 FunctionType::get(IRB.getVoidTy(), {PtrLong->getType()}, false), 883 "brk #" + itostr(0x900 + (AccessInfo & HWASanAccessInfo::RuntimeMask)), 884 "{x0}", 885 /*hasSideEffects=*/true); 886 break; 887 default: 888 report_fatal_error("unsupported architecture"); 889 } 890 IRB.CreateCall(Asm, PtrLong); 891 if (Recover) 892 cast<BranchInst>(CheckFailTerm)->setSuccessor(0, CheckTerm->getParent()); 893 } 894 895 bool HWAddressSanitizer::ignoreMemIntrinsic(MemIntrinsic *MI) { 896 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { 897 return (!ClInstrumentWrites || ignoreAccess(MTI, MTI->getDest())) && 898 (!ClInstrumentReads || ignoreAccess(MTI, MTI->getSource())); 899 } 900 if (isa<MemSetInst>(MI)) 901 return !ClInstrumentWrites || ignoreAccess(MI, MI->getDest()); 902 return false; 903 } 904 905 void HWAddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) { 906 IRBuilder<> IRB(MI); 907 if (isa<MemTransferInst>(MI)) { 908 IRB.CreateCall( 909 isa<MemMoveInst>(MI) ? HWAsanMemmove : HWAsanMemcpy, 910 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()), 911 IRB.CreatePointerCast(MI->getOperand(1), IRB.getInt8PtrTy()), 912 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)}); 913 } else if (isa<MemSetInst>(MI)) { 914 IRB.CreateCall( 915 HWAsanMemset, 916 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()), 917 IRB.CreateIntCast(MI->getOperand(1), IRB.getInt32Ty(), false), 918 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)}); 919 } 920 MI->eraseFromParent(); 921 } 922 923 bool HWAddressSanitizer::instrumentMemAccess(InterestingMemoryOperand &O) { 924 Value *Addr = O.getPtr(); 925 926 LLVM_DEBUG(dbgs() << "Instrumenting: " << O.getInsn() << "\n"); 927 928 if (O.MaybeMask) 929 return false; // FIXME 930 931 IRBuilder<> IRB(O.getInsn()); 932 if (isPowerOf2_64(O.TypeSize) && 933 (O.TypeSize / 8 <= (1ULL << (kNumberOfAccessSizes - 1))) && 934 (!O.Alignment || *O.Alignment >= (1ULL << Mapping.Scale) || 935 *O.Alignment >= O.TypeSize / 8)) { 936 size_t AccessSizeIndex = TypeSizeToSizeIndex(O.TypeSize); 937 if (InstrumentWithCalls) { 938 IRB.CreateCall(HwasanMemoryAccessCallback[O.IsWrite][AccessSizeIndex], 939 IRB.CreatePointerCast(Addr, IntptrTy)); 940 } else if (OutlinedChecks) { 941 instrumentMemAccessOutline(Addr, O.IsWrite, AccessSizeIndex, O.getInsn()); 942 } else { 943 instrumentMemAccessInline(Addr, O.IsWrite, AccessSizeIndex, O.getInsn()); 944 } 945 } else { 946 IRB.CreateCall(HwasanMemoryAccessCallbackSized[O.IsWrite], 947 {IRB.CreatePointerCast(Addr, IntptrTy), 948 ConstantInt::get(IntptrTy, O.TypeSize / 8)}); 949 } 950 untagPointerOperand(O.getInsn(), Addr); 951 952 return true; 953 } 954 955 void HWAddressSanitizer::tagAlloca(IRBuilder<> &IRB, AllocaInst *AI, Value *Tag, 956 size_t Size) { 957 size_t AlignedSize = alignTo(Size, Mapping.getObjectAlignment()); 958 if (!UseShortGranules) 959 Size = AlignedSize; 960 961 Value *JustTag = IRB.CreateTrunc(Tag, IRB.getInt8Ty()); 962 if (InstrumentWithCalls) { 963 IRB.CreateCall(HwasanTagMemoryFunc, 964 {IRB.CreatePointerCast(AI, Int8PtrTy), JustTag, 965 ConstantInt::get(IntptrTy, AlignedSize)}); 966 } else { 967 size_t ShadowSize = Size >> Mapping.Scale; 968 Value *ShadowPtr = memToShadow(IRB.CreatePointerCast(AI, IntptrTy), IRB); 969 // If this memset is not inlined, it will be intercepted in the hwasan 970 // runtime library. That's OK, because the interceptor skips the checks if 971 // the address is in the shadow region. 972 // FIXME: the interceptor is not as fast as real memset. Consider lowering 973 // llvm.memset right here into either a sequence of stores, or a call to 974 // hwasan_tag_memory. 975 if (ShadowSize) 976 IRB.CreateMemSet(ShadowPtr, JustTag, ShadowSize, Align(1)); 977 if (Size != AlignedSize) { 978 IRB.CreateStore( 979 ConstantInt::get(Int8Ty, Size % Mapping.getObjectAlignment()), 980 IRB.CreateConstGEP1_32(Int8Ty, ShadowPtr, ShadowSize)); 981 IRB.CreateStore(JustTag, IRB.CreateConstGEP1_32( 982 Int8Ty, IRB.CreateBitCast(AI, Int8PtrTy), 983 AlignedSize - 1)); 984 } 985 } 986 } 987 988 unsigned HWAddressSanitizer::retagMask(unsigned AllocaNo) { 989 if (TargetTriple.getArch() == Triple::x86_64) 990 return AllocaNo & TagMaskByte; 991 992 // A list of 8-bit numbers that have at most one run of non-zero bits. 993 // x = x ^ (mask << 56) can be encoded as a single armv8 instruction for these 994 // masks. 995 // The list does not include the value 255, which is used for UAR. 996 // 997 // Because we are more likely to use earlier elements of this list than later 998 // ones, it is sorted in increasing order of probability of collision with a 999 // mask allocated (temporally) nearby. The program that generated this list 1000 // can be found at: 1001 // https://github.com/google/sanitizers/blob/master/hwaddress-sanitizer/sort_masks.py 1002 static unsigned FastMasks[] = {0, 128, 64, 192, 32, 96, 224, 112, 240, 1003 48, 16, 120, 248, 56, 24, 8, 124, 252, 1004 60, 28, 12, 4, 126, 254, 62, 30, 14, 1005 6, 2, 127, 63, 31, 15, 7, 3, 1}; 1006 return FastMasks[AllocaNo % (sizeof(FastMasks) / sizeof(FastMasks[0]))]; 1007 } 1008 1009 Value *HWAddressSanitizer::applyTagMask(IRBuilder<> &IRB, Value *OldTag) { 1010 if (TargetTriple.getArch() == Triple::x86_64) { 1011 Constant *TagMask = ConstantInt::get(IntptrTy, TagMaskByte); 1012 Value *NewTag = IRB.CreateAnd(OldTag, TagMask); 1013 return NewTag; 1014 } 1015 // aarch64 uses 8-bit tags, so no mask is needed. 1016 return OldTag; 1017 } 1018 1019 Value *HWAddressSanitizer::getNextTagWithCall(IRBuilder<> &IRB) { 1020 return IRB.CreateZExt(IRB.CreateCall(HwasanGenerateTagFunc), IntptrTy); 1021 } 1022 1023 Value *HWAddressSanitizer::getStackBaseTag(IRBuilder<> &IRB) { 1024 if (ClGenerateTagsWithCalls) 1025 return getNextTagWithCall(IRB); 1026 if (StackBaseTag) 1027 return StackBaseTag; 1028 // Extract some entropy from the stack pointer for the tags. 1029 // Take bits 20..28 (ASLR entropy) and xor with bits 0..8 (these differ 1030 // between functions). 1031 Value *StackPointerLong = getSP(IRB); 1032 Value *StackTag = 1033 applyTagMask(IRB, IRB.CreateXor(StackPointerLong, 1034 IRB.CreateLShr(StackPointerLong, 20))); 1035 StackTag->setName("hwasan.stack.base.tag"); 1036 return StackTag; 1037 } 1038 1039 Value *HWAddressSanitizer::getAllocaTag(IRBuilder<> &IRB, Value *StackTag, 1040 AllocaInst *AI, unsigned AllocaNo) { 1041 if (ClGenerateTagsWithCalls) 1042 return getNextTagWithCall(IRB); 1043 return IRB.CreateXor(StackTag, 1044 ConstantInt::get(IntptrTy, retagMask(AllocaNo))); 1045 } 1046 1047 Value *HWAddressSanitizer::getUARTag(IRBuilder<> &IRB, Value *StackTag) { 1048 if (ClUARRetagToZero) 1049 return ConstantInt::get(IntptrTy, 0); 1050 if (ClGenerateTagsWithCalls) 1051 return getNextTagWithCall(IRB); 1052 return IRB.CreateXor(StackTag, ConstantInt::get(IntptrTy, TagMaskByte)); 1053 } 1054 1055 // Add a tag to an address. 1056 Value *HWAddressSanitizer::tagPointer(IRBuilder<> &IRB, Type *Ty, 1057 Value *PtrLong, Value *Tag) { 1058 assert(!UsePageAliases); 1059 Value *TaggedPtrLong; 1060 if (CompileKernel) { 1061 // Kernel addresses have 0xFF in the most significant byte. 1062 Value *ShiftedTag = 1063 IRB.CreateOr(IRB.CreateShl(Tag, PointerTagShift), 1064 ConstantInt::get(IntptrTy, (1ULL << PointerTagShift) - 1)); 1065 TaggedPtrLong = IRB.CreateAnd(PtrLong, ShiftedTag); 1066 } else { 1067 // Userspace can simply do OR (tag << PointerTagShift); 1068 Value *ShiftedTag = IRB.CreateShl(Tag, PointerTagShift); 1069 TaggedPtrLong = IRB.CreateOr(PtrLong, ShiftedTag); 1070 } 1071 return IRB.CreateIntToPtr(TaggedPtrLong, Ty); 1072 } 1073 1074 // Remove tag from an address. 1075 Value *HWAddressSanitizer::untagPointer(IRBuilder<> &IRB, Value *PtrLong) { 1076 assert(!UsePageAliases); 1077 Value *UntaggedPtrLong; 1078 if (CompileKernel) { 1079 // Kernel addresses have 0xFF in the most significant byte. 1080 UntaggedPtrLong = 1081 IRB.CreateOr(PtrLong, ConstantInt::get(PtrLong->getType(), 1082 0xFFULL << PointerTagShift)); 1083 } else { 1084 // Userspace addresses have 0x00. 1085 UntaggedPtrLong = 1086 IRB.CreateAnd(PtrLong, ConstantInt::get(PtrLong->getType(), 1087 ~(0xFFULL << PointerTagShift))); 1088 } 1089 return UntaggedPtrLong; 1090 } 1091 1092 Value *HWAddressSanitizer::getHwasanThreadSlotPtr(IRBuilder<> &IRB, Type *Ty) { 1093 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1094 if (TargetTriple.isAArch64() && TargetTriple.isAndroid()) { 1095 // Android provides a fixed TLS slot for sanitizers. See TLS_SLOT_SANITIZER 1096 // in Bionic's libc/private/bionic_tls.h. 1097 Function *ThreadPointerFunc = 1098 Intrinsic::getDeclaration(M, Intrinsic::thread_pointer); 1099 Value *SlotPtr = IRB.CreatePointerCast( 1100 IRB.CreateConstGEP1_32(IRB.getInt8Ty(), 1101 IRB.CreateCall(ThreadPointerFunc), 0x30), 1102 Ty->getPointerTo(0)); 1103 return SlotPtr; 1104 } 1105 if (ThreadPtrGlobal) 1106 return ThreadPtrGlobal; 1107 1108 return nullptr; 1109 } 1110 1111 Value *HWAddressSanitizer::getPC(IRBuilder<> &IRB) { 1112 if (TargetTriple.getArch() == Triple::aarch64) 1113 return readRegister(IRB, "pc"); 1114 else 1115 return IRB.CreatePtrToInt(IRB.GetInsertBlock()->getParent(), IntptrTy); 1116 } 1117 1118 Value *HWAddressSanitizer::getSP(IRBuilder<> &IRB) { 1119 if (!CachedSP) { 1120 // FIXME: use addressofreturnaddress (but implement it in aarch64 backend 1121 // first). 1122 Function *F = IRB.GetInsertBlock()->getParent(); 1123 Module *M = F->getParent(); 1124 auto GetStackPointerFn = Intrinsic::getDeclaration( 1125 M, Intrinsic::frameaddress, 1126 IRB.getInt8PtrTy(M->getDataLayout().getAllocaAddrSpace())); 1127 CachedSP = IRB.CreatePtrToInt( 1128 IRB.CreateCall(GetStackPointerFn, 1129 {Constant::getNullValue(IRB.getInt32Ty())}), 1130 IntptrTy); 1131 } 1132 return CachedSP; 1133 } 1134 1135 void HWAddressSanitizer::emitPrologue(IRBuilder<> &IRB, bool WithFrameRecord) { 1136 if (!Mapping.InTls) 1137 ShadowBase = getShadowNonTls(IRB); 1138 else if (!WithFrameRecord && TargetTriple.isAndroid()) 1139 ShadowBase = getDynamicShadowIfunc(IRB); 1140 1141 if (!WithFrameRecord && ShadowBase) 1142 return; 1143 1144 Value *SlotPtr = getHwasanThreadSlotPtr(IRB, IntptrTy); 1145 assert(SlotPtr); 1146 1147 Value *ThreadLong = IRB.CreateLoad(IntptrTy, SlotPtr); 1148 // Extract the address field from ThreadLong. Unnecessary on AArch64 with TBI. 1149 Value *ThreadLongMaybeUntagged = 1150 TargetTriple.isAArch64() ? ThreadLong : untagPointer(IRB, ThreadLong); 1151 1152 if (WithFrameRecord) { 1153 StackBaseTag = IRB.CreateAShr(ThreadLong, 3); 1154 1155 // Prepare ring buffer data. 1156 Value *PC = getPC(IRB); 1157 Value *SP = getSP(IRB); 1158 1159 // Mix SP and PC. 1160 // Assumptions: 1161 // PC is 0x0000PPPPPPPPPPPP (48 bits are meaningful, others are zero) 1162 // SP is 0xsssssssssssSSSS0 (4 lower bits are zero) 1163 // We only really need ~20 lower non-zero bits (SSSS), so we mix like this: 1164 // 0xSSSSPPPPPPPPPPPP 1165 SP = IRB.CreateShl(SP, 44); 1166 1167 // Store data to ring buffer. 1168 Value *RecordPtr = 1169 IRB.CreateIntToPtr(ThreadLongMaybeUntagged, IntptrTy->getPointerTo(0)); 1170 IRB.CreateStore(IRB.CreateOr(PC, SP), RecordPtr); 1171 1172 // Update the ring buffer. Top byte of ThreadLong defines the size of the 1173 // buffer in pages, it must be a power of two, and the start of the buffer 1174 // must be aligned by twice that much. Therefore wrap around of the ring 1175 // buffer is simply Addr &= ~((ThreadLong >> 56) << 12). 1176 // The use of AShr instead of LShr is due to 1177 // https://bugs.llvm.org/show_bug.cgi?id=39030 1178 // Runtime library makes sure not to use the highest bit. 1179 Value *WrapMask = IRB.CreateXor( 1180 IRB.CreateShl(IRB.CreateAShr(ThreadLong, 56), 12, "", true, true), 1181 ConstantInt::get(IntptrTy, (uint64_t)-1)); 1182 Value *ThreadLongNew = IRB.CreateAnd( 1183 IRB.CreateAdd(ThreadLong, ConstantInt::get(IntptrTy, 8)), WrapMask); 1184 IRB.CreateStore(ThreadLongNew, SlotPtr); 1185 } 1186 1187 if (!ShadowBase) { 1188 // Get shadow base address by aligning RecordPtr up. 1189 // Note: this is not correct if the pointer is already aligned. 1190 // Runtime library will make sure this never happens. 1191 ShadowBase = IRB.CreateAdd( 1192 IRB.CreateOr( 1193 ThreadLongMaybeUntagged, 1194 ConstantInt::get(IntptrTy, (1ULL << kShadowBaseAlignment) - 1)), 1195 ConstantInt::get(IntptrTy, 1), "hwasan.shadow"); 1196 ShadowBase = IRB.CreateIntToPtr(ShadowBase, Int8PtrTy); 1197 } 1198 } 1199 1200 Value *HWAddressSanitizer::readRegister(IRBuilder<> &IRB, StringRef Name) { 1201 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1202 Function *ReadRegister = 1203 Intrinsic::getDeclaration(M, Intrinsic::read_register, IntptrTy); 1204 MDNode *MD = MDNode::get(*C, {MDString::get(*C, Name)}); 1205 Value *Args[] = {MetadataAsValue::get(*C, MD)}; 1206 return IRB.CreateCall(ReadRegister, Args); 1207 } 1208 1209 bool HWAddressSanitizer::instrumentLandingPads( 1210 SmallVectorImpl<Instruction *> &LandingPadVec) { 1211 for (auto *LP : LandingPadVec) { 1212 IRBuilder<> IRB(LP->getNextNode()); 1213 IRB.CreateCall( 1214 HWAsanHandleVfork, 1215 {readRegister(IRB, (TargetTriple.getArch() == Triple::x86_64) ? "rsp" 1216 : "sp")}); 1217 } 1218 return true; 1219 } 1220 1221 static bool isLifetimeIntrinsic(Value *V) { 1222 auto *II = dyn_cast<IntrinsicInst>(V); 1223 return II && II->isLifetimeStartOrEnd(); 1224 } 1225 1226 bool HWAddressSanitizer::instrumentStack(memtag::StackInfo &SInfo, 1227 Value *StackTag, 1228 const DominatorTree &DT, 1229 const PostDominatorTree &PDT, 1230 const LoopInfo &LI) { 1231 // Ideally, we want to calculate tagged stack base pointer, and rewrite all 1232 // alloca addresses using that. Unfortunately, offsets are not known yet 1233 // (unless we use ASan-style mega-alloca). Instead we keep the base tag in a 1234 // temp, shift-OR it into each alloca address and xor with the retag mask. 1235 // This generates one extra instruction per alloca use. 1236 unsigned int I = 0; 1237 1238 for (auto &KV : SInfo.AllocasToInstrument) { 1239 auto N = I++; 1240 auto *AI = KV.first; 1241 memtag::AllocaInfo &Info = KV.second; 1242 IRBuilder<> IRB(AI->getNextNode()); 1243 1244 // Replace uses of the alloca with tagged address. 1245 Value *Tag = getAllocaTag(IRB, StackTag, AI, N); 1246 Value *AILong = IRB.CreatePointerCast(AI, IntptrTy); 1247 Value *Replacement = tagPointer(IRB, AI->getType(), AILong, Tag); 1248 std::string Name = 1249 AI->hasName() ? AI->getName().str() : "alloca." + itostr(N); 1250 Replacement->setName(Name + ".hwasan"); 1251 1252 size_t Size = memtag::getAllocaSizeInBytes(*AI); 1253 size_t AlignedSize = alignTo(Size, Mapping.getObjectAlignment()); 1254 1255 Value *AICast = IRB.CreatePointerCast(AI, Int8PtrTy); 1256 1257 auto HandleLifetime = [&](IntrinsicInst *II) { 1258 // Set the lifetime intrinsic to cover the whole alloca. This reduces the 1259 // set of assumptions we need to make about the lifetime. Without this we 1260 // would need to ensure that we can track the lifetime pointer to a 1261 // constant offset from the alloca, and would still need to change the 1262 // size to include the extra alignment we use for the untagging to make 1263 // the size consistent. 1264 // 1265 // The check for standard lifetime below makes sure that we have exactly 1266 // one set of start / end in any execution (i.e. the ends are not 1267 // reachable from each other), so this will not cause any problems. 1268 II->setArgOperand(0, ConstantInt::get(Int64Ty, AlignedSize)); 1269 II->setArgOperand(1, AICast); 1270 }; 1271 llvm::for_each(Info.LifetimeStart, HandleLifetime); 1272 llvm::for_each(Info.LifetimeEnd, HandleLifetime); 1273 1274 AI->replaceUsesWithIf(Replacement, [AICast, AILong](Use &U) { 1275 auto *User = U.getUser(); 1276 return User != AILong && User != AICast && !isLifetimeIntrinsic(User); 1277 }); 1278 1279 for (auto *DDI : Info.DbgVariableIntrinsics) { 1280 // Prepend "tag_offset, N" to the dwarf expression. 1281 // Tag offset logically applies to the alloca pointer, and it makes sense 1282 // to put it at the beginning of the expression. 1283 SmallVector<uint64_t, 8> NewOps = {dwarf::DW_OP_LLVM_tag_offset, 1284 retagMask(N)}; 1285 for (size_t LocNo = 0; LocNo < DDI->getNumVariableLocationOps(); ++LocNo) 1286 if (DDI->getVariableLocationOp(LocNo) == AI) 1287 DDI->setExpression(DIExpression::appendOpsToArg(DDI->getExpression(), 1288 NewOps, LocNo)); 1289 } 1290 1291 auto TagEnd = [&](Instruction *Node) { 1292 IRB.SetInsertPoint(Node); 1293 Value *UARTag = getUARTag(IRB, StackTag); 1294 // When untagging, use the `AlignedSize` because we need to set the tags 1295 // for the entire alloca to zero. If we used `Size` here, we would 1296 // keep the last granule tagged, and store zero in the last byte of the 1297 // last granule, due to how short granules are implemented. 1298 tagAlloca(IRB, AI, UARTag, AlignedSize); 1299 }; 1300 // Calls to functions that may return twice (e.g. setjmp) confuse the 1301 // postdominator analysis, and will leave us to keep memory tagged after 1302 // function return. Work around this by always untagging at every return 1303 // statement if return_twice functions are called. 1304 bool StandardLifetime = 1305 SInfo.UnrecognizedLifetimes.empty() && 1306 memtag::isStandardLifetime(Info.LifetimeStart, Info.LifetimeEnd, &DT, 1307 &LI, ClMaxLifetimes) && 1308 !SInfo.CallsReturnTwice; 1309 if (DetectUseAfterScope && StandardLifetime) { 1310 IntrinsicInst *Start = Info.LifetimeStart[0]; 1311 IRB.SetInsertPoint(Start->getNextNode()); 1312 tagAlloca(IRB, AI, Tag, Size); 1313 if (!memtag::forAllReachableExits(DT, PDT, LI, Start, Info.LifetimeEnd, 1314 SInfo.RetVec, TagEnd)) { 1315 for (auto *End : Info.LifetimeEnd) 1316 End->eraseFromParent(); 1317 } 1318 } else { 1319 tagAlloca(IRB, AI, Tag, Size); 1320 for (auto *RI : SInfo.RetVec) 1321 TagEnd(RI); 1322 // We inserted tagging outside of the lifetimes, so we have to remove 1323 // them. 1324 for (auto &II : Info.LifetimeStart) 1325 II->eraseFromParent(); 1326 for (auto &II : Info.LifetimeEnd) 1327 II->eraseFromParent(); 1328 } 1329 memtag::alignAndPadAlloca(Info, Align(Mapping.getObjectAlignment())); 1330 } 1331 for (auto &I : SInfo.UnrecognizedLifetimes) 1332 I->eraseFromParent(); 1333 return true; 1334 } 1335 1336 bool HWAddressSanitizer::isInterestingAlloca(const AllocaInst &AI) { 1337 return (AI.getAllocatedType()->isSized() && 1338 // FIXME: instrument dynamic allocas, too 1339 AI.isStaticAlloca() && 1340 // alloca() may be called with 0 size, ignore it. 1341 memtag::getAllocaSizeInBytes(AI) > 0 && 1342 // We are only interested in allocas not promotable to registers. 1343 // Promotable allocas are common under -O0. 1344 !isAllocaPromotable(&AI) && 1345 // inalloca allocas are not treated as static, and we don't want 1346 // dynamic alloca instrumentation for them as well. 1347 !AI.isUsedWithInAlloca() && 1348 // swifterror allocas are register promoted by ISel 1349 !AI.isSwiftError()) && 1350 // safe allocas are not interesting 1351 !(SSI && SSI->isSafe(AI)); 1352 } 1353 1354 bool HWAddressSanitizer::sanitizeFunction(Function &F, 1355 FunctionAnalysisManager &FAM) { 1356 if (&F == HwasanCtorFunction) 1357 return false; 1358 1359 if (!F.hasFnAttribute(Attribute::SanitizeHWAddress)) 1360 return false; 1361 1362 LLVM_DEBUG(dbgs() << "Function: " << F.getName() << "\n"); 1363 1364 SmallVector<InterestingMemoryOperand, 16> OperandsToInstrument; 1365 SmallVector<MemIntrinsic *, 16> IntrinToInstrument; 1366 SmallVector<Instruction *, 8> LandingPadVec; 1367 1368 memtag::StackInfoBuilder SIB( 1369 [this](const AllocaInst &AI) { return isInterestingAlloca(AI); }); 1370 for (auto &Inst : instructions(F)) { 1371 if (InstrumentStack) { 1372 SIB.visit(Inst); 1373 } 1374 1375 if (InstrumentLandingPads && isa<LandingPadInst>(Inst)) 1376 LandingPadVec.push_back(&Inst); 1377 1378 getInterestingMemoryOperands(&Inst, OperandsToInstrument); 1379 1380 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&Inst)) 1381 if (!ignoreMemIntrinsic(MI)) 1382 IntrinToInstrument.push_back(MI); 1383 } 1384 1385 memtag::StackInfo &SInfo = SIB.get(); 1386 1387 initializeCallbacks(*F.getParent()); 1388 1389 bool Changed = false; 1390 1391 if (!LandingPadVec.empty()) 1392 Changed |= instrumentLandingPads(LandingPadVec); 1393 1394 if (SInfo.AllocasToInstrument.empty() && F.hasPersonalityFn() && 1395 F.getPersonalityFn()->getName() == kHwasanPersonalityThunkName) { 1396 // __hwasan_personality_thunk is a no-op for functions without an 1397 // instrumented stack, so we can drop it. 1398 F.setPersonalityFn(nullptr); 1399 Changed = true; 1400 } 1401 1402 if (SInfo.AllocasToInstrument.empty() && OperandsToInstrument.empty() && 1403 IntrinToInstrument.empty()) 1404 return Changed; 1405 1406 assert(!ShadowBase); 1407 1408 Instruction *InsertPt = &*F.getEntryBlock().begin(); 1409 IRBuilder<> EntryIRB(InsertPt); 1410 emitPrologue(EntryIRB, 1411 /*WithFrameRecord*/ ClRecordStackHistory && 1412 Mapping.WithFrameRecord && 1413 !SInfo.AllocasToInstrument.empty()); 1414 1415 if (!SInfo.AllocasToInstrument.empty()) { 1416 const DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(F); 1417 const PostDominatorTree &PDT = FAM.getResult<PostDominatorTreeAnalysis>(F); 1418 const LoopInfo &LI = FAM.getResult<LoopAnalysis>(F); 1419 Value *StackTag = 1420 ClGenerateTagsWithCalls ? nullptr : getStackBaseTag(EntryIRB); 1421 instrumentStack(SInfo, StackTag, DT, PDT, LI); 1422 } 1423 1424 // If we split the entry block, move any allocas that were originally in the 1425 // entry block back into the entry block so that they aren't treated as 1426 // dynamic allocas. 1427 if (EntryIRB.GetInsertBlock() != &F.getEntryBlock()) { 1428 InsertPt = &*F.getEntryBlock().begin(); 1429 for (Instruction &I : 1430 llvm::make_early_inc_range(*EntryIRB.GetInsertBlock())) { 1431 if (auto *AI = dyn_cast<AllocaInst>(&I)) 1432 if (isa<ConstantInt>(AI->getArraySize())) 1433 I.moveBefore(InsertPt); 1434 } 1435 } 1436 1437 for (auto &Operand : OperandsToInstrument) 1438 instrumentMemAccess(Operand); 1439 1440 if (ClInstrumentMemIntrinsics && !IntrinToInstrument.empty()) { 1441 for (auto Inst : IntrinToInstrument) 1442 instrumentMemIntrinsic(cast<MemIntrinsic>(Inst)); 1443 } 1444 1445 ShadowBase = nullptr; 1446 StackBaseTag = nullptr; 1447 CachedSP = nullptr; 1448 1449 return true; 1450 } 1451 1452 void HWAddressSanitizer::instrumentGlobal(GlobalVariable *GV, uint8_t Tag) { 1453 assert(!UsePageAliases); 1454 Constant *Initializer = GV->getInitializer(); 1455 uint64_t SizeInBytes = 1456 M.getDataLayout().getTypeAllocSize(Initializer->getType()); 1457 uint64_t NewSize = alignTo(SizeInBytes, Mapping.getObjectAlignment()); 1458 if (SizeInBytes != NewSize) { 1459 // Pad the initializer out to the next multiple of 16 bytes and add the 1460 // required short granule tag. 1461 std::vector<uint8_t> Init(NewSize - SizeInBytes, 0); 1462 Init.back() = Tag; 1463 Constant *Padding = ConstantDataArray::get(*C, Init); 1464 Initializer = ConstantStruct::getAnon({Initializer, Padding}); 1465 } 1466 1467 auto *NewGV = new GlobalVariable(M, Initializer->getType(), GV->isConstant(), 1468 GlobalValue::ExternalLinkage, Initializer, 1469 GV->getName() + ".hwasan"); 1470 NewGV->copyAttributesFrom(GV); 1471 NewGV->setLinkage(GlobalValue::PrivateLinkage); 1472 NewGV->copyMetadata(GV, 0); 1473 NewGV->setAlignment( 1474 MaybeAlign(std::max(GV->getAlignment(), Mapping.getObjectAlignment()))); 1475 1476 // It is invalid to ICF two globals that have different tags. In the case 1477 // where the size of the global is a multiple of the tag granularity the 1478 // contents of the globals may be the same but the tags (i.e. symbol values) 1479 // may be different, and the symbols are not considered during ICF. In the 1480 // case where the size is not a multiple of the granularity, the short granule 1481 // tags would discriminate two globals with different tags, but there would 1482 // otherwise be nothing stopping such a global from being incorrectly ICF'd 1483 // with an uninstrumented (i.e. tag 0) global that happened to have the short 1484 // granule tag in the last byte. 1485 NewGV->setUnnamedAddr(GlobalValue::UnnamedAddr::None); 1486 1487 // Descriptor format (assuming little-endian): 1488 // bytes 0-3: relative address of global 1489 // bytes 4-6: size of global (16MB ought to be enough for anyone, but in case 1490 // it isn't, we create multiple descriptors) 1491 // byte 7: tag 1492 auto *DescriptorTy = StructType::get(Int32Ty, Int32Ty); 1493 const uint64_t MaxDescriptorSize = 0xfffff0; 1494 for (uint64_t DescriptorPos = 0; DescriptorPos < SizeInBytes; 1495 DescriptorPos += MaxDescriptorSize) { 1496 auto *Descriptor = 1497 new GlobalVariable(M, DescriptorTy, true, GlobalValue::PrivateLinkage, 1498 nullptr, GV->getName() + ".hwasan.descriptor"); 1499 auto *GVRelPtr = ConstantExpr::getTrunc( 1500 ConstantExpr::getAdd( 1501 ConstantExpr::getSub( 1502 ConstantExpr::getPtrToInt(NewGV, Int64Ty), 1503 ConstantExpr::getPtrToInt(Descriptor, Int64Ty)), 1504 ConstantInt::get(Int64Ty, DescriptorPos)), 1505 Int32Ty); 1506 uint32_t Size = std::min(SizeInBytes - DescriptorPos, MaxDescriptorSize); 1507 auto *SizeAndTag = ConstantInt::get(Int32Ty, Size | (uint32_t(Tag) << 24)); 1508 Descriptor->setComdat(NewGV->getComdat()); 1509 Descriptor->setInitializer(ConstantStruct::getAnon({GVRelPtr, SizeAndTag})); 1510 Descriptor->setSection("hwasan_globals"); 1511 Descriptor->setMetadata(LLVMContext::MD_associated, 1512 MDNode::get(*C, ValueAsMetadata::get(NewGV))); 1513 appendToCompilerUsed(M, Descriptor); 1514 } 1515 1516 Constant *Aliasee = ConstantExpr::getIntToPtr( 1517 ConstantExpr::getAdd( 1518 ConstantExpr::getPtrToInt(NewGV, Int64Ty), 1519 ConstantInt::get(Int64Ty, uint64_t(Tag) << PointerTagShift)), 1520 GV->getType()); 1521 auto *Alias = GlobalAlias::create(GV->getValueType(), GV->getAddressSpace(), 1522 GV->getLinkage(), "", Aliasee, &M); 1523 Alias->setVisibility(GV->getVisibility()); 1524 Alias->takeName(GV); 1525 GV->replaceAllUsesWith(Alias); 1526 GV->eraseFromParent(); 1527 } 1528 1529 void HWAddressSanitizer::instrumentGlobals() { 1530 std::vector<GlobalVariable *> Globals; 1531 for (GlobalVariable &GV : M.globals()) { 1532 if (GV.hasSanitizerMetadata() && GV.getSanitizerMetadata().NoHWAddress) 1533 continue; 1534 1535 if (GV.isDeclarationForLinker() || GV.getName().startswith("llvm.") || 1536 GV.isThreadLocal()) 1537 continue; 1538 1539 // Common symbols can't have aliases point to them, so they can't be tagged. 1540 if (GV.hasCommonLinkage()) 1541 continue; 1542 1543 // Globals with custom sections may be used in __start_/__stop_ enumeration, 1544 // which would be broken both by adding tags and potentially by the extra 1545 // padding/alignment that we insert. 1546 if (GV.hasSection()) 1547 continue; 1548 1549 Globals.push_back(&GV); 1550 } 1551 1552 MD5 Hasher; 1553 Hasher.update(M.getSourceFileName()); 1554 MD5::MD5Result Hash; 1555 Hasher.final(Hash); 1556 uint8_t Tag = Hash[0]; 1557 1558 for (GlobalVariable *GV : Globals) { 1559 Tag &= TagMaskByte; 1560 // Skip tag 0 in order to avoid collisions with untagged memory. 1561 if (Tag == 0) 1562 Tag = 1; 1563 instrumentGlobal(GV, Tag++); 1564 } 1565 } 1566 1567 void HWAddressSanitizer::instrumentPersonalityFunctions() { 1568 // We need to untag stack frames as we unwind past them. That is the job of 1569 // the personality function wrapper, which either wraps an existing 1570 // personality function or acts as a personality function on its own. Each 1571 // function that has a personality function or that can be unwound past has 1572 // its personality function changed to a thunk that calls the personality 1573 // function wrapper in the runtime. 1574 MapVector<Constant *, std::vector<Function *>> PersonalityFns; 1575 for (Function &F : M) { 1576 if (F.isDeclaration() || !F.hasFnAttribute(Attribute::SanitizeHWAddress)) 1577 continue; 1578 1579 if (F.hasPersonalityFn()) { 1580 PersonalityFns[F.getPersonalityFn()->stripPointerCasts()].push_back(&F); 1581 } else if (!F.hasFnAttribute(Attribute::NoUnwind)) { 1582 PersonalityFns[nullptr].push_back(&F); 1583 } 1584 } 1585 1586 if (PersonalityFns.empty()) 1587 return; 1588 1589 FunctionCallee HwasanPersonalityWrapper = M.getOrInsertFunction( 1590 "__hwasan_personality_wrapper", Int32Ty, Int32Ty, Int32Ty, Int64Ty, 1591 Int8PtrTy, Int8PtrTy, Int8PtrTy, Int8PtrTy, Int8PtrTy); 1592 FunctionCallee UnwindGetGR = M.getOrInsertFunction("_Unwind_GetGR", VoidTy); 1593 FunctionCallee UnwindGetCFA = M.getOrInsertFunction("_Unwind_GetCFA", VoidTy); 1594 1595 for (auto &P : PersonalityFns) { 1596 std::string ThunkName = kHwasanPersonalityThunkName; 1597 if (P.first) 1598 ThunkName += ("." + P.first->getName()).str(); 1599 FunctionType *ThunkFnTy = FunctionType::get( 1600 Int32Ty, {Int32Ty, Int32Ty, Int64Ty, Int8PtrTy, Int8PtrTy}, false); 1601 bool IsLocal = P.first && (!isa<GlobalValue>(P.first) || 1602 cast<GlobalValue>(P.first)->hasLocalLinkage()); 1603 auto *ThunkFn = Function::Create(ThunkFnTy, 1604 IsLocal ? GlobalValue::InternalLinkage 1605 : GlobalValue::LinkOnceODRLinkage, 1606 ThunkName, &M); 1607 if (!IsLocal) { 1608 ThunkFn->setVisibility(GlobalValue::HiddenVisibility); 1609 ThunkFn->setComdat(M.getOrInsertComdat(ThunkName)); 1610 } 1611 1612 auto *BB = BasicBlock::Create(*C, "entry", ThunkFn); 1613 IRBuilder<> IRB(BB); 1614 CallInst *WrapperCall = IRB.CreateCall( 1615 HwasanPersonalityWrapper, 1616 {ThunkFn->getArg(0), ThunkFn->getArg(1), ThunkFn->getArg(2), 1617 ThunkFn->getArg(3), ThunkFn->getArg(4), 1618 P.first ? IRB.CreateBitCast(P.first, Int8PtrTy) 1619 : Constant::getNullValue(Int8PtrTy), 1620 IRB.CreateBitCast(UnwindGetGR.getCallee(), Int8PtrTy), 1621 IRB.CreateBitCast(UnwindGetCFA.getCallee(), Int8PtrTy)}); 1622 WrapperCall->setTailCall(); 1623 IRB.CreateRet(WrapperCall); 1624 1625 for (Function *F : P.second) 1626 F->setPersonalityFn(ThunkFn); 1627 } 1628 } 1629 1630 void HWAddressSanitizer::ShadowMapping::init(Triple &TargetTriple, 1631 bool InstrumentWithCalls) { 1632 Scale = kDefaultShadowScale; 1633 if (TargetTriple.isOSFuchsia()) { 1634 // Fuchsia is always PIE, which means that the beginning of the address 1635 // space is always available. 1636 InGlobal = false; 1637 InTls = false; 1638 Offset = 0; 1639 WithFrameRecord = true; 1640 } else if (ClMappingOffset.getNumOccurrences() > 0) { 1641 InGlobal = false; 1642 InTls = false; 1643 Offset = ClMappingOffset; 1644 WithFrameRecord = false; 1645 } else if (ClEnableKhwasan || InstrumentWithCalls) { 1646 InGlobal = false; 1647 InTls = false; 1648 Offset = 0; 1649 WithFrameRecord = false; 1650 } else if (ClWithIfunc) { 1651 InGlobal = true; 1652 InTls = false; 1653 Offset = kDynamicShadowSentinel; 1654 WithFrameRecord = false; 1655 } else if (ClWithTls) { 1656 InGlobal = false; 1657 InTls = true; 1658 Offset = kDynamicShadowSentinel; 1659 WithFrameRecord = true; 1660 } else { 1661 InGlobal = false; 1662 InTls = false; 1663 Offset = kDynamicShadowSentinel; 1664 WithFrameRecord = false; 1665 } 1666 } 1667