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 Value *getFrameRecordInfo(IRBuilder<> &IRB); 317 318 void instrumentPersonalityFunctions(); 319 320 private: 321 LLVMContext *C; 322 Module &M; 323 const StackSafetyGlobalInfo *SSI; 324 Triple TargetTriple; 325 FunctionCallee HWAsanMemmove, HWAsanMemcpy, HWAsanMemset; 326 FunctionCallee HWAsanHandleVfork; 327 328 /// This struct defines the shadow mapping using the rule: 329 /// shadow = (mem >> Scale) + Offset. 330 /// If InGlobal is true, then 331 /// extern char __hwasan_shadow[]; 332 /// shadow = (mem >> Scale) + &__hwasan_shadow 333 /// If InTls is true, then 334 /// extern char *__hwasan_tls; 335 /// shadow = (mem>>Scale) + align_up(__hwasan_shadow, kShadowBaseAlignment) 336 /// 337 /// If WithFrameRecord is true, then __hwasan_tls will be used to access the 338 /// ring buffer for storing stack allocations on targets that support it. 339 struct ShadowMapping { 340 int Scale; 341 uint64_t Offset; 342 bool InGlobal; 343 bool InTls; 344 bool WithFrameRecord; 345 346 void init(Triple &TargetTriple, bool InstrumentWithCalls); 347 uint64_t getObjectAlignment() const { return 1ULL << Scale; } 348 }; 349 350 ShadowMapping Mapping; 351 352 Type *VoidTy = Type::getVoidTy(M.getContext()); 353 Type *IntptrTy; 354 Type *Int8PtrTy; 355 Type *Int8Ty; 356 Type *Int32Ty; 357 Type *Int64Ty = Type::getInt64Ty(M.getContext()); 358 359 bool CompileKernel; 360 bool Recover; 361 bool OutlinedChecks; 362 bool UseShortGranules; 363 bool InstrumentLandingPads; 364 bool InstrumentWithCalls; 365 bool InstrumentStack; 366 bool DetectUseAfterScope; 367 bool UsePageAliases; 368 369 bool HasMatchAllTag = false; 370 uint8_t MatchAllTag = 0; 371 372 unsigned PointerTagShift; 373 uint64_t TagMaskByte; 374 375 Function *HwasanCtorFunction; 376 377 FunctionCallee HwasanMemoryAccessCallback[2][kNumberOfAccessSizes]; 378 FunctionCallee HwasanMemoryAccessCallbackSized[2]; 379 380 FunctionCallee HwasanTagMemoryFunc; 381 FunctionCallee HwasanGenerateTagFunc; 382 383 Constant *ShadowGlobal; 384 385 Value *ShadowBase = nullptr; 386 Value *StackBaseTag = nullptr; 387 Value *CachedSP = nullptr; 388 GlobalValue *ThreadPtrGlobal = nullptr; 389 }; 390 391 } // end anonymous namespace 392 393 PreservedAnalyses HWAddressSanitizerPass::run(Module &M, 394 ModuleAnalysisManager &MAM) { 395 const StackSafetyGlobalInfo *SSI = nullptr; 396 auto TargetTriple = llvm::Triple(M.getTargetTriple()); 397 if (shouldUseStackSafetyAnalysis(TargetTriple, Options.DisableOptimization)) 398 SSI = &MAM.getResult<StackSafetyGlobalAnalysis>(M); 399 400 HWAddressSanitizer HWASan(M, Options.CompileKernel, Options.Recover, SSI); 401 bool Modified = false; 402 auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 403 for (Function &F : M) 404 Modified |= HWASan.sanitizeFunction(F, FAM); 405 if (Modified) 406 return PreservedAnalyses::none(); 407 return PreservedAnalyses::all(); 408 } 409 void HWAddressSanitizerPass::printPipeline( 410 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { 411 static_cast<PassInfoMixin<HWAddressSanitizerPass> *>(this)->printPipeline( 412 OS, MapClassName2PassName); 413 OS << "<"; 414 if (Options.CompileKernel) 415 OS << "kernel;"; 416 if (Options.Recover) 417 OS << "recover"; 418 OS << ">"; 419 } 420 421 void HWAddressSanitizer::createHwasanCtorComdat() { 422 std::tie(HwasanCtorFunction, std::ignore) = 423 getOrCreateSanitizerCtorAndInitFunctions( 424 M, kHwasanModuleCtorName, kHwasanInitName, 425 /*InitArgTypes=*/{}, 426 /*InitArgs=*/{}, 427 // This callback is invoked when the functions are created the first 428 // time. Hook them into the global ctors list in that case: 429 [&](Function *Ctor, FunctionCallee) { 430 Comdat *CtorComdat = M.getOrInsertComdat(kHwasanModuleCtorName); 431 Ctor->setComdat(CtorComdat); 432 appendToGlobalCtors(M, Ctor, 0, Ctor); 433 }); 434 435 // Create a note that contains pointers to the list of global 436 // descriptors. Adding a note to the output file will cause the linker to 437 // create a PT_NOTE program header pointing to the note that we can use to 438 // find the descriptor list starting from the program headers. A function 439 // provided by the runtime initializes the shadow memory for the globals by 440 // accessing the descriptor list via the note. The dynamic loader needs to 441 // call this function whenever a library is loaded. 442 // 443 // The reason why we use a note for this instead of a more conventional 444 // approach of having a global constructor pass a descriptor list pointer to 445 // the runtime is because of an order of initialization problem. With 446 // constructors we can encounter the following problematic scenario: 447 // 448 // 1) library A depends on library B and also interposes one of B's symbols 449 // 2) B's constructors are called before A's (as required for correctness) 450 // 3) during construction, B accesses one of its "own" globals (actually 451 // interposed by A) and triggers a HWASAN failure due to the initialization 452 // for A not having happened yet 453 // 454 // Even without interposition it is possible to run into similar situations in 455 // cases where two libraries mutually depend on each other. 456 // 457 // We only need one note per binary, so put everything for the note in a 458 // comdat. This needs to be a comdat with an .init_array section to prevent 459 // newer versions of lld from discarding the note. 460 // 461 // Create the note even if we aren't instrumenting globals. This ensures that 462 // binaries linked from object files with both instrumented and 463 // non-instrumented globals will end up with a note, even if a comdat from an 464 // object file with non-instrumented globals is selected. The note is harmless 465 // if the runtime doesn't support it, since it will just be ignored. 466 Comdat *NoteComdat = M.getOrInsertComdat(kHwasanModuleCtorName); 467 468 Type *Int8Arr0Ty = ArrayType::get(Int8Ty, 0); 469 auto Start = 470 new GlobalVariable(M, Int8Arr0Ty, true, GlobalVariable::ExternalLinkage, 471 nullptr, "__start_hwasan_globals"); 472 Start->setVisibility(GlobalValue::HiddenVisibility); 473 Start->setDSOLocal(true); 474 auto Stop = 475 new GlobalVariable(M, Int8Arr0Ty, true, GlobalVariable::ExternalLinkage, 476 nullptr, "__stop_hwasan_globals"); 477 Stop->setVisibility(GlobalValue::HiddenVisibility); 478 Stop->setDSOLocal(true); 479 480 // Null-terminated so actually 8 bytes, which are required in order to align 481 // the note properly. 482 auto *Name = ConstantDataArray::get(*C, "LLVM\0\0\0"); 483 484 auto *NoteTy = StructType::get(Int32Ty, Int32Ty, Int32Ty, Name->getType(), 485 Int32Ty, Int32Ty); 486 auto *Note = 487 new GlobalVariable(M, NoteTy, /*isConstant=*/true, 488 GlobalValue::PrivateLinkage, nullptr, kHwasanNoteName); 489 Note->setSection(".note.hwasan.globals"); 490 Note->setComdat(NoteComdat); 491 Note->setAlignment(Align(4)); 492 Note->setDSOLocal(true); 493 494 // The pointers in the note need to be relative so that the note ends up being 495 // placed in rodata, which is the standard location for notes. 496 auto CreateRelPtr = [&](Constant *Ptr) { 497 return ConstantExpr::getTrunc( 498 ConstantExpr::getSub(ConstantExpr::getPtrToInt(Ptr, Int64Ty), 499 ConstantExpr::getPtrToInt(Note, Int64Ty)), 500 Int32Ty); 501 }; 502 Note->setInitializer(ConstantStruct::getAnon( 503 {ConstantInt::get(Int32Ty, 8), // n_namesz 504 ConstantInt::get(Int32Ty, 8), // n_descsz 505 ConstantInt::get(Int32Ty, ELF::NT_LLVM_HWASAN_GLOBALS), // n_type 506 Name, CreateRelPtr(Start), CreateRelPtr(Stop)})); 507 appendToCompilerUsed(M, Note); 508 509 // Create a zero-length global in hwasan_globals so that the linker will 510 // always create start and stop symbols. 511 auto Dummy = new GlobalVariable( 512 M, Int8Arr0Ty, /*isConstantGlobal*/ true, GlobalVariable::PrivateLinkage, 513 Constant::getNullValue(Int8Arr0Ty), "hwasan.dummy.global"); 514 Dummy->setSection("hwasan_globals"); 515 Dummy->setComdat(NoteComdat); 516 Dummy->setMetadata(LLVMContext::MD_associated, 517 MDNode::get(*C, ValueAsMetadata::get(Note))); 518 appendToCompilerUsed(M, Dummy); 519 } 520 521 /// Module-level initialization. 522 /// 523 /// inserts a call to __hwasan_init to the module's constructor list. 524 void HWAddressSanitizer::initializeModule() { 525 LLVM_DEBUG(dbgs() << "Init " << M.getName() << "\n"); 526 auto &DL = M.getDataLayout(); 527 528 TargetTriple = Triple(M.getTargetTriple()); 529 530 // x86_64 currently has two modes: 531 // - Intel LAM (default) 532 // - pointer aliasing (heap only) 533 bool IsX86_64 = TargetTriple.getArch() == Triple::x86_64; 534 UsePageAliases = shouldUsePageAliases(TargetTriple); 535 InstrumentWithCalls = shouldInstrumentWithCalls(TargetTriple); 536 InstrumentStack = shouldInstrumentStack(TargetTriple); 537 DetectUseAfterScope = shouldDetectUseAfterScope(TargetTriple); 538 PointerTagShift = IsX86_64 ? 57 : 56; 539 TagMaskByte = IsX86_64 ? 0x3F : 0xFF; 540 541 Mapping.init(TargetTriple, InstrumentWithCalls); 542 543 C = &(M.getContext()); 544 IRBuilder<> IRB(*C); 545 IntptrTy = IRB.getIntPtrTy(DL); 546 Int8PtrTy = IRB.getInt8PtrTy(); 547 Int8Ty = IRB.getInt8Ty(); 548 Int32Ty = IRB.getInt32Ty(); 549 550 HwasanCtorFunction = nullptr; 551 552 // Older versions of Android do not have the required runtime support for 553 // short granules, global or personality function instrumentation. On other 554 // platforms we currently require using the latest version of the runtime. 555 bool NewRuntime = 556 !TargetTriple.isAndroid() || !TargetTriple.isAndroidVersionLT(30); 557 558 UseShortGranules = 559 ClUseShortGranules.getNumOccurrences() ? ClUseShortGranules : NewRuntime; 560 OutlinedChecks = 561 TargetTriple.isAArch64() && TargetTriple.isOSBinFormatELF() && 562 (ClInlineAllChecks.getNumOccurrences() ? !ClInlineAllChecks : !Recover); 563 564 if (ClMatchAllTag.getNumOccurrences()) { 565 if (ClMatchAllTag != -1) { 566 HasMatchAllTag = true; 567 MatchAllTag = ClMatchAllTag & 0xFF; 568 } 569 } else if (CompileKernel) { 570 HasMatchAllTag = true; 571 MatchAllTag = 0xFF; 572 } 573 574 // If we don't have personality function support, fall back to landing pads. 575 InstrumentLandingPads = ClInstrumentLandingPads.getNumOccurrences() 576 ? ClInstrumentLandingPads 577 : !NewRuntime; 578 579 if (!CompileKernel) { 580 createHwasanCtorComdat(); 581 bool InstrumentGlobals = 582 ClGlobals.getNumOccurrences() ? ClGlobals : NewRuntime; 583 584 if (InstrumentGlobals && !UsePageAliases) 585 instrumentGlobals(); 586 587 bool InstrumentPersonalityFunctions = 588 ClInstrumentPersonalityFunctions.getNumOccurrences() 589 ? ClInstrumentPersonalityFunctions 590 : NewRuntime; 591 if (InstrumentPersonalityFunctions) 592 instrumentPersonalityFunctions(); 593 } 594 595 if (!TargetTriple.isAndroid()) { 596 Constant *C = M.getOrInsertGlobal("__hwasan_tls", IntptrTy, [&] { 597 auto *GV = new GlobalVariable(M, IntptrTy, /*isConstant=*/false, 598 GlobalValue::ExternalLinkage, nullptr, 599 "__hwasan_tls", nullptr, 600 GlobalVariable::InitialExecTLSModel); 601 appendToCompilerUsed(M, GV); 602 return GV; 603 }); 604 ThreadPtrGlobal = cast<GlobalVariable>(C); 605 } 606 } 607 608 void HWAddressSanitizer::initializeCallbacks(Module &M) { 609 IRBuilder<> IRB(*C); 610 for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) { 611 const std::string TypeStr = AccessIsWrite ? "store" : "load"; 612 const std::string EndingStr = Recover ? "_noabort" : ""; 613 614 HwasanMemoryAccessCallbackSized[AccessIsWrite] = M.getOrInsertFunction( 615 ClMemoryAccessCallbackPrefix + TypeStr + "N" + EndingStr, 616 FunctionType::get(IRB.getVoidTy(), {IntptrTy, IntptrTy}, false)); 617 618 for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes; 619 AccessSizeIndex++) { 620 HwasanMemoryAccessCallback[AccessIsWrite][AccessSizeIndex] = 621 M.getOrInsertFunction( 622 ClMemoryAccessCallbackPrefix + TypeStr + 623 itostr(1ULL << AccessSizeIndex) + EndingStr, 624 FunctionType::get(IRB.getVoidTy(), {IntptrTy}, false)); 625 } 626 } 627 628 HwasanTagMemoryFunc = M.getOrInsertFunction( 629 "__hwasan_tag_memory", IRB.getVoidTy(), Int8PtrTy, Int8Ty, IntptrTy); 630 HwasanGenerateTagFunc = 631 M.getOrInsertFunction("__hwasan_generate_tag", Int8Ty); 632 633 ShadowGlobal = M.getOrInsertGlobal("__hwasan_shadow", 634 ArrayType::get(IRB.getInt8Ty(), 0)); 635 636 const std::string MemIntrinCallbackPrefix = 637 (CompileKernel && !ClKasanMemIntrinCallbackPrefix) 638 ? std::string("") 639 : ClMemoryAccessCallbackPrefix; 640 HWAsanMemmove = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memmove", 641 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 642 IRB.getInt8PtrTy(), IntptrTy); 643 HWAsanMemcpy = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memcpy", 644 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 645 IRB.getInt8PtrTy(), IntptrTy); 646 HWAsanMemset = M.getOrInsertFunction(MemIntrinCallbackPrefix + "memset", 647 IRB.getInt8PtrTy(), IRB.getInt8PtrTy(), 648 IRB.getInt32Ty(), IntptrTy); 649 650 HWAsanHandleVfork = 651 M.getOrInsertFunction("__hwasan_handle_vfork", IRB.getVoidTy(), IntptrTy); 652 } 653 654 Value *HWAddressSanitizer::getOpaqueNoopCast(IRBuilder<> &IRB, Value *Val) { 655 // An empty inline asm with input reg == output reg. 656 // An opaque no-op cast, basically. 657 // This prevents code bloat as a result of rematerializing trivial definitions 658 // such as constants or global addresses at every load and store. 659 InlineAsm *Asm = 660 InlineAsm::get(FunctionType::get(Int8PtrTy, {Val->getType()}, false), 661 StringRef(""), StringRef("=r,0"), 662 /*hasSideEffects=*/false); 663 return IRB.CreateCall(Asm, {Val}, ".hwasan.shadow"); 664 } 665 666 Value *HWAddressSanitizer::getDynamicShadowIfunc(IRBuilder<> &IRB) { 667 return getOpaqueNoopCast(IRB, ShadowGlobal); 668 } 669 670 Value *HWAddressSanitizer::getShadowNonTls(IRBuilder<> &IRB) { 671 if (Mapping.Offset != kDynamicShadowSentinel) 672 return getOpaqueNoopCast( 673 IRB, ConstantExpr::getIntToPtr( 674 ConstantInt::get(IntptrTy, Mapping.Offset), Int8PtrTy)); 675 676 if (Mapping.InGlobal) { 677 return getDynamicShadowIfunc(IRB); 678 } else { 679 Value *GlobalDynamicAddress = 680 IRB.GetInsertBlock()->getParent()->getParent()->getOrInsertGlobal( 681 kHwasanShadowMemoryDynamicAddress, Int8PtrTy); 682 return IRB.CreateLoad(Int8PtrTy, GlobalDynamicAddress); 683 } 684 } 685 686 bool HWAddressSanitizer::ignoreAccess(Instruction *Inst, Value *Ptr) { 687 // Do not instrument acesses from different address spaces; we cannot deal 688 // with them. 689 Type *PtrTy = cast<PointerType>(Ptr->getType()->getScalarType()); 690 if (PtrTy->getPointerAddressSpace() != 0) 691 return true; 692 693 // Ignore swifterror addresses. 694 // swifterror memory addresses are mem2reg promoted by instruction 695 // selection. As such they cannot have regular uses like an instrumentation 696 // function and it makes no sense to track them as memory. 697 if (Ptr->isSwiftError()) 698 return true; 699 700 if (findAllocaForValue(Ptr)) { 701 if (!InstrumentStack) 702 return true; 703 if (SSI && SSI->stackAccessIsSafe(*Inst)) 704 return true; 705 } 706 return false; 707 } 708 709 void HWAddressSanitizer::getInterestingMemoryOperands( 710 Instruction *I, SmallVectorImpl<InterestingMemoryOperand> &Interesting) { 711 // Skip memory accesses inserted by another instrumentation. 712 if (I->hasMetadata(LLVMContext::MD_nosanitize)) 713 return; 714 715 // Do not instrument the load fetching the dynamic shadow address. 716 if (ShadowBase == I) 717 return; 718 719 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 720 if (!ClInstrumentReads || ignoreAccess(I, LI->getPointerOperand())) 721 return; 722 Interesting.emplace_back(I, LI->getPointerOperandIndex(), false, 723 LI->getType(), LI->getAlign()); 724 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) { 725 if (!ClInstrumentWrites || ignoreAccess(I, SI->getPointerOperand())) 726 return; 727 Interesting.emplace_back(I, SI->getPointerOperandIndex(), true, 728 SI->getValueOperand()->getType(), SI->getAlign()); 729 } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) { 730 if (!ClInstrumentAtomics || ignoreAccess(I, RMW->getPointerOperand())) 731 return; 732 Interesting.emplace_back(I, RMW->getPointerOperandIndex(), true, 733 RMW->getValOperand()->getType(), None); 734 } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) { 735 if (!ClInstrumentAtomics || ignoreAccess(I, XCHG->getPointerOperand())) 736 return; 737 Interesting.emplace_back(I, XCHG->getPointerOperandIndex(), true, 738 XCHG->getCompareOperand()->getType(), None); 739 } else if (auto CI = dyn_cast<CallInst>(I)) { 740 for (unsigned ArgNo = 0; ArgNo < CI->arg_size(); ArgNo++) { 741 if (!ClInstrumentByval || !CI->isByValArgument(ArgNo) || 742 ignoreAccess(I, CI->getArgOperand(ArgNo))) 743 continue; 744 Type *Ty = CI->getParamByValType(ArgNo); 745 Interesting.emplace_back(I, ArgNo, false, Ty, Align(1)); 746 } 747 } 748 } 749 750 static unsigned getPointerOperandIndex(Instruction *I) { 751 if (LoadInst *LI = dyn_cast<LoadInst>(I)) 752 return LI->getPointerOperandIndex(); 753 if (StoreInst *SI = dyn_cast<StoreInst>(I)) 754 return SI->getPointerOperandIndex(); 755 if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) 756 return RMW->getPointerOperandIndex(); 757 if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) 758 return XCHG->getPointerOperandIndex(); 759 report_fatal_error("Unexpected instruction"); 760 return -1; 761 } 762 763 static size_t TypeSizeToSizeIndex(uint32_t TypeSize) { 764 size_t Res = countTrailingZeros(TypeSize / 8); 765 assert(Res < kNumberOfAccessSizes); 766 return Res; 767 } 768 769 void HWAddressSanitizer::untagPointerOperand(Instruction *I, Value *Addr) { 770 if (TargetTriple.isAArch64() || TargetTriple.getArch() == Triple::x86_64) 771 return; 772 773 IRBuilder<> IRB(I); 774 Value *AddrLong = IRB.CreatePointerCast(Addr, IntptrTy); 775 Value *UntaggedPtr = 776 IRB.CreateIntToPtr(untagPointer(IRB, AddrLong), Addr->getType()); 777 I->setOperand(getPointerOperandIndex(I), UntaggedPtr); 778 } 779 780 Value *HWAddressSanitizer::memToShadow(Value *Mem, IRBuilder<> &IRB) { 781 // Mem >> Scale 782 Value *Shadow = IRB.CreateLShr(Mem, Mapping.Scale); 783 if (Mapping.Offset == 0) 784 return IRB.CreateIntToPtr(Shadow, Int8PtrTy); 785 // (Mem >> Scale) + Offset 786 return IRB.CreateGEP(Int8Ty, ShadowBase, Shadow); 787 } 788 789 int64_t HWAddressSanitizer::getAccessInfo(bool IsWrite, 790 unsigned AccessSizeIndex) { 791 return (CompileKernel << HWASanAccessInfo::CompileKernelShift) + 792 (HasMatchAllTag << HWASanAccessInfo::HasMatchAllShift) + 793 (MatchAllTag << HWASanAccessInfo::MatchAllShift) + 794 (Recover << HWASanAccessInfo::RecoverShift) + 795 (IsWrite << HWASanAccessInfo::IsWriteShift) + 796 (AccessSizeIndex << HWASanAccessInfo::AccessSizeShift); 797 } 798 799 void HWAddressSanitizer::instrumentMemAccessOutline(Value *Ptr, bool IsWrite, 800 unsigned AccessSizeIndex, 801 Instruction *InsertBefore) { 802 assert(!UsePageAliases); 803 const int64_t AccessInfo = getAccessInfo(IsWrite, AccessSizeIndex); 804 IRBuilder<> IRB(InsertBefore); 805 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 806 Ptr = IRB.CreateBitCast(Ptr, Int8PtrTy); 807 IRB.CreateCall(Intrinsic::getDeclaration( 808 M, UseShortGranules 809 ? Intrinsic::hwasan_check_memaccess_shortgranules 810 : Intrinsic::hwasan_check_memaccess), 811 {ShadowBase, Ptr, ConstantInt::get(Int32Ty, AccessInfo)}); 812 } 813 814 void HWAddressSanitizer::instrumentMemAccessInline(Value *Ptr, bool IsWrite, 815 unsigned AccessSizeIndex, 816 Instruction *InsertBefore) { 817 assert(!UsePageAliases); 818 const int64_t AccessInfo = getAccessInfo(IsWrite, AccessSizeIndex); 819 IRBuilder<> IRB(InsertBefore); 820 821 Value *PtrLong = IRB.CreatePointerCast(Ptr, IntptrTy); 822 Value *PtrTag = IRB.CreateTrunc(IRB.CreateLShr(PtrLong, PointerTagShift), 823 IRB.getInt8Ty()); 824 Value *AddrLong = untagPointer(IRB, PtrLong); 825 Value *Shadow = memToShadow(AddrLong, IRB); 826 Value *MemTag = IRB.CreateLoad(Int8Ty, Shadow); 827 Value *TagMismatch = IRB.CreateICmpNE(PtrTag, MemTag); 828 829 if (HasMatchAllTag) { 830 Value *TagNotIgnored = IRB.CreateICmpNE( 831 PtrTag, ConstantInt::get(PtrTag->getType(), MatchAllTag)); 832 TagMismatch = IRB.CreateAnd(TagMismatch, TagNotIgnored); 833 } 834 835 Instruction *CheckTerm = 836 SplitBlockAndInsertIfThen(TagMismatch, InsertBefore, false, 837 MDBuilder(*C).createBranchWeights(1, 100000)); 838 839 IRB.SetInsertPoint(CheckTerm); 840 Value *OutOfShortGranuleTagRange = 841 IRB.CreateICmpUGT(MemTag, ConstantInt::get(Int8Ty, 15)); 842 Instruction *CheckFailTerm = 843 SplitBlockAndInsertIfThen(OutOfShortGranuleTagRange, CheckTerm, !Recover, 844 MDBuilder(*C).createBranchWeights(1, 100000)); 845 846 IRB.SetInsertPoint(CheckTerm); 847 Value *PtrLowBits = IRB.CreateTrunc(IRB.CreateAnd(PtrLong, 15), Int8Ty); 848 PtrLowBits = IRB.CreateAdd( 849 PtrLowBits, ConstantInt::get(Int8Ty, (1 << AccessSizeIndex) - 1)); 850 Value *PtrLowBitsOOB = IRB.CreateICmpUGE(PtrLowBits, MemTag); 851 SplitBlockAndInsertIfThen(PtrLowBitsOOB, CheckTerm, false, 852 MDBuilder(*C).createBranchWeights(1, 100000), 853 (DomTreeUpdater *)nullptr, nullptr, 854 CheckFailTerm->getParent()); 855 856 IRB.SetInsertPoint(CheckTerm); 857 Value *InlineTagAddr = IRB.CreateOr(AddrLong, 15); 858 InlineTagAddr = IRB.CreateIntToPtr(InlineTagAddr, Int8PtrTy); 859 Value *InlineTag = IRB.CreateLoad(Int8Ty, InlineTagAddr); 860 Value *InlineTagMismatch = IRB.CreateICmpNE(PtrTag, InlineTag); 861 SplitBlockAndInsertIfThen(InlineTagMismatch, CheckTerm, false, 862 MDBuilder(*C).createBranchWeights(1, 100000), 863 (DomTreeUpdater *)nullptr, nullptr, 864 CheckFailTerm->getParent()); 865 866 IRB.SetInsertPoint(CheckFailTerm); 867 InlineAsm *Asm; 868 switch (TargetTriple.getArch()) { 869 case Triple::x86_64: 870 // The signal handler will find the data address in rdi. 871 Asm = InlineAsm::get( 872 FunctionType::get(IRB.getVoidTy(), {PtrLong->getType()}, false), 873 "int3\nnopl " + 874 itostr(0x40 + (AccessInfo & HWASanAccessInfo::RuntimeMask)) + 875 "(%rax)", 876 "{rdi}", 877 /*hasSideEffects=*/true); 878 break; 879 case Triple::aarch64: 880 case Triple::aarch64_be: 881 // The signal handler will find the data address in x0. 882 Asm = InlineAsm::get( 883 FunctionType::get(IRB.getVoidTy(), {PtrLong->getType()}, false), 884 "brk #" + itostr(0x900 + (AccessInfo & HWASanAccessInfo::RuntimeMask)), 885 "{x0}", 886 /*hasSideEffects=*/true); 887 break; 888 default: 889 report_fatal_error("unsupported architecture"); 890 } 891 IRB.CreateCall(Asm, PtrLong); 892 if (Recover) 893 cast<BranchInst>(CheckFailTerm)->setSuccessor(0, CheckTerm->getParent()); 894 } 895 896 bool HWAddressSanitizer::ignoreMemIntrinsic(MemIntrinsic *MI) { 897 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { 898 return (!ClInstrumentWrites || ignoreAccess(MTI, MTI->getDest())) && 899 (!ClInstrumentReads || ignoreAccess(MTI, MTI->getSource())); 900 } 901 if (isa<MemSetInst>(MI)) 902 return !ClInstrumentWrites || ignoreAccess(MI, MI->getDest()); 903 return false; 904 } 905 906 void HWAddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) { 907 IRBuilder<> IRB(MI); 908 if (isa<MemTransferInst>(MI)) { 909 IRB.CreateCall( 910 isa<MemMoveInst>(MI) ? HWAsanMemmove : HWAsanMemcpy, 911 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()), 912 IRB.CreatePointerCast(MI->getOperand(1), IRB.getInt8PtrTy()), 913 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)}); 914 } else if (isa<MemSetInst>(MI)) { 915 IRB.CreateCall( 916 HWAsanMemset, 917 {IRB.CreatePointerCast(MI->getOperand(0), IRB.getInt8PtrTy()), 918 IRB.CreateIntCast(MI->getOperand(1), IRB.getInt32Ty(), false), 919 IRB.CreateIntCast(MI->getOperand(2), IntptrTy, false)}); 920 } 921 MI->eraseFromParent(); 922 } 923 924 bool HWAddressSanitizer::instrumentMemAccess(InterestingMemoryOperand &O) { 925 Value *Addr = O.getPtr(); 926 927 LLVM_DEBUG(dbgs() << "Instrumenting: " << O.getInsn() << "\n"); 928 929 if (O.MaybeMask) 930 return false; // FIXME 931 932 IRBuilder<> IRB(O.getInsn()); 933 if (isPowerOf2_64(O.TypeSize) && 934 (O.TypeSize / 8 <= (1ULL << (kNumberOfAccessSizes - 1))) && 935 (!O.Alignment || *O.Alignment >= (1ULL << Mapping.Scale) || 936 *O.Alignment >= O.TypeSize / 8)) { 937 size_t AccessSizeIndex = TypeSizeToSizeIndex(O.TypeSize); 938 if (InstrumentWithCalls) { 939 IRB.CreateCall(HwasanMemoryAccessCallback[O.IsWrite][AccessSizeIndex], 940 IRB.CreatePointerCast(Addr, IntptrTy)); 941 } else if (OutlinedChecks) { 942 instrumentMemAccessOutline(Addr, O.IsWrite, AccessSizeIndex, O.getInsn()); 943 } else { 944 instrumentMemAccessInline(Addr, O.IsWrite, AccessSizeIndex, O.getInsn()); 945 } 946 } else { 947 IRB.CreateCall(HwasanMemoryAccessCallbackSized[O.IsWrite], 948 {IRB.CreatePointerCast(Addr, IntptrTy), 949 ConstantInt::get(IntptrTy, O.TypeSize / 8)}); 950 } 951 untagPointerOperand(O.getInsn(), Addr); 952 953 return true; 954 } 955 956 void HWAddressSanitizer::tagAlloca(IRBuilder<> &IRB, AllocaInst *AI, Value *Tag, 957 size_t Size) { 958 size_t AlignedSize = alignTo(Size, Mapping.getObjectAlignment()); 959 if (!UseShortGranules) 960 Size = AlignedSize; 961 962 Value *JustTag = IRB.CreateTrunc(Tag, IRB.getInt8Ty()); 963 if (InstrumentWithCalls) { 964 IRB.CreateCall(HwasanTagMemoryFunc, 965 {IRB.CreatePointerCast(AI, Int8PtrTy), JustTag, 966 ConstantInt::get(IntptrTy, AlignedSize)}); 967 } else { 968 size_t ShadowSize = Size >> Mapping.Scale; 969 Value *ShadowPtr = memToShadow(IRB.CreatePointerCast(AI, IntptrTy), IRB); 970 // If this memset is not inlined, it will be intercepted in the hwasan 971 // runtime library. That's OK, because the interceptor skips the checks if 972 // the address is in the shadow region. 973 // FIXME: the interceptor is not as fast as real memset. Consider lowering 974 // llvm.memset right here into either a sequence of stores, or a call to 975 // hwasan_tag_memory. 976 if (ShadowSize) 977 IRB.CreateMemSet(ShadowPtr, JustTag, ShadowSize, Align(1)); 978 if (Size != AlignedSize) { 979 IRB.CreateStore( 980 ConstantInt::get(Int8Ty, Size % Mapping.getObjectAlignment()), 981 IRB.CreateConstGEP1_32(Int8Ty, ShadowPtr, ShadowSize)); 982 IRB.CreateStore(JustTag, IRB.CreateConstGEP1_32( 983 Int8Ty, IRB.CreateBitCast(AI, Int8PtrTy), 984 AlignedSize - 1)); 985 } 986 } 987 } 988 989 unsigned HWAddressSanitizer::retagMask(unsigned AllocaNo) { 990 if (TargetTriple.getArch() == Triple::x86_64) 991 return AllocaNo & TagMaskByte; 992 993 // A list of 8-bit numbers that have at most one run of non-zero bits. 994 // x = x ^ (mask << 56) can be encoded as a single armv8 instruction for these 995 // masks. 996 // The list does not include the value 255, which is used for UAR. 997 // 998 // Because we are more likely to use earlier elements of this list than later 999 // ones, it is sorted in increasing order of probability of collision with a 1000 // mask allocated (temporally) nearby. The program that generated this list 1001 // can be found at: 1002 // https://github.com/google/sanitizers/blob/master/hwaddress-sanitizer/sort_masks.py 1003 static unsigned FastMasks[] = {0, 128, 64, 192, 32, 96, 224, 112, 240, 1004 48, 16, 120, 248, 56, 24, 8, 124, 252, 1005 60, 28, 12, 4, 126, 254, 62, 30, 14, 1006 6, 2, 127, 63, 31, 15, 7, 3, 1}; 1007 return FastMasks[AllocaNo % (sizeof(FastMasks) / sizeof(FastMasks[0]))]; 1008 } 1009 1010 Value *HWAddressSanitizer::applyTagMask(IRBuilder<> &IRB, Value *OldTag) { 1011 if (TargetTriple.getArch() == Triple::x86_64) { 1012 Constant *TagMask = ConstantInt::get(IntptrTy, TagMaskByte); 1013 Value *NewTag = IRB.CreateAnd(OldTag, TagMask); 1014 return NewTag; 1015 } 1016 // aarch64 uses 8-bit tags, so no mask is needed. 1017 return OldTag; 1018 } 1019 1020 Value *HWAddressSanitizer::getNextTagWithCall(IRBuilder<> &IRB) { 1021 return IRB.CreateZExt(IRB.CreateCall(HwasanGenerateTagFunc), IntptrTy); 1022 } 1023 1024 Value *HWAddressSanitizer::getStackBaseTag(IRBuilder<> &IRB) { 1025 if (ClGenerateTagsWithCalls) 1026 return getNextTagWithCall(IRB); 1027 if (StackBaseTag) 1028 return StackBaseTag; 1029 // Extract some entropy from the stack pointer for the tags. 1030 // Take bits 20..28 (ASLR entropy) and xor with bits 0..8 (these differ 1031 // between functions). 1032 Value *StackPointerLong = getSP(IRB); 1033 Value *StackTag = 1034 applyTagMask(IRB, IRB.CreateXor(StackPointerLong, 1035 IRB.CreateLShr(StackPointerLong, 20))); 1036 StackTag->setName("hwasan.stack.base.tag"); 1037 return StackTag; 1038 } 1039 1040 Value *HWAddressSanitizer::getAllocaTag(IRBuilder<> &IRB, Value *StackTag, 1041 AllocaInst *AI, unsigned AllocaNo) { 1042 if (ClGenerateTagsWithCalls) 1043 return getNextTagWithCall(IRB); 1044 return IRB.CreateXor(StackTag, 1045 ConstantInt::get(IntptrTy, retagMask(AllocaNo))); 1046 } 1047 1048 Value *HWAddressSanitizer::getUARTag(IRBuilder<> &IRB, Value *StackTag) { 1049 if (ClUARRetagToZero) 1050 return ConstantInt::get(IntptrTy, 0); 1051 if (ClGenerateTagsWithCalls) 1052 return getNextTagWithCall(IRB); 1053 return IRB.CreateXor(StackTag, ConstantInt::get(IntptrTy, TagMaskByte)); 1054 } 1055 1056 // Add a tag to an address. 1057 Value *HWAddressSanitizer::tagPointer(IRBuilder<> &IRB, Type *Ty, 1058 Value *PtrLong, Value *Tag) { 1059 assert(!UsePageAliases); 1060 Value *TaggedPtrLong; 1061 if (CompileKernel) { 1062 // Kernel addresses have 0xFF in the most significant byte. 1063 Value *ShiftedTag = 1064 IRB.CreateOr(IRB.CreateShl(Tag, PointerTagShift), 1065 ConstantInt::get(IntptrTy, (1ULL << PointerTagShift) - 1)); 1066 TaggedPtrLong = IRB.CreateAnd(PtrLong, ShiftedTag); 1067 } else { 1068 // Userspace can simply do OR (tag << PointerTagShift); 1069 Value *ShiftedTag = IRB.CreateShl(Tag, PointerTagShift); 1070 TaggedPtrLong = IRB.CreateOr(PtrLong, ShiftedTag); 1071 } 1072 return IRB.CreateIntToPtr(TaggedPtrLong, Ty); 1073 } 1074 1075 // Remove tag from an address. 1076 Value *HWAddressSanitizer::untagPointer(IRBuilder<> &IRB, Value *PtrLong) { 1077 assert(!UsePageAliases); 1078 Value *UntaggedPtrLong; 1079 if (CompileKernel) { 1080 // Kernel addresses have 0xFF in the most significant byte. 1081 UntaggedPtrLong = 1082 IRB.CreateOr(PtrLong, ConstantInt::get(PtrLong->getType(), 1083 0xFFULL << PointerTagShift)); 1084 } else { 1085 // Userspace addresses have 0x00. 1086 UntaggedPtrLong = 1087 IRB.CreateAnd(PtrLong, ConstantInt::get(PtrLong->getType(), 1088 ~(0xFFULL << PointerTagShift))); 1089 } 1090 return UntaggedPtrLong; 1091 } 1092 1093 Value *HWAddressSanitizer::getHwasanThreadSlotPtr(IRBuilder<> &IRB, Type *Ty) { 1094 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1095 if (TargetTriple.isAArch64() && TargetTriple.isAndroid()) { 1096 // Android provides a fixed TLS slot for sanitizers. See TLS_SLOT_SANITIZER 1097 // in Bionic's libc/private/bionic_tls.h. 1098 Function *ThreadPointerFunc = 1099 Intrinsic::getDeclaration(M, Intrinsic::thread_pointer); 1100 Value *SlotPtr = IRB.CreatePointerCast( 1101 IRB.CreateConstGEP1_32(IRB.getInt8Ty(), 1102 IRB.CreateCall(ThreadPointerFunc), 0x30), 1103 Ty->getPointerTo(0)); 1104 return SlotPtr; 1105 } 1106 if (ThreadPtrGlobal) 1107 return ThreadPtrGlobal; 1108 1109 return nullptr; 1110 } 1111 1112 Value *HWAddressSanitizer::getPC(IRBuilder<> &IRB) { 1113 if (TargetTriple.getArch() == Triple::aarch64) 1114 return readRegister(IRB, "pc"); 1115 else 1116 return IRB.CreatePtrToInt(IRB.GetInsertBlock()->getParent(), IntptrTy); 1117 } 1118 1119 Value *HWAddressSanitizer::getSP(IRBuilder<> &IRB) { 1120 if (!CachedSP) { 1121 // FIXME: use addressofreturnaddress (but implement it in aarch64 backend 1122 // first). 1123 Function *F = IRB.GetInsertBlock()->getParent(); 1124 Module *M = F->getParent(); 1125 auto GetStackPointerFn = Intrinsic::getDeclaration( 1126 M, Intrinsic::frameaddress, 1127 IRB.getInt8PtrTy(M->getDataLayout().getAllocaAddrSpace())); 1128 CachedSP = IRB.CreatePtrToInt( 1129 IRB.CreateCall(GetStackPointerFn, 1130 {Constant::getNullValue(IRB.getInt32Ty())}), 1131 IntptrTy); 1132 } 1133 return CachedSP; 1134 } 1135 1136 Value *HWAddressSanitizer::getFrameRecordInfo(IRBuilder<> &IRB) { 1137 // Prepare ring buffer data. 1138 Value *PC = getPC(IRB); 1139 Value *SP = getSP(IRB); 1140 1141 // Mix SP and PC. 1142 // Assumptions: 1143 // PC is 0x0000PPPPPPPPPPPP (48 bits are meaningful, others are zero) 1144 // SP is 0xsssssssssssSSSS0 (4 lower bits are zero) 1145 // We only really need ~20 lower non-zero bits (SSSS), so we mix like this: 1146 // 0xSSSSPPPPPPPPPPPP 1147 SP = IRB.CreateShl(SP, 44); 1148 return IRB.CreateOr(PC, SP); 1149 } 1150 1151 void HWAddressSanitizer::emitPrologue(IRBuilder<> &IRB, bool WithFrameRecord) { 1152 if (!Mapping.InTls) 1153 ShadowBase = getShadowNonTls(IRB); 1154 else if (!WithFrameRecord && TargetTriple.isAndroid()) 1155 ShadowBase = getDynamicShadowIfunc(IRB); 1156 1157 if (!WithFrameRecord && ShadowBase) 1158 return; 1159 1160 Value *SlotPtr = getHwasanThreadSlotPtr(IRB, IntptrTy); 1161 assert(SlotPtr); 1162 1163 Value *ThreadLong = IRB.CreateLoad(IntptrTy, SlotPtr); 1164 // Extract the address field from ThreadLong. Unnecessary on AArch64 with TBI. 1165 Value *ThreadLongMaybeUntagged = 1166 TargetTriple.isAArch64() ? ThreadLong : untagPointer(IRB, ThreadLong); 1167 1168 if (WithFrameRecord) { 1169 StackBaseTag = IRB.CreateAShr(ThreadLong, 3); 1170 1171 // Store data to ring buffer. 1172 Value *FrameRecordInfo = getFrameRecordInfo(IRB); 1173 Value *RecordPtr = 1174 IRB.CreateIntToPtr(ThreadLongMaybeUntagged, IntptrTy->getPointerTo(0)); 1175 IRB.CreateStore(FrameRecordInfo, RecordPtr); 1176 1177 // Update the ring buffer. Top byte of ThreadLong defines the size of the 1178 // buffer in pages, it must be a power of two, and the start of the buffer 1179 // must be aligned by twice that much. Therefore wrap around of the ring 1180 // buffer is simply Addr &= ~((ThreadLong >> 56) << 12). 1181 // The use of AShr instead of LShr is due to 1182 // https://bugs.llvm.org/show_bug.cgi?id=39030 1183 // Runtime library makes sure not to use the highest bit. 1184 Value *WrapMask = IRB.CreateXor( 1185 IRB.CreateShl(IRB.CreateAShr(ThreadLong, 56), 12, "", true, true), 1186 ConstantInt::get(IntptrTy, (uint64_t)-1)); 1187 Value *ThreadLongNew = IRB.CreateAnd( 1188 IRB.CreateAdd(ThreadLong, ConstantInt::get(IntptrTy, 8)), WrapMask); 1189 IRB.CreateStore(ThreadLongNew, SlotPtr); 1190 } 1191 1192 if (!ShadowBase) { 1193 // Get shadow base address by aligning RecordPtr up. 1194 // Note: this is not correct if the pointer is already aligned. 1195 // Runtime library will make sure this never happens. 1196 ShadowBase = IRB.CreateAdd( 1197 IRB.CreateOr( 1198 ThreadLongMaybeUntagged, 1199 ConstantInt::get(IntptrTy, (1ULL << kShadowBaseAlignment) - 1)), 1200 ConstantInt::get(IntptrTy, 1), "hwasan.shadow"); 1201 ShadowBase = IRB.CreateIntToPtr(ShadowBase, Int8PtrTy); 1202 } 1203 } 1204 1205 Value *HWAddressSanitizer::readRegister(IRBuilder<> &IRB, StringRef Name) { 1206 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1207 Function *ReadRegister = 1208 Intrinsic::getDeclaration(M, Intrinsic::read_register, IntptrTy); 1209 MDNode *MD = MDNode::get(*C, {MDString::get(*C, Name)}); 1210 Value *Args[] = {MetadataAsValue::get(*C, MD)}; 1211 return IRB.CreateCall(ReadRegister, Args); 1212 } 1213 1214 bool HWAddressSanitizer::instrumentLandingPads( 1215 SmallVectorImpl<Instruction *> &LandingPadVec) { 1216 for (auto *LP : LandingPadVec) { 1217 IRBuilder<> IRB(LP->getNextNode()); 1218 IRB.CreateCall( 1219 HWAsanHandleVfork, 1220 {readRegister(IRB, (TargetTriple.getArch() == Triple::x86_64) ? "rsp" 1221 : "sp")}); 1222 } 1223 return true; 1224 } 1225 1226 static bool isLifetimeIntrinsic(Value *V) { 1227 auto *II = dyn_cast<IntrinsicInst>(V); 1228 return II && II->isLifetimeStartOrEnd(); 1229 } 1230 1231 bool HWAddressSanitizer::instrumentStack(memtag::StackInfo &SInfo, 1232 Value *StackTag, 1233 const DominatorTree &DT, 1234 const PostDominatorTree &PDT, 1235 const LoopInfo &LI) { 1236 // Ideally, we want to calculate tagged stack base pointer, and rewrite all 1237 // alloca addresses using that. Unfortunately, offsets are not known yet 1238 // (unless we use ASan-style mega-alloca). Instead we keep the base tag in a 1239 // temp, shift-OR it into each alloca address and xor with the retag mask. 1240 // This generates one extra instruction per alloca use. 1241 unsigned int I = 0; 1242 1243 for (auto &KV : SInfo.AllocasToInstrument) { 1244 auto N = I++; 1245 auto *AI = KV.first; 1246 memtag::AllocaInfo &Info = KV.second; 1247 IRBuilder<> IRB(AI->getNextNode()); 1248 1249 // Replace uses of the alloca with tagged address. 1250 Value *Tag = getAllocaTag(IRB, StackTag, AI, N); 1251 Value *AILong = IRB.CreatePointerCast(AI, IntptrTy); 1252 Value *Replacement = tagPointer(IRB, AI->getType(), AILong, Tag); 1253 std::string Name = 1254 AI->hasName() ? AI->getName().str() : "alloca." + itostr(N); 1255 Replacement->setName(Name + ".hwasan"); 1256 1257 size_t Size = memtag::getAllocaSizeInBytes(*AI); 1258 size_t AlignedSize = alignTo(Size, Mapping.getObjectAlignment()); 1259 1260 Value *AICast = IRB.CreatePointerCast(AI, Int8PtrTy); 1261 1262 auto HandleLifetime = [&](IntrinsicInst *II) { 1263 // Set the lifetime intrinsic to cover the whole alloca. This reduces the 1264 // set of assumptions we need to make about the lifetime. Without this we 1265 // would need to ensure that we can track the lifetime pointer to a 1266 // constant offset from the alloca, and would still need to change the 1267 // size to include the extra alignment we use for the untagging to make 1268 // the size consistent. 1269 // 1270 // The check for standard lifetime below makes sure that we have exactly 1271 // one set of start / end in any execution (i.e. the ends are not 1272 // reachable from each other), so this will not cause any problems. 1273 II->setArgOperand(0, ConstantInt::get(Int64Ty, AlignedSize)); 1274 II->setArgOperand(1, AICast); 1275 }; 1276 llvm::for_each(Info.LifetimeStart, HandleLifetime); 1277 llvm::for_each(Info.LifetimeEnd, HandleLifetime); 1278 1279 AI->replaceUsesWithIf(Replacement, [AICast, AILong](Use &U) { 1280 auto *User = U.getUser(); 1281 return User != AILong && User != AICast && !isLifetimeIntrinsic(User); 1282 }); 1283 1284 for (auto *DDI : Info.DbgVariableIntrinsics) { 1285 // Prepend "tag_offset, N" to the dwarf expression. 1286 // Tag offset logically applies to the alloca pointer, and it makes sense 1287 // to put it at the beginning of the expression. 1288 SmallVector<uint64_t, 8> NewOps = {dwarf::DW_OP_LLVM_tag_offset, 1289 retagMask(N)}; 1290 for (size_t LocNo = 0; LocNo < DDI->getNumVariableLocationOps(); ++LocNo) 1291 if (DDI->getVariableLocationOp(LocNo) == AI) 1292 DDI->setExpression(DIExpression::appendOpsToArg(DDI->getExpression(), 1293 NewOps, LocNo)); 1294 } 1295 1296 auto TagEnd = [&](Instruction *Node) { 1297 IRB.SetInsertPoint(Node); 1298 Value *UARTag = getUARTag(IRB, StackTag); 1299 // When untagging, use the `AlignedSize` because we need to set the tags 1300 // for the entire alloca to zero. If we used `Size` here, we would 1301 // keep the last granule tagged, and store zero in the last byte of the 1302 // last granule, due to how short granules are implemented. 1303 tagAlloca(IRB, AI, UARTag, AlignedSize); 1304 }; 1305 // Calls to functions that may return twice (e.g. setjmp) confuse the 1306 // postdominator analysis, and will leave us to keep memory tagged after 1307 // function return. Work around this by always untagging at every return 1308 // statement if return_twice functions are called. 1309 bool StandardLifetime = 1310 SInfo.UnrecognizedLifetimes.empty() && 1311 memtag::isStandardLifetime(Info.LifetimeStart, Info.LifetimeEnd, &DT, 1312 &LI, ClMaxLifetimes) && 1313 !SInfo.CallsReturnTwice; 1314 if (DetectUseAfterScope && StandardLifetime) { 1315 IntrinsicInst *Start = Info.LifetimeStart[0]; 1316 IRB.SetInsertPoint(Start->getNextNode()); 1317 tagAlloca(IRB, AI, Tag, Size); 1318 if (!memtag::forAllReachableExits(DT, PDT, LI, Start, Info.LifetimeEnd, 1319 SInfo.RetVec, TagEnd)) { 1320 for (auto *End : Info.LifetimeEnd) 1321 End->eraseFromParent(); 1322 } 1323 } else { 1324 tagAlloca(IRB, AI, Tag, Size); 1325 for (auto *RI : SInfo.RetVec) 1326 TagEnd(RI); 1327 // We inserted tagging outside of the lifetimes, so we have to remove 1328 // them. 1329 for (auto &II : Info.LifetimeStart) 1330 II->eraseFromParent(); 1331 for (auto &II : Info.LifetimeEnd) 1332 II->eraseFromParent(); 1333 } 1334 memtag::alignAndPadAlloca(Info, Align(Mapping.getObjectAlignment())); 1335 } 1336 for (auto &I : SInfo.UnrecognizedLifetimes) 1337 I->eraseFromParent(); 1338 return true; 1339 } 1340 1341 bool HWAddressSanitizer::isInterestingAlloca(const AllocaInst &AI) { 1342 return (AI.getAllocatedType()->isSized() && 1343 // FIXME: instrument dynamic allocas, too 1344 AI.isStaticAlloca() && 1345 // alloca() may be called with 0 size, ignore it. 1346 memtag::getAllocaSizeInBytes(AI) > 0 && 1347 // We are only interested in allocas not promotable to registers. 1348 // Promotable allocas are common under -O0. 1349 !isAllocaPromotable(&AI) && 1350 // inalloca allocas are not treated as static, and we don't want 1351 // dynamic alloca instrumentation for them as well. 1352 !AI.isUsedWithInAlloca() && 1353 // swifterror allocas are register promoted by ISel 1354 !AI.isSwiftError()) && 1355 // safe allocas are not interesting 1356 !(SSI && SSI->isSafe(AI)); 1357 } 1358 1359 bool HWAddressSanitizer::sanitizeFunction(Function &F, 1360 FunctionAnalysisManager &FAM) { 1361 if (&F == HwasanCtorFunction) 1362 return false; 1363 1364 if (!F.hasFnAttribute(Attribute::SanitizeHWAddress)) 1365 return false; 1366 1367 LLVM_DEBUG(dbgs() << "Function: " << F.getName() << "\n"); 1368 1369 SmallVector<InterestingMemoryOperand, 16> OperandsToInstrument; 1370 SmallVector<MemIntrinsic *, 16> IntrinToInstrument; 1371 SmallVector<Instruction *, 8> LandingPadVec; 1372 1373 memtag::StackInfoBuilder SIB( 1374 [this](const AllocaInst &AI) { return isInterestingAlloca(AI); }); 1375 for (auto &Inst : instructions(F)) { 1376 if (InstrumentStack) { 1377 SIB.visit(Inst); 1378 } 1379 1380 if (InstrumentLandingPads && isa<LandingPadInst>(Inst)) 1381 LandingPadVec.push_back(&Inst); 1382 1383 getInterestingMemoryOperands(&Inst, OperandsToInstrument); 1384 1385 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(&Inst)) 1386 if (!ignoreMemIntrinsic(MI)) 1387 IntrinToInstrument.push_back(MI); 1388 } 1389 1390 memtag::StackInfo &SInfo = SIB.get(); 1391 1392 initializeCallbacks(*F.getParent()); 1393 1394 bool Changed = false; 1395 1396 if (!LandingPadVec.empty()) 1397 Changed |= instrumentLandingPads(LandingPadVec); 1398 1399 if (SInfo.AllocasToInstrument.empty() && F.hasPersonalityFn() && 1400 F.getPersonalityFn()->getName() == kHwasanPersonalityThunkName) { 1401 // __hwasan_personality_thunk is a no-op for functions without an 1402 // instrumented stack, so we can drop it. 1403 F.setPersonalityFn(nullptr); 1404 Changed = true; 1405 } 1406 1407 if (SInfo.AllocasToInstrument.empty() && OperandsToInstrument.empty() && 1408 IntrinToInstrument.empty()) 1409 return Changed; 1410 1411 assert(!ShadowBase); 1412 1413 Instruction *InsertPt = &*F.getEntryBlock().begin(); 1414 IRBuilder<> EntryIRB(InsertPt); 1415 emitPrologue(EntryIRB, 1416 /*WithFrameRecord*/ ClRecordStackHistory && 1417 Mapping.WithFrameRecord && 1418 !SInfo.AllocasToInstrument.empty()); 1419 1420 if (!SInfo.AllocasToInstrument.empty()) { 1421 const DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(F); 1422 const PostDominatorTree &PDT = FAM.getResult<PostDominatorTreeAnalysis>(F); 1423 const LoopInfo &LI = FAM.getResult<LoopAnalysis>(F); 1424 Value *StackTag = 1425 ClGenerateTagsWithCalls ? nullptr : getStackBaseTag(EntryIRB); 1426 instrumentStack(SInfo, StackTag, DT, PDT, LI); 1427 } 1428 1429 // If we split the entry block, move any allocas that were originally in the 1430 // entry block back into the entry block so that they aren't treated as 1431 // dynamic allocas. 1432 if (EntryIRB.GetInsertBlock() != &F.getEntryBlock()) { 1433 InsertPt = &*F.getEntryBlock().begin(); 1434 for (Instruction &I : 1435 llvm::make_early_inc_range(*EntryIRB.GetInsertBlock())) { 1436 if (auto *AI = dyn_cast<AllocaInst>(&I)) 1437 if (isa<ConstantInt>(AI->getArraySize())) 1438 I.moveBefore(InsertPt); 1439 } 1440 } 1441 1442 for (auto &Operand : OperandsToInstrument) 1443 instrumentMemAccess(Operand); 1444 1445 if (ClInstrumentMemIntrinsics && !IntrinToInstrument.empty()) { 1446 for (auto Inst : IntrinToInstrument) 1447 instrumentMemIntrinsic(cast<MemIntrinsic>(Inst)); 1448 } 1449 1450 ShadowBase = nullptr; 1451 StackBaseTag = nullptr; 1452 CachedSP = nullptr; 1453 1454 return true; 1455 } 1456 1457 void HWAddressSanitizer::instrumentGlobal(GlobalVariable *GV, uint8_t Tag) { 1458 assert(!UsePageAliases); 1459 Constant *Initializer = GV->getInitializer(); 1460 uint64_t SizeInBytes = 1461 M.getDataLayout().getTypeAllocSize(Initializer->getType()); 1462 uint64_t NewSize = alignTo(SizeInBytes, Mapping.getObjectAlignment()); 1463 if (SizeInBytes != NewSize) { 1464 // Pad the initializer out to the next multiple of 16 bytes and add the 1465 // required short granule tag. 1466 std::vector<uint8_t> Init(NewSize - SizeInBytes, 0); 1467 Init.back() = Tag; 1468 Constant *Padding = ConstantDataArray::get(*C, Init); 1469 Initializer = ConstantStruct::getAnon({Initializer, Padding}); 1470 } 1471 1472 auto *NewGV = new GlobalVariable(M, Initializer->getType(), GV->isConstant(), 1473 GlobalValue::ExternalLinkage, Initializer, 1474 GV->getName() + ".hwasan"); 1475 NewGV->copyAttributesFrom(GV); 1476 NewGV->setLinkage(GlobalValue::PrivateLinkage); 1477 NewGV->copyMetadata(GV, 0); 1478 NewGV->setAlignment( 1479 MaybeAlign(std::max(GV->getAlignment(), Mapping.getObjectAlignment()))); 1480 1481 // It is invalid to ICF two globals that have different tags. In the case 1482 // where the size of the global is a multiple of the tag granularity the 1483 // contents of the globals may be the same but the tags (i.e. symbol values) 1484 // may be different, and the symbols are not considered during ICF. In the 1485 // case where the size is not a multiple of the granularity, the short granule 1486 // tags would discriminate two globals with different tags, but there would 1487 // otherwise be nothing stopping such a global from being incorrectly ICF'd 1488 // with an uninstrumented (i.e. tag 0) global that happened to have the short 1489 // granule tag in the last byte. 1490 NewGV->setUnnamedAddr(GlobalValue::UnnamedAddr::None); 1491 1492 // Descriptor format (assuming little-endian): 1493 // bytes 0-3: relative address of global 1494 // bytes 4-6: size of global (16MB ought to be enough for anyone, but in case 1495 // it isn't, we create multiple descriptors) 1496 // byte 7: tag 1497 auto *DescriptorTy = StructType::get(Int32Ty, Int32Ty); 1498 const uint64_t MaxDescriptorSize = 0xfffff0; 1499 for (uint64_t DescriptorPos = 0; DescriptorPos < SizeInBytes; 1500 DescriptorPos += MaxDescriptorSize) { 1501 auto *Descriptor = 1502 new GlobalVariable(M, DescriptorTy, true, GlobalValue::PrivateLinkage, 1503 nullptr, GV->getName() + ".hwasan.descriptor"); 1504 auto *GVRelPtr = ConstantExpr::getTrunc( 1505 ConstantExpr::getAdd( 1506 ConstantExpr::getSub( 1507 ConstantExpr::getPtrToInt(NewGV, Int64Ty), 1508 ConstantExpr::getPtrToInt(Descriptor, Int64Ty)), 1509 ConstantInt::get(Int64Ty, DescriptorPos)), 1510 Int32Ty); 1511 uint32_t Size = std::min(SizeInBytes - DescriptorPos, MaxDescriptorSize); 1512 auto *SizeAndTag = ConstantInt::get(Int32Ty, Size | (uint32_t(Tag) << 24)); 1513 Descriptor->setComdat(NewGV->getComdat()); 1514 Descriptor->setInitializer(ConstantStruct::getAnon({GVRelPtr, SizeAndTag})); 1515 Descriptor->setSection("hwasan_globals"); 1516 Descriptor->setMetadata(LLVMContext::MD_associated, 1517 MDNode::get(*C, ValueAsMetadata::get(NewGV))); 1518 appendToCompilerUsed(M, Descriptor); 1519 } 1520 1521 Constant *Aliasee = ConstantExpr::getIntToPtr( 1522 ConstantExpr::getAdd( 1523 ConstantExpr::getPtrToInt(NewGV, Int64Ty), 1524 ConstantInt::get(Int64Ty, uint64_t(Tag) << PointerTagShift)), 1525 GV->getType()); 1526 auto *Alias = GlobalAlias::create(GV->getValueType(), GV->getAddressSpace(), 1527 GV->getLinkage(), "", Aliasee, &M); 1528 Alias->setVisibility(GV->getVisibility()); 1529 Alias->takeName(GV); 1530 GV->replaceAllUsesWith(Alias); 1531 GV->eraseFromParent(); 1532 } 1533 1534 void HWAddressSanitizer::instrumentGlobals() { 1535 std::vector<GlobalVariable *> Globals; 1536 for (GlobalVariable &GV : M.globals()) { 1537 if (GV.hasSanitizerMetadata() && GV.getSanitizerMetadata().NoHWAddress) 1538 continue; 1539 1540 if (GV.isDeclarationForLinker() || GV.getName().startswith("llvm.") || 1541 GV.isThreadLocal()) 1542 continue; 1543 1544 // Common symbols can't have aliases point to them, so they can't be tagged. 1545 if (GV.hasCommonLinkage()) 1546 continue; 1547 1548 // Globals with custom sections may be used in __start_/__stop_ enumeration, 1549 // which would be broken both by adding tags and potentially by the extra 1550 // padding/alignment that we insert. 1551 if (GV.hasSection()) 1552 continue; 1553 1554 Globals.push_back(&GV); 1555 } 1556 1557 MD5 Hasher; 1558 Hasher.update(M.getSourceFileName()); 1559 MD5::MD5Result Hash; 1560 Hasher.final(Hash); 1561 uint8_t Tag = Hash[0]; 1562 1563 for (GlobalVariable *GV : Globals) { 1564 Tag &= TagMaskByte; 1565 // Skip tag 0 in order to avoid collisions with untagged memory. 1566 if (Tag == 0) 1567 Tag = 1; 1568 instrumentGlobal(GV, Tag++); 1569 } 1570 } 1571 1572 void HWAddressSanitizer::instrumentPersonalityFunctions() { 1573 // We need to untag stack frames as we unwind past them. That is the job of 1574 // the personality function wrapper, which either wraps an existing 1575 // personality function or acts as a personality function on its own. Each 1576 // function that has a personality function or that can be unwound past has 1577 // its personality function changed to a thunk that calls the personality 1578 // function wrapper in the runtime. 1579 MapVector<Constant *, std::vector<Function *>> PersonalityFns; 1580 for (Function &F : M) { 1581 if (F.isDeclaration() || !F.hasFnAttribute(Attribute::SanitizeHWAddress)) 1582 continue; 1583 1584 if (F.hasPersonalityFn()) { 1585 PersonalityFns[F.getPersonalityFn()->stripPointerCasts()].push_back(&F); 1586 } else if (!F.hasFnAttribute(Attribute::NoUnwind)) { 1587 PersonalityFns[nullptr].push_back(&F); 1588 } 1589 } 1590 1591 if (PersonalityFns.empty()) 1592 return; 1593 1594 FunctionCallee HwasanPersonalityWrapper = M.getOrInsertFunction( 1595 "__hwasan_personality_wrapper", Int32Ty, Int32Ty, Int32Ty, Int64Ty, 1596 Int8PtrTy, Int8PtrTy, Int8PtrTy, Int8PtrTy, Int8PtrTy); 1597 FunctionCallee UnwindGetGR = M.getOrInsertFunction("_Unwind_GetGR", VoidTy); 1598 FunctionCallee UnwindGetCFA = M.getOrInsertFunction("_Unwind_GetCFA", VoidTy); 1599 1600 for (auto &P : PersonalityFns) { 1601 std::string ThunkName = kHwasanPersonalityThunkName; 1602 if (P.first) 1603 ThunkName += ("." + P.first->getName()).str(); 1604 FunctionType *ThunkFnTy = FunctionType::get( 1605 Int32Ty, {Int32Ty, Int32Ty, Int64Ty, Int8PtrTy, Int8PtrTy}, false); 1606 bool IsLocal = P.first && (!isa<GlobalValue>(P.first) || 1607 cast<GlobalValue>(P.first)->hasLocalLinkage()); 1608 auto *ThunkFn = Function::Create(ThunkFnTy, 1609 IsLocal ? GlobalValue::InternalLinkage 1610 : GlobalValue::LinkOnceODRLinkage, 1611 ThunkName, &M); 1612 if (!IsLocal) { 1613 ThunkFn->setVisibility(GlobalValue::HiddenVisibility); 1614 ThunkFn->setComdat(M.getOrInsertComdat(ThunkName)); 1615 } 1616 1617 auto *BB = BasicBlock::Create(*C, "entry", ThunkFn); 1618 IRBuilder<> IRB(BB); 1619 CallInst *WrapperCall = IRB.CreateCall( 1620 HwasanPersonalityWrapper, 1621 {ThunkFn->getArg(0), ThunkFn->getArg(1), ThunkFn->getArg(2), 1622 ThunkFn->getArg(3), ThunkFn->getArg(4), 1623 P.first ? IRB.CreateBitCast(P.first, Int8PtrTy) 1624 : Constant::getNullValue(Int8PtrTy), 1625 IRB.CreateBitCast(UnwindGetGR.getCallee(), Int8PtrTy), 1626 IRB.CreateBitCast(UnwindGetCFA.getCallee(), Int8PtrTy)}); 1627 WrapperCall->setTailCall(); 1628 IRB.CreateRet(WrapperCall); 1629 1630 for (Function *F : P.second) 1631 F->setPersonalityFn(ThunkFn); 1632 } 1633 } 1634 1635 void HWAddressSanitizer::ShadowMapping::init(Triple &TargetTriple, 1636 bool InstrumentWithCalls) { 1637 Scale = kDefaultShadowScale; 1638 if (TargetTriple.isOSFuchsia()) { 1639 // Fuchsia is always PIE, which means that the beginning of the address 1640 // space is always available. 1641 InGlobal = false; 1642 InTls = false; 1643 Offset = 0; 1644 WithFrameRecord = true; 1645 } else if (ClMappingOffset.getNumOccurrences() > 0) { 1646 InGlobal = false; 1647 InTls = false; 1648 Offset = ClMappingOffset; 1649 WithFrameRecord = false; 1650 } else if (ClEnableKhwasan || InstrumentWithCalls) { 1651 InGlobal = false; 1652 InTls = false; 1653 Offset = 0; 1654 WithFrameRecord = false; 1655 } else if (ClWithIfunc) { 1656 InGlobal = true; 1657 InTls = false; 1658 Offset = kDynamicShadowSentinel; 1659 WithFrameRecord = false; 1660 } else if (ClWithTls) { 1661 InGlobal = false; 1662 InTls = true; 1663 Offset = kDynamicShadowSentinel; 1664 WithFrameRecord = true; 1665 } else { 1666 InGlobal = false; 1667 InTls = false; 1668 Offset = kDynamicShadowSentinel; 1669 WithFrameRecord = false; 1670 } 1671 } 1672