| 1 | //===- HWAddressSanitizer.cpp - memory access error detector --------------===// |
| 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/Statistic.h" |
| 19 | #include "llvm/ADT/StringExtras.h" |
| 20 | #include "llvm/ADT/StringRef.h" |
| 21 | #include "llvm/Analysis/BlockFrequencyInfo.h" |
| 22 | #include "llvm/Analysis/DomTreeUpdater.h" |
| 23 | #include "llvm/Analysis/GlobalsModRef.h" |
| 24 | #include "llvm/Analysis/OptimizationRemarkEmitter.h" |
| 25 | #include "llvm/Analysis/PostDominators.h" |
| 26 | #include "llvm/Analysis/ProfileSummaryInfo.h" |
| 27 | #include "llvm/Analysis/StackSafetyAnalysis.h" |
| 28 | #include "llvm/Analysis/TargetLibraryInfo.h" |
| 29 | #include "llvm/Analysis/ValueTracking.h" |
| 30 | #include "llvm/BinaryFormat/Dwarf.h" |
| 31 | #include "llvm/BinaryFormat/ELF.h" |
| 32 | #include "llvm/IR/Attributes.h" |
| 33 | #include "llvm/IR/BasicBlock.h" |
| 34 | #include "llvm/IR/Constant.h" |
| 35 | #include "llvm/IR/Constants.h" |
| 36 | #include "llvm/IR/DataLayout.h" |
| 37 | #include "llvm/IR/DerivedTypes.h" |
| 38 | #include "llvm/IR/Dominators.h" |
| 39 | #include "llvm/IR/Function.h" |
| 40 | #include "llvm/IR/IRBuilder.h" |
| 41 | #include "llvm/IR/InlineAsm.h" |
| 42 | #include "llvm/IR/InstIterator.h" |
| 43 | #include "llvm/IR/Instruction.h" |
| 44 | #include "llvm/IR/Instructions.h" |
| 45 | #include "llvm/IR/IntrinsicInst.h" |
| 46 | #include "llvm/IR/Intrinsics.h" |
| 47 | #include "llvm/IR/LLVMContext.h" |
| 48 | #include "llvm/IR/MDBuilder.h" |
| 49 | #include "llvm/IR/Module.h" |
| 50 | #include "llvm/IR/Type.h" |
| 51 | #include "llvm/IR/Value.h" |
| 52 | #include "llvm/Support/Casting.h" |
| 53 | #include "llvm/Support/CommandLine.h" |
| 54 | #include "llvm/Support/Debug.h" |
| 55 | #include "llvm/Support/MD5.h" |
| 56 | #include "llvm/Support/RandomNumberGenerator.h" |
| 57 | #include "llvm/Support/raw_ostream.h" |
| 58 | #include "llvm/TargetParser/Triple.h" |
| 59 | #include "llvm/Transforms/Instrumentation/AddressSanitizerCommon.h" |
| 60 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
| 61 | #include "llvm/Transforms/Utils/Instrumentation.h" |
| 62 | #include "llvm/Transforms/Utils/Local.h" |
| 63 | #include "llvm/Transforms/Utils/MemoryTaggingSupport.h" |
| 64 | #include "llvm/Transforms/Utils/ModuleUtils.h" |
| 65 | #include "llvm/Transforms/Utils/PromoteMemToReg.h" |
| 66 | #include <optional> |
| 67 | #include <random> |
| 68 | |
| 69 | using namespace llvm; |
| 70 | |
| 71 | #define DEBUG_TYPE "hwasan" |
| 72 | |
| 73 | const char kHwasanModuleCtorName[] = "hwasan.module_ctor" ; |
| 74 | const char kHwasanNoteName[] = "hwasan.note" ; |
| 75 | const char kHwasanInitName[] = "__hwasan_init" ; |
| 76 | const char kHwasanPersonalityThunkName[] = "__hwasan_personality_thunk" ; |
| 77 | |
| 78 | const char kHwasanShadowMemoryDynamicAddress[] = |
| 79 | "__hwasan_shadow_memory_dynamic_address" ; |
| 80 | |
| 81 | // Accesses sizes are powers of two: 1, 2, 4, 8, 16. |
| 82 | static const size_t kNumberOfAccessSizes = 5; |
| 83 | |
| 84 | static const size_t kDefaultShadowScale = 4; |
| 85 | |
| 86 | static const unsigned kShadowBaseAlignment = 32; |
| 87 | |
| 88 | namespace { |
| 89 | enum class OffsetKind { |
| 90 | kFixed = 0, |
| 91 | kGlobal, |
| 92 | kIfunc, |
| 93 | kTls, |
| 94 | }; |
| 95 | } |
| 96 | |
| 97 | static cl::opt<std::string> |
| 98 | ClMemoryAccessCallbackPrefix("hwasan-memory-access-callback-prefix" , |
| 99 | cl::desc("Prefix for memory access callbacks" ), |
| 100 | cl::Hidden, cl::init(Val: "__hwasan_" )); |
| 101 | |
| 102 | static cl::opt<bool> ClKasanMemIntrinCallbackPrefix( |
| 103 | "hwasan-kernel-mem-intrinsic-prefix" , |
| 104 | cl::desc("Use prefix for memory intrinsics in KASAN mode" ), cl::Hidden, |
| 105 | cl::init(Val: false)); |
| 106 | |
| 107 | static cl::opt<bool> ClInstrumentWithCalls( |
| 108 | "hwasan-instrument-with-calls" , |
| 109 | cl::desc("instrument reads and writes with callbacks" ), cl::Hidden, |
| 110 | cl::init(Val: false)); |
| 111 | |
| 112 | static cl::opt<bool> ClInstrumentReads("hwasan-instrument-reads" , |
| 113 | cl::desc("instrument read instructions" ), |
| 114 | cl::Hidden, cl::init(Val: true)); |
| 115 | |
| 116 | static cl::opt<bool> |
| 117 | ClInstrumentWrites("hwasan-instrument-writes" , |
| 118 | cl::desc("instrument write instructions" ), cl::Hidden, |
| 119 | cl::init(Val: true)); |
| 120 | |
| 121 | static cl::opt<bool> ClInstrumentAtomics( |
| 122 | "hwasan-instrument-atomics" , |
| 123 | cl::desc("instrument atomic instructions (rmw, cmpxchg)" ), cl::Hidden, |
| 124 | cl::init(Val: true)); |
| 125 | |
| 126 | static cl::opt<bool> ClInstrumentByval("hwasan-instrument-byval" , |
| 127 | cl::desc("instrument byval arguments" ), |
| 128 | cl::Hidden, cl::init(Val: true)); |
| 129 | |
| 130 | static cl::opt<bool> |
| 131 | ClRecover("hwasan-recover" , |
| 132 | cl::desc("Enable recovery mode (continue-after-error)." ), |
| 133 | cl::Hidden, cl::init(Val: false)); |
| 134 | |
| 135 | static cl::opt<bool> ClInstrumentStack("hwasan-instrument-stack" , |
| 136 | cl::desc("instrument stack (allocas)" ), |
| 137 | cl::Hidden, cl::init(Val: true)); |
| 138 | |
| 139 | static cl::opt<bool> |
| 140 | ClUseStackSafety("hwasan-use-stack-safety" , cl::Hidden, cl::init(Val: true), |
| 141 | cl::Hidden, cl::desc("Use Stack Safety analysis results" ), |
| 142 | cl::Optional); |
| 143 | |
| 144 | static cl::opt<size_t> ClMaxLifetimes( |
| 145 | "hwasan-max-lifetimes-for-alloca" , cl::Hidden, cl::init(Val: 3), |
| 146 | cl::ReallyHidden, |
| 147 | cl::desc("How many lifetime ends to handle for a single alloca." ), |
| 148 | cl::Optional); |
| 149 | |
| 150 | static cl::opt<bool> |
| 151 | ClUseAfterScope("hwasan-use-after-scope" , |
| 152 | cl::desc("detect use after scope within function" ), |
| 153 | cl::Hidden, cl::init(Val: true)); |
| 154 | |
| 155 | static cl::opt<bool> ClGenerateTagsWithCalls( |
| 156 | "hwasan-generate-tags-with-calls" , |
| 157 | cl::desc("generate new tags with runtime library calls" ), cl::Hidden, |
| 158 | cl::init(Val: false)); |
| 159 | |
| 160 | static cl::opt<bool> ClGlobals("hwasan-globals" , cl::desc("Instrument globals" ), |
| 161 | cl::Hidden, cl::init(Val: false)); |
| 162 | |
| 163 | static cl::opt<int> ClMatchAllTag( |
| 164 | "hwasan-match-all-tag" , |
| 165 | cl::desc("don't report bad accesses via pointers with this tag" ), |
| 166 | cl::Hidden, cl::init(Val: -1)); |
| 167 | |
| 168 | static cl::opt<bool> |
| 169 | ClEnableKhwasan("hwasan-kernel" , |
| 170 | cl::desc("Enable KernelHWAddressSanitizer instrumentation" ), |
| 171 | cl::Hidden, cl::init(Val: false)); |
| 172 | |
| 173 | // These flags allow to change the shadow mapping and control how shadow memory |
| 174 | // is accessed. The shadow mapping looks like: |
| 175 | // Shadow = (Mem >> scale) + offset |
| 176 | |
| 177 | static cl::opt<uint64_t> |
| 178 | ClMappingOffset("hwasan-mapping-offset" , |
| 179 | cl::desc("HWASan shadow mapping offset [EXPERIMENTAL]" ), |
| 180 | cl::Hidden); |
| 181 | |
| 182 | static cl::opt<OffsetKind> ClMappingOffsetDynamic( |
| 183 | "hwasan-mapping-offset-dynamic" , |
| 184 | cl::desc("HWASan shadow mapping dynamic offset location" ), cl::Hidden, |
| 185 | cl::values(clEnumValN(OffsetKind::kGlobal, "global" , "Use global" ), |
| 186 | clEnumValN(OffsetKind::kIfunc, "ifunc" , "Use ifunc global" ), |
| 187 | clEnumValN(OffsetKind::kTls, "tls" , "Use TLS" ))); |
| 188 | |
| 189 | static cl::opt<bool> |
| 190 | ClFrameRecords("hwasan-with-frame-record" , |
| 191 | cl::desc("Use ring buffer for stack allocations" ), |
| 192 | cl::Hidden); |
| 193 | |
| 194 | static cl::opt<int> ClHotPercentileCutoff("hwasan-percentile-cutoff-hot" , |
| 195 | cl::desc("Hot percentile cutoff." )); |
| 196 | |
| 197 | static cl::opt<float> |
| 198 | ClRandomKeepRate("hwasan-random-rate" , |
| 199 | cl::desc("Probability value in the range [0.0, 1.0] " |
| 200 | "to keep instrumentation of a function. " |
| 201 | "Note: instrumentation can be skipped randomly " |
| 202 | "OR because of the hot percentile cutoff, if " |
| 203 | "both are supplied." )); |
| 204 | |
| 205 | STATISTIC(NumTotalFuncs, "Number of total funcs" ); |
| 206 | STATISTIC(NumInstrumentedFuncs, "Number of instrumented funcs" ); |
| 207 | STATISTIC(NumNoProfileSummaryFuncs, "Number of funcs without PS" ); |
| 208 | |
| 209 | // Mode for selecting how to insert frame record info into the stack ring |
| 210 | // buffer. |
| 211 | enum RecordStackHistoryMode { |
| 212 | // Do not record frame record info. |
| 213 | none, |
| 214 | |
| 215 | // Insert instructions into the prologue for storing into the stack ring |
| 216 | // buffer directly. |
| 217 | instr, |
| 218 | |
| 219 | // Add a call to __hwasan_add_frame_record in the runtime. |
| 220 | libcall, |
| 221 | }; |
| 222 | |
| 223 | static cl::opt<RecordStackHistoryMode> ClRecordStackHistory( |
| 224 | "hwasan-record-stack-history" , |
| 225 | cl::desc("Record stack frames with tagged allocations in a thread-local " |
| 226 | "ring buffer" ), |
| 227 | cl::values(clEnumVal(none, "Do not record stack ring history" ), |
| 228 | clEnumVal(instr, "Insert instructions into the prologue for " |
| 229 | "storing into the stack ring buffer directly" ), |
| 230 | clEnumVal(libcall, "Add a call to __hwasan_add_frame_record for " |
| 231 | "storing into the stack ring buffer" )), |
| 232 | cl::Hidden, cl::init(Val: instr)); |
| 233 | |
| 234 | static cl::opt<bool> |
| 235 | ClInstrumentMemIntrinsics("hwasan-instrument-mem-intrinsics" , |
| 236 | cl::desc("instrument memory intrinsics" ), |
| 237 | cl::Hidden, cl::init(Val: true)); |
| 238 | |
| 239 | static cl::opt<bool> |
| 240 | ClInstrumentLandingPads("hwasan-instrument-landing-pads" , |
| 241 | cl::desc("instrument landing pads" ), cl::Hidden, |
| 242 | cl::init(Val: false)); |
| 243 | |
| 244 | static cl::opt<bool> ClUseShortGranules( |
| 245 | "hwasan-use-short-granules" , |
| 246 | cl::desc("use short granules in allocas and outlined checks" ), cl::Hidden, |
| 247 | cl::init(Val: false)); |
| 248 | |
| 249 | static cl::opt<bool> ClInstrumentPersonalityFunctions( |
| 250 | "hwasan-instrument-personality-functions" , |
| 251 | cl::desc("instrument personality functions" ), cl::Hidden); |
| 252 | |
| 253 | static cl::opt<bool> ClInlineAllChecks("hwasan-inline-all-checks" , |
| 254 | cl::desc("inline all checks" ), |
| 255 | cl::Hidden, cl::init(Val: false)); |
| 256 | |
| 257 | static cl::opt<bool> ClInlineFastPathChecks("hwasan-inline-fast-path-checks" , |
| 258 | cl::desc("inline all checks" ), |
| 259 | cl::Hidden, cl::init(Val: false)); |
| 260 | |
| 261 | // Enabled from clang by "-fsanitize-hwaddress-experimental-aliasing". |
| 262 | static cl::opt<bool> ClUsePageAliases("hwasan-experimental-use-page-aliases" , |
| 263 | cl::desc("Use page aliasing in HWASan" ), |
| 264 | cl::Hidden, cl::init(Val: false)); |
| 265 | |
| 266 | namespace { |
| 267 | |
| 268 | template <typename T> T optOr(cl::opt<T> &Opt, T Other) { |
| 269 | return Opt.getNumOccurrences() ? Opt : Other; |
| 270 | } |
| 271 | |
| 272 | bool shouldUsePageAliases(const Triple &TargetTriple) { |
| 273 | return ClUsePageAliases && TargetTriple.getArch() == Triple::x86_64; |
| 274 | } |
| 275 | |
| 276 | bool shouldInstrumentStack(const Triple &TargetTriple) { |
| 277 | return !shouldUsePageAliases(TargetTriple) && ClInstrumentStack; |
| 278 | } |
| 279 | |
| 280 | bool shouldInstrumentWithCalls(const Triple &TargetTriple) { |
| 281 | return optOr(Opt&: ClInstrumentWithCalls, Other: TargetTriple.getArch() == Triple::x86_64); |
| 282 | } |
| 283 | |
| 284 | bool mightUseStackSafetyAnalysis(bool DisableOptimization) { |
| 285 | return optOr(Opt&: ClUseStackSafety, Other: !DisableOptimization); |
| 286 | } |
| 287 | |
| 288 | bool shouldUseStackSafetyAnalysis(const Triple &TargetTriple, |
| 289 | bool DisableOptimization) { |
| 290 | return shouldInstrumentStack(TargetTriple) && |
| 291 | mightUseStackSafetyAnalysis(DisableOptimization); |
| 292 | } |
| 293 | |
| 294 | bool shouldDetectUseAfterScope(const Triple &TargetTriple) { |
| 295 | return ClUseAfterScope && shouldInstrumentStack(TargetTriple); |
| 296 | } |
| 297 | |
| 298 | /// An instrumentation pass implementing detection of addressability bugs |
| 299 | /// using tagged pointers. |
| 300 | class HWAddressSanitizer { |
| 301 | public: |
| 302 | HWAddressSanitizer(Module &M, bool CompileKernel, bool Recover, |
| 303 | const StackSafetyGlobalInfo *SSI) |
| 304 | : M(M), SSI(SSI) { |
| 305 | this->Recover = optOr(Opt&: ClRecover, Other: Recover); |
| 306 | this->CompileKernel = optOr(Opt&: ClEnableKhwasan, Other: CompileKernel); |
| 307 | this->Rng = ClRandomKeepRate.getNumOccurrences() ? M.createRNG(DEBUG_TYPE) |
| 308 | : nullptr; |
| 309 | |
| 310 | initializeModule(); |
| 311 | } |
| 312 | |
| 313 | void sanitizeFunction(Function &F, FunctionAnalysisManager &FAM); |
| 314 | |
| 315 | private: |
| 316 | struct ShadowTagCheckInfo { |
| 317 | Instruction *TagMismatchTerm = nullptr; |
| 318 | Value *PtrLong = nullptr; |
| 319 | Value *AddrLong = nullptr; |
| 320 | Value *PtrTag = nullptr; |
| 321 | Value *MemTag = nullptr; |
| 322 | }; |
| 323 | |
| 324 | bool selectiveInstrumentationShouldSkip(Function &F, |
| 325 | FunctionAnalysisManager &FAM) const; |
| 326 | void initializeModule(); |
| 327 | void createHwasanCtorComdat(); |
| 328 | |
| 329 | void initializeCallbacks(Module &M); |
| 330 | |
| 331 | Value *getOpaqueNoopCast(IRBuilder<> &IRB, Value *Val); |
| 332 | |
| 333 | Value *getDynamicShadowIfunc(IRBuilder<> &IRB); |
| 334 | Value *getShadowNonTls(IRBuilder<> &IRB); |
| 335 | |
| 336 | void untagPointerOperand(Instruction *I, Value *Addr); |
| 337 | Value *memToShadow(Value *Shadow, IRBuilder<> &IRB); |
| 338 | |
| 339 | int64_t getAccessInfo(bool IsWrite, unsigned AccessSizeIndex); |
| 340 | ShadowTagCheckInfo insertShadowTagCheck(Value *Ptr, Instruction *InsertBefore, |
| 341 | DomTreeUpdater &DTU, LoopInfo *LI); |
| 342 | void instrumentMemAccessOutline(Value *Ptr, bool IsWrite, |
| 343 | unsigned AccessSizeIndex, |
| 344 | Instruction *InsertBefore, |
| 345 | DomTreeUpdater &DTU, LoopInfo *LI); |
| 346 | void instrumentMemAccessInline(Value *Ptr, bool IsWrite, |
| 347 | unsigned AccessSizeIndex, |
| 348 | Instruction *InsertBefore, DomTreeUpdater &DTU, |
| 349 | LoopInfo *LI); |
| 350 | bool ignoreMemIntrinsic(OptimizationRemarkEmitter &ORE, MemIntrinsic *MI); |
| 351 | void instrumentMemIntrinsic(MemIntrinsic *MI); |
| 352 | bool instrumentMemAccess(InterestingMemoryOperand &O, DomTreeUpdater &DTU, |
| 353 | LoopInfo *LI, const DataLayout &DL); |
| 354 | bool ignoreAccessWithoutRemark(Instruction *Inst, Value *Ptr); |
| 355 | bool ignoreAccess(OptimizationRemarkEmitter &ORE, Instruction *Inst, |
| 356 | Value *Ptr); |
| 357 | |
| 358 | void getInterestingMemoryOperands( |
| 359 | OptimizationRemarkEmitter &ORE, Instruction *I, |
| 360 | const TargetLibraryInfo &TLI, |
| 361 | SmallVectorImpl<InterestingMemoryOperand> &Interesting); |
| 362 | |
| 363 | void tagAlloca(IRBuilder<> &IRB, AllocaInst *AI, Value *Tag, size_t Size); |
| 364 | Value *tagPointer(IRBuilder<> &IRB, Type *Ty, Value *PtrLong, Value *Tag); |
| 365 | Value *untagPointer(IRBuilder<> &IRB, Value *PtrLong); |
| 366 | bool instrumentStack(memtag::StackInfo &Info, Value *StackTag, Value *UARTag, |
| 367 | const DominatorTree &DT, const PostDominatorTree &PDT, |
| 368 | const LoopInfo &LI); |
| 369 | bool instrumentLandingPads(SmallVectorImpl<Instruction *> &RetVec); |
| 370 | Value *getNextTagWithCall(IRBuilder<> &IRB); |
| 371 | Value *getStackBaseTag(IRBuilder<> &IRB); |
| 372 | Value *getAllocaTag(IRBuilder<> &IRB, Value *StackTag, unsigned AllocaNo); |
| 373 | Value *getUARTag(IRBuilder<> &IRB); |
| 374 | |
| 375 | Value *getHwasanThreadSlotPtr(IRBuilder<> &IRB); |
| 376 | Value *applyTagMask(IRBuilder<> &IRB, Value *OldTag); |
| 377 | unsigned retagMask(unsigned AllocaNo); |
| 378 | |
| 379 | void emitPrologue(IRBuilder<> &IRB, bool WithFrameRecord); |
| 380 | |
| 381 | void instrumentGlobal(GlobalVariable *GV, uint8_t Tag); |
| 382 | void instrumentGlobals(); |
| 383 | |
| 384 | Value *getCachedFP(IRBuilder<> &IRB); |
| 385 | Value *getFrameRecordInfo(IRBuilder<> &IRB); |
| 386 | |
| 387 | void instrumentPersonalityFunctions(); |
| 388 | |
| 389 | LLVMContext *C; |
| 390 | Module &M; |
| 391 | const StackSafetyGlobalInfo *SSI; |
| 392 | Triple TargetTriple; |
| 393 | std::unique_ptr<RandomNumberGenerator> Rng; |
| 394 | |
| 395 | /// This struct defines the shadow mapping using the rule: |
| 396 | /// If `kFixed`, then |
| 397 | /// shadow = (mem >> Scale) + Offset. |
| 398 | /// If `kGlobal`, then |
| 399 | /// extern char* __hwasan_shadow_memory_dynamic_address; |
| 400 | /// shadow = (mem >> Scale) + __hwasan_shadow_memory_dynamic_address |
| 401 | /// If `kIfunc`, then |
| 402 | /// extern char __hwasan_shadow[]; |
| 403 | /// shadow = (mem >> Scale) + &__hwasan_shadow |
| 404 | /// If `kTls`, then |
| 405 | /// extern char *__hwasan_tls; |
| 406 | /// shadow = (mem>>Scale) + align_up(__hwasan_shadow, kShadowBaseAlignment) |
| 407 | /// |
| 408 | /// If WithFrameRecord is true, then __hwasan_tls will be used to access the |
| 409 | /// ring buffer for storing stack allocations on targets that support it. |
| 410 | class ShadowMapping { |
| 411 | OffsetKind Kind; |
| 412 | uint64_t Offset; |
| 413 | uint8_t Scale; |
| 414 | bool WithFrameRecord; |
| 415 | |
| 416 | void SetFixed(uint64_t O) { |
| 417 | Kind = OffsetKind::kFixed; |
| 418 | Offset = O; |
| 419 | } |
| 420 | |
| 421 | public: |
| 422 | void init(Triple &TargetTriple, bool InstrumentWithCalls, |
| 423 | bool CompileKernel); |
| 424 | Align getObjectAlignment() const { return Align(1ULL << Scale); } |
| 425 | bool isInGlobal() const { return Kind == OffsetKind::kGlobal; } |
| 426 | bool isInIfunc() const { return Kind == OffsetKind::kIfunc; } |
| 427 | bool isInTls() const { return Kind == OffsetKind::kTls; } |
| 428 | bool isFixed() const { return Kind == OffsetKind::kFixed; } |
| 429 | uint8_t scale() const { return Scale; }; |
| 430 | uint64_t offset() const { |
| 431 | assert(isFixed()); |
| 432 | return Offset; |
| 433 | }; |
| 434 | bool withFrameRecord() const { return WithFrameRecord; }; |
| 435 | }; |
| 436 | |
| 437 | ShadowMapping Mapping; |
| 438 | |
| 439 | Type *VoidTy = Type::getVoidTy(C&: M.getContext()); |
| 440 | Type *IntptrTy = M.getDataLayout().getIntPtrType(C&: M.getContext()); |
| 441 | PointerType *PtrTy = PointerType::getUnqual(C&: M.getContext()); |
| 442 | Type *Int8Ty = Type::getInt8Ty(C&: M.getContext()); |
| 443 | Type *Int32Ty = Type::getInt32Ty(C&: M.getContext()); |
| 444 | Type *Int64Ty = Type::getInt64Ty(C&: M.getContext()); |
| 445 | |
| 446 | bool CompileKernel; |
| 447 | bool Recover; |
| 448 | bool OutlinedChecks; |
| 449 | bool InlineFastPath; |
| 450 | bool UseShortGranules; |
| 451 | bool InstrumentLandingPads; |
| 452 | bool InstrumentWithCalls; |
| 453 | bool InstrumentStack; |
| 454 | bool InstrumentGlobals; |
| 455 | bool DetectUseAfterScope; |
| 456 | bool UsePageAliases; |
| 457 | bool UseMatchAllCallback; |
| 458 | |
| 459 | std::optional<uint8_t> MatchAllTag; |
| 460 | |
| 461 | unsigned PointerTagShift; |
| 462 | uint64_t TagMaskByte; |
| 463 | |
| 464 | Function *HwasanCtorFunction; |
| 465 | |
| 466 | FunctionCallee HwasanMemoryAccessCallback[2][kNumberOfAccessSizes]; |
| 467 | FunctionCallee HwasanMemoryAccessCallbackSized[2]; |
| 468 | |
| 469 | FunctionCallee HwasanMemmove, HwasanMemcpy, HwasanMemset; |
| 470 | FunctionCallee HwasanHandleVfork; |
| 471 | |
| 472 | FunctionCallee HwasanTagMemoryFunc; |
| 473 | FunctionCallee HwasanGenerateTagFunc; |
| 474 | FunctionCallee HwasanRecordFrameRecordFunc; |
| 475 | |
| 476 | Constant *ShadowGlobal; |
| 477 | |
| 478 | Value *ShadowBase = nullptr; |
| 479 | Value *StackBaseTag = nullptr; |
| 480 | Value *CachedFP = nullptr; |
| 481 | GlobalValue *ThreadPtrGlobal = nullptr; |
| 482 | }; |
| 483 | |
| 484 | } // end anonymous namespace |
| 485 | |
| 486 | PreservedAnalyses HWAddressSanitizerPass::run(Module &M, |
| 487 | ModuleAnalysisManager &MAM) { |
| 488 | // Return early if nosanitize_hwaddress module flag is present for the module. |
| 489 | if (checkIfAlreadyInstrumented(M, Flag: "nosanitize_hwaddress" )) |
| 490 | return PreservedAnalyses::all(); |
| 491 | const StackSafetyGlobalInfo *SSI = nullptr; |
| 492 | const Triple &TargetTriple = M.getTargetTriple(); |
| 493 | if (shouldUseStackSafetyAnalysis(TargetTriple, DisableOptimization: Options.DisableOptimization)) |
| 494 | SSI = &MAM.getResult<StackSafetyGlobalAnalysis>(IR&: M); |
| 495 | |
| 496 | HWAddressSanitizer HWASan(M, Options.CompileKernel, Options.Recover, SSI); |
| 497 | auto &FAM = MAM.getResult<FunctionAnalysisManagerModuleProxy>(IR&: M).getManager(); |
| 498 | for (Function &F : M) |
| 499 | HWASan.sanitizeFunction(F, FAM); |
| 500 | |
| 501 | PreservedAnalyses PA = PreservedAnalyses::none(); |
| 502 | // DominatorTreeAnalysis, PostDominatorTreeAnalysis, and LoopAnalysis |
| 503 | // are incrementally updated throughout this pass whenever |
| 504 | // SplitBlockAndInsertIfThen is called. |
| 505 | PA.preserve<DominatorTreeAnalysis>(); |
| 506 | PA.preserve<PostDominatorTreeAnalysis>(); |
| 507 | PA.preserve<LoopAnalysis>(); |
| 508 | // GlobalsAA is considered stateless and does not get invalidated unless |
| 509 | // explicitly invalidated; PreservedAnalyses::none() is not enough. Sanitizers |
| 510 | // make changes that require GlobalsAA to be invalidated. |
| 511 | PA.abandon<GlobalsAA>(); |
| 512 | return PA; |
| 513 | } |
| 514 | void HWAddressSanitizerPass::printPipeline( |
| 515 | raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { |
| 516 | static_cast<PassInfoMixin<HWAddressSanitizerPass> *>(this)->printPipeline( |
| 517 | OS, MapClassName2PassName); |
| 518 | OS << '<'; |
| 519 | if (Options.CompileKernel) |
| 520 | OS << "kernel;" ; |
| 521 | if (Options.Recover) |
| 522 | OS << "recover" ; |
| 523 | OS << '>'; |
| 524 | } |
| 525 | |
| 526 | void HWAddressSanitizer::createHwasanCtorComdat() { |
| 527 | std::tie(args&: HwasanCtorFunction, args: std::ignore) = |
| 528 | getOrCreateSanitizerCtorAndInitFunctions( |
| 529 | M, CtorName: kHwasanModuleCtorName, InitName: kHwasanInitName, |
| 530 | /*InitArgTypes=*/{}, |
| 531 | /*InitArgs=*/{}, |
| 532 | // This callback is invoked when the functions are created the first |
| 533 | // time. Hook them into the global ctors list in that case: |
| 534 | FunctionsCreatedCallback: [&](Function *Ctor, FunctionCallee) { |
| 535 | Comdat *CtorComdat = M.getOrInsertComdat(Name: kHwasanModuleCtorName); |
| 536 | Ctor->setComdat(CtorComdat); |
| 537 | appendToGlobalCtors(M, F: Ctor, Priority: 0, Data: Ctor); |
| 538 | }); |
| 539 | |
| 540 | // Create a note that contains pointers to the list of global |
| 541 | // descriptors. Adding a note to the output file will cause the linker to |
| 542 | // create a PT_NOTE program header pointing to the note that we can use to |
| 543 | // find the descriptor list starting from the program headers. A function |
| 544 | // provided by the runtime initializes the shadow memory for the globals by |
| 545 | // accessing the descriptor list via the note. The dynamic loader needs to |
| 546 | // call this function whenever a library is loaded. |
| 547 | // |
| 548 | // The reason why we use a note for this instead of a more conventional |
| 549 | // approach of having a global constructor pass a descriptor list pointer to |
| 550 | // the runtime is because of an order of initialization problem. With |
| 551 | // constructors we can encounter the following problematic scenario: |
| 552 | // |
| 553 | // 1) library A depends on library B and also interposes one of B's symbols |
| 554 | // 2) B's constructors are called before A's (as required for correctness) |
| 555 | // 3) during construction, B accesses one of its "own" globals (actually |
| 556 | // interposed by A) and triggers a HWASAN failure due to the initialization |
| 557 | // for A not having happened yet |
| 558 | // |
| 559 | // Even without interposition it is possible to run into similar situations in |
| 560 | // cases where two libraries mutually depend on each other. |
| 561 | // |
| 562 | // We only need one note per binary, so put everything for the note in a |
| 563 | // comdat. This needs to be a comdat with an .init_array section to prevent |
| 564 | // newer versions of lld from discarding the note. |
| 565 | // |
| 566 | // Create the note even if we aren't instrumenting globals. This ensures that |
| 567 | // binaries linked from object files with both instrumented and |
| 568 | // non-instrumented globals will end up with a note, even if a comdat from an |
| 569 | // object file with non-instrumented globals is selected. The note is harmless |
| 570 | // if the runtime doesn't support it, since it will just be ignored. |
| 571 | Comdat *NoteComdat = M.getOrInsertComdat(Name: kHwasanModuleCtorName); |
| 572 | |
| 573 | Type *Int8Arr0Ty = ArrayType::get(ElementType: Int8Ty, NumElements: 0); |
| 574 | auto *Start = |
| 575 | new GlobalVariable(M, Int8Arr0Ty, true, GlobalVariable::ExternalLinkage, |
| 576 | nullptr, "__start_hwasan_globals" ); |
| 577 | Start->setVisibility(GlobalValue::HiddenVisibility); |
| 578 | auto *Stop = |
| 579 | new GlobalVariable(M, Int8Arr0Ty, true, GlobalVariable::ExternalLinkage, |
| 580 | nullptr, "__stop_hwasan_globals" ); |
| 581 | Stop->setVisibility(GlobalValue::HiddenVisibility); |
| 582 | |
| 583 | // Null-terminated so actually 8 bytes, which are required in order to align |
| 584 | // the note properly. |
| 585 | auto *Name = ConstantDataArray::get(Context&: *C, Elts: "LLVM\0\0\0" ); |
| 586 | |
| 587 | auto *NoteTy = StructType::get(elt1: Int32Ty, elts: Int32Ty, elts: Int32Ty, elts: Name->getType(), |
| 588 | elts: Int32Ty, elts: Int32Ty); |
| 589 | auto *Note = |
| 590 | new GlobalVariable(M, NoteTy, /*isConstant=*/true, |
| 591 | GlobalValue::PrivateLinkage, nullptr, kHwasanNoteName); |
| 592 | Note->setSection(".note.hwasan.globals" ); |
| 593 | Note->setComdat(NoteComdat); |
| 594 | Note->setAlignment(Align(4)); |
| 595 | |
| 596 | // The pointers in the note need to be relative so that the note ends up being |
| 597 | // placed in rodata, which is the standard location for notes. |
| 598 | auto CreateRelPtr = [&](Constant *Ptr) { |
| 599 | return ConstantExpr::getTrunc( |
| 600 | C: ConstantExpr::getSub(C1: ConstantExpr::getPtrToInt(C: Ptr, Ty: Int64Ty), |
| 601 | C2: ConstantExpr::getPtrToInt(C: Note, Ty: Int64Ty)), |
| 602 | Ty: Int32Ty); |
| 603 | }; |
| 604 | Note->setInitializer(ConstantStruct::getAnon( |
| 605 | V: {ConstantInt::get(Ty: Int32Ty, V: 8), // n_namesz |
| 606 | ConstantInt::get(Ty: Int32Ty, V: 8), // n_descsz |
| 607 | ConstantInt::get(Ty: Int32Ty, V: ELF::NT_LLVM_HWASAN_GLOBALS), // n_type |
| 608 | Name, CreateRelPtr(Start), CreateRelPtr(Stop)})); |
| 609 | appendToCompilerUsed(M, Values: Note); |
| 610 | |
| 611 | // Create a zero-length global in hwasan_globals so that the linker will |
| 612 | // always create start and stop symbols. |
| 613 | auto *Dummy = new GlobalVariable( |
| 614 | M, Int8Arr0Ty, /*isConstantGlobal*/ true, GlobalVariable::PrivateLinkage, |
| 615 | Constant::getNullValue(Ty: Int8Arr0Ty), "hwasan.dummy.global" ); |
| 616 | Dummy->setSection("hwasan_globals" ); |
| 617 | Dummy->setComdat(NoteComdat); |
| 618 | Dummy->setMetadata(KindID: LLVMContext::MD_associated, |
| 619 | Node: MDNode::get(Context&: *C, MDs: ValueAsMetadata::get(V: Note))); |
| 620 | appendToCompilerUsed(M, Values: Dummy); |
| 621 | } |
| 622 | |
| 623 | /// Module-level initialization. |
| 624 | /// |
| 625 | /// inserts a call to __hwasan_init to the module's constructor list. |
| 626 | void HWAddressSanitizer::initializeModule() { |
| 627 | LLVM_DEBUG(dbgs() << "Init " << M.getName() << "\n" ); |
| 628 | TargetTriple = M.getTargetTriple(); |
| 629 | |
| 630 | // HWASan may do short granule checks on function arguments read from the |
| 631 | // argument memory (last byte of the granule), which invalidates writeonly. |
| 632 | for (Function &F : M.functions()) |
| 633 | removeASanIncompatibleFnAttributes(F, /*ReadsArgMem=*/true); |
| 634 | |
| 635 | // x86_64 currently has two modes: |
| 636 | // - Intel LAM (default) |
| 637 | // - pointer aliasing (heap only) |
| 638 | bool IsX86_64 = TargetTriple.getArch() == Triple::x86_64; |
| 639 | UsePageAliases = shouldUsePageAliases(TargetTriple); |
| 640 | InstrumentWithCalls = shouldInstrumentWithCalls(TargetTriple); |
| 641 | InstrumentStack = shouldInstrumentStack(TargetTriple); |
| 642 | DetectUseAfterScope = shouldDetectUseAfterScope(TargetTriple); |
| 643 | PointerTagShift = IsX86_64 ? 57 : 56; |
| 644 | TagMaskByte = IsX86_64 ? 0x3F : 0xFF; |
| 645 | |
| 646 | Mapping.init(TargetTriple, InstrumentWithCalls, CompileKernel); |
| 647 | |
| 648 | C = &(M.getContext()); |
| 649 | IRBuilder<> IRB(*C); |
| 650 | |
| 651 | HwasanCtorFunction = nullptr; |
| 652 | |
| 653 | // Older versions of Android do not have the required runtime support for |
| 654 | // short granules, global or personality function instrumentation. On other |
| 655 | // platforms we currently require using the latest version of the runtime. |
| 656 | bool NewRuntime = |
| 657 | !TargetTriple.isAndroid() || !TargetTriple.isAndroidVersionLT(Major: 30); |
| 658 | |
| 659 | UseShortGranules = optOr(Opt&: ClUseShortGranules, Other: NewRuntime); |
| 660 | OutlinedChecks = (TargetTriple.isAArch64() || TargetTriple.isRISCV64()) && |
| 661 | TargetTriple.isOSBinFormatELF() && |
| 662 | !optOr(Opt&: ClInlineAllChecks, Other: Recover); |
| 663 | |
| 664 | // These platforms may prefer less inlining to reduce binary size. |
| 665 | InlineFastPath = optOr(Opt&: ClInlineFastPathChecks, Other: !(TargetTriple.isAndroid() || |
| 666 | TargetTriple.isOSFuchsia())); |
| 667 | |
| 668 | if (ClMatchAllTag.getNumOccurrences()) { |
| 669 | if (ClMatchAllTag != -1) { |
| 670 | MatchAllTag = ClMatchAllTag & 0xFF; |
| 671 | } |
| 672 | } else if (CompileKernel) { |
| 673 | MatchAllTag = 0xFF; |
| 674 | } |
| 675 | UseMatchAllCallback = !CompileKernel && MatchAllTag.has_value(); |
| 676 | |
| 677 | // If we don't have personality function support, fall back to landing pads. |
| 678 | InstrumentLandingPads = optOr(Opt&: ClInstrumentLandingPads, Other: !NewRuntime); |
| 679 | |
| 680 | InstrumentGlobals = |
| 681 | !CompileKernel && !UsePageAliases && optOr(Opt&: ClGlobals, Other: NewRuntime); |
| 682 | |
| 683 | if (!CompileKernel) { |
| 684 | createHwasanCtorComdat(); |
| 685 | |
| 686 | if (InstrumentGlobals) |
| 687 | instrumentGlobals(); |
| 688 | |
| 689 | bool InstrumentPersonalityFunctions = |
| 690 | optOr(Opt&: ClInstrumentPersonalityFunctions, Other: NewRuntime); |
| 691 | if (InstrumentPersonalityFunctions) |
| 692 | instrumentPersonalityFunctions(); |
| 693 | } |
| 694 | |
| 695 | if (!TargetTriple.isAndroid()) { |
| 696 | ThreadPtrGlobal = M.getOrInsertGlobal(Name: "__hwasan_tls" , Ty: IntptrTy, CreateGlobalCallback: [&] { |
| 697 | auto *GV = new GlobalVariable(M, IntptrTy, /*isConstant=*/false, |
| 698 | GlobalValue::ExternalLinkage, nullptr, |
| 699 | "__hwasan_tls" , nullptr, |
| 700 | GlobalVariable::InitialExecTLSModel); |
| 701 | appendToCompilerUsed(M, Values: GV); |
| 702 | return GV; |
| 703 | }); |
| 704 | } |
| 705 | } |
| 706 | |
| 707 | void HWAddressSanitizer::initializeCallbacks(Module &M) { |
| 708 | IRBuilder<> IRB(*C); |
| 709 | const std::string MatchAllStr = UseMatchAllCallback ? "_match_all" : "" ; |
| 710 | FunctionType *HwasanMemoryAccessCallbackSizedFnTy, |
| 711 | *HwasanMemoryAccessCallbackFnTy, *HwasanMemTransferFnTy, |
| 712 | *HwasanMemsetFnTy; |
| 713 | if (UseMatchAllCallback) { |
| 714 | HwasanMemoryAccessCallbackSizedFnTy = |
| 715 | FunctionType::get(Result: VoidTy, Params: {IntptrTy, IntptrTy, Int8Ty}, isVarArg: false); |
| 716 | HwasanMemoryAccessCallbackFnTy = |
| 717 | FunctionType::get(Result: VoidTy, Params: {IntptrTy, Int8Ty}, isVarArg: false); |
| 718 | HwasanMemTransferFnTy = |
| 719 | FunctionType::get(Result: PtrTy, Params: {PtrTy, PtrTy, IntptrTy, Int8Ty}, isVarArg: false); |
| 720 | HwasanMemsetFnTy = |
| 721 | FunctionType::get(Result: PtrTy, Params: {PtrTy, Int32Ty, IntptrTy, Int8Ty}, isVarArg: false); |
| 722 | } else { |
| 723 | HwasanMemoryAccessCallbackSizedFnTy = |
| 724 | FunctionType::get(Result: VoidTy, Params: {IntptrTy, IntptrTy}, isVarArg: false); |
| 725 | HwasanMemoryAccessCallbackFnTy = |
| 726 | FunctionType::get(Result: VoidTy, Params: {IntptrTy}, isVarArg: false); |
| 727 | HwasanMemTransferFnTy = |
| 728 | FunctionType::get(Result: PtrTy, Params: {PtrTy, PtrTy, IntptrTy}, isVarArg: false); |
| 729 | HwasanMemsetFnTy = |
| 730 | FunctionType::get(Result: PtrTy, Params: {PtrTy, Int32Ty, IntptrTy}, isVarArg: false); |
| 731 | } |
| 732 | |
| 733 | for (size_t AccessIsWrite = 0; AccessIsWrite <= 1; AccessIsWrite++) { |
| 734 | const std::string TypeStr = AccessIsWrite ? "store" : "load" ; |
| 735 | const std::string EndingStr = Recover ? "_noabort" : "" ; |
| 736 | |
| 737 | HwasanMemoryAccessCallbackSized[AccessIsWrite] = M.getOrInsertFunction( |
| 738 | Name: ClMemoryAccessCallbackPrefix + TypeStr + "N" + MatchAllStr + EndingStr, |
| 739 | T: HwasanMemoryAccessCallbackSizedFnTy); |
| 740 | |
| 741 | for (size_t AccessSizeIndex = 0; AccessSizeIndex < kNumberOfAccessSizes; |
| 742 | AccessSizeIndex++) { |
| 743 | HwasanMemoryAccessCallback[AccessIsWrite][AccessSizeIndex] = |
| 744 | M.getOrInsertFunction(Name: ClMemoryAccessCallbackPrefix + TypeStr + |
| 745 | itostr(X: 1ULL << AccessSizeIndex) + |
| 746 | MatchAllStr + EndingStr, |
| 747 | T: HwasanMemoryAccessCallbackFnTy); |
| 748 | } |
| 749 | } |
| 750 | |
| 751 | const std::string MemIntrinCallbackPrefix = |
| 752 | (CompileKernel && !ClKasanMemIntrinCallbackPrefix) |
| 753 | ? std::string("" ) |
| 754 | : ClMemoryAccessCallbackPrefix; |
| 755 | |
| 756 | HwasanMemmove = M.getOrInsertFunction( |
| 757 | Name: MemIntrinCallbackPrefix + "memmove" + MatchAllStr, T: HwasanMemTransferFnTy); |
| 758 | HwasanMemcpy = M.getOrInsertFunction( |
| 759 | Name: MemIntrinCallbackPrefix + "memcpy" + MatchAllStr, T: HwasanMemTransferFnTy); |
| 760 | HwasanMemset = M.getOrInsertFunction( |
| 761 | Name: MemIntrinCallbackPrefix + "memset" + MatchAllStr, T: HwasanMemsetFnTy); |
| 762 | |
| 763 | HwasanTagMemoryFunc = M.getOrInsertFunction(Name: "__hwasan_tag_memory" , RetTy: VoidTy, |
| 764 | Args: PtrTy, Args: Int8Ty, Args: IntptrTy); |
| 765 | HwasanGenerateTagFunc = |
| 766 | M.getOrInsertFunction(Name: "__hwasan_generate_tag" , RetTy: Int8Ty); |
| 767 | |
| 768 | HwasanRecordFrameRecordFunc = |
| 769 | M.getOrInsertFunction(Name: "__hwasan_add_frame_record" , RetTy: VoidTy, Args: Int64Ty); |
| 770 | |
| 771 | ShadowGlobal = |
| 772 | M.getOrInsertGlobal(Name: "__hwasan_shadow" , Ty: ArrayType::get(ElementType: Int8Ty, NumElements: 0)); |
| 773 | |
| 774 | HwasanHandleVfork = |
| 775 | M.getOrInsertFunction(Name: "__hwasan_handle_vfork" , RetTy: VoidTy, Args: IntptrTy); |
| 776 | } |
| 777 | |
| 778 | Value *HWAddressSanitizer::getOpaqueNoopCast(IRBuilder<> &IRB, Value *Val) { |
| 779 | // An empty inline asm with input reg == output reg. |
| 780 | // An opaque no-op cast, basically. |
| 781 | // This prevents code bloat as a result of rematerializing trivial definitions |
| 782 | // such as constants or global addresses at every load and store. |
| 783 | InlineAsm *Asm = |
| 784 | InlineAsm::get(Ty: FunctionType::get(Result: PtrTy, Params: {Val->getType()}, isVarArg: false), |
| 785 | AsmString: StringRef("" ), Constraints: StringRef("=r,0" ), |
| 786 | /*hasSideEffects=*/false); |
| 787 | return IRB.CreateCall(Callee: Asm, Args: {Val}, Name: ".hwasan.shadow" ); |
| 788 | } |
| 789 | |
| 790 | Value *HWAddressSanitizer::getDynamicShadowIfunc(IRBuilder<> &IRB) { |
| 791 | return getOpaqueNoopCast(IRB, Val: ShadowGlobal); |
| 792 | } |
| 793 | |
| 794 | Value *HWAddressSanitizer::getShadowNonTls(IRBuilder<> &IRB) { |
| 795 | if (Mapping.isFixed()) { |
| 796 | return getOpaqueNoopCast( |
| 797 | IRB, Val: ConstantExpr::getIntToPtr( |
| 798 | C: ConstantInt::get(Ty: IntptrTy, V: Mapping.offset()), Ty: PtrTy)); |
| 799 | } |
| 800 | |
| 801 | if (Mapping.isInIfunc()) |
| 802 | return getDynamicShadowIfunc(IRB); |
| 803 | |
| 804 | Value *GlobalDynamicAddress = |
| 805 | IRB.GetInsertBlock()->getParent()->getParent()->getOrInsertGlobal( |
| 806 | Name: kHwasanShadowMemoryDynamicAddress, Ty: PtrTy); |
| 807 | return IRB.CreateLoad(Ty: PtrTy, Ptr: GlobalDynamicAddress); |
| 808 | } |
| 809 | |
| 810 | bool HWAddressSanitizer::(Instruction *Inst, |
| 811 | Value *Ptr) { |
| 812 | // Do not instrument accesses from different address spaces; we cannot deal |
| 813 | // with them. |
| 814 | Type *PtrTy = cast<PointerType>(Val: Ptr->getType()->getScalarType()); |
| 815 | if (PtrTy->getPointerAddressSpace() != 0) |
| 816 | return true; |
| 817 | |
| 818 | // Ignore swifterror addresses. |
| 819 | // swifterror memory addresses are mem2reg promoted by instruction |
| 820 | // selection. As such they cannot have regular uses like an instrumentation |
| 821 | // function and it makes no sense to track them as memory. |
| 822 | if (Ptr->isSwiftError()) |
| 823 | return true; |
| 824 | |
| 825 | if (findAllocaForValue(V: Ptr)) { |
| 826 | if (!InstrumentStack) |
| 827 | return true; |
| 828 | if (SSI && SSI->stackAccessIsSafe(I: *Inst)) |
| 829 | return true; |
| 830 | } |
| 831 | |
| 832 | if (isa<GlobalVariable>(Val: getUnderlyingObject(V: Ptr))) { |
| 833 | if (!InstrumentGlobals) |
| 834 | return true; |
| 835 | // TODO: Optimize inbound global accesses, like Asan `instrumentMop`. |
| 836 | } |
| 837 | |
| 838 | return false; |
| 839 | } |
| 840 | |
| 841 | bool HWAddressSanitizer::(OptimizationRemarkEmitter &ORE, |
| 842 | Instruction *Inst, Value *Ptr) { |
| 843 | bool Ignored = ignoreAccessWithoutRemark(Inst, Ptr); |
| 844 | if (Ignored) { |
| 845 | ORE.emit( |
| 846 | RemarkBuilder: [&]() { return OptimizationRemark(DEBUG_TYPE, "ignoreAccess" , Inst); }); |
| 847 | } else { |
| 848 | ORE.emit(RemarkBuilder: [&]() { |
| 849 | return OptimizationRemarkMissed(DEBUG_TYPE, "ignoreAccess" , Inst); |
| 850 | }); |
| 851 | } |
| 852 | return Ignored; |
| 853 | } |
| 854 | |
| 855 | void HWAddressSanitizer::getInterestingMemoryOperands( |
| 856 | OptimizationRemarkEmitter &ORE, Instruction *I, |
| 857 | const TargetLibraryInfo &TLI, |
| 858 | SmallVectorImpl<InterestingMemoryOperand> &Interesting) { |
| 859 | // Skip memory accesses inserted by another instrumentation. |
| 860 | if (I->hasMetadata(KindID: LLVMContext::MD_nosanitize)) |
| 861 | return; |
| 862 | |
| 863 | // Do not instrument the load fetching the dynamic shadow address. |
| 864 | if (ShadowBase == I) |
| 865 | return; |
| 866 | |
| 867 | if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) { |
| 868 | if (!ClInstrumentReads || ignoreAccess(ORE, Inst: I, Ptr: LI->getPointerOperand())) |
| 869 | return; |
| 870 | Interesting.emplace_back(Args&: I, Args: LI->getPointerOperandIndex(), Args: false, |
| 871 | Args: LI->getType(), Args: LI->getAlign()); |
| 872 | } else if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) { |
| 873 | if (!ClInstrumentWrites || ignoreAccess(ORE, Inst: I, Ptr: SI->getPointerOperand())) |
| 874 | return; |
| 875 | Interesting.emplace_back(Args&: I, Args: SI->getPointerOperandIndex(), Args: true, |
| 876 | Args: SI->getValueOperand()->getType(), Args: SI->getAlign()); |
| 877 | } else if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: I)) { |
| 878 | if (!ClInstrumentAtomics || ignoreAccess(ORE, Inst: I, Ptr: RMW->getPointerOperand())) |
| 879 | return; |
| 880 | Interesting.emplace_back(Args&: I, Args: RMW->getPointerOperandIndex(), Args: true, |
| 881 | Args: RMW->getValOperand()->getType(), Args: std::nullopt); |
| 882 | } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(Val: I)) { |
| 883 | if (!ClInstrumentAtomics || ignoreAccess(ORE, Inst: I, Ptr: XCHG->getPointerOperand())) |
| 884 | return; |
| 885 | Interesting.emplace_back(Args&: I, Args: XCHG->getPointerOperandIndex(), Args: true, |
| 886 | Args: XCHG->getCompareOperand()->getType(), |
| 887 | Args: std::nullopt); |
| 888 | } else if (auto *CI = dyn_cast<CallInst>(Val: I)) { |
| 889 | for (unsigned ArgNo = 0; ArgNo < CI->arg_size(); ArgNo++) { |
| 890 | if (!ClInstrumentByval || !CI->isByValArgument(ArgNo) || |
| 891 | ignoreAccess(ORE, Inst: I, Ptr: CI->getArgOperand(i: ArgNo))) |
| 892 | continue; |
| 893 | Type *Ty = CI->getParamByValType(ArgNo); |
| 894 | Interesting.emplace_back(Args&: I, Args&: ArgNo, Args: false, Args&: Ty, Args: Align(1)); |
| 895 | } |
| 896 | maybeMarkSanitizerLibraryCallNoBuiltin(CI, TLI: &TLI); |
| 897 | } |
| 898 | } |
| 899 | |
| 900 | static unsigned getPointerOperandIndex(Instruction *I) { |
| 901 | if (LoadInst *LI = dyn_cast<LoadInst>(Val: I)) |
| 902 | return LI->getPointerOperandIndex(); |
| 903 | if (StoreInst *SI = dyn_cast<StoreInst>(Val: I)) |
| 904 | return SI->getPointerOperandIndex(); |
| 905 | if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(Val: I)) |
| 906 | return RMW->getPointerOperandIndex(); |
| 907 | if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(Val: I)) |
| 908 | return XCHG->getPointerOperandIndex(); |
| 909 | report_fatal_error(reason: "Unexpected instruction" ); |
| 910 | return -1; |
| 911 | } |
| 912 | |
| 913 | static size_t TypeSizeToSizeIndex(uint32_t TypeSize) { |
| 914 | size_t Res = llvm::countr_zero(Val: TypeSize / 8); |
| 915 | assert(Res < kNumberOfAccessSizes); |
| 916 | return Res; |
| 917 | } |
| 918 | |
| 919 | void HWAddressSanitizer::untagPointerOperand(Instruction *I, Value *Addr) { |
| 920 | if (TargetTriple.isAArch64() || TargetTriple.getArch() == Triple::x86_64 || |
| 921 | TargetTriple.isRISCV64()) |
| 922 | return; |
| 923 | |
| 924 | IRBuilder<> IRB(I); |
| 925 | Value *AddrLong = IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy); |
| 926 | Value *UntaggedPtr = |
| 927 | IRB.CreateIntToPtr(V: untagPointer(IRB, PtrLong: AddrLong), DestTy: Addr->getType()); |
| 928 | I->setOperand(i: getPointerOperandIndex(I), Val: UntaggedPtr); |
| 929 | } |
| 930 | |
| 931 | Value *HWAddressSanitizer::memToShadow(Value *Mem, IRBuilder<> &IRB) { |
| 932 | // Mem >> Scale |
| 933 | Value *Shadow = IRB.CreateLShr(LHS: Mem, RHS: Mapping.scale()); |
| 934 | if (Mapping.isFixed() && Mapping.offset() == 0) |
| 935 | return IRB.CreateIntToPtr(V: Shadow, DestTy: PtrTy); |
| 936 | // (Mem >> Scale) + Offset |
| 937 | return IRB.CreatePtrAdd(Ptr: ShadowBase, Offset: Shadow); |
| 938 | } |
| 939 | |
| 940 | int64_t HWAddressSanitizer::getAccessInfo(bool IsWrite, |
| 941 | unsigned AccessSizeIndex) { |
| 942 | return (CompileKernel << HWASanAccessInfo::CompileKernelShift) | |
| 943 | (MatchAllTag.has_value() << HWASanAccessInfo::HasMatchAllShift) | |
| 944 | (MatchAllTag.value_or(u: 0) << HWASanAccessInfo::MatchAllShift) | |
| 945 | (Recover << HWASanAccessInfo::RecoverShift) | |
| 946 | (IsWrite << HWASanAccessInfo::IsWriteShift) | |
| 947 | (AccessSizeIndex << HWASanAccessInfo::AccessSizeShift); |
| 948 | } |
| 949 | |
| 950 | HWAddressSanitizer::ShadowTagCheckInfo |
| 951 | HWAddressSanitizer::insertShadowTagCheck(Value *Ptr, Instruction *InsertBefore, |
| 952 | DomTreeUpdater &DTU, LoopInfo *LI) { |
| 953 | ShadowTagCheckInfo R; |
| 954 | |
| 955 | IRBuilder<> IRB(InsertBefore); |
| 956 | |
| 957 | R.PtrLong = IRB.CreatePointerCast(V: Ptr, DestTy: IntptrTy); |
| 958 | R.PtrTag = |
| 959 | IRB.CreateTrunc(V: IRB.CreateLShr(LHS: R.PtrLong, RHS: PointerTagShift), DestTy: Int8Ty); |
| 960 | R.AddrLong = untagPointer(IRB, PtrLong: R.PtrLong); |
| 961 | Value *Shadow = memToShadow(Mem: R.AddrLong, IRB); |
| 962 | R.MemTag = IRB.CreateLoad(Ty: Int8Ty, Ptr: Shadow); |
| 963 | Value *TagMismatch = IRB.CreateICmpNE(LHS: R.PtrTag, RHS: R.MemTag); |
| 964 | |
| 965 | if (MatchAllTag.has_value()) { |
| 966 | Value *TagNotIgnored = IRB.CreateICmpNE( |
| 967 | LHS: R.PtrTag, RHS: ConstantInt::get(Ty: R.PtrTag->getType(), V: *MatchAllTag)); |
| 968 | TagMismatch = IRB.CreateAnd(LHS: TagMismatch, RHS: TagNotIgnored); |
| 969 | } |
| 970 | |
| 971 | R.TagMismatchTerm = SplitBlockAndInsertIfThen( |
| 972 | Cond: TagMismatch, SplitBefore: InsertBefore, Unreachable: false, |
| 973 | BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights(), DTU: &DTU, LI); |
| 974 | |
| 975 | return R; |
| 976 | } |
| 977 | |
| 978 | void HWAddressSanitizer::instrumentMemAccessOutline(Value *Ptr, bool IsWrite, |
| 979 | unsigned AccessSizeIndex, |
| 980 | Instruction *InsertBefore, |
| 981 | DomTreeUpdater &DTU, |
| 982 | LoopInfo *LI) { |
| 983 | assert(!UsePageAliases); |
| 984 | const int64_t AccessInfo = getAccessInfo(IsWrite, AccessSizeIndex); |
| 985 | |
| 986 | if (InlineFastPath) |
| 987 | InsertBefore = |
| 988 | insertShadowTagCheck(Ptr, InsertBefore, DTU, LI).TagMismatchTerm; |
| 989 | |
| 990 | IRBuilder<> IRB(InsertBefore); |
| 991 | bool UseFixedShadowIntrinsic = false; |
| 992 | // The memaccess fixed shadow intrinsic is only supported on AArch64, |
| 993 | // which allows a 16-bit immediate to be left-shifted by 32. |
| 994 | // Since kShadowBaseAlignment == 32, and Linux by default will not |
| 995 | // mmap above 48-bits, practically any valid shadow offset is |
| 996 | // representable. |
| 997 | // In particular, an offset of 4TB (1024 << 32) is representable, and |
| 998 | // ought to be good enough for anybody. |
| 999 | if (TargetTriple.isAArch64() && Mapping.isFixed()) { |
| 1000 | uint16_t OffsetShifted = Mapping.offset() >> 32; |
| 1001 | UseFixedShadowIntrinsic = |
| 1002 | static_cast<uint64_t>(OffsetShifted) << 32 == Mapping.offset(); |
| 1003 | } |
| 1004 | |
| 1005 | if (UseFixedShadowIntrinsic) { |
| 1006 | IRB.CreateIntrinsic( |
| 1007 | ID: UseShortGranules |
| 1008 | ? Intrinsic::hwasan_check_memaccess_shortgranules_fixedshadow |
| 1009 | : Intrinsic::hwasan_check_memaccess_fixedshadow, |
| 1010 | Args: {Ptr, ConstantInt::get(Ty: Int32Ty, V: AccessInfo), |
| 1011 | ConstantInt::get(Ty: Int64Ty, V: Mapping.offset())}); |
| 1012 | } else { |
| 1013 | IRB.CreateIntrinsic( |
| 1014 | ID: UseShortGranules ? Intrinsic::hwasan_check_memaccess_shortgranules |
| 1015 | : Intrinsic::hwasan_check_memaccess, |
| 1016 | Args: {ShadowBase, Ptr, ConstantInt::get(Ty: Int32Ty, V: AccessInfo)}); |
| 1017 | } |
| 1018 | } |
| 1019 | |
| 1020 | void HWAddressSanitizer::instrumentMemAccessInline(Value *Ptr, bool IsWrite, |
| 1021 | unsigned AccessSizeIndex, |
| 1022 | Instruction *InsertBefore, |
| 1023 | DomTreeUpdater &DTU, |
| 1024 | LoopInfo *LI) { |
| 1025 | assert(!UsePageAliases); |
| 1026 | const int64_t AccessInfo = getAccessInfo(IsWrite, AccessSizeIndex); |
| 1027 | |
| 1028 | ShadowTagCheckInfo TCI = insertShadowTagCheck(Ptr, InsertBefore, DTU, LI); |
| 1029 | |
| 1030 | IRBuilder<> IRB(TCI.TagMismatchTerm); |
| 1031 | Value *OutOfShortGranuleTagRange = |
| 1032 | IRB.CreateICmpUGT(LHS: TCI.MemTag, RHS: ConstantInt::get(Ty: Int8Ty, V: 15)); |
| 1033 | Instruction *CheckFailTerm = SplitBlockAndInsertIfThen( |
| 1034 | Cond: OutOfShortGranuleTagRange, SplitBefore: TCI.TagMismatchTerm, Unreachable: !Recover, |
| 1035 | BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights(), DTU: &DTU, LI); |
| 1036 | |
| 1037 | IRB.SetInsertPoint(TCI.TagMismatchTerm); |
| 1038 | Value *PtrLowBits = IRB.CreateTrunc(V: IRB.CreateAnd(LHS: TCI.PtrLong, RHS: 15), DestTy: Int8Ty); |
| 1039 | PtrLowBits = IRB.CreateAdd( |
| 1040 | LHS: PtrLowBits, RHS: ConstantInt::get(Ty: Int8Ty, V: (1 << AccessSizeIndex) - 1)); |
| 1041 | Value *PtrLowBitsOOB = IRB.CreateICmpUGE(LHS: PtrLowBits, RHS: TCI.MemTag); |
| 1042 | SplitBlockAndInsertIfThen(Cond: PtrLowBitsOOB, SplitBefore: TCI.TagMismatchTerm, Unreachable: false, |
| 1043 | BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights(), DTU: &DTU, |
| 1044 | LI, ThenBlock: CheckFailTerm->getParent()); |
| 1045 | |
| 1046 | IRB.SetInsertPoint(TCI.TagMismatchTerm); |
| 1047 | Value *InlineTagAddr = IRB.CreateOr(LHS: TCI.AddrLong, RHS: 15); |
| 1048 | InlineTagAddr = IRB.CreateIntToPtr(V: InlineTagAddr, DestTy: PtrTy); |
| 1049 | Value *InlineTag = IRB.CreateLoad(Ty: Int8Ty, Ptr: InlineTagAddr); |
| 1050 | Value *InlineTagMismatch = IRB.CreateICmpNE(LHS: TCI.PtrTag, RHS: InlineTag); |
| 1051 | SplitBlockAndInsertIfThen(Cond: InlineTagMismatch, SplitBefore: TCI.TagMismatchTerm, Unreachable: false, |
| 1052 | BranchWeights: MDBuilder(*C).createUnlikelyBranchWeights(), DTU: &DTU, |
| 1053 | LI, ThenBlock: CheckFailTerm->getParent()); |
| 1054 | |
| 1055 | IRB.SetInsertPoint(CheckFailTerm); |
| 1056 | InlineAsm *Asm; |
| 1057 | switch (TargetTriple.getArch()) { |
| 1058 | case Triple::x86_64: |
| 1059 | // The signal handler will find the data address in rdi. |
| 1060 | Asm = InlineAsm::get( |
| 1061 | Ty: FunctionType::get(Result: VoidTy, Params: {TCI.PtrLong->getType()}, isVarArg: false), |
| 1062 | AsmString: "int3\nnopl " + |
| 1063 | itostr(X: 0x40 + (AccessInfo & HWASanAccessInfo::RuntimeMask)) + |
| 1064 | "(%rax)" , |
| 1065 | Constraints: "{rdi}" , |
| 1066 | /*hasSideEffects=*/true); |
| 1067 | break; |
| 1068 | case Triple::aarch64: |
| 1069 | case Triple::aarch64_be: |
| 1070 | // The signal handler will find the data address in x0. |
| 1071 | Asm = InlineAsm::get( |
| 1072 | Ty: FunctionType::get(Result: VoidTy, Params: {TCI.PtrLong->getType()}, isVarArg: false), |
| 1073 | AsmString: "brk #" + itostr(X: 0x900 + (AccessInfo & HWASanAccessInfo::RuntimeMask)), |
| 1074 | Constraints: "{x0}" , |
| 1075 | /*hasSideEffects=*/true); |
| 1076 | break; |
| 1077 | case Triple::riscv64: |
| 1078 | // The signal handler will find the data address in x10. |
| 1079 | Asm = InlineAsm::get( |
| 1080 | Ty: FunctionType::get(Result: VoidTy, Params: {TCI.PtrLong->getType()}, isVarArg: false), |
| 1081 | AsmString: "ebreak\naddiw x0, x11, " + |
| 1082 | itostr(X: 0x40 + (AccessInfo & HWASanAccessInfo::RuntimeMask)), |
| 1083 | Constraints: "{x10}" , |
| 1084 | /*hasSideEffects=*/true); |
| 1085 | break; |
| 1086 | default: |
| 1087 | report_fatal_error(reason: "unsupported architecture" ); |
| 1088 | } |
| 1089 | IRB.CreateCall(Callee: Asm, Args: TCI.PtrLong); |
| 1090 | if (Recover) |
| 1091 | cast<BranchInst>(Val: CheckFailTerm) |
| 1092 | ->setSuccessor(idx: 0, NewSucc: TCI.TagMismatchTerm->getParent()); |
| 1093 | } |
| 1094 | |
| 1095 | bool HWAddressSanitizer::(OptimizationRemarkEmitter &ORE, |
| 1096 | MemIntrinsic *MI) { |
| 1097 | if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(Val: MI)) { |
| 1098 | return (!ClInstrumentWrites || ignoreAccess(ORE, Inst: MTI, Ptr: MTI->getDest())) && |
| 1099 | (!ClInstrumentReads || ignoreAccess(ORE, Inst: MTI, Ptr: MTI->getSource())); |
| 1100 | } |
| 1101 | if (isa<MemSetInst>(Val: MI)) |
| 1102 | return !ClInstrumentWrites || ignoreAccess(ORE, Inst: MI, Ptr: MI->getDest()); |
| 1103 | return false; |
| 1104 | } |
| 1105 | |
| 1106 | void HWAddressSanitizer::instrumentMemIntrinsic(MemIntrinsic *MI) { |
| 1107 | IRBuilder<> IRB(MI); |
| 1108 | if (isa<MemTransferInst>(Val: MI)) { |
| 1109 | SmallVector<Value *, 4> Args{ |
| 1110 | MI->getOperand(i_nocapture: 0), MI->getOperand(i_nocapture: 1), |
| 1111 | IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 2), DestTy: IntptrTy, isSigned: false)}; |
| 1112 | |
| 1113 | if (UseMatchAllCallback) |
| 1114 | Args.emplace_back(Args: ConstantInt::get(Ty: Int8Ty, V: *MatchAllTag)); |
| 1115 | IRB.CreateCall(Callee: isa<MemMoveInst>(Val: MI) ? HwasanMemmove : HwasanMemcpy, Args); |
| 1116 | } else if (isa<MemSetInst>(Val: MI)) { |
| 1117 | SmallVector<Value *, 4> Args{ |
| 1118 | MI->getOperand(i_nocapture: 0), |
| 1119 | IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 1), DestTy: IRB.getInt32Ty(), isSigned: false), |
| 1120 | IRB.CreateIntCast(V: MI->getOperand(i_nocapture: 2), DestTy: IntptrTy, isSigned: false)}; |
| 1121 | if (UseMatchAllCallback) |
| 1122 | Args.emplace_back(Args: ConstantInt::get(Ty: Int8Ty, V: *MatchAllTag)); |
| 1123 | IRB.CreateCall(Callee: HwasanMemset, Args); |
| 1124 | } |
| 1125 | MI->eraseFromParent(); |
| 1126 | } |
| 1127 | |
| 1128 | bool HWAddressSanitizer::instrumentMemAccess(InterestingMemoryOperand &O, |
| 1129 | DomTreeUpdater &DTU, LoopInfo *LI, |
| 1130 | const DataLayout &DL) { |
| 1131 | Value *Addr = O.getPtr(); |
| 1132 | |
| 1133 | LLVM_DEBUG(dbgs() << "Instrumenting: " << O.getInsn() << "\n" ); |
| 1134 | |
| 1135 | // If the pointer is statically known to be zero, the tag check will pass |
| 1136 | // since: |
| 1137 | // 1) it has a zero tag |
| 1138 | // 2) the shadow memory corresponding to address 0 is initialized to zero and |
| 1139 | // never updated. |
| 1140 | // We can therefore elide the tag check. |
| 1141 | llvm::KnownBits Known(DL.getPointerTypeSizeInBits(Addr->getType())); |
| 1142 | llvm::computeKnownBits(V: Addr, Known, DL); |
| 1143 | if (Known.isZero()) |
| 1144 | return false; |
| 1145 | |
| 1146 | if (O.MaybeMask) |
| 1147 | return false; // FIXME |
| 1148 | |
| 1149 | IRBuilder<> IRB(O.getInsn()); |
| 1150 | if (!O.TypeStoreSize.isScalable() && isPowerOf2_64(Value: O.TypeStoreSize) && |
| 1151 | (O.TypeStoreSize / 8 <= (1ULL << (kNumberOfAccessSizes - 1))) && |
| 1152 | (!O.Alignment || *O.Alignment >= Mapping.getObjectAlignment() || |
| 1153 | *O.Alignment >= O.TypeStoreSize / 8)) { |
| 1154 | size_t AccessSizeIndex = TypeSizeToSizeIndex(TypeSize: O.TypeStoreSize); |
| 1155 | if (InstrumentWithCalls) { |
| 1156 | SmallVector<Value *, 2> Args{IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy)}; |
| 1157 | if (UseMatchAllCallback) |
| 1158 | Args.emplace_back(Args: ConstantInt::get(Ty: Int8Ty, V: *MatchAllTag)); |
| 1159 | IRB.CreateCall(Callee: HwasanMemoryAccessCallback[O.IsWrite][AccessSizeIndex], |
| 1160 | Args); |
| 1161 | } else if (OutlinedChecks) { |
| 1162 | instrumentMemAccessOutline(Ptr: Addr, IsWrite: O.IsWrite, AccessSizeIndex, InsertBefore: O.getInsn(), |
| 1163 | DTU, LI); |
| 1164 | } else { |
| 1165 | instrumentMemAccessInline(Ptr: Addr, IsWrite: O.IsWrite, AccessSizeIndex, InsertBefore: O.getInsn(), |
| 1166 | DTU, LI); |
| 1167 | } |
| 1168 | } else { |
| 1169 | SmallVector<Value *, 3> Args{ |
| 1170 | IRB.CreatePointerCast(V: Addr, DestTy: IntptrTy), |
| 1171 | IRB.CreateUDiv(LHS: IRB.CreateTypeSize(Ty: IntptrTy, Size: O.TypeStoreSize), |
| 1172 | RHS: ConstantInt::get(Ty: IntptrTy, V: 8))}; |
| 1173 | if (UseMatchAllCallback) |
| 1174 | Args.emplace_back(Args: ConstantInt::get(Ty: Int8Ty, V: *MatchAllTag)); |
| 1175 | IRB.CreateCall(Callee: HwasanMemoryAccessCallbackSized[O.IsWrite], Args); |
| 1176 | } |
| 1177 | untagPointerOperand(I: O.getInsn(), Addr); |
| 1178 | |
| 1179 | return true; |
| 1180 | } |
| 1181 | |
| 1182 | void HWAddressSanitizer::tagAlloca(IRBuilder<> &IRB, AllocaInst *AI, Value *Tag, |
| 1183 | size_t Size) { |
| 1184 | size_t AlignedSize = alignTo(Size, A: Mapping.getObjectAlignment()); |
| 1185 | if (!UseShortGranules) |
| 1186 | Size = AlignedSize; |
| 1187 | |
| 1188 | Tag = IRB.CreateTrunc(V: Tag, DestTy: Int8Ty); |
| 1189 | if (InstrumentWithCalls) { |
| 1190 | IRB.CreateCall(Callee: HwasanTagMemoryFunc, |
| 1191 | Args: {IRB.CreatePointerCast(V: AI, DestTy: PtrTy), Tag, |
| 1192 | ConstantInt::get(Ty: IntptrTy, V: AlignedSize)}); |
| 1193 | } else { |
| 1194 | size_t ShadowSize = Size >> Mapping.scale(); |
| 1195 | Value *AddrLong = untagPointer(IRB, PtrLong: IRB.CreatePointerCast(V: AI, DestTy: IntptrTy)); |
| 1196 | Value *ShadowPtr = memToShadow(Mem: AddrLong, IRB); |
| 1197 | // If this memset is not inlined, it will be intercepted in the hwasan |
| 1198 | // runtime library. That's OK, because the interceptor skips the checks if |
| 1199 | // the address is in the shadow region. |
| 1200 | // FIXME: the interceptor is not as fast as real memset. Consider lowering |
| 1201 | // llvm.memset right here into either a sequence of stores, or a call to |
| 1202 | // hwasan_tag_memory. |
| 1203 | if (ShadowSize) |
| 1204 | IRB.CreateMemSet(Ptr: ShadowPtr, Val: Tag, Size: ShadowSize, Align: Align(1)); |
| 1205 | if (Size != AlignedSize) { |
| 1206 | const uint8_t SizeRemainder = Size % Mapping.getObjectAlignment().value(); |
| 1207 | IRB.CreateStore(Val: ConstantInt::get(Ty: Int8Ty, V: SizeRemainder), |
| 1208 | Ptr: IRB.CreateConstGEP1_32(Ty: Int8Ty, Ptr: ShadowPtr, Idx0: ShadowSize)); |
| 1209 | IRB.CreateStore( |
| 1210 | Val: Tag, Ptr: IRB.CreateConstGEP1_32(Ty: Int8Ty, Ptr: IRB.CreatePointerCast(V: AI, DestTy: PtrTy), |
| 1211 | Idx0: AlignedSize - 1)); |
| 1212 | } |
| 1213 | } |
| 1214 | } |
| 1215 | |
| 1216 | unsigned HWAddressSanitizer::retagMask(unsigned AllocaNo) { |
| 1217 | if (TargetTriple.getArch() == Triple::x86_64) |
| 1218 | return AllocaNo & TagMaskByte; |
| 1219 | |
| 1220 | // A list of 8-bit numbers that have at most one run of non-zero bits. |
| 1221 | // x = x ^ (mask << 56) can be encoded as a single armv8 instruction for these |
| 1222 | // masks. |
| 1223 | // The list does not include the value 255, which is used for UAR. |
| 1224 | // |
| 1225 | // Because we are more likely to use earlier elements of this list than later |
| 1226 | // ones, it is sorted in increasing order of probability of collision with a |
| 1227 | // mask allocated (temporally) nearby. The program that generated this list |
| 1228 | // can be found at: |
| 1229 | // https://github.com/google/sanitizers/blob/master/hwaddress-sanitizer/sort_masks.py |
| 1230 | static const unsigned FastMasks[] = { |
| 1231 | 0, 128, 64, 192, 32, 96, 224, 112, 240, 48, 16, 120, |
| 1232 | 248, 56, 24, 8, 124, 252, 60, 28, 12, 4, 126, 254, |
| 1233 | 62, 30, 14, 6, 2, 127, 63, 31, 15, 7, 3, 1}; |
| 1234 | return FastMasks[AllocaNo % std::size(FastMasks)]; |
| 1235 | } |
| 1236 | |
| 1237 | Value *HWAddressSanitizer::applyTagMask(IRBuilder<> &IRB, Value *OldTag) { |
| 1238 | if (TagMaskByte == 0xFF) |
| 1239 | return OldTag; // No need to clear the tag byte. |
| 1240 | return IRB.CreateAnd(LHS: OldTag, |
| 1241 | RHS: ConstantInt::get(Ty: OldTag->getType(), V: TagMaskByte)); |
| 1242 | } |
| 1243 | |
| 1244 | Value *HWAddressSanitizer::getNextTagWithCall(IRBuilder<> &IRB) { |
| 1245 | return IRB.CreateZExt(V: IRB.CreateCall(Callee: HwasanGenerateTagFunc), DestTy: IntptrTy); |
| 1246 | } |
| 1247 | |
| 1248 | Value *HWAddressSanitizer::getStackBaseTag(IRBuilder<> &IRB) { |
| 1249 | if (ClGenerateTagsWithCalls) |
| 1250 | return nullptr; |
| 1251 | if (StackBaseTag) |
| 1252 | return StackBaseTag; |
| 1253 | // Extract some entropy from the stack pointer for the tags. |
| 1254 | // Take bits 20..28 (ASLR entropy) and xor with bits 0..8 (these differ |
| 1255 | // between functions). |
| 1256 | Value *FramePointerLong = getCachedFP(IRB); |
| 1257 | Value *StackTag = |
| 1258 | applyTagMask(IRB, OldTag: IRB.CreateXor(LHS: FramePointerLong, |
| 1259 | RHS: IRB.CreateLShr(LHS: FramePointerLong, RHS: 20))); |
| 1260 | StackTag->setName("hwasan.stack.base.tag" ); |
| 1261 | return StackTag; |
| 1262 | } |
| 1263 | |
| 1264 | Value *HWAddressSanitizer::getAllocaTag(IRBuilder<> &IRB, Value *StackTag, |
| 1265 | unsigned AllocaNo) { |
| 1266 | if (ClGenerateTagsWithCalls) |
| 1267 | return getNextTagWithCall(IRB); |
| 1268 | return IRB.CreateXor( |
| 1269 | LHS: StackTag, RHS: ConstantInt::get(Ty: StackTag->getType(), V: retagMask(AllocaNo))); |
| 1270 | } |
| 1271 | |
| 1272 | Value *HWAddressSanitizer::getUARTag(IRBuilder<> &IRB) { |
| 1273 | Value *FramePointerLong = getCachedFP(IRB); |
| 1274 | Value *UARTag = |
| 1275 | applyTagMask(IRB, OldTag: IRB.CreateLShr(LHS: FramePointerLong, RHS: PointerTagShift)); |
| 1276 | |
| 1277 | UARTag->setName("hwasan.uar.tag" ); |
| 1278 | return UARTag; |
| 1279 | } |
| 1280 | |
| 1281 | // Add a tag to an address. |
| 1282 | Value *HWAddressSanitizer::tagPointer(IRBuilder<> &IRB, Type *Ty, |
| 1283 | Value *PtrLong, Value *Tag) { |
| 1284 | assert(!UsePageAliases); |
| 1285 | Value *TaggedPtrLong; |
| 1286 | if (CompileKernel) { |
| 1287 | // Kernel addresses have 0xFF in the most significant byte. |
| 1288 | Value *ShiftedTag = |
| 1289 | IRB.CreateOr(LHS: IRB.CreateShl(LHS: Tag, RHS: PointerTagShift), |
| 1290 | RHS: ConstantInt::get(Ty: IntptrTy, V: (1ULL << PointerTagShift) - 1)); |
| 1291 | TaggedPtrLong = IRB.CreateAnd(LHS: PtrLong, RHS: ShiftedTag); |
| 1292 | } else { |
| 1293 | // Userspace can simply do OR (tag << PointerTagShift); |
| 1294 | Value *ShiftedTag = IRB.CreateShl(LHS: Tag, RHS: PointerTagShift); |
| 1295 | TaggedPtrLong = IRB.CreateOr(LHS: PtrLong, RHS: ShiftedTag); |
| 1296 | } |
| 1297 | return IRB.CreateIntToPtr(V: TaggedPtrLong, DestTy: Ty); |
| 1298 | } |
| 1299 | |
| 1300 | // Remove tag from an address. |
| 1301 | Value *HWAddressSanitizer::untagPointer(IRBuilder<> &IRB, Value *PtrLong) { |
| 1302 | assert(!UsePageAliases); |
| 1303 | Value *UntaggedPtrLong; |
| 1304 | if (CompileKernel) { |
| 1305 | // Kernel addresses have 0xFF in the most significant byte. |
| 1306 | UntaggedPtrLong = |
| 1307 | IRB.CreateOr(LHS: PtrLong, RHS: ConstantInt::get(Ty: PtrLong->getType(), |
| 1308 | V: TagMaskByte << PointerTagShift)); |
| 1309 | } else { |
| 1310 | // Userspace addresses have 0x00. |
| 1311 | UntaggedPtrLong = IRB.CreateAnd( |
| 1312 | LHS: PtrLong, RHS: ConstantInt::get(Ty: PtrLong->getType(), |
| 1313 | V: ~(TagMaskByte << PointerTagShift))); |
| 1314 | } |
| 1315 | return UntaggedPtrLong; |
| 1316 | } |
| 1317 | |
| 1318 | Value *HWAddressSanitizer::getHwasanThreadSlotPtr(IRBuilder<> &IRB) { |
| 1319 | // Android provides a fixed TLS slot for sanitizers. See TLS_SLOT_SANITIZER |
| 1320 | // in Bionic's libc/platform/bionic/tls_defines.h. |
| 1321 | constexpr int SanitizerSlot = 6; |
| 1322 | if (TargetTriple.isAArch64() && TargetTriple.isAndroid()) |
| 1323 | return memtag::getAndroidSlotPtr(IRB, Slot: SanitizerSlot); |
| 1324 | return ThreadPtrGlobal; |
| 1325 | } |
| 1326 | |
| 1327 | Value *HWAddressSanitizer::getCachedFP(IRBuilder<> &IRB) { |
| 1328 | if (!CachedFP) |
| 1329 | CachedFP = memtag::getFP(IRB); |
| 1330 | return CachedFP; |
| 1331 | } |
| 1332 | |
| 1333 | Value *HWAddressSanitizer::getFrameRecordInfo(IRBuilder<> &IRB) { |
| 1334 | // Prepare ring buffer data. |
| 1335 | Value *PC = memtag::getPC(TargetTriple, IRB); |
| 1336 | Value *FP = getCachedFP(IRB); |
| 1337 | |
| 1338 | // Mix FP and PC. |
| 1339 | // Assumptions: |
| 1340 | // PC is 0x0000PPPPPPPPPPPP (48 bits are meaningful, others are zero) |
| 1341 | // FP is 0xfffffffffffFFFF0 (4 lower bits are zero) |
| 1342 | // We only really need ~20 lower non-zero bits (FFFF), so we mix like this: |
| 1343 | // 0xFFFFPPPPPPPPPPPP |
| 1344 | // |
| 1345 | // FP works because in AArch64FrameLowering::getFrameIndexReference, we |
| 1346 | // prefer FP-relative offsets for functions compiled with HWASan. |
| 1347 | FP = IRB.CreateShl(LHS: FP, RHS: 44); |
| 1348 | return IRB.CreateOr(LHS: PC, RHS: FP); |
| 1349 | } |
| 1350 | |
| 1351 | void HWAddressSanitizer::emitPrologue(IRBuilder<> &IRB, bool WithFrameRecord) { |
| 1352 | if (!Mapping.isInTls()) |
| 1353 | ShadowBase = getShadowNonTls(IRB); |
| 1354 | else if (!WithFrameRecord && TargetTriple.isAndroid()) |
| 1355 | ShadowBase = getDynamicShadowIfunc(IRB); |
| 1356 | |
| 1357 | if (!WithFrameRecord && ShadowBase) |
| 1358 | return; |
| 1359 | |
| 1360 | Value *SlotPtr = nullptr; |
| 1361 | Value *ThreadLong = nullptr; |
| 1362 | Value *ThreadLongMaybeUntagged = nullptr; |
| 1363 | |
| 1364 | auto getThreadLongMaybeUntagged = [&]() { |
| 1365 | if (!SlotPtr) |
| 1366 | SlotPtr = getHwasanThreadSlotPtr(IRB); |
| 1367 | if (!ThreadLong) |
| 1368 | ThreadLong = IRB.CreateLoad(Ty: IntptrTy, Ptr: SlotPtr); |
| 1369 | // Extract the address field from ThreadLong. Unnecessary on AArch64 with |
| 1370 | // TBI. |
| 1371 | return TargetTriple.isAArch64() ? ThreadLong |
| 1372 | : untagPointer(IRB, PtrLong: ThreadLong); |
| 1373 | }; |
| 1374 | |
| 1375 | if (WithFrameRecord) { |
| 1376 | switch (ClRecordStackHistory) { |
| 1377 | case libcall: { |
| 1378 | // Emit a runtime call into hwasan rather than emitting instructions for |
| 1379 | // recording stack history. |
| 1380 | Value *FrameRecordInfo = getFrameRecordInfo(IRB); |
| 1381 | IRB.CreateCall(Callee: HwasanRecordFrameRecordFunc, Args: {FrameRecordInfo}); |
| 1382 | break; |
| 1383 | } |
| 1384 | case instr: { |
| 1385 | ThreadLongMaybeUntagged = getThreadLongMaybeUntagged(); |
| 1386 | |
| 1387 | StackBaseTag = IRB.CreateAShr(LHS: ThreadLong, RHS: 3); |
| 1388 | |
| 1389 | // Store data to ring buffer. |
| 1390 | Value *FrameRecordInfo = getFrameRecordInfo(IRB); |
| 1391 | Value *RecordPtr = |
| 1392 | IRB.CreateIntToPtr(V: ThreadLongMaybeUntagged, DestTy: IRB.getPtrTy(AddrSpace: 0)); |
| 1393 | IRB.CreateStore(Val: FrameRecordInfo, Ptr: RecordPtr); |
| 1394 | |
| 1395 | IRB.CreateStore(Val: memtag::incrementThreadLong(IRB, ThreadLong, Inc: 8), Ptr: SlotPtr); |
| 1396 | break; |
| 1397 | } |
| 1398 | case none: { |
| 1399 | llvm_unreachable( |
| 1400 | "A stack history recording mode should've been selected." ); |
| 1401 | } |
| 1402 | } |
| 1403 | } |
| 1404 | |
| 1405 | if (!ShadowBase) { |
| 1406 | if (!ThreadLongMaybeUntagged) |
| 1407 | ThreadLongMaybeUntagged = getThreadLongMaybeUntagged(); |
| 1408 | |
| 1409 | // Get shadow base address by aligning RecordPtr up. |
| 1410 | // Note: this is not correct if the pointer is already aligned. |
| 1411 | // Runtime library will make sure this never happens. |
| 1412 | ShadowBase = IRB.CreateAdd( |
| 1413 | LHS: IRB.CreateOr( |
| 1414 | LHS: ThreadLongMaybeUntagged, |
| 1415 | RHS: ConstantInt::get(Ty: IntptrTy, V: (1ULL << kShadowBaseAlignment) - 1)), |
| 1416 | RHS: ConstantInt::get(Ty: IntptrTy, V: 1), Name: "hwasan.shadow" ); |
| 1417 | ShadowBase = IRB.CreateIntToPtr(V: ShadowBase, DestTy: PtrTy); |
| 1418 | } |
| 1419 | } |
| 1420 | |
| 1421 | bool HWAddressSanitizer::instrumentLandingPads( |
| 1422 | SmallVectorImpl<Instruction *> &LandingPadVec) { |
| 1423 | for (auto *LP : LandingPadVec) { |
| 1424 | IRBuilder<> IRB(LP->getNextNonDebugInstruction()); |
| 1425 | IRB.CreateCall( |
| 1426 | Callee: HwasanHandleVfork, |
| 1427 | Args: {memtag::readRegister( |
| 1428 | IRB, Name: (TargetTriple.getArch() == Triple::x86_64) ? "rsp" : "sp" )}); |
| 1429 | } |
| 1430 | return true; |
| 1431 | } |
| 1432 | |
| 1433 | bool HWAddressSanitizer::instrumentStack(memtag::StackInfo &SInfo, |
| 1434 | Value *StackTag, Value *UARTag, |
| 1435 | const DominatorTree &DT, |
| 1436 | const PostDominatorTree &PDT, |
| 1437 | const LoopInfo &LI) { |
| 1438 | // Ideally, we want to calculate tagged stack base pointer, and rewrite all |
| 1439 | // alloca addresses using that. Unfortunately, offsets are not known yet |
| 1440 | // (unless we use ASan-style mega-alloca). Instead we keep the base tag in a |
| 1441 | // temp, shift-OR it into each alloca address and xor with the retag mask. |
| 1442 | // This generates one extra instruction per alloca use. |
| 1443 | unsigned int I = 0; |
| 1444 | |
| 1445 | for (auto &KV : SInfo.AllocasToInstrument) { |
| 1446 | auto N = I++; |
| 1447 | auto *AI = KV.first; |
| 1448 | memtag::AllocaInfo &Info = KV.second; |
| 1449 | IRBuilder<> IRB(AI->getNextNonDebugInstruction()); |
| 1450 | |
| 1451 | // Replace uses of the alloca with tagged address. |
| 1452 | Value *Tag = getAllocaTag(IRB, StackTag, AllocaNo: N); |
| 1453 | Value *AILong = IRB.CreatePointerCast(V: AI, DestTy: IntptrTy); |
| 1454 | Value *AINoTagLong = untagPointer(IRB, PtrLong: AILong); |
| 1455 | Value *Replacement = tagPointer(IRB, Ty: AI->getType(), PtrLong: AINoTagLong, Tag); |
| 1456 | std::string Name = |
| 1457 | AI->hasName() ? AI->getName().str() : "alloca." + itostr(X: N); |
| 1458 | Replacement->setName(Name + ".hwasan" ); |
| 1459 | |
| 1460 | size_t Size = memtag::getAllocaSizeInBytes(AI: *AI); |
| 1461 | size_t AlignedSize = alignTo(Size, A: Mapping.getObjectAlignment()); |
| 1462 | |
| 1463 | Value *AICast = IRB.CreatePointerCast(V: AI, DestTy: PtrTy); |
| 1464 | |
| 1465 | auto HandleLifetime = [&](IntrinsicInst *II) { |
| 1466 | // Set the lifetime intrinsic to cover the whole alloca. This reduces the |
| 1467 | // set of assumptions we need to make about the lifetime. Without this we |
| 1468 | // would need to ensure that we can track the lifetime pointer to a |
| 1469 | // constant offset from the alloca, and would still need to change the |
| 1470 | // size to include the extra alignment we use for the untagging to make |
| 1471 | // the size consistent. |
| 1472 | // |
| 1473 | // The check for standard lifetime below makes sure that we have exactly |
| 1474 | // one set of start / end in any execution (i.e. the ends are not |
| 1475 | // reachable from each other), so this will not cause any problems. |
| 1476 | II->setArgOperand(i: 0, v: ConstantInt::get(Ty: Int64Ty, V: AlignedSize)); |
| 1477 | II->setArgOperand(i: 1, v: AICast); |
| 1478 | }; |
| 1479 | llvm::for_each(Range&: Info.LifetimeStart, F: HandleLifetime); |
| 1480 | llvm::for_each(Range&: Info.LifetimeEnd, F: HandleLifetime); |
| 1481 | |
| 1482 | AI->replaceUsesWithIf(New: Replacement, ShouldReplace: [AICast, AILong](const Use &U) { |
| 1483 | auto *User = U.getUser(); |
| 1484 | return User != AILong && User != AICast && !isa<LifetimeIntrinsic>(Val: User); |
| 1485 | }); |
| 1486 | |
| 1487 | memtag::annotateDebugRecords(Info, Tag: retagMask(AllocaNo: N)); |
| 1488 | |
| 1489 | auto TagEnd = [&](Instruction *Node) { |
| 1490 | IRB.SetInsertPoint(Node); |
| 1491 | // When untagging, use the `AlignedSize` because we need to set the tags |
| 1492 | // for the entire alloca to original. If we used `Size` here, we would |
| 1493 | // keep the last granule tagged, and store zero in the last byte of the |
| 1494 | // last granule, due to how short granules are implemented. |
| 1495 | tagAlloca(IRB, AI, Tag: UARTag, Size: AlignedSize); |
| 1496 | }; |
| 1497 | // Calls to functions that may return twice (e.g. setjmp) confuse the |
| 1498 | // postdominator analysis, and will leave us to keep memory tagged after |
| 1499 | // function return. Work around this by always untagging at every return |
| 1500 | // statement if return_twice functions are called. |
| 1501 | bool StandardLifetime = |
| 1502 | !SInfo.CallsReturnTwice && |
| 1503 | SInfo.UnrecognizedLifetimes.empty() && |
| 1504 | memtag::isStandardLifetime(LifetimeStart: Info.LifetimeStart, LifetimeEnd: Info.LifetimeEnd, DT: &DT, |
| 1505 | LI: &LI, MaxLifetimes: ClMaxLifetimes); |
| 1506 | if (DetectUseAfterScope && StandardLifetime) { |
| 1507 | IntrinsicInst *Start = Info.LifetimeStart[0]; |
| 1508 | IRB.SetInsertPoint(Start->getNextNode()); |
| 1509 | tagAlloca(IRB, AI, Tag, Size); |
| 1510 | if (!memtag::forAllReachableExits(DT, PDT, LI, Start, Ends: Info.LifetimeEnd, |
| 1511 | RetVec: SInfo.RetVec, Callback: TagEnd)) { |
| 1512 | for (auto *End : Info.LifetimeEnd) |
| 1513 | End->eraseFromParent(); |
| 1514 | } |
| 1515 | } else { |
| 1516 | tagAlloca(IRB, AI, Tag, Size); |
| 1517 | for (auto *RI : SInfo.RetVec) |
| 1518 | TagEnd(RI); |
| 1519 | // We inserted tagging outside of the lifetimes, so we have to remove |
| 1520 | // them. |
| 1521 | for (auto &II : Info.LifetimeStart) |
| 1522 | II->eraseFromParent(); |
| 1523 | for (auto &II : Info.LifetimeEnd) |
| 1524 | II->eraseFromParent(); |
| 1525 | } |
| 1526 | memtag::alignAndPadAlloca(Info, Align: Mapping.getObjectAlignment()); |
| 1527 | } |
| 1528 | for (auto &I : SInfo.UnrecognizedLifetimes) |
| 1529 | I->eraseFromParent(); |
| 1530 | return true; |
| 1531 | } |
| 1532 | |
| 1533 | static void (const Function &F, OptimizationRemarkEmitter &ORE, |
| 1534 | bool Skip) { |
| 1535 | if (Skip) { |
| 1536 | ORE.emit(RemarkBuilder: [&]() { |
| 1537 | return OptimizationRemark(DEBUG_TYPE, "Skip" , &F) |
| 1538 | << "Skipped: F=" << ore::NV("Function" , &F); |
| 1539 | }); |
| 1540 | } else { |
| 1541 | ORE.emit(RemarkBuilder: [&]() { |
| 1542 | return OptimizationRemarkMissed(DEBUG_TYPE, "Sanitize" , &F) |
| 1543 | << "Sanitized: F=" << ore::NV("Function" , &F); |
| 1544 | }); |
| 1545 | } |
| 1546 | } |
| 1547 | |
| 1548 | bool HWAddressSanitizer::selectiveInstrumentationShouldSkip( |
| 1549 | Function &F, FunctionAnalysisManager &FAM) const { |
| 1550 | auto SkipHot = [&]() { |
| 1551 | if (!ClHotPercentileCutoff.getNumOccurrences()) |
| 1552 | return false; |
| 1553 | auto &MAMProxy = FAM.getResult<ModuleAnalysisManagerFunctionProxy>(IR&: F); |
| 1554 | ProfileSummaryInfo *PSI = |
| 1555 | MAMProxy.getCachedResult<ProfileSummaryAnalysis>(IR&: *F.getParent()); |
| 1556 | if (!PSI || !PSI->hasProfileSummary()) { |
| 1557 | ++NumNoProfileSummaryFuncs; |
| 1558 | return false; |
| 1559 | } |
| 1560 | return PSI->isFunctionHotInCallGraphNthPercentile( |
| 1561 | PercentileCutoff: ClHotPercentileCutoff, F: &F, BFI&: FAM.getResult<BlockFrequencyAnalysis>(IR&: F)); |
| 1562 | }; |
| 1563 | |
| 1564 | auto SkipRandom = [&]() { |
| 1565 | if (!ClRandomKeepRate.getNumOccurrences()) |
| 1566 | return false; |
| 1567 | std::bernoulli_distribution D(ClRandomKeepRate); |
| 1568 | return !D(*Rng); |
| 1569 | }; |
| 1570 | |
| 1571 | bool Skip = SkipRandom() || SkipHot(); |
| 1572 | emitRemark(F, ORE&: FAM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F), Skip); |
| 1573 | return Skip; |
| 1574 | } |
| 1575 | |
| 1576 | void HWAddressSanitizer::sanitizeFunction(Function &F, |
| 1577 | FunctionAnalysisManager &FAM) { |
| 1578 | if (&F == HwasanCtorFunction) |
| 1579 | return; |
| 1580 | |
| 1581 | // Do not apply any instrumentation for naked functions. |
| 1582 | if (F.hasFnAttribute(Kind: Attribute::Naked)) |
| 1583 | return; |
| 1584 | |
| 1585 | if (!F.hasFnAttribute(Kind: Attribute::SanitizeHWAddress)) |
| 1586 | return; |
| 1587 | |
| 1588 | if (F.empty()) |
| 1589 | return; |
| 1590 | |
| 1591 | NumTotalFuncs++; |
| 1592 | |
| 1593 | OptimizationRemarkEmitter &ORE = |
| 1594 | FAM.getResult<OptimizationRemarkEmitterAnalysis>(IR&: F); |
| 1595 | |
| 1596 | if (selectiveInstrumentationShouldSkip(F, FAM)) |
| 1597 | return; |
| 1598 | |
| 1599 | NumInstrumentedFuncs++; |
| 1600 | |
| 1601 | LLVM_DEBUG(dbgs() << "Function: " << F.getName() << "\n" ); |
| 1602 | |
| 1603 | SmallVector<InterestingMemoryOperand, 16> OperandsToInstrument; |
| 1604 | SmallVector<MemIntrinsic *, 16> IntrinToInstrument; |
| 1605 | SmallVector<Instruction *, 8> LandingPadVec; |
| 1606 | const TargetLibraryInfo &TLI = FAM.getResult<TargetLibraryAnalysis>(IR&: F); |
| 1607 | |
| 1608 | memtag::StackInfoBuilder SIB(SSI, DEBUG_TYPE); |
| 1609 | for (auto &Inst : instructions(F)) { |
| 1610 | if (InstrumentStack) { |
| 1611 | SIB.visit(ORE, Inst); |
| 1612 | } |
| 1613 | |
| 1614 | if (InstrumentLandingPads && isa<LandingPadInst>(Val: Inst)) |
| 1615 | LandingPadVec.push_back(Elt: &Inst); |
| 1616 | |
| 1617 | getInterestingMemoryOperands(ORE, I: &Inst, TLI, Interesting&: OperandsToInstrument); |
| 1618 | |
| 1619 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Val: &Inst)) |
| 1620 | if (!ignoreMemIntrinsic(ORE, MI)) |
| 1621 | IntrinToInstrument.push_back(Elt: MI); |
| 1622 | } |
| 1623 | |
| 1624 | memtag::StackInfo &SInfo = SIB.get(); |
| 1625 | |
| 1626 | initializeCallbacks(M&: *F.getParent()); |
| 1627 | |
| 1628 | if (!LandingPadVec.empty()) |
| 1629 | instrumentLandingPads(LandingPadVec); |
| 1630 | |
| 1631 | if (SInfo.AllocasToInstrument.empty() && F.hasPersonalityFn() && |
| 1632 | F.getPersonalityFn()->getName() == kHwasanPersonalityThunkName) { |
| 1633 | // __hwasan_personality_thunk is a no-op for functions without an |
| 1634 | // instrumented stack, so we can drop it. |
| 1635 | F.setPersonalityFn(nullptr); |
| 1636 | } |
| 1637 | |
| 1638 | if (SInfo.AllocasToInstrument.empty() && OperandsToInstrument.empty() && |
| 1639 | IntrinToInstrument.empty()) |
| 1640 | return; |
| 1641 | |
| 1642 | assert(!ShadowBase); |
| 1643 | |
| 1644 | BasicBlock::iterator InsertPt = F.getEntryBlock().begin(); |
| 1645 | IRBuilder<> EntryIRB(&F.getEntryBlock(), InsertPt); |
| 1646 | emitPrologue(IRB&: EntryIRB, |
| 1647 | /*WithFrameRecord*/ ClRecordStackHistory != none && |
| 1648 | Mapping.withFrameRecord() && |
| 1649 | !SInfo.AllocasToInstrument.empty()); |
| 1650 | |
| 1651 | if (!SInfo.AllocasToInstrument.empty()) { |
| 1652 | const DominatorTree &DT = FAM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 1653 | const PostDominatorTree &PDT = FAM.getResult<PostDominatorTreeAnalysis>(IR&: F); |
| 1654 | const LoopInfo &LI = FAM.getResult<LoopAnalysis>(IR&: F); |
| 1655 | Value *StackTag = getStackBaseTag(IRB&: EntryIRB); |
| 1656 | Value *UARTag = getUARTag(IRB&: EntryIRB); |
| 1657 | instrumentStack(SInfo, StackTag, UARTag, DT, PDT, LI); |
| 1658 | } |
| 1659 | |
| 1660 | // If we split the entry block, move any allocas that were originally in the |
| 1661 | // entry block back into the entry block so that they aren't treated as |
| 1662 | // dynamic allocas. |
| 1663 | if (EntryIRB.GetInsertBlock() != &F.getEntryBlock()) { |
| 1664 | InsertPt = F.getEntryBlock().begin(); |
| 1665 | for (Instruction &I : |
| 1666 | llvm::make_early_inc_range(Range&: *EntryIRB.GetInsertBlock())) { |
| 1667 | if (auto *AI = dyn_cast<AllocaInst>(Val: &I)) |
| 1668 | if (isa<ConstantInt>(Val: AI->getArraySize())) |
| 1669 | I.moveBefore(BB&: F.getEntryBlock(), I: InsertPt); |
| 1670 | } |
| 1671 | } |
| 1672 | |
| 1673 | DominatorTree *DT = FAM.getCachedResult<DominatorTreeAnalysis>(IR&: F); |
| 1674 | PostDominatorTree *PDT = FAM.getCachedResult<PostDominatorTreeAnalysis>(IR&: F); |
| 1675 | LoopInfo *LI = FAM.getCachedResult<LoopAnalysis>(IR&: F); |
| 1676 | DomTreeUpdater DTU(DT, PDT, DomTreeUpdater::UpdateStrategy::Lazy); |
| 1677 | const DataLayout &DL = F.getDataLayout(); |
| 1678 | for (auto &Operand : OperandsToInstrument) |
| 1679 | instrumentMemAccess(O&: Operand, DTU, LI, DL); |
| 1680 | DTU.flush(); |
| 1681 | |
| 1682 | if (ClInstrumentMemIntrinsics && !IntrinToInstrument.empty()) { |
| 1683 | for (auto *Inst : IntrinToInstrument) |
| 1684 | instrumentMemIntrinsic(MI: Inst); |
| 1685 | } |
| 1686 | |
| 1687 | ShadowBase = nullptr; |
| 1688 | StackBaseTag = nullptr; |
| 1689 | CachedFP = nullptr; |
| 1690 | } |
| 1691 | |
| 1692 | void HWAddressSanitizer::instrumentGlobal(GlobalVariable *GV, uint8_t Tag) { |
| 1693 | assert(!UsePageAliases); |
| 1694 | Constant *Initializer = GV->getInitializer(); |
| 1695 | uint64_t SizeInBytes = |
| 1696 | M.getDataLayout().getTypeAllocSize(Ty: Initializer->getType()); |
| 1697 | uint64_t NewSize = alignTo(Size: SizeInBytes, A: Mapping.getObjectAlignment()); |
| 1698 | if (SizeInBytes != NewSize) { |
| 1699 | // Pad the initializer out to the next multiple of 16 bytes and add the |
| 1700 | // required short granule tag. |
| 1701 | std::vector<uint8_t> Init(NewSize - SizeInBytes, 0); |
| 1702 | Init.back() = Tag; |
| 1703 | Constant *Padding = ConstantDataArray::get(Context&: *C, Elts&: Init); |
| 1704 | Initializer = ConstantStruct::getAnon(V: {Initializer, Padding}); |
| 1705 | } |
| 1706 | |
| 1707 | auto *NewGV = new GlobalVariable(M, Initializer->getType(), GV->isConstant(), |
| 1708 | GlobalValue::ExternalLinkage, Initializer, |
| 1709 | GV->getName() + ".hwasan" ); |
| 1710 | NewGV->copyAttributesFrom(Src: GV); |
| 1711 | NewGV->setLinkage(GlobalValue::PrivateLinkage); |
| 1712 | NewGV->copyMetadata(Src: GV, Offset: 0); |
| 1713 | NewGV->setAlignment( |
| 1714 | std::max(a: GV->getAlign().valueOrOne(), b: Mapping.getObjectAlignment())); |
| 1715 | |
| 1716 | // It is invalid to ICF two globals that have different tags. In the case |
| 1717 | // where the size of the global is a multiple of the tag granularity the |
| 1718 | // contents of the globals may be the same but the tags (i.e. symbol values) |
| 1719 | // may be different, and the symbols are not considered during ICF. In the |
| 1720 | // case where the size is not a multiple of the granularity, the short granule |
| 1721 | // tags would discriminate two globals with different tags, but there would |
| 1722 | // otherwise be nothing stopping such a global from being incorrectly ICF'd |
| 1723 | // with an uninstrumented (i.e. tag 0) global that happened to have the short |
| 1724 | // granule tag in the last byte. |
| 1725 | NewGV->setUnnamedAddr(GlobalValue::UnnamedAddr::None); |
| 1726 | |
| 1727 | // Descriptor format (assuming little-endian): |
| 1728 | // bytes 0-3: relative address of global |
| 1729 | // bytes 4-6: size of global (16MB ought to be enough for anyone, but in case |
| 1730 | // it isn't, we create multiple descriptors) |
| 1731 | // byte 7: tag |
| 1732 | auto *DescriptorTy = StructType::get(elt1: Int32Ty, elts: Int32Ty); |
| 1733 | const uint64_t MaxDescriptorSize = 0xfffff0; |
| 1734 | for (uint64_t DescriptorPos = 0; DescriptorPos < SizeInBytes; |
| 1735 | DescriptorPos += MaxDescriptorSize) { |
| 1736 | auto *Descriptor = |
| 1737 | new GlobalVariable(M, DescriptorTy, true, GlobalValue::PrivateLinkage, |
| 1738 | nullptr, GV->getName() + ".hwasan.descriptor" ); |
| 1739 | auto *GVRelPtr = ConstantExpr::getTrunc( |
| 1740 | C: ConstantExpr::getAdd( |
| 1741 | C1: ConstantExpr::getSub( |
| 1742 | C1: ConstantExpr::getPtrToInt(C: NewGV, Ty: Int64Ty), |
| 1743 | C2: ConstantExpr::getPtrToInt(C: Descriptor, Ty: Int64Ty)), |
| 1744 | C2: ConstantInt::get(Ty: Int64Ty, V: DescriptorPos)), |
| 1745 | Ty: Int32Ty); |
| 1746 | uint32_t Size = std::min(a: SizeInBytes - DescriptorPos, b: MaxDescriptorSize); |
| 1747 | auto *SizeAndTag = ConstantInt::get(Ty: Int32Ty, V: Size | (uint32_t(Tag) << 24)); |
| 1748 | Descriptor->setComdat(NewGV->getComdat()); |
| 1749 | Descriptor->setInitializer(ConstantStruct::getAnon(V: {GVRelPtr, SizeAndTag})); |
| 1750 | Descriptor->setSection("hwasan_globals" ); |
| 1751 | Descriptor->setMetadata(KindID: LLVMContext::MD_associated, |
| 1752 | Node: MDNode::get(Context&: *C, MDs: ValueAsMetadata::get(V: NewGV))); |
| 1753 | appendToCompilerUsed(M, Values: Descriptor); |
| 1754 | } |
| 1755 | |
| 1756 | Constant *Aliasee = ConstantExpr::getIntToPtr( |
| 1757 | C: ConstantExpr::getAdd( |
| 1758 | C1: ConstantExpr::getPtrToInt(C: NewGV, Ty: Int64Ty), |
| 1759 | C2: ConstantInt::get(Ty: Int64Ty, V: uint64_t(Tag) << PointerTagShift)), |
| 1760 | Ty: GV->getType()); |
| 1761 | auto *Alias = GlobalAlias::create(Ty: GV->getValueType(), AddressSpace: GV->getAddressSpace(), |
| 1762 | Linkage: GV->getLinkage(), Name: "" , Aliasee, Parent: &M); |
| 1763 | Alias->setVisibility(GV->getVisibility()); |
| 1764 | Alias->takeName(V: GV); |
| 1765 | GV->replaceAllUsesWith(V: Alias); |
| 1766 | GV->eraseFromParent(); |
| 1767 | } |
| 1768 | |
| 1769 | void HWAddressSanitizer::instrumentGlobals() { |
| 1770 | std::vector<GlobalVariable *> Globals; |
| 1771 | for (GlobalVariable &GV : M.globals()) { |
| 1772 | if (GV.hasSanitizerMetadata() && GV.getSanitizerMetadata().NoHWAddress) |
| 1773 | continue; |
| 1774 | |
| 1775 | if (GV.isDeclarationForLinker() || GV.getName().starts_with(Prefix: "llvm." ) || |
| 1776 | GV.isThreadLocal()) |
| 1777 | continue; |
| 1778 | |
| 1779 | // Common symbols can't have aliases point to them, so they can't be tagged. |
| 1780 | if (GV.hasCommonLinkage()) |
| 1781 | continue; |
| 1782 | |
| 1783 | // Globals with custom sections may be used in __start_/__stop_ enumeration, |
| 1784 | // which would be broken both by adding tags and potentially by the extra |
| 1785 | // padding/alignment that we insert. |
| 1786 | if (GV.hasSection()) |
| 1787 | continue; |
| 1788 | |
| 1789 | Globals.push_back(x: &GV); |
| 1790 | } |
| 1791 | |
| 1792 | MD5 Hasher; |
| 1793 | Hasher.update(Str: M.getSourceFileName()); |
| 1794 | MD5::MD5Result Hash; |
| 1795 | Hasher.final(Result&: Hash); |
| 1796 | uint8_t Tag = Hash[0]; |
| 1797 | |
| 1798 | assert(TagMaskByte >= 16); |
| 1799 | |
| 1800 | for (GlobalVariable *GV : Globals) { |
| 1801 | // Don't allow globals to be tagged with something that looks like a |
| 1802 | // short-granule tag, otherwise we lose inter-granule overflow detection, as |
| 1803 | // the fast path shadow-vs-address check succeeds. |
| 1804 | if (Tag < 16 || Tag > TagMaskByte) |
| 1805 | Tag = 16; |
| 1806 | instrumentGlobal(GV, Tag: Tag++); |
| 1807 | } |
| 1808 | } |
| 1809 | |
| 1810 | void HWAddressSanitizer::instrumentPersonalityFunctions() { |
| 1811 | // We need to untag stack frames as we unwind past them. That is the job of |
| 1812 | // the personality function wrapper, which either wraps an existing |
| 1813 | // personality function or acts as a personality function on its own. Each |
| 1814 | // function that has a personality function or that can be unwound past has |
| 1815 | // its personality function changed to a thunk that calls the personality |
| 1816 | // function wrapper in the runtime. |
| 1817 | MapVector<Constant *, std::vector<Function *>> PersonalityFns; |
| 1818 | for (Function &F : M) { |
| 1819 | if (F.isDeclaration() || !F.hasFnAttribute(Kind: Attribute::SanitizeHWAddress)) |
| 1820 | continue; |
| 1821 | |
| 1822 | if (F.hasPersonalityFn()) { |
| 1823 | PersonalityFns[F.getPersonalityFn()->stripPointerCasts()].push_back(x: &F); |
| 1824 | } else if (!F.hasFnAttribute(Kind: Attribute::NoUnwind)) { |
| 1825 | PersonalityFns[nullptr].push_back(x: &F); |
| 1826 | } |
| 1827 | } |
| 1828 | |
| 1829 | if (PersonalityFns.empty()) |
| 1830 | return; |
| 1831 | |
| 1832 | FunctionCallee HwasanPersonalityWrapper = M.getOrInsertFunction( |
| 1833 | Name: "__hwasan_personality_wrapper" , RetTy: Int32Ty, Args: Int32Ty, Args: Int32Ty, Args: Int64Ty, Args: PtrTy, |
| 1834 | Args: PtrTy, Args: PtrTy, Args: PtrTy, Args: PtrTy); |
| 1835 | FunctionCallee UnwindGetGR = M.getOrInsertFunction(Name: "_Unwind_GetGR" , RetTy: VoidTy); |
| 1836 | FunctionCallee UnwindGetCFA = M.getOrInsertFunction(Name: "_Unwind_GetCFA" , RetTy: VoidTy); |
| 1837 | |
| 1838 | for (auto &P : PersonalityFns) { |
| 1839 | std::string ThunkName = kHwasanPersonalityThunkName; |
| 1840 | if (P.first) |
| 1841 | ThunkName += ("." + P.first->getName()).str(); |
| 1842 | FunctionType *ThunkFnTy = FunctionType::get( |
| 1843 | Result: Int32Ty, Params: {Int32Ty, Int32Ty, Int64Ty, PtrTy, PtrTy}, isVarArg: false); |
| 1844 | bool IsLocal = P.first && (!isa<GlobalValue>(Val: P.first) || |
| 1845 | cast<GlobalValue>(Val: P.first)->hasLocalLinkage()); |
| 1846 | auto *ThunkFn = Function::Create(Ty: ThunkFnTy, |
| 1847 | Linkage: IsLocal ? GlobalValue::InternalLinkage |
| 1848 | : GlobalValue::LinkOnceODRLinkage, |
| 1849 | N: ThunkName, M: &M); |
| 1850 | // TODO: think about other attributes as well. |
| 1851 | if (any_of(Range&: P.second, P: [](const Function *F) { |
| 1852 | return F->hasFnAttribute(Kind: "branch-target-enforcement" ); |
| 1853 | })) { |
| 1854 | ThunkFn->addFnAttr(Kind: "branch-target-enforcement" ); |
| 1855 | } |
| 1856 | if (!IsLocal) { |
| 1857 | ThunkFn->setVisibility(GlobalValue::HiddenVisibility); |
| 1858 | ThunkFn->setComdat(M.getOrInsertComdat(Name: ThunkName)); |
| 1859 | } |
| 1860 | |
| 1861 | auto *BB = BasicBlock::Create(Context&: *C, Name: "entry" , Parent: ThunkFn); |
| 1862 | IRBuilder<> IRB(BB); |
| 1863 | CallInst *WrapperCall = IRB.CreateCall( |
| 1864 | Callee: HwasanPersonalityWrapper, |
| 1865 | Args: {ThunkFn->getArg(i: 0), ThunkFn->getArg(i: 1), ThunkFn->getArg(i: 2), |
| 1866 | ThunkFn->getArg(i: 3), ThunkFn->getArg(i: 4), |
| 1867 | P.first ? P.first : Constant::getNullValue(Ty: PtrTy), |
| 1868 | UnwindGetGR.getCallee(), UnwindGetCFA.getCallee()}); |
| 1869 | WrapperCall->setTailCall(); |
| 1870 | IRB.CreateRet(V: WrapperCall); |
| 1871 | |
| 1872 | for (Function *F : P.second) |
| 1873 | F->setPersonalityFn(ThunkFn); |
| 1874 | } |
| 1875 | } |
| 1876 | |
| 1877 | void HWAddressSanitizer::ShadowMapping::init(Triple &TargetTriple, |
| 1878 | bool InstrumentWithCalls, |
| 1879 | bool CompileKernel) { |
| 1880 | // Start with defaults. |
| 1881 | Scale = kDefaultShadowScale; |
| 1882 | Kind = OffsetKind::kTls; |
| 1883 | WithFrameRecord = true; |
| 1884 | |
| 1885 | // Tune for the target. |
| 1886 | if (TargetTriple.isOSFuchsia()) { |
| 1887 | // Fuchsia is always PIE, which means that the beginning of the address |
| 1888 | // space is always available. |
| 1889 | SetFixed(0); |
| 1890 | } else if (CompileKernel || InstrumentWithCalls) { |
| 1891 | SetFixed(0); |
| 1892 | WithFrameRecord = false; |
| 1893 | } |
| 1894 | |
| 1895 | WithFrameRecord = optOr(Opt&: ClFrameRecords, Other: WithFrameRecord); |
| 1896 | |
| 1897 | // Apply the last of ClMappingOffset and ClMappingOffsetDynamic. |
| 1898 | Kind = optOr(Opt&: ClMappingOffsetDynamic, Other: Kind); |
| 1899 | if (ClMappingOffset.getNumOccurrences() > 0 && |
| 1900 | !(ClMappingOffsetDynamic.getNumOccurrences() > 0 && |
| 1901 | ClMappingOffsetDynamic.getPosition() > ClMappingOffset.getPosition())) { |
| 1902 | SetFixed(ClMappingOffset); |
| 1903 | } |
| 1904 | } |
| 1905 | |