1 //===- StackProtector.cpp - Stack Protector Insertion ---------------------===// 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 // This pass inserts stack protectors into functions which need them. A variable 10 // with a random value in it is stored onto the stack before the local variables 11 // are allocated. Upon exiting the block, the stored value is checked. If it's 12 // changed, then there was some sort of violation and the program aborts. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "llvm/CodeGen/StackProtector.h" 17 #include "llvm/ADT/SmallPtrSet.h" 18 #include "llvm/ADT/Statistic.h" 19 #include "llvm/Analysis/BranchProbabilityInfo.h" 20 #include "llvm/Analysis/EHPersonalities.h" 21 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 22 #include "llvm/CodeGen/Passes.h" 23 #include "llvm/CodeGen/TargetLowering.h" 24 #include "llvm/CodeGen/TargetPassConfig.h" 25 #include "llvm/CodeGen/TargetSubtargetInfo.h" 26 #include "llvm/IR/Attributes.h" 27 #include "llvm/IR/BasicBlock.h" 28 #include "llvm/IR/Constants.h" 29 #include "llvm/IR/DataLayout.h" 30 #include "llvm/IR/DebugInfo.h" 31 #include "llvm/IR/DebugLoc.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/Instruction.h" 37 #include "llvm/IR/Instructions.h" 38 #include "llvm/IR/IntrinsicInst.h" 39 #include "llvm/IR/Intrinsics.h" 40 #include "llvm/IR/MDBuilder.h" 41 #include "llvm/IR/Module.h" 42 #include "llvm/IR/Type.h" 43 #include "llvm/IR/User.h" 44 #include "llvm/Pass.h" 45 #include "llvm/Support/Casting.h" 46 #include "llvm/Support/CommandLine.h" 47 #include "llvm/Target/TargetMachine.h" 48 #include "llvm/Target/TargetOptions.h" 49 #include <utility> 50 51 using namespace llvm; 52 53 #define DEBUG_TYPE "stack-protector" 54 55 STATISTIC(NumFunProtected, "Number of functions protected"); 56 STATISTIC(NumAddrTaken, "Number of local variables that have their address" 57 " taken."); 58 59 static cl::opt<bool> EnableSelectionDAGSP("enable-selectiondag-sp", 60 cl::init(true), cl::Hidden); 61 62 char StackProtector::ID = 0; 63 64 INITIALIZE_PASS_BEGIN(StackProtector, DEBUG_TYPE, 65 "Insert stack protectors", false, true) 66 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) 67 INITIALIZE_PASS_END(StackProtector, DEBUG_TYPE, 68 "Insert stack protectors", false, true) 69 70 FunctionPass *llvm::createStackProtectorPass() { return new StackProtector(); } 71 72 void StackProtector::getAnalysisUsage(AnalysisUsage &AU) const { 73 AU.addRequired<TargetPassConfig>(); 74 AU.addPreserved<DominatorTreeWrapperPass>(); 75 } 76 77 bool StackProtector::runOnFunction(Function &Fn) { 78 F = &Fn; 79 M = F->getParent(); 80 DominatorTreeWrapperPass *DTWP = 81 getAnalysisIfAvailable<DominatorTreeWrapperPass>(); 82 DT = DTWP ? &DTWP->getDomTree() : nullptr; 83 TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>(); 84 Trip = TM->getTargetTriple(); 85 TLI = TM->getSubtargetImpl(Fn)->getTargetLowering(); 86 HasPrologue = false; 87 HasIRCheck = false; 88 89 Attribute Attr = Fn.getFnAttribute("stack-protector-buffer-size"); 90 if (Attr.isStringAttribute() && 91 Attr.getValueAsString().getAsInteger(10, SSPBufferSize)) 92 return false; // Invalid integer string 93 94 if (!RequiresStackProtector()) 95 return false; 96 97 // TODO(etienneb): Functions with funclets are not correctly supported now. 98 // Do nothing if this is funclet-based personality. 99 if (Fn.hasPersonalityFn()) { 100 EHPersonality Personality = classifyEHPersonality(Fn.getPersonalityFn()); 101 if (isFuncletEHPersonality(Personality)) 102 return false; 103 } 104 105 ++NumFunProtected; 106 return InsertStackProtectors(); 107 } 108 109 /// \param [out] IsLarge is set to true if a protectable array is found and 110 /// it is "large" ( >= ssp-buffer-size). In the case of a structure with 111 /// multiple arrays, this gets set if any of them is large. 112 bool StackProtector::ContainsProtectableArray(Type *Ty, bool &IsLarge, 113 bool Strong, 114 bool InStruct) const { 115 if (!Ty) 116 return false; 117 if (ArrayType *AT = dyn_cast<ArrayType>(Ty)) { 118 if (!AT->getElementType()->isIntegerTy(8)) { 119 // If we're on a non-Darwin platform or we're inside of a structure, don't 120 // add stack protectors unless the array is a character array. 121 // However, in strong mode any array, regardless of type and size, 122 // triggers a protector. 123 if (!Strong && (InStruct || !Trip.isOSDarwin())) 124 return false; 125 } 126 127 // If an array has more than SSPBufferSize bytes of allocated space, then we 128 // emit stack protectors. 129 if (SSPBufferSize <= M->getDataLayout().getTypeAllocSize(AT)) { 130 IsLarge = true; 131 return true; 132 } 133 134 if (Strong) 135 // Require a protector for all arrays in strong mode 136 return true; 137 } 138 139 const StructType *ST = dyn_cast<StructType>(Ty); 140 if (!ST) 141 return false; 142 143 bool NeedsProtector = false; 144 for (StructType::element_iterator I = ST->element_begin(), 145 E = ST->element_end(); 146 I != E; ++I) 147 if (ContainsProtectableArray(*I, IsLarge, Strong, true)) { 148 // If the element is a protectable array and is large (>= SSPBufferSize) 149 // then we are done. If the protectable array is not large, then 150 // keep looking in case a subsequent element is a large array. 151 if (IsLarge) 152 return true; 153 NeedsProtector = true; 154 } 155 156 return NeedsProtector; 157 } 158 159 bool StackProtector::HasAddressTaken(const Instruction *AI, 160 SmallPtrSetImpl<const PHINode *> &VisitedPHIs) { 161 for (const User *U : AI->users()) { 162 const auto *I = cast<Instruction>(U); 163 switch (I->getOpcode()) { 164 case Instruction::Store: 165 if (AI == cast<StoreInst>(I)->getValueOperand()) 166 return true; 167 break; 168 case Instruction::AtomicCmpXchg: 169 // cmpxchg conceptually includes both a load and store from the same 170 // location. So, like store, the value being stored is what matters. 171 if (AI == cast<AtomicCmpXchgInst>(I)->getNewValOperand()) 172 return true; 173 break; 174 case Instruction::PtrToInt: 175 if (AI == cast<PtrToIntInst>(I)->getOperand(0)) 176 return true; 177 break; 178 case Instruction::Call: { 179 // Ignore intrinsics that do not become real instructions. 180 // TODO: Narrow this to intrinsics that have store-like effects. 181 const auto *CI = cast<CallInst>(I); 182 if (!isa<DbgInfoIntrinsic>(CI) && !CI->isLifetimeStartOrEnd()) 183 return true; 184 break; 185 } 186 case Instruction::Invoke: 187 return true; 188 case Instruction::BitCast: 189 case Instruction::GetElementPtr: 190 case Instruction::Select: 191 case Instruction::AddrSpaceCast: 192 if (HasAddressTaken(I, VisitedPHIs)) 193 return true; 194 break; 195 case Instruction::PHI: { 196 // Keep track of what PHI nodes we have already visited to ensure 197 // they are only visited once. 198 const auto *PN = cast<PHINode>(I); 199 if (VisitedPHIs.insert(PN).second) 200 if (HasAddressTaken(PN, VisitedPHIs)) 201 return true; 202 break; 203 } 204 case Instruction::Load: 205 case Instruction::AtomicRMW: 206 case Instruction::Ret: 207 // These instructions take an address operand, but have load-like or 208 // other innocuous behavior that should not trigger a stack protector. 209 // atomicrmw conceptually has both load and store semantics, but the 210 // value being stored must be integer; so if a pointer is being stored, 211 // we'll catch it in the PtrToInt case above. 212 break; 213 default: 214 // Conservatively return true for any instruction that takes an address 215 // operand, but is not handled above. 216 return true; 217 } 218 } 219 return false; 220 } 221 222 /// Search for the first call to the llvm.stackprotector intrinsic and return it 223 /// if present. 224 static const CallInst *findStackProtectorIntrinsic(Function &F) { 225 for (const BasicBlock &BB : F) 226 for (const Instruction &I : BB) 227 if (const CallInst *CI = dyn_cast<CallInst>(&I)) 228 if (CI->getCalledFunction() == 229 Intrinsic::getDeclaration(F.getParent(), Intrinsic::stackprotector)) 230 return CI; 231 return nullptr; 232 } 233 234 /// Check whether or not this function needs a stack protector based 235 /// upon the stack protector level. 236 /// 237 /// We use two heuristics: a standard (ssp) and strong (sspstrong). 238 /// The standard heuristic which will add a guard variable to functions that 239 /// call alloca with a either a variable size or a size >= SSPBufferSize, 240 /// functions with character buffers larger than SSPBufferSize, and functions 241 /// with aggregates containing character buffers larger than SSPBufferSize. The 242 /// strong heuristic will add a guard variables to functions that call alloca 243 /// regardless of size, functions with any buffer regardless of type and size, 244 /// functions with aggregates that contain any buffer regardless of type and 245 /// size, and functions that contain stack-based variables that have had their 246 /// address taken. 247 bool StackProtector::RequiresStackProtector() { 248 bool Strong = false; 249 bool NeedsProtector = false; 250 HasPrologue = findStackProtectorIntrinsic(*F); 251 252 if (F->hasFnAttribute(Attribute::SafeStack)) 253 return false; 254 255 // We are constructing the OptimizationRemarkEmitter on the fly rather than 256 // using the analysis pass to avoid building DominatorTree and LoopInfo which 257 // are not available this late in the IR pipeline. 258 OptimizationRemarkEmitter ORE(F); 259 260 if (F->hasFnAttribute(Attribute::StackProtectReq)) { 261 ORE.emit([&]() { 262 return OptimizationRemark(DEBUG_TYPE, "StackProtectorRequested", F) 263 << "Stack protection applied to function " 264 << ore::NV("Function", F) 265 << " due to a function attribute or command-line switch"; 266 }); 267 NeedsProtector = true; 268 Strong = true; // Use the same heuristic as strong to determine SSPLayout 269 } else if (F->hasFnAttribute(Attribute::StackProtectStrong)) 270 Strong = true; 271 else if (HasPrologue) 272 NeedsProtector = true; 273 else if (!F->hasFnAttribute(Attribute::StackProtect)) 274 return false; 275 276 /// VisitedPHIs - The set of PHI nodes visited when determining 277 /// if a variable's reference has been taken. This set 278 /// is maintained to ensure we don't visit the same PHI node multiple 279 /// times. 280 SmallPtrSet<const PHINode *, 16> VisitedPHIs; 281 282 for (const BasicBlock &BB : *F) { 283 for (const Instruction &I : BB) { 284 if (const AllocaInst *AI = dyn_cast<AllocaInst>(&I)) { 285 if (AI->isArrayAllocation()) { 286 auto RemarkBuilder = [&]() { 287 return OptimizationRemark(DEBUG_TYPE, "StackProtectorAllocaOrArray", 288 &I) 289 << "Stack protection applied to function " 290 << ore::NV("Function", F) 291 << " due to a call to alloca or use of a variable length " 292 "array"; 293 }; 294 if (const auto *CI = dyn_cast<ConstantInt>(AI->getArraySize())) { 295 if (CI->getLimitedValue(SSPBufferSize) >= SSPBufferSize) { 296 // A call to alloca with size >= SSPBufferSize requires 297 // stack protectors. 298 Layout.insert(std::make_pair(AI, 299 MachineFrameInfo::SSPLK_LargeArray)); 300 ORE.emit(RemarkBuilder); 301 NeedsProtector = true; 302 } else if (Strong) { 303 // Require protectors for all alloca calls in strong mode. 304 Layout.insert(std::make_pair(AI, 305 MachineFrameInfo::SSPLK_SmallArray)); 306 ORE.emit(RemarkBuilder); 307 NeedsProtector = true; 308 } 309 } else { 310 // A call to alloca with a variable size requires protectors. 311 Layout.insert(std::make_pair(AI, 312 MachineFrameInfo::SSPLK_LargeArray)); 313 ORE.emit(RemarkBuilder); 314 NeedsProtector = true; 315 } 316 continue; 317 } 318 319 bool IsLarge = false; 320 if (ContainsProtectableArray(AI->getAllocatedType(), IsLarge, Strong)) { 321 Layout.insert(std::make_pair(AI, IsLarge 322 ? MachineFrameInfo::SSPLK_LargeArray 323 : MachineFrameInfo::SSPLK_SmallArray)); 324 ORE.emit([&]() { 325 return OptimizationRemark(DEBUG_TYPE, "StackProtectorBuffer", &I) 326 << "Stack protection applied to function " 327 << ore::NV("Function", F) 328 << " due to a stack allocated buffer or struct containing a " 329 "buffer"; 330 }); 331 NeedsProtector = true; 332 continue; 333 } 334 335 if (Strong && HasAddressTaken(AI, VisitedPHIs)) { 336 ++NumAddrTaken; 337 Layout.insert(std::make_pair(AI, MachineFrameInfo::SSPLK_AddrOf)); 338 ORE.emit([&]() { 339 return OptimizationRemark(DEBUG_TYPE, "StackProtectorAddressTaken", 340 &I) 341 << "Stack protection applied to function " 342 << ore::NV("Function", F) 343 << " due to the address of a local variable being taken"; 344 }); 345 NeedsProtector = true; 346 } 347 } 348 } 349 } 350 351 return NeedsProtector; 352 } 353 354 /// Create a stack guard loading and populate whether SelectionDAG SSP is 355 /// supported. 356 static Value *getStackGuard(const TargetLoweringBase *TLI, Module *M, 357 IRBuilder<> &B, 358 bool *SupportsSelectionDAGSP = nullptr) { 359 if (Value *Guard = TLI->getIRStackGuard(B)) 360 return B.CreateLoad(B.getInt8PtrTy(), Guard, true, "StackGuard"); 361 362 // Use SelectionDAG SSP handling, since there isn't an IR guard. 363 // 364 // This is more or less weird, since we optionally output whether we 365 // should perform a SelectionDAG SP here. The reason is that it's strictly 366 // defined as !TLI->getIRStackGuard(B), where getIRStackGuard is also 367 // mutating. There is no way to get this bit without mutating the IR, so 368 // getting this bit has to happen in this right time. 369 // 370 // We could have define a new function TLI::supportsSelectionDAGSP(), but that 371 // will put more burden on the backends' overriding work, especially when it 372 // actually conveys the same information getIRStackGuard() already gives. 373 if (SupportsSelectionDAGSP) 374 *SupportsSelectionDAGSP = true; 375 TLI->insertSSPDeclarations(*M); 376 return B.CreateCall(Intrinsic::getDeclaration(M, Intrinsic::stackguard)); 377 } 378 379 /// Insert code into the entry block that stores the stack guard 380 /// variable onto the stack: 381 /// 382 /// entry: 383 /// StackGuardSlot = alloca i8* 384 /// StackGuard = <stack guard> 385 /// call void @llvm.stackprotector(StackGuard, StackGuardSlot) 386 /// 387 /// Returns true if the platform/triple supports the stackprotectorcreate pseudo 388 /// node. 389 static bool CreatePrologue(Function *F, Module *M, ReturnInst *RI, 390 const TargetLoweringBase *TLI, AllocaInst *&AI) { 391 bool SupportsSelectionDAGSP = false; 392 IRBuilder<> B(&F->getEntryBlock().front()); 393 PointerType *PtrTy = Type::getInt8PtrTy(RI->getContext()); 394 AI = B.CreateAlloca(PtrTy, nullptr, "StackGuardSlot"); 395 396 Value *GuardSlot = getStackGuard(TLI, M, B, &SupportsSelectionDAGSP); 397 B.CreateCall(Intrinsic::getDeclaration(M, Intrinsic::stackprotector), 398 {GuardSlot, AI}); 399 return SupportsSelectionDAGSP; 400 } 401 402 /// InsertStackProtectors - Insert code into the prologue and epilogue of the 403 /// function. 404 /// 405 /// - The prologue code loads and stores the stack guard onto the stack. 406 /// - The epilogue checks the value stored in the prologue against the original 407 /// value. It calls __stack_chk_fail if they differ. 408 bool StackProtector::InsertStackProtectors() { 409 // If the target wants to XOR the frame pointer into the guard value, it's 410 // impossible to emit the check in IR, so the target *must* support stack 411 // protection in SDAG. 412 bool SupportsSelectionDAGSP = 413 TLI->useStackGuardXorFP() || 414 (EnableSelectionDAGSP && !TM->Options.EnableFastISel && 415 !TM->Options.EnableGlobalISel); 416 AllocaInst *AI = nullptr; // Place on stack that stores the stack guard. 417 418 for (Function::iterator I = F->begin(), E = F->end(); I != E;) { 419 BasicBlock *BB = &*I++; 420 ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator()); 421 if (!RI) 422 continue; 423 424 // Generate prologue instrumentation if not already generated. 425 if (!HasPrologue) { 426 HasPrologue = true; 427 SupportsSelectionDAGSP &= CreatePrologue(F, M, RI, TLI, AI); 428 } 429 430 // SelectionDAG based code generation. Nothing else needs to be done here. 431 // The epilogue instrumentation is postponed to SelectionDAG. 432 if (SupportsSelectionDAGSP) 433 break; 434 435 // Find the stack guard slot if the prologue was not created by this pass 436 // itself via a previous call to CreatePrologue(). 437 if (!AI) { 438 const CallInst *SPCall = findStackProtectorIntrinsic(*F); 439 assert(SPCall && "Call to llvm.stackprotector is missing"); 440 AI = cast<AllocaInst>(SPCall->getArgOperand(1)); 441 } 442 443 // Set HasIRCheck to true, so that SelectionDAG will not generate its own 444 // version. SelectionDAG called 'shouldEmitSDCheck' to check whether 445 // instrumentation has already been generated. 446 HasIRCheck = true; 447 448 // Generate epilogue instrumentation. The epilogue intrumentation can be 449 // function-based or inlined depending on which mechanism the target is 450 // providing. 451 if (Function *GuardCheck = TLI->getSSPStackGuardCheck(*M)) { 452 // Generate the function-based epilogue instrumentation. 453 // The target provides a guard check function, generate a call to it. 454 IRBuilder<> B(RI); 455 LoadInst *Guard = B.CreateLoad(B.getInt8PtrTy(), AI, true, "Guard"); 456 CallInst *Call = B.CreateCall(GuardCheck, {Guard}); 457 Call->setAttributes(GuardCheck->getAttributes()); 458 Call->setCallingConv(GuardCheck->getCallingConv()); 459 } else { 460 // Generate the epilogue with inline instrumentation. 461 // If we do not support SelectionDAG based tail calls, generate IR level 462 // tail calls. 463 // 464 // For each block with a return instruction, convert this: 465 // 466 // return: 467 // ... 468 // ret ... 469 // 470 // into this: 471 // 472 // return: 473 // ... 474 // %1 = <stack guard> 475 // %2 = load StackGuardSlot 476 // %3 = cmp i1 %1, %2 477 // br i1 %3, label %SP_return, label %CallStackCheckFailBlk 478 // 479 // SP_return: 480 // ret ... 481 // 482 // CallStackCheckFailBlk: 483 // call void @__stack_chk_fail() 484 // unreachable 485 486 // Create the FailBB. We duplicate the BB every time since the MI tail 487 // merge pass will merge together all of the various BB into one including 488 // fail BB generated by the stack protector pseudo instruction. 489 BasicBlock *FailBB = CreateFailBB(); 490 491 // Split the basic block before the return instruction. 492 BasicBlock *NewBB = BB->splitBasicBlock(RI->getIterator(), "SP_return"); 493 494 // Update the dominator tree if we need to. 495 if (DT && DT->isReachableFromEntry(BB)) { 496 DT->addNewBlock(NewBB, BB); 497 DT->addNewBlock(FailBB, BB); 498 } 499 500 // Remove default branch instruction to the new BB. 501 BB->getTerminator()->eraseFromParent(); 502 503 // Move the newly created basic block to the point right after the old 504 // basic block so that it's in the "fall through" position. 505 NewBB->moveAfter(BB); 506 507 // Generate the stack protector instructions in the old basic block. 508 IRBuilder<> B(BB); 509 Value *Guard = getStackGuard(TLI, M, B); 510 LoadInst *LI2 = B.CreateLoad(B.getInt8PtrTy(), AI, true); 511 Value *Cmp = B.CreateICmpEQ(Guard, LI2); 512 auto SuccessProb = 513 BranchProbabilityInfo::getBranchProbStackProtector(true); 514 auto FailureProb = 515 BranchProbabilityInfo::getBranchProbStackProtector(false); 516 MDNode *Weights = MDBuilder(F->getContext()) 517 .createBranchWeights(SuccessProb.getNumerator(), 518 FailureProb.getNumerator()); 519 B.CreateCondBr(Cmp, NewBB, FailBB, Weights); 520 } 521 } 522 523 // Return if we didn't modify any basic blocks. i.e., there are no return 524 // statements in the function. 525 return HasPrologue; 526 } 527 528 /// CreateFailBB - Create a basic block to jump to when the stack protector 529 /// check fails. 530 BasicBlock *StackProtector::CreateFailBB() { 531 LLVMContext &Context = F->getContext(); 532 BasicBlock *FailBB = BasicBlock::Create(Context, "CallStackCheckFailBlk", F); 533 IRBuilder<> B(FailBB); 534 B.SetCurrentDebugLocation(DebugLoc::get(0, 0, F->getSubprogram())); 535 if (Trip.isOSOpenBSD()) { 536 FunctionCallee StackChkFail = M->getOrInsertFunction( 537 "__stack_smash_handler", Type::getVoidTy(Context), 538 Type::getInt8PtrTy(Context)); 539 540 B.CreateCall(StackChkFail, B.CreateGlobalStringPtr(F->getName(), "SSH")); 541 } else { 542 FunctionCallee StackChkFail = 543 M->getOrInsertFunction("__stack_chk_fail", Type::getVoidTy(Context)); 544 545 B.CreateCall(StackChkFail, {}); 546 } 547 B.CreateUnreachable(); 548 return FailBB; 549 } 550 551 bool StackProtector::shouldEmitSDCheck(const BasicBlock &BB) const { 552 return HasPrologue && !HasIRCheck && isa<ReturnInst>(BB.getTerminator()); 553 } 554 555 void StackProtector::copyToMachineFrameInfo(MachineFrameInfo &MFI) const { 556 if (Layout.empty()) 557 return; 558 559 for (int I = 0, E = MFI.getObjectIndexEnd(); I != E; ++I) { 560 if (MFI.isDeadObjectIndex(I)) 561 continue; 562 563 const AllocaInst *AI = MFI.getObjectAllocation(I); 564 if (!AI) 565 continue; 566 567 SSPLayoutMap::const_iterator LI = Layout.find(AI); 568 if (LI == Layout.end()) 569 continue; 570 571 MFI.setObjectSSPLayout(I, LI->second); 572 } 573 } 574