1 | //===- ARMSLSHardening.cpp - Harden Straight Line Missspeculation ---------===// |
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 file contains a pass to insert code to mitigate against side channel |
10 | // vulnerabilities that may happen under straight line miss-speculation. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "ARM.h" |
15 | #include "ARMInstrInfo.h" |
16 | #include "ARMSubtarget.h" |
17 | #include "llvm/CodeGen/IndirectThunks.h" |
18 | #include "llvm/CodeGen/MachineBasicBlock.h" |
19 | #include "llvm/CodeGen/MachineFunction.h" |
20 | #include "llvm/CodeGen/MachineFunctionPass.h" |
21 | #include "llvm/CodeGen/MachineInstr.h" |
22 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
23 | #include "llvm/CodeGen/MachineOperand.h" |
24 | #include "llvm/IR/DebugLoc.h" |
25 | #include <cassert> |
26 | |
27 | using namespace llvm; |
28 | |
29 | #define DEBUG_TYPE "arm-sls-hardening" |
30 | |
31 | #define ARM_SLS_HARDENING_NAME "ARM sls hardening pass" |
32 | |
33 | namespace { |
34 | |
35 | class ARMSLSHardening : public MachineFunctionPass { |
36 | public: |
37 | const TargetInstrInfo *TII; |
38 | const ARMSubtarget *ST; |
39 | |
40 | static char ID; |
41 | |
42 | ARMSLSHardening() : MachineFunctionPass(ID) { |
43 | initializeARMSLSHardeningPass(*PassRegistry::getPassRegistry()); |
44 | } |
45 | |
46 | bool runOnMachineFunction(MachineFunction &Fn) override; |
47 | |
48 | StringRef getPassName() const override { return ARM_SLS_HARDENING_NAME; } |
49 | |
50 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
51 | AU.setPreservesCFG(); |
52 | MachineFunctionPass::getAnalysisUsage(AU); |
53 | } |
54 | |
55 | private: |
56 | bool hardenReturnsAndBRs(MachineBasicBlock &MBB) const; |
57 | bool hardenIndirectCalls(MachineBasicBlock &MBB) const; |
58 | MachineBasicBlock & |
59 | ConvertIndirectCallToIndirectJump(MachineBasicBlock &MBB, |
60 | MachineBasicBlock::iterator) const; |
61 | }; |
62 | |
63 | } // end anonymous namespace |
64 | |
65 | char ARMSLSHardening::ID = 0; |
66 | |
67 | INITIALIZE_PASS(ARMSLSHardening, "arm-sls-hardening" , |
68 | ARM_SLS_HARDENING_NAME, false, false) |
69 | |
70 | static void insertSpeculationBarrier(const ARMSubtarget *ST, |
71 | MachineBasicBlock &MBB, |
72 | MachineBasicBlock::iterator MBBI, |
73 | DebugLoc DL, |
74 | bool AlwaysUseISBDSB = false) { |
75 | assert(MBBI != MBB.begin() && |
76 | "Must not insert SpeculationBarrierEndBB as only instruction in MBB." ); |
77 | assert(std::prev(MBBI)->isBarrier() && |
78 | "SpeculationBarrierEndBB must only follow unconditional control flow " |
79 | "instructions." ); |
80 | assert(std::prev(MBBI)->isTerminator() && |
81 | "SpeculationBarrierEndBB must only follow terminators." ); |
82 | const TargetInstrInfo *TII = ST->getInstrInfo(); |
83 | assert(ST->hasDataBarrier() || ST->hasSB()); |
84 | bool ProduceSB = ST->hasSB() && !AlwaysUseISBDSB; |
85 | unsigned BarrierOpc = |
86 | ProduceSB ? (ST->isThumb() ? ARM::t2SpeculationBarrierSBEndBB |
87 | : ARM::SpeculationBarrierSBEndBB) |
88 | : (ST->isThumb() ? ARM::t2SpeculationBarrierISBDSBEndBB |
89 | : ARM::SpeculationBarrierISBDSBEndBB); |
90 | if (MBBI == MBB.end() || !isSpeculationBarrierEndBBOpcode(Opc: MBBI->getOpcode())) |
91 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: BarrierOpc)); |
92 | } |
93 | |
94 | bool ARMSLSHardening::runOnMachineFunction(MachineFunction &MF) { |
95 | ST = &MF.getSubtarget<ARMSubtarget>(); |
96 | TII = MF.getSubtarget().getInstrInfo(); |
97 | |
98 | bool Modified = false; |
99 | for (auto &MBB : MF) { |
100 | Modified |= hardenReturnsAndBRs(MBB); |
101 | Modified |= hardenIndirectCalls(MBB); |
102 | } |
103 | |
104 | return Modified; |
105 | } |
106 | |
107 | bool ARMSLSHardening::hardenReturnsAndBRs(MachineBasicBlock &MBB) const { |
108 | if (!ST->hardenSlsRetBr()) |
109 | return false; |
110 | assert(!ST->isThumb1Only()); |
111 | bool Modified = false; |
112 | MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator(), E = MBB.end(); |
113 | MachineBasicBlock::iterator NextMBBI; |
114 | for (; MBBI != E; MBBI = NextMBBI) { |
115 | MachineInstr &MI = *MBBI; |
116 | NextMBBI = std::next(x: MBBI); |
117 | if (isIndirectControlFlowNotComingBack(MI)) { |
118 | assert(MI.isTerminator()); |
119 | assert(!TII->isPredicated(MI)); |
120 | insertSpeculationBarrier(ST, MBB, MBBI: std::next(x: MBBI), DL: MI.getDebugLoc()); |
121 | Modified = true; |
122 | } |
123 | } |
124 | return Modified; |
125 | } |
126 | |
127 | static const char SLSBLRNamePrefix[] = "__llvm_slsblr_thunk_" ; |
128 | |
129 | static const struct ThunkNameRegMode { |
130 | const char* Name; |
131 | Register Reg; |
132 | bool isThumb; |
133 | } SLSBLRThunks[] = { |
134 | {.Name: "__llvm_slsblr_thunk_arm_r0" , .Reg: ARM::R0, .isThumb: false}, |
135 | {.Name: "__llvm_slsblr_thunk_arm_r1" , .Reg: ARM::R1, .isThumb: false}, |
136 | {.Name: "__llvm_slsblr_thunk_arm_r2" , .Reg: ARM::R2, .isThumb: false}, |
137 | {.Name: "__llvm_slsblr_thunk_arm_r3" , .Reg: ARM::R3, .isThumb: false}, |
138 | {.Name: "__llvm_slsblr_thunk_arm_r4" , .Reg: ARM::R4, .isThumb: false}, |
139 | {.Name: "__llvm_slsblr_thunk_arm_r5" , .Reg: ARM::R5, .isThumb: false}, |
140 | {.Name: "__llvm_slsblr_thunk_arm_r6" , .Reg: ARM::R6, .isThumb: false}, |
141 | {.Name: "__llvm_slsblr_thunk_arm_r7" , .Reg: ARM::R7, .isThumb: false}, |
142 | {.Name: "__llvm_slsblr_thunk_arm_r8" , .Reg: ARM::R8, .isThumb: false}, |
143 | {.Name: "__llvm_slsblr_thunk_arm_r9" , .Reg: ARM::R9, .isThumb: false}, |
144 | {.Name: "__llvm_slsblr_thunk_arm_r10" , .Reg: ARM::R10, .isThumb: false}, |
145 | {.Name: "__llvm_slsblr_thunk_arm_r11" , .Reg: ARM::R11, .isThumb: false}, |
146 | {.Name: "__llvm_slsblr_thunk_arm_sp" , .Reg: ARM::SP, .isThumb: false}, |
147 | {.Name: "__llvm_slsblr_thunk_arm_pc" , .Reg: ARM::PC, .isThumb: false}, |
148 | {.Name: "__llvm_slsblr_thunk_thumb_r0" , .Reg: ARM::R0, .isThumb: true}, |
149 | {.Name: "__llvm_slsblr_thunk_thumb_r1" , .Reg: ARM::R1, .isThumb: true}, |
150 | {.Name: "__llvm_slsblr_thunk_thumb_r2" , .Reg: ARM::R2, .isThumb: true}, |
151 | {.Name: "__llvm_slsblr_thunk_thumb_r3" , .Reg: ARM::R3, .isThumb: true}, |
152 | {.Name: "__llvm_slsblr_thunk_thumb_r4" , .Reg: ARM::R4, .isThumb: true}, |
153 | {.Name: "__llvm_slsblr_thunk_thumb_r5" , .Reg: ARM::R5, .isThumb: true}, |
154 | {.Name: "__llvm_slsblr_thunk_thumb_r6" , .Reg: ARM::R6, .isThumb: true}, |
155 | {.Name: "__llvm_slsblr_thunk_thumb_r7" , .Reg: ARM::R7, .isThumb: true}, |
156 | {.Name: "__llvm_slsblr_thunk_thumb_r8" , .Reg: ARM::R8, .isThumb: true}, |
157 | {.Name: "__llvm_slsblr_thunk_thumb_r9" , .Reg: ARM::R9, .isThumb: true}, |
158 | {.Name: "__llvm_slsblr_thunk_thumb_r10" , .Reg: ARM::R10, .isThumb: true}, |
159 | {.Name: "__llvm_slsblr_thunk_thumb_r11" , .Reg: ARM::R11, .isThumb: true}, |
160 | {.Name: "__llvm_slsblr_thunk_thumb_sp" , .Reg: ARM::SP, .isThumb: true}, |
161 | {.Name: "__llvm_slsblr_thunk_thumb_pc" , .Reg: ARM::PC, .isThumb: true}, |
162 | }; |
163 | |
164 | // An enum for tracking whether Arm and Thumb thunks have been inserted into the |
165 | // current module so far. |
166 | enum ArmInsertedThunks { NoThunk = 0, ArmThunk = 1, ThumbThunk = 2 }; |
167 | |
168 | inline ArmInsertedThunks &operator|=(ArmInsertedThunks &X, |
169 | ArmInsertedThunks Y) { |
170 | return X = static_cast<ArmInsertedThunks>(X | Y); |
171 | } |
172 | |
173 | namespace { |
174 | struct SLSBLRThunkInserter |
175 | : ThunkInserter<SLSBLRThunkInserter, ArmInsertedThunks> { |
176 | const char *getThunkPrefix() { return SLSBLRNamePrefix; } |
177 | bool mayUseThunk(const MachineFunction &MF) { |
178 | ComdatThunks &= !MF.getSubtarget<ARMSubtarget>().hardenSlsNoComdat(); |
179 | return MF.getSubtarget<ARMSubtarget>().hardenSlsBlr(); |
180 | } |
181 | ArmInsertedThunks insertThunks(MachineModuleInfo &MMI, MachineFunction &MF, |
182 | ArmInsertedThunks InsertedThunks); |
183 | void populateThunk(MachineFunction &MF); |
184 | |
185 | private: |
186 | bool ComdatThunks = true; |
187 | }; |
188 | } // namespace |
189 | |
190 | ArmInsertedThunks |
191 | SLSBLRThunkInserter::insertThunks(MachineModuleInfo &MMI, MachineFunction &MF, |
192 | ArmInsertedThunks InsertedThunks) { |
193 | if ((InsertedThunks & ArmThunk && |
194 | !MF.getSubtarget<ARMSubtarget>().isThumb()) || |
195 | (InsertedThunks & ThumbThunk && |
196 | MF.getSubtarget<ARMSubtarget>().isThumb())) |
197 | return NoThunk; |
198 | // FIXME: It probably would be possible to filter which thunks to produce |
199 | // based on which registers are actually used in indirect calls in this |
200 | // function. But would that be a worthwhile optimization? |
201 | const ARMSubtarget *ST = &MF.getSubtarget<ARMSubtarget>(); |
202 | for (auto T : SLSBLRThunks) |
203 | if (ST->isThumb() == T.isThumb) |
204 | createThunkFunction(MMI, Name: T.Name, Comdat: ComdatThunks, |
205 | TargetAttrs: T.isThumb ? "+thumb-mode" : "" ); |
206 | return ST->isThumb() ? ThumbThunk : ArmThunk; |
207 | } |
208 | |
209 | void SLSBLRThunkInserter::populateThunk(MachineFunction &MF) { |
210 | assert(MF.getFunction().hasComdat() == ComdatThunks && |
211 | "ComdatThunks value changed since MF creation" ); |
212 | // FIXME: How to better communicate Register number, rather than through |
213 | // name and lookup table? |
214 | assert(MF.getName().starts_with(getThunkPrefix())); |
215 | auto ThunkIt = llvm::find_if( |
216 | Range: SLSBLRThunks, P: [&MF](auto T) { return T.Name == MF.getName(); }); |
217 | assert(ThunkIt != std::end(SLSBLRThunks)); |
218 | Register ThunkReg = ThunkIt->Reg; |
219 | bool isThumb = ThunkIt->isThumb; |
220 | |
221 | const TargetInstrInfo *TII = MF.getSubtarget<ARMSubtarget>().getInstrInfo(); |
222 | MachineBasicBlock *Entry = &MF.front(); |
223 | Entry->clear(); |
224 | |
225 | // These thunks need to consist of the following instructions: |
226 | // __llvm_slsblr_thunk_(arm/thumb)_rN: |
227 | // bx rN |
228 | // barrierInsts |
229 | Entry->addLiveIn(PhysReg: ThunkReg); |
230 | if (isThumb) |
231 | BuildMI(BB: Entry, MIMD: DebugLoc(), MCID: TII->get(Opcode: ARM::tBX)) |
232 | .addReg(RegNo: ThunkReg) |
233 | .add(MOs: predOps(Pred: ARMCC::AL)); |
234 | else |
235 | BuildMI(BB: Entry, MIMD: DebugLoc(), MCID: TII->get(Opcode: ARM::BX)) |
236 | .addReg(RegNo: ThunkReg); |
237 | |
238 | // Make sure the thunks do not make use of the SB extension in case there is |
239 | // a function somewhere that will call to it that for some reason disabled |
240 | // the SB extension locally on that function, even though it's enabled for |
241 | // the module otherwise. Therefore set AlwaysUseISBSDB to true. |
242 | insertSpeculationBarrier(ST: &MF.getSubtarget<ARMSubtarget>(), MBB&: *Entry, |
243 | MBBI: Entry->end(), DL: DebugLoc(), AlwaysUseISBDSB: true /*AlwaysUseISBDSB*/); |
244 | } |
245 | |
246 | MachineBasicBlock &ARMSLSHardening::ConvertIndirectCallToIndirectJump( |
247 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI) const { |
248 | // Transform an indirect call to an indirect jump as follows: |
249 | // Before: |
250 | // |-----------------------------| |
251 | // | ... | |
252 | // | instI | |
253 | // | BLX rN | |
254 | // | instJ | |
255 | // | ... | |
256 | // |-----------------------------| |
257 | // |
258 | // After: |
259 | // |---------- -------------------------| |
260 | // | ... | |
261 | // | instI | |
262 | // | *call* __llvm_slsblr_thunk_mode_xN | |
263 | // | instJ | |
264 | // | ... | |
265 | // |--------------------------------------| |
266 | // |
267 | // __llvm_slsblr_thunk_mode_xN: |
268 | // |-----------------------------| |
269 | // | BX rN | |
270 | // | barrierInsts | |
271 | // |-----------------------------| |
272 | // |
273 | // The __llvm_slsblr_thunk_mode_xN thunks are created by the |
274 | // SLSBLRThunkInserter. |
275 | // This function merely needs to transform an indirect call to a direct call |
276 | // to __llvm_slsblr_thunk_xN. |
277 | MachineInstr &IndirectCall = *MBBI; |
278 | assert(isIndirectCall(IndirectCall) && !IndirectCall.isReturn()); |
279 | int RegOpIdxOnIndirectCall = -1; |
280 | bool isThumb; |
281 | switch (IndirectCall.getOpcode()) { |
282 | case ARM::BLX: // !isThumb2 |
283 | case ARM::BLX_noip: // !isThumb2 |
284 | isThumb = false; |
285 | RegOpIdxOnIndirectCall = 0; |
286 | break; |
287 | case ARM::tBLXr: // isThumb2 |
288 | case ARM::tBLXr_noip: // isThumb2 |
289 | isThumb = true; |
290 | RegOpIdxOnIndirectCall = 2; |
291 | break; |
292 | default: |
293 | llvm_unreachable("unhandled Indirect Call" ); |
294 | } |
295 | |
296 | Register Reg = IndirectCall.getOperand(i: RegOpIdxOnIndirectCall).getReg(); |
297 | // Since linkers are allowed to clobber R12 on function calls, the above |
298 | // mitigation only works if the original indirect call instruction was not |
299 | // using R12. Code generation before must make sure that no indirect call |
300 | // using R12 was produced if the mitigation is enabled. |
301 | // Also, the transformation is incorrect if the indirect call uses LR, so |
302 | // also have to avoid that. |
303 | assert(Reg != ARM::R12 && Reg != ARM::LR); |
304 | bool RegIsKilled = IndirectCall.getOperand(i: RegOpIdxOnIndirectCall).isKill(); |
305 | |
306 | DebugLoc DL = IndirectCall.getDebugLoc(); |
307 | |
308 | MachineFunction &MF = *MBBI->getMF(); |
309 | auto ThunkIt = llvm::find_if(Range: SLSBLRThunks, P: [Reg, isThumb](auto T) { |
310 | return T.Reg == Reg && T.isThumb == isThumb; |
311 | }); |
312 | assert(ThunkIt != std::end(SLSBLRThunks)); |
313 | Module *M = MF.getFunction().getParent(); |
314 | const GlobalValue *GV = cast<GlobalValue>(Val: M->getNamedValue(Name: ThunkIt->Name)); |
315 | |
316 | MachineInstr *BL = |
317 | isThumb ? BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: ARM::tBL)) |
318 | .addImm(Val: IndirectCall.getOperand(i: 0).getImm()) |
319 | .addReg(RegNo: IndirectCall.getOperand(i: 1).getReg()) |
320 | .addGlobalAddress(GV) |
321 | : BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: ARM::BL)).addGlobalAddress(GV); |
322 | |
323 | // Now copy the implicit operands from IndirectCall to BL and copy other |
324 | // necessary info. |
325 | // However, both IndirectCall and BL instructions implictly use SP and |
326 | // implicitly define LR. Blindly copying implicit operands would result in SP |
327 | // and LR operands to be present multiple times. While this may not be too |
328 | // much of an issue, let's avoid that for cleanliness, by removing those |
329 | // implicit operands from the BL created above before we copy over all |
330 | // implicit operands from the IndirectCall. |
331 | int ImpLROpIdx = -1; |
332 | int ImpSPOpIdx = -1; |
333 | for (unsigned OpIdx = BL->getNumExplicitOperands(); |
334 | OpIdx < BL->getNumOperands(); OpIdx++) { |
335 | MachineOperand Op = BL->getOperand(i: OpIdx); |
336 | if (!Op.isReg()) |
337 | continue; |
338 | if (Op.getReg() == ARM::LR && Op.isDef()) |
339 | ImpLROpIdx = OpIdx; |
340 | if (Op.getReg() == ARM::SP && !Op.isDef()) |
341 | ImpSPOpIdx = OpIdx; |
342 | } |
343 | assert(ImpLROpIdx != -1); |
344 | assert(ImpSPOpIdx != -1); |
345 | int FirstOpIdxToRemove = std::max(a: ImpLROpIdx, b: ImpSPOpIdx); |
346 | int SecondOpIdxToRemove = std::min(a: ImpLROpIdx, b: ImpSPOpIdx); |
347 | BL->removeOperand(OpNo: FirstOpIdxToRemove); |
348 | BL->removeOperand(OpNo: SecondOpIdxToRemove); |
349 | // Now copy over the implicit operands from the original IndirectCall |
350 | BL->copyImplicitOps(MF, MI: IndirectCall); |
351 | MF.moveCallSiteInfo(Old: &IndirectCall, New: BL); |
352 | // Also add the register called in the IndirectCall as being used in the |
353 | // called thunk. |
354 | BL->addOperand(Op: MachineOperand::CreateReg(Reg, isDef: false /*isDef*/, isImp: true /*isImp*/, |
355 | isKill: RegIsKilled /*isKill*/)); |
356 | // Remove IndirectCallinstruction |
357 | MBB.erase(I: MBBI); |
358 | return MBB; |
359 | } |
360 | |
361 | bool ARMSLSHardening::hardenIndirectCalls(MachineBasicBlock &MBB) const { |
362 | if (!ST->hardenSlsBlr()) |
363 | return false; |
364 | bool Modified = false; |
365 | MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); |
366 | MachineBasicBlock::iterator NextMBBI; |
367 | for (; MBBI != E; MBBI = NextMBBI) { |
368 | MachineInstr &MI = *MBBI; |
369 | NextMBBI = std::next(x: MBBI); |
370 | // Tail calls are both indirect calls and "returns". |
371 | // They are also indirect jumps, so should be handled by sls-harden-retbr, |
372 | // rather than sls-harden-blr. |
373 | if (isIndirectCall(MI) && !MI.isReturn()) { |
374 | ConvertIndirectCallToIndirectJump(MBB, MBBI); |
375 | Modified = true; |
376 | } |
377 | } |
378 | return Modified; |
379 | } |
380 | |
381 | |
382 | |
383 | FunctionPass *llvm::createARMSLSHardeningPass() { |
384 | return new ARMSLSHardening(); |
385 | } |
386 | |
387 | namespace { |
388 | class ARMIndirectThunks : public ThunkInserterPass<SLSBLRThunkInserter> { |
389 | public: |
390 | static char ID; |
391 | |
392 | ARMIndirectThunks() : ThunkInserterPass(ID) {} |
393 | |
394 | StringRef getPassName() const override { return "ARM Indirect Thunks" ; } |
395 | }; |
396 | } // end anonymous namespace |
397 | |
398 | char ARMIndirectThunks::ID = 0; |
399 | |
400 | FunctionPass *llvm::createARMIndirectThunks() { |
401 | return new ARMIndirectThunks(); |
402 | } |
403 | |