1//==-- X86LoadValueInjectionLoadHardening.cpp - LVI load hardening for x86 --=//
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/// Description: This pass finds Load Value Injection (LVI) gadgets consisting
10/// of a load from memory (i.e., SOURCE), and any operation that may transmit
11/// the value loaded from memory over a covert channel, or use the value loaded
12/// from memory to determine a branch/call target (i.e., SINK). After finding
13/// all such gadgets in a given function, the pass minimally inserts LFENCE
14/// instructions in such a manner that the following property is satisfied: for
15/// all SOURCE+SINK pairs, all paths in the CFG from SOURCE to SINK contain at
16/// least one LFENCE instruction. The algorithm that implements this minimal
17/// insertion is influenced by an academic paper that minimally inserts memory
18/// fences for high-performance concurrent programs:
19/// http://www.cs.ucr.edu/~lesani/companion/oopsla15/OOPSLA15.pdf
20/// The algorithm implemented in this pass is as follows:
21/// 1. Build a condensed CFG (i.e., a GadgetGraph) consisting only of the
22/// following components:
23/// - SOURCE instructions (also includes function arguments)
24/// - SINK instructions
25/// - Basic block entry points
26/// - Basic block terminators
27/// - LFENCE instructions
28/// 2. Analyze the GadgetGraph to determine which SOURCE+SINK pairs (i.e.,
29/// gadgets) are already mitigated by existing LFENCEs. If all gadgets have been
30/// mitigated, go to step 6.
31/// 3. Use a heuristic or plugin to approximate minimal LFENCE insertion.
32/// 4. Insert one LFENCE along each CFG edge that was cut in step 3.
33/// 5. Go to step 2.
34/// 6. If any LFENCEs were inserted, return `true` from runOnMachineFunction()
35/// to tell LLVM that the function was modified.
36///
37//===----------------------------------------------------------------------===//
38
39#include "ImmutableGraph.h"
40#include "X86.h"
41#include "X86Subtarget.h"
42#include "X86TargetMachine.h"
43#include "llvm/ADT/DenseMap.h"
44#include "llvm/ADT/STLExtras.h"
45#include "llvm/ADT/SmallSet.h"
46#include "llvm/ADT/Statistic.h"
47#include "llvm/ADT/StringRef.h"
48#include "llvm/CodeGen/MachineBasicBlock.h"
49#include "llvm/CodeGen/MachineDominanceFrontier.h"
50#include "llvm/CodeGen/MachineDominators.h"
51#include "llvm/CodeGen/MachineFunction.h"
52#include "llvm/CodeGen/MachineFunctionPass.h"
53#include "llvm/CodeGen/MachineInstr.h"
54#include "llvm/CodeGen/MachineInstrBuilder.h"
55#include "llvm/CodeGen/MachineLoopInfo.h"
56#include "llvm/CodeGen/MachineRegisterInfo.h"
57#include "llvm/CodeGen/RDFGraph.h"
58#include "llvm/CodeGen/RDFLiveness.h"
59#include "llvm/InitializePasses.h"
60#include "llvm/Support/CommandLine.h"
61#include "llvm/Support/DOTGraphTraits.h"
62#include "llvm/Support/Debug.h"
63#include "llvm/Support/DynamicLibrary.h"
64#include "llvm/Support/GraphWriter.h"
65#include "llvm/Support/raw_ostream.h"
66
67using namespace llvm;
68
69#define PASS_KEY "x86-lvi-load"
70#define DEBUG_TYPE PASS_KEY
71
72STATISTIC(NumFences, "Number of LFENCEs inserted for LVI mitigation");
73STATISTIC(NumFunctionsConsidered, "Number of functions analyzed");
74STATISTIC(NumFunctionsMitigated, "Number of functions for which mitigations "
75 "were deployed");
76STATISTIC(NumGadgets, "Number of LVI gadgets detected during analysis");
77
78static cl::opt<std::string> OptimizePluginPath(
79 PASS_KEY "-opt-plugin",
80 cl::desc("Specify a plugin to optimize LFENCE insertion"), cl::Hidden);
81
82static cl::opt<bool> NoConditionalBranches(
83 PASS_KEY "-no-cbranch",
84 cl::desc("Don't treat conditional branches as disclosure gadgets. This "
85 "may improve performance, at the cost of security."),
86 cl::init(Val: false), cl::Hidden);
87
88static cl::opt<bool> EmitDot(
89 PASS_KEY "-dot",
90 cl::desc(
91 "For each function, emit a dot graph depicting potential LVI gadgets"),
92 cl::init(Val: false), cl::Hidden);
93
94static cl::opt<bool> EmitDotOnly(
95 PASS_KEY "-dot-only",
96 cl::desc("For each function, emit a dot graph depicting potential LVI "
97 "gadgets, and do not insert any fences"),
98 cl::init(Val: false), cl::Hidden);
99
100static cl::opt<bool> EmitDotVerify(
101 PASS_KEY "-dot-verify",
102 cl::desc("For each function, emit a dot graph to stdout depicting "
103 "potential LVI gadgets, used for testing purposes only"),
104 cl::init(Val: false), cl::Hidden);
105
106static llvm::sys::DynamicLibrary OptimizeDL;
107typedef int (*OptimizeCutT)(unsigned int *Nodes, unsigned int NodesSize,
108 unsigned int *Edges, int *EdgeValues,
109 int *CutEdges /* out */, unsigned int EdgesSize);
110static OptimizeCutT OptimizeCut = nullptr;
111
112namespace {
113
114struct MachineGadgetGraph : ImmutableGraph<MachineInstr *, int> {
115 static constexpr int GadgetEdgeSentinel = -1;
116 static constexpr MachineInstr *const ArgNodeSentinel = nullptr;
117
118 using GraphT = ImmutableGraph<MachineInstr *, int>;
119 using Node = typename GraphT::Node;
120 using Edge = typename GraphT::Edge;
121 using size_type = typename GraphT::size_type;
122 MachineGadgetGraph(std::unique_ptr<Node[]> Nodes,
123 std::unique_ptr<Edge[]> Edges, size_type NodesSize,
124 size_type EdgesSize, int NumFences = 0, int NumGadgets = 0)
125 : GraphT(std::move(Nodes), std::move(Edges), NodesSize, EdgesSize),
126 NumFences(NumFences), NumGadgets(NumGadgets) {}
127 static inline bool isCFGEdge(const Edge &E) {
128 return E.getValue() != GadgetEdgeSentinel;
129 }
130 static inline bool isGadgetEdge(const Edge &E) {
131 return E.getValue() == GadgetEdgeSentinel;
132 }
133 int NumFences;
134 int NumGadgets;
135};
136
137class X86LoadValueInjectionLoadHardeningPass : public MachineFunctionPass {
138public:
139 X86LoadValueInjectionLoadHardeningPass() : MachineFunctionPass(ID) {}
140
141 StringRef getPassName() const override {
142 return "X86 Load Value Injection (LVI) Load Hardening";
143 }
144 void getAnalysisUsage(AnalysisUsage &AU) const override;
145 bool runOnMachineFunction(MachineFunction &MF) override;
146
147 static char ID;
148
149private:
150 using GraphBuilder = ImmutableGraphBuilder<MachineGadgetGraph>;
151 using Edge = MachineGadgetGraph::Edge;
152 using Node = MachineGadgetGraph::Node;
153 using EdgeSet = MachineGadgetGraph::EdgeSet;
154 using NodeSet = MachineGadgetGraph::NodeSet;
155
156 const X86Subtarget *STI = nullptr;
157 const TargetInstrInfo *TII = nullptr;
158 const TargetRegisterInfo *TRI = nullptr;
159
160 std::unique_ptr<MachineGadgetGraph>
161 getGadgetGraph(MachineFunction &MF, const MachineLoopInfo &MLI,
162 const MachineDominatorTree &MDT,
163 const MachineDominanceFrontier &MDF) const;
164 int hardenLoadsWithPlugin(MachineFunction &MF,
165 std::unique_ptr<MachineGadgetGraph> Graph) const;
166 int hardenLoadsWithHeuristic(MachineFunction &MF,
167 std::unique_ptr<MachineGadgetGraph> Graph) const;
168 int elimMitigatedEdgesAndNodes(MachineGadgetGraph &G,
169 EdgeSet &ElimEdges /* in, out */,
170 NodeSet &ElimNodes /* in, out */) const;
171 std::unique_ptr<MachineGadgetGraph>
172 trimMitigatedEdges(std::unique_ptr<MachineGadgetGraph> Graph) const;
173 int insertFences(MachineFunction &MF, MachineGadgetGraph &G,
174 EdgeSet &CutEdges /* in, out */) const;
175 bool instrUsesRegToAccessMemory(const MachineInstr &I, unsigned Reg) const;
176 bool instrUsesRegToBranch(const MachineInstr &I, unsigned Reg) const;
177 inline bool isFence(const MachineInstr *MI) const {
178 return MI && (MI->getOpcode() == X86::LFENCE ||
179 (STI->useLVIControlFlowIntegrity() && MI->isCall()));
180 }
181};
182
183} // end anonymous namespace
184
185namespace llvm {
186
187template <>
188struct GraphTraits<MachineGadgetGraph *>
189 : GraphTraits<ImmutableGraph<MachineInstr *, int> *> {};
190
191template <>
192struct DOTGraphTraits<MachineGadgetGraph *> : DefaultDOTGraphTraits {
193 using GraphType = MachineGadgetGraph;
194 using Traits = llvm::GraphTraits<GraphType *>;
195 using NodeRef = typename Traits::NodeRef;
196 using EdgeRef = typename Traits::EdgeRef;
197 using ChildIteratorType = typename Traits::ChildIteratorType;
198 using ChildEdgeIteratorType = typename Traits::ChildEdgeIteratorType;
199
200 DOTGraphTraits(bool IsSimple = false) : DefaultDOTGraphTraits(IsSimple) {}
201
202 std::string getNodeLabel(NodeRef Node, GraphType *) {
203 if (Node->getValue() == MachineGadgetGraph::ArgNodeSentinel)
204 return "ARGS";
205
206 std::string Str;
207 raw_string_ostream OS(Str);
208 OS << *Node->getValue();
209 return OS.str();
210 }
211
212 static std::string getNodeAttributes(NodeRef Node, GraphType *) {
213 MachineInstr *MI = Node->getValue();
214 if (MI == MachineGadgetGraph::ArgNodeSentinel)
215 return "color = blue";
216 if (MI->getOpcode() == X86::LFENCE)
217 return "color = green";
218 return "";
219 }
220
221 static std::string getEdgeAttributes(NodeRef, ChildIteratorType E,
222 GraphType *) {
223 int EdgeVal = (*E.getCurrent()).getValue();
224 return EdgeVal >= 0 ? "label = " + std::to_string(val: EdgeVal)
225 : "color = red, style = \"dashed\"";
226 }
227};
228
229} // end namespace llvm
230
231constexpr MachineInstr *MachineGadgetGraph::ArgNodeSentinel;
232constexpr int MachineGadgetGraph::GadgetEdgeSentinel;
233
234char X86LoadValueInjectionLoadHardeningPass::ID = 0;
235
236void X86LoadValueInjectionLoadHardeningPass::getAnalysisUsage(
237 AnalysisUsage &AU) const {
238 MachineFunctionPass::getAnalysisUsage(AU);
239 AU.addRequired<MachineLoopInfoWrapperPass>();
240 AU.addRequired<MachineDominatorTreeWrapperPass>();
241 AU.addRequired<MachineDominanceFrontier>();
242 AU.setPreservesCFG();
243}
244
245static void writeGadgetGraph(raw_ostream &OS, MachineFunction &MF,
246 MachineGadgetGraph *G) {
247 WriteGraph(O&: OS, G, /*ShortNames*/ false,
248 Title: "Speculative gadgets for \"" + MF.getName() + "\" function");
249}
250
251bool X86LoadValueInjectionLoadHardeningPass::runOnMachineFunction(
252 MachineFunction &MF) {
253 LLVM_DEBUG(dbgs() << "***** " << getPassName() << " : " << MF.getName()
254 << " *****\n");
255 STI = &MF.getSubtarget<X86Subtarget>();
256 if (!STI->useLVILoadHardening())
257 return false;
258
259 // FIXME: support 32-bit
260 if (!STI->is64Bit())
261 report_fatal_error(reason: "LVI load hardening is only supported on 64-bit", gen_crash_diag: false);
262
263 // Don't skip functions with the "optnone" attr but participate in opt-bisect.
264 const Function &F = MF.getFunction();
265 if (!F.hasOptNone() && skipFunction(F))
266 return false;
267
268 ++NumFunctionsConsidered;
269 TII = STI->getInstrInfo();
270 TRI = STI->getRegisterInfo();
271 LLVM_DEBUG(dbgs() << "Building gadget graph...\n");
272 const auto &MLI = getAnalysis<MachineLoopInfoWrapperPass>().getLI();
273 const auto &MDT = getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree();
274 const auto &MDF = getAnalysis<MachineDominanceFrontier>();
275 std::unique_ptr<MachineGadgetGraph> Graph = getGadgetGraph(MF, MLI, MDT, MDF);
276 LLVM_DEBUG(dbgs() << "Building gadget graph... Done\n");
277 if (Graph == nullptr)
278 return false; // didn't find any gadgets
279
280 if (EmitDotVerify) {
281 writeGadgetGraph(OS&: outs(), MF, G: Graph.get());
282 return false;
283 }
284
285 if (EmitDot || EmitDotOnly) {
286 LLVM_DEBUG(dbgs() << "Emitting gadget graph...\n");
287 std::error_code FileError;
288 std::string FileName = "lvi.";
289 FileName += MF.getName();
290 FileName += ".dot";
291 raw_fd_ostream FileOut(FileName, FileError);
292 if (FileError)
293 errs() << FileError.message();
294 writeGadgetGraph(OS&: FileOut, MF, G: Graph.get());
295 FileOut.close();
296 LLVM_DEBUG(dbgs() << "Emitting gadget graph... Done\n");
297 if (EmitDotOnly)
298 return false;
299 }
300
301 int FencesInserted;
302 if (!OptimizePluginPath.empty()) {
303 if (!OptimizeDL.isValid()) {
304 std::string ErrorMsg;
305 OptimizeDL = llvm::sys::DynamicLibrary::getPermanentLibrary(
306 filename: OptimizePluginPath.c_str(), errMsg: &ErrorMsg);
307 if (!ErrorMsg.empty())
308 report_fatal_error(reason: Twine("Failed to load opt plugin: \"") + ErrorMsg +
309 "\"");
310 OptimizeCut = (OptimizeCutT)OptimizeDL.getAddressOfSymbol(symbolName: "optimize_cut");
311 if (!OptimizeCut)
312 report_fatal_error(reason: "Invalid optimization plugin");
313 }
314 FencesInserted = hardenLoadsWithPlugin(MF, Graph: std::move(Graph));
315 } else { // Use the default greedy heuristic
316 FencesInserted = hardenLoadsWithHeuristic(MF, Graph: std::move(Graph));
317 }
318
319 if (FencesInserted > 0)
320 ++NumFunctionsMitigated;
321 NumFences += FencesInserted;
322 return (FencesInserted > 0);
323}
324
325std::unique_ptr<MachineGadgetGraph>
326X86LoadValueInjectionLoadHardeningPass::getGadgetGraph(
327 MachineFunction &MF, const MachineLoopInfo &MLI,
328 const MachineDominatorTree &MDT,
329 const MachineDominanceFrontier &MDF) const {
330 using namespace rdf;
331
332 // Build the Register Dataflow Graph using the RDF framework
333 DataFlowGraph DFG{MF, *TII, *TRI, MDT, MDF};
334 DFG.build();
335 Liveness L{MF.getRegInfo(), DFG};
336 L.computePhiInfo();
337
338 GraphBuilder Builder;
339 using GraphIter = typename GraphBuilder::BuilderNodeRef;
340 DenseMap<MachineInstr *, GraphIter> NodeMap;
341 int FenceCount = 0, GadgetCount = 0;
342 auto MaybeAddNode = [&NodeMap, &Builder](MachineInstr *MI) {
343 auto Ref = NodeMap.find(Val: MI);
344 if (Ref == NodeMap.end()) {
345 auto I = Builder.addVertex(V: MI);
346 NodeMap[MI] = I;
347 return std::pair<GraphIter, bool>{I, true};
348 }
349 return std::pair<GraphIter, bool>{Ref->getSecond(), false};
350 };
351
352 // The `Transmitters` map memoizes transmitters found for each def. If a def
353 // has not yet been analyzed, then it will not appear in the map. If a def
354 // has been analyzed and was determined not to have any transmitters, then
355 // its list of transmitters will be empty.
356 DenseMap<NodeId, std::vector<NodeId>> Transmitters;
357
358 // Analyze all machine instructions to find gadgets and LFENCEs, adding
359 // each interesting value to `Nodes`
360 auto AnalyzeDef = [&](NodeAddr<DefNode *> SourceDef) {
361 SmallSet<NodeId, 8> UsesVisited, DefsVisited;
362 std::function<void(NodeAddr<DefNode *>)> AnalyzeDefUseChain =
363 [&](NodeAddr<DefNode *> Def) {
364 if (Transmitters.contains(Val: Def.Id))
365 return; // Already analyzed `Def`
366
367 // Use RDF to find all the uses of `Def`
368 rdf::NodeSet Uses;
369 RegisterRef DefReg = Def.Addr->getRegRef(G: DFG);
370 for (auto UseID : L.getAllReachedUses(RefRR: DefReg, DefA: Def)) {
371 auto Use = DFG.addr<UseNode *>(N: UseID);
372 if (Use.Addr->getFlags() & NodeAttrs::PhiRef) { // phi node
373 NodeAddr<PhiNode *> Phi = Use.Addr->getOwner(G: DFG);
374 for (const auto& I : L.getRealUses(P: Phi.Id)) {
375 if (DFG.getPRI().alias(RA: RegisterRef(I.first), RB: DefReg)) {
376 for (const auto &UA : I.second)
377 Uses.emplace(args: UA.first);
378 }
379 }
380 } else { // not a phi node
381 Uses.emplace(args&: UseID);
382 }
383 }
384
385 // For each use of `Def`, we want to know whether:
386 // (1) The use can leak the Def'ed value,
387 // (2) The use can further propagate the Def'ed value to more defs
388 for (auto UseID : Uses) {
389 if (!UsesVisited.insert(V: UseID).second)
390 continue; // Already visited this use of `Def`
391
392 auto Use = DFG.addr<UseNode *>(N: UseID);
393 assert(!(Use.Addr->getFlags() & NodeAttrs::PhiRef));
394 MachineOperand &UseMO = Use.Addr->getOp();
395 MachineInstr &UseMI = *UseMO.getParent();
396 assert(UseMO.isReg());
397
398 // We naively assume that an instruction propagates any loaded
399 // uses to all defs unless the instruction is a call, in which
400 // case all arguments will be treated as gadget sources during
401 // analysis of the callee function.
402 if (UseMI.isCall())
403 continue;
404
405 // Check whether this use can transmit (leak) its value.
406 if (instrUsesRegToAccessMemory(I: UseMI, Reg: UseMO.getReg()) ||
407 (!NoConditionalBranches &&
408 instrUsesRegToBranch(I: UseMI, Reg: UseMO.getReg()))) {
409 Transmitters[Def.Id].push_back(x: Use.Addr->getOwner(G: DFG).Id);
410 if (UseMI.mayLoad())
411 continue; // Found a transmitting load -- no need to continue
412 // traversing its defs (i.e., this load will become
413 // a new gadget source anyways).
414 }
415
416 // Check whether the use propagates to more defs.
417 NodeAddr<InstrNode *> Owner{Use.Addr->getOwner(G: DFG)};
418 rdf::NodeList AnalyzedChildDefs;
419 for (const auto &ChildDef :
420 Owner.Addr->members_if(P: DataFlowGraph::IsDef, G: DFG)) {
421 if (!DefsVisited.insert(V: ChildDef.Id).second)
422 continue; // Already visited this def
423 if (Def.Addr->getAttrs() & NodeAttrs::Dead)
424 continue;
425 if (Def.Id == ChildDef.Id)
426 continue; // `Def` uses itself (e.g., increment loop counter)
427
428 AnalyzeDefUseChain(ChildDef);
429
430 // `Def` inherits all of its child defs' transmitters.
431 for (auto TransmitterId : Transmitters[ChildDef.Id])
432 Transmitters[Def.Id].push_back(x: TransmitterId);
433 }
434 }
435
436 // Note that this statement adds `Def.Id` to the map if no
437 // transmitters were found for `Def`.
438 auto &DefTransmitters = Transmitters[Def.Id];
439
440 // Remove duplicate transmitters
441 llvm::sort(C&: DefTransmitters);
442 DefTransmitters.erase(first: llvm::unique(R&: DefTransmitters),
443 last: DefTransmitters.end());
444 };
445
446 // Find all of the transmitters
447 AnalyzeDefUseChain(SourceDef);
448 auto &SourceDefTransmitters = Transmitters[SourceDef.Id];
449 if (SourceDefTransmitters.empty())
450 return; // No transmitters for `SourceDef`
451
452 MachineInstr *Source = SourceDef.Addr->getFlags() & NodeAttrs::PhiRef
453 ? MachineGadgetGraph::ArgNodeSentinel
454 : SourceDef.Addr->getOp().getParent();
455 auto GadgetSource = MaybeAddNode(Source);
456 // Each transmitter is a sink for `SourceDef`.
457 for (auto TransmitterId : SourceDefTransmitters) {
458 MachineInstr *Sink = DFG.addr<StmtNode *>(N: TransmitterId).Addr->getCode();
459 auto GadgetSink = MaybeAddNode(Sink);
460 // Add the gadget edge to the graph.
461 Builder.addEdge(E: MachineGadgetGraph::GadgetEdgeSentinel,
462 From: GadgetSource.first, To: GadgetSink.first);
463 ++GadgetCount;
464 }
465 };
466
467 LLVM_DEBUG(dbgs() << "Analyzing def-use chains to find gadgets\n");
468 // Analyze function arguments
469 NodeAddr<BlockNode *> EntryBlock = DFG.getFunc().Addr->getEntryBlock(G: DFG);
470 for (NodeAddr<PhiNode *> ArgPhi :
471 EntryBlock.Addr->members_if(P: DataFlowGraph::IsPhi, G: DFG)) {
472 NodeList Defs = ArgPhi.Addr->members_if(P: DataFlowGraph::IsDef, G: DFG);
473 llvm::for_each(Range&: Defs, F: AnalyzeDef);
474 }
475 // Analyze every instruction in MF
476 for (NodeAddr<BlockNode *> BA : DFG.getFunc().Addr->members(G: DFG)) {
477 for (NodeAddr<StmtNode *> SA :
478 BA.Addr->members_if(P: DataFlowGraph::IsCode<NodeAttrs::Stmt>, G: DFG)) {
479 MachineInstr *MI = SA.Addr->getCode();
480 if (isFence(MI)) {
481 MaybeAddNode(MI);
482 ++FenceCount;
483 } else if (MI->mayLoad()) {
484 NodeList Defs = SA.Addr->members_if(P: DataFlowGraph::IsDef, G: DFG);
485 llvm::for_each(Range&: Defs, F: AnalyzeDef);
486 }
487 }
488 }
489 LLVM_DEBUG(dbgs() << "Found " << FenceCount << " fences\n");
490 LLVM_DEBUG(dbgs() << "Found " << GadgetCount << " gadgets\n");
491 if (GadgetCount == 0)
492 return nullptr;
493 NumGadgets += GadgetCount;
494
495 // Traverse CFG to build the rest of the graph
496 SmallSet<MachineBasicBlock *, 8> BlocksVisited;
497 std::function<void(MachineBasicBlock *, GraphIter, unsigned)> TraverseCFG =
498 [&](MachineBasicBlock *MBB, GraphIter GI, unsigned ParentDepth) {
499 unsigned LoopDepth = MLI.getLoopDepth(BB: MBB);
500 if (!MBB->empty()) {
501 // Always add the first instruction in each block
502 auto NI = MBB->begin();
503 auto BeginBB = MaybeAddNode(&*NI);
504 Builder.addEdge(E: ParentDepth, From: GI, To: BeginBB.first);
505 if (!BlocksVisited.insert(Ptr: MBB).second)
506 return;
507
508 // Add any instructions within the block that are gadget components
509 GI = BeginBB.first;
510 while (++NI != MBB->end()) {
511 auto Ref = NodeMap.find(Val: &*NI);
512 if (Ref != NodeMap.end()) {
513 Builder.addEdge(E: LoopDepth, From: GI, To: Ref->getSecond());
514 GI = Ref->getSecond();
515 }
516 }
517
518 // Always add the terminator instruction, if one exists
519 auto T = MBB->getFirstTerminator();
520 if (T != MBB->end()) {
521 auto EndBB = MaybeAddNode(&*T);
522 if (EndBB.second)
523 Builder.addEdge(E: LoopDepth, From: GI, To: EndBB.first);
524 GI = EndBB.first;
525 }
526 }
527 for (MachineBasicBlock *Succ : MBB->successors())
528 TraverseCFG(Succ, GI, LoopDepth);
529 };
530 // ArgNodeSentinel is a pseudo-instruction that represents MF args in the
531 // GadgetGraph
532 GraphIter ArgNode = MaybeAddNode(MachineGadgetGraph::ArgNodeSentinel).first;
533 TraverseCFG(&MF.front(), ArgNode, 0);
534 std::unique_ptr<MachineGadgetGraph> G{Builder.get(Args&: FenceCount, Args&: GadgetCount)};
535 LLVM_DEBUG(dbgs() << "Found " << G->nodes_size() << " nodes\n");
536 return G;
537}
538
539// Returns the number of remaining gadget edges that could not be eliminated
540int X86LoadValueInjectionLoadHardeningPass::elimMitigatedEdgesAndNodes(
541 MachineGadgetGraph &G, EdgeSet &ElimEdges /* in, out */,
542 NodeSet &ElimNodes /* in, out */) const {
543 if (G.NumFences > 0) {
544 // Eliminate fences and CFG edges that ingress and egress the fence, as
545 // they are trivially mitigated.
546 for (const Edge &E : G.edges()) {
547 const Node *Dest = E.getDest();
548 if (isFence(MI: Dest->getValue())) {
549 ElimNodes.insert(N: *Dest);
550 ElimEdges.insert(E);
551 for (const Edge &DE : Dest->edges())
552 ElimEdges.insert(E: DE);
553 }
554 }
555 }
556
557 // Find and eliminate gadget edges that have been mitigated.
558 int RemainingGadgets = 0;
559 NodeSet ReachableNodes{G};
560 for (const Node &RootN : G.nodes()) {
561 if (llvm::none_of(Range: RootN.edges(), P: MachineGadgetGraph::isGadgetEdge))
562 continue; // skip this node if it isn't a gadget source
563
564 // Find all of the nodes that are CFG-reachable from RootN using DFS
565 ReachableNodes.clear();
566 std::function<void(const Node *, bool)> FindReachableNodes =
567 [&](const Node *N, bool FirstNode) {
568 if (!FirstNode)
569 ReachableNodes.insert(N: *N);
570 for (const Edge &E : N->edges()) {
571 const Node *Dest = E.getDest();
572 if (MachineGadgetGraph::isCFGEdge(E) && !ElimEdges.contains(E) &&
573 !ReachableNodes.contains(N: *Dest))
574 FindReachableNodes(Dest, false);
575 }
576 };
577 FindReachableNodes(&RootN, true);
578
579 // Any gadget whose sink is unreachable has been mitigated
580 for (const Edge &E : RootN.edges()) {
581 if (MachineGadgetGraph::isGadgetEdge(E)) {
582 if (ReachableNodes.contains(N: *E.getDest())) {
583 // This gadget's sink is reachable
584 ++RemainingGadgets;
585 } else { // This gadget's sink is unreachable, and therefore mitigated
586 ElimEdges.insert(E);
587 }
588 }
589 }
590 }
591 return RemainingGadgets;
592}
593
594std::unique_ptr<MachineGadgetGraph>
595X86LoadValueInjectionLoadHardeningPass::trimMitigatedEdges(
596 std::unique_ptr<MachineGadgetGraph> Graph) const {
597 NodeSet ElimNodes{*Graph};
598 EdgeSet ElimEdges{*Graph};
599 int RemainingGadgets =
600 elimMitigatedEdgesAndNodes(G&: *Graph, ElimEdges, ElimNodes);
601 if (ElimEdges.empty() && ElimNodes.empty()) {
602 Graph->NumFences = 0;
603 Graph->NumGadgets = RemainingGadgets;
604 } else {
605 Graph = GraphBuilder::trim(G: *Graph, TrimNodes: ElimNodes, TrimEdges: ElimEdges, Args: 0 /* NumFences */,
606 Args&: RemainingGadgets);
607 }
608 return Graph;
609}
610
611int X86LoadValueInjectionLoadHardeningPass::hardenLoadsWithPlugin(
612 MachineFunction &MF, std::unique_ptr<MachineGadgetGraph> Graph) const {
613 int FencesInserted = 0;
614
615 do {
616 LLVM_DEBUG(dbgs() << "Eliminating mitigated paths...\n");
617 Graph = trimMitigatedEdges(Graph: std::move(Graph));
618 LLVM_DEBUG(dbgs() << "Eliminating mitigated paths... Done\n");
619 if (Graph->NumGadgets == 0)
620 break;
621
622 LLVM_DEBUG(dbgs() << "Cutting edges...\n");
623 EdgeSet CutEdges{*Graph};
624 auto Nodes = std::make_unique<unsigned int[]>(num: Graph->nodes_size() +
625 1 /* terminator node */);
626 auto Edges = std::make_unique<unsigned int[]>(num: Graph->edges_size());
627 auto EdgeCuts = std::make_unique<int[]>(num: Graph->edges_size());
628 auto EdgeValues = std::make_unique<int[]>(num: Graph->edges_size());
629 for (const Node &N : Graph->nodes()) {
630 Nodes[Graph->getNodeIndex(N)] = Graph->getEdgeIndex(E: *N.edges_begin());
631 }
632 Nodes[Graph->nodes_size()] = Graph->edges_size(); // terminator node
633 for (const Edge &E : Graph->edges()) {
634 Edges[Graph->getEdgeIndex(E)] = Graph->getNodeIndex(N: *E.getDest());
635 EdgeValues[Graph->getEdgeIndex(E)] = E.getValue();
636 }
637 OptimizeCut(Nodes.get(), Graph->nodes_size(), Edges.get(), EdgeValues.get(),
638 EdgeCuts.get(), Graph->edges_size());
639 for (int I = 0; I < Graph->edges_size(); ++I)
640 if (EdgeCuts[I])
641 CutEdges.set(I);
642 LLVM_DEBUG(dbgs() << "Cutting edges... Done\n");
643 LLVM_DEBUG(dbgs() << "Cut " << CutEdges.count() << " edges\n");
644
645 LLVM_DEBUG(dbgs() << "Inserting LFENCEs...\n");
646 FencesInserted += insertFences(MF, G&: *Graph, CutEdges);
647 LLVM_DEBUG(dbgs() << "Inserting LFENCEs... Done\n");
648 LLVM_DEBUG(dbgs() << "Inserted " << FencesInserted << " fences\n");
649
650 Graph = GraphBuilder::trim(G: *Graph, TrimNodes: NodeSet{*Graph}, TrimEdges: CutEdges);
651 } while (true);
652
653 return FencesInserted;
654}
655
656int X86LoadValueInjectionLoadHardeningPass::hardenLoadsWithHeuristic(
657 MachineFunction &MF, std::unique_ptr<MachineGadgetGraph> Graph) const {
658 // If `MF` does not have any fences, then no gadgets would have been
659 // mitigated at this point.
660 if (Graph->NumFences > 0) {
661 LLVM_DEBUG(dbgs() << "Eliminating mitigated paths...\n");
662 Graph = trimMitigatedEdges(Graph: std::move(Graph));
663 LLVM_DEBUG(dbgs() << "Eliminating mitigated paths... Done\n");
664 }
665
666 if (Graph->NumGadgets == 0)
667 return 0;
668
669 LLVM_DEBUG(dbgs() << "Cutting edges...\n");
670 EdgeSet CutEdges{*Graph};
671
672 // Begin by collecting all ingress CFG edges for each node
673 DenseMap<const Node *, SmallVector<const Edge *, 2>> IngressEdgeMap;
674 for (const Edge &E : Graph->edges())
675 if (MachineGadgetGraph::isCFGEdge(E))
676 IngressEdgeMap[E.getDest()].push_back(Elt: &E);
677
678 // For each gadget edge, make cuts that guarantee the gadget will be
679 // mitigated. A computationally efficient way to achieve this is to either:
680 // (a) cut all egress CFG edges from the gadget source, or
681 // (b) cut all ingress CFG edges to the gadget sink.
682 //
683 // Moreover, the algorithm tries not to make a cut into a loop by preferring
684 // to make a (b)-type cut if the gadget source resides at a greater loop depth
685 // than the gadget sink, or an (a)-type cut otherwise.
686 for (const Node &N : Graph->nodes()) {
687 for (const Edge &E : N.edges()) {
688 if (!MachineGadgetGraph::isGadgetEdge(E))
689 continue;
690
691 SmallVector<const Edge *, 2> EgressEdges;
692 SmallVector<const Edge *, 2> &IngressEdges = IngressEdgeMap[E.getDest()];
693 for (const Edge &EgressEdge : N.edges())
694 if (MachineGadgetGraph::isCFGEdge(E: EgressEdge))
695 EgressEdges.push_back(Elt: &EgressEdge);
696
697 int EgressCutCost = 0, IngressCutCost = 0;
698 for (const Edge *EgressEdge : EgressEdges)
699 if (!CutEdges.contains(E: *EgressEdge))
700 EgressCutCost += EgressEdge->getValue();
701 for (const Edge *IngressEdge : IngressEdges)
702 if (!CutEdges.contains(E: *IngressEdge))
703 IngressCutCost += IngressEdge->getValue();
704
705 auto &EdgesToCut =
706 IngressCutCost < EgressCutCost ? IngressEdges : EgressEdges;
707 for (const Edge *E : EdgesToCut)
708 CutEdges.insert(E: *E);
709 }
710 }
711 LLVM_DEBUG(dbgs() << "Cutting edges... Done\n");
712 LLVM_DEBUG(dbgs() << "Cut " << CutEdges.count() << " edges\n");
713
714 LLVM_DEBUG(dbgs() << "Inserting LFENCEs...\n");
715 int FencesInserted = insertFences(MF, G&: *Graph, CutEdges);
716 LLVM_DEBUG(dbgs() << "Inserting LFENCEs... Done\n");
717 LLVM_DEBUG(dbgs() << "Inserted " << FencesInserted << " fences\n");
718
719 return FencesInserted;
720}
721
722int X86LoadValueInjectionLoadHardeningPass::insertFences(
723 MachineFunction &MF, MachineGadgetGraph &G,
724 EdgeSet &CutEdges /* in, out */) const {
725 int FencesInserted = 0;
726 for (const Node &N : G.nodes()) {
727 for (const Edge &E : N.edges()) {
728 if (CutEdges.contains(E)) {
729 MachineInstr *MI = N.getValue(), *Prev;
730 MachineBasicBlock *MBB; // Insert an LFENCE in this MBB
731 MachineBasicBlock::iterator InsertionPt; // ...at this point
732 if (MI == MachineGadgetGraph::ArgNodeSentinel) {
733 // insert LFENCE at beginning of entry block
734 MBB = &MF.front();
735 InsertionPt = MBB->begin();
736 Prev = nullptr;
737 } else if (MI->isBranch()) { // insert the LFENCE before the branch
738 MBB = MI->getParent();
739 InsertionPt = MI;
740 Prev = MI->getPrevNode();
741 // Remove all egress CFG edges from this branch because the inserted
742 // LFENCE prevents gadgets from crossing the branch.
743 for (const Edge &E : N.edges()) {
744 if (MachineGadgetGraph::isCFGEdge(E))
745 CutEdges.insert(E);
746 }
747 } else { // insert the LFENCE after the instruction
748 MBB = MI->getParent();
749 InsertionPt = MI->getNextNode() ? MI->getNextNode() : MBB->end();
750 Prev = InsertionPt == MBB->end()
751 ? (MBB->empty() ? nullptr : &MBB->back())
752 : InsertionPt->getPrevNode();
753 }
754 // Ensure this insertion is not redundant (two LFENCEs in sequence).
755 if ((InsertionPt == MBB->end() || !isFence(MI: &*InsertionPt)) &&
756 (!Prev || !isFence(MI: Prev))) {
757 BuildMI(BB&: *MBB, I: InsertionPt, MIMD: DebugLoc(), MCID: TII->get(Opcode: X86::LFENCE));
758 ++FencesInserted;
759 }
760 }
761 }
762 }
763 return FencesInserted;
764}
765
766bool X86LoadValueInjectionLoadHardeningPass::instrUsesRegToAccessMemory(
767 const MachineInstr &MI, unsigned Reg) const {
768 if (!MI.mayLoadOrStore() || MI.getOpcode() == X86::MFENCE ||
769 MI.getOpcode() == X86::SFENCE || MI.getOpcode() == X86::LFENCE)
770 return false;
771
772 const int MemRefBeginIdx = X86::getFirstAddrOperandIdx(MI);
773 if (MemRefBeginIdx < 0) {
774 LLVM_DEBUG(dbgs() << "Warning: unable to obtain memory operand for loading "
775 "instruction:\n";
776 MI.print(dbgs()); dbgs() << '\n';);
777 return false;
778 }
779
780 const MachineOperand &BaseMO =
781 MI.getOperand(i: MemRefBeginIdx + X86::AddrBaseReg);
782 const MachineOperand &IndexMO =
783 MI.getOperand(i: MemRefBeginIdx + X86::AddrIndexReg);
784 return (BaseMO.isReg() && BaseMO.getReg() != X86::NoRegister &&
785 TRI->regsOverlap(RegA: BaseMO.getReg(), RegB: Reg)) ||
786 (IndexMO.isReg() && IndexMO.getReg() != X86::NoRegister &&
787 TRI->regsOverlap(RegA: IndexMO.getReg(), RegB: Reg));
788}
789
790bool X86LoadValueInjectionLoadHardeningPass::instrUsesRegToBranch(
791 const MachineInstr &MI, unsigned Reg) const {
792 if (!MI.isConditionalBranch())
793 return false;
794 for (const MachineOperand &Use : MI.uses())
795 if (Use.isReg() && Use.getReg() == Reg)
796 return true;
797 return false;
798}
799
800INITIALIZE_PASS_BEGIN(X86LoadValueInjectionLoadHardeningPass, PASS_KEY,
801 "X86 LVI load hardening", false, false)
802INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass)
803INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
804INITIALIZE_PASS_DEPENDENCY(MachineDominanceFrontier)
805INITIALIZE_PASS_END(X86LoadValueInjectionLoadHardeningPass, PASS_KEY,
806 "X86 LVI load hardening", false, false)
807
808FunctionPass *llvm::createX86LoadValueInjectionLoadHardeningPass() {
809 return new X86LoadValueInjectionLoadHardeningPass();
810}
811