[cfi-verify] Support cross-DSO
authorJoel Galenson <jgalenson@google.com>
Fri, 24 Aug 2018 15:21:58 +0000 (15:21 +0000)
committerJoel Galenson <jgalenson@google.com>
Fri, 24 Aug 2018 15:21:58 +0000 (15:21 +0000)
When used in cross-DSO mode, CFI will generate calls to special functions rather than trap instructions.  For example, instead of generating

if (!InlinedFastCheck(f))
  abort();
call *f

CFI generates

if (!InlinedFastCheck(f))
  __cfi_slowpath(CallSiteTypeId, f);
call *f

This patch teaches cfi-verify to recognize calls to __cfi_slowpath and abort and treat them as trap functions.

In addition to normal symbols, we also parse the dynamic relocations to handle cross-DSO calls in libraries.

We also extend cfi-verify to recognize other patterns that occur using cross-DSO.  For example, some indirect calls are not guarded by a branch to a trap but instead follow a call to __cfi_slowpath.  For example:

if (!InlinedFastCheck(f))
  call *f
else {
  __cfi_slowpath(CallSiteTypeId, f);
  call *f
}

In this case, the second call to f is not marked as protected by the current code.  We thus recognize if indirect calls directly follow a call to a function that will trap on CFI violations and treat them as protected.

We also ignore indirect calls in the PLT, since on AArch64 each entry contains an indirect call that should not be protected by CFI, and these are labeled incorrectly when debug information is not present.

Differential Revision: https://reviews.llvm.org/D49383

llvm-svn: 340612

llvm/test/tools/llvm-cfi-verify/AArch64/Inputs/function-only-check.o [new file with mode: 0644]
llvm/test/tools/llvm-cfi-verify/AArch64/function-only-check.s [new file with mode: 0644]
llvm/test/tools/llvm-cfi-verify/X86/Inputs/function-only-check.o [new file with mode: 0644]
llvm/test/tools/llvm-cfi-verify/X86/function-only-check.s [new file with mode: 0644]
llvm/tools/llvm-cfi-verify/lib/FileAnalysis.cpp
llvm/tools/llvm-cfi-verify/lib/FileAnalysis.h
llvm/tools/llvm-cfi-verify/lib/GraphBuilder.cpp

diff --git a/llvm/test/tools/llvm-cfi-verify/AArch64/Inputs/function-only-check.o b/llvm/test/tools/llvm-cfi-verify/AArch64/Inputs/function-only-check.o
new file mode 100644 (file)
index 0000000..49142ff
Binary files /dev/null and b/llvm/test/tools/llvm-cfi-verify/AArch64/Inputs/function-only-check.o differ
diff --git a/llvm/test/tools/llvm-cfi-verify/AArch64/function-only-check.s b/llvm/test/tools/llvm-cfi-verify/AArch64/function-only-check.s
new file mode 100644 (file)
index 0000000..30460c8
--- /dev/null
@@ -0,0 +1,9 @@
+# RUN: llvm-cfi-verify -search-length-undef=6 %S/Inputs/function-only-check.o | FileCheck %s
+
+# CHECK-LABEL: {{^Instruction: .* \(PROTECTED\)}}
+# CHECK-NEXT: tiny.cc:9
+
+# CHECK: Expected Protected: 1 (100.00%)
+# CHECK: Unexpected Protected: 0 (0.00%)
+# CHECK: Expected Unprotected: 0 (0.00%)
+# CHECK: Unexpected Unprotected (BAD): 0 (0.00%)
diff --git a/llvm/test/tools/llvm-cfi-verify/X86/Inputs/function-only-check.o b/llvm/test/tools/llvm-cfi-verify/X86/Inputs/function-only-check.o
new file mode 100644 (file)
index 0000000..099076a
Binary files /dev/null and b/llvm/test/tools/llvm-cfi-verify/X86/Inputs/function-only-check.o differ
diff --git a/llvm/test/tools/llvm-cfi-verify/X86/function-only-check.s b/llvm/test/tools/llvm-cfi-verify/X86/function-only-check.s
new file mode 100644 (file)
index 0000000..250f5dc
--- /dev/null
@@ -0,0 +1,8 @@
+# RUN: llvm-cfi-verify %S/Inputs/function-only-check.o | FileCheck %s
+
+# CHECK-LABEL: {{^Instruction: .* \(PROTECTED\)}}
+
+# CHECK: Expected Protected: 1 (100.00%)
+# CHECK: Unexpected Protected: 0 (0.00%)
+# CHECK: Expected Unprotected: 0 (0.00%)
+# CHECK: Unexpected Unprotected (BAD): 0 (0.00%)
index 29819d8..8dc0802 100644 (file)
@@ -105,6 +105,9 @@ Expected<FileAnalysis> FileAnalysis::Create(StringRef Filename) {
   if (auto SectionParseResponse = Analysis.parseCodeSections())
     return std::move(SectionParseResponse);
 
+  if (auto SymbolTableParseResponse = Analysis.parseSymbolTable())
+    return std::move(SymbolTableParseResponse);
+
   return std::move(Analysis);
 }
 
@@ -165,7 +168,18 @@ const Instr &FileAnalysis::getInstructionOrDie(uint64_t Address) const {
 
 bool FileAnalysis::isCFITrap(const Instr &InstrMeta) const {
   const auto &InstrDesc = MII->get(InstrMeta.Instruction.getOpcode());
-  return InstrDesc.isTrap();
+  return InstrDesc.isTrap() || willTrapOnCFIViolation(InstrMeta);
+}
+
+bool FileAnalysis::willTrapOnCFIViolation(const Instr &InstrMeta) const {
+  const auto &InstrDesc = MII->get(InstrMeta.Instruction.getOpcode());
+  if (!InstrDesc.isCall())
+    return false;
+  uint64_t Target;
+  if (!MIA->evaluateBranch(InstrMeta.Instruction, InstrMeta.VMAddress,
+                           InstrMeta.InstructionSize, Target))
+    return false;
+  return TrapOnFailFunctionAddresses.count(Target) > 0;
 }
 
 bool FileAnalysis::canFallThrough(const Instr &InstrMeta) const {
@@ -431,6 +445,12 @@ Error FileAnalysis::parseCodeSections() {
     if (!(object::ELFSectionRef(Section).getFlags() & ELF::SHF_EXECINSTR))
       continue;
 
+    // Avoid checking the PLT since it produces spurious failures on AArch64
+    // when ignoring DWARF data.
+    StringRef SectionName;
+    if (!Section.getName(SectionName) && SectionName == ".plt")
+      continue;
+
     StringRef SectionContents;
     if (Section.getContents(SectionContents))
       return make_error<StringError>("Failed to retrieve section contents",
@@ -518,6 +538,40 @@ void FileAnalysis::addInstruction(const Instr &Instruction) {
   }
 }
 
+Error FileAnalysis::parseSymbolTable() {
+  // Functions that will trap on CFI violations.
+  SmallSet<StringRef, 4> TrapOnFailFunctions;
+  TrapOnFailFunctions.insert("__cfi_slowpath");
+  TrapOnFailFunctions.insert("__cfi_slowpath_diag");
+  TrapOnFailFunctions.insert("abort");
+
+  // Look through the list of symbols for functions that will trap on CFI
+  // violations.
+  for (auto &Sym : Object->symbols()) {
+    auto SymNameOrErr = Sym.getName();
+    if (!SymNameOrErr)
+      consumeError(SymNameOrErr.takeError());
+    else if (TrapOnFailFunctions.count(*SymNameOrErr) > 0) {
+      auto AddrOrErr = Sym.getAddress();
+      if (!AddrOrErr)
+        consumeError(AddrOrErr.takeError());
+      else
+        TrapOnFailFunctionAddresses.insert(*AddrOrErr);
+    }
+  }
+  if (auto *ElfObject = dyn_cast<object::ELFObjectFileBase>(Object)) {
+    for (const auto &Addr : ElfObject->getPltAddresses()) {
+      object::SymbolRef Sym(Addr.first, Object);
+      auto SymNameOrErr = Sym.getName();
+      if (!SymNameOrErr)
+        consumeError(SymNameOrErr.takeError());
+      else if (TrapOnFailFunctions.count(*SymNameOrErr) > 0)
+        TrapOnFailFunctionAddresses.insert(Addr.second);
+    }
+  }
+  return Error::success();
+}
+
 UnsupportedDisassembly::UnsupportedDisassembly(StringRef Text) : Text(Text) {}
 
 char UnsupportedDisassembly::ID;
index 3f0a704..67421e8 100644 (file)
@@ -11,6 +11,7 @@
 #define LLVM_CFI_VERIFY_FILE_ANALYSIS_H
 
 #include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallSet.h"
 #include "llvm/BinaryFormat/ELF.h"
 #include "llvm/DebugInfo/Symbolize/Symbolize.h"
 #include "llvm/MC/MCAsmInfo.h"
@@ -110,6 +111,10 @@ public:
   // Returns whether this instruction is used by CFI to trap the program.
   bool isCFITrap(const Instr &InstrMeta) const;
 
+  // Returns whether this instruction is a call to a function that will trap on
+  // CFI violations (i.e., it serves as a trap in this instance).
+  bool willTrapOnCFIViolation(const Instr &InstrMeta) const;
+
   // Returns whether this function can fall through to the next instruction.
   // Undefined (and bad) instructions cannot fall through, and instruction that
   // modify the control flow can only fall through if they are conditional
@@ -183,6 +188,10 @@ protected:
   // internal members. Should only be called once by Create().
   Error parseCodeSections();
 
+  // Parses the symbol table to look for the addresses of functions that will
+  // trap on CFI violations.
+  Error parseSymbolTable();
+
 private:
   // Members that describe the input file.
   object::OwningBinary<object::Binary> Binary;
@@ -218,6 +227,9 @@ private:
 
   // A list of addresses of indirect control flow instructions.
   std::set<uint64_t> IndirectInstructions;
+
+  // The addresses of functions that will trap on CFI violations.
+  SmallSet<uint64_t, 4> TrapOnFailFunctionAddresses;
 };
 
 class UnsupportedDisassembly : public ErrorInfo<UnsupportedDisassembly> {
index 4153b5f..5b2bc6f 100644 (file)
@@ -311,6 +311,24 @@ void GraphBuilder::buildFlowGraphImpl(const FileAnalysis &Analysis,
     Result.ConditionalBranchNodes.push_back(BranchNode);
   }
 
+  // When using cross-DSO, some indirect calls are not guarded by a branch to a
+  // trap but instead follow a call to __cfi_slowpath.  For example:
+  // if (!InlinedFastCheck(f))
+  //    call *f
+  //  else {
+  //    __cfi_slowpath(CallSiteTypeId, f);
+  //    call *f
+  //  }
+  // To mark the second call as protected, we recognize indirect calls that
+  // directly follow calls to functions that will trap on CFI violations.
+  if (CFCrossRefs.empty()) {
+    const Instr *PrevInstr = Analysis.getPrevInstructionSequential(ChildMeta);
+    if (PrevInstr && Analysis.willTrapOnCFIViolation(*PrevInstr)) {
+      Result.IntermediateNodes[PrevInstr->VMAddress] = Address;
+      HasValidCrossRef = true;
+    }
+  }
+
   if (!HasValidCrossRef)
     Result.OrphanedNodes.push_back(Address);