[X86] Enable ibt-seal optimization when LTO is used in Kernel

[joaomoreira]

Summary

Intel's CET/IBT requires every indirect branch target to be an ENDBR instruction. Because of that, the compiler needs to correctly emit these instruction on function's prologues. Because this is a security feature, it is desirable that only actual indirect-branch-targeted functions are emitted with ENDBRs. While it is possible to identify address-taken functions through LTO, minimizing these ENDBR instructions remains a hard task for user-space binaries because exported functions may end being reachable through PLT entries, that will use an indirect branch for such. Because this cannot be determined during compilation-time, the compiler currently emits ENDBRs to every non-local-linkage function.

Despite the challenge presented for user-space, the kernel landscape is different as no PLTs are used. With the intent of providing the most fit ENDBR emission for the kernel, kernel developers proposed an optimization named "ibt-seal" which replaces the ENDBRs for NOPs directly in the binary. The discussion of this feature can be seen in [1].

This diff brings the enablement of the flag -mibt-seal, which in combination with LTO enforces a different policy for ENDBR placement in when the code-model is set to "kernel". In this scenario, the compiler will only emit ENDBRs to address taken functions, ignoring non-address taken functions that are don't have local linkage.

A comparison between an LTO-compiled kernel binaries without and with the -mibt-seal feature enabled shows that when -mibt-seal was used, the number of ENDBRs in the vmlinux.o binary patched by objtool decreased from 44383 to 33192, and that the number of superfluous ENDBR instructions nopped-out decreased from 11730 to 540.

The 540 missed superfluous ENDBRs need to be investigated further, but hypotheses are: assembly code not being taken care of by the compiler, kernel exported symbols mechanisms creating bogus address taken situations or even these being removed due to other binary optimizations like kernel's static_calls. For now, I assume that the large drop in the number of ENDBR instructions already justifies the feature being merged.

[1] - https://lkml.org/lkml/2021/11/22/591

[joaomoreira]

Modularized needsPrologueENDBR function and removed missed comment.

[xiangzhangllvm]

Thanks for this optimization. The logic is clear on base of "the kernel landscape is different as no PLTs are used"

llvm/lib/Target/X86/X86IndirectBranchTracking.cpp

124

Not sure what the action if a function both has local and external call.
Can this be replace with "hasExternalLinkage" ?

[joaomoreira]

llvm/lib/Target/X86/X86IndirectBranchTracking.cpp

124

This is remaining from the same logic which was in place before this patch. Still, in my understanding, it cannot be replaced by "hasExternalLinkage" as static functions may have their address taken and still have local linkage.

[xiangzhangllvm]

LGTM, Let me accept later if no others opps

[joaomoreira]

Hi, is anything still preventing this from being merged?

[joaomoreira]

@pengfei or @craig.topper perhaps one of you could merge this? please, let me know if I'm missing any obvious requirement or detail.

[pengfei]

In D116070#3259639, @joaomoreira wrote:

@pengfei or @craig.topper perhaps one of you could merge this? please, let me know if I'm missing any obvious requirement or detail.

Sure, I can do it.

[MaskRay]

The module flags metadata usage "Override" was wrong. I fixed it by using "Min".

[MaskRay]

I have a question why the option is needed.

  if (IBTSeal) {
    return F.hasAddressTaken();
  }
  // if !IBTSeal, fall into default case.
  LLVM_FALLTHROUGH;
// Address taken or externally linked functions may be reachable.
default:
  return (F.hasAddressTaken() || !F.hasLocalLinkage());

• If the link mixes bitcode files and ELF relocatable files, for a function in a bitcode file, F.hasAddressTaken() doesn't indicate that its address is not taken by an ELF relocatable file.
• For ThinLTO, a function may have false F.hasAddressTaken() for the definition in one module and true F.hasAddressTaken() for a reference in another module.

In both cases, return F.hasAddressTaken(); isn't sufficient.

If we don't consider these issue: I think -mibt-seal can be removed if we merge unnamed_addr/local_unnamed_addr for ThinLTO (but we don't). If we do,
F.hasAddressTaken() || (F.hasHiddenVisibility() && !F.hasLocalLinkage()) can be used as a proxy.

[MaskRay]

(I substantially edited my previous comment.) It now says this:

If the link mixes bitcode files and ELF relocatable files, for a function in a bitcode file, F.hasAddressTaken() doesn't indicate that its address is not taken by an ELF relocatable file.

[MaskRay]

Sent D140363 to remove this incorrectly implemented option.

llvm/test/CodeGen/X86/ibtseal-small.ll

1

It's not necessary to create 3 identical test files. We can use 3 RUN lines instead of using the DejaGnu style approach:)

clang/include/clang/Basic/CodeGenOptions.def

 CODEGENOPT(InstrumentFunctionsAfterInlining , 1, 0) ///< Set when
                           ///< -finstrument-functions-after-inlining is enabled.
 CODEGENOPT(InstrumentFunctionEntryBare , 1, 0) ///< Set when
                                ///< -finstrument-function-entry-bare is enabled.
 CODEGENOPT(CFProtectionReturn , 1, 0) ///< if -fcf-protection is
                                       ///< set to full or return.
 CODEGENOPT(CFProtectionBranch , 1, 0) ///< if -fcf-protection is
                                       ///< set to full or branch.
+CODEGENOPT(IBTSeal, 1, 0)             ///< set to optimize CFProtectionBranch.
+
 CODEGENOPT(XRayInstrumentFunctions , 1, 0) ///< Set when -fxray-instrument is
                                            ///< enabled.
 CODEGENOPT(StackSizeSection  , 1, 0) ///< Set when -fstack-size-section is enabled.
 CODEGENOPT(ForceDwarfFrameSection , 1, 0) ///< Set when -fforce-dwarf-frame is
                                           ///< enabled.
 
 ///< Set when -fxray-always-emit-customevents is enabled.
 CODEGENOPT(XRayAlwaysEmitCustomEvents , 1, 0)

clang/include/clang/Driver/Options.td

 def finstrument_function_entry_bare : Flag<["-"], "finstrument-function-entry-bare">, Group<f_Group>, Flags<[CC1Option]>,
   HelpText<"Instrument function entry only, after inlining, without arguments to the instrumentation call">,
   MarshallingInfoFlag<CodeGenOpts<"InstrumentFunctionEntryBare">>;
 def fcf_protection_EQ : Joined<["-"], "fcf-protection=">, Flags<[CoreOption, CC1Option]>, Group<f_Group>,
   HelpText<"Instrument control-flow architecture protection. Options: return, branch, full, none.">, Values<"return,branch,full,none">;
 def fcf_protection : Flag<["-"], "fcf-protection">, Group<f_Group>, Flags<[CoreOption, CC1Option]>,
   Alias<fcf_protection_EQ>, AliasArgs<["full"]>,
   HelpText<"Enable cf-protection in 'full' mode">;
+def mibt_seal : Flag<["-"], "mibt-seal">, Group<m_Group>, Flags<[CoreOption, CC1Option]>,
+  HelpText<"Optimize fcf-protection=branch/full (requires LTO).">;
 
 defm xray_instrument : BoolFOption<"xray-instrument",
   LangOpts<"XRayInstrument">, DefaultFalse,
   PosFlag<SetTrue, [CC1Option], "Generate XRay instrumentation sleds on function entry and exit">,
   NegFlag<SetFalse>>;
 
 def fxray_instruction_threshold_EQ :
   JoinedOrSeparate<["-"], "fxray-instruction-threshold=">,

clang/lib/CodeGen/CodeGenModule.cpp

void CodeGenModule::Release() {
 
   if (CodeGenOpts.CFProtectionBranch &&
       Target.checkCFProtectionBranchSupported(getDiags())) {
     // Indicate that we want to instrument branch control flow protection.
     getModule().addModuleFlag(llvm::Module::Override, "cf-protection-branch",
                               1);
   }
 
+  if (CodeGenOpts.IBTSeal)
+    getModule().addModuleFlag(llvm::Module::Override, "ibt-seal", 1);
+
   // Add module metadata for return address signing (ignoring
   // non-leaf/all) and stack tagging. These are actually turned on by function
   // attributes, but we use module metadata to emit build attributes. This is
   // needed for LTO, where the function attributes are inside bitcode
   // serialised into a global variable by the time build attributes are
   // emitted, so we can't access them.
   if (Context.getTargetInfo().hasFeature("ptrauth") &&
       LangOpts.getSignReturnAddressScope() !=

clang/lib/Driver/ToolChains/Clang.cpp

if (IsCuda || IsHIP) {
     }
   }
 
   if (Arg *A = Args.getLastArg(options::OPT_fcf_protection_EQ)) {
     CmdArgs.push_back(
         Args.MakeArgString(Twine("-fcf-protection=") + A->getValue()));
   }
 
+  if (IsUsingLTO)
+    Args.AddLastArg(CmdArgs, options::OPT_mibt_seal);
+
   // Forward -f options with positive and negative forms; we translate these by
   // hand.  Do not propagate PGO options to the GPU-side compilations as the
   // profile info is for the host-side compilation only.
   if (!(IsCudaDevice || IsHIPDevice)) {
     if (Arg *A = getLastProfileSampleUseArg(Args)) {
       auto *PGOArg = Args.getLastArg(
           options::OPT_fprofile_generate, options::OPT_fprofile_generate_EQ,
           options::OPT_fcs_profile_generate,

clang/lib/Frontend/CompilerInvocation.cpp

if (const Arg *A = Args.getLastArg(OPT_fcf_protection_EQ)) {
     } else if (Name == "return")
       Opts.CFProtectionReturn = 1;
     else if (Name == "branch")
       Opts.CFProtectionBranch = 1;
     else if (Name != "none")
       Diags.Report(diag::err_drv_invalid_value) << A->getAsString(Args) << Name;
   }
 
+  if (Opts.PrepareForLTO && Args.hasArg(OPT_mibt_seal))
+    Opts.IBTSeal = 1;
+
   for (auto *A :
        Args.filtered(OPT_mlink_bitcode_file, OPT_mlink_builtin_bitcode)) {
     CodeGenOptions::BitcodeFileToLink F;
     F.Filename = A->getValue();
     if (A->getOption().matches(OPT_mlink_builtin_bitcode)) {
       F.LinkFlags = llvm::Linker::Flags::LinkOnlyNeeded;
       // When linking CUDA bitcode, propagate function attributes so that
       // e.g. libdevice gets fast-math attrs if we're building with fast-math.

llvm/lib/Target/X86/X86IndirectBranchTracking.cpp

if (!MOp.isGlobal())
     return false;
   auto *CalleeFn = dyn_cast<Function>(MOp.getGlobal());
   if (!CalleeFn)
     return false;
   AttributeList Attrs = CalleeFn->getAttributes();
   return Attrs.hasFnAttr(Attribute::ReturnsTwice);
 }
 
+// Checks if function should have an ENDBR in its prologue
+static bool needsPrologueENDBR(MachineFunction &MF, const Module *M) {
+  Function &F = MF.getFunction();
+
+  if (F.doesNoCfCheck())
+    return false;
+
+  const X86TargetMachine *TM =
+      static_cast<const X86TargetMachine *>(&MF.getTarget());
+  Metadata *IBTSeal = M->getModuleFlag("ibt-seal");
+
+  switch (TM->getCodeModel()) {
+  // Large code model functions always reachable through indirect calls.
+  case CodeModel::Large:
+    return true;
+  // Only address taken functions in LTO'ed kernel are reachable indirectly.
+  // IBTSeal implies LTO, thus only check if function is address taken.
+  case CodeModel::Kernel:
+    // Check if ibt-seal was enabled (implies LTO is being used).
+    if (IBTSeal) {
+      return F.hasAddressTaken();
+    }
+    // if !IBTSeal, fall into default case.
+    LLVM_FALLTHROUGH;
+  // Address taken or externally linked functions may be reachable.
+  default:
+    return (F.hasAddressTaken() || !F.hasLocalLinkage());
+  }
+}
+
 bool X86IndirectBranchTrackingPass::runOnMachineFunction(MachineFunction &MF) {
   const X86Subtarget &SubTarget = MF.getSubtarget<X86Subtarget>();
 
+  const Module *M = MF.getMMI().getModule();
   // Check that the cf-protection-branch is enabled.
-  Metadata *isCFProtectionSupported =
-      MF.getMMI().getModule()->getModuleFlag("cf-protection-branch");
-  // NB: We need to enable IBT in jitted code if JIT compiler is CET
-  // enabled.
+  Metadata *isCFProtectionSupported = M->getModuleFlag("cf-protection-branch");
+
+  //  NB: We need to enable IBT in jitted code if JIT compiler is CET
+  //  enabled.
   const X86TargetMachine *TM =
       static_cast<const X86TargetMachine *>(&MF.getTarget());
 #ifdef __CET__
   bool isJITwithCET = TM->isJIT();
 #else
   bool isJITwithCET = false;
 #endif
   if (!isCFProtectionSupported && !IndirectBranchTracking && !isJITwithCET)
     return false;
 
   // True if the current MF was changed and false otherwise.
   bool Changed = false;
 
   TII = SubTarget.getInstrInfo();
   EndbrOpcode = SubTarget.is64Bit() ? X86::ENDBR64 : X86::ENDBR32;
 
-  // Large code model, non-internal function or function whose address
-  // was taken, can be accessed through indirect calls. Mark the first
-  // BB with ENDBR instruction unless nocf_check attribute is used.
-  if ((TM->getCodeModel() == CodeModel::Large ||
-       MF.getFunction().hasAddressTaken() ||
-       !MF.getFunction().hasLocalLinkage()) &&
-      !MF.getFunction().doesNoCfCheck()) {
+  // If function is reachable indirectly, mark the first BB with ENDBR.
+  if (needsPrologueENDBR(MF, M)) {
     auto MBB = MF.begin();
     Changed |= addENDBR(*MBB, MBB->begin());
   }
 
   for (auto &MBB : MF) {
     // Find all basic blocks that their address was taken (for example
     // in the case of indirect jump) and add ENDBR instruction.
     if (MBB.hasAddressTaken())

llvm/test/CodeGen/X86/ibtseal-kernel.ll

+; RUN: llc < %s -O2 -mtriple=x86_64-unknown-linux-gnu -x86-indirect-branch-tracking --code-model=kernel | FileCheck %s --check-prefix=CHECK-KERNEL-IBTSEAL
+
+; CHECK-KERNEL-IBTSEAL: foo:
+; CHECK-KERNEL-IBTSEAL: endbr
+; CHECK-KERNEL-IBTSEAL: bar:
+; CHECK-KERNEL-IBTSEAL-NOT: endbr
+
+target triple = "x86_64-unknown-linux-gnu"
+
+define dso_local void @foo() {
+  ret void
+}
+
+define dso_local i8* @bar() {
+  ret i8* bitcast (void ()* @foo to i8*)
+}
+
+!llvm.module.flags = !{!1}
+!1 = !{i32 4, !"ibt-seal", i32 1}

llvm/test/CodeGen/X86/ibtseal-large.ll

+; RUN: llc < %s -O2 -mtriple=x86_64-unknown-linux-gnu -x86-indirect-branch-tracking --code-model=large | FileCheck %s --check-prefix=CHECK-LARGE-IBTSEAL
+
+; CHECK-LARGE-IBTSEAL: foo:
+; CHECK-LARGE-IBTSEAL: endbr
+; CHECK-LARGE-IBTSEAL: bar:
+; CHECK-LARGE-IBTSEAL: endbr
+
+target triple = "x86_64-unknown-linux-gnu"
+
+define dso_local void @foo() {
+  ret void
+}
+
+define dso_local i8* @bar() {
+  ret i8* bitcast (void ()* @foo to i8*)
+}
+
+!llvm.module.flags = !{!1}
+!1 = !{i32 4, !"ibt-seal", i32 1}

llvm/test/CodeGen/X86/ibtseal-small.ll

+; RUN: llc < %s -O2 -mtriple=x86_64-unknown-linux-gnu -x86-indirect-branch-tracking --code-model=small | FileCheck %s --check-prefix=CHECK-SMALL-IBTSEAL
+
+; CHECK-SMALL-IBTSEAL: foo:
+; CHECK-SMALL-IBTSEAL: endbr
+; CHECK-SMALL-IBTSEAL: bar:
+; CHECK-SMALL-IBTSEAL: endbr
+
+target triple = "x86_64-unknown-linux-gnu"
+
+define dso_local void @foo() {
+  ret void
+}
+
+define dso_local i8* @bar() {
+  ret i8* bitcast (void ()* @foo to i8*)
+}
+
+!llvm.module.flags = !{!1}
+!1 = !{i32 4, !"ibt-seal", i32 1}