Re: [PATCH v2 2/7] kcfi: Add core Kernel Control Flow Integrity infrastructure

[Kees Cook <kees@kernel.org>]

On Mon, Sep 08, 2025 at 05:32:58PM +0200, Peter Zijlstra wrote:
> On Fri, Sep 05, 2025 at 09:19:29AM -0700, Kees Cook wrote:
> > On Fri, Sep 05, 2025 at 10:51:03AM +0200, Peter Zijlstra wrote:
> > > On Thu, Sep 04, 2025 at 05:24:10PM -0700, Kees Cook wrote:
> > > > +- The check-call instruction sequence must be treated a single unit: it
> > > > +  cannot be rearranged or split or optimized. The pattern is that
> > > > +  indirect calls, "call *$target", get converted into:
> > > > +
> > > > +    mov $target_expression, %target ; only present if the expression was
> > > > +                                    ; not already %target register
> > > > +    load -$offset(%target), %tmp    ; load the typeid hash at target
> > > > +    cmp $hash, %tmp                 ; compare expected typeid with loaded
> > > > +    je .Lcheck_passed               ; jump to the indirect call
> > > > +  .Lkcfi_trap$N:                    ; label of trap insn
> > > > +    trap                            ; trap on failure, but arranged so
> > > > +                                    ; "permissive mode" falls through
> > > > +  .Lkcfi_call$N:                    ; label of call insn
> > > > +    call *%target                   ; actual indirect call
> > > > +
> > > > +  This pattern of call immediately after trap provides for the
> > > > +  "permissive" checking mode automatically: the trap gets handled,
> > > > +  a warning emitted, and then execution continues after the trap to
> > > > +  the call.
> > > 
> > > I know it is far too late to do anything here. But I've recently dug
> > > through a bunch of optimization manual and the like and that Jcc is
> > > about as bad as it gets :/
> > > 
> > > The old optimization manual states that forward jumps are assumed
> > > not-taken; while backward jumps are assumed taken.
> > > 
> > > The new wisdom is that any Jcc must be assumed not-taken; that is, the
> > > fallthrough case has the best throughput.
> > 
> > I would expect the cmp to be the slowest part of this sequence, and I
> > figured the both the trap and the call to be speculation barriers? I'm
> > not sure, though. Is changing the sequence actually useful?
> 
> The load can miss, in which case it is definitely the most expensive
> thing around.
> 
> > > Here we have a forward branch which is assumed taken :-(
> > 
> > The constraints we have are:
> > 
> > - Linux x86 KCFI trap handler decodes the instructions from the trap
> >   backwards, but it uses exact offsets (-12 and -6).
> > - Control flow following the trap must make the call (for warn-only mode)
> > 
> > If we change this, we'd need to make the insn decoder smarter to likey
> > look at the insn AFTER the trap ("is it a direct jump?")
> > 
> > And then use this, which is ugly, but matches second constraint:
> > 
> > 	cmp $hash %tmp
> > 	jne .Ltrap
> > .Lcall:
> > 	call *%target
> > 	jmp .Ldone
> > .Ltrap:
> > 	ud2
> > 	jmp .Lcall
> > .Ldone:
> 
> Ah, you can do something like:
> 
> 	cmp $hash, %tmp
> 	jne +3
> 	nopl -42(%rax)
> 	call *%target
> 
> which is only 2 bytes longer. Notably, that nopl is 4 bytes and the 4th
> byte is 0xd6 (aka UDB). This is an effective UDcc instruction based
> around a forward non-taken branch.

Oh right, I forgot about the nop encodings.

> But yeah, I don't know if it is worth changing this. Its just that I've
> been staring at these things far too much of late :-)

To do this we'd need to change the Linux trap handler and Clang's
implementation, so yeah, I'm inclined to just leave it as-is until we
have a stronger reason to change it.

-- 
Kees Cook