Re: [PATCH 2/2] x86/retpoline,kprobes: Avoid treating rethunk as an indirect jump

[Google]

On Wed, 5 Jul 2023 16:50:17 +0200
Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:

> On Wed, Jul 05, 2023 at 11:20:38PM +0900, Masami Hiramatsu wrote:
> > On Wed, 5 Jul 2023 10:58:57 +0200
> > Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > 
> > > On Wed, Jul 05, 2023 at 10:15:47AM +0200, Petr Pavlu wrote:
> > > > Functions can_optimize() and insn_is_indirect_jump() consider jumps to
> > > > the range [__indirect_thunk_start, __indirect_thunk_end] as indirect
> > > > jumps and prevent use of optprobes in functions containing them.
> > > 
> > > Why ?!? I mean, doing an opt-probe of an indirect jump/call instruction
> > > itself doesn't really make sense and I can see why you'd want to not do
> > > that. But why disallow an opt-probe if there's one in the function as a
> > > whole, but not the probe target?
> > 
> > Here we need to clarify the reason why functions which have indirect jumps
> > are not allowed to use opt-probe. Since optprobe can replace multiple 
> > instructions with a jump, if any jmp (is used for jump inside same function)
> > jumps to the second and subsequent instructions replaced by optprobe's jump,
> > that target instruction can not be optimized.
> > 
> > The problem of indirect jump (which jumps to the same function) is that
> > we don't know which addresses will be the target of the indirect jump.
> > So, for safety, I disallow optprobe for such function. In that case, normal
> > kprobe is used because it replaces only one instruction.
> 
> Ah, you're worried about jump-tables; you don't want to optimize across
> a jump-table target because then things go *boom*.
> 
> There's two things:
> 
>  - when X86_KERNEL_IBT=y any indirect jump target should be an ENDBR
>    instruction, so jump-table targets can be easily detected.
> 
>  - when RETPOLINE=y || X86_KERNEL_IBT=y we have jump-tables disabled,
>    search for -fno-jump-table in arch/x86/Makefile.
> 
> At some point in the future we should be able to allow jump-tables for
> RETPOLINE=n && IBT=y builds (provided the compilers behave), but we
> currently don't bother to find out.
> 
> Therefore, when either CONFIG option is found, you can assume that any
> indirect jump will be to another function.

OK, I confirmed that '-fno-jump-tables' is set when X86_KERNEL_IBT=y || RETPOLINE=y
so we can skip this indirect jump check. That makes things simpler.

> 
> > If I understand correctly, all indirect jump will be replaced with JMP_NOSPEC.
> > If you read the insn_jump_into_range, I onlu jecks the jump code, not call.
> > So the functions only have indirect call still allow optprobe.
> 
> With the introduction of kCFI JMP_NOSPEC is no longer an equivalent to a
> C indirect jump.

If I understand correctly, kCFI is enabled by CFI_CLANG, and clang is not
using jump-tables by default, so we can focus on gcc. In that case
current check still work, correct?

Thank you,

-- 
Masami Hiramatsu (Google) <mhiramat@xxxxxxxxxx>