Re: [PATCH 4/9] x86/alternative: Implement .retpoline_sites support

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 13/10/2021 16:12, Peter Zijlstra wrote:
> On Wed, Oct 13, 2021 at 03:38:27PM +0100, Andrew Cooper wrote:
>> On 13/10/2021 13:22, Peter Zijlstra wrote:
>>> +/*
>>> + * Rewrite the compiler generated retpoline thunk calls.
>>> + *
>>> + * For spectre_v2=off (!X86_FEATURE_RETPOLINE), rewrite them into immediate
>>> + * indirect instructions, avoiding the extra indirection.
>>> + *
>>> + * For example, convert:
>>> + *
>>> + *   CALL __x86_indirect_thunk_\reg
>>> + *
>>> + * into:
>>> + *
>>> + *   CALL *%\reg
>>> + *
>>> + */
>>> +static int patch_retpoline(void *addr, struct insn *insn, u8 *bytes)
>>> +{
>>> +	void (*target)(void);
>>> +	int reg, i = 0;
>>> +
>>> +	if (cpu_feature_enabled(X86_FEATURE_RETPOLINE))
>>> +		return -1;
>>> +
>>> +	target = addr + insn->length + insn->immediate.value;
>>> +	reg = (target - &__x86_indirect_thunk_rax) /
>>> +	      (&__x86_indirect_thunk_rcx - &__x86_indirect_thunk_rax);
>> This is equal measures beautiful and terrifying.
> Thanks! :-)
>
>> Something around here really wants to BUG_ON(reg == 4), because
>> literally nothing good can come from selecting %rsp.
> Ack, I had to add rsp to get the offsets right, but indeed, if anything
> ever selects that we're in trouble.

Actually, all you need is space for the RSP thunk, not an actual RSP
thunk, and it's probably a wise move not to write one out.

You can fill it with 0xcc's, and make sure not to make it an exported
symbol.

>
>> Also, it might be a good idea (previous patch perhaps) to have some
>> linker assertions to confirm that the symbols are laid out safely to do
>> this calculation.
> I was hoping that since all this is in .S it would be immune from crazy
> things like a compiler and do as told. But I suppose carzy stuff like
> LTO (or worse BOLT) can totaly wreck this still (then BOLT won't care
> about linker script assertions either).
>
> I'll see if I can come up with something.

Another cross check could be something like:

unsigned long reg_to_thunk[] = {
    &__x86_indirec_thunk_rax,
    ...
};

because then BUG_ON(target != reg_to_thunk[reg]) will catch any errors
from layout issues.

Using 0 for rsp could then subsume the individual check.

~Andrew