Re: [PATCH 3/3] x86/alternative: Patch a single alternative location only once

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Sep 29, 2025 at 01:29:47PM +0200, Juergen Gross wrote:
> Instead of patching a single location potentially multiple times in
> case of nested ALTERNATIVE()s, do the patching only after having
> evaluated all alt_instr instances for that location.
> 
> This has multiple advantages:
> 
> - In case of replacing an indirect with a direct call using the
>   ALT_FLAG_DIRECT_CALL flag, there is no longer the need to have that
>   instance before any other instances at the same location (the
>   original instruction is needed for finding the target of the direct
>   call).
> 
> - In case of nested ALTERNATIVE()s there is no intermediate replacement
>   visible. This avoids any problems in case e.g. an interrupt is
>   happening between the single instances and the patched location is
>   used during handling the interrupt.
> 
> Signed-off-by: Juergen Gross <jgross@suse.com>
> ---
> V2:
> - new patch
> ---
>  arch/x86/kernel/alternative.c | 27 +++++++++++++++++++--------
>  1 file changed, 19 insertions(+), 8 deletions(-)
> 
> diff --git a/arch/x86/kernel/alternative.c b/arch/x86/kernel/alternative.c
> index 735cc017f2d3..ccf07131cd47 100644
> --- a/arch/x86/kernel/alternative.c
> +++ b/arch/x86/kernel/alternative.c
> @@ -648,6 +648,8 @@ void __init_or_module noinline apply_alternatives(struct alt_instr *start,
>  	u8 insn_buff[MAX_PATCH_LEN];
>  	u8 *instr;
>  	struct alt_instr *a, *b;
> +	unsigned int instances = 0;
> +	bool patched = false;
>  
>  	DPRINTK(ALT, "alt table %px, -> %px", start, end);
>  
> @@ -677,9 +679,13 @@ void __init_or_module noinline apply_alternatives(struct alt_instr *start,
>  		 * padding for all alt_instr entries for this site (nested
>  		 * alternatives result in consecutive entries).
>  		 */
> -		for (b = a+1; b < end && instr_va(b) == instr_va(a); b++) {
> -			u8 len = max(a->instrlen, b->instrlen);
> -			a->instrlen = b->instrlen = len;
> +		if (!instances) {
> +			for (b = a+1; b < end && instr_va(b) == instr_va(a); b++) {
> +				u8 len = max(a->instrlen, b->instrlen);
> +				a->instrlen = b->instrlen = len;
> +			}
> +			instances = b - a;
> +			patched = false;
>  		}
>  
>  		instr = instr_va(a);
> @@ -692,14 +698,19 @@ void __init_or_module noinline apply_alternatives(struct alt_instr *start,
>  		 * - feature not present but ALT_FLAG_NOT is set to mean,
>  		 *   patch if feature is *NOT* present.
>  		 */
> -		if (!boot_cpu_has(a->cpuid) == !(a->flags & ALT_FLAG_NOT)) {
> -			memcpy(insn_buff, instr, a->instrlen);
> -			optimize_nops(instr, insn_buff, a->instrlen);
> -		} else {
> +		if (!boot_cpu_has(a->cpuid) != !(a->flags & ALT_FLAG_NOT)) {
>  			apply_one_alternative(instr, insn_buff, a);
> +			patched = true;
>  		}
>  
> -		text_poke_early(instr, insn_buff, a->instrlen);
> +		instances--;
> +		if (!instances) {
> +			if (!patched) {
> +				memcpy(insn_buff, instr, a->instrlen);
> +				optimize_nops(instr, insn_buff, a->instrlen);
> +			}
> +			text_poke_early(instr, insn_buff, a->instrlen);
> +		}
>  	}
>  
>  	kasan_enable_current();

I think you lost the optimize_nops() call for the patched case.

That is, note how apply_one_alternative() does 0x90 padding, but then
you only do optimize_nops() when !patched.