Re: [PATCH v6 2/9] x86/bhi: Make clear_bhb_loop() effective on newer CPUs

[Pawan Gupta]

First of all, apologies for not responding to this and many other emails I
still need to read. (For the past few months I was off-work and have been
dealing with a personal emergency. Now thats over, I am catching up with
the pending stuff.)

On Sat, Jan 24, 2026 at 08:34:18PM +0100, Borislav Petkov wrote:
> On Mon, Dec 01, 2025 at 10:19:14PM -0800, Pawan Gupta wrote:
> > diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S
> > index 886f86790b4467347031bc27d3d761d5cc286da1..9f6f4a7c5baf1fe4e3ab18b11e25e2fbcc77489d 100644
> > --- a/arch/x86/entry/entry_64.S
> > +++ b/arch/x86/entry/entry_64.S
> > @@ -1536,7 +1536,11 @@ SYM_FUNC_START(clear_bhb_loop)
> >  	ANNOTATE_NOENDBR
> >  	push	%rbp
> >  	mov	%rsp, %rbp
> > -	movl	$5, %ecx
> > +
> > +	/* loop count differs based on BHI_CTRL, see Intel's BHI guidance */
> > +	ALTERNATIVE "movl $5,  %ecx; movl $5, %edx",	\
> > +		    "movl $12, %ecx; movl $7, %edx", X86_FEATURE_BHI_CTRL
> 
> Why isn't this written like this:
> 
> in C:
> 
> clear_bhb_loop:
> 
> 	if (cpu_feature_enabled(X86_FEATURE_BHI_CTRL))
> 		__clear_bhb_loop(12, 7);
> 	else
> 		__clear_bhb_loop(5, 5);
> 
> and then the __-version is asm and it gets those two arguments from %rdi, and
> %rsi instead of more hard-coded, error-prone registers diddling alternative
> gunk?

This would require CLEAR_BRANCH_HISTORY to move the hard-coded arguments to
the register, which isn't looking pretty:

.macro CLEAR_BRANCH_HISTORY
	ALTERNATIVE "movq $5,  %rdi; movq $5, %rsi",		\
		    "movq $12, %rdi; movq $7, %rsi", X86_FEATURE_BHI_CTRL

	ALTERNATIVE "", "call clear_bhb_loop; lfence", X86_FEATURE_CLEAR_BHB_LOOP
.endm

I don't think we can avoid the register diddling one way or the other. Also
it is best if the loop count stays within clear_bhb_loop(), so that atleast
the callsites can stay clean and don't have to worry about the magic number
arguments.