Re: [RFC PATCH 03/15] x86/cpu/intel: Fix init_intel() checks for extended family numbers

[Sohil Mehta]

On 12/20/2024 3:27 PM, Dave Hansen wrote:
> On 12/20/24 13:36, Sohil Mehta wrote:
>> X86_FEATURE_REP_GOOD is only set for family 6 processors.  Extend the
>> check to family numbers beyond 15.
> 
> Could you explain why, please?
> 

To answer this I was trying to understand where Fast string
(X86_FEATURE_REP_GOOD) is used. It looks like copy_page() in
lib/copy_page_64.S is the only place it really matters. clear_page() in
include/asm/page_64.h would likely use Enhanced fast strings (ERMS) if
available.

Would it be correct to say that copy_page() and potentially clear_page()
would be slower on Family 18/19 CPUs without the fix?

>> It is uncertain whether the Pentium 4s (family 15) should set the
>> feature flag as well. Commit 185f3b9da24c ("x86: make intel.c have
>> 64-bit support code") that originally set X86_FEATURE_REP_GOOD also set
>> the x86_cache_alignment preference for family 15 processors. The
>> omission of the family 15 seems intentional.
>>

Analyzing more, it seems the below check in early_init_intel() is not
really effective.

/*
 * If fast string is not enabled in IA32_MISC_ENABLE for any reason,
 * clear the fast string and enhanced fast string CPU capabilities.
 */
if (c->x86_vfm >= INTEL_PENTIUM_M_DOTHAN) {
	rdmsrl(MSR_IA32_MISC_ENABLE, misc_enable);
	if (!(misc_enable & MSR_IA32_MISC_ENABLE_FAST_STRING)) {
		pr_info("Disabled fast string operations\n");
		setup_clear_cpu_cap(X86_FEATURE_REP_GOOD);
		setup_clear_cpu_cap(X86_FEATURE_ERMS);
	}
}


It gets overridden later in intel_init() with the below code:

if (c->x86 == 6)
	set_cpu_cap(c, X86_FEATURE_REP_GOOD);

Shouldn't the order of this be the other way around?

>> Also, the 32-bit user copy alignment preference is only set for family 6
>> and 15 processors. Extend the preference to family numbers beyond 15.
> 
> Can you please provide some more context so it's clear which hunk this
> refers to?  Alternatively, can you break this out into a separate patch?
> 

This is referring to the below chunk. Separating it seems like a better
idea to avoid ambiguity.

> -
>  #ifdef CONFIG_X86_INTEL_USERCOPY
>  	/*
>  	 * Set up the preferred alignment for movsl bulk memory moves
> +	 * Family 4 - 486: untested
> +	 * Family 5 - Pentium: untested
> +	 * Family 6 - PII/PIII only like movsl with 8-byte alignment
> +	 * Family 15 - P4 is OK down to 8-byte alignment
>  	 */
> -	switch (c->x86) {
> -	case 4:		/* 486: untested */
> -		break;
> -	case 5:		/* Old Pentia: untested */
> -		break;
> -	case 6:		/* PII/PIII only like movsl with 8-byte alignment */
> -		movsl_mask.mask = 7;
> -		break;
> -	case 15:	/* P4 is OK down to 8-byte alignment */
> +	if (c->x86_vfm >= INTEL_PENTIUM_PRO)
>  		movsl_mask.mask = 7;
> -		break;
> -	}
>  #endif