Re: [x86/cpuid] Question regarding historic leaf 0x80000000 code

[Andrew Cooper]

On 20/03/2025 7:33 pm, Ahmed S. Darwish wrote:
> Hi hpa,
>
> As you probably know by now, we're working on a centralized "CPUID table"
> data model, on top of the cleanups at:
>
>     https://lore.kernel.org/x86-cpuid/20250304085152.51092-1-darwi@xxxxxxxxxxxxx
>     https://lore.kernel.org/x86-cpuid/20250319122137.4004-1-darwi@xxxxxxxxxxxxx
>     https://lore.kernel.org/x86-cpuid/20250312143738.458507-1-darwi@xxxxxxxxxxxxx
>
> The idea is to remove all the direct CPUID queries from the x86 code, and
> access a pre-filled table instead (with caveats outside the scope of this
> question.)
>
> While changing the CPUID queries one by one, I've stumbled with:
>
>     => arch/x86/kernel/cpu/common.c
>     void get_cpu_cap(struct cpuinfo_x86 *c)
>     {
> 	...
> 	/* AMD-defined flags: level 0x80000001 */
> 	eax = cpuid_eax(0x80000000);
> 	c->extended_cpuid_level = eax;
>
> 	if ((eax & 0xffff0000) == 0x80000000) {
> 		if (eax >= 0x80000001) {
> 			cpuid(0x80000001, &eax, &ebx, &ecx, &edx);
>
> 			c->x86_capability[CPUID_8000_0001_ECX] = ecx;
> 			c->x86_capability[CPUID_8000_0001_EDX] = edx;
> 		}
> 	}
> 	...
>     }
>
> You've contributed that snippet here:
>
>     https://git.kernel.org/pub/scm/linux/kernel/git/history/history.git/commit/?h=2.4.0-test11pre5&id=67ad24e6d39c3
>
> Do you remember what was the rationale for the "if (eax & 0xffff0000) ==
> 0x80000000" check?
>
> You've also contriubted a similar check to head_32.S:
>
>     https://lore.kernel.org/r/1258154897-6770-2-git-send-email-hpa@xxxxxxxxx
>
>     /* Check if extended functions are implemented */
>     movl $0x80000000, %eax
>     cpuid
>     /* Value must be in the range 0x80000001 to 0x8000ffff */
>     subl $0x80000001, %eax
>     cmpl $(0x8000ffff-0x80000001), %eax
>     ja .Lenable_paging
>
> So I would assume, it would be safe to have a similar check in our
> centralized "system CPUID table" scanner.
>
> In all cases, if you know more details, it would be great to know, so
> that I add it in the centralized CPUID table patch queue changelog (in a
> separate patch.)

The problem is that for an out-of-range leaf, Intel returns the data
from the maximum in-range leaf, rather than zeroes.

On pre-AMD64-capable Intel CPUs, you'll get what amounts to junk in a
query for leaf 0x80000000, where it's probably leaf 5 or so.

Checking for the upper half of the output matching the input is a way of
distinguishing Intel behaviour from all the other vendors.

~Andrew