Re: [PATCH 4/5] x86/cpuid: Standardize on u32 in <asm/cpuid/api.h>

[Andrew Cooper]

On 18/03/2025 7:25 pm, H. Peter Anvin wrote:
> On March 18, 2025 12:09:59 PM PDT, Andrew Cooper <andrew.cooper3@xxxxxxxxxx> wrote:
>> On 18/03/2025 6:48 pm, H. Peter Anvin wrote:
>>> One more thing is that we ought to be able to make cpuid a const function, allowing the compiler to elide multiple calls. (Slight warning for feature-enabling MSRs changing CPUID), but that would require changing the API to returning a structure, since a pure or const structure can't return values by reference.
>> It's not only the feature-enabling MSRs.  It's also OSXSAVE/OSPKE/etc in
>> CR4, and on Intel CPUs, the CPUID instruction still has a side effect
>> for microcode patch revision MSR.
>>
>> There are a few too many side effects to call it const/pure.
>>
>> That said, when experimenting with the same in Xen, there was nothing
>> interesting the compiler could do with const/pure because of how the
>> existing logic is laid out.  Removing volatile and the memory clobber
>> however did allow the compiler to make slightly better code.
>>
>> ~Andrew
> Well, I guess I lump CRs, DRs and MSRs together. There is also CPUID for serialization, which is really a totally different use for the same instruction.

Andy Luto got rid of all CPUID serialisation ages back.  It's about the
worst of the available options, even on native.  It's IRET-to-self
(doesn't exit under any virt), or SERIALISE on bleeding edge CPUs.

> tglx has suggested that we should cache or even preload the cpuid data (the latter would have the potential advantage of making the memory data structures a little easier to manage, given the very large potential space.)
>
> The biggest issue is that there is no general mechanism for detecting which cpuid leaves have subleaves, and if they do, how many. I *believe* all existing subleaf sets are dense, but one could at least hypothetically see a vendor or VM define a CPUID leaf with a sparse subleaf set.

XSTATE is sparse in general; AVX-512 being the notable absence on
Intel's client line, and AMD has LWP at subleaf 62.

But yes, these are all problems trying to maintain an in-memory copy of
the CPUID state.  Maintenance of this in Xen leaves a lot to be desired.

~Andrew