Re: [PATCH 3/5] x86/kvm: Convert some slow-path static_cpu_has() callers to boot_cpu_has()

[Paolo Bonzini]

On 31/03/19 17:12, Borislav Petkov wrote:
> On Sun, Mar 31, 2019 at 04:20:11PM +0200, Paolo Bonzini wrote:
>> These are not slow path.
> 
> Those functions do a *lot* of stuff like a bunch of MSR reads which are
> tens of cycles each at least.

The MSR reads and writes are not done in the common case.  Also, you
cannot really expect boot_cpu_data to be in L1 in these functions since
they run after the guest---or if they do, each L1 line you fill in with
host data is one line you "steal" from the guest.

Paolo

> I don't think a RIP-relative MOV and a BT:
> 
>         movq    boot_cpu_data+20(%rip), %rax    # MEM[(const long unsigned int *)&boot_cpu_data + 20B], _45
>         btq     $59, %rax       #, _45
> 
> are at all noticeable.
> 
> On latest AMD and Intel uarch those are 2-4 cycles, according to
> 
> https://agner.org/optimize/instruction_tables.ods
>