Re: [RFC PATCH v2 21/34] x86/msr: Utilize the alternatives mechanism to write MSR

[Jürgen Groß <jgross@xxxxxxxx>]

On 22.04.25 10:22, Xin Li (Intel) wrote:
> The story started from tglx's reply in [1]:
>
>    For actual performance relevant code the current PV ops mechanics
>    are a horrorshow when the op defaults to the native instruction.
>
>    look at wrmsrl():
>
>    wrmsrl(msr, val
>     wrmsr(msr, (u32)val, (u32)val >> 32))
>      paravirt_write_msr(msr, low, high)
>        PVOP_VCALL3(cpu.write_msr, msr, low, high)
>
>    Which results in
>
>         mov     $msr, %edi
>         mov     $val, %rdx
>         mov     %edx, %esi
>         shr     $0x20, %rdx
>         call    native_write_msr
>
>    and native_write_msr() does at minimum:
>
>         mov    %edi,%ecx
>         mov    %esi,%eax
>         wrmsr
>         ret
>
>    In the worst case 'ret' is going through the return thunk. Not to
>    talk about function prologues and whatever.
>
>    This becomes even more silly for trivial instructions like STI/CLI
>    or in the worst case paravirt_nop().
This is nonsense.

In the non-Xen case the initial indirect call is directly replaced with
STI/CLI via alternative patching, while for Xen it is replaced by a direct
call.

The paravirt_nop() case is handled in alt_replace_call() by replacing the
indirect call with a nop in case the target of the call was paravirt_nop()
(which is in fact no_func()).

>    The call makes only sense, when the native default is an actual
>    function, but for the trivial cases it's a blatant engineering
>    trainwreck.
The trivial cases are all handled as stated above: a direct replacement
instruction is placed at the indirect call position.

> Later a consensus was reached to utilize the alternatives mechanism to
> eliminate the indirect call overhead introduced by the pv_ops APIs:
>
>      1) When built with !CONFIG_XEN_PV, X86_FEATURE_XENPV becomes a
>         disabled feature, preventing the Xen code from being built
>         and ensuring the native code is executed unconditionally.
This is the case today already. There is no need for any change to have
this in place.

>      2) When built with CONFIG_XEN_PV:
>
>         2.1) If not running on the Xen hypervisor (!X86_FEATURE_XENPV),
>              the kernel runtime binary is patched to unconditionally
>              jump to the native MSR write code.
>
>         2.2) If running on the Xen hypervisor (X86_FEATURE_XENPV), the
>              kernel runtime binary is patched to unconditionally jump
>              to the Xen MSR write code.
I can't see what is different here compared to today's state.

> The alternatives mechanism is also used to choose the new immediate
> form MSR write instruction when it's available.
Yes, this needs to be added.

> Consequently, remove the pv_ops MSR write APIs and the Xen callbacks.
I still don't see a major difference to today's solution.

Only the "paravirt" term has been eliminated.


Juergen

Attachment: OpenPGP_0xB0DE9DD628BF132F.asc
Description: OpenPGP public key

Attachment: OpenPGP_signature.asc
Description: OpenPGP digital signature