Re: [PATCH v1 2/4] x86/fred: Write to FRED MSRs with wrmsrns()

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 03/07/2024 5:06 pm, H. Peter Anvin wrote:
> On July 3, 2024 9:00:53 AM PDT, Andrew Cooper <andrew.cooper3@citrix.com> wrote:
>> On 03/07/2024 4:54 pm, Borislav Petkov wrote:
>>> Dave Hansen <dave.hansen@intel.com> wrote:
>>>> On 7/3/24 01:54, Xin Li (Intel) wrote:
>>>> &gt; Do FRED MSR writes with wrmsrns() rather than wrmsrl().
>>>>
>>>> A longer changelog would be appreciated here.  The wrmsrns() is
>>>> presumably to avoid the WRMSR serialization overhead and the CR4 write
>>>> provides all of the serialization that we need.
>>> Also, all those wrmsrns() writes better be behind a CPUID check.
>> They're not, in Linux.
>>
>> For the $N'th time, here is the primitive that Linux wants to stea^w
>> borrow for this to be sane.
>>
>> /* Non-serialising WRMSR, when available.  Falls back to a serialising
>> WRMSR. */
>> static inline void wrmsrns(uint32_t msr, uint32_t lo, uint32_t hi)
>> {
>>     /*
>>      * WRMSR is 2 bytes.  WRMSRNS is 3 bytes.  Pad WRMSR with a redundant CS
>>      * prefix to avoid a trailing NOP.
>>      */
>>     alternative_input(".byte 0x2e; wrmsr",
>>                       ".byte 0x0f,0x01,0xc6", X86_FEATURE_WRMSRNS,
>>                       "c" (msr), "a" (lo), "d" (hi));
>> }
>>
>> ~Andrew
> I believe tglx declared to use them unconditionally since FRED depends on WRMSRNS (and the kernel enforces that.)

I know that Linux has chosen to have this as a software-enforced
requirement.

The dependency does not exist architecturally, and just because it
happens to be true on Intel processors doesn't mean it's true of other
implementations.

~Andrew