Re: [PATCH v1 37/40] x86/cacheinfo: Extract out cache self-snoop checks

[Andrew Cooper]

On 05/03/2025 6:40 pm, Ahmed S. Darwish wrote:
> Hi Andrew,
>
> On Tue, 04 Mar 2025, Andrew Cooper wrote:
>> On 04/03/2025 8:51 am, Ahmed S. Darwish wrote:
>>> The logic of not doing a cache flush if the CPU declares cache self
>>> snooping support is repeated across the x86/cacheinfo code.  Extract it
>>> into its own function.
>>>
>>> Signed-off-by: Ahmed S. Darwish <darwi@xxxxxxxxxxxxx>
>> I know you're just refactoring code, but the SDM has basically reverted
>> this statement about it being safe to skip WBINVD based on SELFSNOOP.
>>
> Still, thanks a lot for sharing :)
>
>> It turns out not to be safe in cases where the underlying physical
>> memory changes from cacheable to unchangeable.  By skipping the WBINVD
>> as part of changing the memory type, you end up with spurious writebacks
>> at a later point when the memory is expected to be UC.  Apparently this
>> is a problem for CLX devices, hence the change in the SDM.
> While writing that refactoring patch, I indeed noticed that there is an
> errata list of CPUs where X86_FEATURE_SELFSNOOP is force disabled, thus
> ensuring WBINVD is never skipped:
>
>     static void check_memory_type_self_snoop_errata(...)
>     {
>      	switch (c->x86_vfm) {
>      	case INTEL_CORE_YONAH:
>      	case INTEL_CORE2_MEROM:
>      	case INTEL_CORE2_MEROM_L:
>      	case INTEL_CORE2_PENRYN:
>      	case INTEL_CORE2_DUNNINGTON:
>      	case INTEL_NEHALEM:
>      	case INTEL_NEHALEM_G:
>      	case INTEL_NEHALEM_EP:
>      	case INTEL_NEHALEM_EX:
>      	case INTEL_WESTMERE:
>      	case INTEL_WESTMERE_EP:
>      	case INTEL_SANDYBRIDGE:
>      		setup_clear_cpu_cap(X86_FEATURE_SELFSNOOP);
>      	}
>     }
>
> That's why I added "CPUs without known erratas" in the comments:
>
>     /*
>      * Cache flushing is the most time-consuming step when programming
>      * the MTRRs.  On many Intel CPUs without known erratas, it can be
>      * skipped if the CPU declares cache self-snooping support.
>      */
>     static void maybe_flush_caches(void)
>     {
>            if (!static_cpu_has(X86_FEATURE_SELFSNOOP))
>                    wbinvd();
>     }
>
> But, interestingly, CLX devices (intel-family.h CASCADELAKE_X /
> SKYLAKE_X) are not part of the kernel's Self Snoop errata list above.

CLX (Cascade Lake) != CXL (Compute eXpress Link).

CXL is the new PCIe.  (So say the CXL consortium at least.)

~Andrew