Re: [Xen-devel] [PATCH] iommu/quirk: disable shared EPT for Sandybridge and earlier processors.

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 25/11/15 10:49, Jan Beulich wrote:
>>>> On 25.11.15 at 11:28, <andrew.cooper3@xxxxxxxxxx> wrote:
>> On 24/11/15 17:41, Jan Beulich wrote:
>>>>>> On 24.11.15 at 18:17, <Anshul Makkar anshul.makkar@xxxxxxxxxx> wrote:
>>>> --- a/xen/drivers/passthrough/vtd/quirks.c
>>>> +++ b/xen/drivers/passthrough/vtd/quirks.c
>>>> @@ -320,6 +320,20 @@ void __init platform_quirks_init(void)
>>>>      /* Tylersburg interrupt remap quirk */
>>>>      if ( iommu_intremap )
>>>>          tylersburg_intremap_quirk();
>>>> +
>>>> +    /*
>>>> +     * Disable shared EPT ("sharept") on Sandybridge and older processors
>>>> +     * by default.
>>>> +     * SandyBridge has no huge page support for IOTLB which leads to 
>>>> fallback
>>>> +     * on 4k pages and leads to performance degradation.
>>>> +     *
>>>> +     * Shared EPT ("sharept") will be disabled only if user has not
>>>> +     * provided explicit choice on the command line thus 
>>>> iommu_hap_pt_share is
>>>> +     * at its initialized value of -1.
>>>> +     */
>>>> +    if ( (boot_cpu_data.x86 == 0x06 && (boot_cpu_data.x86_model <= 0x2F ||
>>>> +          boot_cpu_data.x86_model == 0x36)) && (iommu_hap_pt_share == -1) 
>>>> )
>>>> +        iommu_hap_pt_share = 0;
>>> If we really want to do this, then I think we should key this on
>>> EPT but not VT-d having 2M support, instead of on CPU models.
>> This check is already performed by vtd_ept_page_compatible()
> Yeah, I realized there would be such a check on the way home.
>
>> The problem is that SandyBridge IOMMUs advertise 2M support and do
>> function with it, but cannot cache 2MB translations in the IOTLBs.
>>
>> As a result, attempting to use 2M translations causes substantially
>> worse performance than 4K translations.
> So commit message and comment should make this more explicit,
> to avoid the impression "IOTLB" isn't just the relatively common
> mis-naming of "IOMMU".
>
> Plus I guess the sharing won't need suppressing if !opt_hap_2mb?
>
> Further the model based check is relatively broad, and includes
> Atoms (0x36 actually is one), which can't be considered "Sandybridge
> or older" imo.
>
> And finally I'm not fully convinced using CPU model info to deduce
> chipset behavior is entirely correct (albeit perhaps in practice it'll
> be fine except maybe when running Xen itself virtualized).

What else would you suggest? I can't think of any better identifying
information.

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel