Re: [RFC] Proposed x86 shadow paging changes

["Teddy Astie" <teddy.astie@xxxxxxxxxx>]

Le 12/06/2025 à 14:12, Jan Beulich a écrit :
> On 12.06.2025 13:45, Teddy Astie wrote:
>> In Xen x86, there is the shadow paging that could be used in some cases
>> (notably with HVM), and is used when HAP is not available. It may also
>> be used in very specific cases for PV guests.
>>
>> With the vast majority of hardware supporting HAP (through EPT or NPT),
>> it's rarely used in practice for HVM, with most downstream projects
>> (XenServer, XCP-ng, QubesOS) are compiling-out its support by default.
>> It also performs very significantly worse than HAP.
>
> Whenever this argument was made, iirc it was in particular George (who
> was involved in the original work aiui) who kept pointing that such a
> statement is not generally true, for numbers being workload dependent.
>

I am a bit curious on specific workloads where shadow paging performs
better than hap. Especially considering that shadow paging doesn't
support PCID and the existence of ASIDs.

>> Therefore, it's not very used in production, and unsurprisingly isn't
>> actively being worked on (most projects being around HAP).
>
> Is it not? Feel free to review this 2 year old series
> (https://lists.xen.org/archives/html/xen-devel/2023-05/msg01140.html),
> to allow it to finally go in. I'm simply reluctant to do any other work
> in that area until such earlier work was properly settled. In fact I may
> have added one or two patches on top in the meantime, but I saw no good
> reason to even post them, given the state of the series.
>
>> One of the issues of Shadow Paging is that it relies on a heavy set of
>> optimizations, with some of them being effectively obselete
>> (optimisations based on heuristics designed for more than 20 years old
>> kernels) or eventually problematic (e.g with modern CPU mitigations e.g
>> L1TF mitigations).
>> Most of these optimizations are hard to reason regarding reliability and
>> security and difficult to debug. Some of these optimizations causes very
>> subtle issues with TLB refactoring for ASID management rework [1].
>> Moreover, it's fair to say performance is no longer the priority for
>> shadow paging.
>>
>> My proposal would be to :
>> - significantly reduce the complexity of shadow paging by dropping most
>> of the complex optimizations
>
> If such can be proven to have no dramatic effect on performance, maybe.
>
>> - consider Shadow Paging as Deprecated
>
> I consider this impossible as long as it's still used for PV migration
> and PV L1TF fallback. In particular ...
>

What about deprecated only for HVM ?

>>   > Functional completeness: Yes
>>   > Functional stability: Quirky
>>   > Interface stability: No (as in, may disappear the next release)
>
> ... I don't see this as being possible to happen any time soon.
>
> Jan
>
>>   > Security supported: Yes
>>
>> Teddy
>>
>> [1]
>> https://lore.kernel.org/xen-devel/9cdb3e67abd01390bcc4cd103ca539d6bf7adbc0.1747312394.git.teddy.astie@xxxxxxxxxx/
>>
>>
>>   | Vates
>>
>> XCP-ng & Xen Orchestra - Vates solutions
>>
>> web: https://vates.tech
>>
>>
>


Teddy Astie | Vates XCP-ng Developer

XCP-ng & Xen Orchestra - Vates solutions

web: https://vates.tech