[RFC PATCH v1 0/4] Reduce cost of ptep_get_lockless on arm64

[David Hildenbrand]

On 15.02.24 13:17, Ryan Roberts wrote:
> This is an RFC for a series that aims to reduce the cost and complexity of
> ptep_get_lockless() for arm64 when supporting transparent contpte mappings [1].
> The approach came from discussion with Mark and David [2].
> 
> It introduces a new helper, ptep_get_lockless_norecency(), which allows the
> access and dirty bits in the returned pte to be incorrect. This relaxation
> permits arm64's implementation to just read the single target pte, and avoids
> having to iterate over the full contpte block to gather the access and dirty
> bits, for the contpte case.
> 
> It turns out that none of the call sites using ptep_get_lockless() require
> accurate access and dirty bit information, so we can also convert those sites.
> Although a couple of places need care (see patches 2 and 3).
> 
> Arguably patch 3 is a bit fragile, given the wide accessibility of
> vmf->orig_pte. So it might make sense to drop this patch and stick to using
> ptep_get_lockless() in the page fault path. I'm keen to hear opinions.

Yes. Especially as we have these pte_same() checks that might just fail 
now because of wrong accessed/dirty bits?

Likely, we just want to read "the real deal" on both sides of the 
pte_same() handling.

> 
> I've chosen the name "recency" because it's shortish and somewhat descriptive,
> and is alredy used in a couple of places to mean similar things (see mglru and
> damon). I'm open to other names if anyone has better ideas.

Not a native speaker; works for me.

> 
> If concensus is that this approach is generally acceptable, I intend to create a
> series in future to do a similar thing with ptep_get() -> ptep_get_norecency().

Yes, sounds good to me.

-- 
Cheers,

David / dhildenb