Re: [PATCH] mm/filemap: Implement fast short reads

[David Hildenbrand]

On 27.10.25 17:48, Linus Torvalds wrote:
> On Mon, 27 Oct 2025 at 09:06, David Hildenbrand <david@xxxxxxxxxx> wrote:
> > So I really wish that we can defer optimizing this to freeing folios
> > under RCU instead.
>
> So just to see, I dug around when we started to do the rcu-protected
> folio lookup (well, it was obviously not a folio at the time).
>
> Mainly because we actually had a lot more of those subtle
> folio_try_get() users than I expected us to have,
>
> It goes back to July 2008 (commit e286781d5f2e: "mm: speculative page
> references" being the first in the series).
>
> I do have to say that the original naming was better: we used to call
> the "try_get" operation "page_cache_get_speculative()", which made it
> very clear that it was doing something speculative and different from
> some of our other rcu patterns, where if it's successful it's all
> good.
>
> Because even when successful, the folio in folio_try_get() is still
> speculative and needs checking.

Right, and we only access its content after verifying that (a) it is the 
one we wanted to access and (b) stabilizing it, so it will stay that way.

> Not all of our current users seem to re-verify the source of the folio
> afterwards (deferred_split_scan() makes me go "Uhh - you seem to rely
> on folio_try_get() as some kind of validity check" for example).

Note that deferred_split_scan() only operates with anon folios and has 
some nasty NASTY interaction with folio freeing. :)

--
Cheers

David / dhildenb