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

[Hugh Dickins]

On Thu, 23 Oct 2025, Yang Shi wrote:
> On Thu, Oct 23, 2025 at 3:34 AM Kiryl Shutsemau <kirill@xxxxxxxxxxxxx> wrote:
> >
> > On Wed, Oct 22, 2025 at 07:28:27PM +0200, David Hildenbrand wrote:
> > > "garbage" as in pointing at something without a direct map, something that's
> > > protected differently (MTE? weird CoCo protection?) or even worse MMIO with
> > > undesired read-effects.
> >
> > Pedro already points to the problem with missing direct mapping.
> > _nofault() copy should help with this.
> >
> > Can direct mapping ever be converted to MMIO? It can be converted to DMA
> > buffer (which is fine), but MMIO? I have not seen it even in virtualized
> > environments.
> >
> > I cannot say for all CoCo protections, but TDX guest shared<->private
> > should be fine.
> >
> > I am not sure about MTE. Is there a way to bypass MTE check for a load?
> > And how does it deal with stray reads from load_unaligned_zeropad()?
> 
> If I remember correctly, _nofault() copy should skip tag check too.
> 
> Thanks,
> Yang

Not a reply to Yang, I'm just tacking on to the latest mail...

I sew no mention of page migration: __folio_migrate_mapping() does
not use page_cache_delete(), so would surely need to do its own
write_seqcount_begin() and _end(), wouldn't it?

It is using folio_ref_freeze() and _unfreeze(), thinking that
keeps everyone else safely away: but not filemap_read_fast_rcu().

And all other users of folio_ref_freeze() (probably not many but
I have not searched) need to be checked too: maybe no others need
changing, but that has to be established.

This makes a fundamental change to speculative page cache assumptions.

My own opinion was expressed very much better than I could,
by Dave Chinner on Oct 21: cognitive load of niche fastpath.

Hugh