Re: [PATCH v1 06/11] mm: support GUP-triggered unsharing via FAULT_FLAG_UNSHARE (!hugetlb)

[Linus Torvalds]

On Thu, Dec 23, 2021 at 4:54 AM Jan Kara <jack@xxxxxxx> wrote:
>
> So forcing pinning for O_DIRECT reads is IMO mandatory.

I don't disagree.

And I do think the eventual aim should be to do it for writes too even
if they don't necessarily require it (since they write _from_ the VM
data, not _to_ the VM data - the "read-vs-write direction has always
been confusing when it comes to GUP").

Partly just for consistency in the IO paths - I think people want to
share as much as possible in there - but also just to make sure that
we're all done with the "wrong-way-cow" kind of issues for good.

If we get to the point where the legacy GUP is used only for very
special things (looking up physical pages for debug and trace purposes
etc), I think that would be lovely.

That may be a pretty long-range goal, though.

              Linus