On Sun, 24 Nov 2024 at 13:50, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>
> Linus, do you see any good reasons to bother with that kind of stuff?
> It's not the first time such metadata update vs. read atomicity comes
> up, maybe we ought to settle that for good and document the decision
> and reasons for it.
>
> This time it's about timestamp (seconds vs. nanoseconds), but there'd
> been mode vs. uid vs. gid mentioned in earlier threads.
I think the only one we ended up really caring about was i_size, which
had the 32-bit tearing problem and we do i_size_read() as a result.
I *do* think that we could perhaps extend (and rename) the
inode->i_size_seqcount to just cover all of the core inode metadata
stuff.
And then - exactly like we already do in practice with
inode->i_size_seqcount - some places might just choose to ignore it
anyway.
But at least using a sequence count shouldn't make things like stat()
any worse in practice.
That said, I don't think this is a real problem in practice. The race
window is too small, and the race effects are too insignificant.
Yes, getting the nanoseconds out of sync with the seconds is a bug,
but when it effectively never happens, and when it *does* happen it
likely has no real downsides, I suspect it's also not something we
should worry over-much about.
So I mention the "rename and extend i_size_seqcount" as a solution
that I suspect might be acceptable if somebody has the motivation and
energy, but honestly I also think "nobody can be bothered" is
acceptable in practice.
Linus
