Re: [PATCH] mm/filemap: do not allocate cache pages beyond end of file at read

[Linus Torvalds <torvalds@linux-foundation.org>]

On Wed, Oct 30, 2019 at 7:50 AM Kirill A. Shutemov <kirill@shutemov.name> wrote:
>
> I don't know much about filesystems, but can't size of file change after
> the open() under network filesystem? Revlidation on read looks like an
> requirement anyway, no?

Requirement? No. But QoS issue, yes.

But note that NFS already does that. Look at nfs_file_read(), and
notice how it's not using generic_file_buffered_read() directly, it's
doing its own thing first with checking for direct-IO, but then doing
that nfs_revalidate_mapping() that checks whether caches should be
re-validated.

It's not just size of the file, the actual cached contents may need
invalidating too etc.

And note how the generic page cache reader doesn't need to care. If
what the generic code does isn't enough, or is the wrong thing, the
filesystem simply shouldn't use it, or, like NFS, do its own thing
first/last.

So I think the "some filesystems may have other rules" is irrelevant.
If they do have other rules, it's _their_ issue, not the issue of the
generic page cache read logic.

             Linus