Re: [PATCH] x86/uaccess: small optimization in unsafe_copy_to_user()

From: Al Viro
Date: Sat Apr 17 2021 - 14:19:27 EST

On Sat, Apr 17, 2021 at 09:27:04AM -0700, Linus Torvalds wrote:
> On Sat, Apr 17, 2021 at 9:08 AM Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > Side note: I'm, looking at the readdir cases that I wrote, and I have
> > to just say that is broken too. So "stones and glass houses" etc, and
> > I'll have to fix that too.
> In particular, the very very old OLD_READDIR interface that only fills
> in one dirent at a time didn't call verify_dirent_name(). Same for the
> compat version.
> This requires a corrupt filesystem to be an issue (and even then,
> most/all would have the length of a directory entry in an 'unsigned
> char', so even corrupt filesystems would generally never have a
> negative name length).
> So I don't think it's an issue in _practice_, but at the same time it
> is very much an example of the same issue that put_cmsg() has in
> net-next: unsafe user copies should be fully guarded and not have some
> "but this would never happen because callers would never do anything
> bad".
> Al - fairly trivial patch applied, comments?

Should be fine... FWIW, I've a patch in the same area, making those suckers
return bool. Seeing that they are only ever called via dir_emit(), dir_emit_dot()
and dir_emit_dotdot() and all of those return ->actor(...) == 0...

Anyway, that'd be trivial to rebase on top of yours.