Re: [PATCH v2] fs: don't let getdents return bogus names

[Jann Horn <jannh@google.com>]

On Tue, Jul 31, 2018 at 6:51 PM Al Viro <viro@zeniv.linux.org.uk> wrote:
>
> On Tue, Jul 31, 2018 at 06:10:27PM +0200, Jann Horn wrote:
> > +/*
> > + * Most filesystems don't filter out bogus directory entry names, and userspace
> > + * can get very confused by such names. Behave as if a low-level IO error had
> > + * happened while reading directory entries.
> > + */
> > +bool bogus_dirent_name(int *errp, const char *name, int namlen,
> > +                    const char *caller)
> > +{
> > +     if (namlen == 0) {
> > +             pr_err_once("%s: filesystem returned bogus empty name\n",
> > +                         caller);
> > +             *errp = -EUCLEAN;
> > +             return true;
> > +     }
> > +     if (memchr(name, '/', namlen)) {
> > +             pr_err_once("%s: filesystem returned bogus name '%*pEhp' (contains slash)\n",
> > +                         caller, namlen, name);
> > +             *errp = -EUCLEAN;
> > +             return true;
> > +     }
> > +     return false;
> > +}
>
> > +     if (bogus_dirent_name(&buf->result, name, namlen, __func__))
> > +             return -EUCLEAN;
>
> These calling conventions st^Ware rather suboptimal.  First of all,
>         * none of ->actor() callbacks will ever get called directly.
>         * there are only 4 callers.  3 of them (all in fs.h) are
> of the form return ....->actor(...) == 0;  The fourth is
>         return orig_ctx->actor(orig_ctx, name, namelen, offset, ino, d_type);
> in ovl_fill_real(), which itself is an ->actor() callback.
>
> So all these "return -E..." in the instances are completely pointless;
> we should just turn filldir_t into pointer-to-function-returning-bool
> and get rid of that boilerplate, rather than adding more to it.

Do you want me to try to write a patch that does that change?

> Furthermore, who the hell cares which callback has stepped into it?
> "The first time it happened from getdents(2) in a 32bit process and
> that's all you'll ever get out of me" seems to be less than helpful...

Yeah, true. Should I just remove the method name from the pr_err_once?
Or should I also use one of the _ratelimited things and print multiple
messages?

> And frankly, I would prefer
>         buf->result = check_dirent_name(name, namelen);
>         if (unlikely(buf->result))
>                 return false;
> making that thing return -EUCLEAN or 0.  Quite possibly - inlining it
> as well...

I guess that would conform to normal kernel coding standards a bit better.

Thanks for the quick feedback!