Re: [PATCH] net: don't unconditionally copy_from_user a struct ifreq for socket ioctls

[Peter Collingbourne <pcc@google.com>]

On Thu, Aug 26, 2021 at 1:59 AM Arnd Bergmann <arnd@kernel.org> wrote:
>
> On Thu, Aug 26, 2021 at 3:28 AM Peter Collingbourne <pcc@google.com> wrote:
>
> > -       } else {
> > +       } else if (is_dev_ioctl_cmd(cmd)) {
> >                 struct ifreq ifr;
> >                 bool need_copyout;
> >                 if (copy_from_user(&ifr, argp, sizeof(struct ifreq)))
> > @@ -1118,6 +1118,8 @@ static long sock_do_ioctl(struct net *net, struct socket *sock,
> >                 if (!err && need_copyout)
> >                         if (copy_to_user(argp, &ifr, sizeof(struct ifreq)))
> >                                 return -EFAULT;
> > +       } else {
> > +               err = -ENOTTY;
> >         }
> >         return err;
> >  }
> > @@ -3306,6 +3308,8 @@ static int compat_ifr_data_ioctl(struct net *net, unsigned int cmd,
> >         struct ifreq ifreq;
> >         u32 data32;
> >
> > +       if (!is_dev_ioctl_cmd(cmd))
> > +               return -ENOTTY;
> >         if (copy_from_user(ifreq.ifr_name, u_ifreq32->ifr_name, IFNAMSIZ))
> >                 return -EFAULT;
> >         if (get_user(data32, &u_ifreq32->ifr_data))
>
> This adds yet another long switch() statement into the socket ioctl
> case, when there
> is already one in compat_sock_ioctl_trans(), one in dev_ifsioc() and one in
> dev_ioctl(), all with roughly the same set of ioctl command codes. If

I think that David's suggestion of using _IOC_TYPE() should be enough
to address this for now.

> any of them
> are called frequently, that makes it all even slower, so I wonder if
> there should
> be a larger rework altogether. Maybe something based on a single lookup table
> that we search through directly from sock_ioctl()/compat_sock_ioctl() to deal
> with the differences in handling (ifreq based, compat handler, proto_ops
> override, dev_load, rtnl_lock, rcu_read_lock, CAP_NET_ADMIN, copyout, ...).
>
> You are also adding the checks into different places for native and compat
> mode, which makes them diverge more when we should be trying to
> make them more common.
>
> I think based on my recent changes, some other simplifications are possible,
> based on how the compat path already enumerates all the dev ioctls.

I think we should leave that for a followup if still necessary.

Peter