Re: [PATCH net 2/4] Revert "kill dev_ifsioc()"

[Johannes Berg <johannes@sipsolutions.net>]

On Sat, 2019-01-26 at 18:49 +0100, Johannes Berg wrote:
> On Sat, 2019-01-26 at 18:45 +0100, Johannes Berg wrote:
> > 
> > Yes and no. It *sometimes* (actually rarely, since we don't really have
> > dev_ioctls that much, afaict) hits this, but it could also just hit
> 
> Actually, no, I'm wrong. We do mostly hit dev_ioctl(), since that's the
> common case for things like SIOCGIFNAME.
> 
> However, e.g. for SIOCGIFADDR we do go into
> 
> > static long sock_do_ioctl(struct net *net, struct socket *sock,
> >                           unsigned int cmd, unsigned long arg)
> > {
> > [...]
> >         err = sock->ops->ioctl(sock, cmd, arg);
> > [...]
> >         if (err != -ENOIOCTLCMD)
> >                 return err;
> 
> This, and like I said, plumbing the whole compat stuff through to the
> sock->ops->ioctl() there doesn't seem like a great idea.

So - discussing this further on IRC I thought we could get away with
making struct ifreq just not include the members that are too big
(ifr_map and ifr_settings), but that's also a non-starter because we
need to copy.

Al also points out that all of these reverts break decnet, because that
does some really messy things in dn_dev_ioctl(). Turns out though that
on 64-bit decnet is already broken anyway, because DN_IFREQ_SIZE is
actually wrong. It should presumably be equivalent to something like

struct ifreq_dn {
	char    ifrn_name[IFNAMSIZ];
	struct  sockaddr_dn ifru_dnaddr;
};

but in fact *isn't* because 

#define DN_IFREQ_SIZE (sizeof(struct ifreq) - sizeof(struct sockaddr) + sizeof(struct sockaddr_dn))

wouldn't be sizeof(struct ifreq_dn), because in this expression
"sizeof(struct ifreq) - sizeof(struct sockaddr)" isn't the same as
"offsetof(struct ifreq, ifr_ifru)" which would be the right thing.

So with these patches we go back to the original state before Al's
patches, but that does mean:

 * decnet doesn't work right on 64-bit (kernel & userland) because it
   will attempt to copy a bigger buffer than the user would presumably
   be expected to provide a struct ifreq_dn like I defined above to
   SIOCGIFADDR, and if this is at the end of a page boundary it faults
 * 32-bit userland on 64-bit kernel is completely broken here for decnet
   ioctls because we copy too little data (struct ifreq, while struct
   ifreq_dn is bigger)

The first of these isn't that hard to fix, just fix the DN_IFREQ_SIZE.

The second one I don't see a good way to fix at all.

johannes