Re: [syzbot] [net?] UBSAN: shift-out-of-bounds in xfrm_selector_match (2)

[Steffen Klassert <steffen.klassert@secunet.com>]

On Wed, Sep 25, 2024 at 01:08:48PM +0200, Sabrina Dubroca wrote:
> 2024-09-24, 14:51:20 -0700, syzbot wrote:
> > syzbot has found a reproducer for the following issue on:
> > 
> > HEAD commit:    151ac45348af net: sparx5: Fix invalid timestamps
> > git tree:       net-next
> > console+strace: https://syzkaller.appspot.com/x/log.txt?x=15808a80580000
> > kernel config:  https://syzkaller.appspot.com/x/.config?x=37c006d80708398d
> > dashboard link: https://syzkaller.appspot.com/bug?extid=cc39f136925517aed571
> > compiler:       Debian clang version 15.0.6, GNU ld (GNU Binutils for Debian) 2.40
> > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=122ad2a9980000
> > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=1387b107980000
> 
> syzbot managed to create an SA with:
> 
> usersa.sel.family = 0
> usersa.sel.prefixlen_s = 128
> usersa.family = AF_INET
> 
> Because of the AF_UNSPEC selector, verify_newsa_info doesn't put
> limits on prefixlen_{s,d}. But then copy_from_user_state sets
> x->sel.family to usersa.family (AF_INET).
> 
> So I think verify_newsa_info should do the same conversion before
> checking prefixlen:
> 
> 
> diff --git a/net/xfrm/xfrm_user.c b/net/xfrm/xfrm_user.c
> index 55f039ec3d59..8d06a37adbd9 100644
> --- a/net/xfrm/xfrm_user.c
> +++ b/net/xfrm/xfrm_user.c
> @@ -201,6 +201,7 @@ static int verify_newsa_info(struct xfrm_usersa_info *p,
>  {
>  	int err;
>  	u8 sa_dir = attrs[XFRMA_SA_DIR] ? nla_get_u8(attrs[XFRMA_SA_DIR]) : 0;
> +	u16 family = p->sel.family;
>  
>  	err = -EINVAL;
>  	switch (p->family) {
> @@ -221,7 +222,10 @@ static int verify_newsa_info(struct xfrm_usersa_info *p,
>  		goto out;
>  	}
>  
> -	switch (p->sel.family) {
> +	if (!family && !(p->flags & XFRM_STATE_AF_UNSPEC))
> +		family = p->family;
> +
> +	switch (family) {
>  	case AF_UNSPEC:
>  		break;
>  
> 
> 
> Steffen, does that make sense?

Yes, it does. Later, in copy_from_user_state() we do

if (!x->sel.family && !(p->flags & XFRM_STATE_AF_UNSPEC))
	x->sel.family = p->family;

anyway.

> Without this, we have prefixlen=128 when we get to addr4_match, which
> does a shift of (32 - prefixlen), so we get
> 
> UBSAN: shift-out-of-bounds in ./include/net/xfrm.h:900:23
> shift exponent -96 is negative
> 
> 
> Maybe a check for prefixlen < 128 would also be useful in the
> XFRM_STATE_AF_UNSPEC case, to avoid the same problems with syzbot
> passing prefixlen=200 for an ipv6 SA. I don't know how
> XFRM_STATE_AF_UNSPEC is used, so I'm not sure what restrictions we can
> put. If we end up with prefixlen = 100 used from ipv4 we'll still have
> the same issues.

I've introduced XFRM_STATE_AF_UNSPEC back in 2008 to make
inter addressfamily tunnels working while maintaining
backwards compatibility to openswan that did not set
the selector family. At least that's what I found in
an E-Mail conversation from back then.

A check for prefixlen <= 128 would make sense in any case.
But not sure if we can restrict that somehow further.

> 
> >  __ip4_datagram_connect+0x96c/0x1260 net/ipv4/datagram.c:49
> >  __ip6_datagram_connect+0x194/0x1230
> >  ip6_datagram_connect net/ipv6/datagram.c:279 [inline]
> >  ip6_datagram_connect_v6_only+0x63/0xa0 net/ipv6/datagram.c:291
> 
> This path also looks a bit dubious. From the reproducer, we have a
> rawv6 socket trying to connect to a v4mapped address, despite having
> ip6_datagram_connect_v6_only as its ->connect.
> 
> pingv6 sockets also use ip6_datagram_connect_v6_only and set
> sk->sk_ipv6only=1 (in net/ipv4/ping.c ping_init_sock), but rawv6 don't
> have this, so __ip6_datagram_connect can end up in
> __ip4_datagram_connect. I guess it would make sense to set it in rawv6
> too. rawv6_bind already rejected v4mapped addresses.
> 
> And then we could add a DEBUG_NET_WARN_ON_ONCE(!ipv6_only_sock(sk)) in
> ip6_datagram_connect_v6_only, or maybe even call ipv6_addr_type to
> reject v4mapped addresses and reject them like the non-AF_INET6 case.

I can't comment on that now, let me have a closer look into it.