On Fri, Nov 4, 2022 at 10:11 AM Pavel Machek <pavel@xxxxxxx> wrote: > > Hi! > > > [ Upstream commit d89d7ff01235f218dad37de84457717f699dee79 ] > > > > Another syzbot report [1] with no reproducer hints > > at a bug in ip6_gre tunnel (dev:ip6gretap0) > > > > Since ipv6 mcast code makes sure to read dev->mtu once > > and applies a sanity check on it (see commit b9b312a7a451 > > "ipv6: mcast: better catch silly mtu values"), a remaining > > possibility is that a layer is able to set dev->mtu to > > an underflowed value (high order bit set). > > > > This could happen indeed in ip6gre_tnl_link_config_route(), > > ip6_tnl_link_config() and ipip6_tunnel_bind_dev() > > > > Make sure to sanitize mtu value in a local variable before > > it is written once on dev->mtu, as lockless readers could > > catch wrong temporary value. > > Ok, but now types seem to be confused: > > > diff --git a/net/ipv6/ip6_tunnel.c b/net/ipv6/ip6_tunnel.c > > index 3a2741569b84..0d4cab94c5dd 100644 > > --- a/net/ipv6/ip6_tunnel.c > > +++ b/net/ipv6/ip6_tunnel.c > > @@ -1476,8 +1476,8 @@ static void ip6_tnl_link_config(struct ip6_tnl *t) > > struct net_device *tdev = NULL; > > struct __ip6_tnl_parm *p = &t->parms; > > struct flowi6 *fl6 = &t->fl.u.ip6; > > - unsigned int mtu; > > int t_hlen; > > + int mtu; > > > > memcpy(dev->dev_addr, &p->laddr, sizeof(struct in6_addr)); > > memcpy(dev->broadcast, &p->raddr, sizeof(struct in6_addr)); > > @@ -1524,12 +1524,13 @@ static void ip6_tnl_link_config(struct ip6_tnl *t) > > dev->hard_header_len = tdev->hard_header_len + t_hlen; > > mtu = min_t(unsigned int, tdev->mtu, IP6_MAX_MTU); > > mtu is now signed, but we still do min_t on unsigned types. > > > - dev->mtu = mtu - t_hlen; > > + mtu = mtu - t_hlen; > > if (!(t->parms.flags & IP6_TNL_F_IGN_ENCAP_LIMIT)) > > - dev->mtu -= 8; > > + mtu -= 8; > > > > I don't see overflow potential right away, but it may be worth fixing. > This was intended ( part of the fix) so that the following check is going to catch 'negative' mtu [1] if (mtu < IPV6_MIN_MTU) mtu = IPV6_MIN_MTU; Otherwise, if a fuzzer succeeds to get mtu = 0xFFFFFFC0, sanity test [1] leaves the problematic mtu in dev->mtu. Thanks.