Re: [4.9.10] ip_route_me_harder() reading off-slab

From: Daniel J Blueman
Date: Sat Feb 18 2017 - 02:05:18 EST


On 17 February 2017 at 13:36, Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote:
> On Fri, 2017-02-17 at 12:36 +0800, Daniel J Blueman wrote:
>> When booting a VM in libvirt/KVM attached to a local bridge and KASAN
>> enabled on 4.9.10, we see a stream of KASAN warnings about off-slab
>> access [1].
>>
>> Let me know if you'd like more debug.
>
> Could you try the following patch ?
>
> Thanks !
>
> diff --git a/net/ipv4/netfilter.c b/net/ipv4/netfilter.c
> index b3cc1335adbc1a20dcd225d0501b0a286d27e3c8..18839e59da849f0988924bcbc9873965a3681eb0 100644
> --- a/net/ipv4/netfilter.c
> +++ b/net/ipv4/netfilter.c
> @@ -23,7 +23,8 @@ int ip_route_me_harder(struct net *net, struct sk_buff *skb, unsigned int addr_t
> struct rtable *rt;
> struct flowi4 fl4 = {};
> __be32 saddr = iph->saddr;
> - __u8 flags = skb->sk ? inet_sk_flowi_flags(skb->sk) : 0;
> + struct sock *sk = skb->sk;
> + __u8 flags = sk && sk_fullsock(sk) ? inet_sk_flowi_flags(sk) : 0;
> struct net_device *dev = skb_dst(skb)->dev;
> unsigned int hh_len;
>
> @@ -40,7 +41,7 @@ int ip_route_me_harder(struct net *net, struct sk_buff *skb, unsigned int addr_t
> fl4.daddr = iph->daddr;
> fl4.saddr = saddr;
> fl4.flowi4_tos = RT_TOS(iph->tos);
> - fl4.flowi4_oif = skb->sk ? skb->sk->sk_bound_dev_if : 0;
> + fl4.flowi4_oif = sk ? sk->sk_bound_dev_if : 0;
> if (!fl4.flowi4_oif)
> fl4.flowi4_oif = l3mdev_master_ifindex(dev);
> fl4.flowi4_mark = skb->mark;
> @@ -61,7 +62,7 @@ int ip_route_me_harder(struct net *net, struct sk_buff *skb, unsigned int addr_t
> xfrm_decode_session(skb, flowi4_to_flowi(&fl4), AF_INET) == 0) {
> struct dst_entry *dst = skb_dst(skb);
> skb_dst_set(skb, NULL);
> - dst = xfrm_lookup(net, dst, flowi4_to_flowi(&fl4), skb->sk, 0);
> + dst = xfrm_lookup(net, dst, flowi4_to_flowi(&fl4), sk, 0);
> if (IS_ERR(dst))
> return PTR_ERR(dst);
> skb_dst_set(skb, dst);

Fine work! This nicely resolves the issue. I'll test Florian's
proposed fix also.

Tested-by: Daniel J Blueman <daniel@xxxxxxxxx>

Thanks,
Dan
--
Daniel J Blueman