[PATCH net] net/sched: sch_frag: fix OOB read while processing IPv4 fragments

[PATCH net] net/sched: sch_frag: fix OOB read while processing IPv4 fragments

[Davide Caratti]

when the Linux kernel fragments a packet that was previously re-assembled
by the 'act_ct' action, the following splat can be seen on KASAN kernels:

 BUG: KASAN: stack-out-of-bounds in ip_do_fragment+0x1b03/0x1f60
 Read of size 1 at addr ffff88887f209574 by task ping/5640

 CPU: 29 PID: 5640 Comm: ping Tainted: G S                5.12.0-rc6+ #413
 Hardware name: Supermicro SYS-6027R-72RF/X9DRH-7TF/7F/iTF/iF, BIOS 3.0  07/26/2013
 Call Trace:
  <IRQ>
  dump_stack+0x92/0xc1
  print_address_description.constprop.7+0x1a/0x150
  kasan_report.cold.17+0x7f/0x111
  ip_do_fragment+0x1b03/0x1f60
  sch_fragment+0x4bf/0xe40
  tcf_mirred_act+0xc3d/0x11a0 [act_mirred]
  tcf_action_exec+0x104/0x3e0
  fl_classify+0x49a/0x5e0 [cls_flower]

for IPv4 packets, sch_fragment() uses a temporary struct dst_entry. Then,
in the following call graph:

  ip_fragment()
    ip_do_fragment()
      ip_skb_dst_mtu()
        ip_dst_mtu_maybe_forward()
          ip_mtu_locked()

a pointer to that struct is casted as pointer to struct rtable, hence the
OOB stack access. Fix this, changing the temporary variable used for IPv4
packets in sch_fragment(), similarly to what is done for IPv6 in the same
function.

Fixes: c129412f74e9 ("net/sched: sch_frag: add generic packet fragment support.")
Reported-by: Shuang Li <shuali@redhat.com>
Signed-off-by: Davide Caratti <dcaratti@redhat.com>
---
 net/sched/sch_frag.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/net/sched/sch_frag.c b/net/sched/sch_frag.c
index e1e77d3fb6c0..8c06381391d6 100644
--- a/net/sched/sch_frag.c
+++ b/net/sched/sch_frag.c
@@ -90,16 +90,16 @@ static int sch_fragment(struct net *net, struct sk_buff *skb,
 	}
 
 	if (skb_protocol(skb, true) == htons(ETH_P_IP)) {
-		struct dst_entry sch_frag_dst;
+		struct rtable sch_frag_rt = { 0 };
 		unsigned long orig_dst;
 
 		sch_frag_prepare_frag(skb, xmit);
-		dst_init(&sch_frag_dst, &sch_frag_dst_ops, NULL, 1,
+		dst_init(&sch_frag_rt.dst, &sch_frag_dst_ops, NULL, 1,
 			 DST_OBSOLETE_NONE, DST_NOCOUNT);
-		sch_frag_dst.dev = skb->dev;
+		sch_frag_rt.dst.dev = skb->dev;
 
 		orig_dst = skb->_skb_refdst;
-		skb_dst_set_noref(skb, &sch_frag_dst);
+		skb_dst_set_noref(skb, &sch_frag_rt.dst);
 		IPCB(skb)->frag_max_size = mru;
 
 		ret = ip_do_fragment(net, skb->sk, skb, sch_frag_xmit);
-- 
2.30.2

Re: [PATCH net] net/sched: sch_frag: fix OOB read while processing IPv4 fragments

[Davide Caratti]

On Fri, 2021-04-16 at 22:29 +0200, Davide Caratti wrote:

[...]
> 
> for IPv4 packets, sch_fragment() uses a temporary struct dst_entry. Then,
> in the following call graph:
> 
>   ip_fragment()

^^ the above line is a typo,

>     ip_do_fragment()
>       ip_skb_dst_mtu()
>         ip_dst_mtu_maybe_forward()
>           ip_mtu_locked()
> 
> a pointer to that struct is casted as pointer to struct rtable, hence the
> OOB stack access. Fix this, changing the temporary variable used for IPv4
> packets in sch_fragment(), similarly to what is done for IPv6 in the same
> function.

and thanks to Eelco's help I just reproduced a similar splat with
openvswitch. Indeed, ovs_fragment() seems to have the same problem [1];
I will follow-up with a series that fixes both data-paths.

-- 
davide

[1] https://elixir.bootlin.com/linux/v5.12-rc8/source/net/openvswitch/actions.c#L813