]> git.baikalelectronics.ru Git - kernel.git/commit
netfilter: ecache: don't look for ecache extension on dying/unconfirmed conntracks
authorFlorian Westphal <fw@strlen.de>
Tue, 22 Oct 2019 16:56:42 +0000 (18:56 +0200)
committerPablo Neira Ayuso <pablo@netfilter.org>
Sat, 26 Oct 2019 10:36:42 +0000 (12:36 +0200)
commit700a769d2f2c61fd2b25e579448771b19c7e9ad1
tree92d573c137954df64547203823621fc66cc8d29f
parent1206b018ab018e0b0040e827a22b89d14caa2bc9
netfilter: ecache: don't look for ecache extension on dying/unconfirmed conntracks

syzbot reported following splat:
BUG: KASAN: use-after-free in __nf_ct_ext_exist
include/net/netfilter/nf_conntrack_extend.h:53 [inline]
BUG: KASAN: use-after-free in nf_ct_deliver_cached_events+0x5c3/0x6d0
net/netfilter/nf_conntrack_ecache.c:205
nf_conntrack_confirm include/net/netfilter/nf_conntrack_core.h:65 [inline]
nf_confirm+0x3d8/0x4d0 net/netfilter/nf_conntrack_proto.c:154
[..]

While there is no reproducer yet, the syzbot report contains one
interesting bit of information:

Freed by task 27585:
[..]
 kfree+0x10a/0x2c0 mm/slab.c:3757
 nf_ct_ext_destroy+0x2ab/0x2e0 net/netfilter/nf_conntrack_extend.c:38
 nf_conntrack_free+0x8f/0xe0 net/netfilter/nf_conntrack_core.c:1418
 destroy_conntrack+0x1a2/0x270 net/netfilter/nf_conntrack_core.c:626
 nf_conntrack_put include/linux/netfilter/nf_conntrack_common.h:31 [inline]
 nf_ct_resolve_clash net/netfilter/nf_conntrack_core.c:915 [inline]
 ^^^^^^^^^^^^^^^^^^^
 __nf_conntrack_confirm+0x21ca/0x2830 net/netfilter/nf_conntrack_core.c:1038
 nf_conntrack_confirm include/net/netfilter/nf_conntrack_core.h:63 [inline]
 nf_confirm+0x3e7/0x4d0 net/netfilter/nf_conntrack_proto.c:154

This is whats happening:

1. a conntrack entry is about to be confirmed (added to hash table).
2. a clash with existing entry is detected.
3. nf_ct_resolve_clash() puts skb->nfct (the "losing" entry).
4. this entry now has a refcount of 0 and is freed to SLAB_TYPESAFE_BY_RCU
   kmem cache.

skb->nfct has been replaced by the one found in the hash.
Problem is that nf_conntrack_confirm() uses the old ct:

static inline int nf_conntrack_confirm(struct sk_buff *skb)
{
 struct nf_conn *ct = (struct nf_conn *)skb_nfct(skb);
 int ret = NF_ACCEPT;

  if (ct) {
    if (!nf_ct_is_confirmed(ct))
       ret = __nf_conntrack_confirm(skb);
    if (likely(ret == NF_ACCEPT))
nf_ct_deliver_cached_events(ct); /* This ct has refcount 0! */
  }
  return ret;
}

As of "netfilter: conntrack: free extension area immediately", we can't
access conntrack extensions in this case.

To fix this, make sure we check the dying bit presence before attempting
to get the eache extension.

Reported-by: syzbot+c7aabc9fe93e7f3637ba@syzkaller.appspotmail.com
Fixes: 0b331c716b8528 ("netfilter: conntrack: free extension area immediately")
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
net/netfilter/nf_conntrack_ecache.c