Re: general protection fault in skb_segment

From: Marcelo Ricardo Leitner
Date: Sat Dec 30 2017 - 19:52:41 EST


On Sat, Dec 30, 2017 at 08:42:41AM +0100, Willem de Bruijn wrote:
> > syzkaller hit the following crash on
> > 37759fa6d0fa9e4d6036d19ac12f555bfc0aeafd
> > git://git.cmpxchg.org/linux-mmots.git/master
> > compiler: gcc (GCC) 7.1.1 20170620
> > .config is attached
> > Raw console output is attached.
> > C reproducer is attached
> > syzkaller reproducer is attached. See https://goo.gl/kgGztJ
> > for information about syzkaller reproducers
>
> Reproduced with the C reproducer on v4.15-rc1 and mainline
> going back at least to v4.8, but not v4.7. SCTP GSO was
> introduced in v4.8-rc1, so a patch in this set is likely the starting
> point. Indeed crashes at 90017accff61 ("sctp: Add GSO support"),
> but not at 90017accff61~4.
>
> The reproducer with its sandbox removed shows this invocation in strace -f
>
> # strace -f ./repro2
> [... skipped ...]
> socket(PF_INET, SOCK_STREAM, IPPROTO_IP) = 3
> open("/dev/net/tun", O_RDONLY) = 4
> fcntl(4, F_DUPFD, 3) = 5
> socket(PF_PACKET, SOCK_RAW|SOCK_CLOEXEC, 8) = 6
> ioctl(4, TUNSETIFF, 0x20e63000) = 0
> ioctl(3, SIOCSIFFLAGS, {ifr_name="syz0",
> ifr_flags=IFF_UP|IFF_PROMISC|IFF_ALLMULTI}) = 0
> setsockopt(6, SOL_PACKET, 0xf /* PACKET_??? */, [4096], 4) = 0
> ioctl(6, SIOCGIFINDEX, {ifr_name="syz0", ifr_index=24}) = 0
> bind(6, {sa_family=AF_PACKET, proto=0000, if24, pkttype=PACKET_HOST,
> addr(6)={1, aaaaaaaaaa00}, 20) = 0
> dup2(6, 5) = 5
> write(5, "\0\201\1\0\350\367\0\0\3\0E\364\0 \0d\0\0\7\2042\342\0\0\0
> \177\0\0\1\0\t"..., 42
>
> where 0xf in setsockopt is PACKET_VNET_HDR
>
> So this is a packet socket writing something that apparently looks
> like an SCTP packet, is only 42 bytes long, but has GSO set in its
> virtio_net_hdr struct.
>
> It crashes in skb_segment seemingly on a NULL list_skb.
>
> (gdb) list *(skb_segment+0x2a4)
> 0xffffffff8167cc24 is in skb_segment (net/core/skbuff.c:3566).
> 3561 if (hsize < 0)
> 3562 hsize = 0;
> 3563 if (hsize > len || !sg)
> 3564 hsize = len;
> 3565
> 3566 if (!hsize && i >= nfrags && skb_headlen(list_skb) &&
> 3567 (skb_headlen(list_skb) == len || sg)) {
> 3568 BUG_ON(skb_headlen(list_skb) > len);
> 3569
> 3570 i = 0;
>
> Likely there is a hidden assumption about SCTP GSO packets that does
> not hold for such packets generated by PF_PACKET.
>
> SCTP GSO introduced the GSO_BY_FRAGS mss value, so the code
> takes a different path for SCTP packets generated by the SCTP stack.
>
> PF_PACKET does not necessarily set gso_size to GSO_BY_FRAGS, so
> does not take the branch that requires list_skb to be non-zero here:
>
> if (unlikely(mss == GSO_BY_FRAGS)) {
> len = list_skb->len;
> } else {
> len = head_skb->len - offset;
> if (len > mss)
> len = mss;
> }
>
> hsize = skb_headlen(head_skb) - offset;
> if (hsize < 0)
> hsize = 0;
> if (hsize > len || !sg)
> hsize = len;
>
> if (!hsize && i >= nfrags && skb_headlen(list_skb) &&
> (skb_headlen(list_skb) == len || sg)) {
>
> Somewhat tangential, but any PF_PACKET socket can set this
> magic gso_size value in its virtio_net_hdr, so if it is assumed to
> be an SCTP GSO specific option, setting it for a TCP GSO packet
> may also cause unexpected results.

It seems virtio_net could use more sanity checks. When PACKET_VNET_HDR
is used, it will end up calling:
tpacket_rcv() {
...
if (do_vnet) {
if (virtio_net_hdr_from_skb(skb, h.raw + macoff -
sizeof(struct virtio_net_hdr),
vio_le(), true)) {
spin_lock(&sk->sk_receive_queue.lock);
goto drop_n_account;
}
}

and virtio_net_hdr_from_skb does:
if (skb_is_gso(skb)) {
...
if (sinfo->gso_type & SKB_GSO_TCPV4)
hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
else if (sinfo->gso_type & SKB_GSO_TCPV6)
hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
else
return -EINVAL;

Meaning that any gso_type other than TCP would be rejected, but this
SCTP one got through. Seems the header contains a sctp header, but the
gso_type set was actually pointing to TCP (otherwise it would have
been rejected). AFAICT if this packet had an ESP header, for example,
it could have hit esp4_gso_segment. Can you please confirm this?

I don't know of anywhere in the stack validating if the gso_type
matches the header that actually is in there.

The fix you mentioned is a good start, we want that one way or
another, but I'm afraid this bug is bigger than sctp.

Marcelo