Re: [PATCH v2 net] netfilter: nft_payload: validate offset for all csum_type paths

[Florian Westphal <fw@strlen.de>]

Siho Lee <25esihoya@gmail.com> wrote:
> For NFT_PAYLOAD_LL_HEADER, offset is computed as:
>     offset = skb_mac_header(skb) - skb->data - vlan_hlen
> which evaluates to -14 (or -18 with VLAN) after eth_type_trans()
> pulls the Ethernet header. This is a valid negative offset that
> refers to the Ethernet header area (used by bridge/vlan rules).
> 
> However, without any bounds check in the csum=NONE path:
> - skb_ensure_writable(skb, max(offset + priv->len, 0)):
>   max() converts negative values to 0, making it a no-op.

Are you sure?

> - skb_store_bits(skb, offset, src, priv->len):
>   A negative offset that exceeds skb headroom writes out of bounds.

Sure, but how can that happen?  This should be explained here,
because I am NOT seeing a bug in the first place.

> Add proper validation after the csum condition block:
> - Negative offsets: ensure they fall within skb_headroom(skb)
>   (bridge/vlan rules legitimately access the Ethernet header)
> - Positive offsets: ensure offset + len does not exceed skb->len

Large offset/len should make skb_copy_bits return an error.

> Also remove the max() wrapper from skb_ensure_writable() since
> the new validation guarantees the offset is within range.

No, this patch still breaks test cases we have.  Please figure out
if there is a bug in this code, and if there is, explain it in a way
that I can understand (e.g. provide broken example that triggers OOB).

Then, git clone https://git.netfilter.org/nftables and make sure the
tests pass.

> +	if (skb_ensure_writable(skb, offset + priv->len) ||

ensure_writable has unsigned arg, so this -14 + 6 will asks for
4GB and this aborts here.