The skbs processed by ip_cmsg_recv() are not guaranteed to be linear e.g. when sending UDP packets over loopback with MSGMORE.
Using csum_partial() on [potentially] the whole skb len is dangerous; instead be on the safe side and use skb_checksum().

Thanks to syzkaller team to detect the issue and provide the reproducer.

v1 -> v2:
- move the variable declaration in a tighter scope

Fixes ad6f939ab193 ("ip: Add offset parameter to ip_cmsg_recv")

Reported-by Andrey Konovalov <[email protected]>

Signed-off-by Paolo Abeni <[email protected]>

Acked-by Eric Dumazet <[email protected]>

Signed-off-by David S. Miller <[email protected]>

Signed-off-by Greg Kroah-Hartman <[email protected]>

---
 net/ipv4/ip_sockglue.c |    8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

net/ipv4/ip_sockglue.c +4 -4

@@ -105,10 +105,10 @@ static void ip_cmsg_recv_checksum(struct
 	if (skb->ip_summed != CHECKSUM_COMPLETE)
 		return;
 
-	if (offset != 0)
-		csum = csum_sub(csum,
-				csum_partial(skb->data + tlen,
-					     offset, 0));
+	if (offset != 0) {
+		int tend_off = skb_transport_offset(skb) + tlen;
+		csum = csum_sub(csum, skb_checksum(skb, tend_off, offset, 0));
+	}
 
 	put_cmsg(msg, SOL_IP, IP_CHECKSUM, sizeof(__wsum), &csum);
 }

Patches currently in stable-queue which might be from [email protected] are

queue-4.4/ip-fix-ip_checksum-handling.patch