tcp: clean ext on tx recycle

Otherwise we will find stray/unexpected/old extensions
value on next iteration.

On tcp_write_xmit() we can end-up splitting an already queued
skb in two parts, via tso_fragment(). The newly created skb
can be allocated via the tx cache and the mptcp stack will not
be aware of it, so nobody set properly the MPTCP ext.

End result, we transmit the skb using an hold MPTCP DSS map
and that confuses the rx side/corrupt the stream. It requires
some concurrent conditions, so it's not deterministic.

Resetting the ext on recycle fixes all the current mptcp self tests
issues.

Apparently only MPTCP has issues with this kind of stray ext,
so an alternative would be add an additional mptcp hook in
tso_fragment() or in sk_stream_alloc_skb() to always init
the ext.

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

include/net/sock.h

@@ -1470,6 +1470,7 @@ static inline void sk_wmem_free_skb(struct sock *sk, struct sk_buff *skb)
 	sk_mem_uncharge(sk, skb->truesize);
 	if (static_branch_unlikely(&tcp_tx_skb_cache_key) &&
 	    !sk->sk_tx_skb_cache && !skb_cloned(skb)) {
+		skb_ext_reset(skb);
 		skb_zcopy_clear(skb, true);
 		sk->sk_tx_skb_cache = skb;
 		return;