sock: correct sk_wmem_queued accounting on efault in tcp zerocopy

Syzkaller hits WARN_ON(sk->sk_wmem_queued) in sk_stream_kill_queues
after triggering an EFAULT in __zerocopy_sg_from_iter.

On this error, skb_zerocopy_stream_iter resets the skb to its state
before the operation with __pskb_trim. It cannot kfree_skb like
datagram callers, as the skb may have data from a previous send call.

__pskb_trim calls skb_condense for unowned skbs, which adjusts their
truesize. These tcp skbuffs are owned and their truesize must add up
to sk_wmem_queued. But they match because their skb->sk is NULL until
tcp_transmit_skb.

Temporarily set skb->sk when calling __pskb_trim to signal that the
skbuffs are owned and avoid the skb_condense path.

Fixes: 52267790 ("sock: add MSG_ZEROCOPY")
Signed-off-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/core/skbuff.c

@@ -1124,9 +1124,13 @@ int skb_zerocopy_iter_stream(struct sock *sk, struct sk_buff *skb,
 
 	err = __zerocopy_sg_from_iter(sk, skb, &msg->msg_iter, len);
 	if (err == -EFAULT || (err == -EMSGSIZE && skb->len == orig_len)) {
+		struct sock *save_sk = skb->sk;
+
 		/* Streams do not free skb on error. Reset to prev state. */
 		msg->msg_iter = orig_iter;
+		skb->sk = sk;
 		___pskb_trim(skb, orig_len);
+		skb->sk = save_sk;
 		return err;
 	}