tcp: consistently check retransmit hint

tcp_simple_retransmit() used for path MTU discovery may not adjust
the retransmit hint properly by deducting retrans_out before checking
it to adjust the hint. This patch fixes this by a correct routine
tcp_mark_skb_lost() already used by the RACK loss detection.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Author: yuchungcheng

net/ipv4/tcp_input.c

@@ -2675,13 +2675,8 @@ void tcp_simple_retransmit(struct sock *sk)
 
 	skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
 		if (tcp_skb_seglen(skb) > mss &&
-		    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
-			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
-				TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
-				tp->retrans_out -= tcp_skb_pcount(skb);
-			}
-			tcp_skb_mark_lost_uncond_verify(tp, skb);
-		}
+		    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
+			tcp_mark_skb_lost(sk, skb);
 	}
 
 	tcp_clear_retrans_hints_partial(tp);

net/ipv4/tcp_recovery.c

@@ -246,6 +246,6 @@ void tcp_newreno_mark_lost(struct sock *sk, bool snd_una_advanced)
 			tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
 				     mss, mss, GFP_ATOMIC);
 
-		tcp_skb_mark_lost_uncond_verify(tp, skb);
+		tcp_mark_skb_lost(sk, skb);
 	}
 }