tcp_recovery.c

#include <linux/tcp.h>
#include <net/tcp.h>

int sysctl_tcp_recovery __read_mostly = TCP_RACK_LOSS_DETECTION;

static void tcp_rack_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
{
	struct tcp_sock *tp = tcp_sk(sk);

	tcp_skb_mark_lost_uncond_verify(tp, skb);
	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
		/* Account for retransmits that are lost again */
		TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
		tp->retrans_out -= tcp_skb_pcount(skb);
		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
			      tcp_skb_pcount(skb));
	}
}

static bool tcp_rack_sent_after(const struct skb_mstamp *t1,
				const struct skb_mstamp *t2,
				u32 seq1, u32 seq2)
{
	return skb_mstamp_after(t1, t2) ||
	       (t1->v64 == t2->v64 && after(seq1, seq2));
}

/* RACK loss detection (IETF draft draft-ietf-tcpm-rack-01):
 *
 * Marks a packet lost, if some packet sent later has been (s)acked.
 * The underlying idea is similar to the traditional dupthresh and FACK
 * but they look at different metrics:
 *
 * dupthresh: 3 OOO packets delivered (packet count)
 * FACK: sequence delta to highest sacked sequence (sequence space)
 * RACK: sent time delta to the latest delivered packet (time domain)
 *
 * The advantage of RACK is it applies to both original and retransmitted
 * packet and therefore is robust against tail losses. Another advantage
 * is being more resilient to reordering by simply allowing some
 * "settling delay", instead of tweaking the dupthresh.
 *
 * When tcp_rack_detect_loss() detects some packets are lost and we
 * are not already in the CA_Recovery state, either tcp_rack_reo_timeout()
 * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will
 * make us enter the CA_Recovery state.
 */
static void tcp_rack_detect_loss(struct sock *sk, const struct skb_mstamp *now,
				 u32 *reo_timeout)
{
	struct tcp_sock *tp = tcp_sk(sk);
	struct sk_buff *skb;
	u32 reo_wnd;

	*reo_timeout = 0;
	/* To be more reordering resilient, allow min_rtt/4 settling delay
	 * (lower-bounded to 1000uS). We use min_rtt instead of the smoothed
	 * RTT because reordering is often a path property and less related
	 * to queuing or delayed ACKs.
	 */
	reo_wnd = 1000;
	if ((tp->rack.reord || !tp->lost_out) && tcp_min_rtt(tp) != ~0U)
		reo_wnd = max(tcp_min_rtt(tp) >> 2, reo_wnd);

	tcp_for_write_queue(skb, sk) {
		struct tcp_skb_cb *scb = TCP_SKB_CB(skb);

		if (skb == tcp_send_head(sk))
			break;