tcp: use sequence to break TS ties for RACK loss detection
authorYuchung Cheng <ycheng@google.com>
Fri, 13 Jan 2017 06:11:34 +0000 (22:11 -0800)
committerDavid S. Miller <davem@davemloft.net>
Sat, 14 Jan 2017 03:37:16 +0000 (22:37 -0500)
The packets inside a jumbo skb (e.g., TSO) share the same skb
timestamp, even though they are sent sequentially on the wire. Since
RACK is based on time, it can not detect some packets inside the
same skb are lost.  However, we can leverage the packet sequence
numbers as extended timestamps to detect losses. Therefore, when
RACK timestamp is identical to skb's timestamp (i.e., one of the
packets of the skb is acked or sacked), we use the sequence numbers
of the acked and unacked packets to break ties.

We can use the same sequence logic to advance RACK xmit time as
well to detect more losses and avoid timeout.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/tcp.h
include/net/tcp.h
net/ipv4/tcp_input.c
net/ipv4/tcp_recovery.c

index 1255c59..970d5f0 100644 (file)
@@ -208,6 +208,7 @@ struct tcp_sock {
        struct tcp_rack {
                struct skb_mstamp mstamp; /* (Re)sent time of the skb */
                u32 rtt_us;  /* Associated RTT */
+               u32 end_seq; /* Ending TCP sequence of the skb */
                u8 advanced; /* mstamp advanced since last lost marking */
                u8 reord;    /* reordering detected */
        } rack;
index 64fcdeb..5fb1e75 100644 (file)
@@ -1867,7 +1867,7 @@ extern int sysctl_tcp_recovery;
 #define TCP_RACK_LOST_RETRANS  0x1
 
 extern void tcp_rack_mark_lost(struct sock *sk, const struct skb_mstamp *now);
-extern void tcp_rack_advance(struct tcp_sock *tp, u8 sacked,
+extern void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
                             const struct skb_mstamp *xmit_time,
                             const struct skb_mstamp *ack_time);
 extern void tcp_rack_reo_timeout(struct sock *sk);
index be11918..e42ca11 100644 (file)
@@ -1218,7 +1218,8 @@ static u8 tcp_sacktag_one(struct sock *sk,
                return sacked;
 
        if (!(sacked & TCPCB_SACKED_ACKED)) {
-               tcp_rack_advance(tp, sacked, xmit_time, &state->ack_time);
+               tcp_rack_advance(tp, sacked, end_seq,
+                                xmit_time, &state->ack_time);
 
                if (sacked & TCPCB_SACKED_RETRANS) {
                        /* If the segment is not tagged as lost,
@@ -3171,7 +3172,7 @@ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
                } else if (tcp_is_sack(tp)) {
                        tp->delivered += acked_pcount;
                        if (!tcp_skb_spurious_retrans(tp, skb))
-                               tcp_rack_advance(tp, sacked,
+                               tcp_rack_advance(tp, sacked, scb->end_seq,
                                                 &skb->skb_mstamp,
                                                 &sack->ack_time);
                }
index eb39b1b..1e330a2 100644 (file)
@@ -16,6 +16,14 @@ static void tcp_rack_mark_skb_lost(struct sock *sk, struct sk_buff *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));
+}
+
 /* 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:
@@ -60,7 +68,8 @@ static void tcp_rack_detect_loss(struct sock *sk, const struct skb_mstamp *now,
                    scb->sacked & TCPCB_SACKED_ACKED)
                        continue;
 
-               if (skb_mstamp_after(&tp->rack.mstamp, &skb->skb_mstamp)) {
+               if (tcp_rack_sent_after(&tp->rack.mstamp, &skb->skb_mstamp,
+                                       tp->rack.end_seq, scb->end_seq)) {
                        /* Step 3 in draft-cheng-tcpm-rack-00.txt:
                         * A packet is lost if its elapsed time is beyond
                         * the recent RTT plus the reordering window.
@@ -113,14 +122,15 @@ void tcp_rack_mark_lost(struct sock *sk, const struct skb_mstamp *now)
  * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from
  * draft-cheng-tcpm-rack-00.txt
  */
-void tcp_rack_advance(struct tcp_sock *tp, u8 sacked,
+void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
                      const struct skb_mstamp *xmit_time,
                      const struct skb_mstamp *ack_time)
 {
        u32 rtt_us;
 
        if (tp->rack.mstamp.v64 &&
-           !skb_mstamp_after(xmit_time, &tp->rack.mstamp))
+           !tcp_rack_sent_after(xmit_time, &tp->rack.mstamp,
+                                end_seq, tp->rack.end_seq))
                return;
 
        rtt_us = skb_mstamp_us_delta(ack_time, xmit_time);
@@ -140,6 +150,7 @@ void tcp_rack_advance(struct tcp_sock *tp, u8 sacked,
        }
        tp->rack.rtt_us = rtt_us;
        tp->rack.mstamp = *xmit_time;
+       tp->rack.end_seq = end_seq;
        tp->rack.advanced = 1;
 }