Merge branch 'tcp-rack'

Yuchung Cheng says:

====================
RACK loss detection

RACK (Recent ACK) loss recovery uses the notion of time instead of
packet sequence (FACK) or counts (dupthresh).

It's inspired by the FACK heuristic in tcp_mark_lost_retrans(): when a
limited transmit (new data packet) is sacked in recovery, then any
retransmission sent before that newly sacked packet was sent must have
been lost, since at least one round trip time has elapsed.

But that existing heuristic from tcp_mark_lost_retrans()
has several limitations:
  1) it can't detect tail drops since it depends on limited transmit
  2) it's disabled upon reordering (assumes no reordering)
  3) it's only enabled in fast recovery but not timeout recovery

RACK addresses these limitations with a core idea: an unacknowledged
packet P1 is deemed lost if a packet P2 that was sent later is is
s/acked, since at least one round trip has passed.

Since RACK cares about the time sequence instead of the data sequence
of packets, it can detect tail drops when a later retransmission is
s/acked, while FACK or dupthresh can't. For reordering RACK uses a
dynamically adjusted reordering window ("reo_wnd") to reduce false
positives on ever (small) degree of reordering, similar to the delayed
Early Retransmit.

In the current patch set RACK is only a supplemental loss detection
and does not trigger fast recovery. However we are developing RACK
to replace or consolidate FACK/dupthresh, early retransmit, and
thin-dupack. These heuristics all implicitly bear the time notion.
For example, the delayed Early Retransmit is simply applying RACK
to trigger the fast recovery with small inflight.

RACK requires measuring the minimum RTT. Tracking a global min is less
robust due to traffic engineering pathing changes. Therefore it uses a
windowed filter by Kathleen Nichols. The min RTT can also be useful
for various other purposes like congestion control or stat monitoring.

This patch has been used on Google servers for well over 1 year. RACK
has also been implemented in the QUIC protocol. We are submitting an
IETF draft as well.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>

1 files changed, 21 insertions, 0 deletions

diff --git a/include/net/tcp.h b/include/net/tcp.h
index eed94fc355c1..11e320412216 100644
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -279,6 +279,7 @@ extern int sysctl_tcp_limit_output_bytes;
 extern int sysctl_tcp_challenge_ack_limit;
 extern unsigned int sysctl_tcp_notsent_lowat;
 extern int sysctl_tcp_min_tso_segs;
+extern int sysctl_tcp_min_rtt_wlen;
 extern int sysctl_tcp_autocorking;
 extern int sysctl_tcp_invalid_ratelimit;
 extern int sysctl_tcp_pacing_ss_ratio;
@@ -566,6 +567,7 @@ void tcp_resume_early_retransmit(struct sock *sk);
 void tcp_rearm_rto(struct sock *sk);
 void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req);
 void tcp_reset(struct sock *sk);
+void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb);
 
 /* tcp_timer.c */
 void tcp_init_xmit_timers(struct sock *);
@@ -671,6 +673,12 @@ static inline bool tcp_ca_dst_locked(const struct dst_entry *dst)
 	return dst_metric_locked(dst, RTAX_CC_ALGO);
 }
 
+/* Minimum RTT in usec. ~0 means not available. */
+static inline u32 tcp_min_rtt(const struct tcp_sock *tp)
+{
+	return tp->rtt_min[0].rtt;
+}
+
 /* Compute the actual receive window we are currently advertising.
  * Rcv_nxt can be after the window if our peer push more data
  * than the offered window.
@@ -1743,6 +1751,19 @@ int tcpv4_offload_init(void);
 void tcp_v4_init(void);
 void tcp_init(void);
 
+/* tcp_recovery.c */
+
+/* Flags to enable various loss recovery features. See below */
+extern int sysctl_tcp_recovery;
+
+/* Use TCP RACK to detect (some) tail and retransmit losses */
+#define TCP_RACK_LOST_RETRANS  0x1
+
+extern int tcp_rack_mark_lost(struct sock *sk);
+
+extern void tcp_rack_advance(struct tcp_sock *tp,
+			     const struct skb_mstamp *xmit_time, u8 sacked);
+
 /*
  * Save and compile IPv4 options, return a pointer to it
  */