Merge branch 'tcp-ts-usec-resolution'

Eric Dumazet says:

====================
tcp: add optional usec resolution to TCP TS

As discussed in various public places in 2016, Google adopted
usec resolution in RFC 7323 TS values, at Van Jacobson suggestion.

Goals were :

1) better observability of delays in networking stacks/fabrics.

2) better disambiguation of events based on TSval/ecr values.

3) building block for congestion control modules needing usec resolution.

Back then we implemented a schem based on private SYN options
to safely negotiate the feature.

For upstream submission, we chose to use a much simpler route
attribute because this feature is probably going to be used
in private networks.

ip route add 10/8 ... features tcp_usec_ts

References:

https://www.ietf.org/proceedings/97/slides/slides-97-tcpm-tcp-options-for-low-latency-00.pdf
https://datatracker.ietf.org/doc/draft-wang-tcpm-low-latency-opt/

First two patches are fixing old minor bugs and might be taken
by stable teams (thanks to appropriate Fixes: tags)
====================

Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

5 files changed, 65 insertions, 25 deletions

diff --git a/include/linux/tcp.h b/include/linux/tcp.h
index e15452df9804..6df715b6e51d 100644
--- a/include/linux/tcp.h
+++ b/include/linux/tcp.h
@@ -152,6 +152,7 @@ struct tcp_request_sock {
 	u64				snt_synack; /* first SYNACK sent time */
 	bool				tfo_listener;
 	bool				is_mptcp;
+	s8				req_usec_ts;
 #if IS_ENABLED(CONFIG_MPTCP)
 	bool				drop_req;
 #endif
@@ -257,7 +258,8 @@ struct tcp_sock {
 	u8	compressed_ack;
 	u8	dup_ack_counter:2,
 		tlp_retrans:1,	/* TLP is a retransmission */
-		unused:5;
+		tcp_usec_ts:1, /* TSval values in usec */
+		unused:4;
 	u32	chrono_start;	/* Start time in jiffies of a TCP chrono */
 	u32	chrono_stat[3];	/* Time in jiffies for chrono_stat stats */
 	u8	chrono_type:2,	/* current chronograph type */
@@ -576,4 +578,9 @@ void tcp_sock_set_quickack(struct sock *sk, int val);
 int tcp_sock_set_syncnt(struct sock *sk, int val);
 int tcp_sock_set_user_timeout(struct sock *sk, int val);
 
+static inline bool dst_tcp_usec_ts(const struct dst_entry *dst)
+{
+	return dst_feature(dst, RTAX_FEATURE_TCP_USEC_TS);
+}
+
 #endif	/* _LINUX_TCP_H */
diff --git a/include/net/inet_timewait_sock.h b/include/net/inet_timewait_sock.h
index 4a8e578405cb..b14999ff55db 100644
--- a/include/net/inet_timewait_sock.h
+++ b/include/net/inet_timewait_sock.h
@@ -67,7 +67,8 @@ struct inet_timewait_sock {
 	/* And these are ours. */
 	unsigned int		tw_transparent  : 1,
 				tw_flowlabel	: 20,
-				tw_pad		: 3,	/* 3 bits hole */
+				tw_usec_ts	: 1,
+				tw_pad		: 2,	/* 2 bits hole */
 				tw_tos		: 8;
 	u32			tw_txhash;
 	u32			tw_priority;
diff --git a/include/net/tcp.h b/include/net/tcp.h
index bad304d173a5..39b731c900dd 100644
--- a/include/net/tcp.h
+++ b/include/net/tcp.h
@@ -166,7 +166,12 @@ static_assert((1 << ATO_BITS) > TCP_DELACK_MAX);
 #define MAX_TCP_KEEPCNT		127
 #define MAX_TCP_SYNCNT		127
 
-#define TCP_PAWS_24DAYS	(60 * 60 * 24 * 24)
+/* Ensure that TCP PAWS checks are relaxed after ~2147 seconds
+ * to avoid overflows. This assumes a clock smaller than 1 Mhz.
+ * Default clock is 1 Khz, tcp_usec_ts uses 1 Mhz.
+ */
+#define TCP_PAWS_WRAP (INT_MAX / USEC_PER_SEC)
+
 #define TCP_PAWS_MSL	60		/* Per-host timestamps are invalidated
 					 * after this time. It should be equal
 					 * (or greater than) TCP_TIMEWAIT_LEN
@@ -798,22 +803,31 @@ static inline u64 tcp_clock_us(void)
 	return div_u64(tcp_clock_ns(), NSEC_PER_USEC);
 }
 
-/* This should only be used in contexts where tp->tcp_mstamp is up to date */
-static inline u32 tcp_time_stamp(const struct tcp_sock *tp)
+static inline u64 tcp_clock_ms(void)
 {
-	return div_u64(tp->tcp_mstamp, USEC_PER_SEC / TCP_TS_HZ);
+	return div_u64(tcp_clock_ns(), NSEC_PER_MSEC);
 }
 
-/* Convert a nsec timestamp into TCP TSval timestamp (ms based currently) */
-static inline u32 tcp_ns_to_ts(u64 ns)
+/* TCP Timestamp included in TS option (RFC 1323) can either use ms
+ * or usec resolution. Each socket carries a flag to select one or other
+ * resolution, as the route attribute could change anytime.
+ * Each flow must stick to initial resolution.
+ */
+static inline u32 tcp_clock_ts(bool usec_ts)
 {
-	return div_u64(ns, NSEC_PER_SEC / TCP_TS_HZ);
+	return usec_ts ? tcp_clock_us() : tcp_clock_ms();
 }
 
-/* Could use tcp_clock_us() / 1000, but this version uses a single divide */
-static inline u32 tcp_time_stamp_raw(void)
+static inline u32 tcp_time_stamp_ms(const struct tcp_sock *tp)
 {
-	return tcp_ns_to_ts(tcp_clock_ns());
+	return div_u64(tp->tcp_mstamp, USEC_PER_MSEC);
+}
+
+static inline u32 tcp_time_stamp_ts(const struct tcp_sock *tp)
+{
+	if (tp->tcp_usec_ts)
+		return tp->tcp_mstamp;
+	return tcp_time_stamp_ms(tp);
 }
 
 void tcp_mstamp_refresh(struct tcp_sock *tp);
@@ -823,17 +837,30 @@ static inline u32 tcp_stamp_us_delta(u64 t1, u64 t0)
 	return max_t(s64, t1 - t0, 0);
 }
 
-static inline u32 tcp_skb_timestamp(const struct sk_buff *skb)
-{
-	return tcp_ns_to_ts(skb->skb_mstamp_ns);
-}
-
 /* provide the departure time in us unit */
 static inline u64 tcp_skb_timestamp_us(const struct sk_buff *skb)
 {
 	return div_u64(skb->skb_mstamp_ns, NSEC_PER_USEC);
 }
 
+/* Provide skb TSval in usec or ms unit */
+static inline u32 tcp_skb_timestamp_ts(bool usec_ts, const struct sk_buff *skb)
+{
+	if (usec_ts)
+		return tcp_skb_timestamp_us(skb);
+
+	return div_u64(skb->skb_mstamp_ns, NSEC_PER_MSEC);
+}
+
+static inline u32 tcp_tw_tsval(const struct tcp_timewait_sock *tcptw)
+{
+	return tcp_clock_ts(tcptw->tw_sk.tw_usec_ts) + tcptw->tw_ts_offset;
+}
+
+static inline u32 tcp_rsk_tsval(const struct tcp_request_sock *treq)
+{
+	return tcp_clock_ts(treq->req_usec_ts) + treq->ts_off;
+}
 
 #define tcp_flag_byte(th) (((u_int8_t *)th)[13])
 
@@ -1599,7 +1626,7 @@ static inline bool tcp_paws_check(const struct tcp_options_received *rx_opt,
 	if ((s32)(rx_opt->ts_recent - rx_opt->rcv_tsval) <= paws_win)
 		return true;
 	if (unlikely(!time_before32(ktime_get_seconds(),
-				    rx_opt->ts_recent_stamp + TCP_PAWS_24DAYS)))
+				    rx_opt->ts_recent_stamp + TCP_PAWS_WRAP)))
 		return true;
 	/*
 	 * Some OSes send SYN and SYNACK messages with tsval=0 tsecr=0,
diff --git a/include/uapi/linux/rtnetlink.h b/include/uapi/linux/rtnetlink.h
index 51c13cf9c5ae..aa2482a0614a 100644
--- a/include/uapi/linux/rtnetlink.h
+++ b/include/uapi/linux/rtnetlink.h
@@ -502,13 +502,17 @@ enum {
 
 #define RTAX_MAX (__RTAX_MAX - 1)
 
-#define RTAX_FEATURE_ECN	(1 << 0)
-#define RTAX_FEATURE_SACK	(1 << 1)
-#define RTAX_FEATURE_TIMESTAMP	(1 << 2)
-#define RTAX_FEATURE_ALLFRAG	(1 << 3)
-
-#define RTAX_FEATURE_MASK	(RTAX_FEATURE_ECN | RTAX_FEATURE_SACK | \
-				 RTAX_FEATURE_TIMESTAMP | RTAX_FEATURE_ALLFRAG)
+#define RTAX_FEATURE_ECN		(1 << 0)
+#define RTAX_FEATURE_SACK		(1 << 1) /* unused */
+#define RTAX_FEATURE_TIMESTAMP		(1 << 2) /* unused */
+#define RTAX_FEATURE_ALLFRAG		(1 << 3)
+#define RTAX_FEATURE_TCP_USEC_TS	(1 << 4)
+
+#define RTAX_FEATURE_MASK	(RTAX_FEATURE_ECN |		\
+				 RTAX_FEATURE_SACK |		\
+				 RTAX_FEATURE_TIMESTAMP |	\
+				 RTAX_FEATURE_ALLFRAG |		\
+				 RTAX_FEATURE_TCP_USEC_TS)
 
 struct rta_session {
 	__u8	proto;
diff --git a/include/uapi/linux/tcp.h b/include/uapi/linux/tcp.h
index d1d08da6331a..8aa3916e14f6 100644
--- a/include/uapi/linux/tcp.h
+++ b/include/uapi/linux/tcp.h
@@ -170,6 +170,7 @@ enum tcp_fastopen_client_fail {
 #define TCPI_OPT_ECN		8 /* ECN was negociated at TCP session init */
 #define TCPI_OPT_ECN_SEEN	16 /* we received at least one packet with ECT */
 #define TCPI_OPT_SYN_DATA	32 /* SYN-ACK acked data in SYN sent or rcvd */
+#define TCPI_OPT_USEC_TS	64 /* usec timestamps */
 
 /*
  * Sender's congestion state indicating normal or abnormal situations