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>

drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c

@@ -2259,7 +2259,7 @@ static void chtls_rx_ack(struct sock *sk, struct sk_buff *skb)
 
 		if (tp->snd_una != snd_una) {
 			tp->snd_una = snd_una;
-			tp->rcv_tstamp = tcp_time_stamp(tp);
+			tp->rcv_tstamp = tcp_jiffies32;
 			if (tp->snd_una == tp->snd_nxt &&
 			    !csk_flag_nochk(csk, CSK_TX_FAILOVER))
 				csk_reset_flag(csk, CSK_TX_WAIT_IDLE);

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 */

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;

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,

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;

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

net/ipv4/syncookies.c

@@ -41,7 +41,6 @@ static siphash_aligned_key_t syncookie_secret[2];
  * requested/supported by the syn/synack exchange.
  */
 #define TSBITS	6
-#define TSMASK	(((__u32)1 << TSBITS) - 1)
 
 static u32 cookie_hash(__be32 saddr, __be32 daddr, __be16 sport, __be16 dport,
 		       u32 count, int c)
@@ -52,6 +51,14 @@ static u32 cookie_hash(__be32 saddr, __be32 daddr, __be16 sport, __be16 dport,
 			    count, &syncookie_secret[c]);
 }
 
+/* Convert one nsec 64bit timestamp to ts (ms or usec resolution) */
+static u64 tcp_ns_to_ts(bool usec_ts, u64 val)
+{
+	if (usec_ts)
+		return div_u64(val, NSEC_PER_USEC);
+
+	return div_u64(val, NSEC_PER_MSEC);
+}
 
 /*
  * when syncookies are in effect and tcp timestamps are enabled we encode
@@ -62,27 +69,24 @@ static u32 cookie_hash(__be32 saddr, __be32 daddr, __be16 sport, __be16 dport,
  */
 u64 cookie_init_timestamp(struct request_sock *req, u64 now)
 {
-	struct inet_request_sock *ireq;
-	u32 ts, ts_now = tcp_ns_to_ts(now);
+	const struct inet_request_sock *ireq = inet_rsk(req);
+	u64 ts, ts_now = tcp_ns_to_ts(false, now);
 	u32 options = 0;
 
-	ireq = inet_rsk(req);
-
 	options = ireq->wscale_ok ? ireq->snd_wscale : TS_OPT_WSCALE_MASK;
 	if (ireq->sack_ok)
 		options |= TS_OPT_SACK;
 	if (ireq->ecn_ok)
 		options |= TS_OPT_ECN;
 
-	ts = ts_now & ~TSMASK;
+	ts = (ts_now >> TSBITS) << TSBITS;
 	ts |= options;
-	if (ts > ts_now) {
-		ts >>= TSBITS;
-		ts--;
-		ts <<= TSBITS;
-		ts |= options;
-	}
-	return (u64)ts * (NSEC_PER_SEC / TCP_TS_HZ);
+	if (ts > ts_now)
+		ts -= (1UL << TSBITS);
+
+	if (tcp_rsk(req)->req_usec_ts)
+		return ts * NSEC_PER_USEC;
+	return ts * NSEC_PER_MSEC;
 }
 
 
@@ -302,6 +306,8 @@ struct request_sock *cookie_tcp_reqsk_alloc(const struct request_sock_ops *ops,
 	treq->af_specific = af_ops;
 
 	treq->syn_tos = TCP_SKB_CB(skb)->ip_dsfield;
+	treq->req_usec_ts = -1;
+
 #if IS_ENABLED(CONFIG_MPTCP)
 	treq->is_mptcp = sk_is_mptcp(sk);
 	if (treq->is_mptcp) {

net/ipv4/tcp.c

@@ -3629,10 +3629,16 @@ int do_tcp_setsockopt(struct sock *sk, int level, int optname,
 			tp->fastopen_no_cookie = val;
 		break;
 	case TCP_TIMESTAMP:
-		if (!tp->repair)
+		if (!tp->repair) {
 			err = -EPERM;
-		else
-			WRITE_ONCE(tp->tsoffset, val - tcp_time_stamp_raw());
+			break;
+		}
+		/* val is an opaque field,
+		 * and low order bit contains usec_ts enable bit.
+		 * Its a best effort, and we do not care if user makes an error.
+		 */
+		tp->tcp_usec_ts = val & 1;
+		WRITE_ONCE(tp->tsoffset, val - tcp_clock_ts(tp->tcp_usec_ts));
 		break;
 	case TCP_REPAIR_WINDOW:
 		err = tcp_repair_set_window(tp, optval, optlen);
@@ -3754,6 +3760,8 @@ void tcp_get_info(struct sock *sk, struct tcp_info *info)
 		info->tcpi_options |= TCPI_OPT_ECN_SEEN;
 	if (tp->syn_data_acked)
 		info->tcpi_options |= TCPI_OPT_SYN_DATA;
+	if (tp->tcp_usec_ts)
+		info->tcpi_options |= TCPI_OPT_USEC_TS;
 
 	info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
 	info->tcpi_ato = jiffies_to_usecs(min_t(u32, icsk->icsk_ack.ato,
@@ -3817,10 +3825,8 @@ void tcp_get_info(struct sock *sk, struct tcp_info *info)
 	info->tcpi_total_rto = tp->total_rto;
 	info->tcpi_total_rto_recoveries = tp->total_rto_recoveries;
 	info->tcpi_total_rto_time = tp->total_rto_time;
-	if (tp->rto_stamp) {
-		info->tcpi_total_rto_time += tcp_time_stamp_raw() -
-						tp->rto_stamp;
-	}
+	if (tp->rto_stamp)
+		info->tcpi_total_rto_time += tcp_clock_ms() - tp->rto_stamp;
 
 	unlock_sock_fast(sk, slow);
 }
@@ -4145,7 +4151,11 @@ int do_tcp_getsockopt(struct sock *sk, int level,
 		break;
 
 	case TCP_TIMESTAMP:
-		val = tcp_time_stamp_raw() + READ_ONCE(tp->tsoffset);
+		val = tcp_clock_ts(tp->tcp_usec_ts) + READ_ONCE(tp->tsoffset);
+		if (tp->tcp_usec_ts)
+			val |= 1;
+		else
+			val &= ~1;
 		break;
 	case TCP_NOTSENT_LOWAT:
 		val = READ_ONCE(tp->notsent_lowat);