Merge branch 'tcp-TCP-TS-option-use-1-ms-clock'

Eric Dumazet says:

====================
tcp: TCP TS option use 1 ms clock

TCP Timestamps option is defined in RFC 7323

Traditionally on linux, it has been tied to the internal
'jiffy' variable, because it had been a cheap and good enough
generator.

Unfortunately some distros use HZ=250 or even HZ=100 leading
to not very useful TCP timestamps.

For TCP flows in the DC, Google has used usec resolution for more
than two years with great success [1].
RCVBUF autotuning is more precise.

This series converts tp->tcp_mstamp to a plain u64 value storing
a 1 usec TCP clock.

This choice will allow us to upstream the 1 usec TS option as
discussed in IETF 97.

Kathleen Nichols [2] and others advocate for 1ms TS clocks for
network analysis. (1ms being the lowest value supported by RFC 7323.)

[1] https://www.ietf.org/proceedings/97/slides/slides-97-tcpm-tcp-options-for-low-latency-00.pdf
[2] http://netseminar.stanford.edu/seminars/02_02_17.pdf
====================

Signed-off-by: David S. Miller <[email protected]>

Author: davem330

include/linux/skbuff.h

@@ -506,66 +506,6 @@ typedef unsigned int sk_buff_data_t;
 typedef unsigned char *sk_buff_data_t;
 #endif
 
-/**
- * struct skb_mstamp - multi resolution time stamps
- * @stamp_us: timestamp in us resolution
- * @stamp_jiffies: timestamp in jiffies
- */
-struct skb_mstamp {
-	union {
-		u64		v64;
-		struct {
-			u32	stamp_us;
-			u32	stamp_jiffies;
-		};
-	};
-};
-
-/**
- * skb_mstamp_get - get current timestamp
- * @cl: place to store timestamps
- */
-static inline void skb_mstamp_get(struct skb_mstamp *cl)
-{
-	u64 val = local_clock();
-
-	do_div(val, NSEC_PER_USEC);
-	cl->stamp_us = (u32)val;
-	cl->stamp_jiffies = (u32)jiffies;
-}
-
-/**
- * skb_mstamp_delta - compute the difference in usec between two skb_mstamp
- * @t1: pointer to newest sample
- * @t0: pointer to oldest sample
- */
-static inline u32 skb_mstamp_us_delta(const struct skb_mstamp *t1,
-				      const struct skb_mstamp *t0)
-{
-	s32 delta_us = t1->stamp_us - t0->stamp_us;
-	u32 delta_jiffies = t1->stamp_jiffies - t0->stamp_jiffies;
-
-	/* If delta_us is negative, this might be because interval is too big,
-	 * or local_clock() drift is too big : fallback using jiffies.
-	 */
-	if (delta_us <= 0 ||
-	    delta_jiffies >= (INT_MAX / (USEC_PER_SEC / HZ)))
-
-		delta_us = jiffies_to_usecs(delta_jiffies);
-
-	return delta_us;
-}
-
-static inline bool skb_mstamp_after(const struct skb_mstamp *t1,
-				    const struct skb_mstamp *t0)
-{
-	s32 diff = t1->stamp_jiffies - t0->stamp_jiffies;
-
-	if (!diff)
-		diff = t1->stamp_us - t0->stamp_us;
-	return diff > 0;
-}
-
 /** 
  *	struct sk_buff - socket buffer
  *	@next: Next buffer in list
@@ -646,7 +586,7 @@ struct sk_buff {
 
 			union {
 				ktime_t		tstamp;
-				struct skb_mstamp skb_mstamp;
+				u64		skb_mstamp;
 			};
 		};
 		struct rb_node	rbnode; /* used in netem & tcp stack */

include/linux/tcp.h

@@ -123,7 +123,7 @@ struct tcp_request_sock_ops;
 struct tcp_request_sock {
 	struct inet_request_sock 	req;
 	const struct tcp_request_sock_ops *af_specific;
-	struct skb_mstamp		snt_synack; /* first SYNACK sent time */
+	u64				snt_synack; /* first SYNACK sent time */
 	bool				tfo_listener;
 	u32				txhash;
 	u32				rcv_isn;
@@ -211,7 +211,7 @@ struct tcp_sock {
 
 	/* Information of the most recently (s)acked skb */
 	struct tcp_rack {
-		struct skb_mstamp mstamp; /* (Re)sent time of the skb */
+		u64 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 */
@@ -240,7 +240,7 @@ struct tcp_sock {
 	u32	tlp_high_seq;	/* snd_nxt at the time of TLP retransmit. */
 
 /* RTT measurement */
-	struct skb_mstamp tcp_mstamp; /* most recent packet received/sent */
+	u64	tcp_mstamp;	/* most recent packet received/sent */
 	u32	srtt_us;	/* smoothed round trip time << 3 in usecs */
 	u32	mdev_us;	/* medium deviation			*/
 	u32	mdev_max_us;	/* maximal mdev for the last rtt period	*/
@@ -280,8 +280,8 @@ struct tcp_sock {
 	u32	delivered;	/* Total data packets delivered incl. rexmits */
 	u32	lost;		/* Total data packets lost incl. rexmits */
 	u32	app_limited;	/* limited until "delivered" reaches this val */
-	struct skb_mstamp first_tx_mstamp;  /* start of window send phase */
-	struct skb_mstamp delivered_mstamp; /* time we reached "delivered" */
+	u64	first_tx_mstamp;  /* start of window send phase */
+	u64	delivered_mstamp; /* time we reached "delivered" */
 	u32	rate_delivered;    /* saved rate sample: packets delivered */
 	u32	rate_interval_us;  /* saved rate sample: time elapsed */
 
@@ -335,16 +335,16 @@ struct tcp_sock {
 
 /* Receiver side RTT estimation */
 	struct {
-		u32		rtt_us;
-		u32		seq;
-		struct skb_mstamp time;
+		u32	rtt_us;
+		u32	seq;
+		u64	time;
 	} rcv_rtt_est;
 
 /* Receiver queue space */
 	struct {
-		int		space;
-		u32		seq;
-		struct skb_mstamp time;
+		int	space;
+		u32	seq;
+		u64	time;
 	} rcvq_space;
 
 /* TCP-specific MTU probe information. */

include/net/tcp.h

@@ -519,7 +519,7 @@ static inline u32 tcp_cookie_time(void)
 u32 __cookie_v4_init_sequence(const struct iphdr *iph, const struct tcphdr *th,
 			      u16 *mssp);
 __u32 cookie_v4_init_sequence(const struct sk_buff *skb, __u16 *mss);
-__u32 cookie_init_timestamp(struct request_sock *req);
+u64 cookie_init_timestamp(struct request_sock *req);
 bool cookie_timestamp_decode(struct tcp_options_received *opt);
 bool cookie_ecn_ok(const struct tcp_options_received *opt,
 		   const struct net *net, const struct dst_entry *dst);
@@ -700,17 +700,61 @@ u32 __tcp_select_window(struct sock *sk);
 
 void tcp_send_window_probe(struct sock *sk);
 
-/* TCP timestamps are only 32-bits, this causes a slight
- * complication on 64-bit systems since we store a snapshot
- * of jiffies in the buffer control blocks below.  We decided
- * to use only the low 32-bits of jiffies and hide the ugly
- * casts with the following macro.
+/* TCP uses 32bit jiffies to save some space.
+ * Note that this is different from tcp_time_stamp, which
+ * historically has been the same until linux-4.13.
  */
-#define tcp_time_stamp		((__u32)(jiffies))
+#define tcp_jiffies32 ((u32)jiffies)
+
+/*
+ * Deliver a 32bit value for TCP timestamp option (RFC 7323)
+ * It is no longer tied to jiffies, but to 1 ms clock.
+ * Note: double check if you want to use tcp_jiffies32 instead of this.
+ */
+#define TCP_TS_HZ	1000
+
+static inline u64 tcp_clock_ns(void)
+{
+	return local_clock();
+}
+
+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)
+{
+	return div_u64(tp->tcp_mstamp, USEC_PER_SEC / TCP_TS_HZ);
+}
+
+/* Could use tcp_clock_us() / 1000, but this version uses a single divide */
+static inline u32 tcp_time_stamp_raw(void)
+{
+	return div_u64(tcp_clock_ns(), NSEC_PER_SEC / TCP_TS_HZ);
+}
+
+
+/* Refresh 1us clock of a TCP socket,
+ * ensuring monotically increasing values.
+ */
+static inline void tcp_mstamp_refresh(struct tcp_sock *tp)
+{
+	u64 val = tcp_clock_us();
+
+	if (val > tp->tcp_mstamp)
+		tp->tcp_mstamp = val;
+}
+
+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 skb->skb_mstamp.stamp_jiffies;
+	return div_u64(skb->skb_mstamp, USEC_PER_SEC / TCP_TS_HZ);
 }
 
 
@@ -775,9 +819,9 @@ struct tcp_skb_cb {
 			/* pkts S/ACKed so far upon tx of skb, incl retrans: */
 			__u32 delivered;
 			/* start of send pipeline phase */
-			struct skb_mstamp first_tx_mstamp;
+			u64 first_tx_mstamp;
 			/* when we reached the "delivered" count */
-			struct skb_mstamp delivered_mstamp;
+			u64 delivered_mstamp;
 		} tx;   /* only used for outgoing skbs */
 		union {
 			struct inet_skb_parm	h4;
@@ -893,7 +937,7 @@ struct ack_sample {
  * A sample is invalid if "delivered" or "interval_us" is negative.
  */
 struct rate_sample {
-	struct	skb_mstamp prior_mstamp; /* starting timestamp for interval */
+	u64  prior_mstamp; /* starting timestamp for interval */
 	u32  prior_delivered;	/* tp->delivered at "prior_mstamp" */
 	s32  delivered;		/* number of packets delivered over interval */
 	long interval_us;	/* time for tp->delivered to incr "delivered" */
@@ -1242,7 +1286,7 @@ static inline void tcp_slow_start_after_idle_check(struct sock *sk)
 	if (!sysctl_tcp_slow_start_after_idle || tp->packets_out ||
 	    ca_ops->cong_control)
 		return;
-	delta = tcp_time_stamp - tp->lsndtime;
+	delta = tcp_jiffies32 - tp->lsndtime;
 	if (delta > inet_csk(sk)->icsk_rto)
 		tcp_cwnd_restart(sk, delta);
 }
@@ -1304,8 +1348,8 @@ static inline u32 keepalive_time_elapsed(const struct tcp_sock *tp)
 {
 	const struct inet_connection_sock *icsk = &tp->inet_conn;
 
-	return min_t(u32, tcp_time_stamp - icsk->icsk_ack.lrcvtime,
-			  tcp_time_stamp - tp->rcv_tstamp);
+	return min_t(u32, tcp_jiffies32 - icsk->icsk_ack.lrcvtime,
+			  tcp_jiffies32 - tp->rcv_tstamp);
 }
 
 static inline int tcp_fin_time(const struct sock *sk)
@@ -1859,7 +1903,7 @@ void tcp_init(void);
 /* tcp_recovery.c */
 extern void tcp_rack_mark_lost(struct sock *sk);
 extern void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
-			     const struct skb_mstamp *xmit_time);
+			     u64 xmit_time);
 extern void tcp_rack_reo_timeout(struct sock *sk);
 
 /*

net/dccp/ccids/ccid2.c

@@ -233,7 +233,7 @@ static void ccid2_hc_tx_packet_sent(struct sock *sk, unsigned int len)
 {
 	struct dccp_sock *dp = dccp_sk(sk);
 	struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk);
-	const u32 now = ccid2_time_stamp;
+	const u32 now = ccid2_jiffies32;
 	struct ccid2_seq *next;
 
 	/* slow-start after idle periods (RFC 2581, RFC 2861) */
@@ -466,7 +466,7 @@ static void ccid2_new_ack(struct sock *sk, struct ccid2_seq *seqp,
 	 * The cleanest solution is to not use the ccid2s_sent field at all
 	 * and instead use DCCP timestamps: requires changes in other places.
 	 */
-	ccid2_rtt_estimator(sk, ccid2_time_stamp - seqp->ccid2s_sent);
+	ccid2_rtt_estimator(sk, ccid2_jiffies32 - seqp->ccid2s_sent);
 }
 
 static void ccid2_congestion_event(struct sock *sk, struct ccid2_seq *seqp)
@@ -478,7 +478,7 @@ static void ccid2_congestion_event(struct sock *sk, struct ccid2_seq *seqp)
 		return;
 	}
 
-	hc->tx_last_cong = ccid2_time_stamp;
+	hc->tx_last_cong = ccid2_jiffies32;
 
 	hc->tx_cwnd      = hc->tx_cwnd / 2 ? : 1U;
 	hc->tx_ssthresh  = max(hc->tx_cwnd, 2U);
@@ -731,7 +731,7 @@ static int ccid2_hc_tx_init(struct ccid *ccid, struct sock *sk)
 
 	hc->tx_rto	 = DCCP_TIMEOUT_INIT;
 	hc->tx_rpdupack  = -1;
-	hc->tx_last_cong = hc->tx_lsndtime = hc->tx_cwnd_stamp = ccid2_time_stamp;
+	hc->tx_last_cong = hc->tx_lsndtime = hc->tx_cwnd_stamp = ccid2_jiffies32;
 	hc->tx_cwnd_used = 0;
 	setup_timer(&hc->tx_rtotimer, ccid2_hc_tx_rto_expire,
 			(unsigned long)sk);

net/dccp/ccids/ccid2.h

@@ -27,7 +27,7 @@
  * CCID-2 timestamping faces the same issues as TCP timestamping.
  * Hence we reuse/share as much of the code as possible.
  */
-#define ccid2_time_stamp	tcp_time_stamp
+#define ccid2_jiffies32	((u32)jiffies)
 
 /* NUMDUPACK parameter from RFC 4341, p. 6 */
 #define NUMDUPACK	3

net/ipv4/syncookies.c

@@ -66,10 +66,10 @@ static u32 cookie_hash(__be32 saddr, __be32 daddr, __be16 sport, __be16 dport,
  * Since subsequent timestamps use the normal tcp_time_stamp value, we
  * must make sure that the resulting initial timestamp is <= tcp_time_stamp.
  */
-__u32 cookie_init_timestamp(struct request_sock *req)
+u64 cookie_init_timestamp(struct request_sock *req)
 {
 	struct inet_request_sock *ireq;
-	u32 ts, ts_now = tcp_time_stamp;
+	u32 ts, ts_now = tcp_time_stamp_raw();
 	u32 options = 0;
 
 	ireq = inet_rsk(req);
@@ -88,7 +88,7 @@ __u32 cookie_init_timestamp(struct request_sock *req)
 		ts <<= TSBITS;
 		ts |= options;
 	}
-	return ts;
+	return (u64)ts * (USEC_PER_SEC / TCP_TS_HZ);
 }
 
 
@@ -343,7 +343,7 @@ struct sock *cookie_v4_check(struct sock *sk, struct sk_buff *skb)
 	ireq->wscale_ok		= tcp_opt.wscale_ok;
 	ireq->tstamp_ok		= tcp_opt.saw_tstamp;
 	req->ts_recent		= tcp_opt.saw_tstamp ? tcp_opt.rcv_tsval : 0;
-	treq->snt_synack.v64	= 0;
+	treq->snt_synack	= 0;
 	treq->tfo_listener	= false;
 
 	ireq->ir_iif = inet_request_bound_dev_if(sk, skb);

net/ipv4/tcp.c

@@ -386,7 +386,7 @@ void tcp_init_sock(struct sock *sk)
 
 	icsk->icsk_rto = TCP_TIMEOUT_INIT;
 	tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
-	minmax_reset(&tp->rtt_min, tcp_time_stamp, ~0U);
+	minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
 
 	/* So many TCP implementations out there (incorrectly) count the
 	 * initial SYN frame in their delayed-ACK and congestion control
@@ -2706,7 +2706,7 @@ static int do_tcp_setsockopt(struct sock *sk, int level,
 		if (!tp->repair)
 			err = -EPERM;
 		else
-			tp->tsoffset = val - tcp_time_stamp;
+			tp->tsoffset = val - tcp_time_stamp_raw();
 		break;
 	case TCP_REPAIR_WINDOW:
 		err = tcp_repair_set_window(tp, optval, optlen);
@@ -2757,7 +2757,7 @@ static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
 	for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
 		stats[i] = tp->chrono_stat[i - 1];
 		if (i == tp->chrono_type)
-			stats[i] += tcp_time_stamp - tp->chrono_start;
+			stats[i] += tcp_jiffies32 - tp->chrono_start;
 		stats[i] *= USEC_PER_SEC / HZ;
 		total += stats[i];
 	}
@@ -2841,7 +2841,7 @@ void tcp_get_info(struct sock *sk, struct tcp_info *info)
 	info->tcpi_retrans = tp->retrans_out;
 	info->tcpi_fackets = tp->fackets_out;
 
-	now = tcp_time_stamp;
+	now = tcp_jiffies32;
 	info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
 	info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
 	info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
@@ -3072,7 +3072,7 @@ static int do_tcp_getsockopt(struct sock *sk, int level,
 		break;
 
 	case TCP_TIMESTAMP:
-		val = tcp_time_stamp + tp->tsoffset;
+		val = tcp_time_stamp_raw() + tp->tsoffset;
 		break;
 	case TCP_NOTSENT_LOWAT:
 		val = tp->notsent_lowat;