Skip to contentSkip to conversation

simplify microsecond rtt sampling

DISCUSSION John Heffner 3 participants 6 messages Sep 27, 2006 – Sep 28, 2006

Subsystems (3)
Networking Network drivers TCP
lore
Share & tools
lore

John Heffner

author
#0 lore raw

About commit 2d2abbab63f6726a147ae61ada39bf2c9ee0db9a:

It looks like this patch bypassed the enforcement of Karn's algorithm in tcp_ack_no_tstamp() for the purposes of usec RTT sampling used by congestion control modules. This will give them bad RTT data when there are retransmits. I haven't actually observed this, but it seems like it would be the case. ;) Please correct me if I'm wrong.

Here's a patch that should be a fix.

Signed-off-by John Heffner <[email protected]>
net/ipv4/tcp_input.c +2 -4
index 159fa3f..725c868 100644
@@ -2306,8 +2306,6 @@ static int tcp_clean_rtx_queue(struct so
 				seq_rtt = -1;
 			} else if (seq_rtt < 0) {
 				seq_rtt = now - scb->when;
-				if (rtt_sample)
-					(*rtt_sample)(sk, tcp_usrtt(skb));
 			}
 			if (sacked & TCPCB_SACKED_ACKED)
 				tp->sacked_out -= tcp_skb_pcount(skb);

@@ -2320,8 +2318,6 @@ static int tcp_clean_rtx_queue(struct so
 			}
 		} else if (seq_rtt < 0) {
 			seq_rtt = now - scb->when;
-			if (rtt_sample)
-				(*rtt_sample)(sk, tcp_usrtt(skb));
 		}
 		tcp_dec_pcount_approx(&tp->fackets_out, skb);
 		tcp_packets_out_dec(tp, skb);

@@ -2333,6 +2329,8 @@ static int tcp_clean_rtx_queue(struct so
 	if (acked&FLAG_ACKED) {
 		tcp_ack_update_rtt(sk, acked, seq_rtt);
 		tcp_ack_packets_out(sk, tp);
+		if (rtt_sample && !(acked & FLAG_RETRANS_DATA_ACKED))
+			(*rtt_sample)(sk, tcp_usrtt(skb));
 
 		if (icsk->icsk_ca_ops->pkts_acked)
 			icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
author
#1 lore raw

Okay, this patch is junk (never trust compile-tested code). Will send something better soon.

-John

Quoted 44 lines · John Heffner

About commit 2d2abbab63f6726a147ae61ada39bf2c9ee0db9a:

It looks like this patch bypassed the enforcement of Karn's algorithm in tcp_ack_no_tstamp() for the purposes of usec RTT sampling used by congestion control modules. This will give them bad RTT data when there are retransmits. I haven't actually observed this, but it seems like it would be the case. ;) Please correct me if I'm wrong.

Here's a patch that should be a fix.

Signed-off-by John Heffner <[email protected]>

------------------------------------------------------------------------

net/ipv4/tcp_input.c +2 -4
index 159fa3f..725c868 100644
@@ -2306,8 +2306,6 @@ static int tcp_clean_rtx_queue(struct so
 				seq_rtt = -1;
 			} else if (seq_rtt < 0) {
 				seq_rtt = now - scb->when;
-				if (rtt_sample)
-					(*rtt_sample)(sk, tcp_usrtt(skb));
 			}
 			if (sacked & TCPCB_SACKED_ACKED)
 				tp->sacked_out -= tcp_skb_pcount(skb);

@@ -2320,8 +2318,6 @@ static int tcp_clean_rtx_queue(struct so
 			}
 		} else if (seq_rtt < 0) {
 			seq_rtt = now - scb->when;
-			if (rtt_sample)
-				(*rtt_sample)(sk, tcp_usrtt(skb));
 		}
 		tcp_dec_pcount_approx(&tp->fackets_out, skb);
 		tcp_packets_out_dec(tp, skb);

@@ -2333,6 +2329,8 @@ static int tcp_clean_rtx_queue(struct so
 	if (acked&FLAG_ACKED) {
 		tcp_ack_update_rtt(sk, acked, seq_rtt);
 		tcp_ack_packets_out(sk, tp);
+		if (rtt_sample && !(acked & FLAG_RETRANS_DATA_ACKED))
+			(*rtt_sample)(sk, tcp_usrtt(skb));
 
 		if (icsk->icsk_ca_ops->pkts_acked)
 			icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
author
#2 lore raw

Here is a corrected patch.

Signed-off-by John Heffner <[email protected]>

-static u32 tcp_usrtt(const struct sk_buff *skb)
+static u32 tcp_usrtt(struct timeval *tv)
{
- struct timeval tv, now;
+ struct timeval now;

do_gettimeofday(&now);
- skb_get_timestamp(skb, &tv);
- return (now.tv_sec - tv.tv_sec) * 1000000 + (now.tv_usec - tv.tv_usec);
+ return (now.tv_sec - tv->tv_sec) * 1000000 + (now.tv_usec - tv->tv_usec);
}

/* Remove acknowledged frames from the retransmission queue. */
@@ -2249,6 +2248,7 @@ static int tcp_clean_rtx_queue(struct so
u32 pkts_acked = 0;
void (*rtt_sample)(struct sock *sk, u32 usrtt)
= icsk->icsk_ca_ops->rtt_sample;
+ struct timeval tv;

while ((skb = skb_peek(&sk->sk_write_queue)) &&
skb != sk->sk_send_head) {
@@ -2297,8 +2297,7 @@ static int tcp_clean_rtx_queue(struct so
seq_rtt = -1;
} else if (seq_rtt < 0) {
seq_rtt = now - scb->when;
- if (rtt_sample)
- (*rtt_sample)(sk, tcp_usrtt(skb));
+ skb_get_timestamp(skb, &tv);
}
if (sacked & TCPCB_SACKED_ACKED)
tp->sacked_out -= tcp_skb_pcount(skb);
@@ -2311,8 +2310,7 @@ static int tcp_clean_rtx_queue(struct so
}
} else if (seq_rtt < 0) {
seq_rtt = now - scb->when;
- if (rtt_sample)
- (*rtt_sample)(sk, tcp_usrtt(skb));
+ skb_get_timestamp(skb, &tv);
}
tcp_dec_pcount_approx(&tp->fackets_out, skb); tcp_packets_out_dec(tp, skb);
@@ -2324,6 +2322,8 @@ static int tcp_clean_rtx_queue(struct so
if (acked&FLAG_ACKED) {
tcp_ack_update_rtt(sk, acked, seq_rtt); tcp_ack_packets_out(sk, tp);
+ if (rtt_sample && !(acked & FLAG_RETRANS_DATA_ACKED))
+ (*rtt_sample)(sk, tcp_usrtt(&tv));

if (icsk->icsk_ca_ops->pkts_acked)
icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);

author
#3 lore raw

Sigh. Here's one that's not corrupted. And for the record:

This changes the microsecond RTT sampling so that samples are taken in the same way that RTT samples are taken for the RTO calculator: on the last segment acknowledged, and only when the segment hasn't been retransmitted.

Signed-off-by John Heffner <[email protected]>
net/ipv4/tcp_input.c +8 -8
index b5521a9..d0f6bd6 100644
@@ -2228,13 +2228,12 @@ static int tcp_tso_acked(struct sock *sk
 	return acked;
 }
 
-static u32 tcp_usrtt(const struct sk_buff *skb)
+static u32 tcp_usrtt(struct timeval *tv)
 {
-	struct timeval tv, now;
+	struct timeval now;
 
 	do_gettimeofday(&now);
-	skb_get_timestamp(skb, &tv);
-	return (now.tv_sec - tv.tv_sec) * 1000000 + (now.tv_usec - tv.tv_usec);
+	return (now.tv_sec - tv->tv_sec) * 1000000 + (now.tv_usec - tv->tv_usec);
 }
 
 /* Remove acknowledged frames from the retransmission queue. */

@@ -2249,6 +2248,7 @@ static int tcp_clean_rtx_queue(struct so
 	u32 pkts_acked = 0;
 	void (*rtt_sample)(struct sock *sk, u32 usrtt)
 		= icsk->icsk_ca_ops->rtt_sample;
+	struct timeval tv;
 
 	while ((skb = skb_peek(&sk->sk_write_queue)) &&
 	       skb != sk->sk_send_head) {

@@ -2297,8 +2297,7 @@ static int tcp_clean_rtx_queue(struct so
 				seq_rtt = -1;
 			} else if (seq_rtt < 0) {
 				seq_rtt = now - scb->when;
-				if (rtt_sample)
-					(*rtt_sample)(sk, tcp_usrtt(skb));
+				skb_get_timestamp(skb, &tv);
 			}
 			if (sacked & TCPCB_SACKED_ACKED)
 				tp->sacked_out -= tcp_skb_pcount(skb);

@@ -2311,8 +2310,7 @@ static int tcp_clean_rtx_queue(struct so
 			}
 		} else if (seq_rtt < 0) {
 			seq_rtt = now - scb->when;
-			if (rtt_sample)
-				(*rtt_sample)(sk, tcp_usrtt(skb));
+			skb_get_timestamp(skb, &tv);
 		}
 		tcp_dec_pcount_approx(&tp->fackets_out, skb);
 		tcp_packets_out_dec(tp, skb);

@@ -2324,6 +2322,8 @@ static int tcp_clean_rtx_queue(struct so
 	if (acked&FLAG_ACKED) {
 		tcp_ack_update_rtt(sk, acked, seq_rtt);
 		tcp_ack_packets_out(sk, tp);
+		if (rtt_sample && !(acked & FLAG_RETRANS_DATA_ACKED))
+			(*rtt_sample)(sk, tcp_usrtt(&tv));
 
 		if (icsk->icsk_ca_ops->pkts_acked)
 			icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked);
#4 lore raw
Quoted 10 lines · On Thu, 28 Sep 2006 13:47:28 +0100 John Heffner <[email protected]> wrote:

Sigh. Here's one that's not corrupted. And for the record:

This changes the microsecond RTT sampling so that samples are taken in the same way that RTT samples are taken for the RTO calculator: on the last segment acknowledged, and only when the segment hasn't been retransmitted.

Signed-off-by John Heffner <[email protected]>

Looks ok. I wish skbuff timestamp was converted to ktime.

maintainer
#5 lore raw

From: Stephen Hemminger <[email protected]>
Date: Thu, 28 Sep 2006 09:36:50 -0700

Quoted 12 lines
Quoted 9 lines · On Thu, 28 Sep 2006 13:47:28 +0100 John Heffner <[email protected]> wrote:

Sigh. Here's one that's not corrupted. And for the record:

This changes the microsecond RTT sampling so that samples are taken in the same way that RTT samples are taken for the RTO calculator: on the last segment acknowledged, and only when the segment hasn't been retransmitted.

Quoted 3 lines
Quoted 1 line
Signed-off-by John Heffner <[email protected]>

Looks ok. I wish skbuff timestamp was converted to ktime.

It looks good to me too, applied.

Since this is really a bug fix, I'm goint to push this into -stable as well.

Thanks a lot.