: Fix Request/Response RTT sampling

[DCCP]: Fix Request/Response RTT sampling

This fixes a problem with the Request/Response RTT sampling:

The current timestamp is taken after the options are processed, which is too late if there is anything in the option processing code which takes more than a trivial amount of time.

One instance is feature negotiation - I found that a full run through all the option negotiation can take up to 38 ms. (Whether this says something about feature negotiation or its implementation is a separate issue.)

The real problem (and it was caused by me) is that one does not know how much time is spent in processing the options - and since DCCP has space for up to 1020 bytes of options, the situation can be even much worse.

Therefore and since the Linux implementation always sends the required Elapsed Time option, the timestamp is taken when the function is called, not later.

---
 net/dccp/input.c |   12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)

@@ -329,6 +329,7 @@ static int dccp_rcv_request_sent_state_p
 	if (dh->dccph_type = DCCP_PKT_RESPONSE) {
 		const struct inet_connection_sock *icsk = inet_csk(sk);
 		struct dccp_sock *dp = dccp_sk(sk);
+		long tstamp = dccp_timestamp();
 
 		/* Stop the REQUEST timer */
 		inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);

@@ -349,13 +350,10 @@ static int dccp_rcv_request_sent_state_p
 		if (dccp_parse_options(sk, skb))
 			goto out_invalid_packet;
 
-		/* Obtain RTT sample from SYN exchange (used by CCID 3) */
-		if (dp->dccps_options_received.dccpor_timestamp_echo) {
-			long d = dccp_timestamp();
-
-			d -= dp->dccps_options_received.dccpor_timestamp_echo;
-			dp->dccps_syn_rtt = dccp_sample_rtt(sk, 10 * d);
-		}
+		/* Obtain usec RTT sample from SYN exchange (used by CCID 3) */
+		if (likely(dp->dccps_options_received.dccpor_timestamp_echo))
+			dp->dccps_syn_rtt = dccp_sample_rtt(sk, 10 * (tstamp -
+			    dp->dccps_options_received.dccpor_timestamp_echo));
 
 		if (dccp_msk(sk)->dccpms_send_ack_vector &&
 		    dccp_ackvec_add(dp->dccps_hc_rx_ackvec, sk,

Quoting Ian McDonald:

|  On 9/13/07, Gerrit Renker <[email protected]> wrote:
|  > [DCCP]: Fix Request/Response RTT sampling
|  >

Thanks for looking at this, but I won't/can't use the sign-off, since I merged this patch with the one which had the error fixed by this patch.

It is attached below, in the test tree the patch is located as follows at the bottom of the patch stack; the order of patches is:

#0: Complete patch of Arnaldo's conversion to ktime_t (combined from the
patches sent to this list)

#1: [DCCP] Reuse ktime_get_real() calls again
As suggested, reduces the number of required timestamps - for CCID3.

#2: [DCCP]: Provide 10s of microsecond timesource
Separates the ktime_get_real() call into a wrapper function so that it can be called whenever a timestamp is needed; 10s of usec resolution, and seeded at boot/module load time (wraparound after 11.9 hours).

#3: [DCCP]: Simplify interface of dccp_sample_rtt
The patch fixed by this (5/5) one; the merged result is shown below.

---------------------> Updated Patch #3, fixed by 5/5 <----------------------- [DCCP]: Simplify interface of dccp_sample_rtt

The third parameter of dccp_sample_rtt now becomes useless and is removed.

Also combined the subtraction of the timestamp echo and the elapsed time. This is safe, since (a) presence of timestamp echo is tested first and (b) elapsed time is either present and non-zero or it is not set and equals 0 due to the memset in dccp_parse_options.

To avoid measuring option-processing time, the timestamp for measuring the initial Request/Response RTT sample is taken directly when the function is called (the Linux implementation always adds a timestamp on the Request, so there is no loss in doing this).

---
 net/dccp/ccids/ccid3.c |    6 +++---
 net/dccp/dccp.h        |    3 +--
 net/dccp/input.c       |   44 +++++++++++++++-----------------------------
 3 files changed, 19 insertions(+), 34 deletions(-)

@@ -297,8 +297,7 @@ extern int	   dccp_v4_connect(struct soc
 extern int	   dccp_send_reset(struct sock *sk, enum dccp_reset_codes code);
 extern void	   dccp_send_close(struct sock *sk, const int active);
 extern int	   dccp_invalid_packet(struct sk_buff *skb);
-extern u32	   dccp_sample_rtt(struct sock *sk, ktime_t t_recv,
-				   ktime_t *t_history);
+extern u32	   dccp_sample_rtt(struct sock *sk, long delta);
 
 static inline int dccp_bad_service_code(const struct sock *sk,
 					const __be32 service)

@@ -448,7 +448,7 @@ static void ccid3_hc_tx_packet_recv(stru
 		 * Calculate new round trip sample as per [RFC 3448, 4.3] by
 		 *	R_sample  =  (now - t_recvdata) - t_elapsed
 		 */
-		r_sample = dccp_sample_rtt(sk, now, &packet->dccphtx_tstamp);
+		r_sample = dccp_sample_rtt(sk, ktime_us_delta(now, packet->dccphtx_tstamp));
 
 		/*
 		 * Update RTT estimate by

@@ -881,9 +881,9 @@ static void ccid3_hc_rx_packet_recv(stru
 	case DCCP_PKT_DATAACK:
 		if (opt_recv->dccpor_timestamp_echo = 0)
 			break;
+		r_sample = dccp_timestamp() - opt_recv->dccpor_timestamp_echo;
 		rtt_prev = hcrx->ccid3hcrx_rtt;
-		now = ktime_get_real();
-		r_sample = dccp_sample_rtt(sk, now, NULL);
+		r_sample = dccp_sample_rtt(sk, 10 * r_sample);
 
 		if (hcrx->ccid3hcrx_state = TFRC_RSTATE_NO_DATA)
 			hcrx->ccid3hcrx_rtt = r_sample;

OK. Reviewed the merged patch now.

Web1: http://wand.net.nz/~iam4/
Web2: http://www.jandi.co.nz
Blog: http://iansblog.jandi.co.nz