Commit 1b7fdd2a authored by Yuchung Cheng's avatar Yuchung Cheng Committed by David S. Miller

tcp: do not use cached RTT for RTT estimation

RTT cached in the TCP metrics are valuable for the initial timeout
because SYN RTT usually does not account for serialization delays
on low BW path.

However using it to seed the RTT estimator maybe disruptive because
other components (e.g., pacing) require the smooth RTT to be obtained
from actual connection.

The solution is to use the higher cached RTT to set the first RTO
conservatively like tcp_rtt_estimator(), but avoid seeding the other
RTT estimator variables such as srtt.  It is also a good idea to
keep RTO conservative to obtain the first RTT sample, and the
performance is insured by TCP loss probe if SYN RTT is available.

To keep the seeding formula consistent across SYN RTT and cached RTT,
the rttvar is twice the cached RTT instead of cached RTTVAR value. The
reason is because cached variation may be too small (near min RTO)
which defeats the purpose of being conservative on first RTO. However
the metrics still keep the RTT variations as they might be useful for
user applications (through ip).
Signed-off-by: default avatarYuchung Cheng <ycheng@google.com>
Signed-off-by: default avatarNeal Cardwell <ncardwell@google.com>
Signed-off-by: default avatarEric Dumazet <edumazet@google.com>
Tested-by: default avatarEric Dumazet <edumazet@google.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent 08f89b98
...@@ -443,7 +443,7 @@ void tcp_init_metrics(struct sock *sk) ...@@ -443,7 +443,7 @@ void tcp_init_metrics(struct sock *sk)
struct dst_entry *dst = __sk_dst_get(sk); struct dst_entry *dst = __sk_dst_get(sk);
struct tcp_sock *tp = tcp_sk(sk); struct tcp_sock *tp = tcp_sk(sk);
struct tcp_metrics_block *tm; struct tcp_metrics_block *tm;
u32 val; u32 val, crtt = 0; /* cached RTT scaled by 8 */
if (dst == NULL) if (dst == NULL)
goto reset; goto reset;
...@@ -478,40 +478,18 @@ void tcp_init_metrics(struct sock *sk) ...@@ -478,40 +478,18 @@ void tcp_init_metrics(struct sock *sk)
tp->reordering = val; tp->reordering = val;
} }
val = tcp_metric_get(tm, TCP_METRIC_RTT); crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
if (val == 0 || tp->srtt == 0) {
rcu_read_unlock(); rcu_read_unlock();
goto reset;
}
/* Initial rtt is determined from SYN,SYN-ACK.
* The segment is small and rtt may appear much
* less than real one. Use per-dst memory
* to make it more realistic.
*
* A bit of theory. RTT is time passed after "normal" sized packet
* is sent until it is ACKed. In normal circumstances sending small
* packets force peer to delay ACKs and calculation is correct too.
* The algorithm is adaptive and, provided we follow specs, it
* NEVER underestimate RTT. BUT! If peer tries to make some clever
* tricks sort of "quick acks" for time long enough to decrease RTT
* to low value, and then abruptly stops to do it and starts to delay
* ACKs, wait for troubles.
*/
val = msecs_to_jiffies(val);
if (val > tp->srtt) {
tp->srtt = val;
tp->rtt_seq = tp->snd_nxt;
}
val = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
if (val > tp->mdev) {
tp->mdev = val;
tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
}
rcu_read_unlock();
tcp_set_rto(sk);
reset: reset:
if (tp->srtt == 0) { if (crtt > tp->srtt) {
/* Initial RTT (tp->srtt) from SYN usually don't measure
* serialization delay on low BW links well so RTO may be
* under-estimated. Stay conservative and seed RTO with
* the RTTs from past data exchanges, using the same seeding
* formula in tcp_rtt_estimator().
*/
inet_csk(sk)->icsk_rto = crtt + max(crtt >> 2, tcp_rto_min(sk));
} else if (tp->srtt == 0) {
/* RFC6298: 5.7 We've failed to get a valid RTT sample from /* RFC6298: 5.7 We've failed to get a valid RTT sample from
* 3WHS. This is most likely due to retransmission, * 3WHS. This is most likely due to retransmission,
* including spurious one. Reset the RTO back to 3secs * including spurious one. Reset the RTO back to 3secs
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment