linux/net/ipv4/tcp_vegas.c
Yuchung Cheng 9f9843a751 tcp: properly handle stretch acks in slow start
Slow start now increases cwnd by 1 if an ACK acknowledges some packets,
regardless the number of packets. Consequently slow start performance
is highly dependent on the degree of the stretch ACKs caused by
receiver or network ACK compression mechanisms (e.g., delayed-ACK,
GRO, etc).  But slow start algorithm is to send twice the amount of
packets of packets left so it should process a stretch ACK of degree
N as if N ACKs of degree 1, then exits when cwnd exceeds ssthresh. A
follow up patch will use the remainder of the N (if greater than 1)
to adjust cwnd in the congestion avoidance phase.

In addition this patch retires the experimental limited slow start
(LSS) feature. LSS has multiple drawbacks but questionable benefit. The
fractional cwnd increase in LSS requires a loop in slow start even
though it's rarely used. Configuring such an increase step via a global
sysctl on different BDPS seems hard. Finally and most importantly the
slow start overshoot concern is now better covered by the Hybrid slow
start (hystart) enabled by default.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2013-11-04 19:57:59 -05:00

341 lines
9.6 KiB
C

/*
* TCP Vegas congestion control
*
* This is based on the congestion detection/avoidance scheme described in
* Lawrence S. Brakmo and Larry L. Peterson.
* "TCP Vegas: End to end congestion avoidance on a global internet."
* IEEE Journal on Selected Areas in Communication, 13(8):1465--1480,
* October 1995. Available from:
* ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps
*
* See http://www.cs.arizona.edu/xkernel/ for their implementation.
* The main aspects that distinguish this implementation from the
* Arizona Vegas implementation are:
* o We do not change the loss detection or recovery mechanisms of
* Linux in any way. Linux already recovers from losses quite well,
* using fine-grained timers, NewReno, and FACK.
* o To avoid the performance penalty imposed by increasing cwnd
* only every-other RTT during slow start, we increase during
* every RTT during slow start, just like Reno.
* o Largely to allow continuous cwnd growth during slow start,
* we use the rate at which ACKs come back as the "actual"
* rate, rather than the rate at which data is sent.
* o To speed convergence to the right rate, we set the cwnd
* to achieve the right ("actual") rate when we exit slow start.
* o To filter out the noise caused by delayed ACKs, we use the
* minimum RTT sample observed during the last RTT to calculate
* the actual rate.
* o When the sender re-starts from idle, it waits until it has
* received ACKs for an entire flight of new data before making
* a cwnd adjustment decision. The original Vegas implementation
* assumed senders never went idle.
*/
#include <linux/mm.h>
#include <linux/module.h>
#include <linux/skbuff.h>
#include <linux/inet_diag.h>
#include <net/tcp.h>
#include "tcp_vegas.h"
static int alpha = 2;
static int beta = 4;
static int gamma = 1;
module_param(alpha, int, 0644);
MODULE_PARM_DESC(alpha, "lower bound of packets in network");
module_param(beta, int, 0644);
MODULE_PARM_DESC(beta, "upper bound of packets in network");
module_param(gamma, int, 0644);
MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)");
/* There are several situations when we must "re-start" Vegas:
*
* o when a connection is established
* o after an RTO
* o after fast recovery
* o when we send a packet and there is no outstanding
* unacknowledged data (restarting an idle connection)
*
* In these circumstances we cannot do a Vegas calculation at the
* end of the first RTT, because any calculation we do is using
* stale info -- both the saved cwnd and congestion feedback are
* stale.
*
* Instead we must wait until the completion of an RTT during
* which we actually receive ACKs.
*/
static void vegas_enable(struct sock *sk)
{
const struct tcp_sock *tp = tcp_sk(sk);
struct vegas *vegas = inet_csk_ca(sk);
/* Begin taking Vegas samples next time we send something. */
vegas->doing_vegas_now = 1;
/* Set the beginning of the next send window. */
vegas->beg_snd_nxt = tp->snd_nxt;
vegas->cntRTT = 0;
vegas->minRTT = 0x7fffffff;
}
/* Stop taking Vegas samples for now. */
static inline void vegas_disable(struct sock *sk)
{
struct vegas *vegas = inet_csk_ca(sk);
vegas->doing_vegas_now = 0;
}
void tcp_vegas_init(struct sock *sk)
{
struct vegas *vegas = inet_csk_ca(sk);
vegas->baseRTT = 0x7fffffff;
vegas_enable(sk);
}
EXPORT_SYMBOL_GPL(tcp_vegas_init);
/* Do RTT sampling needed for Vegas.
* Basically we:
* o min-filter RTT samples from within an RTT to get the current
* propagation delay + queuing delay (we are min-filtering to try to
* avoid the effects of delayed ACKs)
* o min-filter RTT samples from a much longer window (forever for now)
* to find the propagation delay (baseRTT)
*/
void tcp_vegas_pkts_acked(struct sock *sk, u32 cnt, s32 rtt_us)
{
struct vegas *vegas = inet_csk_ca(sk);
u32 vrtt;
if (rtt_us < 0)
return;
/* Never allow zero rtt or baseRTT */
vrtt = rtt_us + 1;
/* Filter to find propagation delay: */
if (vrtt < vegas->baseRTT)
vegas->baseRTT = vrtt;
/* Find the min RTT during the last RTT to find
* the current prop. delay + queuing delay:
*/
vegas->minRTT = min(vegas->minRTT, vrtt);
vegas->cntRTT++;
}
EXPORT_SYMBOL_GPL(tcp_vegas_pkts_acked);
void tcp_vegas_state(struct sock *sk, u8 ca_state)
{
if (ca_state == TCP_CA_Open)
vegas_enable(sk);
else
vegas_disable(sk);
}
EXPORT_SYMBOL_GPL(tcp_vegas_state);
/*
* If the connection is idle and we are restarting,
* then we don't want to do any Vegas calculations
* until we get fresh RTT samples. So when we
* restart, we reset our Vegas state to a clean
* slate. After we get acks for this flight of
* packets, _then_ we can make Vegas calculations
* again.
*/
void tcp_vegas_cwnd_event(struct sock *sk, enum tcp_ca_event event)
{
if (event == CA_EVENT_CWND_RESTART ||
event == CA_EVENT_TX_START)
tcp_vegas_init(sk);
}
EXPORT_SYMBOL_GPL(tcp_vegas_cwnd_event);
static inline u32 tcp_vegas_ssthresh(struct tcp_sock *tp)
{
return min(tp->snd_ssthresh, tp->snd_cwnd-1);
}
static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 acked,
u32 in_flight)
{
struct tcp_sock *tp = tcp_sk(sk);
struct vegas *vegas = inet_csk_ca(sk);
if (!vegas->doing_vegas_now) {
tcp_reno_cong_avoid(sk, ack, acked, in_flight);
return;
}
if (after(ack, vegas->beg_snd_nxt)) {
/* Do the Vegas once-per-RTT cwnd adjustment. */
/* Save the extent of the current window so we can use this
* at the end of the next RTT.
*/
vegas->beg_snd_nxt = tp->snd_nxt;
/* We do the Vegas calculations only if we got enough RTT
* samples that we can be reasonably sure that we got
* at least one RTT sample that wasn't from a delayed ACK.
* If we only had 2 samples total,
* then that means we're getting only 1 ACK per RTT, which
* means they're almost certainly delayed ACKs.
* If we have 3 samples, we should be OK.
*/
if (vegas->cntRTT <= 2) {
/* We don't have enough RTT samples to do the Vegas
* calculation, so we'll behave like Reno.
*/
tcp_reno_cong_avoid(sk, ack, acked, in_flight);
} else {
u32 rtt, diff;
u64 target_cwnd;
/* We have enough RTT samples, so, using the Vegas
* algorithm, we determine if we should increase or
* decrease cwnd, and by how much.
*/
/* Pluck out the RTT we are using for the Vegas
* calculations. This is the min RTT seen during the
* last RTT. Taking the min filters out the effects
* of delayed ACKs, at the cost of noticing congestion
* a bit later.
*/
rtt = vegas->minRTT;
/* Calculate the cwnd we should have, if we weren't
* going too fast.
*
* This is:
* (actual rate in segments) * baseRTT
*/
target_cwnd = tp->snd_cwnd * vegas->baseRTT / rtt;
/* Calculate the difference between the window we had,
* and the window we would like to have. This quantity
* is the "Diff" from the Arizona Vegas papers.
*/
diff = tp->snd_cwnd * (rtt-vegas->baseRTT) / vegas->baseRTT;
if (diff > gamma && tp->snd_cwnd <= tp->snd_ssthresh) {
/* Going too fast. Time to slow down
* and switch to congestion avoidance.
*/
/* Set cwnd to match the actual rate
* exactly:
* cwnd = (actual rate) * baseRTT
* Then we add 1 because the integer
* truncation robs us of full link
* utilization.
*/
tp->snd_cwnd = min(tp->snd_cwnd, (u32)target_cwnd+1);
tp->snd_ssthresh = tcp_vegas_ssthresh(tp);
} else if (tp->snd_cwnd <= tp->snd_ssthresh) {
/* Slow start. */
tcp_slow_start(tp, acked);
} else {
/* Congestion avoidance. */
/* Figure out where we would like cwnd
* to be.
*/
if (diff > beta) {
/* The old window was too fast, so
* we slow down.
*/
tp->snd_cwnd--;
tp->snd_ssthresh
= tcp_vegas_ssthresh(tp);
} else if (diff < alpha) {
/* We don't have enough extra packets
* in the network, so speed up.
*/
tp->snd_cwnd++;
} else {
/* Sending just as fast as we
* should be.
*/
}
}
if (tp->snd_cwnd < 2)
tp->snd_cwnd = 2;
else if (tp->snd_cwnd > tp->snd_cwnd_clamp)
tp->snd_cwnd = tp->snd_cwnd_clamp;
tp->snd_ssthresh = tcp_current_ssthresh(sk);
}
/* Wipe the slate clean for the next RTT. */
vegas->cntRTT = 0;
vegas->minRTT = 0x7fffffff;
}
/* Use normal slow start */
else if (tp->snd_cwnd <= tp->snd_ssthresh)
tcp_slow_start(tp, acked);
}
/* Extract info for Tcp socket info provided via netlink. */
void tcp_vegas_get_info(struct sock *sk, u32 ext, struct sk_buff *skb)
{
const struct vegas *ca = inet_csk_ca(sk);
if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
struct tcpvegas_info info = {
.tcpv_enabled = ca->doing_vegas_now,
.tcpv_rttcnt = ca->cntRTT,
.tcpv_rtt = ca->baseRTT,
.tcpv_minrtt = ca->minRTT,
};
nla_put(skb, INET_DIAG_VEGASINFO, sizeof(info), &info);
}
}
EXPORT_SYMBOL_GPL(tcp_vegas_get_info);
static struct tcp_congestion_ops tcp_vegas __read_mostly = {
.flags = TCP_CONG_RTT_STAMP,
.init = tcp_vegas_init,
.ssthresh = tcp_reno_ssthresh,
.cong_avoid = tcp_vegas_cong_avoid,
.min_cwnd = tcp_reno_min_cwnd,
.pkts_acked = tcp_vegas_pkts_acked,
.set_state = tcp_vegas_state,
.cwnd_event = tcp_vegas_cwnd_event,
.get_info = tcp_vegas_get_info,
.owner = THIS_MODULE,
.name = "vegas",
};
static int __init tcp_vegas_register(void)
{
BUILD_BUG_ON(sizeof(struct vegas) > ICSK_CA_PRIV_SIZE);
tcp_register_congestion_control(&tcp_vegas);
return 0;
}
static void __exit tcp_vegas_unregister(void)
{
tcp_unregister_congestion_control(&tcp_vegas);
}
module_init(tcp_vegas_register);
module_exit(tcp_vegas_unregister);
MODULE_AUTHOR("Stephen Hemminger");
MODULE_LICENSE("GPL");
MODULE_DESCRIPTION("TCP Vegas");