linux/net/ipv4/tcp_yeah.c
Neal Cardwell e73ebb0881 tcp: stretch ACK fixes prep
LRO, GRO, delayed ACKs, and middleboxes can cause "stretch ACKs" that
cover more than the RFC-specified maximum of 2 packets. These stretch
ACKs can cause serious performance shortfalls in common congestion
control algorithms that were designed and tuned years ago with
receiver hosts that were not using LRO or GRO, and were instead
politely ACKing every other packet.

This patch series fixes Reno and CUBIC to handle stretch ACKs.

This patch prepares for the upcoming stretch ACK bug fix patches. It
adds an "acked" parameter to tcp_cong_avoid_ai() to allow for future
fixes to tcp_cong_avoid_ai() to correctly handle stretch ACKs, and
changes all congestion control algorithms to pass in 1 for the ACKed
count. It also changes tcp_slow_start() to return the number of packet
ACK "credits" that were not processed in slow start mode, and can be
processed by the congestion control module in additive increase mode.

In future patches we will fix tcp_cong_avoid_ai() to handle stretch
ACKs, and fix Reno and CUBIC handling of stretch ACKs in slow start
and additive increase mode.

Reported-by: Eyal Perry <eyalpe@mellanox.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2015-01-28 22:18:37 -08:00

255 lines
6.9 KiB
C

/*
*
* YeAH TCP
*
* For further details look at:
* https://web.archive.org/web/20080316215752/http://wil.cs.caltech.edu/pfldnet2007/paper/YeAH_TCP.pdf
*
*/
#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"
#define TCP_YEAH_ALPHA 80 /* number of packets queued at the bottleneck */
#define TCP_YEAH_GAMMA 1 /* fraction of queue to be removed per rtt */
#define TCP_YEAH_DELTA 3 /* log minimum fraction of cwnd to be removed on loss */
#define TCP_YEAH_EPSILON 1 /* log maximum fraction to be removed on early decongestion */
#define TCP_YEAH_PHY 8 /* maximum delta from base */
#define TCP_YEAH_RHO 16 /* minimum number of consecutive rtt to consider competition on loss */
#define TCP_YEAH_ZETA 50 /* minimum number of state switches to reset reno_count */
#define TCP_SCALABLE_AI_CNT 100U
/* YeAH variables */
struct yeah {
struct vegas vegas; /* must be first */
/* YeAH */
u32 lastQ;
u32 doing_reno_now;
u32 reno_count;
u32 fast_count;
u32 pkts_acked;
};
static void tcp_yeah_init(struct sock *sk)
{
struct tcp_sock *tp = tcp_sk(sk);
struct yeah *yeah = inet_csk_ca(sk);
tcp_vegas_init(sk);
yeah->doing_reno_now = 0;
yeah->lastQ = 0;
yeah->reno_count = 2;
/* Ensure the MD arithmetic works. This is somewhat pedantic,
* since I don't think we will see a cwnd this large. :) */
tp->snd_cwnd_clamp = min_t(u32, tp->snd_cwnd_clamp, 0xffffffff/128);
}
static void tcp_yeah_pkts_acked(struct sock *sk, u32 pkts_acked, s32 rtt_us)
{
const struct inet_connection_sock *icsk = inet_csk(sk);
struct yeah *yeah = inet_csk_ca(sk);
if (icsk->icsk_ca_state == TCP_CA_Open)
yeah->pkts_acked = pkts_acked;
tcp_vegas_pkts_acked(sk, pkts_acked, rtt_us);
}
static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, u32 acked)
{
struct tcp_sock *tp = tcp_sk(sk);
struct yeah *yeah = inet_csk_ca(sk);
if (!tcp_is_cwnd_limited(sk))
return;
if (tp->snd_cwnd <= tp->snd_ssthresh)
tcp_slow_start(tp, acked);
else if (!yeah->doing_reno_now) {
/* Scalable */
tp->snd_cwnd_cnt += yeah->pkts_acked;
if (tp->snd_cwnd_cnt > min(tp->snd_cwnd, TCP_SCALABLE_AI_CNT)) {
if (tp->snd_cwnd < tp->snd_cwnd_clamp)
tp->snd_cwnd++;
tp->snd_cwnd_cnt = 0;
}
yeah->pkts_acked = 1;
} else {
/* Reno */
tcp_cong_avoid_ai(tp, tp->snd_cwnd, 1);
}
/* The key players are v_vegas.beg_snd_una and v_beg_snd_nxt.
*
* These are so named because they represent the approximate values
* of snd_una and snd_nxt at the beginning of the current RTT. More
* precisely, they represent the amount of data sent during the RTT.
* At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
* we will calculate that (v_beg_snd_nxt - v_vegas.beg_snd_una) outstanding
* bytes of data have been ACKed during the course of the RTT, giving
* an "actual" rate of:
*
* (v_beg_snd_nxt - v_vegas.beg_snd_una) / (rtt duration)
*
* Unfortunately, v_vegas.beg_snd_una is not exactly equal to snd_una,
* because delayed ACKs can cover more than one segment, so they
* don't line up yeahly with the boundaries of RTTs.
*
* Another unfortunate fact of life is that delayed ACKs delay the
* advance of the left edge of our send window, so that the number
* of bytes we send in an RTT is often less than our cwnd will allow.
* So we keep track of our cwnd separately, in v_beg_snd_cwnd.
*/
if (after(ack, yeah->vegas.beg_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 (yeah->vegas.cntRTT > 2) {
u32 rtt, queue;
u64 bw;
/* 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 = yeah->vegas.minRTT;
/* Compute excess number of packets above bandwidth
* Avoid doing full 64 bit divide.
*/
bw = tp->snd_cwnd;
bw *= rtt - yeah->vegas.baseRTT;
do_div(bw, rtt);
queue = bw;
if (queue > TCP_YEAH_ALPHA ||
rtt - yeah->vegas.baseRTT > (yeah->vegas.baseRTT / TCP_YEAH_PHY)) {
if (queue > TCP_YEAH_ALPHA &&
tp->snd_cwnd > yeah->reno_count) {
u32 reduction = min(queue / TCP_YEAH_GAMMA ,
tp->snd_cwnd >> TCP_YEAH_EPSILON);
tp->snd_cwnd -= reduction;
tp->snd_cwnd = max(tp->snd_cwnd,
yeah->reno_count);
tp->snd_ssthresh = tp->snd_cwnd;
}
if (yeah->reno_count <= 2)
yeah->reno_count = max(tp->snd_cwnd>>1, 2U);
else
yeah->reno_count++;
yeah->doing_reno_now = min(yeah->doing_reno_now + 1,
0xffffffU);
} else {
yeah->fast_count++;
if (yeah->fast_count > TCP_YEAH_ZETA) {
yeah->reno_count = 2;
yeah->fast_count = 0;
}
yeah->doing_reno_now = 0;
}
yeah->lastQ = queue;
}
/* Save the extent of the current window so we can use this
* at the end of the next RTT.
*/
yeah->vegas.beg_snd_una = yeah->vegas.beg_snd_nxt;
yeah->vegas.beg_snd_nxt = tp->snd_nxt;
yeah->vegas.beg_snd_cwnd = tp->snd_cwnd;
/* Wipe the slate clean for the next RTT. */
yeah->vegas.cntRTT = 0;
yeah->vegas.minRTT = 0x7fffffff;
}
}
static u32 tcp_yeah_ssthresh(struct sock *sk)
{
const struct tcp_sock *tp = tcp_sk(sk);
struct yeah *yeah = inet_csk_ca(sk);
u32 reduction;
if (yeah->doing_reno_now < TCP_YEAH_RHO) {
reduction = yeah->lastQ;
reduction = min(reduction, max(tp->snd_cwnd>>1, 2U));
reduction = max(reduction, tp->snd_cwnd >> TCP_YEAH_DELTA);
} else
reduction = max(tp->snd_cwnd>>1, 2U);
yeah->fast_count = 0;
yeah->reno_count = max(yeah->reno_count>>1, 2U);
return tp->snd_cwnd - reduction;
}
static struct tcp_congestion_ops tcp_yeah __read_mostly = {
.init = tcp_yeah_init,
.ssthresh = tcp_yeah_ssthresh,
.cong_avoid = tcp_yeah_cong_avoid,
.set_state = tcp_vegas_state,
.cwnd_event = tcp_vegas_cwnd_event,
.get_info = tcp_vegas_get_info,
.pkts_acked = tcp_yeah_pkts_acked,
.owner = THIS_MODULE,
.name = "yeah",
};
static int __init tcp_yeah_register(void)
{
BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE);
tcp_register_congestion_control(&tcp_yeah);
return 0;
}
static void __exit tcp_yeah_unregister(void)
{
tcp_unregister_congestion_control(&tcp_yeah);
}
module_init(tcp_yeah_register);
module_exit(tcp_yeah_unregister);
MODULE_AUTHOR("Angelo P. Castellani");
MODULE_LICENSE("GPL");
MODULE_DESCRIPTION("YeAH TCP");