5 * For further details look at:
6 * http://wil.cs.caltech.edu/pfldnet2007/paper/YeAH_TCP.pdf
10 #include <linux/module.h>
11 #include <linux/skbuff.h>
12 #include <linux/inet_diag.h>
16 #include "tcp_vegas.h"
18 #define TCP_YEAH_ALPHA 80 //lin number of packets queued at the bottleneck
19 #define TCP_YEAH_GAMMA 1 //lin fraction of queue to be removed per rtt
20 #define TCP_YEAH_DELTA 3 //log minimum fraction of cwnd to be removed on loss
21 #define TCP_YEAH_EPSILON 1 //log maximum fraction to be removed on early decongestion
22 #define TCP_YEAH_PHY 8 //lin maximum delta from base
23 #define TCP_YEAH_RHO 16 //lin minumum number of consecutive rtt to consider competition on loss
24 #define TCP_YEAH_ZETA 50 //lin minimum number of state switchs to reset reno_count
26 #define TCP_SCALABLE_AI_CNT 100U
30 struct vegas vegas; /* must be first */
42 static void tcp_yeah_init(struct sock *sk)
44 struct tcp_sock *tp = tcp_sk(sk);
45 struct yeah *yeah = inet_csk_ca(sk);
49 yeah->doing_reno_now = 0;
54 /* Ensure the MD arithmetic works. This is somewhat pedantic,
55 * since I don't think we will see a cwnd this large. :) */
56 tp->snd_cwnd_clamp = min_t(u32, tp->snd_cwnd_clamp, 0xffffffff/128);
61 static void tcp_yeah_pkts_acked(struct sock *sk, u32 pkts_acked, s32 rtt_us)
63 const struct inet_connection_sock *icsk = inet_csk(sk);
64 struct yeah *yeah = inet_csk_ca(sk);
66 if (icsk->icsk_ca_state == TCP_CA_Open)
67 yeah->pkts_acked = pkts_acked;
69 tcp_vegas_pkts_acked(sk, pkts_acked, rtt_us);
72 static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
74 struct tcp_sock *tp = tcp_sk(sk);
75 struct yeah *yeah = inet_csk_ca(sk);
77 if (!tcp_is_cwnd_limited(sk, in_flight))
80 if (tp->snd_cwnd <= tp->snd_ssthresh)
83 else if (!yeah->doing_reno_now) {
86 tp->snd_cwnd_cnt += yeah->pkts_acked;
87 if (tp->snd_cwnd_cnt > min(tp->snd_cwnd, TCP_SCALABLE_AI_CNT)){
88 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
97 tcp_cong_avoid_ai(tp, tp->snd_cwnd);
100 /* The key players are v_vegas.beg_snd_una and v_beg_snd_nxt.
102 * These are so named because they represent the approximate values
103 * of snd_una and snd_nxt at the beginning of the current RTT. More
104 * precisely, they represent the amount of data sent during the RTT.
105 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
106 * we will calculate that (v_beg_snd_nxt - v_vegas.beg_snd_una) outstanding
107 * bytes of data have been ACKed during the course of the RTT, giving
108 * an "actual" rate of:
110 * (v_beg_snd_nxt - v_vegas.beg_snd_una) / (rtt duration)
112 * Unfortunately, v_vegas.beg_snd_una is not exactly equal to snd_una,
113 * because delayed ACKs can cover more than one segment, so they
114 * don't line up yeahly with the boundaries of RTTs.
116 * Another unfortunate fact of life is that delayed ACKs delay the
117 * advance of the left edge of our send window, so that the number
118 * of bytes we send in an RTT is often less than our cwnd will allow.
119 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
122 if (after(ack, yeah->vegas.beg_snd_nxt)) {
124 /* We do the Vegas calculations only if we got enough RTT
125 * samples that we can be reasonably sure that we got
126 * at least one RTT sample that wasn't from a delayed ACK.
127 * If we only had 2 samples total,
128 * then that means we're getting only 1 ACK per RTT, which
129 * means they're almost certainly delayed ACKs.
130 * If we have 3 samples, we should be OK.
133 if (yeah->vegas.cntRTT > 2) {
137 /* We have enough RTT samples, so, using the Vegas
138 * algorithm, we determine if we should increase or
139 * decrease cwnd, and by how much.
142 /* Pluck out the RTT we are using for the Vegas
143 * calculations. This is the min RTT seen during the
144 * last RTT. Taking the min filters out the effects
145 * of delayed ACKs, at the cost of noticing congestion
148 rtt = yeah->vegas.minRTT;
150 /* Compute excess number of packets above bandwidth
151 * Avoid doing full 64 bit divide.
154 bw *= rtt - yeah->vegas.baseRTT;
158 if (queue > TCP_YEAH_ALPHA ||
159 rtt - yeah->vegas.baseRTT > (yeah->vegas.baseRTT / TCP_YEAH_PHY)) {
160 if (queue > TCP_YEAH_ALPHA
161 && tp->snd_cwnd > yeah->reno_count) {
162 u32 reduction = min(queue / TCP_YEAH_GAMMA ,
163 tp->snd_cwnd >> TCP_YEAH_EPSILON);
165 tp->snd_cwnd -= reduction;
167 tp->snd_cwnd = max(tp->snd_cwnd,
170 tp->snd_ssthresh = tp->snd_cwnd;
173 if (yeah->reno_count <= 2)
174 yeah->reno_count = max(tp->snd_cwnd>>1, 2U);
178 yeah->doing_reno_now = min(yeah->doing_reno_now + 1,
183 if (yeah->fast_count > TCP_YEAH_ZETA) {
184 yeah->reno_count = 2;
185 yeah->fast_count = 0;
188 yeah->doing_reno_now = 0;
195 /* Save the extent of the current window so we can use this
196 * at the end of the next RTT.
198 yeah->vegas.beg_snd_una = yeah->vegas.beg_snd_nxt;
199 yeah->vegas.beg_snd_nxt = tp->snd_nxt;
200 yeah->vegas.beg_snd_cwnd = tp->snd_cwnd;
202 /* Wipe the slate clean for the next RTT. */
203 yeah->vegas.cntRTT = 0;
204 yeah->vegas.minRTT = 0x7fffffff;
208 static u32 tcp_yeah_ssthresh(struct sock *sk) {
209 const struct tcp_sock *tp = tcp_sk(sk);
210 struct yeah *yeah = inet_csk_ca(sk);
213 if (yeah->doing_reno_now < TCP_YEAH_RHO) {
214 reduction = yeah->lastQ;
216 reduction = min( reduction, max(tp->snd_cwnd>>1, 2U) );
218 reduction = max( reduction, tp->snd_cwnd >> TCP_YEAH_DELTA);
220 reduction = max(tp->snd_cwnd>>1, 2U);
222 yeah->fast_count = 0;
223 yeah->reno_count = max(yeah->reno_count>>1, 2U);
225 return tp->snd_cwnd - reduction;
228 static struct tcp_congestion_ops tcp_yeah = {
229 .flags = TCP_CONG_RTT_STAMP,
230 .init = tcp_yeah_init,
231 .ssthresh = tcp_yeah_ssthresh,
232 .cong_avoid = tcp_yeah_cong_avoid,
233 .min_cwnd = tcp_reno_min_cwnd,
234 .set_state = tcp_vegas_state,
235 .cwnd_event = tcp_vegas_cwnd_event,
236 .get_info = tcp_vegas_get_info,
237 .pkts_acked = tcp_yeah_pkts_acked,
239 .owner = THIS_MODULE,
243 static int __init tcp_yeah_register(void)
245 BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE);
246 tcp_register_congestion_control(&tcp_yeah);
250 static void __exit tcp_yeah_unregister(void)
252 tcp_unregister_congestion_control(&tcp_yeah);
255 module_init(tcp_yeah_register);
256 module_exit(tcp_yeah_unregister);
258 MODULE_AUTHOR("Angelo P. Castellani");
259 MODULE_LICENSE("GPL");
260 MODULE_DESCRIPTION("YeAH TCP");