5 * For further details look at:
6 * http://wil.cs.caltech.edu/pfldnet2007/paper/YeAH_TCP.pdf
12 /* Default values of the Vegas variables, in fixed-point representation
13 * with V_PARAM_SHIFT bits to the right of the binary point.
15 #define V_PARAM_SHIFT 1
17 #define TCP_YEAH_ALPHA 80 //lin number of packets queued at the bottleneck
18 #define TCP_YEAH_GAMMA 1 //lin fraction of queue to be removed per rtt
19 #define TCP_YEAH_DELTA 3 //log minimum fraction of cwnd to be removed on loss
20 #define TCP_YEAH_EPSILON 1 //log maximum fraction to be removed on early decongestion
21 #define TCP_YEAH_PHY 8 //lin maximum delta from base
22 #define TCP_YEAH_RHO 16 //lin minumum number of consecutive rtt to consider competition on loss
23 #define TCP_YEAH_ZETA 50 //lin minimum number of state switchs to reset reno_count
25 #define TCP_SCALABLE_AI_CNT 100U
30 u32 beg_snd_nxt; /* right edge during last RTT */
31 u32 beg_snd_una; /* left edge during last RTT */
32 u32 beg_snd_cwnd; /* saves the size of the cwnd */
33 u8 doing_vegas_now;/* if true, do vegas for this RTT */
34 u16 cntRTT; /* # of RTTs measured within last RTT */
35 u32 minRTT; /* min of RTTs measured within last RTT (in usec) */
36 u32 baseRTT; /* the min of all Vegas RTT measurements seen (in usec) */
48 static void tcp_yeah_init(struct sock *sk)
50 struct tcp_sock *tp = tcp_sk(sk);
51 struct yeah *yeah = inet_csk_ca(sk);
55 yeah->doing_reno_now = 0;
60 /* Ensure the MD arithmetic works. This is somewhat pedantic,
61 * since I don't think we will see a cwnd this large. :) */
62 tp->snd_cwnd_clamp = min_t(u32, tp->snd_cwnd_clamp, 0xffffffff/128);
67 static void tcp_yeah_pkts_acked(struct sock *sk, u32 pkts_acked, ktime_t last)
69 const struct inet_connection_sock *icsk = inet_csk(sk);
70 struct yeah *yeah = inet_csk_ca(sk);
72 if (icsk->icsk_ca_state == TCP_CA_Open)
73 yeah->pkts_acked = pkts_acked;
75 tcp_vegas_pkts_acked(sk, pkts_acked, last);
78 static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack,
79 u32 seq_rtt, u32 in_flight, int flag)
81 struct tcp_sock *tp = tcp_sk(sk);
82 struct yeah *yeah = inet_csk_ca(sk);
84 if (!tcp_is_cwnd_limited(sk, in_flight))
87 if (tp->snd_cwnd <= tp->snd_ssthresh) {
89 } else if (!yeah->doing_reno_now) {
92 tp->snd_cwnd_cnt+=yeah->pkts_acked;
93 if (tp->snd_cwnd_cnt > min(tp->snd_cwnd, TCP_SCALABLE_AI_CNT)){
94 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
104 if (tp->snd_cwnd_cnt < tp->snd_cwnd)
107 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
109 tp->snd_cwnd_cnt = 0;
113 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
115 * These are so named because they represent the approximate values
116 * of snd_una and snd_nxt at the beginning of the current RTT. More
117 * precisely, they represent the amount of data sent during the RTT.
118 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
119 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
120 * bytes of data have been ACKed during the course of the RTT, giving
121 * an "actual" rate of:
123 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
125 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
126 * because delayed ACKs can cover more than one segment, so they
127 * don't line up yeahly with the boundaries of RTTs.
129 * Another unfortunate fact of life is that delayed ACKs delay the
130 * advance of the left edge of our send window, so that the number
131 * of bytes we send in an RTT is often less than our cwnd will allow.
132 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
135 if (after(ack, yeah->beg_snd_nxt)) {
137 /* We do the Vegas calculations only if we got enough RTT
138 * samples that we can be reasonably sure that we got
139 * at least one RTT sample that wasn't from a delayed ACK.
140 * If we only had 2 samples total,
141 * then that means we're getting only 1 ACK per RTT, which
142 * means they're almost certainly delayed ACKs.
143 * If we have 3 samples, we should be OK.
146 if (yeah->cntRTT > 2) {
150 /* We have enough RTT samples, so, using the Vegas
151 * algorithm, we determine if we should increase or
152 * decrease cwnd, and by how much.
155 /* Pluck out the RTT we are using for the Vegas
156 * calculations. This is the min RTT seen during the
157 * last RTT. Taking the min filters out the effects
158 * of delayed ACKs, at the cost of noticing congestion
163 /* Compute excess number of packets above bandwidth
164 * Avoid doing full 64 bit divide.
167 bw *= rtt - yeah->baseRTT;
171 if (queue > TCP_YEAH_ALPHA ||
172 rtt - yeah->baseRTT > (yeah->baseRTT / TCP_YEAH_PHY)) {
173 if (queue > TCP_YEAH_ALPHA
174 && tp->snd_cwnd > yeah->reno_count) {
175 u32 reduction = min(queue / TCP_YEAH_GAMMA ,
176 tp->snd_cwnd >> TCP_YEAH_EPSILON);
178 tp->snd_cwnd -= reduction;
180 tp->snd_cwnd = max(tp->snd_cwnd,
183 tp->snd_ssthresh = tp->snd_cwnd;
186 if (yeah->reno_count <= 2)
187 yeah->reno_count = max(tp->snd_cwnd>>1, 2U);
191 yeah->doing_reno_now = min(yeah->doing_reno_now + 1,
196 if (yeah->fast_count > TCP_YEAH_ZETA) {
197 yeah->reno_count = 2;
198 yeah->fast_count = 0;
201 yeah->doing_reno_now = 0;
208 /* Save the extent of the current window so we can use this
209 * at the end of the next RTT.
211 yeah->beg_snd_una = yeah->beg_snd_nxt;
212 yeah->beg_snd_nxt = tp->snd_nxt;
213 yeah->beg_snd_cwnd = tp->snd_cwnd;
215 /* Wipe the slate clean for the next RTT. */
217 yeah->minRTT = 0x7fffffff;
221 static u32 tcp_yeah_ssthresh(struct sock *sk) {
222 const struct tcp_sock *tp = tcp_sk(sk);
223 struct yeah *yeah = inet_csk_ca(sk);
226 if (yeah->doing_reno_now < TCP_YEAH_RHO) {
227 reduction = yeah->lastQ;
229 reduction = min( reduction, max(tp->snd_cwnd>>1, 2U) );
231 reduction = max( reduction, tp->snd_cwnd >> TCP_YEAH_DELTA);
233 reduction = max(tp->snd_cwnd>>1,2U);
235 yeah->fast_count = 0;
236 yeah->reno_count = max(yeah->reno_count>>1, 2U);
238 return tp->snd_cwnd - reduction;
241 static struct tcp_congestion_ops tcp_yeah = {
242 .flags = TCP_CONG_RTT_STAMP,
243 .init = tcp_yeah_init,
244 .ssthresh = tcp_yeah_ssthresh,
245 .cong_avoid = tcp_yeah_cong_avoid,
246 .min_cwnd = tcp_reno_min_cwnd,
247 .set_state = tcp_vegas_state,
248 .cwnd_event = tcp_vegas_cwnd_event,
249 .get_info = tcp_vegas_get_info,
250 .pkts_acked = tcp_yeah_pkts_acked,
252 .owner = THIS_MODULE,
256 static int __init tcp_yeah_register(void)
258 BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE);
259 tcp_register_congestion_control(&tcp_yeah);
263 static void __exit tcp_yeah_unregister(void)
265 tcp_unregister_congestion_control(&tcp_yeah);
268 module_init(tcp_yeah_register);
269 module_exit(tcp_yeah_unregister);
271 MODULE_AUTHOR("Angelo P. Castellani");
272 MODULE_LICENSE("GPL");
273 MODULE_DESCRIPTION("YeAH TCP");