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
)
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
;
76 static void tcp_yeah_cong_avoid(struct sock
*sk
, u32 ack
,
77 u32 seq_rtt
, u32 in_flight
, int flag
)
79 struct tcp_sock
*tp
= tcp_sk(sk
);
80 struct yeah
*yeah
= inet_csk_ca(sk
);
82 if (!tcp_is_cwnd_limited(sk
, in_flight
))
85 if (tp
->snd_cwnd
<= tp
->snd_ssthresh
) {
87 } else if (!yeah
->doing_reno_now
) {
90 tp
->snd_cwnd_cnt
+=yeah
->pkts_acked
;
91 if (tp
->snd_cwnd_cnt
> min(tp
->snd_cwnd
, TCP_SCALABLE_AI_CNT
)){
92 if (tp
->snd_cwnd
< tp
->snd_cwnd_clamp
)
102 if (tp
->snd_cwnd_cnt
< tp
->snd_cwnd
)
105 if (tp
->snd_cwnd_cnt
>= tp
->snd_cwnd
) {
107 tp
->snd_cwnd_cnt
= 0;
111 /* The key players are v_beg_snd_una and v_beg_snd_nxt.
113 * These are so named because they represent the approximate values
114 * of snd_una and snd_nxt at the beginning of the current RTT. More
115 * precisely, they represent the amount of data sent during the RTT.
116 * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
117 * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
118 * bytes of data have been ACKed during the course of the RTT, giving
119 * an "actual" rate of:
121 * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
123 * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
124 * because delayed ACKs can cover more than one segment, so they
125 * don't line up yeahly with the boundaries of RTTs.
127 * Another unfortunate fact of life is that delayed ACKs delay the
128 * advance of the left edge of our send window, so that the number
129 * of bytes we send in an RTT is often less than our cwnd will allow.
130 * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
133 if (after(ack
, yeah
->beg_snd_nxt
)) {
135 /* We do the Vegas calculations only if we got enough RTT
136 * samples that we can be reasonably sure that we got
137 * at least one RTT sample that wasn't from a delayed ACK.
138 * If we only had 2 samples total,
139 * then that means we're getting only 1 ACK per RTT, which
140 * means they're almost certainly delayed ACKs.
141 * If we have 3 samples, we should be OK.
144 if (yeah
->cntRTT
> 2) {
148 /* We have enough RTT samples, so, using the Vegas
149 * algorithm, we determine if we should increase or
150 * decrease cwnd, and by how much.
153 /* Pluck out the RTT we are using for the Vegas
154 * calculations. This is the min RTT seen during the
155 * last RTT. Taking the min filters out the effects
156 * of delayed ACKs, at the cost of noticing congestion
161 /* Compute excess number of packets above bandwidth
162 * Avoid doing full 64 bit divide.
165 bw
*= rtt
- yeah
->baseRTT
;
169 if (queue
> TCP_YEAH_ALPHA
||
170 rtt
- yeah
->baseRTT
> (yeah
->baseRTT
/ TCP_YEAH_PHY
)) {
171 if (queue
> TCP_YEAH_ALPHA
172 && tp
->snd_cwnd
> yeah
->reno_count
) {
173 u32 reduction
= min(queue
/ TCP_YEAH_GAMMA
,
174 tp
->snd_cwnd
>> TCP_YEAH_EPSILON
);
176 tp
->snd_cwnd
-= reduction
;
178 tp
->snd_cwnd
= max(tp
->snd_cwnd
,
181 tp
->snd_ssthresh
= tp
->snd_cwnd
;
184 if (yeah
->reno_count
<= 2)
185 yeah
->reno_count
= max(tp
->snd_cwnd
>>1, 2U);
189 yeah
->doing_reno_now
= min(yeah
->doing_reno_now
+ 1,
194 if (yeah
->fast_count
> TCP_YEAH_ZETA
) {
195 yeah
->reno_count
= 2;
196 yeah
->fast_count
= 0;
199 yeah
->doing_reno_now
= 0;
206 /* Save the extent of the current window so we can use this
207 * at the end of the next RTT.
209 yeah
->beg_snd_una
= yeah
->beg_snd_nxt
;
210 yeah
->beg_snd_nxt
= tp
->snd_nxt
;
211 yeah
->beg_snd_cwnd
= tp
->snd_cwnd
;
213 /* Wipe the slate clean for the next RTT. */
215 yeah
->minRTT
= 0x7fffffff;
219 static u32
tcp_yeah_ssthresh(struct sock
*sk
) {
220 const struct tcp_sock
*tp
= tcp_sk(sk
);
221 struct yeah
*yeah
= inet_csk_ca(sk
);
224 if (yeah
->doing_reno_now
< TCP_YEAH_RHO
) {
225 reduction
= yeah
->lastQ
;
227 reduction
= min( reduction
, max(tp
->snd_cwnd
>>1, 2U) );
229 reduction
= max( reduction
, tp
->snd_cwnd
>> TCP_YEAH_DELTA
);
231 reduction
= max(tp
->snd_cwnd
>>1,2U);
233 yeah
->fast_count
= 0;
234 yeah
->reno_count
= max(yeah
->reno_count
>>1, 2U);
236 return tp
->snd_cwnd
- reduction
;
239 static struct tcp_congestion_ops tcp_yeah
= {
240 .init
= tcp_yeah_init
,
241 .ssthresh
= tcp_yeah_ssthresh
,
242 .cong_avoid
= tcp_yeah_cong_avoid
,
243 .min_cwnd
= tcp_reno_min_cwnd
,
244 .rtt_sample
= tcp_vegas_rtt_calc
,
245 .set_state
= tcp_vegas_state
,
246 .cwnd_event
= tcp_vegas_cwnd_event
,
247 .get_info
= tcp_vegas_get_info
,
248 .pkts_acked
= tcp_yeah_pkts_acked
,
250 .owner
= THIS_MODULE
,
254 static int __init
tcp_yeah_register(void)
256 BUG_ON(sizeof(struct yeah
) > ICSK_CA_PRIV_SIZE
);
257 tcp_register_congestion_control(&tcp_yeah
);
261 static void __exit
tcp_yeah_unregister(void)
263 tcp_unregister_congestion_control(&tcp_yeah
);
266 module_init(tcp_yeah_register
);
267 module_exit(tcp_yeah_unregister
);
269 MODULE_AUTHOR("Angelo P. Castellani");
270 MODULE_LICENSE("GPL");
271 MODULE_DESCRIPTION("YeAH TCP");