[IPv4] RAW: Compact the API for the kernel
[linux-2.6] / net / ipv4 / tcp_yeah.c
1 /*
2  *
3  *   YeAH TCP
4  *
5  * For further details look at:
6  *    http://wil.cs.caltech.edu/pfldnet2007/paper/YeAH_TCP.pdf
7  *
8  */
9 #include <linux/mm.h>
10 #include <linux/module.h>
11 #include <linux/skbuff.h>
12 #include <linux/inet_diag.h>
13
14 #include <net/tcp.h>
15
16 #include "tcp_vegas.h"
17
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
25
26 #define TCP_SCALABLE_AI_CNT      100U
27
28 /* YeAH variables */
29 struct yeah {
30         struct vegas vegas;     /* must be first */
31
32         /* YeAH */
33         u32 lastQ;
34         u32 doing_reno_now;
35
36         u32 reno_count;
37         u32 fast_count;
38
39         u32 pkts_acked;
40 };
41
42 static void tcp_yeah_init(struct sock *sk)
43 {
44         struct tcp_sock *tp = tcp_sk(sk);
45         struct yeah *yeah = inet_csk_ca(sk);
46
47         tcp_vegas_init(sk);
48
49         yeah->doing_reno_now = 0;
50         yeah->lastQ = 0;
51
52         yeah->reno_count = 2;
53
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);
57
58 }
59
60
61 static void tcp_yeah_pkts_acked(struct sock *sk, u32 pkts_acked, s32 rtt_us)
62 {
63         const struct inet_connection_sock *icsk = inet_csk(sk);
64         struct yeah *yeah = inet_csk_ca(sk);
65
66         if (icsk->icsk_ca_state == TCP_CA_Open)
67                 yeah->pkts_acked = pkts_acked;
68
69         tcp_vegas_pkts_acked(sk, pkts_acked, rtt_us);
70 }
71
72 static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack,
73                                 u32 in_flight, int flag)
74 {
75         struct tcp_sock *tp = tcp_sk(sk);
76         struct yeah *yeah = inet_csk_ca(sk);
77
78         if (!tcp_is_cwnd_limited(sk, in_flight))
79                 return;
80
81         if (tp->snd_cwnd <= tp->snd_ssthresh)
82                 tcp_slow_start(tp);
83
84         else if (!yeah->doing_reno_now) {
85                 /* Scalable */
86
87                 tp->snd_cwnd_cnt+=yeah->pkts_acked;
88                 if (tp->snd_cwnd_cnt > min(tp->snd_cwnd, TCP_SCALABLE_AI_CNT)){
89                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
90                                 tp->snd_cwnd++;
91                         tp->snd_cwnd_cnt = 0;
92                 }
93
94                 yeah->pkts_acked = 1;
95
96         } else {
97                 /* Reno */
98
99                 if (tp->snd_cwnd_cnt < tp->snd_cwnd)
100                         tp->snd_cwnd_cnt++;
101
102                 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
103                         tp->snd_cwnd++;
104                         tp->snd_cwnd_cnt = 0;
105                 }
106         }
107
108         /* The key players are v_vegas.beg_snd_una and v_beg_snd_nxt.
109          *
110          * These are so named because they represent the approximate values
111          * of snd_una and snd_nxt at the beginning of the current RTT. More
112          * precisely, they represent the amount of data sent during the RTT.
113          * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
114          * we will calculate that (v_beg_snd_nxt - v_vegas.beg_snd_una) outstanding
115          * bytes of data have been ACKed during the course of the RTT, giving
116          * an "actual" rate of:
117          *
118          *     (v_beg_snd_nxt - v_vegas.beg_snd_una) / (rtt duration)
119          *
120          * Unfortunately, v_vegas.beg_snd_una is not exactly equal to snd_una,
121          * because delayed ACKs can cover more than one segment, so they
122          * don't line up yeahly with the boundaries of RTTs.
123          *
124          * Another unfortunate fact of life is that delayed ACKs delay the
125          * advance of the left edge of our send window, so that the number
126          * of bytes we send in an RTT is often less than our cwnd will allow.
127          * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
128          */
129
130         if (after(ack, yeah->vegas.beg_snd_nxt)) {
131
132                 /* We do the Vegas calculations only if we got enough RTT
133                  * samples that we can be reasonably sure that we got
134                  * at least one RTT sample that wasn't from a delayed ACK.
135                  * If we only had 2 samples total,
136                  * then that means we're getting only 1 ACK per RTT, which
137                  * means they're almost certainly delayed ACKs.
138                  * If  we have 3 samples, we should be OK.
139                  */
140
141                 if (yeah->vegas.cntRTT > 2) {
142                         u32 rtt, queue;
143                         u64 bw;
144
145                         /* We have enough RTT samples, so, using the Vegas
146                          * algorithm, we determine if we should increase or
147                          * decrease cwnd, and by how much.
148                          */
149
150                         /* Pluck out the RTT we are using for the Vegas
151                          * calculations. This is the min RTT seen during the
152                          * last RTT. Taking the min filters out the effects
153                          * of delayed ACKs, at the cost of noticing congestion
154                          * a bit later.
155                          */
156                         rtt = yeah->vegas.minRTT;
157
158                         /* Compute excess number of packets above bandwidth
159                          * Avoid doing full 64 bit divide.
160                          */
161                         bw = tp->snd_cwnd;
162                         bw *= rtt - yeah->vegas.baseRTT;
163                         do_div(bw, rtt);
164                         queue = bw;
165
166                         if (queue > TCP_YEAH_ALPHA ||
167                             rtt - yeah->vegas.baseRTT > (yeah->vegas.baseRTT / TCP_YEAH_PHY)) {
168                                 if (queue > TCP_YEAH_ALPHA
169                                     && tp->snd_cwnd > yeah->reno_count) {
170                                         u32 reduction = min(queue / TCP_YEAH_GAMMA ,
171                                                             tp->snd_cwnd >> TCP_YEAH_EPSILON);
172
173                                         tp->snd_cwnd -= reduction;
174
175                                         tp->snd_cwnd = max(tp->snd_cwnd,
176                                                            yeah->reno_count);
177
178                                         tp->snd_ssthresh = tp->snd_cwnd;
179                                 }
180
181                                 if (yeah->reno_count <= 2)
182                                         yeah->reno_count = max(tp->snd_cwnd>>1, 2U);
183                                 else
184                                         yeah->reno_count++;
185
186                                 yeah->doing_reno_now = min(yeah->doing_reno_now + 1,
187                                                            0xffffffU);
188                         } else {
189                                 yeah->fast_count++;
190
191                                 if (yeah->fast_count > TCP_YEAH_ZETA) {
192                                         yeah->reno_count = 2;
193                                         yeah->fast_count = 0;
194                                 }
195
196                                 yeah->doing_reno_now = 0;
197                         }
198
199                         yeah->lastQ = queue;
200
201                 }
202
203                 /* Save the extent of the current window so we can use this
204                  * at the end of the next RTT.
205                  */
206                 yeah->vegas.beg_snd_una  = yeah->vegas.beg_snd_nxt;
207                 yeah->vegas.beg_snd_nxt  = tp->snd_nxt;
208                 yeah->vegas.beg_snd_cwnd = tp->snd_cwnd;
209
210                 /* Wipe the slate clean for the next RTT. */
211                 yeah->vegas.cntRTT = 0;
212                 yeah->vegas.minRTT = 0x7fffffff;
213         }
214 }
215
216 static u32 tcp_yeah_ssthresh(struct sock *sk) {
217         const struct tcp_sock *tp = tcp_sk(sk);
218         struct yeah *yeah = inet_csk_ca(sk);
219         u32 reduction;
220
221         if (yeah->doing_reno_now < TCP_YEAH_RHO) {
222                 reduction = yeah->lastQ;
223
224                 reduction = min( reduction, max(tp->snd_cwnd>>1, 2U) );
225
226                 reduction = max( reduction, tp->snd_cwnd >> TCP_YEAH_DELTA);
227         } else
228                 reduction = max(tp->snd_cwnd>>1,2U);
229
230         yeah->fast_count = 0;
231         yeah->reno_count = max(yeah->reno_count>>1, 2U);
232
233         return tp->snd_cwnd - reduction;
234 }
235
236 static struct tcp_congestion_ops tcp_yeah = {
237         .flags          = TCP_CONG_RTT_STAMP,
238         .init           = tcp_yeah_init,
239         .ssthresh       = tcp_yeah_ssthresh,
240         .cong_avoid     = tcp_yeah_cong_avoid,
241         .min_cwnd       = tcp_reno_min_cwnd,
242         .set_state      = tcp_vegas_state,
243         .cwnd_event     = tcp_vegas_cwnd_event,
244         .get_info       = tcp_vegas_get_info,
245         .pkts_acked     = tcp_yeah_pkts_acked,
246
247         .owner          = THIS_MODULE,
248         .name           = "yeah",
249 };
250
251 static int __init tcp_yeah_register(void)
252 {
253         BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE);
254         tcp_register_congestion_control(&tcp_yeah);
255         return 0;
256 }
257
258 static void __exit tcp_yeah_unregister(void)
259 {
260         tcp_unregister_congestion_control(&tcp_yeah);
261 }
262
263 module_init(tcp_yeah_register);
264 module_exit(tcp_yeah_unregister);
265
266 MODULE_AUTHOR("Angelo P. Castellani");
267 MODULE_LICENSE("GPL");
268 MODULE_DESCRIPTION("YeAH TCP");