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