Documentation: move nfsroot.txt to filesystems/
[linux-2.6] / net / ipv4 / tcp_bic.c
1 /*
2  * Binary Increase Congestion control for TCP
3  * Home page:
4  *      http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
5  * This is from the implementation of BICTCP in
6  * Lison-Xu, Kahaled Harfoush, and Injong Rhee.
7  *  "Binary Increase Congestion Control for Fast, Long Distance
8  *  Networks" in InfoComm 2004
9  * Available from:
10  *  http://netsrv.csc.ncsu.edu/export/bitcp.pdf
11  *
12  * Unless BIC is enabled and congestion window is large
13  * this behaves the same as the original Reno.
14  */
15
16 #include <linux/mm.h>
17 #include <linux/module.h>
18 #include <net/tcp.h>
19
20
21 #define BICTCP_BETA_SCALE    1024       /* Scale factor beta calculation
22                                          * max_cwnd = snd_cwnd * beta
23                                          */
24 #define BICTCP_B                4        /*
25                                           * In binary search,
26                                           * go to point (max+min)/N
27                                           */
28
29 static int fast_convergence = 1;
30 static int max_increment = 16;
31 static int low_window = 14;
32 static int beta = 819;          /* = 819/1024 (BICTCP_BETA_SCALE) */
33 static int initial_ssthresh;
34 static int smooth_part = 20;
35
36 module_param(fast_convergence, int, 0644);
37 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
38 module_param(max_increment, int, 0644);
39 MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
40 module_param(low_window, int, 0644);
41 MODULE_PARM_DESC(low_window, "lower bound on congestion window (for TCP friendliness)");
42 module_param(beta, int, 0644);
43 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
44 module_param(initial_ssthresh, int, 0644);
45 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
46 module_param(smooth_part, int, 0644);
47 MODULE_PARM_DESC(smooth_part, "log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax");
48
49
50 /* BIC TCP Parameters */
51 struct bictcp {
52         u32     cnt;            /* increase cwnd by 1 after ACKs */
53         u32     last_max_cwnd;  /* last maximum snd_cwnd */
54         u32     loss_cwnd;      /* congestion window at last loss */
55         u32     last_cwnd;      /* the last snd_cwnd */
56         u32     last_time;      /* time when updated last_cwnd */
57         u32     epoch_start;    /* beginning of an epoch */
58 #define ACK_RATIO_SHIFT 4
59         u32     delayed_ack;    /* estimate the ratio of Packets/ACKs << 4 */
60 };
61
62 static inline void bictcp_reset(struct bictcp *ca)
63 {
64         ca->cnt = 0;
65         ca->last_max_cwnd = 0;
66         ca->loss_cwnd = 0;
67         ca->last_cwnd = 0;
68         ca->last_time = 0;
69         ca->epoch_start = 0;
70         ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
71 }
72
73 static void bictcp_init(struct sock *sk)
74 {
75         bictcp_reset(inet_csk_ca(sk));
76         if (initial_ssthresh)
77                 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
78 }
79
80 /*
81  * Compute congestion window to use.
82  */
83 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
84 {
85         if (ca->last_cwnd == cwnd &&
86             (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
87                 return;
88
89         ca->last_cwnd = cwnd;
90         ca->last_time = tcp_time_stamp;
91
92         if (ca->epoch_start == 0) /* record the beginning of an epoch */
93                 ca->epoch_start = tcp_time_stamp;
94
95         /* start off normal */
96         if (cwnd <= low_window) {
97                 ca->cnt = cwnd;
98                 return;
99         }
100
101         /* binary increase */
102         if (cwnd < ca->last_max_cwnd) {
103                 __u32   dist = (ca->last_max_cwnd - cwnd)
104                         / BICTCP_B;
105
106                 if (dist > max_increment)
107                         /* linear increase */
108                         ca->cnt = cwnd / max_increment;
109                 else if (dist <= 1U)
110                         /* binary search increase */
111                         ca->cnt = (cwnd * smooth_part) / BICTCP_B;
112                 else
113                         /* binary search increase */
114                         ca->cnt = cwnd / dist;
115         } else {
116                 /* slow start AMD linear increase */
117                 if (cwnd < ca->last_max_cwnd + BICTCP_B)
118                         /* slow start */
119                         ca->cnt = (cwnd * smooth_part) / BICTCP_B;
120                 else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
121                         /* slow start */
122                         ca->cnt = (cwnd * (BICTCP_B-1))
123                                 / (cwnd - ca->last_max_cwnd);
124                 else
125                         /* linear increase */
126                         ca->cnt = cwnd / max_increment;
127         }
128
129         /* if in slow start or link utilization is very low */
130         if (ca->loss_cwnd == 0) {
131                 if (ca->cnt > 20) /* increase cwnd 5% per RTT */
132                         ca->cnt = 20;
133         }
134
135         ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
136         if (ca->cnt == 0)                       /* cannot be zero */
137                 ca->cnt = 1;
138 }
139
140 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
141 {
142         struct tcp_sock *tp = tcp_sk(sk);
143         struct bictcp *ca = inet_csk_ca(sk);
144
145         if (!tcp_is_cwnd_limited(sk, in_flight))
146                 return;
147
148         if (tp->snd_cwnd <= tp->snd_ssthresh)
149                 tcp_slow_start(tp);
150         else {
151                 bictcp_update(ca, tp->snd_cwnd);
152
153                 /* In dangerous area, increase slowly.
154                  * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
155                  */
156                 if (tp->snd_cwnd_cnt >= ca->cnt) {
157                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
158                                 tp->snd_cwnd++;
159                         tp->snd_cwnd_cnt = 0;
160                 } else
161                         tp->snd_cwnd_cnt++;
162         }
163
164 }
165
166 /*
167  *      behave like Reno until low_window is reached,
168  *      then increase congestion window slowly
169  */
170 static u32 bictcp_recalc_ssthresh(struct sock *sk)
171 {
172         const struct tcp_sock *tp = tcp_sk(sk);
173         struct bictcp *ca = inet_csk_ca(sk);
174
175         ca->epoch_start = 0;    /* end of epoch */
176
177         /* Wmax and fast convergence */
178         if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
179                 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
180                         / (2 * BICTCP_BETA_SCALE);
181         else
182                 ca->last_max_cwnd = tp->snd_cwnd;
183
184         ca->loss_cwnd = tp->snd_cwnd;
185
186
187         if (tp->snd_cwnd <= low_window)
188                 return max(tp->snd_cwnd >> 1U, 2U);
189         else
190                 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
191 }
192
193 static u32 bictcp_undo_cwnd(struct sock *sk)
194 {
195         const struct tcp_sock *tp = tcp_sk(sk);
196         const struct bictcp *ca = inet_csk_ca(sk);
197         return max(tp->snd_cwnd, ca->last_max_cwnd);
198 }
199
200 static void bictcp_state(struct sock *sk, u8 new_state)
201 {
202         if (new_state == TCP_CA_Loss)
203                 bictcp_reset(inet_csk_ca(sk));
204 }
205
206 /* Track delayed acknowledgment ratio using sliding window
207  * ratio = (15*ratio + sample) / 16
208  */
209 static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt)
210 {
211         const struct inet_connection_sock *icsk = inet_csk(sk);
212
213         if (icsk->icsk_ca_state == TCP_CA_Open) {
214                 struct bictcp *ca = inet_csk_ca(sk);
215                 cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
216                 ca->delayed_ack += cnt;
217         }
218 }
219
220
221 static struct tcp_congestion_ops bictcp = {
222         .init           = bictcp_init,
223         .ssthresh       = bictcp_recalc_ssthresh,
224         .cong_avoid     = bictcp_cong_avoid,
225         .set_state      = bictcp_state,
226         .undo_cwnd      = bictcp_undo_cwnd,
227         .pkts_acked     = bictcp_acked,
228         .owner          = THIS_MODULE,
229         .name           = "bic",
230 };
231
232 static int __init bictcp_register(void)
233 {
234         BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
235         return tcp_register_congestion_control(&bictcp);
236 }
237
238 static void __exit bictcp_unregister(void)
239 {
240         tcp_unregister_congestion_control(&bictcp);
241 }
242
243 module_init(bictcp_register);
244 module_exit(bictcp_unregister);
245
246 MODULE_AUTHOR("Stephen Hemminger");
247 MODULE_LICENSE("GPL");
248 MODULE_DESCRIPTION("BIC TCP");