Merge branch 'master'
[linux-2.6] / net / ipv4 / tcp_bic.c
1 /*
2  * Binary Increase Congestion control for TCP
3  *
4  * This is from the implementation of BICTCP in
5  * Lison-Xu, Kahaled Harfoush, and Injong Rhee.
6  *  "Binary Increase Congestion Control for Fast, Long Distance
7  *  Networks" in InfoComm 2004
8  * Available from:
9  *  http://www.csc.ncsu.edu/faculty/rhee/export/bitcp.pdf
10  *
11  * Unless BIC is enabled and congestion window is large
12  * this behaves the same as the original Reno.
13  */
14
15 #include <linux/config.h>
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 low_utilization_threshold = 153;
34 static int low_utilization_period = 2;
35 static int initial_ssthresh = 100;
36 static int smooth_part = 20;
37
38 module_param(fast_convergence, int, 0644);
39 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
40 module_param(max_increment, int, 0644);
41 MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
42 module_param(low_window, int, 0644);
43 MODULE_PARM_DESC(low_window, "lower bound on congestion window (for TCP friendliness)");
44 module_param(beta, int, 0644);
45 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
46 module_param(low_utilization_threshold, int, 0644);
47 MODULE_PARM_DESC(low_utilization_threshold, "percent (scaled by 1024) for low utilization mode");
48 module_param(low_utilization_period, int, 0644);
49 MODULE_PARM_DESC(low_utilization_period, "if average delay exceeds then goto to low utilization mode (seconds)");
50 module_param(initial_ssthresh, int, 0644);
51 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
52 module_param(smooth_part, int, 0644);
53 MODULE_PARM_DESC(smooth_part, "log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax");
54
55
56 /* BIC TCP Parameters */
57 struct bictcp {
58         u32     cnt;            /* increase cwnd by 1 after ACKs */
59         u32     last_max_cwnd;  /* last maximum snd_cwnd */
60         u32     loss_cwnd;      /* congestion window at last loss */
61         u32     last_cwnd;      /* the last snd_cwnd */
62         u32     last_time;      /* time when updated last_cwnd */
63         u32     delay_min;      /* min delay */
64         u32     delay_max;      /* max delay */
65         u32     last_delay;
66         u8      low_utilization;/* 0: high; 1: low */
67         u32     low_utilization_start;  /* starting time of low utilization detection*/
68         u32     epoch_start;    /* beginning of an epoch */
69 #define ACK_RATIO_SHIFT 4
70         u32     delayed_ack;    /* estimate the ratio of Packets/ACKs << 4 */
71 };
72
73 static inline void bictcp_reset(struct bictcp *ca)
74 {
75         ca->cnt = 0;
76         ca->last_max_cwnd = 0;
77         ca->loss_cwnd = 0;
78         ca->last_cwnd = 0;
79         ca->last_time = 0;
80         ca->delay_min = 0;
81         ca->delay_max = 0;
82         ca->last_delay = 0;
83         ca->low_utilization = 0;
84         ca->low_utilization_start = 0;
85         ca->epoch_start = 0;
86         ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
87 }
88
89 static void bictcp_init(struct sock *sk)
90 {
91         bictcp_reset(inet_csk_ca(sk));
92         if (initial_ssthresh)
93                 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
94 }
95
96 /*
97  * Compute congestion window to use.
98  */
99 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
100 {
101         if (ca->last_cwnd == cwnd &&
102             (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
103                 return;
104
105         ca->last_cwnd = cwnd;
106         ca->last_time = tcp_time_stamp;
107
108         if (ca->epoch_start == 0) /* record the beginning of an epoch */
109                 ca->epoch_start = tcp_time_stamp;
110
111         /* start off normal */
112         if (cwnd <= low_window) {
113                 ca->cnt = cwnd;
114                 return;
115         }
116
117         /* binary increase */
118         if (cwnd < ca->last_max_cwnd) {
119                 __u32   dist = (ca->last_max_cwnd - cwnd)
120                         / BICTCP_B;
121
122                 if (dist > max_increment)
123                         /* linear increase */
124                         ca->cnt = cwnd / max_increment;
125                 else if (dist <= 1U)
126                         /* binary search increase */
127                         ca->cnt = (cwnd * smooth_part) / BICTCP_B;
128                 else
129                         /* binary search increase */
130                         ca->cnt = cwnd / dist;
131         } else {
132                 /* slow start AMD linear increase */
133                 if (cwnd < ca->last_max_cwnd + BICTCP_B)
134                         /* slow start */
135                         ca->cnt = (cwnd * smooth_part) / BICTCP_B;
136                 else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
137                         /* slow start */
138                         ca->cnt = (cwnd * (BICTCP_B-1))
139                                 / (cwnd - ca->last_max_cwnd);
140                 else
141                         /* linear increase */
142                         ca->cnt = cwnd / max_increment;
143         }
144
145         /* if in slow start or link utilization is very low */
146         if ( ca->loss_cwnd == 0 ||
147              (cwnd > ca->loss_cwnd && ca->low_utilization)) {
148                 if (ca->cnt > 20) /* increase cwnd 5% per RTT */
149                         ca->cnt = 20;
150         }
151
152         ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
153         if (ca->cnt == 0)                       /* cannot be zero */
154                 ca->cnt = 1;
155 }
156
157
158 /* Detect low utilization in congestion avoidance */
159 static inline void bictcp_low_utilization(struct sock *sk, int flag)
160 {
161         const struct tcp_sock *tp = tcp_sk(sk);
162         struct bictcp *ca = inet_csk_ca(sk);
163         u32 dist, delay;
164
165         /* No time stamp */
166         if (!(tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) ||
167              /* Discard delay samples right after fast recovery */
168              tcp_time_stamp < ca->epoch_start + HZ ||
169              /* this delay samples may not be accurate */
170              flag == 0) {
171                 ca->last_delay = 0;
172                 goto notlow;
173         }
174
175         delay = ca->last_delay<<3;      /* use the same scale as tp->srtt*/
176         ca->last_delay = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
177         if (delay == 0)                 /* no previous delay sample */
178                 goto notlow;
179
180         /* first time call or link delay decreases */
181         if (ca->delay_min == 0 || ca->delay_min > delay) {
182                 ca->delay_min = ca->delay_max = delay;
183                 goto notlow;
184         }
185
186         if (ca->delay_max < delay)
187                 ca->delay_max = delay;
188
189         /* utilization is low, if avg delay < dist*threshold
190            for checking_period time */
191         dist = ca->delay_max - ca->delay_min;
192         if (dist <= ca->delay_min>>6 ||
193             tp->srtt - ca->delay_min >=  (dist*low_utilization_threshold)>>10)
194                 goto notlow;
195
196         if (ca->low_utilization_start == 0) {
197                 ca->low_utilization = 0;
198                 ca->low_utilization_start = tcp_time_stamp;
199         } else if ((s32)(tcp_time_stamp - ca->low_utilization_start)
200                         > low_utilization_period*HZ) {
201                 ca->low_utilization = 1;
202         }
203
204         return;
205
206  notlow:
207         ca->low_utilization = 0;
208         ca->low_utilization_start = 0;
209
210 }
211
212 static void bictcp_cong_avoid(struct sock *sk, u32 ack,
213                               u32 seq_rtt, u32 in_flight, int data_acked)
214 {
215         struct tcp_sock *tp = tcp_sk(sk);
216         struct bictcp *ca = inet_csk_ca(sk);
217
218         bictcp_low_utilization(sk, data_acked);
219
220         if (!tcp_is_cwnd_limited(sk, in_flight))
221                 return;
222
223         if (tp->snd_cwnd <= tp->snd_ssthresh)
224                 tcp_slow_start(tp);
225         else {
226                 bictcp_update(ca, tp->snd_cwnd);
227
228                 /* In dangerous area, increase slowly.
229                  * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
230                  */
231                 if (tp->snd_cwnd_cnt >= ca->cnt) {
232                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
233                                 tp->snd_cwnd++;
234                         tp->snd_cwnd_cnt = 0;
235                 } else
236                         tp->snd_cwnd_cnt++;
237         }
238
239 }
240
241 /*
242  *      behave like Reno until low_window is reached,
243  *      then increase congestion window slowly
244  */
245 static u32 bictcp_recalc_ssthresh(struct sock *sk)
246 {
247         const struct tcp_sock *tp = tcp_sk(sk);
248         struct bictcp *ca = inet_csk_ca(sk);
249
250         ca->epoch_start = 0;    /* end of epoch */
251
252         /* in case of wrong delay_max*/
253         if (ca->delay_min > 0 && ca->delay_max > ca->delay_min)
254                 ca->delay_max = ca->delay_min
255                         + ((ca->delay_max - ca->delay_min)* 90) / 100;
256
257         /* Wmax and fast convergence */
258         if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
259                 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
260                         / (2 * BICTCP_BETA_SCALE);
261         else
262                 ca->last_max_cwnd = tp->snd_cwnd;
263
264         ca->loss_cwnd = tp->snd_cwnd;
265
266
267         if (tp->snd_cwnd <= low_window)
268                 return max(tp->snd_cwnd >> 1U, 2U);
269         else
270                 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
271 }
272
273 static u32 bictcp_undo_cwnd(struct sock *sk)
274 {
275         const struct tcp_sock *tp = tcp_sk(sk);
276         const struct bictcp *ca = inet_csk_ca(sk);
277         return max(tp->snd_cwnd, ca->last_max_cwnd);
278 }
279
280 static u32 bictcp_min_cwnd(struct sock *sk)
281 {
282         const struct tcp_sock *tp = tcp_sk(sk);
283         return tp->snd_ssthresh;
284 }
285
286 static void bictcp_state(struct sock *sk, u8 new_state)
287 {
288         if (new_state == TCP_CA_Loss)
289                 bictcp_reset(inet_csk_ca(sk));
290 }
291
292 /* Track delayed acknowledgement ratio using sliding window
293  * ratio = (15*ratio + sample) / 16
294  */
295 static void bictcp_acked(struct sock *sk, u32 cnt)
296 {
297         const struct inet_connection_sock *icsk = inet_csk(sk);
298
299         if (cnt > 0 &&  icsk->icsk_ca_state == TCP_CA_Open) {
300                 struct bictcp *ca = inet_csk_ca(sk);
301                 cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
302                 ca->delayed_ack += cnt;
303         }
304 }
305
306
307 static struct tcp_congestion_ops bictcp = {
308         .init           = bictcp_init,
309         .ssthresh       = bictcp_recalc_ssthresh,
310         .cong_avoid     = bictcp_cong_avoid,
311         .set_state      = bictcp_state,
312         .undo_cwnd      = bictcp_undo_cwnd,
313         .min_cwnd       = bictcp_min_cwnd,
314         .pkts_acked     = bictcp_acked,
315         .owner          = THIS_MODULE,
316         .name           = "bic",
317 };
318
319 static int __init bictcp_register(void)
320 {
321         BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
322         return tcp_register_congestion_control(&bictcp);
323 }
324
325 static void __exit bictcp_unregister(void)
326 {
327         tcp_unregister_congestion_control(&bictcp);
328 }
329
330 module_init(bictcp_register);
331 module_exit(bictcp_unregister);
332
333 MODULE_AUTHOR("Stephen Hemminger");
334 MODULE_LICENSE("GPL");
335 MODULE_DESCRIPTION("BIC TCP");