1 // SPDX-License-Identifier: GPL-2.0-only <<
2 /* 1 /*
3 * Binary Increase Congestion control for TCP 2 * Binary Increase Congestion control for TCP
4 * Home page: 3 * Home page:
5 * http://netsrv.csc.ncsu.edu/twiki/bin/v 4 * http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
6 * This is from the implementation of BICTCP i 5 * This is from the implementation of BICTCP in
7 * Lison-Xu, Kahaled Harfoush, and Injong Rhee 6 * Lison-Xu, Kahaled Harfoush, and Injong Rhee.
8 * "Binary Increase Congestion Control for Fa 7 * "Binary Increase Congestion Control for Fast, Long Distance
9 * Networks" in InfoComm 2004 8 * Networks" in InfoComm 2004
10 * Available from: 9 * Available from:
11 * http://netsrv.csc.ncsu.edu/export/bitcp.pd 10 * http://netsrv.csc.ncsu.edu/export/bitcp.pdf
12 * 11 *
13 * Unless BIC is enabled and congestion window 12 * Unless BIC is enabled and congestion window is large
14 * this behaves the same as the original Reno. 13 * this behaves the same as the original Reno.
15 */ 14 */
16 15
17 #include <linux/mm.h> 16 #include <linux/mm.h>
18 #include <linux/module.h> 17 #include <linux/module.h>
19 #include <net/tcp.h> 18 #include <net/tcp.h>
20 19
21 #define BICTCP_BETA_SCALE 1024 /* Sca 20 #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation
22 * max 21 * max_cwnd = snd_cwnd * beta
23 */ 22 */
24 #define BICTCP_B 4 /* 23 #define BICTCP_B 4 /*
25 * In 24 * In binary search,
26 * go 25 * go to point (max+min)/N
27 */ 26 */
28 27
29 static int fast_convergence = 1; 28 static int fast_convergence = 1;
30 static int max_increment = 16; 29 static int max_increment = 16;
31 static int low_window = 14; 30 static int low_window = 14;
32 static int beta = 819; /* = 819/1024 31 static int beta = 819; /* = 819/1024 (BICTCP_BETA_SCALE) */
33 static int initial_ssthresh; 32 static int initial_ssthresh;
34 static int smooth_part = 20; 33 static int smooth_part = 20;
35 34
36 module_param(fast_convergence, int, 0644); 35 module_param(fast_convergence, int, 0644);
37 MODULE_PARM_DESC(fast_convergence, "turn on/of 36 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
38 module_param(max_increment, int, 0644); 37 module_param(max_increment, int, 0644);
39 MODULE_PARM_DESC(max_increment, "Limit on incr 38 MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
40 module_param(low_window, int, 0644); 39 module_param(low_window, int, 0644);
41 MODULE_PARM_DESC(low_window, "lower bound on c 40 MODULE_PARM_DESC(low_window, "lower bound on congestion window (for TCP friendliness)");
42 module_param(beta, int, 0644); 41 module_param(beta, int, 0644);
43 MODULE_PARM_DESC(beta, "beta for multiplicativ 42 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
44 module_param(initial_ssthresh, int, 0644); 43 module_param(initial_ssthresh, int, 0644);
45 MODULE_PARM_DESC(initial_ssthresh, "initial va 44 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
46 module_param(smooth_part, int, 0644); 45 module_param(smooth_part, int, 0644);
47 MODULE_PARM_DESC(smooth_part, "log(B/(B*Smin)) 46 MODULE_PARM_DESC(smooth_part, "log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax");
48 47
49 /* BIC TCP Parameters */ 48 /* BIC TCP Parameters */
50 struct bictcp { 49 struct bictcp {
51 u32 cnt; /* increase cw 50 u32 cnt; /* increase cwnd by 1 after ACKs */
52 u32 last_max_cwnd; /* last maximu 51 u32 last_max_cwnd; /* last maximum snd_cwnd */
>> 52 u32 loss_cwnd; /* congestion window at last loss */
53 u32 last_cwnd; /* the last sn 53 u32 last_cwnd; /* the last snd_cwnd */
54 u32 last_time; /* time when u 54 u32 last_time; /* time when updated last_cwnd */
55 u32 epoch_start; /* beginning o 55 u32 epoch_start; /* beginning of an epoch */
56 #define ACK_RATIO_SHIFT 4 56 #define ACK_RATIO_SHIFT 4
57 u32 delayed_ack; /* estimate th 57 u32 delayed_ack; /* estimate the ratio of Packets/ACKs << 4 */
58 }; 58 };
59 59
60 static inline void bictcp_reset(struct bictcp 60 static inline void bictcp_reset(struct bictcp *ca)
61 { 61 {
62 ca->cnt = 0; 62 ca->cnt = 0;
63 ca->last_max_cwnd = 0; 63 ca->last_max_cwnd = 0;
64 ca->last_cwnd = 0; 64 ca->last_cwnd = 0;
65 ca->last_time = 0; 65 ca->last_time = 0;
66 ca->epoch_start = 0; 66 ca->epoch_start = 0;
67 ca->delayed_ack = 2 << ACK_RATIO_SHIFT 67 ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
68 } 68 }
69 69
70 static void bictcp_init(struct sock *sk) 70 static void bictcp_init(struct sock *sk)
71 { 71 {
72 struct bictcp *ca = inet_csk_ca(sk); 72 struct bictcp *ca = inet_csk_ca(sk);
73 73
74 bictcp_reset(ca); 74 bictcp_reset(ca);
>> 75 ca->loss_cwnd = 0;
75 76
76 if (initial_ssthresh) 77 if (initial_ssthresh)
77 tcp_sk(sk)->snd_ssthresh = ini 78 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
78 } 79 }
79 80
80 /* 81 /*
81 * Compute congestion window to use. 82 * Compute congestion window to use.
82 */ 83 */
83 static inline void bictcp_update(struct bictcp 84 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
84 { 85 {
85 if (ca->last_cwnd == cwnd && 86 if (ca->last_cwnd == cwnd &&
86 (s32)(tcp_jiffies32 - ca->last_tim 87 (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32)
87 return; 88 return;
88 89
89 ca->last_cwnd = cwnd; 90 ca->last_cwnd = cwnd;
90 ca->last_time = tcp_jiffies32; 91 ca->last_time = tcp_jiffies32;
91 92
92 if (ca->epoch_start == 0) /* record th 93 if (ca->epoch_start == 0) /* record the beginning of an epoch */
93 ca->epoch_start = tcp_jiffies3 94 ca->epoch_start = tcp_jiffies32;
94 95
95 /* start off normal */ 96 /* start off normal */
96 if (cwnd <= low_window) { 97 if (cwnd <= low_window) {
97 ca->cnt = cwnd; 98 ca->cnt = cwnd;
98 return; 99 return;
99 } 100 }
100 101
101 /* binary increase */ 102 /* binary increase */
102 if (cwnd < ca->last_max_cwnd) { 103 if (cwnd < ca->last_max_cwnd) {
103 __u32 dist = (ca->last_max_c 104 __u32 dist = (ca->last_max_cwnd - cwnd)
104 / BICTCP_B; 105 / BICTCP_B;
105 106
106 if (dist > max_increment) 107 if (dist > max_increment)
107 /* linear increase */ 108 /* linear increase */
108 ca->cnt = cwnd / max_i 109 ca->cnt = cwnd / max_increment;
109 else if (dist <= 1U) 110 else if (dist <= 1U)
110 /* binary search incre 111 /* binary search increase */
111 ca->cnt = (cwnd * smoo 112 ca->cnt = (cwnd * smooth_part) / BICTCP_B;
112 else 113 else
113 /* binary search incre 114 /* binary search increase */
114 ca->cnt = cwnd / dist; 115 ca->cnt = cwnd / dist;
115 } else { 116 } else {
116 /* slow start AMD linear incre 117 /* slow start AMD linear increase */
117 if (cwnd < ca->last_max_cwnd + 118 if (cwnd < ca->last_max_cwnd + BICTCP_B)
118 /* slow start */ 119 /* slow start */
119 ca->cnt = (cwnd * smoo 120 ca->cnt = (cwnd * smooth_part) / BICTCP_B;
120 else if (cwnd < ca->last_max_c 121 else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
121 /* slow start */ 122 /* slow start */
122 ca->cnt = (cwnd * (BIC 123 ca->cnt = (cwnd * (BICTCP_B-1))
123 / (cwnd - ca-> 124 / (cwnd - ca->last_max_cwnd);
124 else 125 else
125 /* linear increase */ 126 /* linear increase */
126 ca->cnt = cwnd / max_i 127 ca->cnt = cwnd / max_increment;
127 } 128 }
128 129
129 /* if in slow start or link utilizatio 130 /* if in slow start or link utilization is very low */
130 if (ca->last_max_cwnd == 0) { 131 if (ca->last_max_cwnd == 0) {
131 if (ca->cnt > 20) /* increase 132 if (ca->cnt > 20) /* increase cwnd 5% per RTT */
132 ca->cnt = 20; 133 ca->cnt = 20;
133 } 134 }
134 135
135 ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) 136 ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
136 if (ca->cnt == 0) 137 if (ca->cnt == 0) /* cannot be zero */
137 ca->cnt = 1; 138 ca->cnt = 1;
138 } 139 }
139 140
140 static void bictcp_cong_avoid(struct sock *sk, 141 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked)
141 { 142 {
142 struct tcp_sock *tp = tcp_sk(sk); 143 struct tcp_sock *tp = tcp_sk(sk);
143 struct bictcp *ca = inet_csk_ca(sk); 144 struct bictcp *ca = inet_csk_ca(sk);
144 145
145 if (!tcp_is_cwnd_limited(sk)) 146 if (!tcp_is_cwnd_limited(sk))
146 return; 147 return;
147 148
148 if (tcp_in_slow_start(tp)) { !! 149 if (tcp_in_slow_start(tp))
149 acked = tcp_slow_start(tp, ack !! 150 tcp_slow_start(tp, acked);
150 if (!acked) !! 151 else {
151 return; !! 152 bictcp_update(ca, tp->snd_cwnd);
>> 153 tcp_cong_avoid_ai(tp, ca->cnt, 1);
152 } 154 }
153 bictcp_update(ca, tcp_snd_cwnd(tp)); <<
154 tcp_cong_avoid_ai(tp, ca->cnt, acked); <<
155 } 155 }
156 156
157 /* 157 /*
158 * behave like Reno until low_window is r 158 * behave like Reno until low_window is reached,
159 * then increase congestion window slowly 159 * then increase congestion window slowly
160 */ 160 */
161 static u32 bictcp_recalc_ssthresh(struct sock 161 static u32 bictcp_recalc_ssthresh(struct sock *sk)
162 { 162 {
163 const struct tcp_sock *tp = tcp_sk(sk) 163 const struct tcp_sock *tp = tcp_sk(sk);
164 struct bictcp *ca = inet_csk_ca(sk); 164 struct bictcp *ca = inet_csk_ca(sk);
165 165
166 ca->epoch_start = 0; /* end of epoc 166 ca->epoch_start = 0; /* end of epoch */
167 167
168 /* Wmax and fast convergence */ 168 /* Wmax and fast convergence */
169 if (tcp_snd_cwnd(tp) < ca->last_max_cw !! 169 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
170 ca->last_max_cwnd = (tcp_snd_c !! 170 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
171 / (2 * BICTCP_BETA_SCA 171 / (2 * BICTCP_BETA_SCALE);
172 else 172 else
173 ca->last_max_cwnd = tcp_snd_cw !! 173 ca->last_max_cwnd = tp->snd_cwnd;
174 174
175 if (tcp_snd_cwnd(tp) <= low_window) !! 175 ca->loss_cwnd = tp->snd_cwnd;
176 return max(tcp_snd_cwnd(tp) >> !! 176
>> 177 if (tp->snd_cwnd <= low_window)
>> 178 return max(tp->snd_cwnd >> 1U, 2U);
177 else 179 else
178 return max((tcp_snd_cwnd(tp) * !! 180 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
>> 181 }
>> 182
>> 183 static u32 bictcp_undo_cwnd(struct sock *sk)
>> 184 {
>> 185 const struct tcp_sock *tp = tcp_sk(sk);
>> 186 const struct bictcp *ca = inet_csk_ca(sk);
>> 187
>> 188 return max(tp->snd_cwnd, ca->loss_cwnd);
179 } 189 }
180 190
181 static void bictcp_state(struct sock *sk, u8 n 191 static void bictcp_state(struct sock *sk, u8 new_state)
182 { 192 {
183 if (new_state == TCP_CA_Loss) 193 if (new_state == TCP_CA_Loss)
184 bictcp_reset(inet_csk_ca(sk)); 194 bictcp_reset(inet_csk_ca(sk));
185 } 195 }
186 196
187 /* Track delayed acknowledgment ratio using sl 197 /* Track delayed acknowledgment ratio using sliding window
188 * ratio = (15*ratio + sample) / 16 198 * ratio = (15*ratio + sample) / 16
189 */ 199 */
190 static void bictcp_acked(struct sock *sk, cons 200 static void bictcp_acked(struct sock *sk, const struct ack_sample *sample)
191 { 201 {
192 const struct inet_connection_sock *ics 202 const struct inet_connection_sock *icsk = inet_csk(sk);
193 203
194 if (icsk->icsk_ca_state == TCP_CA_Open 204 if (icsk->icsk_ca_state == TCP_CA_Open) {
195 struct bictcp *ca = inet_csk_c 205 struct bictcp *ca = inet_csk_ca(sk);
196 206
197 ca->delayed_ack += sample->pkt 207 ca->delayed_ack += sample->pkts_acked -
198 (ca->delayed_ack >> AC 208 (ca->delayed_ack >> ACK_RATIO_SHIFT);
199 } 209 }
200 } 210 }
201 211
202 static struct tcp_congestion_ops bictcp __read 212 static struct tcp_congestion_ops bictcp __read_mostly = {
203 .init = bictcp_init, 213 .init = bictcp_init,
204 .ssthresh = bictcp_recalc_ssthre 214 .ssthresh = bictcp_recalc_ssthresh,
205 .cong_avoid = bictcp_cong_avoid, 215 .cong_avoid = bictcp_cong_avoid,
206 .set_state = bictcp_state, 216 .set_state = bictcp_state,
207 .undo_cwnd = tcp_reno_undo_cwnd, !! 217 .undo_cwnd = bictcp_undo_cwnd,
208 .pkts_acked = bictcp_acked, 218 .pkts_acked = bictcp_acked,
209 .owner = THIS_MODULE, 219 .owner = THIS_MODULE,
210 .name = "bic", 220 .name = "bic",
211 }; 221 };
212 222
213 static int __init bictcp_register(void) 223 static int __init bictcp_register(void)
214 { 224 {
215 BUILD_BUG_ON(sizeof(struct bictcp) > I 225 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
216 return tcp_register_congestion_control 226 return tcp_register_congestion_control(&bictcp);
217 } 227 }
218 228
219 static void __exit bictcp_unregister(void) 229 static void __exit bictcp_unregister(void)
220 { 230 {
221 tcp_unregister_congestion_control(&bic 231 tcp_unregister_congestion_control(&bictcp);
222 } 232 }
223 233
224 module_init(bictcp_register); 234 module_init(bictcp_register);
225 module_exit(bictcp_unregister); 235 module_exit(bictcp_unregister);
226 236
227 MODULE_AUTHOR("Stephen Hemminger"); 237 MODULE_AUTHOR("Stephen Hemminger");
228 MODULE_LICENSE("GPL"); 238 MODULE_LICENSE("GPL");
229 MODULE_DESCRIPTION("BIC TCP"); 239 MODULE_DESCRIPTION("BIC TCP");
230 240
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.