1/* 2 * TCP Illinois congestion control. 3 * Home page: 4 * http://www.ews.uiuc.edu/~shaoliu/tcpillinois/index.html 5 * 6 * The algorithm is described in: 7 * "TCP-Illinois: A Loss and Delay-Based Congestion Control Algorithm 8 * for High-Speed Networks" 9 * http://www.ifp.illinois.edu/~srikant/Papers/liubassri06perf.pdf 10 * 11 * Implemented from description in paper and ns-2 simulation. 12 * Copyright (C) 2007 Stephen Hemminger <shemminger@linux-foundation.org> 13 */ 14 15#include <linux/module.h> 16#include <linux/skbuff.h> 17#include <linux/inet_diag.h> 18#include <asm/div64.h> 19#include <net/tcp.h> 20 21#define ALPHA_SHIFT 7 22#define ALPHA_SCALE (1u<<ALPHA_SHIFT) 23#define ALPHA_MIN ((3*ALPHA_SCALE)/10) /* ~0.3 */ 24#define ALPHA_MAX (10*ALPHA_SCALE) /* 10.0 */ 25#define ALPHA_BASE ALPHA_SCALE /* 1.0 */ 26#define RTT_MAX (U32_MAX / ALPHA_MAX) /* 3.3 secs */ 27 28#define BETA_SHIFT 6 29#define BETA_SCALE (1u<<BETA_SHIFT) 30#define BETA_MIN (BETA_SCALE/8) /* 0.125 */ 31#define BETA_MAX (BETA_SCALE/2) /* 0.5 */ 32#define BETA_BASE BETA_MAX 33 34static int win_thresh __read_mostly = 15; 35module_param(win_thresh, int, 0); 36MODULE_PARM_DESC(win_thresh, "Window threshold for starting adaptive sizing"); 37 38static int theta __read_mostly = 5; 39module_param(theta, int, 0); 40MODULE_PARM_DESC(theta, "# of fast RTT's before full growth"); 41 42/* TCP Illinois Parameters */ 43struct illinois { 44 u64 sum_rtt; /* sum of rtt's measured within last rtt */ 45 u16 cnt_rtt; /* # of rtts measured within last rtt */ 46 u32 base_rtt; /* min of all rtt in usec */ 47 u32 max_rtt; /* max of all rtt in usec */ 48 u32 end_seq; /* right edge of current RTT */ 49 u32 alpha; /* Additive increase */ 50 u32 beta; /* Muliplicative decrease */ 51 u16 acked; /* # packets acked by current ACK */ 52 u8 rtt_above; /* average rtt has gone above threshold */ 53 u8 rtt_low; /* # of rtts measurements below threshold */ 54}; 55 56static void rtt_reset(struct sock *sk) 57{ 58 struct tcp_sock *tp = tcp_sk(sk); 59 struct illinois *ca = inet_csk_ca(sk); 60 61 ca->end_seq = tp->snd_nxt; 62 ca->cnt_rtt = 0; 63 ca->sum_rtt = 0; 64 65 /* TODO: age max_rtt? */ 66} 67 68static void tcp_illinois_init(struct sock *sk) 69{ 70 struct illinois *ca = inet_csk_ca(sk); 71 72 ca->alpha = ALPHA_MAX; 73 ca->beta = BETA_BASE; 74 ca->base_rtt = 0x7fffffff; 75 ca->max_rtt = 0; 76 77 ca->acked = 0; 78 ca->rtt_low = 0; 79 ca->rtt_above = 0; 80 81 rtt_reset(sk); 82} 83 84/* Measure RTT for each ack. */ 85static void tcp_illinois_acked(struct sock *sk, u32 pkts_acked, s32 rtt) 86{ 87 struct illinois *ca = inet_csk_ca(sk); 88 89 ca->acked = pkts_acked; 90 91 /* dup ack, no rtt sample */ 92 if (rtt < 0) 93 return; 94 95 /* ignore bogus values, this prevents wraparound in alpha math */ 96 if (rtt > RTT_MAX) 97 rtt = RTT_MAX; 98 99 /* keep track of minimum RTT seen so far */ 100 if (ca->base_rtt > rtt) 101 ca->base_rtt = rtt; 102 103 /* and max */ 104 if (ca->max_rtt < rtt) 105 ca->max_rtt = rtt; 106 107 ++ca->cnt_rtt; 108 ca->sum_rtt += rtt; 109} 110 111/* Maximum queuing delay */ 112static inline u32 max_delay(const struct illinois *ca) 113{ 114 return ca->max_rtt - ca->base_rtt; 115} 116 117/* Average queuing delay */ 118static inline u32 avg_delay(const struct illinois *ca) 119{ 120 u64 t = ca->sum_rtt; 121 122 do_div(t, ca->cnt_rtt); 123 return t - ca->base_rtt; 124} 125 126/* 127 * Compute value of alpha used for additive increase. 128 * If small window then use 1.0, equivalent to Reno. 129 * 130 * For larger windows, adjust based on average delay. 131 * A. If average delay is at minimum (we are uncongested), 132 * then use large alpha (10.0) to increase faster. 133 * B. If average delay is at maximum (getting congested) 134 * then use small alpha (0.3) 135 * 136 * The result is a convex window growth curve. 137 */ 138static u32 alpha(struct illinois *ca, u32 da, u32 dm) 139{ 140 u32 d1 = dm / 100; /* Low threshold */ 141 142 if (da <= d1) { 143 /* If never got out of low delay zone, then use max */ 144 if (!ca->rtt_above) 145 return ALPHA_MAX; 146 147 /* Wait for 5 good RTT's before allowing alpha to go alpha max. 148 * This prevents one good RTT from causing sudden window increase. 149 */ 150 if (++ca->rtt_low < theta) 151 return ca->alpha; 152 153 ca->rtt_low = 0; 154 ca->rtt_above = 0; 155 return ALPHA_MAX; 156 } 157 158 ca->rtt_above = 1; 159 160 /* 161 * Based on: 162 * 163 * (dm - d1) amin amax 164 * k1 = ------------------- 165 * amax - amin 166 * 167 * (dm - d1) amin 168 * k2 = ---------------- - d1 169 * amax - amin 170 * 171 * k1 172 * alpha = ---------- 173 * k2 + da 174 */ 175 176 dm -= d1; 177 da -= d1; 178 return (dm * ALPHA_MAX) / 179 (dm + (da * (ALPHA_MAX - ALPHA_MIN)) / ALPHA_MIN); 180} 181 182/* 183 * Beta used for multiplicative decrease. 184 * For small window sizes returns same value as Reno (0.5) 185 * 186 * If delay is small (10% of max) then beta = 1/8 187 * If delay is up to 80% of max then beta = 1/2 188 * In between is a linear function 189 */ 190static u32 beta(u32 da, u32 dm) 191{ 192 u32 d2, d3; 193 194 d2 = dm / 10; 195 if (da <= d2) 196 return BETA_MIN; 197 198 d3 = (8 * dm) / 10; 199 if (da >= d3 || d3 <= d2) 200 return BETA_MAX; 201 202 /* 203 * Based on: 204 * 205 * bmin d3 - bmax d2 206 * k3 = ------------------- 207 * d3 - d2 208 * 209 * bmax - bmin 210 * k4 = ------------- 211 * d3 - d2 212 * 213 * b = k3 + k4 da 214 */ 215 return (BETA_MIN * d3 - BETA_MAX * d2 + (BETA_MAX - BETA_MIN) * da) 216 / (d3 - d2); 217} 218 219/* Update alpha and beta values once per RTT */ 220static void update_params(struct sock *sk) 221{ 222 struct tcp_sock *tp = tcp_sk(sk); 223 struct illinois *ca = inet_csk_ca(sk); 224 225 if (tp->snd_cwnd < win_thresh) { 226 ca->alpha = ALPHA_BASE; 227 ca->beta = BETA_BASE; 228 } else if (ca->cnt_rtt > 0) { 229 u32 dm = max_delay(ca); 230 u32 da = avg_delay(ca); 231 232 ca->alpha = alpha(ca, da, dm); 233 ca->beta = beta(da, dm); 234 } 235 236 rtt_reset(sk); 237} 238 239/* 240 * In case of loss, reset to default values 241 */ 242static void tcp_illinois_state(struct sock *sk, u8 new_state) 243{ 244 struct illinois *ca = inet_csk_ca(sk); 245 246 if (new_state == TCP_CA_Loss) { 247 ca->alpha = ALPHA_BASE; 248 ca->beta = BETA_BASE; 249 ca->rtt_low = 0; 250 ca->rtt_above = 0; 251 rtt_reset(sk); 252 } 253} 254 255/* 256 * Increase window in response to successful acknowledgment. 257 */ 258static void tcp_illinois_cong_avoid(struct sock *sk, u32 ack, u32 acked) 259{ 260 struct tcp_sock *tp = tcp_sk(sk); 261 struct illinois *ca = inet_csk_ca(sk); 262 263 if (after(ack, ca->end_seq)) 264 update_params(sk); 265 266 /* RFC2861 only increase cwnd if fully utilized */ 267 if (!tcp_is_cwnd_limited(sk)) 268 return; 269 270 /* In slow start */ 271 if (tcp_in_slow_start(tp)) 272 tcp_slow_start(tp, acked); 273 274 else { 275 u32 delta; 276 277 /* snd_cwnd_cnt is # of packets since last cwnd increment */ 278 tp->snd_cwnd_cnt += ca->acked; 279 ca->acked = 1; 280 281 /* This is close approximation of: 282 * tp->snd_cwnd += alpha/tp->snd_cwnd 283 */ 284 delta = (tp->snd_cwnd_cnt * ca->alpha) >> ALPHA_SHIFT; 285 if (delta >= tp->snd_cwnd) { 286 tp->snd_cwnd = min(tp->snd_cwnd + delta / tp->snd_cwnd, 287 (u32)tp->snd_cwnd_clamp); 288 tp->snd_cwnd_cnt = 0; 289 } 290 } 291} 292 293static u32 tcp_illinois_ssthresh(struct sock *sk) 294{ 295 struct tcp_sock *tp = tcp_sk(sk); 296 struct illinois *ca = inet_csk_ca(sk); 297 298 /* Multiplicative decrease */ 299 return max(tp->snd_cwnd - ((tp->snd_cwnd * ca->beta) >> BETA_SHIFT), 2U); 300} 301 302/* Extract info for Tcp socket info provided via netlink. */ 303static size_t tcp_illinois_info(struct sock *sk, u32 ext, int *attr, 304 union tcp_cc_info *info) 305{ 306 const struct illinois *ca = inet_csk_ca(sk); 307 308 if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) { 309 info->vegas.tcpv_enabled = 1; 310 info->vegas.tcpv_rttcnt = ca->cnt_rtt; 311 info->vegas.tcpv_minrtt = ca->base_rtt; 312 info->vegas.tcpv_rtt = 0; 313 314 if (info->vegas.tcpv_rttcnt > 0) { 315 u64 t = ca->sum_rtt; 316 317 do_div(t, info->vegas.tcpv_rttcnt); 318 info->vegas.tcpv_rtt = t; 319 } 320 *attr = INET_DIAG_VEGASINFO; 321 return sizeof(struct tcpvegas_info); 322 } 323 return 0; 324} 325 326static struct tcp_congestion_ops tcp_illinois __read_mostly = { 327 .init = tcp_illinois_init, 328 .ssthresh = tcp_illinois_ssthresh, 329 .cong_avoid = tcp_illinois_cong_avoid, 330 .set_state = tcp_illinois_state, 331 .get_info = tcp_illinois_info, 332 .pkts_acked = tcp_illinois_acked, 333 334 .owner = THIS_MODULE, 335 .name = "illinois", 336}; 337 338static int __init tcp_illinois_register(void) 339{ 340 BUILD_BUG_ON(sizeof(struct illinois) > ICSK_CA_PRIV_SIZE); 341 return tcp_register_congestion_control(&tcp_illinois); 342} 343 344static void __exit tcp_illinois_unregister(void) 345{ 346 tcp_unregister_congestion_control(&tcp_illinois); 347} 348 349module_init(tcp_illinois_register); 350module_exit(tcp_illinois_unregister); 351 352MODULE_AUTHOR("Stephen Hemminger, Shao Liu"); 353MODULE_LICENSE("GPL"); 354MODULE_DESCRIPTION("TCP Illinois"); 355MODULE_VERSION("1.0"); 356