1 /* 2 * H-TCP congestion control. The algorithm is detailed in: 3 * R.N.Shorten, D.J.Leith: 4 * "H-TCP: TCP for high-speed and long-distance networks" 5 * Proc. PFLDnet, Argonne, 2004. 6 * http://www.hamilton.ie/net/htcp3.pdf 7 */ 8 9 #include <linux/config.h> 10 #include <linux/mm.h> 11 #include <linux/module.h> 12 #include <net/tcp.h> 13 14 #define ALPHA_BASE (1<<7) /* 1.0 with shift << 7 */ 15 #define BETA_MIN (1<<6) /* 0.5 with shift << 7 */ 16 #define BETA_MAX 102 /* 0.8 with shift << 7 */ 17 18 static int use_rtt_scaling = 1; 19 module_param(use_rtt_scaling, int, 0644); 20 MODULE_PARM_DESC(use_rtt_scaling, "turn on/off RTT scaling"); 21 22 static int use_bandwidth_switch = 1; 23 module_param(use_bandwidth_switch, int, 0644); 24 MODULE_PARM_DESC(use_bandwidth_switch, "turn on/off bandwidth switcher"); 25 26 struct htcp { 27 u16 alpha; /* Fixed point arith, << 7 */ 28 u8 beta; /* Fixed point arith, << 7 */ 29 u8 modeswitch; /* Delay modeswitch until we had at least one congestion event */ 30 u8 ccount; /* Number of RTTs since last congestion event */ 31 u8 undo_ccount; 32 u16 packetcount; 33 u32 minRTT; 34 u32 maxRTT; 35 u32 snd_cwnd_cnt2; 36 37 u32 undo_maxRTT; 38 u32 undo_old_maxB; 39 40 /* Bandwidth estimation */ 41 u32 minB; 42 u32 maxB; 43 u32 old_maxB; 44 u32 Bi; 45 u32 lasttime; 46 }; 47 48 static inline void htcp_reset(struct htcp *ca) 49 { 50 ca->undo_ccount = ca->ccount; 51 ca->undo_maxRTT = ca->maxRTT; 52 ca->undo_old_maxB = ca->old_maxB; 53 54 ca->ccount = 0; 55 ca->snd_cwnd_cnt2 = 0; 56 } 57 58 static u32 htcp_cwnd_undo(struct sock *sk) 59 { 60 const struct tcp_sock *tp = tcp_sk(sk); 61 struct htcp *ca = inet_csk_ca(sk); 62 ca->ccount = ca->undo_ccount; 63 ca->maxRTT = ca->undo_maxRTT; 64 ca->old_maxB = ca->undo_old_maxB; 65 return max(tp->snd_cwnd, (tp->snd_ssthresh<<7)/ca->beta); 66 } 67 68 static inline void measure_rtt(struct sock *sk) 69 { 70 const struct inet_connection_sock *icsk = inet_csk(sk); 71 const struct tcp_sock *tp = tcp_sk(sk); 72 struct htcp *ca = inet_csk_ca(sk); 73 u32 srtt = tp->srtt>>3; 74 75 /* keep track of minimum RTT seen so far, minRTT is zero at first */ 76 if (ca->minRTT > srtt || !ca->minRTT) 77 ca->minRTT = srtt; 78 79 /* max RTT */ 80 if (icsk->icsk_ca_state == TCP_CA_Open && tp->snd_ssthresh < 0xFFFF && ca->ccount > 3) { 81 if (ca->maxRTT < ca->minRTT) 82 ca->maxRTT = ca->minRTT; 83 if (ca->maxRTT < srtt && srtt <= ca->maxRTT+HZ/50) 84 ca->maxRTT = srtt; 85 } 86 } 87 88 static void measure_achieved_throughput(struct sock *sk, u32 pkts_acked) 89 { 90 const struct inet_connection_sock *icsk = inet_csk(sk); 91 const struct tcp_sock *tp = tcp_sk(sk); 92 struct htcp *ca = inet_csk_ca(sk); 93 u32 now = tcp_time_stamp; 94 95 /* achieved throughput calculations */ 96 if (icsk->icsk_ca_state != TCP_CA_Open && 97 icsk->icsk_ca_state != TCP_CA_Disorder) { 98 ca->packetcount = 0; 99 ca->lasttime = now; 100 return; 101 } 102 103 ca->packetcount += pkts_acked; 104 105 if (ca->packetcount >= tp->snd_cwnd - (ca->alpha>>7? : 1) 106 && now - ca->lasttime >= ca->minRTT 107 && ca->minRTT > 0) { 108 __u32 cur_Bi = ca->packetcount*HZ/(now - ca->lasttime); 109 if (ca->ccount <= 3) { 110 /* just after backoff */ 111 ca->minB = ca->maxB = ca->Bi = cur_Bi; 112 } else { 113 ca->Bi = (3*ca->Bi + cur_Bi)/4; 114 if (ca->Bi > ca->maxB) 115 ca->maxB = ca->Bi; 116 if (ca->minB > ca->maxB) 117 ca->minB = ca->maxB; 118 } 119 ca->packetcount = 0; 120 ca->lasttime = now; 121 } 122 } 123 124 static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT) 125 { 126 if (use_bandwidth_switch) { 127 u32 maxB = ca->maxB; 128 u32 old_maxB = ca->old_maxB; 129 ca->old_maxB = ca->maxB; 130 131 if (!between(5*maxB, 4*old_maxB, 6*old_maxB)) { 132 ca->beta = BETA_MIN; 133 ca->modeswitch = 0; 134 return; 135 } 136 } 137 138 if (ca->modeswitch && minRTT > max(HZ/100, 1) && maxRTT) { 139 ca->beta = (minRTT<<7)/maxRTT; 140 if (ca->beta < BETA_MIN) 141 ca->beta = BETA_MIN; 142 else if (ca->beta > BETA_MAX) 143 ca->beta = BETA_MAX; 144 } else { 145 ca->beta = BETA_MIN; 146 ca->modeswitch = 1; 147 } 148 } 149 150 static inline void htcp_alpha_update(struct htcp *ca) 151 { 152 u32 minRTT = ca->minRTT; 153 u32 factor = 1; 154 u32 diff = ca->ccount * minRTT; /* time since last backoff */ 155 156 if (diff > HZ) { 157 diff -= HZ; 158 factor = 1+ ( 10*diff + ((diff/2)*(diff/2)/HZ) )/HZ; 159 } 160 161 if (use_rtt_scaling && minRTT) { 162 u32 scale = (HZ<<3)/(10*minRTT); 163 scale = min(max(scale, 1U<<2), 10U<<3); /* clamping ratio to interval [0.5,10]<<3 */ 164 factor = (factor<<3)/scale; 165 if (!factor) 166 factor = 1; 167 } 168 169 ca->alpha = 2*factor*((1<<7)-ca->beta); 170 if (!ca->alpha) 171 ca->alpha = ALPHA_BASE; 172 } 173 174 /* After we have the rtt data to calculate beta, we'd still prefer to wait one 175 * rtt before we adjust our beta to ensure we are working from a consistent 176 * data. 177 * 178 * This function should be called when we hit a congestion event since only at 179 * that point do we really have a real sense of maxRTT (the queues en route 180 * were getting just too full now). 181 */ 182 static void htcp_param_update(struct sock *sk) 183 { 184 struct htcp *ca = inet_csk_ca(sk); 185 u32 minRTT = ca->minRTT; 186 u32 maxRTT = ca->maxRTT; 187 188 htcp_beta_update(ca, minRTT, maxRTT); 189 htcp_alpha_update(ca); 190 191 /* add slowly fading memory for maxRTT to accommodate routing changes etc */ 192 if (minRTT > 0 && maxRTT > minRTT) 193 ca->maxRTT = minRTT + ((maxRTT-minRTT)*95)/100; 194 } 195 196 static u32 htcp_recalc_ssthresh(struct sock *sk) 197 { 198 const struct tcp_sock *tp = tcp_sk(sk); 199 const struct htcp *ca = inet_csk_ca(sk); 200 htcp_param_update(sk); 201 return max((tp->snd_cwnd * ca->beta) >> 7, 2U); 202 } 203 204 static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt, 205 u32 in_flight, int data_acked) 206 { 207 struct tcp_sock *tp = tcp_sk(sk); 208 struct htcp *ca = inet_csk_ca(sk); 209 210 if (in_flight < tp->snd_cwnd) 211 return; 212 213 if (tp->snd_cwnd <= tp->snd_ssthresh) { 214 /* In "safe" area, increase. */ 215 if (tp->snd_cwnd < tp->snd_cwnd_clamp) 216 tp->snd_cwnd++; 217 } else { 218 measure_rtt(sk); 219 220 /* keep track of number of round-trip times since last backoff event */ 221 if (ca->snd_cwnd_cnt2++ > tp->snd_cwnd) { 222 ca->ccount++; 223 ca->snd_cwnd_cnt2 = 0; 224 htcp_alpha_update(ca); 225 } 226 227 /* In dangerous area, increase slowly. 228 * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd 229 */ 230 if ((tp->snd_cwnd_cnt++ * ca->alpha)>>7 >= tp->snd_cwnd) { 231 if (tp->snd_cwnd < tp->snd_cwnd_clamp) 232 tp->snd_cwnd++; 233 tp->snd_cwnd_cnt = 0; 234 ca->ccount++; 235 } 236 } 237 } 238 239 /* Lower bound on congestion window. */ 240 static u32 htcp_min_cwnd(struct sock *sk) 241 { 242 const struct tcp_sock *tp = tcp_sk(sk); 243 return tp->snd_ssthresh; 244 } 245 246 247 static void htcp_init(struct sock *sk) 248 { 249 struct htcp *ca = inet_csk_ca(sk); 250 251 memset(ca, 0, sizeof(struct htcp)); 252 ca->alpha = ALPHA_BASE; 253 ca->beta = BETA_MIN; 254 } 255 256 static void htcp_state(struct sock *sk, u8 new_state) 257 { 258 switch (new_state) { 259 case TCP_CA_CWR: 260 case TCP_CA_Recovery: 261 case TCP_CA_Loss: 262 htcp_reset(inet_csk_ca(sk)); 263 break; 264 } 265 } 266 267 static struct tcp_congestion_ops htcp = { 268 .init = htcp_init, 269 .ssthresh = htcp_recalc_ssthresh, 270 .min_cwnd = htcp_min_cwnd, 271 .cong_avoid = htcp_cong_avoid, 272 .set_state = htcp_state, 273 .undo_cwnd = htcp_cwnd_undo, 274 .pkts_acked = measure_achieved_throughput, 275 .owner = THIS_MODULE, 276 .name = "htcp", 277 }; 278 279 static int __init htcp_register(void) 280 { 281 BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE); 282 BUILD_BUG_ON(BETA_MIN >= BETA_MAX); 283 if (!use_bandwidth_switch) 284 htcp.pkts_acked = NULL; 285 return tcp_register_congestion_control(&htcp); 286 } 287 288 static void __exit htcp_unregister(void) 289 { 290 tcp_unregister_congestion_control(&htcp); 291 } 292 293 module_init(htcp_register); 294 module_exit(htcp_unregister); 295 296 MODULE_AUTHOR("Baruch Even"); 297 MODULE_LICENSE("GPL"); 298 MODULE_DESCRIPTION("H-TCP"); 299