net/ipv4/tcp_yeah.c

   1 /*
   2  *
   3  *   YeAH TCP
   4  *
   5  * For further details look at:
   6  *    http://wil.cs.caltech.edu/pfldnet2007/paper/YeAH_TCP.pdf
   7  *
   8  */
   9
  10 #include "tcp_yeah.h"
  11
  12 /* Default values of the Vegas variables, in fixed-point representation
  13  * with V_PARAM_SHIFT bits to the right of the binary point.
  14  */
  15 #define V_PARAM_SHIFT 1
  16
  17 #define TCP_YEAH_ALPHA       80 //lin number of packets queued at the bottleneck
  18 #define TCP_YEAH_GAMMA        1 //lin fraction of queue to be removed per rtt
  19 #define TCP_YEAH_DELTA        3 //log minimum fraction of cwnd to be removed on loss
  20 #define TCP_YEAH_EPSILON      1 //log maximum fraction to be removed on early decongestion
  21 #define TCP_YEAH_PHY          8 //lin maximum delta from base
  22 #define TCP_YEAH_RHO         16 //lin minumum number of consecutive rtt to consider competition on loss
  23 #define TCP_YEAH_ZETA        50 //lin minimum number of state switchs to reset reno_count
  24
  25 #define TCP_SCALABLE_AI_CNT      100U
  26
  27 /* YeAH variables */
  28 struct yeah {
  29         /* Vegas */
  30         u32     beg_snd_nxt;    /* right edge during last RTT */
  31         u32     beg_snd_una;    /* left edge  during last RTT */
  32         u32     beg_snd_cwnd;   /* saves the size of the cwnd */
  33         u8      doing_vegas_now;/* if true, do vegas for this RTT */
  34         u16     cntRTT;         /* # of RTTs measured within last RTT */
  35         u32     minRTT;         /* min of RTTs measured within last RTT (in usec) */
  36         u32     baseRTT;        /* the min of all Vegas RTT measurements seen (in usec) */
  37
  38         /* YeAH */
  39         u32 lastQ;
  40         u32 doing_reno_now;
  41
  42         u32 reno_count;
  43         u32 fast_count;
  44
  45         u32 pkts_acked;
  46 };
  47
  48 static void tcp_yeah_init(struct sock *sk)
  49 {
  50         struct tcp_sock *tp = tcp_sk(sk);
  51         struct yeah *yeah = inet_csk_ca(sk);
  52
  53         tcp_vegas_init(sk);
  54
  55         yeah->doing_reno_now = 0;
  56         yeah->lastQ = 0;
  57
  58         yeah->reno_count = 2;
  59
  60         /* Ensure the MD arithmetic works.  This is somewhat pedantic,
  61          * since I don't think we will see a cwnd this large. :) */
  62         tp->snd_cwnd_clamp = min_t(u32, tp->snd_cwnd_clamp, 0xffffffff/128);
  63
  64 }
  65
  66
  67 static void tcp_yeah_pkts_acked(struct sock *sk, u32 pkts_acked, ktime_t last)
  68 {
  69         const struct inet_connection_sock *icsk = inet_csk(sk);
  70         struct yeah *yeah = inet_csk_ca(sk);
  71
  72         if (icsk->icsk_ca_state == TCP_CA_Open)
  73                 yeah->pkts_acked = pkts_acked;
  74
  75         tcp_vegas_pkts_acked(sk, pkts_acked, last);
  76 }
  77
  78 static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack,
  79                                 u32 seq_rtt, u32 in_flight, int flag)
  80 {
  81         struct tcp_sock *tp = tcp_sk(sk);
  82         struct yeah *yeah = inet_csk_ca(sk);
  83
  84         if (!tcp_is_cwnd_limited(sk, in_flight))
  85                 return;
  86
  87         if (tp->snd_cwnd <= tp->snd_ssthresh) {
  88                 tcp_slow_start(tp);
  89         } else if (!yeah->doing_reno_now) {
  90                 /* Scalable */
  91
  92                 tp->snd_cwnd_cnt+=yeah->pkts_acked;
  93                 if (tp->snd_cwnd_cnt > min(tp->snd_cwnd, TCP_SCALABLE_AI_CNT)){
  94                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
  95                                 tp->snd_cwnd++;
  96                         tp->snd_cwnd_cnt = 0;
  97                 }
  98
  99                 yeah->pkts_acked = 1;
 100
 101         } else {
 102                 /* Reno */
 103
 104                 if (tp->snd_cwnd_cnt < tp->snd_cwnd)
 105                         tp->snd_cwnd_cnt++;
 106
 107                 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
 108                         tp->snd_cwnd++;
 109                         tp->snd_cwnd_cnt = 0;
 110                 }
 111         }
 112
 113         /* The key players are v_beg_snd_una and v_beg_snd_nxt.
 114          *
 115          * These are so named because they represent the approximate values
 116          * of snd_una and snd_nxt at the beginning of the current RTT. More
 117          * precisely, they represent the amount of data sent during the RTT.
 118          * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
 119          * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
 120          * bytes of data have been ACKed during the course of the RTT, giving
 121          * an "actual" rate of:
 122          *
 123          *     (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
 124          *
 125          * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
 126          * because delayed ACKs can cover more than one segment, so they
 127          * don't line up yeahly with the boundaries of RTTs.
 128          *
 129          * Another unfortunate fact of life is that delayed ACKs delay the
 130          * advance of the left edge of our send window, so that the number
 131          * of bytes we send in an RTT is often less than our cwnd will allow.
 132          * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
 133          */
 134
 135         if (after(ack, yeah->beg_snd_nxt)) {
 136
 137                 /* We do the Vegas calculations only if we got enough RTT
 138                  * samples that we can be reasonably sure that we got
 139                  * at least one RTT sample that wasn't from a delayed ACK.
 140                  * If we only had 2 samples total,
 141                  * then that means we're getting only 1 ACK per RTT, which
 142                  * means they're almost certainly delayed ACKs.
 143                  * If  we have 3 samples, we should be OK.
 144                  */
 145
 146                 if (yeah->cntRTT > 2) {
 147                         u32 rtt, queue;
 148                         u64 bw;
 149
 150                         /* We have enough RTT samples, so, using the Vegas
 151                          * algorithm, we determine if we should increase or
 152                          * decrease cwnd, and by how much.
 153                          */
 154
 155                         /* Pluck out the RTT we are using for the Vegas
 156                          * calculations. This is the min RTT seen during the
 157                          * last RTT. Taking the min filters out the effects
 158                          * of delayed ACKs, at the cost of noticing congestion
 159                          * a bit later.
 160                          */
 161                         rtt = yeah->minRTT;
 162
 163                         /* Compute excess number of packets above bandwidth
 164                          * Avoid doing full 64 bit divide.
 165                          */
 166                         bw = tp->snd_cwnd;
 167                         bw *= rtt - yeah->baseRTT;
 168                         do_div(bw, rtt);
 169                         queue = bw;
 170
 171                         if (queue > TCP_YEAH_ALPHA ||
 172                             rtt - yeah->baseRTT > (yeah->baseRTT / TCP_YEAH_PHY)) {
 173                                 if (queue > TCP_YEAH_ALPHA
 174                                     && tp->snd_cwnd > yeah->reno_count) {
 175                                         u32 reduction = min(queue / TCP_YEAH_GAMMA ,
 176                                                             tp->snd_cwnd >> TCP_YEAH_EPSILON);
 177
 178                                         tp->snd_cwnd -= reduction;
 179
 180                                         tp->snd_cwnd = max(tp->snd_cwnd,
 181                                                            yeah->reno_count);
 182
 183                                         tp->snd_ssthresh = tp->snd_cwnd;
 184                                 }
 185
 186                                 if (yeah->reno_count <= 2)
 187                                         yeah->reno_count = max(tp->snd_cwnd>>1, 2U);
 188                                 else
 189                                         yeah->reno_count++;
 190
 191                                 yeah->doing_reno_now = min(yeah->doing_reno_now + 1,
 192                                                            0xffffffU);
 193                         } else {
 194                                 yeah->fast_count++;
 195
 196                                 if (yeah->fast_count > TCP_YEAH_ZETA) {
 197                                         yeah->reno_count = 2;
 198                                         yeah->fast_count = 0;
 199                                 }
 200
 201                                 yeah->doing_reno_now = 0;
 202                         }
 203
 204                         yeah->lastQ = queue;
 205
 206                 }
 207
 208                 /* Save the extent of the current window so we can use this
 209                  * at the end of the next RTT.
 210                  */
 211                 yeah->beg_snd_una  = yeah->beg_snd_nxt;
 212                 yeah->beg_snd_nxt  = tp->snd_nxt;
 213                 yeah->beg_snd_cwnd = tp->snd_cwnd;
 214
 215                 /* Wipe the slate clean for the next RTT. */
 216                 yeah->cntRTT = 0;
 217                 yeah->minRTT = 0x7fffffff;
 218         }
 219 }
 220
 221 static u32 tcp_yeah_ssthresh(struct sock *sk) {
 222         const struct tcp_sock *tp = tcp_sk(sk);
 223         struct yeah *yeah = inet_csk_ca(sk);
 224         u32 reduction;
 225
 226         if (yeah->doing_reno_now < TCP_YEAH_RHO) {
 227                 reduction = yeah->lastQ;
 228
 229                 reduction = min( reduction, max(tp->snd_cwnd>>1, 2U) );
 230
 231                 reduction = max( reduction, tp->snd_cwnd >> TCP_YEAH_DELTA);
 232         } else
 233                 reduction = max(tp->snd_cwnd>>1,2U);
 234
 235         yeah->fast_count = 0;
 236         yeah->reno_count = max(yeah->reno_count>>1, 2U);
 237
 238         return tp->snd_cwnd - reduction;
 239 }
 240
 241 static struct tcp_congestion_ops tcp_yeah = {
 242         .flags          = TCP_CONG_RTT_STAMP,
 243         .init           = tcp_yeah_init,
 244         .ssthresh       = tcp_yeah_ssthresh,
 245         .cong_avoid     = tcp_yeah_cong_avoid,
 246         .min_cwnd       = tcp_reno_min_cwnd,
 247         .set_state      = tcp_vegas_state,
 248         .cwnd_event     = tcp_vegas_cwnd_event,
 249         .get_info       = tcp_vegas_get_info,
 250         .pkts_acked     = tcp_yeah_pkts_acked,
 251
 252         .owner          = THIS_MODULE,
 253         .name           = "yeah",
 254 };
 255
 256 static int __init tcp_yeah_register(void)
 257 {
 258         BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE);
 259         tcp_register_congestion_control(&tcp_yeah);
 260         return 0;
 261 }
 262
 263 static void __exit tcp_yeah_unregister(void)
 264 {
 265         tcp_unregister_congestion_control(&tcp_yeah);
 266 }
 267
 268 module_init(tcp_yeah_register);
 269 module_exit(tcp_yeah_unregister);
 270
 271 MODULE_AUTHOR("Angelo P. Castellani");
 272 MODULE_LICENSE("GPL");
 273 MODULE_DESCRIPTION("YeAH TCP");