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 #include <linux/mm.h>
  10 #include <linux/module.h>
  11 #include <linux/skbuff.h>
  12 #include <linux/inet_diag.h>
  13
  14 #include <net/tcp.h>
  15
  16 #include "tcp_vegas.h"
  17
  18 #define TCP_YEAH_ALPHA       80 //lin number of packets queued at the bottleneck
  19 #define TCP_YEAH_GAMMA        1 //lin fraction of queue to be removed per rtt
  20 #define TCP_YEAH_DELTA        3 //log minimum fraction of cwnd to be removed on loss
  21 #define TCP_YEAH_EPSILON      1 //log maximum fraction to be removed on early decongestion
  22 #define TCP_YEAH_PHY          8 //lin maximum delta from base
  23 #define TCP_YEAH_RHO         16 //lin minimum number of consecutive rtt to consider competition on loss
  24 #define TCP_YEAH_ZETA        50 //lin minimum number of state switchs to reset reno_count
  25
  26 #define TCP_SCALABLE_AI_CNT      100U
  27
  28 /* YeAH variables */
  29 struct yeah {
  30         struct vegas vegas;     /* must be first */
  31
  32         /* YeAH */
  33         u32 lastQ;
  34         u32 doing_reno_now;
  35
  36         u32 reno_count;
  37         u32 fast_count;
  38
  39         u32 pkts_acked;
  40 };
  41
  42 static void tcp_yeah_init(struct sock *sk)
  43 {
  44         struct tcp_sock *tp = tcp_sk(sk);
  45         struct yeah *yeah = inet_csk_ca(sk);
  46
  47         tcp_vegas_init(sk);
  48
  49         yeah->doing_reno_now = 0;
  50         yeah->lastQ = 0;
  51
  52         yeah->reno_count = 2;
  53
  54         /* Ensure the MD arithmetic works.  This is somewhat pedantic,
  55          * since I don't think we will see a cwnd this large. :) */
  56         tp->snd_cwnd_clamp = min_t(u32, tp->snd_cwnd_clamp, 0xffffffff/128);
  57
  58 }
  59
  60
  61 static void tcp_yeah_pkts_acked(struct sock *sk, u32 pkts_acked, s32 rtt_us)
  62 {
  63         const struct inet_connection_sock *icsk = inet_csk(sk);
  64         struct yeah *yeah = inet_csk_ca(sk);
  65
  66         if (icsk->icsk_ca_state == TCP_CA_Open)
  67                 yeah->pkts_acked = pkts_acked;
  68
  69         tcp_vegas_pkts_acked(sk, pkts_acked, rtt_us);
  70 }
  71
  72 static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
  73 {
  74         struct tcp_sock *tp = tcp_sk(sk);
  75         struct yeah *yeah = inet_csk_ca(sk);
  76
  77         if (!tcp_is_cwnd_limited(sk, in_flight))
  78                 return;
  79
  80         if (tp->snd_cwnd <= tp->snd_ssthresh)
  81                 tcp_slow_start(tp);
  82
  83         else if (!yeah->doing_reno_now) {
  84                 /* Scalable */
  85
  86                 tp->snd_cwnd_cnt += yeah->pkts_acked;
  87                 if (tp->snd_cwnd_cnt > min(tp->snd_cwnd, TCP_SCALABLE_AI_CNT)){
  88                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
  89                                 tp->snd_cwnd++;
  90                         tp->snd_cwnd_cnt = 0;
  91                 }
  92
  93                 yeah->pkts_acked = 1;
  94
  95         } else {
  96                 /* Reno */
  97                 tcp_cong_avoid_ai(tp, tp->snd_cwnd);
  98         }
  99
 100         /* The key players are v_vegas.beg_snd_una and v_beg_snd_nxt.
 101          *
 102          * These are so named because they represent the approximate values
 103          * of snd_una and snd_nxt at the beginning of the current RTT. More
 104          * precisely, they represent the amount of data sent during the RTT.
 105          * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
 106          * we will calculate that (v_beg_snd_nxt - v_vegas.beg_snd_una) outstanding
 107          * bytes of data have been ACKed during the course of the RTT, giving
 108          * an "actual" rate of:
 109          *
 110          *     (v_beg_snd_nxt - v_vegas.beg_snd_una) / (rtt duration)
 111          *
 112          * Unfortunately, v_vegas.beg_snd_una is not exactly equal to snd_una,
 113          * because delayed ACKs can cover more than one segment, so they
 114          * don't line up yeahly with the boundaries of RTTs.
 115          *
 116          * Another unfortunate fact of life is that delayed ACKs delay the
 117          * advance of the left edge of our send window, so that the number
 118          * of bytes we send in an RTT is often less than our cwnd will allow.
 119          * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
 120          */
 121
 122         if (after(ack, yeah->vegas.beg_snd_nxt)) {
 123
 124                 /* We do the Vegas calculations only if we got enough RTT
 125                  * samples that we can be reasonably sure that we got
 126                  * at least one RTT sample that wasn't from a delayed ACK.
 127                  * If we only had 2 samples total,
 128                  * then that means we're getting only 1 ACK per RTT, which
 129                  * means they're almost certainly delayed ACKs.
 130                  * If  we have 3 samples, we should be OK.
 131                  */
 132
 133                 if (yeah->vegas.cntRTT > 2) {
 134                         u32 rtt, queue;
 135                         u64 bw;
 136
 137                         /* We have enough RTT samples, so, using the Vegas
 138                          * algorithm, we determine if we should increase or
 139                          * decrease cwnd, and by how much.
 140                          */
 141
 142                         /* Pluck out the RTT we are using for the Vegas
 143                          * calculations. This is the min RTT seen during the
 144                          * last RTT. Taking the min filters out the effects
 145                          * of delayed ACKs, at the cost of noticing congestion
 146                          * a bit later.
 147                          */
 148                         rtt = yeah->vegas.minRTT;
 149
 150                         /* Compute excess number of packets above bandwidth
 151                          * Avoid doing full 64 bit divide.
 152                          */
 153                         bw = tp->snd_cwnd;
 154                         bw *= rtt - yeah->vegas.baseRTT;
 155                         do_div(bw, rtt);
 156                         queue = bw;
 157
 158                         if (queue > TCP_YEAH_ALPHA ||
 159                             rtt - yeah->vegas.baseRTT > (yeah->vegas.baseRTT / TCP_YEAH_PHY)) {
 160                                 if (queue > TCP_YEAH_ALPHA &&
 161                                     tp->snd_cwnd > yeah->reno_count) {
 162                                         u32 reduction = min(queue / TCP_YEAH_GAMMA ,
 163                                                             tp->snd_cwnd >> TCP_YEAH_EPSILON);
 164
 165                                         tp->snd_cwnd -= reduction;
 166
 167                                         tp->snd_cwnd = max(tp->snd_cwnd,
 168                                                            yeah->reno_count);
 169
 170                                         tp->snd_ssthresh = tp->snd_cwnd;
 171                                 }
 172
 173                                 if (yeah->reno_count <= 2)
 174                                         yeah->reno_count = max(tp->snd_cwnd>>1, 2U);
 175                                 else
 176                                         yeah->reno_count++;
 177
 178                                 yeah->doing_reno_now = min(yeah->doing_reno_now + 1,
 179                                                            0xffffffU);
 180                         } else {
 181                                 yeah->fast_count++;
 182
 183                                 if (yeah->fast_count > TCP_YEAH_ZETA) {
 184                                         yeah->reno_count = 2;
 185                                         yeah->fast_count = 0;
 186                                 }
 187
 188                                 yeah->doing_reno_now = 0;
 189                         }
 190
 191                         yeah->lastQ = queue;
 192
 193                 }
 194
 195                 /* Save the extent of the current window so we can use this
 196                  * at the end of the next RTT.
 197                  */
 198                 yeah->vegas.beg_snd_una  = yeah->vegas.beg_snd_nxt;
 199                 yeah->vegas.beg_snd_nxt  = tp->snd_nxt;
 200                 yeah->vegas.beg_snd_cwnd = tp->snd_cwnd;
 201
 202                 /* Wipe the slate clean for the next RTT. */
 203                 yeah->vegas.cntRTT = 0;
 204                 yeah->vegas.minRTT = 0x7fffffff;
 205         }
 206 }
 207
 208 static u32 tcp_yeah_ssthresh(struct sock *sk) {
 209         const struct tcp_sock *tp = tcp_sk(sk);
 210         struct yeah *yeah = inet_csk_ca(sk);
 211         u32 reduction;
 212
 213         if (yeah->doing_reno_now < TCP_YEAH_RHO) {
 214                 reduction = yeah->lastQ;
 215
 216                 reduction = min( reduction, max(tp->snd_cwnd>>1, 2U) );
 217
 218                 reduction = max( reduction, tp->snd_cwnd >> TCP_YEAH_DELTA);
 219         } else
 220                 reduction = max(tp->snd_cwnd>>1, 2U);
 221
 222         yeah->fast_count = 0;
 223         yeah->reno_count = max(yeah->reno_count>>1, 2U);
 224
 225         return tp->snd_cwnd - reduction;
 226 }
 227
 228 static struct tcp_congestion_ops tcp_yeah __read_mostly = {
 229         .flags          = TCP_CONG_RTT_STAMP,
 230         .init           = tcp_yeah_init,
 231         .ssthresh       = tcp_yeah_ssthresh,
 232         .cong_avoid     = tcp_yeah_cong_avoid,
 233         .min_cwnd       = tcp_reno_min_cwnd,
 234         .set_state      = tcp_vegas_state,
 235         .cwnd_event     = tcp_vegas_cwnd_event,
 236         .get_info       = tcp_vegas_get_info,
 237         .pkts_acked     = tcp_yeah_pkts_acked,
 238
 239         .owner          = THIS_MODULE,
 240         .name           = "yeah",
 241 };
 242
 243 static int __init tcp_yeah_register(void)
 244 {
 245         BUG_ON(sizeof(struct yeah) > ICSK_CA_PRIV_SIZE);
 246         tcp_register_congestion_control(&tcp_yeah);
 247         return 0;
 248 }
 249
 250 static void __exit tcp_yeah_unregister(void)
 251 {
 252         tcp_unregister_congestion_control(&tcp_yeah);
 253 }
 254
 255 module_init(tcp_yeah_register);
 256 module_exit(tcp_yeah_unregister);
 257
 258 MODULE_AUTHOR("Angelo P. Castellani");
 259 MODULE_LICENSE("GPL");
 260 MODULE_DESCRIPTION("YeAH TCP");