linux/net/ipv4/tcp_yeah.c
<<
>>
Prefs
   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 */
  29struct 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
  42static 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
  61static 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
  72static 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
 208static 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
 228static 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
 243static 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
 250static void __exit tcp_yeah_unregister(void)
 251{
 252        tcp_unregister_congestion_control(&tcp_yeah);
 253}
 254
 255module_init(tcp_yeah_register);
 256module_exit(tcp_yeah_unregister);
 257
 258MODULE_AUTHOR("Angelo P. Castellani");
 259MODULE_LICENSE("GPL");
 260MODULE_DESCRIPTION("YeAH TCP");
 261
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.