net/ipv4/tcp_veno.c

   1 /*
   2  * TCP Veno congestion control
   3  *
   4  * This is based on the congestion detection/avoidance scheme described in
   5  *    C. P. Fu, S. C. Liew.
   6  *    "TCP Veno: TCP Enhancement for Transmission over Wireless Access Networks."
   7  *    IEEE Journal on Selected Areas in Communication,
   8  *    Feb. 2003.
   9  *      See http://www.ie.cuhk.edu.hk/fileadmin/staff_upload/soung/Journal/J3.pdf
  10  */
  11
  12 #include <linux/mm.h>
  13 #include <linux/module.h>
  14 #include <linux/skbuff.h>
  15 #include <linux/inet_diag.h>
  16
  17 #include <net/tcp.h>
  18
  19 /* Default values of the Veno variables, in fixed-point representation
  20  * with V_PARAM_SHIFT bits to the right of the binary point.
  21  */
  22 #define V_PARAM_SHIFT 1
  23 static const int beta = 3 << V_PARAM_SHIFT;
  24
  25 /* Veno variables */
  26 struct veno {
  27         u8 doing_veno_now;      /* if true, do veno for this rtt */
  28         u16 cntrtt;             /* # of rtts measured within last rtt */
  29         u32 minrtt;             /* min of rtts measured within last rtt (in usec) */
  30         u32 basertt;            /* the min of all Veno rtt measurements seen (in usec) */
  31         u32 inc;                /* decide whether to increase cwnd */
  32         u32 diff;               /* calculate the diff rate */
  33         u32 loss_cwnd;          /* cwnd when loss occured */
  34 };
  35
  36 /* There are several situations when we must "re-start" Veno:
  37  *
  38  *  o when a connection is established
  39  *  o after an RTO
  40  *  o after fast recovery
  41  *  o when we send a packet and there is no outstanding
  42  *    unacknowledged data (restarting an idle connection)
  43  *
  44  */
  45 static inline void veno_enable(struct sock *sk)
  46 {
  47         struct veno *veno = inet_csk_ca(sk);
  48
  49         /* turn on Veno */
  50         veno->doing_veno_now = 1;
  51
  52         veno->minrtt = 0x7fffffff;
  53 }
  54
  55 static inline void veno_disable(struct sock *sk)
  56 {
  57         struct veno *veno = inet_csk_ca(sk);
  58
  59         /* turn off Veno */
  60         veno->doing_veno_now = 0;
  61 }
  62
  63 static void tcp_veno_init(struct sock *sk)
  64 {
  65         struct veno *veno = inet_csk_ca(sk);
  66
  67         veno->basertt = 0x7fffffff;
  68         veno->inc = 1;
  69         veno_enable(sk);
  70 }
  71
  72 /* Do rtt sampling needed for Veno. */
  73 static void tcp_veno_pkts_acked(struct sock *sk,
  74                                 const struct ack_sample *sample)
  75 {
  76         struct veno *veno = inet_csk_ca(sk);
  77         u32 vrtt;
  78
  79         if (sample->rtt_us < 0)
  80                 return;
  81
  82         /* Never allow zero rtt or baseRTT */
  83         vrtt = sample->rtt_us + 1;
  84
  85         /* Filter to find propagation delay: */
  86         if (vrtt < veno->basertt)
  87                 veno->basertt = vrtt;
  88
  89         /* Find the min rtt during the last rtt to find
  90          * the current prop. delay + queuing delay:
  91          */
  92         veno->minrtt = min(veno->minrtt, vrtt);
  93         veno->cntrtt++;
  94 }
  95
  96 static void tcp_veno_state(struct sock *sk, u8 ca_state)
  97 {
  98         if (ca_state == TCP_CA_Open)
  99                 veno_enable(sk);
 100         else
 101                 veno_disable(sk);
 102 }
 103
 104 /*
 105  * If the connection is idle and we are restarting,
 106  * then we don't want to do any Veno calculations
 107  * until we get fresh rtt samples.  So when we
 108  * restart, we reset our Veno state to a clean
 109  * state. After we get acks for this flight of
 110  * packets, _then_ we can make Veno calculations
 111  * again.
 112  */
 113 static void tcp_veno_cwnd_event(struct sock *sk, enum tcp_ca_event event)
 114 {
 115         if (event == CA_EVENT_CWND_RESTART || event == CA_EVENT_TX_START)
 116                 tcp_veno_init(sk);
 117 }
 118
 119 static void tcp_veno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
 120 {
 121         struct tcp_sock *tp = tcp_sk(sk);
 122         struct veno *veno = inet_csk_ca(sk);
 123
 124         if (!veno->doing_veno_now) {
 125                 tcp_reno_cong_avoid(sk, ack, acked);
 126                 return;
 127         }
 128
 129         /* limited by applications */
 130         if (!tcp_is_cwnd_limited(sk))
 131                 return;
 132
 133         /* We do the Veno calculations only if we got enough rtt samples */
 134         if (veno->cntrtt <= 2) {
 135                 /* We don't have enough rtt samples to do the Veno
 136                  * calculation, so we'll behave like Reno.
 137                  */
 138                 tcp_reno_cong_avoid(sk, ack, acked);
 139         } else {
 140                 u64 target_cwnd;
 141                 u32 rtt;
 142
 143                 /* We have enough rtt samples, so, using the Veno
 144                  * algorithm, we determine the state of the network.
 145                  */
 146
 147                 rtt = veno->minrtt;
 148
 149                 target_cwnd = (u64)tp->snd_cwnd * veno->basertt;
 150                 target_cwnd <<= V_PARAM_SHIFT;
 151                 do_div(target_cwnd, rtt);
 152
 153                 veno->diff = (tp->snd_cwnd << V_PARAM_SHIFT) - target_cwnd;
 154
 155                 if (tcp_in_slow_start(tp)) {
 156                         /* Slow start.  */
 157                         tcp_slow_start(tp, acked);
 158                 } else {
 159                         /* Congestion avoidance. */
 160                         if (veno->diff < beta) {
 161                                 /* In the "non-congestive state", increase cwnd
 162                                  *  every rtt.
 163                                  */
 164                                 tcp_cong_avoid_ai(tp, tp->snd_cwnd, 1);
 165                         } else {
 166                                 /* In the "congestive state", increase cwnd
 167                                  * every other rtt.
 168                                  */
 169                                 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
 170                                         if (veno->inc &&
 171                                             tp->snd_cwnd < tp->snd_cwnd_clamp) {
 172                                                 tp->snd_cwnd++;
 173                                                 veno->inc = 0;
 174                                         } else
 175                                                 veno->inc = 1;
 176                                         tp->snd_cwnd_cnt = 0;
 177                                 } else
 178                                         tp->snd_cwnd_cnt++;
 179                         }
 180                 }
 181                 if (tp->snd_cwnd < 2)
 182                         tp->snd_cwnd = 2;
 183                 else if (tp->snd_cwnd > tp->snd_cwnd_clamp)
 184                         tp->snd_cwnd = tp->snd_cwnd_clamp;
 185         }
 186         /* Wipe the slate clean for the next rtt. */
 187         /* veno->cntrtt = 0; */
 188         veno->minrtt = 0x7fffffff;
 189 }
 190
 191 /* Veno MD phase */
 192 static u32 tcp_veno_ssthresh(struct sock *sk)
 193 {
 194         const struct tcp_sock *tp = tcp_sk(sk);
 195         struct veno *veno = inet_csk_ca(sk);
 196
 197         veno->loss_cwnd = tp->snd_cwnd;
 198         if (veno->diff < beta)
 199                 /* in "non-congestive state", cut cwnd by 1/5 */
 200                 return max(tp->snd_cwnd * 4 / 5, 2U);
 201         else
 202                 /* in "congestive state", cut cwnd by 1/2 */
 203                 return max(tp->snd_cwnd >> 1U, 2U);
 204 }
 205
 206 static u32 tcp_veno_cwnd_undo(struct sock *sk)
 207 {
 208         const struct veno *veno = inet_csk_ca(sk);
 209
 210         return max(tcp_sk(sk)->snd_cwnd, veno->loss_cwnd);
 211 }
 212
 213 static struct tcp_congestion_ops tcp_veno __read_mostly = {
 214         .init           = tcp_veno_init,
 215         .ssthresh       = tcp_veno_ssthresh,
 216         .undo_cwnd      = tcp_veno_cwnd_undo,
 217         .cong_avoid     = tcp_veno_cong_avoid,
 218         .pkts_acked     = tcp_veno_pkts_acked,
 219         .set_state      = tcp_veno_state,
 220         .cwnd_event     = tcp_veno_cwnd_event,
 221
 222         .owner          = THIS_MODULE,
 223         .name           = "veno",
 224 };
 225
 226 static int __init tcp_veno_register(void)
 227 {
 228         BUILD_BUG_ON(sizeof(struct veno) > ICSK_CA_PRIV_SIZE);
 229         tcp_register_congestion_control(&tcp_veno);
 230         return 0;
 231 }
 232
 233 static void __exit tcp_veno_unregister(void)
 234 {
 235         tcp_unregister_congestion_control(&tcp_veno);
 236 }
 237
 238 module_init(tcp_veno_register);
 239 module_exit(tcp_veno_unregister);
 240
 241 MODULE_AUTHOR("Bin Zhou, Cheng Peng Fu");
 242 MODULE_LICENSE("GPL");
 243 MODULE_DESCRIPTION("TCP Veno");