net/ipv4/tcp_westwood.c

   1 /*
   2  * TCP Westwood+
   3  *
   4  *      Angelo Dell'Aera:       TCP Westwood+ support
   5  */
   6
   7 #include <linux/config.h>
   8 #include <linux/mm.h>
   9 #include <linux/module.h>
  10 #include <linux/skbuff.h>
  11 #include <linux/inet_diag.h>
  12 #include <net/tcp.h>
  13
  14 /* TCP Westwood structure */
  15 struct westwood {
  16         u32    bw_ns_est;        /* first bandwidth estimation..not too smoothed 8) */
  17         u32    bw_est;           /* bandwidth estimate */
  18         u32    rtt_win_sx;       /* here starts a new evaluation... */
  19         u32    bk;
  20         u32    snd_una;          /* used for evaluating the number of acked bytes */
  21         u32    cumul_ack;
  22         u32    accounted;
  23         u32    rtt;
  24         u32    rtt_min;          /* minimum observed RTT */
  25         u8     first_ack;        /* flag which infers that this is the first ack */
  26 };
  27
  28
  29 /* TCP Westwood functions and constants */
  30 #define TCP_WESTWOOD_RTT_MIN   (HZ/20)  /* 50ms */
  31 #define TCP_WESTWOOD_INIT_RTT  (20*HZ)  /* maybe too conservative?! */
  32
  33 /*
  34  * @tcp_westwood_create
  35  * This function initializes fields used in TCP Westwood+,
  36  * it is called after the initial SYN, so the sequence numbers
  37  * are correct but new passive connections we have no
  38  * information about RTTmin at this time so we simply set it to
  39  * TCP_WESTWOOD_INIT_RTT. This value was chosen to be too conservative
  40  * since in this way we're sure it will be updated in a consistent
  41  * way as soon as possible. It will reasonably happen within the first
  42  * RTT period of the connection lifetime.
  43  */
  44 static void tcp_westwood_init(struct sock *sk)
  45 {
  46         struct westwood *w = inet_csk_ca(sk);
  47
  48         w->bk = 0;
  49         w->bw_ns_est = 0;
  50         w->bw_est = 0;
  51         w->accounted = 0;
  52         w->cumul_ack = 0;
  53         w->rtt_min = w->rtt = TCP_WESTWOOD_INIT_RTT;
  54         w->rtt_win_sx = tcp_time_stamp;
  55         w->snd_una = tcp_sk(sk)->snd_una;
  56         w->first_ack = 1;
  57 }
  58
  59 /*
  60  * @westwood_do_filter
  61  * Low-pass filter. Implemented using constant coefficients.
  62  */
  63 static inline u32 westwood_do_filter(u32 a, u32 b)
  64 {
  65         return (((7 * a) + b) >> 3);
  66 }
  67
  68 static inline void westwood_filter(struct westwood *w, u32 delta)
  69 {
  70         w->bw_ns_est = westwood_do_filter(w->bw_ns_est, w->bk / delta);
  71         w->bw_est = westwood_do_filter(w->bw_est, w->bw_ns_est);
  72 }
  73
  74 /*
  75  * @westwood_pkts_acked
  76  * Called after processing group of packets.
  77  * but all westwood needs is the last sample of srtt.
  78  */
  79 static void tcp_westwood_pkts_acked(struct sock *sk, u32 cnt)
  80 {
  81         struct westwood *w = inet_csk_ca(sk);
  82         if (cnt > 0)
  83                 w->rtt = tcp_sk(sk)->srtt >> 3;
  84 }
  85
  86 /*
  87  * @westwood_update_window
  88  * It updates RTT evaluation window if it is the right moment to do
  89  * it. If so it calls filter for evaluating bandwidth.
  90  */
  91 static void westwood_update_window(struct sock *sk)
  92 {
  93         struct westwood *w = inet_csk_ca(sk);
  94         s32 delta = tcp_time_stamp - w->rtt_win_sx;
  95
  96         /* Initialise w->snd_una with the first acked sequence number in order
  97          * to fix mismatch between tp->snd_una and w->snd_una for the first
  98          * bandwidth sample
  99          */
 100         if (w->first_ack) {
 101                 w->snd_una = tcp_sk(sk)->snd_una;
 102                 w->first_ack = 0;
 103         }
 104
 105         /*
 106          * See if a RTT-window has passed.
 107          * Be careful since if RTT is less than
 108          * 50ms we don't filter but we continue 'building the sample'.
 109          * This minimum limit was chosen since an estimation on small
 110          * time intervals is better to avoid...
 111          * Obviously on a LAN we reasonably will always have
 112          * right_bound = left_bound + WESTWOOD_RTT_MIN
 113          */
 114         if (w->rtt && delta > max_t(u32, w->rtt, TCP_WESTWOOD_RTT_MIN)) {
 115                 westwood_filter(w, delta);
 116
 117                 w->bk = 0;
 118                 w->rtt_win_sx = tcp_time_stamp;
 119         }
 120 }
 121
 122 /*
 123  * @westwood_fast_bw
 124  * It is called when we are in fast path. In particular it is called when
 125  * header prediction is successful. In such case in fact update is
 126  * straight forward and doesn't need any particular care.
 127  */
 128 static inline void westwood_fast_bw(struct sock *sk)
 129 {
 130         const struct tcp_sock *tp = tcp_sk(sk);
 131         struct westwood *w = inet_csk_ca(sk);
 132
 133         westwood_update_window(sk);
 134
 135         w->bk += tp->snd_una - w->snd_una;
 136         w->snd_una = tp->snd_una;
 137         w->rtt_min = min(w->rtt, w->rtt_min);
 138 }
 139
 140 /*
 141  * @westwood_acked_count
 142  * This function evaluates cumul_ack for evaluating bk in case of
 143  * delayed or partial acks.
 144  */
 145 static inline u32 westwood_acked_count(struct sock *sk)
 146 {
 147         const struct tcp_sock *tp = tcp_sk(sk);
 148         struct westwood *w = inet_csk_ca(sk);
 149
 150         w->cumul_ack = tp->snd_una - w->snd_una;
 151
 152         /* If cumul_ack is 0 this is a dupack since it's not moving
 153          * tp->snd_una.
 154          */
 155         if (!w->cumul_ack) {
 156                 w->accounted += tp->mss_cache;
 157                 w->cumul_ack = tp->mss_cache;
 158         }
 159
 160         if (w->cumul_ack > tp->mss_cache) {
 161                 /* Partial or delayed ack */
 162                 if (w->accounted >= w->cumul_ack) {
 163                         w->accounted -= w->cumul_ack;
 164                         w->cumul_ack = tp->mss_cache;
 165                 } else {
 166                         w->cumul_ack -= w->accounted;
 167                         w->accounted = 0;
 168                 }
 169         }
 170
 171         w->snd_una = tp->snd_una;
 172
 173         return w->cumul_ack;
 174 }
 175
 176
 177 /*
 178  * TCP Westwood
 179  * Here limit is evaluated as Bw estimation*RTTmin (for obtaining it
 180  * in packets we use mss_cache). Rttmin is guaranteed to be >= 2
 181  * so avoids ever returning 0.
 182  */
 183 static u32 tcp_westwood_bw_rttmin(const struct sock *sk)
 184 {
 185         const struct tcp_sock *tp = tcp_sk(sk);
 186         const struct westwood *w = inet_csk_ca(sk);
 187         return max_t(u32, (w->bw_est * w->rtt_min) / tp->mss_cache, 2);
 188 }
 189
 190 static void tcp_westwood_event(struct sock *sk, enum tcp_ca_event event)
 191 {
 192         struct tcp_sock *tp = tcp_sk(sk);
 193         struct westwood *w = inet_csk_ca(sk);
 194
 195         switch(event) {
 196         case CA_EVENT_FAST_ACK:
 197                 westwood_fast_bw(sk);
 198                 break;
 199
 200         case CA_EVENT_COMPLETE_CWR:
 201                 tp->snd_cwnd = tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
 202                 break;
 203
 204         case CA_EVENT_FRTO:
 205                 tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
 206                 break;
 207
 208         case CA_EVENT_SLOW_ACK:
 209                 westwood_update_window(sk);
 210                 w->bk += westwood_acked_count(sk);
 211                 w->rtt_min = min(w->rtt, w->rtt_min);
 212                 break;
 213
 214         default:
 215                 /* don't care */
 216                 break;
 217         }
 218 }
 219
 220
 221 /* Extract info for Tcp socket info provided via netlink. */
 222 static void tcp_westwood_info(struct sock *sk, u32 ext,
 223                               struct sk_buff *skb)
 224 {
 225         const struct westwood *ca = inet_csk_ca(sk);
 226         if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
 227                 struct rtattr *rta;
 228                 struct tcpvegas_info *info;
 229
 230                 rta = __RTA_PUT(skb, INET_DIAG_VEGASINFO, sizeof(*info));
 231                 info = RTA_DATA(rta);
 232                 info->tcpv_enabled = 1;
 233                 info->tcpv_rttcnt = 0;
 234                 info->tcpv_rtt = jiffies_to_usecs(ca->rtt);
 235                 info->tcpv_minrtt = jiffies_to_usecs(ca->rtt_min);
 236         rtattr_failure: ;
 237         }
 238 }
 239
 240
 241 static struct tcp_congestion_ops tcp_westwood = {
 242         .init           = tcp_westwood_init,
 243         .ssthresh       = tcp_reno_ssthresh,
 244         .cong_avoid     = tcp_reno_cong_avoid,
 245         .min_cwnd       = tcp_westwood_bw_rttmin,
 246         .cwnd_event     = tcp_westwood_event,
 247         .get_info       = tcp_westwood_info,
 248         .pkts_acked     = tcp_westwood_pkts_acked,
 249
 250         .owner          = THIS_MODULE,
 251         .name           = "westwood"
 252 };
 253
 254 static int __init tcp_westwood_register(void)
 255 {
 256         BUG_ON(sizeof(struct westwood) > ICSK_CA_PRIV_SIZE);
 257         return tcp_register_congestion_control(&tcp_westwood);
 258 }
 259
 260 static void __exit tcp_westwood_unregister(void)
 261 {
 262         tcp_unregister_congestion_control(&tcp_westwood);
 263 }
 264
 265 module_init(tcp_westwood_register);
 266 module_exit(tcp_westwood_unregister);
 267
 268 MODULE_AUTHOR("Stephen Hemminger, Angelo Dell'Aera");
 269 MODULE_LICENSE("GPL");
 270 MODULE_DESCRIPTION("TCP Westwood+");