net/ipv4/tcp_westwood.c

   1 /*
   2  * TCP Westwood+: end-to-end bandwidth estimation for TCP
   3  *
   4  *      Angelo Dell'Aera: author of the first version of TCP Westwood+ in Linux 2.4
   5  *
   6  * Support at http://c3lab.poliba.it/index.php/Westwood
   7  * Main references in literature:
   8  *
   9  * - Mascolo S, Casetti, M. Gerla et al.
  10  *   "TCP Westwood: bandwidth estimation for TCP" Proc. ACM Mobicom 2001
  11  *
  12  * - A. Grieco, s. Mascolo
  13  *   "Performance evaluation of New Reno, Vegas, Westwood+ TCP" ACM Computer
  14  *     Comm. Review, 2004
  15  *
  16  * - A. Dell'Aera, L. Grieco, S. Mascolo.
  17  *   "Linux 2.4 Implementation of Westwood+ TCP with Rate-Halving :
  18  *    A Performance Evaluation Over the Internet" (ICC 2004), Paris, June 2004
  19  *
  20  * Westwood+ employs end-to-end bandwidth measurement to set cwnd and
  21  * ssthresh after packet loss. The probing phase is as the original Reno.
  22  */
  23
  24 #include <linux/mm.h>
  25 #include <linux/module.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/inet_diag.h>
  28 #include <net/tcp.h>
  29
  30 /* TCP Westwood structure */
  31 struct westwood {
  32         u32    bw_ns_est;        /* first bandwidth estimation..not too smoothed 8) */
  33         u32    bw_est;           /* bandwidth estimate */
  34         u32    rtt_win_sx;       /* here starts a new evaluation... */
  35         u32    bk;
  36         u32    snd_una;          /* used for evaluating the number of acked bytes */
  37         u32    cumul_ack;
  38         u32    accounted;
  39         u32    rtt;
  40         u32    rtt_min;          /* minimum observed RTT */
  41         u8     first_ack;        /* flag which infers that this is the first ack */
  42         u8     reset_rtt_min;    /* Reset RTT min to next RTT sample*/
  43 };
  44
  45
  46 /* TCP Westwood functions and constants */
  47 #define TCP_WESTWOOD_RTT_MIN   (HZ/20)  /* 50ms */
  48 #define TCP_WESTWOOD_INIT_RTT  (20*HZ)  /* maybe too conservative?! */
  49
  50 /*
  51  * @tcp_westwood_create
  52  * This function initializes fields used in TCP Westwood+,
  53  * it is called after the initial SYN, so the sequence numbers
  54  * are correct but new passive connections we have no
  55  * information about RTTmin at this time so we simply set it to
  56  * TCP_WESTWOOD_INIT_RTT. This value was chosen to be too conservative
  57  * since in this way we're sure it will be updated in a consistent
  58  * way as soon as possible. It will reasonably happen within the first
  59  * RTT period of the connection lifetime.
  60  */
  61 static void tcp_westwood_init(struct sock *sk)
  62 {
  63         struct westwood *w = inet_csk_ca(sk);
  64
  65         w->bk = 0;
  66         w->bw_ns_est = 0;
  67         w->bw_est = 0;
  68         w->accounted = 0;
  69         w->cumul_ack = 0;
  70         w->reset_rtt_min = 1;
  71         w->rtt_min = w->rtt = TCP_WESTWOOD_INIT_RTT;
  72         w->rtt_win_sx = tcp_time_stamp;
  73         w->snd_una = tcp_sk(sk)->snd_una;
  74         w->first_ack = 1;
  75 }
  76
  77 /*
  78  * @westwood_do_filter
  79  * Low-pass filter. Implemented using constant coefficients.
  80  */
  81 static inline u32 westwood_do_filter(u32 a, u32 b)
  82 {
  83         return (((7 * a) + b) >> 3);
  84 }
  85
  86 static void westwood_filter(struct westwood *w, u32 delta)
  87 {
  88         /* If the filter is empty fill it with the first sample of bandwidth  */
  89         if (w->bw_ns_est == 0 && w->bw_est == 0) {
  90                 w->bw_ns_est = w->bk / delta;
  91                 w->bw_est = w->bw_ns_est;
  92         } else {
  93                 w->bw_ns_est = westwood_do_filter(w->bw_ns_est, w->bk / delta);
  94                 w->bw_est = westwood_do_filter(w->bw_est, w->bw_ns_est);
  95         }
  96 }
  97
  98 /*
  99  * @westwood_pkts_acked
 100  * Called after processing group of packets.
 101  * but all westwood needs is the last sample of srtt.
 102  */
 103 static void tcp_westwood_pkts_acked(struct sock *sk, u32 cnt, ktime_t last)
 104 {
 105         struct westwood *w = inet_csk_ca(sk);
 106         if (cnt > 0)
 107                 w->rtt = tcp_sk(sk)->srtt >> 3;
 108 }
 109
 110 /*
 111  * @westwood_update_window
 112  * It updates RTT evaluation window if it is the right moment to do
 113  * it. If so it calls filter for evaluating bandwidth.
 114  */
 115 static void westwood_update_window(struct sock *sk)
 116 {
 117         struct westwood *w = inet_csk_ca(sk);
 118         s32 delta = tcp_time_stamp - w->rtt_win_sx;
 119
 120         /* Initialize w->snd_una with the first acked sequence number in order
 121          * to fix mismatch between tp->snd_una and w->snd_una for the first
 122          * bandwidth sample
 123          */
 124         if (w->first_ack) {
 125                 w->snd_una = tcp_sk(sk)->snd_una;
 126                 w->first_ack = 0;
 127         }
 128
 129         /*
 130          * See if a RTT-window has passed.
 131          * Be careful since if RTT is less than
 132          * 50ms we don't filter but we continue 'building the sample'.
 133          * This minimum limit was chosen since an estimation on small
 134          * time intervals is better to avoid...
 135          * Obviously on a LAN we reasonably will always have
 136          * right_bound = left_bound + WESTWOOD_RTT_MIN
 137          */
 138         if (w->rtt && delta > max_t(u32, w->rtt, TCP_WESTWOOD_RTT_MIN)) {
 139                 westwood_filter(w, delta);
 140
 141                 w->bk = 0;
 142                 w->rtt_win_sx = tcp_time_stamp;
 143         }
 144 }
 145
 146 static inline void update_rtt_min(struct westwood *w)
 147 {
 148         if (w->reset_rtt_min) {
 149                 w->rtt_min = w->rtt;
 150                 w->reset_rtt_min = 0;
 151         } else
 152                 w->rtt_min = min(w->rtt, w->rtt_min);
 153 }
 154
 155
 156 /*
 157  * @westwood_fast_bw
 158  * It is called when we are in fast path. In particular it is called when
 159  * header prediction is successful. In such case in fact update is
 160  * straight forward and doesn't need any particular care.
 161  */
 162 static inline void westwood_fast_bw(struct sock *sk)
 163 {
 164         const struct tcp_sock *tp = tcp_sk(sk);
 165         struct westwood *w = inet_csk_ca(sk);
 166
 167         westwood_update_window(sk);
 168
 169         w->bk += tp->snd_una - w->snd_una;
 170         w->snd_una = tp->snd_una;
 171         update_rtt_min(w);
 172 }
 173
 174 /*
 175  * @westwood_acked_count
 176  * This function evaluates cumul_ack for evaluating bk in case of
 177  * delayed or partial acks.
 178  */
 179 static inline u32 westwood_acked_count(struct sock *sk)
 180 {
 181         const struct tcp_sock *tp = tcp_sk(sk);
 182         struct westwood *w = inet_csk_ca(sk);
 183
 184         w->cumul_ack = tp->snd_una - w->snd_una;
 185
 186         /* If cumul_ack is 0 this is a dupack since it's not moving
 187          * tp->snd_una.
 188          */
 189         if (!w->cumul_ack) {
 190                 w->accounted += tp->mss_cache;
 191                 w->cumul_ack = tp->mss_cache;
 192         }
 193
 194         if (w->cumul_ack > tp->mss_cache) {
 195                 /* Partial or delayed ack */
 196                 if (w->accounted >= w->cumul_ack) {
 197                         w->accounted -= w->cumul_ack;
 198                         w->cumul_ack = tp->mss_cache;
 199                 } else {
 200                         w->cumul_ack -= w->accounted;
 201                         w->accounted = 0;
 202                 }
 203         }
 204
 205         w->snd_una = tp->snd_una;
 206
 207         return w->cumul_ack;
 208 }
 209
 210
 211 /*
 212  * TCP Westwood
 213  * Here limit is evaluated as Bw estimation*RTTmin (for obtaining it
 214  * in packets we use mss_cache). Rttmin is guaranteed to be >= 2
 215  * so avoids ever returning 0.
 216  */
 217 static u32 tcp_westwood_bw_rttmin(const struct sock *sk)
 218 {
 219         const struct tcp_sock *tp = tcp_sk(sk);
 220         const struct westwood *w = inet_csk_ca(sk);
 221         return max_t(u32, (w->bw_est * w->rtt_min) / tp->mss_cache, 2);
 222 }
 223
 224 static void tcp_westwood_event(struct sock *sk, enum tcp_ca_event event)
 225 {
 226         struct tcp_sock *tp = tcp_sk(sk);
 227         struct westwood *w = inet_csk_ca(sk);
 228
 229         switch (event) {
 230         case CA_EVENT_FAST_ACK:
 231                 westwood_fast_bw(sk);
 232                 break;
 233
 234         case CA_EVENT_COMPLETE_CWR:
 235                 tp->snd_cwnd = tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
 236                 break;
 237
 238         case CA_EVENT_FRTO:
 239                 tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
 240                 /* Update RTT_min when next ack arrives */
 241                 w->reset_rtt_min = 1;
 242                 break;
 243
 244         case CA_EVENT_SLOW_ACK:
 245                 westwood_update_window(sk);
 246                 w->bk += westwood_acked_count(sk);
 247                 update_rtt_min(w);
 248                 break;
 249
 250         default:
 251                 /* don't care */
 252                 break;
 253         }
 254 }
 255
 256
 257 /* Extract info for Tcp socket info provided via netlink. */
 258 static void tcp_westwood_info(struct sock *sk, u32 ext,
 259                               struct sk_buff *skb)
 260 {
 261         const struct westwood *ca = inet_csk_ca(sk);
 262         if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
 263                 struct tcpvegas_info info = {
 264                         .tcpv_enabled = 1,
 265                         .tcpv_rtt = jiffies_to_usecs(ca->rtt),
 266                         .tcpv_minrtt = jiffies_to_usecs(ca->rtt_min),
 267                 };
 268
 269                 nla_put(skb, INET_DIAG_VEGASINFO, sizeof(info), &info);
 270         }
 271 }
 272
 273
 274 static struct tcp_congestion_ops tcp_westwood = {
 275         .init           = tcp_westwood_init,
 276         .ssthresh       = tcp_reno_ssthresh,
 277         .cong_avoid     = tcp_reno_cong_avoid,
 278         .min_cwnd       = tcp_westwood_bw_rttmin,
 279         .cwnd_event     = tcp_westwood_event,
 280         .get_info       = tcp_westwood_info,
 281         .pkts_acked     = tcp_westwood_pkts_acked,
 282
 283         .owner          = THIS_MODULE,
 284         .name           = "westwood"
 285 };
 286
 287 static int __init tcp_westwood_register(void)
 288 {
 289         BUILD_BUG_ON(sizeof(struct westwood) > ICSK_CA_PRIV_SIZE);
 290         return tcp_register_congestion_control(&tcp_westwood);
 291 }
 292
 293 static void __exit tcp_westwood_unregister(void)
 294 {
 295         tcp_unregister_congestion_control(&tcp_westwood);
 296 }
 297
 298 module_init(tcp_westwood_register);
 299 module_exit(tcp_westwood_unregister);
 300
 301 MODULE_AUTHOR("Stephen Hemminger, Angelo Dell'Aera");
 302 MODULE_LICENSE("GPL");
 303 MODULE_DESCRIPTION("TCP Westwood+");