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/config.h>
  25 #include <linux/mm.h>
  26 #include <linux/module.h>
  27 #include <linux/skbuff.h>
  28 #include <linux/inet_diag.h>
  29 #include <net/tcp.h>
  30
  31 /* TCP Westwood structure */
  32 struct westwood {
  33         u32    bw_ns_est;        /* first bandwidth estimation..not too smoothed 8) */
  34         u32    bw_est;           /* bandwidth estimate */
  35         u32    rtt_win_sx;       /* here starts a new evaluation... */
  36         u32    bk;
  37         u32    snd_una;          /* used for evaluating the number of acked bytes */
  38         u32    cumul_ack;
  39         u32    accounted;
  40         u32    rtt;
  41         u32    rtt_min;          /* minimum observed RTT */
  42         u8     first_ack;        /* flag which infers that this is the first ack */
  43         u8     reset_rtt_min;    /* Reset RTT min to next RTT sample*/
  44 };
  45
  46
  47 /* TCP Westwood functions and constants */
  48 #define TCP_WESTWOOD_RTT_MIN   (HZ/20)  /* 50ms */
  49 #define TCP_WESTWOOD_INIT_RTT  (20*HZ)  /* maybe too conservative?! */
  50
  51 /*
  52  * @tcp_westwood_create
  53  * This function initializes fields used in TCP Westwood+,
  54  * it is called after the initial SYN, so the sequence numbers
  55  * are correct but new passive connections we have no
  56  * information about RTTmin at this time so we simply set it to
  57  * TCP_WESTWOOD_INIT_RTT. This value was chosen to be too conservative
  58  * since in this way we're sure it will be updated in a consistent
  59  * way as soon as possible. It will reasonably happen within the first
  60  * RTT period of the connection lifetime.
  61  */
  62 static void tcp_westwood_init(struct sock *sk)
  63 {
  64         struct westwood *w = inet_csk_ca(sk);
  65
  66         w->bk = 0;
  67         w->bw_ns_est = 0;
  68         w->bw_est = 0;
  69         w->accounted = 0;
  70         w->cumul_ack = 0;
  71         w->reset_rtt_min = 1;
  72         w->rtt_min = w->rtt = TCP_WESTWOOD_INIT_RTT;
  73         w->rtt_win_sx = tcp_time_stamp;
  74         w->snd_una = tcp_sk(sk)->snd_una;
  75         w->first_ack = 1;
  76 }
  77
  78 /*
  79  * @westwood_do_filter
  80  * Low-pass filter. Implemented using constant coefficients.
  81  */
  82 static inline u32 westwood_do_filter(u32 a, u32 b)
  83 {
  84         return (((7 * a) + b) >> 3);
  85 }
  86
  87 static void westwood_filter(struct westwood *w, u32 delta)
  88 {
  89         /* If the filter is empty fill it with the first sample of bandwidth  */
  90         if (w->bw_ns_est == 0 && w->bw_est == 0) {
  91                 w->bw_ns_est = w->bk / delta;
  92                 w->bw_est = w->bw_ns_est;
  93         } else {
  94                 w->bw_ns_est = westwood_do_filter(w->bw_ns_est, w->bk / delta);
  95                 w->bw_est = westwood_do_filter(w->bw_est, w->bw_ns_est);
  96         }
  97 }
  98
  99 /*
 100  * @westwood_pkts_acked
 101  * Called after processing group of packets.
 102  * but all westwood needs is the last sample of srtt.
 103  */
 104 static void tcp_westwood_pkts_acked(struct sock *sk, u32 cnt)
 105 {
 106         struct westwood *w = inet_csk_ca(sk);
 107         if (cnt > 0)
 108                 w->rtt = tcp_sk(sk)->srtt >> 3;
 109 }
 110
 111 /*
 112  * @westwood_update_window
 113  * It updates RTT evaluation window if it is the right moment to do
 114  * it. If so it calls filter for evaluating bandwidth.
 115  */
 116 static void westwood_update_window(struct sock *sk)
 117 {
 118         struct westwood *w = inet_csk_ca(sk);
 119         s32 delta = tcp_time_stamp - w->rtt_win_sx;
 120
 121         /* Initialize w->snd_una with the first acked sequence number in order
 122          * to fix mismatch between tp->snd_una and w->snd_una for the first
 123          * bandwidth sample
 124          */
 125         if (w->first_ack) {
 126                 w->snd_una = tcp_sk(sk)->snd_una;
 127                 w->first_ack = 0;
 128         }
 129
 130         /*
 131          * See if a RTT-window has passed.
 132          * Be careful since if RTT is less than
 133          * 50ms we don't filter but we continue 'building the sample'.
 134          * This minimum limit was chosen since an estimation on small
 135          * time intervals is better to avoid...
 136          * Obviously on a LAN we reasonably will always have
 137          * right_bound = left_bound + WESTWOOD_RTT_MIN
 138          */
 139         if (w->rtt && delta > max_t(u32, w->rtt, TCP_WESTWOOD_RTT_MIN)) {
 140                 westwood_filter(w, delta);
 141
 142                 w->bk = 0;
 143                 w->rtt_win_sx = tcp_time_stamp;
 144         }
 145 }
 146
 147 static inline void update_rtt_min(struct westwood *w)
 148 {
 149         if (w->reset_rtt_min) {
 150                 w->rtt_min = w->rtt;
 151                 w->reset_rtt_min = 0;
 152         } else
 153                 w->rtt_min = min(w->rtt, w->rtt_min);
 154 }
 155
 156
 157 /*
 158  * @westwood_fast_bw
 159  * It is called when we are in fast path. In particular it is called when
 160  * header prediction is successful. In such case in fact update is
 161  * straight forward and doesn't need any particular care.
 162  */
 163 static inline void westwood_fast_bw(struct sock *sk)
 164 {
 165         const struct tcp_sock *tp = tcp_sk(sk);
 166         struct westwood *w = inet_csk_ca(sk);
 167
 168         westwood_update_window(sk);
 169
 170         w->bk += tp->snd_una - w->snd_una;
 171         w->snd_una = tp->snd_una;
 172         update_rtt_min(w);
 173 }
 174
 175 /*
 176  * @westwood_acked_count
 177  * This function evaluates cumul_ack for evaluating bk in case of
 178  * delayed or partial acks.
 179  */
 180 static inline u32 westwood_acked_count(struct sock *sk)
 181 {
 182         const struct tcp_sock *tp = tcp_sk(sk);
 183         struct westwood *w = inet_csk_ca(sk);
 184
 185         w->cumul_ack = tp->snd_una - w->snd_una;
 186
 187         /* If cumul_ack is 0 this is a dupack since it's not moving
 188          * tp->snd_una.
 189          */
 190         if (!w->cumul_ack) {
 191                 w->accounted += tp->mss_cache;
 192                 w->cumul_ack = tp->mss_cache;
 193         }
 194
 195         if (w->cumul_ack > tp->mss_cache) {
 196                 /* Partial or delayed ack */
 197                 if (w->accounted >= w->cumul_ack) {
 198                         w->accounted -= w->cumul_ack;
 199                         w->cumul_ack = tp->mss_cache;
 200                 } else {
 201                         w->cumul_ack -= w->accounted;
 202                         w->accounted = 0;
 203                 }
 204         }
 205
 206         w->snd_una = tp->snd_una;
 207
 208         return w->cumul_ack;
 209 }
 210
 211
 212 /*
 213  * TCP Westwood
 214  * Here limit is evaluated as Bw estimation*RTTmin (for obtaining it
 215  * in packets we use mss_cache). Rttmin is guaranteed to be >= 2
 216  * so avoids ever returning 0.
 217  */
 218 static u32 tcp_westwood_bw_rttmin(const struct sock *sk)
 219 {
 220         const struct tcp_sock *tp = tcp_sk(sk);
 221         const struct westwood *w = inet_csk_ca(sk);
 222         return max_t(u32, (w->bw_est * w->rtt_min) / tp->mss_cache, 2);
 223 }
 224
 225 static void tcp_westwood_event(struct sock *sk, enum tcp_ca_event event)
 226 {
 227         struct tcp_sock *tp = tcp_sk(sk);
 228         struct westwood *w = inet_csk_ca(sk);
 229
 230         switch(event) {
 231         case CA_EVENT_FAST_ACK:
 232                 westwood_fast_bw(sk);
 233                 break;
 234
 235         case CA_EVENT_COMPLETE_CWR:
 236                 tp->snd_cwnd = tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
 237                 break;
 238
 239         case CA_EVENT_FRTO:
 240                 tp->snd_ssthresh = tcp_westwood_bw_rttmin(sk);
 241                 /* Update RTT_min when next ack arrives */
 242                 w->reset_rtt_min = 1;
 243                 break;
 244
 245         case CA_EVENT_SLOW_ACK:
 246                 westwood_update_window(sk);
 247                 w->bk += westwood_acked_count(sk);
 248                 update_rtt_min(w);
 249                 break;
 250
 251         default:
 252                 /* don't care */
 253                 break;
 254         }
 255 }
 256
 257
 258 /* Extract info for Tcp socket info provided via netlink. */
 259 static void tcp_westwood_info(struct sock *sk, u32 ext,
 260                               struct sk_buff *skb)
 261 {
 262         const struct westwood *ca = inet_csk_ca(sk);
 263         if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
 264                 struct rtattr *rta;
 265                 struct tcpvegas_info *info;
 266
 267                 rta = __RTA_PUT(skb, INET_DIAG_VEGASINFO, sizeof(*info));
 268                 info = RTA_DATA(rta);
 269                 info->tcpv_enabled = 1;
 270                 info->tcpv_rttcnt = 0;
 271                 info->tcpv_rtt = jiffies_to_usecs(ca->rtt);
 272                 info->tcpv_minrtt = jiffies_to_usecs(ca->rtt_min);
 273         rtattr_failure: ;
 274         }
 275 }
 276
 277
 278 static struct tcp_congestion_ops tcp_westwood = {
 279         .init           = tcp_westwood_init,
 280         .ssthresh       = tcp_reno_ssthresh,
 281         .cong_avoid     = tcp_reno_cong_avoid,
 282         .min_cwnd       = tcp_westwood_bw_rttmin,
 283         .cwnd_event     = tcp_westwood_event,
 284         .get_info       = tcp_westwood_info,
 285         .pkts_acked     = tcp_westwood_pkts_acked,
 286
 287         .owner          = THIS_MODULE,
 288         .name           = "westwood"
 289 };
 290
 291 static int __init tcp_westwood_register(void)
 292 {
 293         BUG_ON(sizeof(struct westwood) > ICSK_CA_PRIV_SIZE);
 294         return tcp_register_congestion_control(&tcp_westwood);
 295 }
 296
 297 static void __exit tcp_westwood_unregister(void)
 298 {
 299         tcp_unregister_congestion_control(&tcp_westwood);
 300 }
 301
 302 module_init(tcp_westwood_register);
 303 module_exit(tcp_westwood_unregister);
 304
 305 MODULE_AUTHOR("Stephen Hemminger, Angelo Dell'Aera");
 306 MODULE_LICENSE("GPL");
 307 MODULE_DESCRIPTION("TCP Westwood+");