secchan/pinsched.c

   1 /*
   2  * Copyright (c) 2008, 2009 Nicira Networks.
   3  *
   4  * Licensed under the Apache License, Version 2.0 (the "License");
   5  * you may not use this file except in compliance with the License.
   6  * You may obtain a copy of the License at:
   7  *
   8  *     http://www.apache.org/licenses/LICENSE-2.0
   9  *
  10  * Unless required by applicable law or agreed to in writing, software
  11  * distributed under the License is distributed on an "AS IS" BASIS,
  12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  * See the License for the specific language governing permissions and
  14  * limitations under the License.
  15  */
  16
  17 #include <config.h>
  18 #include "pinsched.h"
  19 #include <arpa/inet.h>
  20 #include <stdlib.h>
  21 #include "ofpbuf.h"
  22 #include "openflow/openflow.h"
  23 #include "poll-loop.h"
  24 #include "port-array.h"
  25 #include "queue.h"
  26 #include "random.h"
  27 #include "rconn.h"
  28 #include "status.h"
  29 #include "timeval.h"
  30 #include "vconn.h"
  31
  32 struct pinsched {
  33     /* Client-supplied parameters. */
  34     int rate_limit;           /* Packets added to bucket per second. */
  35     int burst_limit;          /* Maximum token bucket size, in packets. */
  36
  37     /* One queue per physical port. */
  38     struct port_array queues;   /* Array of "struct ovs_queue *". */
  39     int n_queued;               /* Sum over queues[*].n. */
  40     unsigned int last_tx_port;  /* Last port checked in round-robin. */
  41
  42     /* Token bucket.
  43      *
  44      * It costs 1000 tokens to send a single packet_in message.  A single token
  45      * per message would be more straightforward, but this choice lets us avoid
  46      * round-off error in refill_bucket()'s calculation of how many tokens to
  47      * add to the bucket, since no division step is needed. */
  48     long long int last_fill;    /* Time at which we last added tokens. */
  49     int tokens;                 /* Current number of tokens. */
  50
  51     /* Transmission queue. */
  52     int n_txq;                  /* No. of packets waiting in rconn for tx. */
  53
  54     /* Statistics reporting. */
  55     unsigned long long n_normal;        /* # txed w/o rate limit queuing. */
  56     unsigned long long n_limited;       /* # queued for rate limiting. */
  57     unsigned long long n_queue_dropped; /* # dropped due to queue overflow. */
  58
  59     /* Switch status. */
  60     struct status_category *ss_cat;
  61 };
  62
  63 static struct ofpbuf *
  64 dequeue_packet(struct pinsched *ps, struct ovs_queue *q,
  65                unsigned int port_no)
  66 {
  67     struct ofpbuf *packet = queue_pop_head(q);
  68     if (!q->n) {
  69         free(q);
  70         port_array_set(&ps->queues, port_no, NULL);
  71     }
  72     ps->n_queued--;
  73     return packet;
  74 }
  75
  76 /* Drop a packet from the longest queue in 'ps'. */
  77 static void
  78 drop_packet(struct pinsched *ps)
  79 {
  80     struct ovs_queue *longest;  /* Queue currently selected as longest. */
  81     int n_longest;              /* # of queues of same length as 'longest'. */
  82     unsigned int longest_port_no;
  83     unsigned int port_no;
  84     struct ovs_queue *q;
  85
  86     ps->n_queue_dropped++;
  87
  88     longest = port_array_first(&ps->queues, &port_no);
  89     longest_port_no = port_no;
  90     n_longest = 1;
  91     while ((q = port_array_next(&ps->queues, &port_no)) != NULL) {
  92         if (longest->n < q->n) {
  93             longest = q;
  94             n_longest = 1;
  95         } else if (longest->n == q->n) {
  96             n_longest++;
  97
  98             /* Randomly select one of the longest queues, with a uniform
  99              * distribution (Knuth algorithm 3.4.2R). */
 100             if (!random_range(n_longest)) {
 101                 longest = q;
 102                 longest_port_no = port_no;
 103             }
 104         }
 105     }
 106
 107     /* FIXME: do we want to pop the tail instead? */
 108     ofpbuf_delete(dequeue_packet(ps, longest, longest_port_no));
 109 }
 110
 111 /* Remove and return the next packet to transmit (in round-robin order). */
 112 static struct ofpbuf *
 113 get_tx_packet(struct pinsched *ps)
 114 {
 115     struct ovs_queue *q = port_array_next(&ps->queues, &ps->last_tx_port);
 116     if (!q) {
 117         q = port_array_first(&ps->queues, &ps->last_tx_port);
 118     }
 119     return dequeue_packet(ps, q, ps->last_tx_port);
 120 }
 121
 122 /* Add tokens to the bucket based on elapsed time. */
 123 static void
 124 refill_bucket(struct pinsched *ps)
 125 {
 126     long long int now = time_msec();
 127     long long int tokens = (now - ps->last_fill) * ps->rate_limit + ps->tokens;
 128     if (tokens >= 1000) {
 129         ps->last_fill = now;
 130         ps->tokens = MIN(tokens, ps->burst_limit * 1000);
 131     }
 132 }
 133
 134 /* Attempts to remove enough tokens from 'ps' to transmit a packet.  Returns
 135  * true if successful, false otherwise.  (In the latter case no tokens are
 136  * removed.) */
 137 static bool
 138 get_token(struct pinsched *ps)
 139 {
 140     if (ps->tokens >= 1000) {
 141         ps->tokens -= 1000;
 142         return true;
 143     } else {
 144         return false;
 145     }
 146 }
 147
 148 void
 149 pinsched_send(struct pinsched *ps, uint16_t port_no,
 150               struct ofpbuf *packet, pinsched_tx_cb *cb, void *aux)
 151 {
 152     if (!ps) {
 153         cb(packet, aux);
 154     } else if (!ps->n_queued && get_token(ps)) {
 155         /* In the common case where we are not constrained by the rate limit,
 156          * let the packet take the normal path. */
 157         ps->n_normal++;
 158         cb(packet, aux);
 159     } else {
 160         /* Otherwise queue it up for the periodic callback to drain out. */
 161         struct ovs_queue *q;
 162
 163         /* We are called with a buffer obtained from dpif_recv() that has much
 164          * more allocated space than actual content most of the time.  Since
 165          * we're going to store the packet for some time, free up that
 166          * otherwise wasted space. */
 167         ofpbuf_trim(packet);
 168
 169         if (ps->n_queued >= ps->burst_limit) {
 170             drop_packet(ps);
 171         }
 172         q = port_array_get(&ps->queues, port_no);
 173         if (!q) {
 174             q = xmalloc(sizeof *q);
 175             queue_init(q);
 176             port_array_set(&ps->queues, port_no, q);
 177         }
 178         queue_push_tail(q, packet);
 179         ps->n_queued++;
 180         ps->n_limited++;
 181     }
 182 }
 183
 184 static void
 185 pinsched_status_cb(struct status_reply *sr, void *ps_)
 186 {
 187     struct pinsched *ps = ps_;
 188
 189     status_reply_put(sr, "normal=%llu", ps->n_normal);
 190     status_reply_put(sr, "limited=%llu", ps->n_limited);
 191     status_reply_put(sr, "queue-dropped=%llu", ps->n_queue_dropped);
 192 }
 193
 194 void
 195 pinsched_run(struct pinsched *ps, pinsched_tx_cb *cb, void *aux)
 196 {
 197     if (ps) {
 198         int i;
 199
 200         /* Drain some packets out of the bucket if possible, but limit the
 201          * number of iterations to allow other code to get work done too. */
 202         refill_bucket(ps);
 203         for (i = 0; ps->n_queued && get_token(ps) && i < 50; i++) {
 204             cb(get_tx_packet(ps), aux);
 205         }
 206     }
 207 }
 208
 209 void
 210 pinsched_wait(struct pinsched *ps)
 211 {
 212     if (ps && ps->n_queued) {
 213         if (ps->tokens >= 1000) {
 214             /* We can transmit more packets as soon as we're called again. */
 215             poll_immediate_wake();
 216         } else {
 217             /* We have to wait for the bucket to re-fill.  We could calculate
 218              * the exact amount of time here for increased smoothness. */
 219             poll_timer_wait(TIME_UPDATE_INTERVAL / 2);
 220         }
 221     }
 222 }
 223
 224 /* Creates and returns a scheduler for sending packet-in messages. */
 225 struct pinsched *
 226 pinsched_create(int rate_limit, int burst_limit, struct switch_status *ss)
 227 {
 228     struct pinsched *ps;
 229
 230     ps = xcalloc(1, sizeof *ps);
 231     port_array_init(&ps->queues);
 232     ps->n_queued = 0;
 233     ps->last_tx_port = PORT_ARRAY_SIZE;
 234     ps->last_fill = time_msec();
 235     ps->tokens = rate_limit * 100;
 236     ps->n_txq = 0;
 237     ps->n_normal = 0;
 238     ps->n_limited = 0;
 239     ps->n_queue_dropped = 0;
 240     pinsched_set_limits(ps, rate_limit, burst_limit);
 241
 242     if (ss) {
 243         ps->ss_cat = switch_status_register(ss, "rate-limit",
 244                                             pinsched_status_cb, ps);
 245     }
 246
 247     return ps;
 248 }
 249
 250 void
 251 pinsched_destroy(struct pinsched *ps)
 252 {
 253     if (ps) {
 254         struct ovs_queue *queue;
 255         unsigned int port_no;
 256
 257         PORT_ARRAY_FOR_EACH (queue, &ps->queues, port_no) {
 258             queue_destroy(queue);
 259             free(queue);
 260         }
 261         port_array_destroy(&ps->queues);
 262         switch_status_unregister(ps->ss_cat);
 263         free(ps);
 264     }
 265 }
 266
 267 void
 268 pinsched_set_limits(struct pinsched *ps, int rate_limit, int burst_limit)
 269 {
 270     if (rate_limit <= 0) {
 271         rate_limit = 1000;
 272     }
 273     if (burst_limit <= 0) {
 274         burst_limit = rate_limit / 4;
 275     }
 276     burst_limit = MAX(burst_limit, 1);
 277     burst_limit = MIN(burst_limit, INT_MAX / 1000);
 278
 279     ps->rate_limit = rate_limit;
 280     ps->burst_limit = burst_limit;
 281     while (ps->n_queued > burst_limit) {
 282         drop_packet(ps);
 283     }
 284 }