2 * Copyright (c) 2008, 2009 Nicira Networks.
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:
8 * http://www.apache.org/licenses/LICENSE-2.0
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.
19 #include <arpa/inet.h>
22 #include "openflow/openflow.h"
23 #include "poll-loop.h"
24 #include "port-array.h"
33 /* Client-supplied parameters. */
34 int rate_limit; /* Packets added to bucket per second. */
35 int burst_limit; /* Maximum token bucket size, in packets. */
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. */
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. */
51 /* Transmission queue. */
52 int n_txq; /* No. of packets waiting in rconn for tx. */
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. */
60 struct status_category *ss_cat;
63 static struct ofpbuf *
64 dequeue_packet(struct pinsched *ps, struct ovs_queue *q,
67 struct ofpbuf *packet = queue_pop_head(q);
70 port_array_set(&ps->queues, port_no, NULL);
76 /* Drop a packet from the longest queue in 'ps'. */
78 drop_packet(struct pinsched *ps)
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;
86 ps->n_queue_dropped++;
88 longest = port_array_first(&ps->queues, &port_no);
89 longest_port_no = port_no;
91 while ((q = port_array_next(&ps->queues, &port_no)) != NULL) {
92 if (longest->n < q->n) {
95 } else if (longest->n == q->n) {
98 /* Randomly select one of the longest queues, with a uniform
99 * distribution (Knuth algorithm 3.4.2R). */
100 if (!random_range(n_longest)) {
102 longest_port_no = port_no;
107 /* FIXME: do we want to pop the tail instead? */
108 ofpbuf_delete(dequeue_packet(ps, longest, longest_port_no));
111 /* Remove and return the next packet to transmit (in round-robin order). */
112 static struct ofpbuf *
113 get_tx_packet(struct pinsched *ps)
115 struct ovs_queue *q = port_array_next(&ps->queues, &ps->last_tx_port);
117 q = port_array_first(&ps->queues, &ps->last_tx_port);
119 return dequeue_packet(ps, q, ps->last_tx_port);
122 /* Add tokens to the bucket based on elapsed time. */
124 refill_bucket(struct pinsched *ps)
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) {
130 ps->tokens = MIN(tokens, ps->burst_limit * 1000);
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
138 get_token(struct pinsched *ps)
140 if (ps->tokens >= 1000) {
149 pinsched_send(struct pinsched *ps, uint16_t port_no,
150 struct ofpbuf *packet, pinsched_tx_cb *cb, void *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. */
160 /* Otherwise queue it up for the periodic callback to drain out. */
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. */
169 if (ps->n_queued >= ps->burst_limit) {
172 q = port_array_get(&ps->queues, port_no);
174 q = xmalloc(sizeof *q);
176 port_array_set(&ps->queues, port_no, q);
178 queue_push_tail(q, packet);
185 pinsched_status_cb(struct status_reply *sr, void *ps_)
187 struct pinsched *ps = ps_;
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);
195 pinsched_run(struct pinsched *ps, pinsched_tx_cb *cb, void *aux)
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. */
203 for (i = 0; ps->n_queued && get_token(ps) && i < 50; i++) {
204 cb(get_tx_packet(ps), aux);
210 pinsched_wait(struct pinsched *ps)
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();
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);
224 /* Creates and returns a scheduler for sending packet-in messages. */
226 pinsched_create(int rate_limit, int burst_limit, struct switch_status *ss)
230 ps = xcalloc(1, sizeof *ps);
231 port_array_init(&ps->queues);
233 ps->last_tx_port = PORT_ARRAY_SIZE;
234 ps->last_fill = time_msec();
235 ps->tokens = rate_limit * 100;
239 ps->n_queue_dropped = 0;
240 pinsched_set_limits(ps, rate_limit, burst_limit);
243 ps->ss_cat = switch_status_register(ss, "rate-limit",
244 pinsched_status_cb, ps);
251 pinsched_destroy(struct pinsched *ps)
254 struct ovs_queue *queue;
255 unsigned int port_no;
257 PORT_ARRAY_FOR_EACH (queue, &ps->queues, port_no) {
258 queue_destroy(queue);
261 port_array_destroy(&ps->queues);
262 switch_status_unregister(ps->ss_cat);
268 pinsched_set_limits(struct pinsched *ps, int rate_limit, int burst_limit)
270 if (rate_limit <= 0) {
273 if (burst_limit <= 0) {
274 burst_limit = rate_limit / 4;
276 burst_limit = MAX(burst_limit, 1);
277 burst_limit = MIN(burst_limit, INT_MAX / 1000);
279 ps->rate_limit = rate_limit;
280 ps->burst_limit = burst_limit;
281 while (ps->n_queued > burst_limit) {