2 * Copyright (c) 2008, 2009, 2010 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 <sys/types.h>
20 #include <netinet/in.h>
21 #include <arpa/inet.h>
25 #include "openflow/openflow.h"
26 #include "poll-loop.h"
27 #include "port-array.h"
35 struct list packets; /* Contains "struct ofpbuf"s. */
36 int n; /* Number of packets in 'packets'. */
40 /* Client-supplied parameters. */
41 int rate_limit; /* Packets added to bucket per second. */
42 int burst_limit; /* Maximum token bucket size, in packets. */
44 /* One queue per physical port. */
45 struct port_array queues; /* Array of "struct pinqueue *"s. */
46 int n_queued; /* Sum over queues[*].n. */
47 unsigned int last_tx_port; /* Last port checked in round-robin. */
51 * It costs 1000 tokens to send a single packet_in message. A single token
52 * per message would be more straightforward, but this choice lets us avoid
53 * round-off error in refill_bucket()'s calculation of how many tokens to
54 * add to the bucket, since no division step is needed. */
55 long long int last_fill; /* Time at which we last added tokens. */
56 int tokens; /* Current number of tokens. */
58 /* Transmission queue. */
59 int n_txq; /* No. of packets waiting in rconn for tx. */
61 /* Statistics reporting. */
62 unsigned long long n_normal; /* # txed w/o rate limit queuing. */
63 unsigned long long n_limited; /* # queued for rate limiting. */
64 unsigned long long n_queue_dropped; /* # dropped due to queue overflow. */
67 struct status_category *ss_cat;
70 static struct ofpbuf *
71 dequeue_packet(struct pinsched *ps, struct pinqueue *q, unsigned int port_no)
73 struct ofpbuf *packet = ofpbuf_from_list(list_pop_front(&q->packets));
76 port_array_delete(&ps->queues, port_no);
82 /* Drop a packet from the longest queue in 'ps'. */
84 drop_packet(struct pinsched *ps)
86 struct pinqueue *longest; /* Queue currently selected as longest. */
87 int n_longest; /* # of queues of same length as 'longest'. */
88 unsigned int longest_port_no;
92 ps->n_queue_dropped++;
94 longest = port_array_first(&ps->queues, &port_no);
95 longest_port_no = port_no;
97 while ((q = port_array_next(&ps->queues, &port_no)) != NULL) {
98 if (longest->n < q->n) {
101 } else if (longest->n == q->n) {
104 /* Randomly select one of the longest queues, with a uniform
105 * distribution (Knuth algorithm 3.4.2R). */
106 if (!random_range(n_longest)) {
108 longest_port_no = port_no;
113 /* FIXME: do we want to pop the tail instead? */
114 ofpbuf_delete(dequeue_packet(ps, longest, longest_port_no));
117 /* Remove and return the next packet to transmit (in round-robin order). */
118 static struct ofpbuf *
119 get_tx_packet(struct pinsched *ps)
121 struct pinqueue *q = port_array_next(&ps->queues, &ps->last_tx_port);
123 q = port_array_first(&ps->queues, &ps->last_tx_port);
125 return dequeue_packet(ps, q, ps->last_tx_port);
128 /* Add tokens to the bucket based on elapsed time. */
130 refill_bucket(struct pinsched *ps)
132 long long int now = time_msec();
133 long long int tokens = (now - ps->last_fill) * ps->rate_limit + ps->tokens;
134 if (tokens >= 1000) {
136 ps->tokens = MIN(tokens, ps->burst_limit * 1000);
140 /* Attempts to remove enough tokens from 'ps' to transmit a packet. Returns
141 * true if successful, false otherwise. (In the latter case no tokens are
144 get_token(struct pinsched *ps)
146 if (ps->tokens >= 1000) {
155 pinsched_send(struct pinsched *ps, uint16_t port_no,
156 struct ofpbuf *packet, pinsched_tx_cb *cb, void *aux)
160 } else if (!ps->n_queued && get_token(ps)) {
161 /* In the common case where we are not constrained by the rate limit,
162 * let the packet take the normal path. */
166 /* Otherwise queue it up for the periodic callback to drain out. */
169 /* We are called with a buffer obtained from dpif_recv() that has much
170 * more allocated space than actual content most of the time. Since
171 * we're going to store the packet for some time, free up that
172 * otherwise wasted space. */
175 if (ps->n_queued >= ps->burst_limit) {
178 q = port_array_get(&ps->queues, port_no);
180 q = xmalloc(sizeof *q);
181 list_init(&q->packets);
183 port_array_set(&ps->queues, port_no, q);
185 list_push_back(&q->packets, &packet->list_node);
193 pinsched_status_cb(struct status_reply *sr, void *ps_)
195 struct pinsched *ps = ps_;
197 status_reply_put(sr, "normal=%llu", ps->n_normal);
198 status_reply_put(sr, "limited=%llu", ps->n_limited);
199 status_reply_put(sr, "queue-dropped=%llu", ps->n_queue_dropped);
203 pinsched_run(struct pinsched *ps, pinsched_tx_cb *cb, void *aux)
208 /* Drain some packets out of the bucket if possible, but limit the
209 * number of iterations to allow other code to get work done too. */
211 for (i = 0; ps->n_queued && get_token(ps) && i < 50; i++) {
212 cb(get_tx_packet(ps), aux);
218 pinsched_wait(struct pinsched *ps)
220 if (ps && ps->n_queued) {
221 if (ps->tokens >= 1000) {
222 /* We can transmit more packets as soon as we're called again. */
223 poll_immediate_wake();
225 /* We have to wait for the bucket to re-fill. We could calculate
226 * the exact amount of time here for increased smoothness. */
227 poll_timer_wait(TIME_UPDATE_INTERVAL / 2);
232 /* Creates and returns a scheduler for sending packet-in messages. */
234 pinsched_create(int rate_limit, int burst_limit, struct switch_status *ss)
238 ps = xzalloc(sizeof *ps);
239 port_array_init(&ps->queues);
241 ps->last_tx_port = PORT_ARRAY_SIZE;
242 ps->last_fill = time_msec();
243 ps->tokens = rate_limit * 100;
247 ps->n_queue_dropped = 0;
248 pinsched_set_limits(ps, rate_limit, burst_limit);
251 ps->ss_cat = switch_status_register(ss, "rate-limit",
252 pinsched_status_cb, ps);
259 pinsched_destroy(struct pinsched *ps)
262 struct pinqueue *queue;
263 unsigned int port_no;
265 PORT_ARRAY_FOR_EACH (queue, &ps->queues, port_no) {
266 ofpbuf_list_delete(&queue->packets);
269 port_array_destroy(&ps->queues);
270 switch_status_unregister(ps->ss_cat);
276 pinsched_get_limits(const struct pinsched *ps,
277 int *rate_limit, int *burst_limit)
279 *rate_limit = ps->rate_limit;
280 *burst_limit = ps->burst_limit;
284 pinsched_set_limits(struct pinsched *ps, int rate_limit, int burst_limit)
286 if (rate_limit <= 0) {
289 if (burst_limit <= 0) {
290 burst_limit = rate_limit / 4;
292 burst_limit = MAX(burst_limit, 1);
293 burst_limit = MIN(burst_limit, INT_MAX / 1000);
295 ps->rate_limit = rate_limit;
296 ps->burst_limit = burst_limit;
297 while (ps->n_queued > burst_limit) {