2 * Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013, 2014 Nicira, Inc.
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>
28 #include "openflow/openflow.h"
29 #include "poll-loop.h"
34 #include "token-bucket.h"
38 struct hmap_node node; /* In struct pinsched's 'queues' hmap. */
39 ofp_port_t port_no; /* Port number. */
40 struct list packets; /* Contains "struct ofpbuf"s. */
41 int n; /* Number of packets in 'packets'. */
45 struct token_bucket token_bucket;
47 /* One queue per physical port. */
48 struct hmap queues; /* Contains "struct pinqueue"s. */
49 unsigned int n_queued; /* Sum over queues[*].n. */
50 struct pinqueue *next_txq; /* Next pinqueue check in round-robin. */
52 /* Statistics reporting. */
53 unsigned long long n_normal; /* # txed w/o rate limit queuing. */
54 unsigned long long n_limited; /* # queued for rate limiting. */
55 unsigned long long n_queue_dropped; /* # dropped due to queue overflow. */
59 advance_txq(struct pinsched *ps)
61 struct hmap_node *next;
64 ? hmap_next(&ps->queues, &ps->next_txq->node)
65 : hmap_first(&ps->queues));
66 ps->next_txq = next ? CONTAINER_OF(next, struct pinqueue, node) : NULL;
69 static struct ofpbuf *
70 dequeue_packet(struct pinsched *ps, struct pinqueue *q)
72 struct ofpbuf *packet = ofpbuf_from_list(list_pop_front(&q->packets));
79 adjust_limits(int *rate_limit, int *burst_limit)
81 if (*rate_limit <= 0) {
84 if (*burst_limit <= 0) {
85 *burst_limit = *rate_limit / 4;
87 if (*burst_limit < 1) {
92 /* Destroys 'q' and removes it from 'ps''s set of queues.
93 * (The caller must ensure that 'q' is empty.) */
95 pinqueue_destroy(struct pinsched *ps, struct pinqueue *q)
97 hmap_remove(&ps->queues, &q->node);
101 static struct pinqueue *
102 pinqueue_get(struct pinsched *ps, ofp_port_t port_no)
104 uint32_t hash = hash_ofp_port(port_no);
107 HMAP_FOR_EACH_IN_BUCKET (q, node, hash, &ps->queues) {
108 if (port_no == q->port_no) {
113 q = xmalloc(sizeof *q);
114 hmap_insert(&ps->queues, &q->node, hash);
115 q->port_no = port_no;
116 list_init(&q->packets);
121 /* Drop a packet from the longest queue in 'ps'. */
123 drop_packet(struct pinsched *ps)
125 struct pinqueue *longest; /* Queue currently selected as longest. */
126 int n_longest = 0; /* # of queues of same length as 'longest'. */
129 ps->n_queue_dropped++;
132 HMAP_FOR_EACH (q, node, &ps->queues) {
133 if (!longest || longest->n < q->n) {
136 } else if (longest->n == q->n) {
139 /* Randomly select one of the longest queues, with a uniform
140 * distribution (Knuth algorithm 3.4.2R). */
141 if (!random_range(n_longest)) {
147 /* FIXME: do we want to pop the tail instead? */
148 ofpbuf_delete(dequeue_packet(ps, longest));
149 if (longest->n == 0) {
150 pinqueue_destroy(ps, longest);
154 /* Remove and return the next packet to transmit (in round-robin order). */
155 static struct ofpbuf *
156 get_tx_packet(struct pinsched *ps)
158 struct ofpbuf *packet;
166 packet = dequeue_packet(ps, q);
169 pinqueue_destroy(ps, q);
175 /* Attempts to remove enough tokens from 'ps' to transmit a packet. Returns
176 * true if successful, false otherwise. (In the latter case no tokens are
179 get_token(struct pinsched *ps)
181 return token_bucket_withdraw(&ps->token_bucket, 1000);
185 pinsched_send(struct pinsched *ps, ofp_port_t port_no,
186 struct ofpbuf *packet, struct list *txq)
190 list_push_back(txq, &packet->list_node);
191 } else if (!ps->n_queued && get_token(ps)) {
192 /* In the common case where we are not constrained by the rate limit,
193 * let the packet take the normal path. */
195 list_push_back(txq, &packet->list_node);
197 /* Otherwise queue it up for the periodic callback to drain out. */
200 /* We might be called with a buffer obtained from dpif_recv() that has
201 * much more allocated space than actual content most of the time.
202 * Since we're going to store the packet for some time, free up that
203 * otherwise wasted space. */
206 if (ps->n_queued * 1000 >= ps->token_bucket.burst) {
209 q = pinqueue_get(ps, port_no);
210 list_push_back(&q->packets, &packet->list_node);
218 pinsched_run(struct pinsched *ps, struct list *txq)
224 /* Drain some packets out of the bucket if possible, but limit the
225 * number of iterations to allow other code to get work done too. */
226 for (i = 0; ps->n_queued && get_token(ps) && i < 50; i++) {
227 struct ofpbuf *packet = get_tx_packet(ps);
228 list_push_back(txq, &packet->list_node);
234 pinsched_wait(struct pinsched *ps)
236 if (ps && ps->n_queued) {
237 token_bucket_wait(&ps->token_bucket, 1000);
241 /* Creates and returns a scheduler for sending packet-in messages. */
243 pinsched_create(int rate_limit, int burst_limit)
247 ps = xzalloc(sizeof *ps);
249 adjust_limits(&rate_limit, &burst_limit);
250 token_bucket_init(&ps->token_bucket,
251 rate_limit, sat_mul(burst_limit, 1000));
253 hmap_init(&ps->queues);
258 ps->n_queue_dropped = 0;
264 pinsched_destroy(struct pinsched *ps)
267 struct pinqueue *q, *next;
269 HMAP_FOR_EACH_SAFE (q, next, node, &ps->queues) {
270 hmap_remove(&ps->queues, &q->node);
271 ofpbuf_list_delete(&q->packets);
274 hmap_destroy(&ps->queues);
280 pinsched_get_limits(const struct pinsched *ps,
281 int *rate_limit, int *burst_limit)
283 *rate_limit = ps->token_bucket.rate;
284 *burst_limit = ps->token_bucket.burst / 1000;
288 pinsched_set_limits(struct pinsched *ps, int rate_limit, int burst_limit)
290 adjust_limits(&rate_limit, &burst_limit);
291 token_bucket_set(&ps->token_bucket,
292 rate_limit, sat_mul(burst_limit, 1000));
293 while (ps->n_queued > burst_limit) {
298 /* Retrieves statistics for 'ps'. The statistics will be all zero if 'ps' is
301 pinsched_get_stats(const struct pinsched *ps, struct pinsched_stats *stats)
304 stats->n_queued = ps->n_queued;
305 stats->n_normal = ps->n_normal;
306 stats->n_limited = ps->n_limited;
307 stats->n_queue_dropped = ps->n_queue_dropped;
309 memset(stats, 0, sizeof *stats);