2 * net/sched/sch_sfq.c Stochastic Fairness Queueing discipline.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru>
12 #include <linux/module.h>
13 #include <linux/types.h>
14 #include <linux/kernel.h>
15 #include <linux/jiffies.h>
16 #include <linux/string.h>
18 #include <linux/errno.h>
19 #include <linux/init.h>
20 #include <linux/ipv6.h>
21 #include <linux/skbuff.h>
22 #include <linux/jhash.h>
24 #include <net/netlink.h>
25 #include <net/pkt_sched.h>
28 /* Stochastic Fairness Queuing algorithm.
29 =======================================
32 Paul E. McKenney "Stochastic Fairness Queuing",
33 IEEE INFOCOMM'90 Proceedings, San Francisco, 1990.
35 Paul E. McKenney "Stochastic Fairness Queuing",
36 "Interworking: Research and Experience", v.2, 1991, p.113-131.
40 M. Shreedhar and George Varghese "Efficient Fair
41 Queuing using Deficit Round Robin", Proc. SIGCOMM 95.
44 This is not the thing that is usually called (W)FQ nowadays.
45 It does not use any timestamp mechanism, but instead
46 processes queues in round-robin order.
50 - It is very cheap. Both CPU and memory requirements are minimal.
54 - "Stochastic" -> It is not 100% fair.
55 When hash collisions occur, several flows are considered as one.
57 - "Round-robin" -> It introduces larger delays than virtual clock
58 based schemes, and should not be used for isolating interactive
59 traffic from non-interactive. It means, that this scheduler
60 should be used as leaf of CBQ or P3, which put interactive traffic
61 to higher priority band.
63 We still need true WFQ for top level CSZ, but using WFQ
64 for the best effort traffic is absolutely pointless:
65 SFQ is superior for this purpose.
68 This implementation limits maximal queue length to 128;
69 maximal mtu to 2^15-1; number of hash buckets to 1024.
70 The only goal of this restrictions was that all data
71 fit into one 4K page :-). Struct sfq_sched_data is
72 organized in anti-cache manner: all the data for a bucket
73 are scattered over different locations. This is not good,
74 but it allowed me to put it into 4K.
76 It is easy to increase these values, but not in flight. */
79 #define SFQ_HASH_DIVISOR 1024
81 /* This type should contain at least SFQ_DEPTH*2 values */
82 typedef unsigned char sfq_index;
94 unsigned quantum; /* Allotment per round: MUST BE >= MTU */
98 struct tcf_proto *filter_list;
99 struct timer_list perturb_timer;
101 sfq_index tail; /* Index of current slot in round */
102 sfq_index max_depth; /* Maximal depth */
104 sfq_index ht[SFQ_HASH_DIVISOR]; /* Hash table */
105 sfq_index next[SFQ_DEPTH]; /* Active slots link */
106 short allot[SFQ_DEPTH]; /* Current allotment per slot */
107 unsigned short hash[SFQ_DEPTH]; /* Hash value indexed by slots */
108 struct sk_buff_head qs[SFQ_DEPTH]; /* Slot queue */
109 struct sfq_head dep[SFQ_DEPTH*2]; /* Linked list of slots, indexed by depth */
112 static __inline__ unsigned sfq_fold_hash(struct sfq_sched_data *q, u32 h, u32 h1)
114 return jhash_2words(h, h1, q->perturbation) & (SFQ_HASH_DIVISOR - 1);
117 static unsigned sfq_hash(struct sfq_sched_data *q, struct sk_buff *skb)
121 switch (skb->protocol) {
122 case __constant_htons(ETH_P_IP):
124 const struct iphdr *iph = ip_hdr(skb);
126 h2 = iph->saddr ^ iph->protocol;
127 if (!(iph->frag_off&htons(IP_MF|IP_OFFSET)) &&
128 (iph->protocol == IPPROTO_TCP ||
129 iph->protocol == IPPROTO_UDP ||
130 iph->protocol == IPPROTO_UDPLITE ||
131 iph->protocol == IPPROTO_SCTP ||
132 iph->protocol == IPPROTO_DCCP ||
133 iph->protocol == IPPROTO_ESP))
134 h2 ^= *(((u32*)iph) + iph->ihl);
137 case __constant_htons(ETH_P_IPV6):
139 struct ipv6hdr *iph = ipv6_hdr(skb);
140 h = iph->daddr.s6_addr32[3];
141 h2 = iph->saddr.s6_addr32[3] ^ iph->nexthdr;
142 if (iph->nexthdr == IPPROTO_TCP ||
143 iph->nexthdr == IPPROTO_UDP ||
144 iph->nexthdr == IPPROTO_UDPLITE ||
145 iph->nexthdr == IPPROTO_SCTP ||
146 iph->nexthdr == IPPROTO_DCCP ||
147 iph->nexthdr == IPPROTO_ESP)
148 h2 ^= *(u32*)&iph[1];
152 h = (unsigned long)skb->dst ^ skb->protocol;
153 h2 = (unsigned long)skb->sk;
156 return sfq_fold_hash(q, h, h2);
159 static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch,
162 struct sfq_sched_data *q = qdisc_priv(sch);
163 struct tcf_result res;
166 if (TC_H_MAJ(skb->priority) == sch->handle &&
167 TC_H_MIN(skb->priority) > 0 &&
168 TC_H_MIN(skb->priority) <= SFQ_HASH_DIVISOR)
169 return TC_H_MIN(skb->priority);
172 return sfq_hash(q, skb) + 1;
174 *qerr = NET_XMIT_BYPASS;
175 result = tc_classify(skb, q->filter_list, &res);
177 #ifdef CONFIG_NET_CLS_ACT
181 *qerr = NET_XMIT_SUCCESS;
186 if (TC_H_MIN(res.classid) <= SFQ_HASH_DIVISOR)
187 return TC_H_MIN(res.classid);
192 static inline void sfq_link(struct sfq_sched_data *q, sfq_index x)
195 int d = q->qs[x].qlen + SFQ_DEPTH;
201 q->dep[p].next = q->dep[n].prev = x;
204 static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x)
213 if (n == p && q->max_depth == q->qs[x].qlen + 1)
219 static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x)
229 if (q->max_depth < d)
235 static unsigned int sfq_drop(struct Qdisc *sch)
237 struct sfq_sched_data *q = qdisc_priv(sch);
238 sfq_index d = q->max_depth;
242 /* Queue is full! Find the longest slot and
243 drop a packet from it */
246 sfq_index x = q->dep[d + SFQ_DEPTH].next;
249 __skb_unlink(skb, &q->qs[x]);
254 sch->qstats.backlog -= len;
259 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
260 d = q->next[q->tail];
261 q->next[q->tail] = q->next[d];
262 q->allot[q->next[d]] += q->quantum;
265 __skb_unlink(skb, &q->qs[d]);
269 q->ht[q->hash[d]] = SFQ_DEPTH;
271 sch->qstats.backlog -= len;
279 sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
281 struct sfq_sched_data *q = qdisc_priv(sch);
286 hash = sfq_classify(skb, sch, &ret);
288 if (ret == NET_XMIT_BYPASS)
296 if (x == SFQ_DEPTH) {
297 q->ht[hash] = x = q->dep[SFQ_DEPTH].next;
301 /* If selected queue has length q->limit, this means that
302 * all another queues are empty and that we do simple tail drop,
303 * i.e. drop _this_ packet.
305 if (q->qs[x].qlen >= q->limit)
306 return qdisc_drop(skb, sch);
308 sch->qstats.backlog += skb->len;
309 __skb_queue_tail(&q->qs[x], skb);
311 if (q->qs[x].qlen == 1) { /* The flow is new */
312 if (q->tail == SFQ_DEPTH) { /* It is the first flow */
315 q->allot[x] = q->quantum;
317 q->next[x] = q->next[q->tail];
318 q->next[q->tail] = x;
322 if (++sch->q.qlen <= q->limit) {
323 sch->bstats.bytes += skb->len;
324 sch->bstats.packets++;
333 sfq_requeue(struct sk_buff *skb, struct Qdisc *sch)
335 struct sfq_sched_data *q = qdisc_priv(sch);
340 hash = sfq_classify(skb, sch, &ret);
342 if (ret == NET_XMIT_BYPASS)
350 if (x == SFQ_DEPTH) {
351 q->ht[hash] = x = q->dep[SFQ_DEPTH].next;
355 sch->qstats.backlog += skb->len;
356 __skb_queue_head(&q->qs[x], skb);
357 /* If selected queue has length q->limit+1, this means that
358 * all another queues are empty and we do simple tail drop.
359 * This packet is still requeued at head of queue, tail packet
362 if (q->qs[x].qlen > q->limit) {
364 __skb_unlink(skb, &q->qs[x]);
366 sch->qstats.backlog -= skb->len;
372 if (q->qs[x].qlen == 1) { /* The flow is new */
373 if (q->tail == SFQ_DEPTH) { /* It is the first flow */
376 q->allot[x] = q->quantum;
378 q->next[x] = q->next[q->tail];
379 q->next[q->tail] = x;
384 if (++sch->q.qlen <= q->limit) {
385 sch->qstats.requeues++;
397 static struct sk_buff *
398 sfq_dequeue(struct Qdisc *sch)
400 struct sfq_sched_data *q = qdisc_priv(sch);
404 /* No active slots */
405 if (q->tail == SFQ_DEPTH)
408 a = old_a = q->next[q->tail];
411 skb = __skb_dequeue(&q->qs[a]);
414 sch->qstats.backlog -= skb->len;
416 /* Is the slot empty? */
417 if (q->qs[a].qlen == 0) {
418 q->ht[q->hash[a]] = SFQ_DEPTH;
424 q->next[q->tail] = a;
425 q->allot[a] += q->quantum;
426 } else if ((q->allot[a] -= skb->len) <= 0) {
429 q->allot[a] += q->quantum;
435 sfq_reset(struct Qdisc *sch)
439 while ((skb = sfq_dequeue(sch)) != NULL)
443 static void sfq_perturbation(unsigned long arg)
445 struct Qdisc *sch = (struct Qdisc *)arg;
446 struct sfq_sched_data *q = qdisc_priv(sch);
448 q->perturbation = net_random();
450 if (q->perturb_period)
451 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
454 static int sfq_change(struct Qdisc *sch, struct nlattr *opt)
456 struct sfq_sched_data *q = qdisc_priv(sch);
457 struct tc_sfq_qopt *ctl = nla_data(opt);
460 if (opt->nla_len < nla_attr_size(sizeof(*ctl)))
464 q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
465 q->perturb_period = ctl->perturb_period * HZ;
467 q->limit = min_t(u32, ctl->limit, SFQ_DEPTH - 1);
470 while (sch->q.qlen > q->limit)
472 qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen);
474 del_timer(&q->perturb_timer);
475 if (q->perturb_period) {
476 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
477 q->perturbation = net_random();
479 sch_tree_unlock(sch);
483 static int sfq_init(struct Qdisc *sch, struct nlattr *opt)
485 struct sfq_sched_data *q = qdisc_priv(sch);
488 q->perturb_timer.function = sfq_perturbation;
489 q->perturb_timer.data = (unsigned long)sch;;
490 init_timer_deferrable(&q->perturb_timer);
492 for (i = 0; i < SFQ_HASH_DIVISOR; i++)
493 q->ht[i] = SFQ_DEPTH;
495 for (i = 0; i < SFQ_DEPTH; i++) {
496 skb_queue_head_init(&q->qs[i]);
497 q->dep[i + SFQ_DEPTH].next = i + SFQ_DEPTH;
498 q->dep[i + SFQ_DEPTH].prev = i + SFQ_DEPTH;
501 q->limit = SFQ_DEPTH - 1;
505 q->quantum = psched_mtu(sch->dev);
506 q->perturb_period = 0;
507 q->perturbation = net_random();
509 int err = sfq_change(sch, opt);
514 for (i = 0; i < SFQ_DEPTH; i++)
519 static void sfq_destroy(struct Qdisc *sch)
521 struct sfq_sched_data *q = qdisc_priv(sch);
523 tcf_destroy_chain(&q->filter_list);
524 q->perturb_period = 0;
525 del_timer_sync(&q->perturb_timer);
528 static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb)
530 struct sfq_sched_data *q = qdisc_priv(sch);
531 unsigned char *b = skb_tail_pointer(skb);
532 struct tc_sfq_qopt opt;
534 opt.quantum = q->quantum;
535 opt.perturb_period = q->perturb_period / HZ;
537 opt.limit = q->limit;
538 opt.divisor = SFQ_HASH_DIVISOR;
539 opt.flows = q->limit;
541 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
550 static int sfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
551 struct nlattr **tca, unsigned long *arg)
556 static unsigned long sfq_get(struct Qdisc *sch, u32 classid)
561 static struct tcf_proto **sfq_find_tcf(struct Qdisc *sch, unsigned long cl)
563 struct sfq_sched_data *q = qdisc_priv(sch);
567 return &q->filter_list;
570 static int sfq_dump_class(struct Qdisc *sch, unsigned long cl,
571 struct sk_buff *skb, struct tcmsg *tcm)
573 tcm->tcm_handle |= TC_H_MIN(cl);
577 static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
580 struct sfq_sched_data *q = qdisc_priv(sch);
581 sfq_index idx = q->ht[cl-1];
582 struct gnet_stats_queue qs = { .qlen = q->qs[idx].qlen };
583 struct tc_sfq_xstats xstats = { .allot = q->allot[idx] };
585 if (gnet_stats_copy_queue(d, &qs) < 0)
587 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
590 static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
592 struct sfq_sched_data *q = qdisc_priv(sch);
598 for (i = 0; i < SFQ_HASH_DIVISOR; i++) {
599 if (q->ht[i] == SFQ_DEPTH ||
600 arg->count < arg->skip) {
604 if (arg->fn(sch, i + 1, arg) < 0) {
612 static const struct Qdisc_class_ops sfq_class_ops = {
614 .change = sfq_change_class,
615 .tcf_chain = sfq_find_tcf,
616 .dump = sfq_dump_class,
617 .dump_stats = sfq_dump_class_stats,
621 static struct Qdisc_ops sfq_qdisc_ops __read_mostly = {
622 .cl_ops = &sfq_class_ops,
624 .priv_size = sizeof(struct sfq_sched_data),
625 .enqueue = sfq_enqueue,
626 .dequeue = sfq_dequeue,
627 .requeue = sfq_requeue,
631 .destroy = sfq_destroy,
634 .owner = THIS_MODULE,
637 static int __init sfq_module_init(void)
639 return register_qdisc(&sfq_qdisc_ops);
641 static void __exit sfq_module_exit(void)
643 unregister_qdisc(&sfq_qdisc_ops);
645 module_init(sfq_module_init)
646 module_exit(sfq_module_exit)
647 MODULE_LICENSE("GPL");