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 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 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_SUCCESS | __NET_XMIT_BYPASS;
175 result = tc_classify(skb, q->filter_list, &res);
177 #ifdef CONFIG_NET_CLS_ACT
181 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
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;
248 len = qdisc_pkt_len(skb);
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;
264 len = qdisc_pkt_len(skb);
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 += qdisc_pkt_len(skb);
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 += qdisc_pkt_len(skb);
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 += qdisc_pkt_len(skb);
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 -= qdisc_pkt_len(skb);
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++;
394 static struct sk_buff *
395 sfq_peek(struct Qdisc *sch)
397 struct sfq_sched_data *q = qdisc_priv(sch);
400 /* No active slots */
401 if (q->tail == SFQ_DEPTH)
404 a = q->next[q->tail];
405 return skb_peek(&q->qs[a]);
408 static struct sk_buff *
409 sfq_dequeue(struct Qdisc *sch)
411 struct sfq_sched_data *q = qdisc_priv(sch);
415 /* No active slots */
416 if (q->tail == SFQ_DEPTH)
419 a = old_a = q->next[q->tail];
422 skb = __skb_dequeue(&q->qs[a]);
425 sch->qstats.backlog -= qdisc_pkt_len(skb);
427 /* Is the slot empty? */
428 if (q->qs[a].qlen == 0) {
429 q->ht[q->hash[a]] = SFQ_DEPTH;
435 q->next[q->tail] = a;
436 q->allot[a] += q->quantum;
437 } else if ((q->allot[a] -= qdisc_pkt_len(skb)) <= 0) {
440 q->allot[a] += q->quantum;
446 sfq_reset(struct Qdisc *sch)
450 while ((skb = sfq_dequeue(sch)) != NULL)
454 static void sfq_perturbation(unsigned long arg)
456 struct Qdisc *sch = (struct Qdisc *)arg;
457 struct sfq_sched_data *q = qdisc_priv(sch);
459 q->perturbation = net_random();
461 if (q->perturb_period)
462 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
465 static int sfq_change(struct Qdisc *sch, struct nlattr *opt)
467 struct sfq_sched_data *q = qdisc_priv(sch);
468 struct tc_sfq_qopt *ctl = nla_data(opt);
471 if (opt->nla_len < nla_attr_size(sizeof(*ctl)))
475 q->quantum = ctl->quantum ? : psched_mtu(qdisc_dev(sch));
476 q->perturb_period = ctl->perturb_period * HZ;
478 q->limit = min_t(u32, ctl->limit, SFQ_DEPTH - 1);
481 while (sch->q.qlen > q->limit)
483 qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen);
485 del_timer(&q->perturb_timer);
486 if (q->perturb_period) {
487 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
488 q->perturbation = net_random();
490 sch_tree_unlock(sch);
494 static int sfq_init(struct Qdisc *sch, struct nlattr *opt)
496 struct sfq_sched_data *q = qdisc_priv(sch);
499 q->perturb_timer.function = sfq_perturbation;
500 q->perturb_timer.data = (unsigned long)sch;;
501 init_timer_deferrable(&q->perturb_timer);
503 for (i = 0; i < SFQ_HASH_DIVISOR; i++)
504 q->ht[i] = SFQ_DEPTH;
506 for (i = 0; i < SFQ_DEPTH; i++) {
507 skb_queue_head_init(&q->qs[i]);
508 q->dep[i + SFQ_DEPTH].next = i + SFQ_DEPTH;
509 q->dep[i + SFQ_DEPTH].prev = i + SFQ_DEPTH;
512 q->limit = SFQ_DEPTH - 1;
516 q->quantum = psched_mtu(qdisc_dev(sch));
517 q->perturb_period = 0;
518 q->perturbation = net_random();
520 int err = sfq_change(sch, opt);
525 for (i = 0; i < SFQ_DEPTH; i++)
530 static void sfq_destroy(struct Qdisc *sch)
532 struct sfq_sched_data *q = qdisc_priv(sch);
534 tcf_destroy_chain(&q->filter_list);
535 q->perturb_period = 0;
536 del_timer_sync(&q->perturb_timer);
539 static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb)
541 struct sfq_sched_data *q = qdisc_priv(sch);
542 unsigned char *b = skb_tail_pointer(skb);
543 struct tc_sfq_qopt opt;
545 opt.quantum = q->quantum;
546 opt.perturb_period = q->perturb_period / HZ;
548 opt.limit = q->limit;
549 opt.divisor = SFQ_HASH_DIVISOR;
550 opt.flows = q->limit;
552 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
561 static int sfq_change_class(struct Qdisc *sch, u32 classid, u32 parentid,
562 struct nlattr **tca, unsigned long *arg)
567 static unsigned long sfq_get(struct Qdisc *sch, u32 classid)
572 static struct tcf_proto **sfq_find_tcf(struct Qdisc *sch, unsigned long cl)
574 struct sfq_sched_data *q = qdisc_priv(sch);
578 return &q->filter_list;
581 static int sfq_dump_class(struct Qdisc *sch, unsigned long cl,
582 struct sk_buff *skb, struct tcmsg *tcm)
584 tcm->tcm_handle |= TC_H_MIN(cl);
588 static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
591 struct sfq_sched_data *q = qdisc_priv(sch);
592 sfq_index idx = q->ht[cl-1];
593 struct gnet_stats_queue qs = { .qlen = q->qs[idx].qlen };
594 struct tc_sfq_xstats xstats = { .allot = q->allot[idx] };
596 if (gnet_stats_copy_queue(d, &qs) < 0)
598 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
601 static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
603 struct sfq_sched_data *q = qdisc_priv(sch);
609 for (i = 0; i < SFQ_HASH_DIVISOR; i++) {
610 if (q->ht[i] == SFQ_DEPTH ||
611 arg->count < arg->skip) {
615 if (arg->fn(sch, i + 1, arg) < 0) {
623 static const struct Qdisc_class_ops sfq_class_ops = {
625 .change = sfq_change_class,
626 .tcf_chain = sfq_find_tcf,
627 .dump = sfq_dump_class,
628 .dump_stats = sfq_dump_class_stats,
632 static struct Qdisc_ops sfq_qdisc_ops __read_mostly = {
633 .cl_ops = &sfq_class_ops,
635 .priv_size = sizeof(struct sfq_sched_data),
636 .enqueue = sfq_enqueue,
637 .dequeue = sfq_dequeue,
639 .requeue = sfq_requeue,
643 .destroy = sfq_destroy,
646 .owner = THIS_MODULE,
649 static int __init sfq_module_init(void)
651 return register_qdisc(&sfq_qdisc_ops);
653 static void __exit sfq_module_exit(void)
655 unregister_qdisc(&sfq_qdisc_ops);
657 module_init(sfq_module_init)
658 module_exit(sfq_module_exit)
659 MODULE_LICENSE("GPL");