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 <asm/uaccess.h>
14 #include <asm/system.h>
15 #include <linux/bitops.h>
16 #include <linux/types.h>
17 #include <linux/kernel.h>
18 #include <linux/jiffies.h>
19 #include <linux/string.h>
21 #include <linux/socket.h>
22 #include <linux/sockios.h>
24 #include <linux/errno.h>
25 #include <linux/interrupt.h>
26 #include <linux/if_ether.h>
27 #include <linux/inet.h>
28 #include <linux/netdevice.h>
29 #include <linux/etherdevice.h>
30 #include <linux/notifier.h>
31 #include <linux/init.h>
33 #include <linux/ipv6.h>
34 #include <net/route.h>
35 #include <linux/skbuff.h>
37 #include <net/pkt_sched.h>
40 /* Stochastic Fairness Queuing algorithm.
41 =======================================
44 Paul E. McKenney "Stochastic Fairness Queuing",
45 IEEE INFOCOMM'90 Proceedings, San Francisco, 1990.
47 Paul E. McKenney "Stochastic Fairness Queuing",
48 "Interworking: Research and Experience", v.2, 1991, p.113-131.
52 M. Shreedhar and George Varghese "Efficient Fair
53 Queuing using Deficit Round Robin", Proc. SIGCOMM 95.
56 This is not the thing that is usually called (W)FQ nowadays.
57 It does not use any timestamp mechanism, but instead
58 processes queues in round-robin order.
62 - It is very cheap. Both CPU and memory requirements are minimal.
66 - "Stochastic" -> It is not 100% fair.
67 When hash collisions occur, several flows are considered as one.
69 - "Round-robin" -> It introduces larger delays than virtual clock
70 based schemes, and should not be used for isolating interactive
71 traffic from non-interactive. It means, that this scheduler
72 should be used as leaf of CBQ or P3, which put interactive traffic
73 to higher priority band.
75 We still need true WFQ for top level CSZ, but using WFQ
76 for the best effort traffic is absolutely pointless:
77 SFQ is superior for this purpose.
80 This implementation limits maximal queue length to 128;
81 maximal mtu to 2^15-1; number of hash buckets to 1024.
82 The only goal of this restrictions was that all data
83 fit into one 4K page :-). Struct sfq_sched_data is
84 organized in anti-cache manner: all the data for a bucket
85 are scattered over different locations. This is not good,
86 but it allowed me to put it into 4K.
88 It is easy to increase these values, but not in flight. */
91 #define SFQ_HASH_DIVISOR 1024
93 /* This type should contain at least SFQ_DEPTH*2 values */
94 typedef unsigned char sfq_index;
102 struct sfq_sched_data
106 unsigned quantum; /* Allotment per round: MUST BE >= MTU */
110 struct timer_list perturb_timer;
112 sfq_index tail; /* Index of current slot in round */
113 sfq_index max_depth; /* Maximal depth */
115 sfq_index ht[SFQ_HASH_DIVISOR]; /* Hash table */
116 sfq_index next[SFQ_DEPTH]; /* Active slots link */
117 short allot[SFQ_DEPTH]; /* Current allotment per slot */
118 unsigned short hash[SFQ_DEPTH]; /* Hash value indexed by slots */
119 struct sk_buff_head qs[SFQ_DEPTH]; /* Slot queue */
120 struct sfq_head dep[SFQ_DEPTH*2]; /* Linked list of slots, indexed by depth */
123 static __inline__ unsigned sfq_fold_hash(struct sfq_sched_data *q, u32 h, u32 h1)
125 int pert = q->perturbation;
127 /* Have we any rotation primitives? If not, WHY? */
128 h ^= (h1<<pert) ^ (h1>>(0x1F - pert));
133 static unsigned sfq_hash(struct sfq_sched_data *q, struct sk_buff *skb)
137 switch (skb->protocol) {
138 case __constant_htons(ETH_P_IP):
140 struct iphdr *iph = skb->nh.iph;
142 h2 = iph->saddr^iph->protocol;
143 if (!(iph->frag_off&htons(IP_MF|IP_OFFSET)) &&
144 (iph->protocol == IPPROTO_TCP ||
145 iph->protocol == IPPROTO_UDP ||
146 iph->protocol == IPPROTO_SCTP ||
147 iph->protocol == IPPROTO_DCCP ||
148 iph->protocol == IPPROTO_ESP))
149 h2 ^= *(((u32*)iph) + iph->ihl);
152 case __constant_htons(ETH_P_IPV6):
154 struct ipv6hdr *iph = skb->nh.ipv6h;
155 h = iph->daddr.s6_addr32[3];
156 h2 = iph->saddr.s6_addr32[3]^iph->nexthdr;
157 if (iph->nexthdr == IPPROTO_TCP ||
158 iph->nexthdr == IPPROTO_UDP ||
159 iph->nexthdr == IPPROTO_SCTP ||
160 iph->nexthdr == IPPROTO_DCCP ||
161 iph->nexthdr == IPPROTO_ESP)
162 h2 ^= *(u32*)&iph[1];
166 h = (u32)(unsigned long)skb->dst^skb->protocol;
167 h2 = (u32)(unsigned long)skb->sk;
169 return sfq_fold_hash(q, h, h2);
172 static inline void sfq_link(struct sfq_sched_data *q, sfq_index x)
175 int d = q->qs[x].qlen + SFQ_DEPTH;
181 q->dep[p].next = q->dep[n].prev = x;
184 static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x)
193 if (n == p && q->max_depth == q->qs[x].qlen + 1)
199 static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x)
209 if (q->max_depth < d)
215 static unsigned int sfq_drop(struct Qdisc *sch)
217 struct sfq_sched_data *q = qdisc_priv(sch);
218 sfq_index d = q->max_depth;
222 /* Queue is full! Find the longest slot and
223 drop a packet from it */
226 sfq_index x = q->dep[d+SFQ_DEPTH].next;
229 __skb_unlink(skb, &q->qs[x]);
234 sch->qstats.backlog -= len;
239 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
240 d = q->next[q->tail];
241 q->next[q->tail] = q->next[d];
242 q->allot[q->next[d]] += q->quantum;
245 __skb_unlink(skb, &q->qs[d]);
249 q->ht[q->hash[d]] = SFQ_DEPTH;
251 sch->qstats.backlog -= len;
259 sfq_enqueue(struct sk_buff *skb, struct Qdisc* sch)
261 struct sfq_sched_data *q = qdisc_priv(sch);
262 unsigned hash = sfq_hash(q, skb);
266 if (x == SFQ_DEPTH) {
267 q->ht[hash] = x = q->dep[SFQ_DEPTH].next;
270 sch->qstats.backlog += skb->len;
271 __skb_queue_tail(&q->qs[x], skb);
273 if (q->qs[x].qlen == 1) { /* The flow is new */
274 if (q->tail == SFQ_DEPTH) { /* It is the first flow */
277 q->allot[x] = q->quantum;
279 q->next[x] = q->next[q->tail];
280 q->next[q->tail] = x;
284 if (++sch->q.qlen < q->limit-1) {
285 sch->bstats.bytes += skb->len;
286 sch->bstats.packets++;
295 sfq_requeue(struct sk_buff *skb, struct Qdisc* sch)
297 struct sfq_sched_data *q = qdisc_priv(sch);
298 unsigned hash = sfq_hash(q, skb);
302 if (x == SFQ_DEPTH) {
303 q->ht[hash] = x = q->dep[SFQ_DEPTH].next;
306 sch->qstats.backlog += skb->len;
307 __skb_queue_head(&q->qs[x], skb);
309 if (q->qs[x].qlen == 1) { /* The flow is new */
310 if (q->tail == SFQ_DEPTH) { /* It is the first flow */
313 q->allot[x] = q->quantum;
315 q->next[x] = q->next[q->tail];
316 q->next[q->tail] = x;
320 if (++sch->q.qlen < q->limit - 1) {
321 sch->qstats.requeues++;
333 static struct sk_buff *
334 sfq_dequeue(struct Qdisc* sch)
336 struct sfq_sched_data *q = qdisc_priv(sch);
340 /* No active slots */
341 if (q->tail == SFQ_DEPTH)
344 a = old_a = q->next[q->tail];
347 skb = __skb_dequeue(&q->qs[a]);
350 sch->qstats.backlog -= skb->len;
352 /* Is the slot empty? */
353 if (q->qs[a].qlen == 0) {
354 q->ht[q->hash[a]] = SFQ_DEPTH;
360 q->next[q->tail] = a;
361 q->allot[a] += q->quantum;
362 } else if ((q->allot[a] -= skb->len) <= 0) {
365 q->allot[a] += q->quantum;
371 sfq_reset(struct Qdisc* sch)
375 while ((skb = sfq_dequeue(sch)) != NULL)
379 static void sfq_perturbation(unsigned long arg)
381 struct Qdisc *sch = (struct Qdisc*)arg;
382 struct sfq_sched_data *q = qdisc_priv(sch);
384 q->perturbation = net_random()&0x1F;
386 if (q->perturb_period) {
387 q->perturb_timer.expires = jiffies + q->perturb_period;
388 add_timer(&q->perturb_timer);
392 static int sfq_change(struct Qdisc *sch, struct rtattr *opt)
394 struct sfq_sched_data *q = qdisc_priv(sch);
395 struct tc_sfq_qopt *ctl = RTA_DATA(opt);
397 if (opt->rta_len < RTA_LENGTH(sizeof(*ctl)))
401 q->quantum = ctl->quantum ? : psched_mtu(sch->dev);
402 q->perturb_period = ctl->perturb_period*HZ;
404 q->limit = min_t(u32, ctl->limit, SFQ_DEPTH);
406 while (sch->q.qlen >= q->limit-1)
409 del_timer(&q->perturb_timer);
410 if (q->perturb_period) {
411 q->perturb_timer.expires = jiffies + q->perturb_period;
412 add_timer(&q->perturb_timer);
414 sch_tree_unlock(sch);
418 static int sfq_init(struct Qdisc *sch, struct rtattr *opt)
420 struct sfq_sched_data *q = qdisc_priv(sch);
423 init_timer(&q->perturb_timer);
424 q->perturb_timer.data = (unsigned long)sch;
425 q->perturb_timer.function = sfq_perturbation;
427 for (i=0; i<SFQ_HASH_DIVISOR; i++)
428 q->ht[i] = SFQ_DEPTH;
429 for (i=0; i<SFQ_DEPTH; i++) {
430 skb_queue_head_init(&q->qs[i]);
431 q->dep[i+SFQ_DEPTH].next = i+SFQ_DEPTH;
432 q->dep[i+SFQ_DEPTH].prev = i+SFQ_DEPTH;
434 q->limit = SFQ_DEPTH;
438 q->quantum = psched_mtu(sch->dev);
439 q->perturb_period = 0;
441 int err = sfq_change(sch, opt);
445 for (i=0; i<SFQ_DEPTH; i++)
450 static void sfq_destroy(struct Qdisc *sch)
452 struct sfq_sched_data *q = qdisc_priv(sch);
453 del_timer(&q->perturb_timer);
456 static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb)
458 struct sfq_sched_data *q = qdisc_priv(sch);
459 unsigned char *b = skb->tail;
460 struct tc_sfq_qopt opt;
462 opt.quantum = q->quantum;
463 opt.perturb_period = q->perturb_period/HZ;
465 opt.limit = q->limit;
466 opt.divisor = SFQ_HASH_DIVISOR;
467 opt.flows = q->limit;
469 RTA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
474 skb_trim(skb, b - skb->data);
478 static struct Qdisc_ops sfq_qdisc_ops = {
482 .priv_size = sizeof(struct sfq_sched_data),
483 .enqueue = sfq_enqueue,
484 .dequeue = sfq_dequeue,
485 .requeue = sfq_requeue,
489 .destroy = sfq_destroy,
492 .owner = THIS_MODULE,
495 static int __init sfq_module_init(void)
497 return register_qdisc(&sfq_qdisc_ops);
499 static void __exit sfq_module_exit(void)
501 unregister_qdisc(&sfq_qdisc_ops);
503 module_init(sfq_module_init)
504 module_exit(sfq_module_exit)
505 MODULE_LICENSE("GPL");