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>
23 #include <linux/slab.h>
25 #include <net/netlink.h>
26 #include <net/pkt_sched.h>
29 /* Stochastic Fairness Queuing algorithm.
30 =======================================
33 Paul E. McKenney "Stochastic Fairness Queuing",
34 IEEE INFOCOMM'90 Proceedings, San Francisco, 1990.
36 Paul E. McKenney "Stochastic Fairness Queuing",
37 "Interworking: Research and Experience", v.2, 1991, p.113-131.
41 M. Shreedhar and George Varghese "Efficient Fair
42 Queuing using Deficit Round Robin", Proc. SIGCOMM 95.
45 This is not the thing that is usually called (W)FQ nowadays.
46 It does not use any timestamp mechanism, but instead
47 processes queues in round-robin order.
51 - It is very cheap. Both CPU and memory requirements are minimal.
55 - "Stochastic" -> It is not 100% fair.
56 When hash collisions occur, several flows are considered as one.
58 - "Round-robin" -> It introduces larger delays than virtual clock
59 based schemes, and should not be used for isolating interactive
60 traffic from non-interactive. It means, that this scheduler
61 should be used as leaf of CBQ or P3, which put interactive traffic
62 to higher priority band.
64 We still need true WFQ for top level CSZ, but using WFQ
65 for the best effort traffic is absolutely pointless:
66 SFQ is superior for this purpose.
69 This implementation limits maximal queue length to 128;
70 max mtu to 2^18-1; max 128 flows, number of hash buckets to 1024.
71 The only goal of this restrictions was that all data
72 fit into one 4K page on 32bit arches.
74 It is easy to increase these values, but not in flight. */
76 #define SFQ_DEPTH 128 /* max number of packets per flow */
77 #define SFQ_SLOTS 128 /* max number of flows */
78 #define SFQ_EMPTY_SLOT 255
79 #define SFQ_HASH_DIVISOR 1024
80 /* We use 16 bits to store allot, and want to handle packets up to 64K
81 * Scale allot by 8 (1<<3) so that no overflow occurs.
83 #define SFQ_ALLOT_SHIFT 3
84 #define SFQ_ALLOT_SIZE(X) DIV_ROUND_UP(X, 1 << SFQ_ALLOT_SHIFT)
86 /* This type should contain at least SFQ_DEPTH + SFQ_SLOTS values */
87 typedef unsigned char sfq_index;
90 * We dont use pointers to save space.
91 * Small indexes [0 ... SFQ_SLOTS - 1] are 'pointers' to slots[] array
92 * while following values [SFQ_SLOTS ... SFQ_SLOTS + SFQ_DEPTH - 1]
93 * are 'pointers' to dep[] array
102 struct sk_buff *skblist_next;
103 struct sk_buff *skblist_prev;
104 sfq_index qlen; /* number of skbs in skblist */
105 sfq_index next; /* next slot in sfq chain */
106 struct sfq_head dep; /* anchor in dep[] chains */
107 unsigned short hash; /* hash value (index in ht[]) */
108 short allot; /* credit for this slot */
111 struct sfq_sched_data
115 unsigned quantum; /* Allotment per round: MUST BE >= MTU */
119 struct tcf_proto *filter_list;
120 struct timer_list perturb_timer;
122 sfq_index cur_depth; /* depth of longest slot */
123 unsigned short scaled_quantum; /* SFQ_ALLOT_SIZE(quantum) */
124 struct sfq_slot *tail; /* current slot in round */
125 sfq_index ht[SFQ_HASH_DIVISOR]; /* Hash table */
126 struct sfq_slot slots[SFQ_SLOTS];
127 struct sfq_head dep[SFQ_DEPTH]; /* Linked list of slots, indexed by depth */
131 * sfq_head are either in a sfq_slot or in dep[] array
133 static inline struct sfq_head *sfq_dep_head(struct sfq_sched_data *q, sfq_index val)
136 return &q->slots[val].dep;
137 return &q->dep[val - SFQ_SLOTS];
140 static __inline__ unsigned sfq_fold_hash(struct sfq_sched_data *q, u32 h, u32 h1)
142 return jhash_2words(h, h1, q->perturbation) & (SFQ_HASH_DIVISOR - 1);
145 static unsigned sfq_hash(struct sfq_sched_data *q, struct sk_buff *skb)
149 switch (skb->protocol) {
150 case htons(ETH_P_IP):
152 const struct iphdr *iph;
155 if (!pskb_network_may_pull(skb, sizeof(*iph)))
158 h = (__force u32)iph->daddr;
159 h2 = (__force u32)iph->saddr ^ iph->protocol;
160 if (iph->frag_off & htons(IP_MF|IP_OFFSET))
162 poff = proto_ports_offset(iph->protocol);
164 pskb_network_may_pull(skb, iph->ihl * 4 + 4 + poff)) {
166 h2 ^= *(u32*)((void *)iph + iph->ihl * 4 + poff);
170 case htons(ETH_P_IPV6):
175 if (!pskb_network_may_pull(skb, sizeof(*iph)))
178 h = (__force u32)iph->daddr.s6_addr32[3];
179 h2 = (__force u32)iph->saddr.s6_addr32[3] ^ iph->nexthdr;
180 poff = proto_ports_offset(iph->nexthdr);
182 pskb_network_may_pull(skb, sizeof(*iph) + 4 + poff)) {
184 h2 ^= *(u32*)((void *)iph + sizeof(*iph) + poff);
190 h = (unsigned long)skb_dst(skb) ^ (__force u32)skb->protocol;
191 h2 = (unsigned long)skb->sk;
194 return sfq_fold_hash(q, h, h2);
197 static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch,
200 struct sfq_sched_data *q = qdisc_priv(sch);
201 struct tcf_result res;
204 if (TC_H_MAJ(skb->priority) == sch->handle &&
205 TC_H_MIN(skb->priority) > 0 &&
206 TC_H_MIN(skb->priority) <= SFQ_HASH_DIVISOR)
207 return TC_H_MIN(skb->priority);
210 return sfq_hash(q, skb) + 1;
212 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
213 result = tc_classify(skb, q->filter_list, &res);
215 #ifdef CONFIG_NET_CLS_ACT
219 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
224 if (TC_H_MIN(res.classid) <= SFQ_HASH_DIVISOR)
225 return TC_H_MIN(res.classid);
231 * x : slot number [0 .. SFQ_SLOTS - 1]
233 static inline void sfq_link(struct sfq_sched_data *q, sfq_index x)
236 int qlen = q->slots[x].qlen;
238 p = qlen + SFQ_SLOTS;
239 n = q->dep[qlen].next;
241 q->slots[x].dep.next = n;
242 q->slots[x].dep.prev = p;
244 q->dep[qlen].next = x; /* sfq_dep_head(q, p)->next = x */
245 sfq_dep_head(q, n)->prev = x;
248 #define sfq_unlink(q, x, n, p) \
249 n = q->slots[x].dep.next; \
250 p = q->slots[x].dep.prev; \
251 sfq_dep_head(q, p)->next = n; \
252 sfq_dep_head(q, n)->prev = p
255 static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x)
260 sfq_unlink(q, x, n, p);
262 d = q->slots[x].qlen--;
263 if (n == p && q->cur_depth == d)
268 static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x)
273 sfq_unlink(q, x, n, p);
275 d = ++q->slots[x].qlen;
276 if (q->cur_depth < d)
281 /* helper functions : might be changed when/if skb use a standard list_head */
283 /* remove one skb from tail of slot queue */
284 static inline struct sk_buff *slot_dequeue_tail(struct sfq_slot *slot)
286 struct sk_buff *skb = slot->skblist_prev;
288 slot->skblist_prev = skb->prev;
289 skb->prev->next = (struct sk_buff *)slot;
290 skb->next = skb->prev = NULL;
294 /* remove one skb from head of slot queue */
295 static inline struct sk_buff *slot_dequeue_head(struct sfq_slot *slot)
297 struct sk_buff *skb = slot->skblist_next;
299 slot->skblist_next = skb->next;
300 skb->next->prev = (struct sk_buff *)slot;
301 skb->next = skb->prev = NULL;
305 static inline void slot_queue_init(struct sfq_slot *slot)
307 slot->skblist_prev = slot->skblist_next = (struct sk_buff *)slot;
310 /* add skb to slot queue (tail add) */
311 static inline void slot_queue_add(struct sfq_slot *slot, struct sk_buff *skb)
313 skb->prev = slot->skblist_prev;
314 skb->next = (struct sk_buff *)slot;
315 slot->skblist_prev->next = skb;
316 slot->skblist_prev = skb;
319 #define slot_queue_walk(slot, skb) \
320 for (skb = slot->skblist_next; \
321 skb != (struct sk_buff *)slot; \
324 static unsigned int sfq_drop(struct Qdisc *sch)
326 struct sfq_sched_data *q = qdisc_priv(sch);
327 sfq_index x, d = q->cur_depth;
330 struct sfq_slot *slot;
332 /* Queue is full! Find the longest slot and drop tail packet from it */
337 skb = slot_dequeue_tail(slot);
338 len = qdisc_pkt_len(skb);
343 sch->qstats.backlog -= len;
348 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
351 q->tail->next = slot->next;
352 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
360 sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
362 struct sfq_sched_data *q = qdisc_priv(sch);
365 struct sfq_slot *slot;
366 int uninitialized_var(ret);
368 hash = sfq_classify(skb, sch, &ret);
370 if (ret & __NET_XMIT_BYPASS)
379 if (x == SFQ_EMPTY_SLOT) {
380 x = q->dep[0].next; /* get a free slot */
386 /* If selected queue has length q->limit, do simple tail drop,
387 * i.e. drop _this_ packet.
389 if (slot->qlen >= q->limit)
390 return qdisc_drop(skb, sch);
392 sch->qstats.backlog += qdisc_pkt_len(skb);
393 slot_queue_add(slot, skb);
395 if (slot->qlen == 1) { /* The flow is new */
396 if (q->tail == NULL) { /* It is the first flow */
399 slot->next = q->tail->next;
403 slot->allot = q->scaled_quantum;
405 if (++sch->q.qlen <= q->limit) {
406 qdisc_bstats_update(sch, skb);
407 return NET_XMIT_SUCCESS;
414 static struct sk_buff *
415 sfq_peek(struct Qdisc *sch)
417 struct sfq_sched_data *q = qdisc_priv(sch);
419 /* No active slots */
423 return q->slots[q->tail->next].skblist_next;
426 static struct sk_buff *
427 sfq_dequeue(struct Qdisc *sch)
429 struct sfq_sched_data *q = qdisc_priv(sch);
432 struct sfq_slot *slot;
434 /* No active slots */
441 if (slot->allot <= 0) {
443 slot->allot += q->scaled_quantum;
446 skb = slot_dequeue_head(slot);
449 sch->qstats.backlog -= qdisc_pkt_len(skb);
451 /* Is the slot empty? */
452 if (slot->qlen == 0) {
453 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
456 q->tail = NULL; /* no more active slots */
459 q->tail->next = next_a;
461 slot->allot -= SFQ_ALLOT_SIZE(qdisc_pkt_len(skb));
467 sfq_reset(struct Qdisc *sch)
471 while ((skb = sfq_dequeue(sch)) != NULL)
475 static void sfq_perturbation(unsigned long arg)
477 struct Qdisc *sch = (struct Qdisc *)arg;
478 struct sfq_sched_data *q = qdisc_priv(sch);
480 q->perturbation = net_random();
482 if (q->perturb_period)
483 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
486 static int sfq_change(struct Qdisc *sch, struct nlattr *opt)
488 struct sfq_sched_data *q = qdisc_priv(sch);
489 struct tc_sfq_qopt *ctl = nla_data(opt);
492 if (opt->nla_len < nla_attr_size(sizeof(*ctl)))
496 q->quantum = ctl->quantum ? : psched_mtu(qdisc_dev(sch));
497 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
498 q->perturb_period = ctl->perturb_period * HZ;
500 q->limit = min_t(u32, ctl->limit, SFQ_DEPTH - 1);
503 while (sch->q.qlen > q->limit)
505 qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen);
507 del_timer(&q->perturb_timer);
508 if (q->perturb_period) {
509 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
510 q->perturbation = net_random();
512 sch_tree_unlock(sch);
516 static int sfq_init(struct Qdisc *sch, struct nlattr *opt)
518 struct sfq_sched_data *q = qdisc_priv(sch);
521 q->perturb_timer.function = sfq_perturbation;
522 q->perturb_timer.data = (unsigned long)sch;
523 init_timer_deferrable(&q->perturb_timer);
525 for (i = 0; i < SFQ_HASH_DIVISOR; i++)
526 q->ht[i] = SFQ_EMPTY_SLOT;
528 for (i = 0; i < SFQ_DEPTH; i++) {
529 q->dep[i].next = i + SFQ_SLOTS;
530 q->dep[i].prev = i + SFQ_SLOTS;
533 q->limit = SFQ_DEPTH - 1;
537 q->quantum = psched_mtu(qdisc_dev(sch));
538 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
539 q->perturb_period = 0;
540 q->perturbation = net_random();
542 int err = sfq_change(sch, opt);
547 for (i = 0; i < SFQ_SLOTS; i++) {
548 slot_queue_init(&q->slots[i]);
554 static void sfq_destroy(struct Qdisc *sch)
556 struct sfq_sched_data *q = qdisc_priv(sch);
558 tcf_destroy_chain(&q->filter_list);
559 q->perturb_period = 0;
560 del_timer_sync(&q->perturb_timer);
563 static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb)
565 struct sfq_sched_data *q = qdisc_priv(sch);
566 unsigned char *b = skb_tail_pointer(skb);
567 struct tc_sfq_qopt opt;
569 opt.quantum = q->quantum;
570 opt.perturb_period = q->perturb_period / HZ;
572 opt.limit = q->limit;
573 opt.divisor = SFQ_HASH_DIVISOR;
574 opt.flows = q->limit;
576 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
585 static struct Qdisc *sfq_leaf(struct Qdisc *sch, unsigned long arg)
590 static unsigned long sfq_get(struct Qdisc *sch, u32 classid)
595 static unsigned long sfq_bind(struct Qdisc *sch, unsigned long parent,
601 static void sfq_put(struct Qdisc *q, unsigned long cl)
605 static struct tcf_proto **sfq_find_tcf(struct Qdisc *sch, unsigned long cl)
607 struct sfq_sched_data *q = qdisc_priv(sch);
611 return &q->filter_list;
614 static int sfq_dump_class(struct Qdisc *sch, unsigned long cl,
615 struct sk_buff *skb, struct tcmsg *tcm)
617 tcm->tcm_handle |= TC_H_MIN(cl);
621 static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
624 struct sfq_sched_data *q = qdisc_priv(sch);
625 sfq_index idx = q->ht[cl - 1];
626 struct gnet_stats_queue qs = { 0 };
627 struct tc_sfq_xstats xstats = { 0 };
630 if (idx != SFQ_EMPTY_SLOT) {
631 const struct sfq_slot *slot = &q->slots[idx];
633 xstats.allot = slot->allot << SFQ_ALLOT_SHIFT;
634 qs.qlen = slot->qlen;
635 slot_queue_walk(slot, skb)
636 qs.backlog += qdisc_pkt_len(skb);
638 if (gnet_stats_copy_queue(d, &qs) < 0)
640 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
643 static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
645 struct sfq_sched_data *q = qdisc_priv(sch);
651 for (i = 0; i < SFQ_HASH_DIVISOR; i++) {
652 if (q->ht[i] == SFQ_EMPTY_SLOT ||
653 arg->count < arg->skip) {
657 if (arg->fn(sch, i + 1, arg) < 0) {
665 static const struct Qdisc_class_ops sfq_class_ops = {
669 .tcf_chain = sfq_find_tcf,
670 .bind_tcf = sfq_bind,
671 .unbind_tcf = sfq_put,
672 .dump = sfq_dump_class,
673 .dump_stats = sfq_dump_class_stats,
677 static struct Qdisc_ops sfq_qdisc_ops __read_mostly = {
678 .cl_ops = &sfq_class_ops,
680 .priv_size = sizeof(struct sfq_sched_data),
681 .enqueue = sfq_enqueue,
682 .dequeue = sfq_dequeue,
687 .destroy = sfq_destroy,
690 .owner = THIS_MODULE,
693 static int __init sfq_module_init(void)
695 return register_qdisc(&sfq_qdisc_ops);
697 static void __exit sfq_module_exit(void)
699 unregister_qdisc(&sfq_qdisc_ops);
701 module_init(sfq_module_init)
702 module_exit(sfq_module_exit)
703 MODULE_LICENSE("GPL");