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
101 struct sk_buff *skblist_next;
102 struct sk_buff *skblist_prev;
103 sfq_index qlen; /* number of skbs in skblist */
104 sfq_index next; /* next slot in sfq chain */
105 struct sfq_head dep; /* anchor in dep[] chains */
106 unsigned short hash; /* hash value (index in ht[]) */
107 short allot; /* credit for this slot */
110 struct sfq_sched_data {
113 unsigned int quantum; /* Allotment per round: MUST BE >= MTU */
117 struct tcf_proto *filter_list;
118 struct timer_list perturb_timer;
120 sfq_index cur_depth; /* depth of longest slot */
121 unsigned short scaled_quantum; /* SFQ_ALLOT_SIZE(quantum) */
122 struct sfq_slot *tail; /* current slot in round */
123 sfq_index ht[SFQ_HASH_DIVISOR]; /* Hash table */
124 struct sfq_slot slots[SFQ_SLOTS];
125 struct sfq_head dep[SFQ_DEPTH]; /* Linked list of slots, indexed by depth */
129 * sfq_head are either in a sfq_slot or in dep[] array
131 static inline struct sfq_head *sfq_dep_head(struct sfq_sched_data *q, sfq_index val)
134 return &q->slots[val].dep;
135 return &q->dep[val - SFQ_SLOTS];
138 static unsigned int sfq_fold_hash(struct sfq_sched_data *q, u32 h, u32 h1)
140 return jhash_2words(h, h1, q->perturbation) & (SFQ_HASH_DIVISOR - 1);
143 static unsigned int sfq_hash(struct sfq_sched_data *q, struct sk_buff *skb)
147 switch (skb->protocol) {
148 case htons(ETH_P_IP):
150 const struct iphdr *iph;
153 if (!pskb_network_may_pull(skb, sizeof(*iph)))
156 h = (__force u32)iph->daddr;
157 h2 = (__force u32)iph->saddr ^ iph->protocol;
158 if (iph->frag_off & htons(IP_MF | IP_OFFSET))
160 poff = proto_ports_offset(iph->protocol);
162 pskb_network_may_pull(skb, iph->ihl * 4 + 4 + poff)) {
164 h2 ^= *(u32 *)((void *)iph + iph->ihl * 4 + poff);
168 case htons(ETH_P_IPV6):
173 if (!pskb_network_may_pull(skb, sizeof(*iph)))
176 h = (__force u32)iph->daddr.s6_addr32[3];
177 h2 = (__force u32)iph->saddr.s6_addr32[3] ^ iph->nexthdr;
178 poff = proto_ports_offset(iph->nexthdr);
180 pskb_network_may_pull(skb, sizeof(*iph) + 4 + poff)) {
182 h2 ^= *(u32 *)((void *)iph + sizeof(*iph) + poff);
188 h = (unsigned long)skb_dst(skb) ^ (__force u32)skb->protocol;
189 h2 = (unsigned long)skb->sk;
192 return sfq_fold_hash(q, h, h2);
195 static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch,
198 struct sfq_sched_data *q = qdisc_priv(sch);
199 struct tcf_result res;
202 if (TC_H_MAJ(skb->priority) == sch->handle &&
203 TC_H_MIN(skb->priority) > 0 &&
204 TC_H_MIN(skb->priority) <= SFQ_HASH_DIVISOR)
205 return TC_H_MIN(skb->priority);
208 return sfq_hash(q, skb) + 1;
210 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
211 result = tc_classify(skb, q->filter_list, &res);
213 #ifdef CONFIG_NET_CLS_ACT
217 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
222 if (TC_H_MIN(res.classid) <= SFQ_HASH_DIVISOR)
223 return TC_H_MIN(res.classid);
229 * x : slot number [0 .. SFQ_SLOTS - 1]
231 static inline void sfq_link(struct sfq_sched_data *q, sfq_index x)
234 int qlen = q->slots[x].qlen;
236 p = qlen + SFQ_SLOTS;
237 n = q->dep[qlen].next;
239 q->slots[x].dep.next = n;
240 q->slots[x].dep.prev = p;
242 q->dep[qlen].next = x; /* sfq_dep_head(q, p)->next = x */
243 sfq_dep_head(q, n)->prev = x;
246 #define sfq_unlink(q, x, n, p) \
247 n = q->slots[x].dep.next; \
248 p = q->slots[x].dep.prev; \
249 sfq_dep_head(q, p)->next = n; \
250 sfq_dep_head(q, n)->prev = p
253 static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x)
258 sfq_unlink(q, x, n, p);
260 d = q->slots[x].qlen--;
261 if (n == p && q->cur_depth == d)
266 static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x)
271 sfq_unlink(q, x, n, p);
273 d = ++q->slots[x].qlen;
274 if (q->cur_depth < d)
279 /* helper functions : might be changed when/if skb use a standard list_head */
281 /* remove one skb from tail of slot queue */
282 static inline struct sk_buff *slot_dequeue_tail(struct sfq_slot *slot)
284 struct sk_buff *skb = slot->skblist_prev;
286 slot->skblist_prev = skb->prev;
287 skb->prev->next = (struct sk_buff *)slot;
288 skb->next = skb->prev = NULL;
292 /* remove one skb from head of slot queue */
293 static inline struct sk_buff *slot_dequeue_head(struct sfq_slot *slot)
295 struct sk_buff *skb = slot->skblist_next;
297 slot->skblist_next = skb->next;
298 skb->next->prev = (struct sk_buff *)slot;
299 skb->next = skb->prev = NULL;
303 static inline void slot_queue_init(struct sfq_slot *slot)
305 slot->skblist_prev = slot->skblist_next = (struct sk_buff *)slot;
308 /* add skb to slot queue (tail add) */
309 static inline void slot_queue_add(struct sfq_slot *slot, struct sk_buff *skb)
311 skb->prev = slot->skblist_prev;
312 skb->next = (struct sk_buff *)slot;
313 slot->skblist_prev->next = skb;
314 slot->skblist_prev = skb;
317 #define slot_queue_walk(slot, skb) \
318 for (skb = slot->skblist_next; \
319 skb != (struct sk_buff *)slot; \
322 static unsigned int sfq_drop(struct Qdisc *sch)
324 struct sfq_sched_data *q = qdisc_priv(sch);
325 sfq_index x, d = q->cur_depth;
328 struct sfq_slot *slot;
330 /* Queue is full! Find the longest slot and drop tail packet from it */
335 skb = slot_dequeue_tail(slot);
336 len = qdisc_pkt_len(skb);
341 sch->qstats.backlog -= len;
346 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
349 q->tail->next = slot->next;
350 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
358 sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
360 struct sfq_sched_data *q = qdisc_priv(sch);
363 struct sfq_slot *slot;
364 int uninitialized_var(ret);
366 hash = sfq_classify(skb, sch, &ret);
368 if (ret & __NET_XMIT_BYPASS)
377 if (x == SFQ_EMPTY_SLOT) {
378 x = q->dep[0].next; /* get a free slot */
384 /* If selected queue has length q->limit, do simple tail drop,
385 * i.e. drop _this_ packet.
387 if (slot->qlen >= q->limit)
388 return qdisc_drop(skb, sch);
390 sch->qstats.backlog += qdisc_pkt_len(skb);
391 slot_queue_add(slot, skb);
393 if (slot->qlen == 1) { /* The flow is new */
394 if (q->tail == NULL) { /* It is the first flow */
397 slot->next = q->tail->next;
401 slot->allot = q->scaled_quantum;
403 if (++sch->q.qlen <= q->limit) {
404 qdisc_bstats_update(sch, skb);
405 return NET_XMIT_SUCCESS;
412 static struct sk_buff *
413 sfq_peek(struct Qdisc *sch)
415 struct sfq_sched_data *q = qdisc_priv(sch);
417 /* No active slots */
421 return q->slots[q->tail->next].skblist_next;
424 static struct sk_buff *
425 sfq_dequeue(struct Qdisc *sch)
427 struct sfq_sched_data *q = qdisc_priv(sch);
430 struct sfq_slot *slot;
432 /* No active slots */
439 if (slot->allot <= 0) {
441 slot->allot += q->scaled_quantum;
444 skb = slot_dequeue_head(slot);
447 sch->qstats.backlog -= qdisc_pkt_len(skb);
449 /* Is the slot empty? */
450 if (slot->qlen == 0) {
451 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
454 q->tail = NULL; /* no more active slots */
457 q->tail->next = next_a;
459 slot->allot -= SFQ_ALLOT_SIZE(qdisc_pkt_len(skb));
465 sfq_reset(struct Qdisc *sch)
469 while ((skb = sfq_dequeue(sch)) != NULL)
473 static void sfq_perturbation(unsigned long arg)
475 struct Qdisc *sch = (struct Qdisc *)arg;
476 struct sfq_sched_data *q = qdisc_priv(sch);
478 q->perturbation = net_random();
480 if (q->perturb_period)
481 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
484 static int sfq_change(struct Qdisc *sch, struct nlattr *opt)
486 struct sfq_sched_data *q = qdisc_priv(sch);
487 struct tc_sfq_qopt *ctl = nla_data(opt);
490 if (opt->nla_len < nla_attr_size(sizeof(*ctl)))
494 q->quantum = ctl->quantum ? : psched_mtu(qdisc_dev(sch));
495 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
496 q->perturb_period = ctl->perturb_period * HZ;
498 q->limit = min_t(u32, ctl->limit, SFQ_DEPTH - 1);
501 while (sch->q.qlen > q->limit)
503 qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen);
505 del_timer(&q->perturb_timer);
506 if (q->perturb_period) {
507 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
508 q->perturbation = net_random();
510 sch_tree_unlock(sch);
514 static int sfq_init(struct Qdisc *sch, struct nlattr *opt)
516 struct sfq_sched_data *q = qdisc_priv(sch);
519 q->perturb_timer.function = sfq_perturbation;
520 q->perturb_timer.data = (unsigned long)sch;
521 init_timer_deferrable(&q->perturb_timer);
523 for (i = 0; i < SFQ_HASH_DIVISOR; i++)
524 q->ht[i] = SFQ_EMPTY_SLOT;
526 for (i = 0; i < SFQ_DEPTH; i++) {
527 q->dep[i].next = i + SFQ_SLOTS;
528 q->dep[i].prev = i + SFQ_SLOTS;
531 q->limit = SFQ_DEPTH - 1;
535 q->quantum = psched_mtu(qdisc_dev(sch));
536 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
537 q->perturb_period = 0;
538 q->perturbation = net_random();
540 int err = sfq_change(sch, opt);
545 for (i = 0; i < SFQ_SLOTS; i++) {
546 slot_queue_init(&q->slots[i]);
552 static void sfq_destroy(struct Qdisc *sch)
554 struct sfq_sched_data *q = qdisc_priv(sch);
556 tcf_destroy_chain(&q->filter_list);
557 q->perturb_period = 0;
558 del_timer_sync(&q->perturb_timer);
561 static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb)
563 struct sfq_sched_data *q = qdisc_priv(sch);
564 unsigned char *b = skb_tail_pointer(skb);
565 struct tc_sfq_qopt opt;
567 opt.quantum = q->quantum;
568 opt.perturb_period = q->perturb_period / HZ;
570 opt.limit = q->limit;
571 opt.divisor = SFQ_HASH_DIVISOR;
572 opt.flows = q->limit;
574 NLA_PUT(skb, TCA_OPTIONS, sizeof(opt), &opt);
583 static struct Qdisc *sfq_leaf(struct Qdisc *sch, unsigned long arg)
588 static unsigned long sfq_get(struct Qdisc *sch, u32 classid)
593 static unsigned long sfq_bind(struct Qdisc *sch, unsigned long parent,
599 static void sfq_put(struct Qdisc *q, unsigned long cl)
603 static struct tcf_proto **sfq_find_tcf(struct Qdisc *sch, unsigned long cl)
605 struct sfq_sched_data *q = qdisc_priv(sch);
609 return &q->filter_list;
612 static int sfq_dump_class(struct Qdisc *sch, unsigned long cl,
613 struct sk_buff *skb, struct tcmsg *tcm)
615 tcm->tcm_handle |= TC_H_MIN(cl);
619 static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
622 struct sfq_sched_data *q = qdisc_priv(sch);
623 sfq_index idx = q->ht[cl - 1];
624 struct gnet_stats_queue qs = { 0 };
625 struct tc_sfq_xstats xstats = { 0 };
628 if (idx != SFQ_EMPTY_SLOT) {
629 const struct sfq_slot *slot = &q->slots[idx];
631 xstats.allot = slot->allot << SFQ_ALLOT_SHIFT;
632 qs.qlen = slot->qlen;
633 slot_queue_walk(slot, skb)
634 qs.backlog += qdisc_pkt_len(skb);
636 if (gnet_stats_copy_queue(d, &qs) < 0)
638 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
641 static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
643 struct sfq_sched_data *q = qdisc_priv(sch);
649 for (i = 0; i < SFQ_HASH_DIVISOR; i++) {
650 if (q->ht[i] == SFQ_EMPTY_SLOT ||
651 arg->count < arg->skip) {
655 if (arg->fn(sch, i + 1, arg) < 0) {
663 static const struct Qdisc_class_ops sfq_class_ops = {
667 .tcf_chain = sfq_find_tcf,
668 .bind_tcf = sfq_bind,
669 .unbind_tcf = sfq_put,
670 .dump = sfq_dump_class,
671 .dump_stats = sfq_dump_class_stats,
675 static struct Qdisc_ops sfq_qdisc_ops __read_mostly = {
676 .cl_ops = &sfq_class_ops,
678 .priv_size = sizeof(struct sfq_sched_data),
679 .enqueue = sfq_enqueue,
680 .dequeue = sfq_dequeue,
685 .destroy = sfq_destroy,
688 .owner = THIS_MODULE,
691 static int __init sfq_module_init(void)
693 return register_qdisc(&sfq_qdisc_ops);
695 static void __exit sfq_module_exit(void)
697 unregister_qdisc(&sfq_qdisc_ops);
699 module_init(sfq_module_init)
700 module_exit(sfq_module_exit)
701 MODULE_LICENSE("GPL");