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/skbuff.h>
21 #include <linux/jhash.h>
22 #include <linux/slab.h>
23 #include <linux/vmalloc.h>
24 #include <net/netlink.h>
25 #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 :
70 - maximal queue length per flow to 127 packets.
73 - number of hash buckets to 65536.
75 It is easy to increase these values, but not in flight. */
77 #define SFQ_MAX_DEPTH 127 /* max number of packets per flow */
78 #define SFQ_DEFAULT_FLOWS 128
79 #define SFQ_MAX_FLOWS (0x10000 - SFQ_MAX_DEPTH - 1) /* max number of flows */
80 #define SFQ_EMPTY_SLOT 0xffff
81 #define SFQ_DEFAULT_HASH_DIVISOR 1024
83 /* We use 16 bits to store allot, and want to handle packets up to 64K
84 * Scale allot by 8 (1<<3) so that no overflow occurs.
86 #define SFQ_ALLOT_SHIFT 3
87 #define SFQ_ALLOT_SIZE(X) DIV_ROUND_UP(X, 1 << SFQ_ALLOT_SHIFT)
89 /* This type should contain at least SFQ_MAX_DEPTH + 1 + SFQ_MAX_FLOWS values */
90 typedef u16 sfq_index;
93 * We dont use pointers to save space.
94 * Small indexes [0 ... SFQ_MAX_FLOWS - 1] are 'pointers' to slots[] array
95 * while following values [SFQ_MAX_FLOWS ... SFQ_MAX_FLOWS + SFQ_MAX_DEPTH]
96 * are 'pointers' to dep[] array
104 struct sk_buff *skblist_next;
105 struct sk_buff *skblist_prev;
106 sfq_index qlen; /* number of skbs in skblist */
107 sfq_index next; /* next slot in sfq RR chain */
108 struct sfq_head dep; /* anchor in dep[] chains */
109 unsigned short hash; /* hash value (index in ht[]) */
110 short allot; /* credit for this slot */
112 unsigned int backlog;
113 struct red_vars vars;
116 struct sfq_sched_data {
117 /* frequently used fields */
118 int limit; /* limit of total number of packets in this qdisc */
119 unsigned int divisor; /* number of slots in hash table */
121 u8 maxdepth; /* limit of packets per flow */
124 u8 cur_depth; /* depth of longest slot */
126 unsigned short scaled_quantum; /* SFQ_ALLOT_SIZE(quantum) */
127 struct tcf_proto __rcu *filter_list;
128 sfq_index *ht; /* Hash table ('divisor' slots) */
129 struct sfq_slot *slots; /* Flows table ('maxflows' entries) */
131 struct red_parms *red_parms;
132 struct tc_sfqred_stats stats;
133 struct sfq_slot *tail; /* current slot in round */
135 struct sfq_head dep[SFQ_MAX_DEPTH + 1];
136 /* Linked lists of slots, indexed by depth
137 * dep[0] : list of unused flows
138 * dep[1] : list of flows with 1 packet
139 * dep[X] : list of flows with X packets
142 unsigned int maxflows; /* number of flows in flows array */
144 unsigned int quantum; /* Allotment per round: MUST BE >= MTU */
145 struct timer_list perturb_timer;
149 * sfq_head are either in a sfq_slot or in dep[] array
151 static inline struct sfq_head *sfq_dep_head(struct sfq_sched_data *q, sfq_index val)
153 if (val < SFQ_MAX_FLOWS)
154 return &q->slots[val].dep;
155 return &q->dep[val - SFQ_MAX_FLOWS];
158 static unsigned int sfq_hash(const struct sfq_sched_data *q,
159 const struct sk_buff *skb)
161 return skb_get_hash_perturb(skb, q->perturbation) & (q->divisor - 1);
164 static unsigned int sfq_classify(struct sk_buff *skb, struct Qdisc *sch,
167 struct sfq_sched_data *q = qdisc_priv(sch);
168 struct tcf_result res;
169 struct tcf_proto *fl;
172 if (TC_H_MAJ(skb->priority) == sch->handle &&
173 TC_H_MIN(skb->priority) > 0 &&
174 TC_H_MIN(skb->priority) <= q->divisor)
175 return TC_H_MIN(skb->priority);
177 fl = rcu_dereference_bh(q->filter_list);
179 return sfq_hash(q, skb) + 1;
181 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
182 result = tc_classify(skb, fl, &res, false);
184 #ifdef CONFIG_NET_CLS_ACT
188 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
193 if (TC_H_MIN(res.classid) <= q->divisor)
194 return TC_H_MIN(res.classid);
200 * x : slot number [0 .. SFQ_MAX_FLOWS - 1]
202 static inline void sfq_link(struct sfq_sched_data *q, sfq_index x)
205 struct sfq_slot *slot = &q->slots[x];
206 int qlen = slot->qlen;
208 p = qlen + SFQ_MAX_FLOWS;
209 n = q->dep[qlen].next;
214 q->dep[qlen].next = x; /* sfq_dep_head(q, p)->next = x */
215 sfq_dep_head(q, n)->prev = x;
218 #define sfq_unlink(q, x, n, p) \
220 n = q->slots[x].dep.next; \
221 p = q->slots[x].dep.prev; \
222 sfq_dep_head(q, p)->next = n; \
223 sfq_dep_head(q, n)->prev = p; \
227 static inline void sfq_dec(struct sfq_sched_data *q, sfq_index x)
232 sfq_unlink(q, x, n, p);
234 d = q->slots[x].qlen--;
235 if (n == p && q->cur_depth == d)
240 static inline void sfq_inc(struct sfq_sched_data *q, sfq_index x)
245 sfq_unlink(q, x, n, p);
247 d = ++q->slots[x].qlen;
248 if (q->cur_depth < d)
253 /* helper functions : might be changed when/if skb use a standard list_head */
255 /* remove one skb from tail of slot queue */
256 static inline struct sk_buff *slot_dequeue_tail(struct sfq_slot *slot)
258 struct sk_buff *skb = slot->skblist_prev;
260 slot->skblist_prev = skb->prev;
261 skb->prev->next = (struct sk_buff *)slot;
262 skb->next = skb->prev = NULL;
266 /* remove one skb from head of slot queue */
267 static inline struct sk_buff *slot_dequeue_head(struct sfq_slot *slot)
269 struct sk_buff *skb = slot->skblist_next;
271 slot->skblist_next = skb->next;
272 skb->next->prev = (struct sk_buff *)slot;
273 skb->next = skb->prev = NULL;
277 static inline void slot_queue_init(struct sfq_slot *slot)
279 memset(slot, 0, sizeof(*slot));
280 slot->skblist_prev = slot->skblist_next = (struct sk_buff *)slot;
283 /* add skb to slot queue (tail add) */
284 static inline void slot_queue_add(struct sfq_slot *slot, struct sk_buff *skb)
286 skb->prev = slot->skblist_prev;
287 skb->next = (struct sk_buff *)slot;
288 slot->skblist_prev->next = skb;
289 slot->skblist_prev = skb;
292 static unsigned int sfq_drop(struct Qdisc *sch)
294 struct sfq_sched_data *q = qdisc_priv(sch);
295 sfq_index x, d = q->cur_depth;
298 struct sfq_slot *slot;
300 /* Queue is full! Find the longest slot and drop tail packet from it */
305 skb = q->headdrop ? slot_dequeue_head(slot) : slot_dequeue_tail(slot);
306 len = qdisc_pkt_len(skb);
307 slot->backlog -= len;
310 qdisc_qstats_drop(sch);
311 qdisc_qstats_backlog_dec(sch, skb);
317 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
320 q->tail->next = slot->next;
321 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
328 /* Is ECN parameter configured */
329 static int sfq_prob_mark(const struct sfq_sched_data *q)
331 return q->flags & TC_RED_ECN;
334 /* Should packets over max threshold just be marked */
335 static int sfq_hard_mark(const struct sfq_sched_data *q)
337 return (q->flags & (TC_RED_ECN | TC_RED_HARDDROP)) == TC_RED_ECN;
340 static int sfq_headdrop(const struct sfq_sched_data *q)
346 sfq_enqueue(struct sk_buff *skb, struct Qdisc *sch)
348 struct sfq_sched_data *q = qdisc_priv(sch);
351 struct sfq_slot *slot;
352 int uninitialized_var(ret);
353 struct sk_buff *head;
356 hash = sfq_classify(skb, sch, &ret);
358 if (ret & __NET_XMIT_BYPASS)
359 qdisc_qstats_drop(sch);
367 if (x == SFQ_EMPTY_SLOT) {
368 x = q->dep[0].next; /* get a free slot */
369 if (x >= SFQ_MAX_FLOWS)
370 return qdisc_drop(skb, sch);
374 slot->backlog = 0; /* should already be 0 anyway... */
375 red_set_vars(&slot->vars);
379 slot->vars.qavg = red_calc_qavg_no_idle_time(q->red_parms,
382 switch (red_action(q->red_parms,
389 qdisc_qstats_overlimit(sch);
390 if (sfq_prob_mark(q)) {
391 /* We know we have at least one packet in queue */
392 if (sfq_headdrop(q) &&
393 INET_ECN_set_ce(slot->skblist_next)) {
394 q->stats.prob_mark_head++;
397 if (INET_ECN_set_ce(skb)) {
398 q->stats.prob_mark++;
402 q->stats.prob_drop++;
403 goto congestion_drop;
406 qdisc_qstats_overlimit(sch);
407 if (sfq_hard_mark(q)) {
408 /* We know we have at least one packet in queue */
409 if (sfq_headdrop(q) &&
410 INET_ECN_set_ce(slot->skblist_next)) {
411 q->stats.forced_mark_head++;
414 if (INET_ECN_set_ce(skb)) {
415 q->stats.forced_mark++;
419 q->stats.forced_drop++;
420 goto congestion_drop;
424 if (slot->qlen >= q->maxdepth) {
426 if (!sfq_headdrop(q))
427 return qdisc_drop(skb, sch);
429 /* We know we have at least one packet in queue */
430 head = slot_dequeue_head(slot);
431 delta = qdisc_pkt_len(head) - qdisc_pkt_len(skb);
432 sch->qstats.backlog -= delta;
433 slot->backlog -= delta;
434 qdisc_drop(head, sch);
436 slot_queue_add(slot, skb);
441 qdisc_qstats_backlog_inc(sch, skb);
442 slot->backlog += qdisc_pkt_len(skb);
443 slot_queue_add(slot, skb);
445 if (slot->qlen == 1) { /* The flow is new */
446 if (q->tail == NULL) { /* It is the first flow */
449 slot->next = q->tail->next;
452 /* We put this flow at the end of our flow list.
453 * This might sound unfair for a new flow to wait after old ones,
454 * but we could endup servicing new flows only, and freeze old ones.
457 /* We could use a bigger initial quantum for new flows */
458 slot->allot = q->scaled_quantum;
460 if (++sch->q.qlen <= q->limit)
461 return NET_XMIT_SUCCESS;
465 /* Return Congestion Notification only if we dropped a packet
468 if (qlen != slot->qlen)
471 /* As we dropped a packet, better let upper stack know this */
472 qdisc_tree_decrease_qlen(sch, 1);
473 return NET_XMIT_SUCCESS;
476 static struct sk_buff *
477 sfq_dequeue(struct Qdisc *sch)
479 struct sfq_sched_data *q = qdisc_priv(sch);
482 struct sfq_slot *slot;
484 /* No active slots */
491 if (slot->allot <= 0) {
493 slot->allot += q->scaled_quantum;
496 skb = slot_dequeue_head(slot);
498 qdisc_bstats_update(sch, skb);
500 qdisc_qstats_backlog_dec(sch, skb);
501 slot->backlog -= qdisc_pkt_len(skb);
502 /* Is the slot empty? */
503 if (slot->qlen == 0) {
504 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
507 q->tail = NULL; /* no more active slots */
510 q->tail->next = next_a;
512 slot->allot -= SFQ_ALLOT_SIZE(qdisc_pkt_len(skb));
518 sfq_reset(struct Qdisc *sch)
522 while ((skb = sfq_dequeue(sch)) != NULL)
527 * When q->perturbation is changed, we rehash all queued skbs
528 * to avoid OOO (Out Of Order) effects.
529 * We dont use sfq_dequeue()/sfq_enqueue() because we dont want to change
532 static void sfq_rehash(struct Qdisc *sch)
534 struct sfq_sched_data *q = qdisc_priv(sch);
537 struct sfq_slot *slot;
538 struct sk_buff_head list;
541 __skb_queue_head_init(&list);
543 for (i = 0; i < q->maxflows; i++) {
548 skb = slot_dequeue_head(slot);
550 __skb_queue_tail(&list, skb);
553 red_set_vars(&slot->vars);
554 q->ht[slot->hash] = SFQ_EMPTY_SLOT;
558 while ((skb = __skb_dequeue(&list)) != NULL) {
559 unsigned int hash = sfq_hash(q, skb);
560 sfq_index x = q->ht[hash];
563 if (x == SFQ_EMPTY_SLOT) {
564 x = q->dep[0].next; /* get a free slot */
565 if (x >= SFQ_MAX_FLOWS) {
567 qdisc_qstats_backlog_dec(sch, skb);
576 if (slot->qlen >= q->maxdepth)
578 slot_queue_add(slot, skb);
580 slot->vars.qavg = red_calc_qavg(q->red_parms,
583 slot->backlog += qdisc_pkt_len(skb);
585 if (slot->qlen == 1) { /* The flow is new */
586 if (q->tail == NULL) { /* It is the first flow */
589 slot->next = q->tail->next;
593 slot->allot = q->scaled_quantum;
596 sch->q.qlen -= dropped;
597 qdisc_tree_decrease_qlen(sch, dropped);
600 static void sfq_perturbation(unsigned long arg)
602 struct Qdisc *sch = (struct Qdisc *)arg;
603 struct sfq_sched_data *q = qdisc_priv(sch);
604 spinlock_t *root_lock = qdisc_lock(qdisc_root_sleeping(sch));
606 spin_lock(root_lock);
607 q->perturbation = prandom_u32();
608 if (!q->filter_list && q->tail)
610 spin_unlock(root_lock);
612 if (q->perturb_period)
613 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
616 static int sfq_change(struct Qdisc *sch, struct nlattr *opt)
618 struct sfq_sched_data *q = qdisc_priv(sch);
619 struct tc_sfq_qopt *ctl = nla_data(opt);
620 struct tc_sfq_qopt_v1 *ctl_v1 = NULL;
622 struct red_parms *p = NULL;
624 if (opt->nla_len < nla_attr_size(sizeof(*ctl)))
626 if (opt->nla_len >= nla_attr_size(sizeof(*ctl_v1)))
627 ctl_v1 = nla_data(opt);
629 (!is_power_of_2(ctl->divisor) || ctl->divisor > 65536))
631 if (ctl_v1 && ctl_v1->qth_min) {
632 p = kmalloc(sizeof(*p), GFP_KERNEL);
638 q->quantum = ctl->quantum;
639 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
641 q->perturb_period = ctl->perturb_period * HZ;
643 q->maxflows = min_t(u32, ctl->flows, SFQ_MAX_FLOWS);
645 q->divisor = ctl->divisor;
646 q->maxflows = min_t(u32, q->maxflows, q->divisor);
650 q->maxdepth = min_t(u32, ctl_v1->depth, SFQ_MAX_DEPTH);
652 swap(q->red_parms, p);
653 red_set_parms(q->red_parms,
654 ctl_v1->qth_min, ctl_v1->qth_max,
656 ctl_v1->Plog, ctl_v1->Scell_log,
660 q->flags = ctl_v1->flags;
661 q->headdrop = ctl_v1->headdrop;
664 q->limit = min_t(u32, ctl->limit, q->maxdepth * q->maxflows);
665 q->maxflows = min_t(u32, q->maxflows, q->limit);
669 while (sch->q.qlen > q->limit)
671 qdisc_tree_decrease_qlen(sch, qlen - sch->q.qlen);
673 del_timer(&q->perturb_timer);
674 if (q->perturb_period) {
675 mod_timer(&q->perturb_timer, jiffies + q->perturb_period);
676 q->perturbation = prandom_u32();
678 sch_tree_unlock(sch);
683 static void *sfq_alloc(size_t sz)
685 void *ptr = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN);
692 static void sfq_free(void *addr)
697 static void sfq_destroy(struct Qdisc *sch)
699 struct sfq_sched_data *q = qdisc_priv(sch);
701 tcf_destroy_chain(&q->filter_list);
702 q->perturb_period = 0;
703 del_timer_sync(&q->perturb_timer);
709 static int sfq_init(struct Qdisc *sch, struct nlattr *opt)
711 struct sfq_sched_data *q = qdisc_priv(sch);
714 q->perturb_timer.function = sfq_perturbation;
715 q->perturb_timer.data = (unsigned long)sch;
716 init_timer_deferrable(&q->perturb_timer);
718 for (i = 0; i < SFQ_MAX_DEPTH + 1; i++) {
719 q->dep[i].next = i + SFQ_MAX_FLOWS;
720 q->dep[i].prev = i + SFQ_MAX_FLOWS;
723 q->limit = SFQ_MAX_DEPTH;
724 q->maxdepth = SFQ_MAX_DEPTH;
727 q->divisor = SFQ_DEFAULT_HASH_DIVISOR;
728 q->maxflows = SFQ_DEFAULT_FLOWS;
729 q->quantum = psched_mtu(qdisc_dev(sch));
730 q->scaled_quantum = SFQ_ALLOT_SIZE(q->quantum);
731 q->perturb_period = 0;
732 q->perturbation = prandom_u32();
735 int err = sfq_change(sch, opt);
740 q->ht = sfq_alloc(sizeof(q->ht[0]) * q->divisor);
741 q->slots = sfq_alloc(sizeof(q->slots[0]) * q->maxflows);
742 if (!q->ht || !q->slots) {
746 for (i = 0; i < q->divisor; i++)
747 q->ht[i] = SFQ_EMPTY_SLOT;
749 for (i = 0; i < q->maxflows; i++) {
750 slot_queue_init(&q->slots[i]);
754 sch->flags |= TCQ_F_CAN_BYPASS;
756 sch->flags &= ~TCQ_F_CAN_BYPASS;
760 static int sfq_dump(struct Qdisc *sch, struct sk_buff *skb)
762 struct sfq_sched_data *q = qdisc_priv(sch);
763 unsigned char *b = skb_tail_pointer(skb);
764 struct tc_sfq_qopt_v1 opt;
765 struct red_parms *p = q->red_parms;
767 memset(&opt, 0, sizeof(opt));
768 opt.v0.quantum = q->quantum;
769 opt.v0.perturb_period = q->perturb_period / HZ;
770 opt.v0.limit = q->limit;
771 opt.v0.divisor = q->divisor;
772 opt.v0.flows = q->maxflows;
773 opt.depth = q->maxdepth;
774 opt.headdrop = q->headdrop;
777 opt.qth_min = p->qth_min >> p->Wlog;
778 opt.qth_max = p->qth_max >> p->Wlog;
781 opt.Scell_log = p->Scell_log;
782 opt.max_P = p->max_P;
784 memcpy(&opt.stats, &q->stats, sizeof(opt.stats));
785 opt.flags = q->flags;
787 if (nla_put(skb, TCA_OPTIONS, sizeof(opt), &opt))
788 goto nla_put_failure;
797 static struct Qdisc *sfq_leaf(struct Qdisc *sch, unsigned long arg)
802 static unsigned long sfq_get(struct Qdisc *sch, u32 classid)
807 static unsigned long sfq_bind(struct Qdisc *sch, unsigned long parent,
810 /* we cannot bypass queue discipline anymore */
811 sch->flags &= ~TCQ_F_CAN_BYPASS;
815 static void sfq_put(struct Qdisc *q, unsigned long cl)
819 static struct tcf_proto __rcu **sfq_find_tcf(struct Qdisc *sch,
822 struct sfq_sched_data *q = qdisc_priv(sch);
826 return &q->filter_list;
829 static int sfq_dump_class(struct Qdisc *sch, unsigned long cl,
830 struct sk_buff *skb, struct tcmsg *tcm)
832 tcm->tcm_handle |= TC_H_MIN(cl);
836 static int sfq_dump_class_stats(struct Qdisc *sch, unsigned long cl,
839 struct sfq_sched_data *q = qdisc_priv(sch);
840 sfq_index idx = q->ht[cl - 1];
841 struct gnet_stats_queue qs = { 0 };
842 struct tc_sfq_xstats xstats = { 0 };
844 if (idx != SFQ_EMPTY_SLOT) {
845 const struct sfq_slot *slot = &q->slots[idx];
847 xstats.allot = slot->allot << SFQ_ALLOT_SHIFT;
848 qs.qlen = slot->qlen;
849 qs.backlog = slot->backlog;
851 if (gnet_stats_copy_queue(d, NULL, &qs, qs.qlen) < 0)
853 return gnet_stats_copy_app(d, &xstats, sizeof(xstats));
856 static void sfq_walk(struct Qdisc *sch, struct qdisc_walker *arg)
858 struct sfq_sched_data *q = qdisc_priv(sch);
864 for (i = 0; i < q->divisor; i++) {
865 if (q->ht[i] == SFQ_EMPTY_SLOT ||
866 arg->count < arg->skip) {
870 if (arg->fn(sch, i + 1, arg) < 0) {
878 static const struct Qdisc_class_ops sfq_class_ops = {
882 .tcf_chain = sfq_find_tcf,
883 .bind_tcf = sfq_bind,
884 .unbind_tcf = sfq_put,
885 .dump = sfq_dump_class,
886 .dump_stats = sfq_dump_class_stats,
890 static struct Qdisc_ops sfq_qdisc_ops __read_mostly = {
891 .cl_ops = &sfq_class_ops,
893 .priv_size = sizeof(struct sfq_sched_data),
894 .enqueue = sfq_enqueue,
895 .dequeue = sfq_dequeue,
896 .peek = qdisc_peek_dequeued,
900 .destroy = sfq_destroy,
903 .owner = THIS_MODULE,
906 static int __init sfq_module_init(void)
908 return register_qdisc(&sfq_qdisc_ops);
910 static void __exit sfq_module_exit(void)
912 unregister_qdisc(&sfq_qdisc_ops);
914 module_init(sfq_module_init)
915 module_exit(sfq_module_exit)
916 MODULE_LICENSE("GPL");