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 maximal mtu to 2^15-1; 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
81 /* This type should contain at least SFQ_DEPTH + SFQ_SLOTS values */
82 typedef unsigned char sfq_index
;
85 * We dont use pointers to save space.
86 * Small indexes [0 ... SFQ_SLOTS - 1] are 'pointers' to slots[] array
87 * while following values [SFQ_SLOTS ... SFQ_SLOTS + SFQ_DEPTH - 1]
88 * are 'pointers' to dep[] array
97 struct sk_buff
*skblist_next
;
98 struct sk_buff
*skblist_prev
;
99 sfq_index qlen
; /* number of skbs in skblist */
100 sfq_index next
; /* next slot in sfq chain */
101 struct sfq_head dep
; /* anchor in dep[] chains */
102 unsigned short hash
; /* hash value (index in ht[]) */
103 short allot
; /* credit for this slot */
106 struct sfq_sched_data
110 unsigned quantum
; /* Allotment per round: MUST BE >= MTU */
114 struct tcf_proto
*filter_list
;
115 struct timer_list perturb_timer
;
117 sfq_index cur_depth
; /* depth of longest slot */
119 struct sfq_slot
*tail
; /* current slot in round */
120 sfq_index ht
[SFQ_HASH_DIVISOR
]; /* Hash table */
121 struct sfq_slot slots
[SFQ_SLOTS
];
122 struct sfq_head dep
[SFQ_DEPTH
]; /* Linked list of slots, indexed by depth */
126 * sfq_head are either in a sfq_slot or in dep[] array
128 static inline struct sfq_head
*sfq_dep_head(struct sfq_sched_data
*q
, sfq_index val
)
131 return &q
->slots
[val
].dep
;
132 return &q
->dep
[val
- SFQ_SLOTS
];
135 static __inline__
unsigned sfq_fold_hash(struct sfq_sched_data
*q
, u32 h
, u32 h1
)
137 return jhash_2words(h
, h1
, q
->perturbation
) & (SFQ_HASH_DIVISOR
- 1);
140 static unsigned sfq_hash(struct sfq_sched_data
*q
, struct sk_buff
*skb
)
144 switch (skb
->protocol
) {
145 case htons(ETH_P_IP
):
147 const struct iphdr
*iph
;
150 if (!pskb_network_may_pull(skb
, sizeof(*iph
)))
153 h
= (__force u32
)iph
->daddr
;
154 h2
= (__force u32
)iph
->saddr
^ iph
->protocol
;
155 if (iph
->frag_off
& htons(IP_MF
|IP_OFFSET
))
157 poff
= proto_ports_offset(iph
->protocol
);
159 pskb_network_may_pull(skb
, iph
->ihl
* 4 + 4 + poff
)) {
161 h2
^= *(u32
*)((void *)iph
+ iph
->ihl
* 4 + poff
);
165 case htons(ETH_P_IPV6
):
170 if (!pskb_network_may_pull(skb
, sizeof(*iph
)))
173 h
= (__force u32
)iph
->daddr
.s6_addr32
[3];
174 h2
= (__force u32
)iph
->saddr
.s6_addr32
[3] ^ iph
->nexthdr
;
175 poff
= proto_ports_offset(iph
->nexthdr
);
177 pskb_network_may_pull(skb
, sizeof(*iph
) + 4 + poff
)) {
179 h2
^= *(u32
*)((void *)iph
+ sizeof(*iph
) + poff
);
185 h
= (unsigned long)skb_dst(skb
) ^ (__force u32
)skb
->protocol
;
186 h2
= (unsigned long)skb
->sk
;
189 return sfq_fold_hash(q
, h
, h2
);
192 static unsigned int sfq_classify(struct sk_buff
*skb
, struct Qdisc
*sch
,
195 struct sfq_sched_data
*q
= qdisc_priv(sch
);
196 struct tcf_result res
;
199 if (TC_H_MAJ(skb
->priority
) == sch
->handle
&&
200 TC_H_MIN(skb
->priority
) > 0 &&
201 TC_H_MIN(skb
->priority
) <= SFQ_HASH_DIVISOR
)
202 return TC_H_MIN(skb
->priority
);
205 return sfq_hash(q
, skb
) + 1;
207 *qerr
= NET_XMIT_SUCCESS
| __NET_XMIT_BYPASS
;
208 result
= tc_classify(skb
, q
->filter_list
, &res
);
210 #ifdef CONFIG_NET_CLS_ACT
214 *qerr
= NET_XMIT_SUCCESS
| __NET_XMIT_STOLEN
;
219 if (TC_H_MIN(res
.classid
) <= SFQ_HASH_DIVISOR
)
220 return TC_H_MIN(res
.classid
);
226 * x : slot number [0 .. SFQ_SLOTS - 1]
228 static inline void sfq_link(struct sfq_sched_data
*q
, sfq_index x
)
231 int qlen
= q
->slots
[x
].qlen
;
233 p
= qlen
+ SFQ_SLOTS
;
234 n
= q
->dep
[qlen
].next
;
236 q
->slots
[x
].dep
.next
= n
;
237 q
->slots
[x
].dep
.prev
= p
;
239 q
->dep
[qlen
].next
= x
; /* sfq_dep_head(q, p)->next = x */
240 sfq_dep_head(q
, n
)->prev
= x
;
243 #define sfq_unlink(q, x, n, p) \
244 n = q->slots[x].dep.next; \
245 p = q->slots[x].dep.prev; \
246 sfq_dep_head(q, p)->next = n; \
247 sfq_dep_head(q, n)->prev = p
250 static inline void sfq_dec(struct sfq_sched_data
*q
, sfq_index x
)
255 sfq_unlink(q
, x
, n
, p
);
257 d
= q
->slots
[x
].qlen
--;
258 if (n
== p
&& q
->cur_depth
== d
)
263 static inline void sfq_inc(struct sfq_sched_data
*q
, sfq_index x
)
268 sfq_unlink(q
, x
, n
, p
);
270 d
= ++q
->slots
[x
].qlen
;
271 if (q
->cur_depth
< d
)
276 /* helper functions : might be changed when/if skb use a standard list_head */
278 /* remove one skb from tail of slot queue */
279 static inline struct sk_buff
*slot_dequeue_tail(struct sfq_slot
*slot
)
281 struct sk_buff
*skb
= slot
->skblist_prev
;
283 slot
->skblist_prev
= skb
->prev
;
284 skb
->next
= skb
->prev
= NULL
;
288 /* remove one skb from head of slot queue */
289 static inline struct sk_buff
*slot_dequeue_head(struct sfq_slot
*slot
)
291 struct sk_buff
*skb
= slot
->skblist_next
;
293 slot
->skblist_next
= skb
->next
;
294 skb
->next
= skb
->prev
= NULL
;
298 static inline void slot_queue_init(struct sfq_slot
*slot
)
300 slot
->skblist_prev
= slot
->skblist_next
= (struct sk_buff
*)slot
;
303 /* add skb to slot queue (tail add) */
304 static inline void slot_queue_add(struct sfq_slot
*slot
, struct sk_buff
*skb
)
306 skb
->prev
= slot
->skblist_prev
;
307 skb
->next
= (struct sk_buff
*)slot
;
308 slot
->skblist_prev
->next
= skb
;
309 slot
->skblist_prev
= skb
;
312 #define slot_queue_walk(slot, skb) \
313 for (skb = slot->skblist_next; \
314 skb != (struct sk_buff *)slot; \
317 static unsigned int sfq_drop(struct Qdisc
*sch
)
319 struct sfq_sched_data
*q
= qdisc_priv(sch
);
320 sfq_index x
, d
= q
->cur_depth
;
323 struct sfq_slot
*slot
;
325 /* Queue is full! Find the longest slot and drop tail packet from it */
330 skb
= slot_dequeue_tail(slot
);
331 len
= qdisc_pkt_len(skb
);
336 sch
->qstats
.backlog
-= len
;
341 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
344 q
->tail
->next
= slot
->next
;
345 q
->ht
[slot
->hash
] = SFQ_EMPTY_SLOT
;
353 sfq_enqueue(struct sk_buff
*skb
, struct Qdisc
*sch
)
355 struct sfq_sched_data
*q
= qdisc_priv(sch
);
358 struct sfq_slot
*slot
;
359 int uninitialized_var(ret
);
361 hash
= sfq_classify(skb
, sch
, &ret
);
363 if (ret
& __NET_XMIT_BYPASS
)
372 if (x
== SFQ_EMPTY_SLOT
) {
373 x
= q
->dep
[0].next
; /* get a free slot */
377 slot_queue_init(slot
);
380 /* If selected queue has length q->limit, do simple tail drop,
381 * i.e. drop _this_ packet.
383 if (slot
->qlen
>= q
->limit
)
384 return qdisc_drop(skb
, sch
);
386 sch
->qstats
.backlog
+= qdisc_pkt_len(skb
);
387 slot_queue_add(slot
, skb
);
389 if (slot
->qlen
== 1) { /* The flow is new */
390 if (q
->tail
== NULL
) { /* It is the first flow */
393 slot
->next
= q
->tail
->next
;
397 slot
->allot
= q
->quantum
;
399 if (++sch
->q
.qlen
<= q
->limit
) {
400 sch
->bstats
.bytes
+= qdisc_pkt_len(skb
);
401 sch
->bstats
.packets
++;
402 return NET_XMIT_SUCCESS
;
409 static struct sk_buff
*
410 sfq_peek(struct Qdisc
*sch
)
412 struct sfq_sched_data
*q
= qdisc_priv(sch
);
414 /* No active slots */
418 return q
->slots
[q
->tail
->next
].skblist_next
;
421 static struct sk_buff
*
422 sfq_dequeue(struct Qdisc
*sch
)
424 struct sfq_sched_data
*q
= qdisc_priv(sch
);
427 struct sfq_slot
*slot
;
429 /* No active slots */
435 skb
= slot_dequeue_head(slot
);
438 sch
->qstats
.backlog
-= qdisc_pkt_len(skb
);
440 /* Is the slot empty? */
441 if (slot
->qlen
== 0) {
442 q
->ht
[slot
->hash
] = SFQ_EMPTY_SLOT
;
445 q
->tail
= NULL
; /* no more active slots */
448 q
->tail
->next
= next_a
;
449 } else if ((slot
->allot
-= qdisc_pkt_len(skb
)) <= 0) {
451 slot
->allot
+= q
->quantum
;
457 sfq_reset(struct Qdisc
*sch
)
461 while ((skb
= sfq_dequeue(sch
)) != NULL
)
465 static void sfq_perturbation(unsigned long arg
)
467 struct Qdisc
*sch
= (struct Qdisc
*)arg
;
468 struct sfq_sched_data
*q
= qdisc_priv(sch
);
470 q
->perturbation
= net_random();
472 if (q
->perturb_period
)
473 mod_timer(&q
->perturb_timer
, jiffies
+ q
->perturb_period
);
476 static int sfq_change(struct Qdisc
*sch
, struct nlattr
*opt
)
478 struct sfq_sched_data
*q
= qdisc_priv(sch
);
479 struct tc_sfq_qopt
*ctl
= nla_data(opt
);
482 if (opt
->nla_len
< nla_attr_size(sizeof(*ctl
)))
486 q
->quantum
= ctl
->quantum
? : psched_mtu(qdisc_dev(sch
));
487 q
->perturb_period
= ctl
->perturb_period
* HZ
;
489 q
->limit
= min_t(u32
, ctl
->limit
, SFQ_DEPTH
- 1);
492 while (sch
->q
.qlen
> q
->limit
)
494 qdisc_tree_decrease_qlen(sch
, qlen
- sch
->q
.qlen
);
496 del_timer(&q
->perturb_timer
);
497 if (q
->perturb_period
) {
498 mod_timer(&q
->perturb_timer
, jiffies
+ q
->perturb_period
);
499 q
->perturbation
= net_random();
501 sch_tree_unlock(sch
);
505 static int sfq_init(struct Qdisc
*sch
, struct nlattr
*opt
)
507 struct sfq_sched_data
*q
= qdisc_priv(sch
);
510 q
->perturb_timer
.function
= sfq_perturbation
;
511 q
->perturb_timer
.data
= (unsigned long)sch
;
512 init_timer_deferrable(&q
->perturb_timer
);
514 for (i
= 0; i
< SFQ_HASH_DIVISOR
; i
++)
515 q
->ht
[i
] = SFQ_EMPTY_SLOT
;
517 for (i
= 0; i
< SFQ_DEPTH
; i
++) {
518 q
->dep
[i
].next
= i
+ SFQ_SLOTS
;
519 q
->dep
[i
].prev
= i
+ SFQ_SLOTS
;
522 q
->limit
= SFQ_DEPTH
- 1;
526 q
->quantum
= psched_mtu(qdisc_dev(sch
));
527 q
->perturb_period
= 0;
528 q
->perturbation
= net_random();
530 int err
= sfq_change(sch
, opt
);
535 for (i
= 0; i
< SFQ_SLOTS
; i
++)
540 static void sfq_destroy(struct Qdisc
*sch
)
542 struct sfq_sched_data
*q
= qdisc_priv(sch
);
544 tcf_destroy_chain(&q
->filter_list
);
545 q
->perturb_period
= 0;
546 del_timer_sync(&q
->perturb_timer
);
549 static int sfq_dump(struct Qdisc
*sch
, struct sk_buff
*skb
)
551 struct sfq_sched_data
*q
= qdisc_priv(sch
);
552 unsigned char *b
= skb_tail_pointer(skb
);
553 struct tc_sfq_qopt opt
;
555 opt
.quantum
= q
->quantum
;
556 opt
.perturb_period
= q
->perturb_period
/ HZ
;
558 opt
.limit
= q
->limit
;
559 opt
.divisor
= SFQ_HASH_DIVISOR
;
560 opt
.flows
= q
->limit
;
562 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
571 static struct Qdisc
*sfq_leaf(struct Qdisc
*sch
, unsigned long arg
)
576 static unsigned long sfq_get(struct Qdisc
*sch
, u32 classid
)
581 static unsigned long sfq_bind(struct Qdisc
*sch
, unsigned long parent
,
587 static void sfq_put(struct Qdisc
*q
, unsigned long cl
)
591 static struct tcf_proto
**sfq_find_tcf(struct Qdisc
*sch
, unsigned long cl
)
593 struct sfq_sched_data
*q
= qdisc_priv(sch
);
597 return &q
->filter_list
;
600 static int sfq_dump_class(struct Qdisc
*sch
, unsigned long cl
,
601 struct sk_buff
*skb
, struct tcmsg
*tcm
)
603 tcm
->tcm_handle
|= TC_H_MIN(cl
);
607 static int sfq_dump_class_stats(struct Qdisc
*sch
, unsigned long cl
,
610 struct sfq_sched_data
*q
= qdisc_priv(sch
);
611 const struct sfq_slot
*slot
= &q
->slots
[q
->ht
[cl
- 1]];
612 struct gnet_stats_queue qs
= { .qlen
= slot
->qlen
};
613 struct tc_sfq_xstats xstats
= { .allot
= slot
->allot
};
616 slot_queue_walk(slot
, skb
)
617 qs
.backlog
+= qdisc_pkt_len(skb
);
619 if (gnet_stats_copy_queue(d
, &qs
) < 0)
621 return gnet_stats_copy_app(d
, &xstats
, sizeof(xstats
));
624 static void sfq_walk(struct Qdisc
*sch
, struct qdisc_walker
*arg
)
626 struct sfq_sched_data
*q
= qdisc_priv(sch
);
632 for (i
= 0; i
< SFQ_HASH_DIVISOR
; i
++) {
633 if (q
->ht
[i
] == SFQ_EMPTY_SLOT
||
634 arg
->count
< arg
->skip
) {
638 if (arg
->fn(sch
, i
+ 1, arg
) < 0) {
646 static const struct Qdisc_class_ops sfq_class_ops
= {
650 .tcf_chain
= sfq_find_tcf
,
651 .bind_tcf
= sfq_bind
,
652 .unbind_tcf
= sfq_put
,
653 .dump
= sfq_dump_class
,
654 .dump_stats
= sfq_dump_class_stats
,
658 static struct Qdisc_ops sfq_qdisc_ops __read_mostly
= {
659 .cl_ops
= &sfq_class_ops
,
661 .priv_size
= sizeof(struct sfq_sched_data
),
662 .enqueue
= sfq_enqueue
,
663 .dequeue
= sfq_dequeue
,
668 .destroy
= sfq_destroy
,
671 .owner
= THIS_MODULE
,
674 static int __init
sfq_module_init(void)
676 return register_qdisc(&sfq_qdisc_ops
);
678 static void __exit
sfq_module_exit(void)
680 unregister_qdisc(&sfq_qdisc_ops
);
682 module_init(sfq_module_init
)
683 module_exit(sfq_module_exit
)
684 MODULE_LICENSE("GPL");