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
= skb
->prev
= NULL
;
304 static inline void slot_queue_init(struct sfq_slot
*slot
)
306 slot
->skblist_prev
= slot
->skblist_next
= (struct sk_buff
*)slot
;
309 /* add skb to slot queue (tail add) */
310 static inline void slot_queue_add(struct sfq_slot
*slot
, struct sk_buff
*skb
)
312 skb
->prev
= slot
->skblist_prev
;
313 skb
->next
= (struct sk_buff
*)slot
;
314 slot
->skblist_prev
->next
= skb
;
315 slot
->skblist_prev
= skb
;
318 #define slot_queue_walk(slot, skb) \
319 for (skb = slot->skblist_next; \
320 skb != (struct sk_buff *)slot; \
323 static unsigned int sfq_drop(struct Qdisc
*sch
)
325 struct sfq_sched_data
*q
= qdisc_priv(sch
);
326 sfq_index x
, d
= q
->cur_depth
;
329 struct sfq_slot
*slot
;
331 /* Queue is full! Find the longest slot and drop tail packet from it */
336 skb
= slot_dequeue_tail(slot
);
337 len
= qdisc_pkt_len(skb
);
342 sch
->qstats
.backlog
-= len
;
347 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
350 q
->tail
->next
= slot
->next
;
351 q
->ht
[slot
->hash
] = SFQ_EMPTY_SLOT
;
359 sfq_enqueue(struct sk_buff
*skb
, struct Qdisc
*sch
)
361 struct sfq_sched_data
*q
= qdisc_priv(sch
);
364 struct sfq_slot
*slot
;
365 int uninitialized_var(ret
);
367 hash
= sfq_classify(skb
, sch
, &ret
);
369 if (ret
& __NET_XMIT_BYPASS
)
378 if (x
== SFQ_EMPTY_SLOT
) {
379 x
= q
->dep
[0].next
; /* get a free slot */
383 slot_queue_init(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 sch
->bstats
.bytes
+= qdisc_pkt_len(skb
);
407 sch
->bstats
.packets
++;
408 return NET_XMIT_SUCCESS
;
415 static struct sk_buff
*
416 sfq_peek(struct Qdisc
*sch
)
418 struct sfq_sched_data
*q
= qdisc_priv(sch
);
420 /* No active slots */
424 return q
->slots
[q
->tail
->next
].skblist_next
;
427 static struct sk_buff
*
428 sfq_dequeue(struct Qdisc
*sch
)
430 struct sfq_sched_data
*q
= qdisc_priv(sch
);
433 struct sfq_slot
*slot
;
435 /* No active slots */
442 if (slot
->allot
<= 0) {
444 slot
->allot
+= q
->scaled_quantum
;
447 skb
= slot_dequeue_head(slot
);
450 sch
->qstats
.backlog
-= qdisc_pkt_len(skb
);
452 /* Is the slot empty? */
453 if (slot
->qlen
== 0) {
454 q
->ht
[slot
->hash
] = SFQ_EMPTY_SLOT
;
457 q
->tail
= NULL
; /* no more active slots */
460 q
->tail
->next
= next_a
;
462 slot
->allot
-= SFQ_ALLOT_SIZE(qdisc_pkt_len(skb
));
468 sfq_reset(struct Qdisc
*sch
)
472 while ((skb
= sfq_dequeue(sch
)) != NULL
)
476 static void sfq_perturbation(unsigned long arg
)
478 struct Qdisc
*sch
= (struct Qdisc
*)arg
;
479 struct sfq_sched_data
*q
= qdisc_priv(sch
);
481 q
->perturbation
= net_random();
483 if (q
->perturb_period
)
484 mod_timer(&q
->perturb_timer
, jiffies
+ q
->perturb_period
);
487 static int sfq_change(struct Qdisc
*sch
, struct nlattr
*opt
)
489 struct sfq_sched_data
*q
= qdisc_priv(sch
);
490 struct tc_sfq_qopt
*ctl
= nla_data(opt
);
493 if (opt
->nla_len
< nla_attr_size(sizeof(*ctl
)))
497 q
->quantum
= ctl
->quantum
? : psched_mtu(qdisc_dev(sch
));
498 q
->scaled_quantum
= SFQ_ALLOT_SIZE(q
->quantum
);
499 q
->perturb_period
= ctl
->perturb_period
* HZ
;
501 q
->limit
= min_t(u32
, ctl
->limit
, SFQ_DEPTH
- 1);
504 while (sch
->q
.qlen
> q
->limit
)
506 qdisc_tree_decrease_qlen(sch
, qlen
- sch
->q
.qlen
);
508 del_timer(&q
->perturb_timer
);
509 if (q
->perturb_period
) {
510 mod_timer(&q
->perturb_timer
, jiffies
+ q
->perturb_period
);
511 q
->perturbation
= net_random();
513 sch_tree_unlock(sch
);
517 static int sfq_init(struct Qdisc
*sch
, struct nlattr
*opt
)
519 struct sfq_sched_data
*q
= qdisc_priv(sch
);
522 q
->perturb_timer
.function
= sfq_perturbation
;
523 q
->perturb_timer
.data
= (unsigned long)sch
;
524 init_timer_deferrable(&q
->perturb_timer
);
526 for (i
= 0; i
< SFQ_HASH_DIVISOR
; i
++)
527 q
->ht
[i
] = SFQ_EMPTY_SLOT
;
529 for (i
= 0; i
< SFQ_DEPTH
; i
++) {
530 q
->dep
[i
].next
= i
+ SFQ_SLOTS
;
531 q
->dep
[i
].prev
= i
+ SFQ_SLOTS
;
534 q
->limit
= SFQ_DEPTH
- 1;
538 q
->quantum
= psched_mtu(qdisc_dev(sch
));
539 q
->scaled_quantum
= SFQ_ALLOT_SIZE(q
->quantum
);
540 q
->perturb_period
= 0;
541 q
->perturbation
= net_random();
543 int err
= sfq_change(sch
, opt
);
548 for (i
= 0; i
< SFQ_SLOTS
; i
++)
553 static void sfq_destroy(struct Qdisc
*sch
)
555 struct sfq_sched_data
*q
= qdisc_priv(sch
);
557 tcf_destroy_chain(&q
->filter_list
);
558 q
->perturb_period
= 0;
559 del_timer_sync(&q
->perturb_timer
);
562 static int sfq_dump(struct Qdisc
*sch
, struct sk_buff
*skb
)
564 struct sfq_sched_data
*q
= qdisc_priv(sch
);
565 unsigned char *b
= skb_tail_pointer(skb
);
566 struct tc_sfq_qopt opt
;
568 opt
.quantum
= q
->quantum
;
569 opt
.perturb_period
= q
->perturb_period
/ HZ
;
571 opt
.limit
= q
->limit
;
572 opt
.divisor
= SFQ_HASH_DIVISOR
;
573 opt
.flows
= q
->limit
;
575 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
584 static struct Qdisc
*sfq_leaf(struct Qdisc
*sch
, unsigned long arg
)
589 static unsigned long sfq_get(struct Qdisc
*sch
, u32 classid
)
594 static unsigned long sfq_bind(struct Qdisc
*sch
, unsigned long parent
,
600 static void sfq_put(struct Qdisc
*q
, unsigned long cl
)
604 static struct tcf_proto
**sfq_find_tcf(struct Qdisc
*sch
, unsigned long cl
)
606 struct sfq_sched_data
*q
= qdisc_priv(sch
);
610 return &q
->filter_list
;
613 static int sfq_dump_class(struct Qdisc
*sch
, unsigned long cl
,
614 struct sk_buff
*skb
, struct tcmsg
*tcm
)
616 tcm
->tcm_handle
|= TC_H_MIN(cl
);
620 static int sfq_dump_class_stats(struct Qdisc
*sch
, unsigned long cl
,
623 struct sfq_sched_data
*q
= qdisc_priv(sch
);
624 sfq_index idx
= q
->ht
[cl
- 1];
625 struct gnet_stats_queue qs
= { 0 };
626 struct tc_sfq_xstats xstats
= { 0 };
629 if (idx
!= SFQ_EMPTY_SLOT
) {
630 const struct sfq_slot
*slot
= &q
->slots
[idx
];
632 xstats
.allot
= slot
->allot
<< SFQ_ALLOT_SHIFT
;
633 qs
.qlen
= slot
->qlen
;
634 slot_queue_walk(slot
, skb
)
635 qs
.backlog
+= qdisc_pkt_len(skb
);
637 if (gnet_stats_copy_queue(d
, &qs
) < 0)
639 return gnet_stats_copy_app(d
, &xstats
, sizeof(xstats
));
642 static void sfq_walk(struct Qdisc
*sch
, struct qdisc_walker
*arg
)
644 struct sfq_sched_data
*q
= qdisc_priv(sch
);
650 for (i
= 0; i
< SFQ_HASH_DIVISOR
; i
++) {
651 if (q
->ht
[i
] == SFQ_EMPTY_SLOT
||
652 arg
->count
< arg
->skip
) {
656 if (arg
->fn(sch
, i
+ 1, arg
) < 0) {
664 static const struct Qdisc_class_ops sfq_class_ops
= {
668 .tcf_chain
= sfq_find_tcf
,
669 .bind_tcf
= sfq_bind
,
670 .unbind_tcf
= sfq_put
,
671 .dump
= sfq_dump_class
,
672 .dump_stats
= sfq_dump_class_stats
,
676 static struct Qdisc_ops sfq_qdisc_ops __read_mostly
= {
677 .cl_ops
= &sfq_class_ops
,
679 .priv_size
= sizeof(struct sfq_sched_data
),
680 .enqueue
= sfq_enqueue
,
681 .dequeue
= sfq_dequeue
,
686 .destroy
= sfq_destroy
,
689 .owner
= THIS_MODULE
,
692 static int __init
sfq_module_init(void)
694 return register_qdisc(&sfq_qdisc_ops
);
696 static void __exit
sfq_module_exit(void)
698 unregister_qdisc(&sfq_qdisc_ops
);
700 module_init(sfq_module_init
)
701 module_exit(sfq_module_exit
)
702 MODULE_LICENSE("GPL");