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; number of hash buckets to 1024.
71 The only goal of this restrictions was that all data
72 fit into one 4K page :-). Struct sfq_sched_data is
73 organized in anti-cache manner: all the data for a bucket
74 are scattered over different locations. This is not good,
75 but it allowed me to put it into 4K.
77 It is easy to increase these values, but not in flight. */
80 #define SFQ_HASH_DIVISOR 1024
82 /* This type should contain at least SFQ_DEPTH*2 values */
83 typedef unsigned char sfq_index
;
95 unsigned quantum
; /* Allotment per round: MUST BE >= MTU */
99 struct tcf_proto
*filter_list
;
100 struct timer_list perturb_timer
;
102 sfq_index tail
; /* Index of current slot in round */
103 sfq_index max_depth
; /* Maximal depth */
105 sfq_index ht
[SFQ_HASH_DIVISOR
]; /* Hash table */
106 sfq_index next
[SFQ_DEPTH
]; /* Active slots link */
107 short allot
[SFQ_DEPTH
]; /* Current allotment per slot */
108 unsigned short hash
[SFQ_DEPTH
]; /* Hash value indexed by slots */
109 struct sk_buff_head qs
[SFQ_DEPTH
]; /* Slot queue */
110 struct sfq_head dep
[SFQ_DEPTH
*2]; /* Linked list of slots, indexed by depth */
113 static __inline__
unsigned sfq_fold_hash(struct sfq_sched_data
*q
, u32 h
, u32 h1
)
115 return jhash_2words(h
, h1
, q
->perturbation
) & (SFQ_HASH_DIVISOR
- 1);
118 static unsigned sfq_hash(struct sfq_sched_data
*q
, struct sk_buff
*skb
)
122 switch (skb
->protocol
) {
123 case htons(ETH_P_IP
):
125 const struct iphdr
*iph
;
127 if (!pskb_network_may_pull(skb
, sizeof(*iph
)))
130 h
= (__force u32
)iph
->daddr
;
131 h2
= (__force u32
)iph
->saddr
^ iph
->protocol
;
132 if (!(iph
->frag_off
&htons(IP_MF
|IP_OFFSET
)) &&
133 (iph
->protocol
== IPPROTO_TCP
||
134 iph
->protocol
== IPPROTO_UDP
||
135 iph
->protocol
== IPPROTO_UDPLITE
||
136 iph
->protocol
== IPPROTO_SCTP
||
137 iph
->protocol
== IPPROTO_DCCP
||
138 iph
->protocol
== IPPROTO_ESP
) &&
139 pskb_network_may_pull(skb
, iph
->ihl
* 4 + 4))
140 h2
^= *(((u32
*)iph
) + iph
->ihl
);
143 case htons(ETH_P_IPV6
):
147 if (!pskb_network_may_pull(skb
, sizeof(*iph
)))
150 h
= (__force u32
)iph
->daddr
.s6_addr32
[3];
151 h2
= (__force u32
)iph
->saddr
.s6_addr32
[3] ^ iph
->nexthdr
;
152 if ((iph
->nexthdr
== IPPROTO_TCP
||
153 iph
->nexthdr
== IPPROTO_UDP
||
154 iph
->nexthdr
== IPPROTO_UDPLITE
||
155 iph
->nexthdr
== IPPROTO_SCTP
||
156 iph
->nexthdr
== IPPROTO_DCCP
||
157 iph
->nexthdr
== IPPROTO_ESP
) &&
158 pskb_network_may_pull(skb
, sizeof(*iph
) + 4))
159 h2
^= *(u32
*)&iph
[1];
164 h
= (unsigned long)skb_dst(skb
) ^ (__force u32
)skb
->protocol
;
165 h2
= (unsigned long)skb
->sk
;
168 return sfq_fold_hash(q
, h
, h2
);
171 static unsigned int sfq_classify(struct sk_buff
*skb
, struct Qdisc
*sch
,
174 struct sfq_sched_data
*q
= qdisc_priv(sch
);
175 struct tcf_result res
;
178 if (TC_H_MAJ(skb
->priority
) == sch
->handle
&&
179 TC_H_MIN(skb
->priority
) > 0 &&
180 TC_H_MIN(skb
->priority
) <= SFQ_HASH_DIVISOR
)
181 return TC_H_MIN(skb
->priority
);
184 return sfq_hash(q
, skb
) + 1;
186 *qerr
= NET_XMIT_SUCCESS
| __NET_XMIT_BYPASS
;
187 result
= tc_classify(skb
, q
->filter_list
, &res
);
189 #ifdef CONFIG_NET_CLS_ACT
193 *qerr
= NET_XMIT_SUCCESS
| __NET_XMIT_STOLEN
;
198 if (TC_H_MIN(res
.classid
) <= SFQ_HASH_DIVISOR
)
199 return TC_H_MIN(res
.classid
);
204 static inline void sfq_link(struct sfq_sched_data
*q
, sfq_index x
)
207 int d
= q
->qs
[x
].qlen
+ SFQ_DEPTH
;
213 q
->dep
[p
].next
= q
->dep
[n
].prev
= x
;
216 static inline void sfq_dec(struct sfq_sched_data
*q
, sfq_index x
)
225 if (n
== p
&& q
->max_depth
== q
->qs
[x
].qlen
+ 1)
231 static inline void sfq_inc(struct sfq_sched_data
*q
, sfq_index x
)
241 if (q
->max_depth
< d
)
247 static unsigned int sfq_drop(struct Qdisc
*sch
)
249 struct sfq_sched_data
*q
= qdisc_priv(sch
);
250 sfq_index d
= q
->max_depth
;
254 /* Queue is full! Find the longest slot and
255 drop a packet from it */
258 sfq_index x
= q
->dep
[d
+ SFQ_DEPTH
].next
;
260 len
= qdisc_pkt_len(skb
);
261 __skb_unlink(skb
, &q
->qs
[x
]);
266 sch
->qstats
.backlog
-= len
;
271 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
272 d
= q
->next
[q
->tail
];
273 q
->next
[q
->tail
] = q
->next
[d
];
274 q
->allot
[q
->next
[d
]] += q
->quantum
;
276 len
= qdisc_pkt_len(skb
);
277 __skb_unlink(skb
, &q
->qs
[d
]);
281 q
->ht
[q
->hash
[d
]] = SFQ_DEPTH
;
283 sch
->qstats
.backlog
-= len
;
291 sfq_enqueue(struct sk_buff
*skb
, struct Qdisc
*sch
)
293 struct sfq_sched_data
*q
= qdisc_priv(sch
);
296 int uninitialized_var(ret
);
298 hash
= sfq_classify(skb
, sch
, &ret
);
300 if (ret
& __NET_XMIT_BYPASS
)
308 if (x
== SFQ_DEPTH
) {
309 q
->ht
[hash
] = x
= q
->dep
[SFQ_DEPTH
].next
;
313 /* If selected queue has length q->limit, this means that
314 * all another queues are empty and that we do simple tail drop,
315 * i.e. drop _this_ packet.
317 if (q
->qs
[x
].qlen
>= q
->limit
)
318 return qdisc_drop(skb
, sch
);
320 sch
->qstats
.backlog
+= qdisc_pkt_len(skb
);
321 __skb_queue_tail(&q
->qs
[x
], skb
);
323 if (q
->qs
[x
].qlen
== 1) { /* The flow is new */
324 if (q
->tail
== SFQ_DEPTH
) { /* It is the first flow */
327 q
->allot
[x
] = q
->quantum
;
329 q
->next
[x
] = q
->next
[q
->tail
];
330 q
->next
[q
->tail
] = x
;
334 if (++sch
->q
.qlen
<= q
->limit
) {
335 sch
->bstats
.bytes
+= qdisc_pkt_len(skb
);
336 sch
->bstats
.packets
++;
337 return NET_XMIT_SUCCESS
;
344 static struct sk_buff
*
345 sfq_peek(struct Qdisc
*sch
)
347 struct sfq_sched_data
*q
= qdisc_priv(sch
);
350 /* No active slots */
351 if (q
->tail
== SFQ_DEPTH
)
354 a
= q
->next
[q
->tail
];
355 return skb_peek(&q
->qs
[a
]);
358 static struct sk_buff
*
359 sfq_dequeue(struct Qdisc
*sch
)
361 struct sfq_sched_data
*q
= qdisc_priv(sch
);
365 /* No active slots */
366 if (q
->tail
== SFQ_DEPTH
)
369 a
= old_a
= q
->next
[q
->tail
];
372 skb
= __skb_dequeue(&q
->qs
[a
]);
375 sch
->qstats
.backlog
-= qdisc_pkt_len(skb
);
377 /* Is the slot empty? */
378 if (q
->qs
[a
].qlen
== 0) {
379 q
->ht
[q
->hash
[a
]] = SFQ_DEPTH
;
385 q
->next
[q
->tail
] = a
;
386 q
->allot
[a
] += q
->quantum
;
387 } else if ((q
->allot
[a
] -= qdisc_pkt_len(skb
)) <= 0) {
390 q
->allot
[a
] += q
->quantum
;
396 sfq_reset(struct Qdisc
*sch
)
400 while ((skb
= sfq_dequeue(sch
)) != NULL
)
404 static void sfq_perturbation(unsigned long arg
)
406 struct Qdisc
*sch
= (struct Qdisc
*)arg
;
407 struct sfq_sched_data
*q
= qdisc_priv(sch
);
409 q
->perturbation
= net_random();
411 if (q
->perturb_period
)
412 mod_timer(&q
->perturb_timer
, jiffies
+ q
->perturb_period
);
415 static int sfq_change(struct Qdisc
*sch
, struct nlattr
*opt
)
417 struct sfq_sched_data
*q
= qdisc_priv(sch
);
418 struct tc_sfq_qopt
*ctl
= nla_data(opt
);
421 if (opt
->nla_len
< nla_attr_size(sizeof(*ctl
)))
425 q
->quantum
= ctl
->quantum
? : psched_mtu(qdisc_dev(sch
));
426 q
->perturb_period
= ctl
->perturb_period
* HZ
;
428 q
->limit
= min_t(u32
, ctl
->limit
, SFQ_DEPTH
- 1);
431 while (sch
->q
.qlen
> q
->limit
)
433 qdisc_tree_decrease_qlen(sch
, qlen
- sch
->q
.qlen
);
435 del_timer(&q
->perturb_timer
);
436 if (q
->perturb_period
) {
437 mod_timer(&q
->perturb_timer
, jiffies
+ q
->perturb_period
);
438 q
->perturbation
= net_random();
440 sch_tree_unlock(sch
);
444 static int sfq_init(struct Qdisc
*sch
, struct nlattr
*opt
)
446 struct sfq_sched_data
*q
= qdisc_priv(sch
);
449 q
->perturb_timer
.function
= sfq_perturbation
;
450 q
->perturb_timer
.data
= (unsigned long)sch
;
451 init_timer_deferrable(&q
->perturb_timer
);
453 for (i
= 0; i
< SFQ_HASH_DIVISOR
; i
++)
454 q
->ht
[i
] = SFQ_DEPTH
;
456 for (i
= 0; i
< SFQ_DEPTH
; i
++) {
457 skb_queue_head_init(&q
->qs
[i
]);
458 q
->dep
[i
+ SFQ_DEPTH
].next
= i
+ SFQ_DEPTH
;
459 q
->dep
[i
+ SFQ_DEPTH
].prev
= i
+ SFQ_DEPTH
;
462 q
->limit
= SFQ_DEPTH
- 1;
466 q
->quantum
= psched_mtu(qdisc_dev(sch
));
467 q
->perturb_period
= 0;
468 q
->perturbation
= net_random();
470 int err
= sfq_change(sch
, opt
);
475 for (i
= 0; i
< SFQ_DEPTH
; i
++)
480 static void sfq_destroy(struct Qdisc
*sch
)
482 struct sfq_sched_data
*q
= qdisc_priv(sch
);
484 tcf_destroy_chain(&q
->filter_list
);
485 q
->perturb_period
= 0;
486 del_timer_sync(&q
->perturb_timer
);
489 static int sfq_dump(struct Qdisc
*sch
, struct sk_buff
*skb
)
491 struct sfq_sched_data
*q
= qdisc_priv(sch
);
492 unsigned char *b
= skb_tail_pointer(skb
);
493 struct tc_sfq_qopt opt
;
495 opt
.quantum
= q
->quantum
;
496 opt
.perturb_period
= q
->perturb_period
/ HZ
;
498 opt
.limit
= q
->limit
;
499 opt
.divisor
= SFQ_HASH_DIVISOR
;
500 opt
.flows
= q
->limit
;
502 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
511 static struct Qdisc
*sfq_leaf(struct Qdisc
*sch
, unsigned long arg
)
516 static unsigned long sfq_get(struct Qdisc
*sch
, u32 classid
)
521 static unsigned long sfq_bind(struct Qdisc
*sch
, unsigned long parent
,
527 static void sfq_put(struct Qdisc
*q
, unsigned long cl
)
531 static struct tcf_proto
**sfq_find_tcf(struct Qdisc
*sch
, unsigned long cl
)
533 struct sfq_sched_data
*q
= qdisc_priv(sch
);
537 return &q
->filter_list
;
540 static int sfq_dump_class(struct Qdisc
*sch
, unsigned long cl
,
541 struct sk_buff
*skb
, struct tcmsg
*tcm
)
543 tcm
->tcm_handle
|= TC_H_MIN(cl
);
547 static int sfq_dump_class_stats(struct Qdisc
*sch
, unsigned long cl
,
550 struct sfq_sched_data
*q
= qdisc_priv(sch
);
551 sfq_index idx
= q
->ht
[cl
-1];
552 struct gnet_stats_queue qs
= { .qlen
= q
->qs
[idx
].qlen
};
553 struct tc_sfq_xstats xstats
= { .allot
= q
->allot
[idx
] };
555 if (gnet_stats_copy_queue(d
, &qs
) < 0)
557 return gnet_stats_copy_app(d
, &xstats
, sizeof(xstats
));
560 static void sfq_walk(struct Qdisc
*sch
, struct qdisc_walker
*arg
)
562 struct sfq_sched_data
*q
= qdisc_priv(sch
);
568 for (i
= 0; i
< SFQ_HASH_DIVISOR
; i
++) {
569 if (q
->ht
[i
] == SFQ_DEPTH
||
570 arg
->count
< arg
->skip
) {
574 if (arg
->fn(sch
, i
+ 1, arg
) < 0) {
582 static const struct Qdisc_class_ops sfq_class_ops
= {
586 .tcf_chain
= sfq_find_tcf
,
587 .bind_tcf
= sfq_bind
,
588 .unbind_tcf
= sfq_put
,
589 .dump
= sfq_dump_class
,
590 .dump_stats
= sfq_dump_class_stats
,
594 static struct Qdisc_ops sfq_qdisc_ops __read_mostly
= {
595 .cl_ops
= &sfq_class_ops
,
597 .priv_size
= sizeof(struct sfq_sched_data
),
598 .enqueue
= sfq_enqueue
,
599 .dequeue
= sfq_dequeue
,
604 .destroy
= sfq_destroy
,
607 .owner
= THIS_MODULE
,
610 static int __init
sfq_module_init(void)
612 return register_qdisc(&sfq_qdisc_ops
);
614 static void __exit
sfq_module_exit(void)
616 unregister_qdisc(&sfq_qdisc_ops
);
618 module_init(sfq_module_init
)
619 module_exit(sfq_module_exit
)
620 MODULE_LICENSE("GPL");