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>
26 #include <net/flow_keys.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_DEFAULT_HASH_DIVISOR 1024
81 /* We use 16 bits to store allot, and want to handle packets up to 64K
82 * Scale allot by 8 (1<<3) so that no overflow occurs.
84 #define SFQ_ALLOT_SHIFT 3
85 #define SFQ_ALLOT_SIZE(X) DIV_ROUND_UP(X, 1 << SFQ_ALLOT_SHIFT)
87 /* This type should contain at least SFQ_DEPTH + SFQ_SLOTS values */
88 typedef unsigned char sfq_index
;
91 * We dont use pointers to save space.
92 * Small indexes [0 ... SFQ_SLOTS - 1] are 'pointers' to slots[] array
93 * while following values [SFQ_SLOTS ... SFQ_SLOTS + SFQ_DEPTH - 1]
94 * 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
{
114 unsigned int quantum
; /* Allotment per round: MUST BE >= MTU */
116 unsigned int divisor
; /* number of slots in hash table */
118 struct tcf_proto
*filter_list
;
119 struct timer_list perturb_timer
;
121 sfq_index cur_depth
; /* depth of longest slot */
122 unsigned short scaled_quantum
; /* SFQ_ALLOT_SIZE(quantum) */
123 struct sfq_slot
*tail
; /* current slot in round */
124 sfq_index
*ht
; /* Hash table (divisor slots) */
125 struct sfq_slot slots
[SFQ_SLOTS
];
126 struct sfq_head dep
[SFQ_DEPTH
]; /* Linked list of slots, indexed by depth */
130 * sfq_head are either in a sfq_slot or in dep[] array
132 static inline struct sfq_head
*sfq_dep_head(struct sfq_sched_data
*q
, sfq_index val
)
135 return &q
->slots
[val
].dep
;
136 return &q
->dep
[val
- SFQ_SLOTS
];
139 static unsigned int sfq_hash(const struct sfq_sched_data
*q
,
140 const struct sk_buff
*skb
)
142 struct flow_keys keys
;
145 skb_flow_dissect(skb
, &keys
);
146 hash
= jhash_3words((__force u32
)keys
.dst
,
147 (__force u32
)keys
.src
^ keys
.ip_proto
,
148 (__force u32
)keys
.ports
, q
->perturbation
);
149 return hash
& (q
->divisor
- 1);
152 static unsigned int sfq_classify(struct sk_buff
*skb
, struct Qdisc
*sch
,
155 struct sfq_sched_data
*q
= qdisc_priv(sch
);
156 struct tcf_result res
;
159 if (TC_H_MAJ(skb
->priority
) == sch
->handle
&&
160 TC_H_MIN(skb
->priority
) > 0 &&
161 TC_H_MIN(skb
->priority
) <= q
->divisor
)
162 return TC_H_MIN(skb
->priority
);
165 return sfq_hash(q
, skb
) + 1;
167 *qerr
= NET_XMIT_SUCCESS
| __NET_XMIT_BYPASS
;
168 result
= tc_classify(skb
, q
->filter_list
, &res
);
170 #ifdef CONFIG_NET_CLS_ACT
174 *qerr
= NET_XMIT_SUCCESS
| __NET_XMIT_STOLEN
;
179 if (TC_H_MIN(res
.classid
) <= q
->divisor
)
180 return TC_H_MIN(res
.classid
);
186 * x : slot number [0 .. SFQ_SLOTS - 1]
188 static inline void sfq_link(struct sfq_sched_data
*q
, sfq_index x
)
191 int qlen
= q
->slots
[x
].qlen
;
193 p
= qlen
+ SFQ_SLOTS
;
194 n
= q
->dep
[qlen
].next
;
196 q
->slots
[x
].dep
.next
= n
;
197 q
->slots
[x
].dep
.prev
= p
;
199 q
->dep
[qlen
].next
= x
; /* sfq_dep_head(q, p)->next = x */
200 sfq_dep_head(q
, n
)->prev
= x
;
203 #define sfq_unlink(q, x, n, p) \
204 n = q->slots[x].dep.next; \
205 p = q->slots[x].dep.prev; \
206 sfq_dep_head(q, p)->next = n; \
207 sfq_dep_head(q, n)->prev = p
210 static inline void sfq_dec(struct sfq_sched_data
*q
, sfq_index x
)
215 sfq_unlink(q
, x
, n
, p
);
217 d
= q
->slots
[x
].qlen
--;
218 if (n
== p
&& q
->cur_depth
== d
)
223 static inline void sfq_inc(struct sfq_sched_data
*q
, sfq_index x
)
228 sfq_unlink(q
, x
, n
, p
);
230 d
= ++q
->slots
[x
].qlen
;
231 if (q
->cur_depth
< d
)
236 /* helper functions : might be changed when/if skb use a standard list_head */
238 /* remove one skb from tail of slot queue */
239 static inline struct sk_buff
*slot_dequeue_tail(struct sfq_slot
*slot
)
241 struct sk_buff
*skb
= slot
->skblist_prev
;
243 slot
->skblist_prev
= skb
->prev
;
244 skb
->prev
->next
= (struct sk_buff
*)slot
;
245 skb
->next
= skb
->prev
= NULL
;
249 /* remove one skb from head of slot queue */
250 static inline struct sk_buff
*slot_dequeue_head(struct sfq_slot
*slot
)
252 struct sk_buff
*skb
= slot
->skblist_next
;
254 slot
->skblist_next
= skb
->next
;
255 skb
->next
->prev
= (struct sk_buff
*)slot
;
256 skb
->next
= skb
->prev
= NULL
;
260 static inline void slot_queue_init(struct sfq_slot
*slot
)
262 slot
->skblist_prev
= slot
->skblist_next
= (struct sk_buff
*)slot
;
265 /* add skb to slot queue (tail add) */
266 static inline void slot_queue_add(struct sfq_slot
*slot
, struct sk_buff
*skb
)
268 skb
->prev
= slot
->skblist_prev
;
269 skb
->next
= (struct sk_buff
*)slot
;
270 slot
->skblist_prev
->next
= skb
;
271 slot
->skblist_prev
= skb
;
274 #define slot_queue_walk(slot, skb) \
275 for (skb = slot->skblist_next; \
276 skb != (struct sk_buff *)slot; \
279 static unsigned int sfq_drop(struct Qdisc
*sch
)
281 struct sfq_sched_data
*q
= qdisc_priv(sch
);
282 sfq_index x
, d
= q
->cur_depth
;
285 struct sfq_slot
*slot
;
287 /* Queue is full! Find the longest slot and drop tail packet from it */
292 skb
= slot_dequeue_tail(slot
);
293 len
= qdisc_pkt_len(skb
);
298 sch
->qstats
.backlog
-= len
;
303 /* It is difficult to believe, but ALL THE SLOTS HAVE LENGTH 1. */
306 q
->tail
->next
= slot
->next
;
307 q
->ht
[slot
->hash
] = SFQ_EMPTY_SLOT
;
315 sfq_enqueue(struct sk_buff
*skb
, struct Qdisc
*sch
)
317 struct sfq_sched_data
*q
= qdisc_priv(sch
);
320 struct sfq_slot
*slot
;
321 int uninitialized_var(ret
);
323 hash
= sfq_classify(skb
, sch
, &ret
);
325 if (ret
& __NET_XMIT_BYPASS
)
334 if (x
== SFQ_EMPTY_SLOT
) {
335 x
= q
->dep
[0].next
; /* get a free slot */
341 /* If selected queue has length q->limit, do simple tail drop,
342 * i.e. drop _this_ packet.
344 if (slot
->qlen
>= q
->limit
)
345 return qdisc_drop(skb
, sch
);
347 sch
->qstats
.backlog
+= qdisc_pkt_len(skb
);
348 slot_queue_add(slot
, skb
);
350 if (slot
->qlen
== 1) { /* The flow is new */
351 if (q
->tail
== NULL
) { /* It is the first flow */
354 slot
->next
= q
->tail
->next
;
358 slot
->allot
= q
->scaled_quantum
;
360 if (++sch
->q
.qlen
<= q
->limit
)
361 return NET_XMIT_SUCCESS
;
365 /* Return Congestion Notification only if we dropped a packet
368 if (qlen
!= slot
->qlen
)
371 /* As we dropped a packet, better let upper stack know this */
372 qdisc_tree_decrease_qlen(sch
, 1);
373 return NET_XMIT_SUCCESS
;
376 static struct sk_buff
*
377 sfq_dequeue(struct Qdisc
*sch
)
379 struct sfq_sched_data
*q
= qdisc_priv(sch
);
382 struct sfq_slot
*slot
;
384 /* No active slots */
391 if (slot
->allot
<= 0) {
393 slot
->allot
+= q
->scaled_quantum
;
396 skb
= slot_dequeue_head(slot
);
398 qdisc_bstats_update(sch
, skb
);
400 sch
->qstats
.backlog
-= qdisc_pkt_len(skb
);
402 /* Is the slot empty? */
403 if (slot
->qlen
== 0) {
404 q
->ht
[slot
->hash
] = SFQ_EMPTY_SLOT
;
407 q
->tail
= NULL
; /* no more active slots */
410 q
->tail
->next
= next_a
;
412 slot
->allot
-= SFQ_ALLOT_SIZE(qdisc_pkt_len(skb
));
418 sfq_reset(struct Qdisc
*sch
)
422 while ((skb
= sfq_dequeue(sch
)) != NULL
)
426 static void sfq_perturbation(unsigned long arg
)
428 struct Qdisc
*sch
= (struct Qdisc
*)arg
;
429 struct sfq_sched_data
*q
= qdisc_priv(sch
);
431 q
->perturbation
= net_random();
433 if (q
->perturb_period
)
434 mod_timer(&q
->perturb_timer
, jiffies
+ q
->perturb_period
);
437 static int sfq_change(struct Qdisc
*sch
, struct nlattr
*opt
)
439 struct sfq_sched_data
*q
= qdisc_priv(sch
);
440 struct tc_sfq_qopt
*ctl
= nla_data(opt
);
443 if (opt
->nla_len
< nla_attr_size(sizeof(*ctl
)))
447 (!is_power_of_2(ctl
->divisor
) || ctl
->divisor
> 65536))
451 q
->quantum
= ctl
->quantum
? : psched_mtu(qdisc_dev(sch
));
452 q
->scaled_quantum
= SFQ_ALLOT_SIZE(q
->quantum
);
453 q
->perturb_period
= ctl
->perturb_period
* HZ
;
455 q
->limit
= min_t(u32
, ctl
->limit
, SFQ_DEPTH
- 1);
457 q
->divisor
= ctl
->divisor
;
459 while (sch
->q
.qlen
> q
->limit
)
461 qdisc_tree_decrease_qlen(sch
, qlen
- sch
->q
.qlen
);
463 del_timer(&q
->perturb_timer
);
464 if (q
->perturb_period
) {
465 mod_timer(&q
->perturb_timer
, jiffies
+ q
->perturb_period
);
466 q
->perturbation
= net_random();
468 sch_tree_unlock(sch
);
472 static int sfq_init(struct Qdisc
*sch
, struct nlattr
*opt
)
474 struct sfq_sched_data
*q
= qdisc_priv(sch
);
478 q
->perturb_timer
.function
= sfq_perturbation
;
479 q
->perturb_timer
.data
= (unsigned long)sch
;
480 init_timer_deferrable(&q
->perturb_timer
);
482 for (i
= 0; i
< SFQ_DEPTH
; i
++) {
483 q
->dep
[i
].next
= i
+ SFQ_SLOTS
;
484 q
->dep
[i
].prev
= i
+ SFQ_SLOTS
;
487 q
->limit
= SFQ_DEPTH
- 1;
490 q
->divisor
= SFQ_DEFAULT_HASH_DIVISOR
;
492 q
->quantum
= psched_mtu(qdisc_dev(sch
));
493 q
->scaled_quantum
= SFQ_ALLOT_SIZE(q
->quantum
);
494 q
->perturb_period
= 0;
495 q
->perturbation
= net_random();
497 int err
= sfq_change(sch
, opt
);
502 sz
= sizeof(q
->ht
[0]) * q
->divisor
;
503 q
->ht
= kmalloc(sz
, GFP_KERNEL
);
504 if (!q
->ht
&& sz
> PAGE_SIZE
)
508 for (i
= 0; i
< q
->divisor
; i
++)
509 q
->ht
[i
] = SFQ_EMPTY_SLOT
;
511 for (i
= 0; i
< SFQ_SLOTS
; i
++) {
512 slot_queue_init(&q
->slots
[i
]);
516 sch
->flags
|= TCQ_F_CAN_BYPASS
;
518 sch
->flags
&= ~TCQ_F_CAN_BYPASS
;
522 static void sfq_destroy(struct Qdisc
*sch
)
524 struct sfq_sched_data
*q
= qdisc_priv(sch
);
526 tcf_destroy_chain(&q
->filter_list
);
527 q
->perturb_period
= 0;
528 del_timer_sync(&q
->perturb_timer
);
529 if (is_vmalloc_addr(q
->ht
))
535 static int sfq_dump(struct Qdisc
*sch
, struct sk_buff
*skb
)
537 struct sfq_sched_data
*q
= qdisc_priv(sch
);
538 unsigned char *b
= skb_tail_pointer(skb
);
539 struct tc_sfq_qopt opt
;
541 opt
.quantum
= q
->quantum
;
542 opt
.perturb_period
= q
->perturb_period
/ HZ
;
544 opt
.limit
= q
->limit
;
545 opt
.divisor
= q
->divisor
;
546 opt
.flows
= q
->limit
;
548 NLA_PUT(skb
, TCA_OPTIONS
, sizeof(opt
), &opt
);
557 static struct Qdisc
*sfq_leaf(struct Qdisc
*sch
, unsigned long arg
)
562 static unsigned long sfq_get(struct Qdisc
*sch
, u32 classid
)
567 static unsigned long sfq_bind(struct Qdisc
*sch
, unsigned long parent
,
570 /* we cannot bypass queue discipline anymore */
571 sch
->flags
&= ~TCQ_F_CAN_BYPASS
;
575 static void sfq_put(struct Qdisc
*q
, unsigned long cl
)
579 static struct tcf_proto
**sfq_find_tcf(struct Qdisc
*sch
, unsigned long cl
)
581 struct sfq_sched_data
*q
= qdisc_priv(sch
);
585 return &q
->filter_list
;
588 static int sfq_dump_class(struct Qdisc
*sch
, unsigned long cl
,
589 struct sk_buff
*skb
, struct tcmsg
*tcm
)
591 tcm
->tcm_handle
|= TC_H_MIN(cl
);
595 static int sfq_dump_class_stats(struct Qdisc
*sch
, unsigned long cl
,
598 struct sfq_sched_data
*q
= qdisc_priv(sch
);
599 sfq_index idx
= q
->ht
[cl
- 1];
600 struct gnet_stats_queue qs
= { 0 };
601 struct tc_sfq_xstats xstats
= { 0 };
604 if (idx
!= SFQ_EMPTY_SLOT
) {
605 const struct sfq_slot
*slot
= &q
->slots
[idx
];
607 xstats
.allot
= slot
->allot
<< SFQ_ALLOT_SHIFT
;
608 qs
.qlen
= slot
->qlen
;
609 slot_queue_walk(slot
, skb
)
610 qs
.backlog
+= qdisc_pkt_len(skb
);
612 if (gnet_stats_copy_queue(d
, &qs
) < 0)
614 return gnet_stats_copy_app(d
, &xstats
, sizeof(xstats
));
617 static void sfq_walk(struct Qdisc
*sch
, struct qdisc_walker
*arg
)
619 struct sfq_sched_data
*q
= qdisc_priv(sch
);
625 for (i
= 0; i
< q
->divisor
; i
++) {
626 if (q
->ht
[i
] == SFQ_EMPTY_SLOT
||
627 arg
->count
< arg
->skip
) {
631 if (arg
->fn(sch
, i
+ 1, arg
) < 0) {
639 static const struct Qdisc_class_ops sfq_class_ops
= {
643 .tcf_chain
= sfq_find_tcf
,
644 .bind_tcf
= sfq_bind
,
645 .unbind_tcf
= sfq_put
,
646 .dump
= sfq_dump_class
,
647 .dump_stats
= sfq_dump_class_stats
,
651 static struct Qdisc_ops sfq_qdisc_ops __read_mostly
= {
652 .cl_ops
= &sfq_class_ops
,
654 .priv_size
= sizeof(struct sfq_sched_data
),
655 .enqueue
= sfq_enqueue
,
656 .dequeue
= sfq_dequeue
,
657 .peek
= qdisc_peek_dequeued
,
661 .destroy
= sfq_destroy
,
664 .owner
= THIS_MODULE
,
667 static int __init
sfq_module_init(void)
669 return register_qdisc(&sfq_qdisc_ops
);
671 static void __exit
sfq_module_exit(void)
673 unregister_qdisc(&sfq_qdisc_ops
);
675 module_init(sfq_module_init
)
676 module_exit(sfq_module_exit
)
677 MODULE_LICENSE("GPL");