2 * net/sched/sch_sfb.c Stochastic Fair Blue
4 * Copyright (c) 2008-2011 Juliusz Chroboczek <jch@pps.jussieu.fr>
5 * Copyright (c) 2011 Eric Dumazet <eric.dumazet@gmail.com>
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * version 2 as published by the Free Software Foundation.
11 * W. Feng, D. Kandlur, D. Saha, K. Shin. Blue:
12 * A New Class of Active Queue Management Algorithms.
13 * U. Michigan CSE-TR-387-99, April 1999.
15 * http://www.thefengs.com/wuchang/blue/CSE-TR-387-99.pdf
19 #include <linux/module.h>
20 #include <linux/types.h>
21 #include <linux/kernel.h>
22 #include <linux/errno.h>
23 #include <linux/skbuff.h>
24 #include <linux/random.h>
25 #include <linux/jhash.h>
27 #include <net/pkt_sched.h>
28 #include <net/inet_ecn.h>
31 * SFB uses two B[l][n] : L x N arrays of bins (L levels, N bins per level)
32 * This implementation uses L = 8 and N = 16
33 * This permits us to split one 32bit hash (provided per packet by rxhash or
34 * external classifier) into 8 subhashes of 4 bits.
36 #define SFB_BUCKET_SHIFT 4
37 #define SFB_NUMBUCKETS (1 << SFB_BUCKET_SHIFT) /* N bins per Level */
38 #define SFB_BUCKET_MASK (SFB_NUMBUCKETS - 1)
39 #define SFB_LEVELS (32 / SFB_BUCKET_SHIFT) /* L */
41 /* SFB algo uses a virtual queue, named "bin" */
43 u16 qlen
; /* length of virtual queue */
44 u16 p_mark
; /* marking probability */
47 /* We use a double buffering right before hash change
48 * (Section 4.4 of SFB reference : moving hash functions)
51 u32 perturbation
; /* jhash perturbation */
52 struct sfb_bucket bins
[SFB_LEVELS
][SFB_NUMBUCKETS
];
55 struct sfb_sched_data
{
57 struct tcf_proto
*filter_list
;
58 unsigned long rehash_interval
;
59 unsigned long warmup_time
; /* double buffering warmup time in jiffies */
61 u32 bin_size
; /* maximum queue length per bin */
62 u32 increment
; /* d1 */
63 u32 decrement
; /* d2 */
64 u32 limit
; /* HARD maximal queue length */
68 unsigned long rehash_time
;
69 unsigned long token_time
;
71 u8 slot
; /* current active bins (0 or 1) */
72 bool double_buffering
;
73 struct sfb_bins bins
[2];
80 u32 childdrop
; /* drops in child qdisc */
81 u32 marked
; /* ECN mark */
86 * Each queued skb might be hashed on one or two bins
87 * We store in skb_cb the two hash values.
88 * (A zero value means double buffering was not used)
94 static inline struct sfb_skb_cb
*sfb_skb_cb(const struct sk_buff
*skb
)
96 qdisc_cb_private_validate(skb
, sizeof(struct sfb_skb_cb
));
97 return (struct sfb_skb_cb
*)qdisc_skb_cb(skb
)->data
;
101 * If using 'internal' SFB flow classifier, hash comes from skb rxhash
102 * If using external classifier, hash comes from the classid.
104 static u32
sfb_hash(const struct sk_buff
*skb
, u32 slot
)
106 return sfb_skb_cb(skb
)->hashes
[slot
];
109 /* Probabilities are coded as Q0.16 fixed-point values,
110 * with 0xFFFF representing 65535/65536 (almost 1.0)
111 * Addition and subtraction are saturating in [0, 65535]
113 static u32
prob_plus(u32 p1
, u32 p2
)
117 return min_t(u32
, res
, SFB_MAX_PROB
);
120 static u32
prob_minus(u32 p1
, u32 p2
)
122 return p1
> p2
? p1
- p2
: 0;
125 static void increment_one_qlen(u32 sfbhash
, u32 slot
, struct sfb_sched_data
*q
)
128 struct sfb_bucket
*b
= &q
->bins
[slot
].bins
[0][0];
130 for (i
= 0; i
< SFB_LEVELS
; i
++) {
131 u32 hash
= sfbhash
& SFB_BUCKET_MASK
;
133 sfbhash
>>= SFB_BUCKET_SHIFT
;
134 if (b
[hash
].qlen
< 0xFFFF)
136 b
+= SFB_NUMBUCKETS
; /* next level */
140 static void increment_qlen(const struct sk_buff
*skb
, struct sfb_sched_data
*q
)
144 sfbhash
= sfb_hash(skb
, 0);
146 increment_one_qlen(sfbhash
, 0, q
);
148 sfbhash
= sfb_hash(skb
, 1);
150 increment_one_qlen(sfbhash
, 1, q
);
153 static void decrement_one_qlen(u32 sfbhash
, u32 slot
,
154 struct sfb_sched_data
*q
)
157 struct sfb_bucket
*b
= &q
->bins
[slot
].bins
[0][0];
159 for (i
= 0; i
< SFB_LEVELS
; i
++) {
160 u32 hash
= sfbhash
& SFB_BUCKET_MASK
;
162 sfbhash
>>= SFB_BUCKET_SHIFT
;
163 if (b
[hash
].qlen
> 0)
165 b
+= SFB_NUMBUCKETS
; /* next level */
169 static void decrement_qlen(const struct sk_buff
*skb
, struct sfb_sched_data
*q
)
173 sfbhash
= sfb_hash(skb
, 0);
175 decrement_one_qlen(sfbhash
, 0, q
);
177 sfbhash
= sfb_hash(skb
, 1);
179 decrement_one_qlen(sfbhash
, 1, q
);
182 static void decrement_prob(struct sfb_bucket
*b
, struct sfb_sched_data
*q
)
184 b
->p_mark
= prob_minus(b
->p_mark
, q
->decrement
);
187 static void increment_prob(struct sfb_bucket
*b
, struct sfb_sched_data
*q
)
189 b
->p_mark
= prob_plus(b
->p_mark
, q
->increment
);
192 static void sfb_zero_all_buckets(struct sfb_sched_data
*q
)
194 memset(&q
->bins
, 0, sizeof(q
->bins
));
198 * compute max qlen, max p_mark, and avg p_mark
200 static u32
sfb_compute_qlen(u32
*prob_r
, u32
*avgpm_r
, const struct sfb_sched_data
*q
)
203 u32 qlen
= 0, prob
= 0, totalpm
= 0;
204 const struct sfb_bucket
*b
= &q
->bins
[q
->slot
].bins
[0][0];
206 for (i
= 0; i
< SFB_LEVELS
* SFB_NUMBUCKETS
; i
++) {
209 totalpm
+= b
->p_mark
;
210 if (prob
< b
->p_mark
)
215 *avgpm_r
= totalpm
/ (SFB_LEVELS
* SFB_NUMBUCKETS
);
220 static void sfb_init_perturbation(u32 slot
, struct sfb_sched_data
*q
)
222 q
->bins
[slot
].perturbation
= net_random();
225 static void sfb_swap_slot(struct sfb_sched_data
*q
)
227 sfb_init_perturbation(q
->slot
, q
);
229 q
->double_buffering
= false;
232 /* Non elastic flows are allowed to use part of the bandwidth, expressed
233 * in "penalty_rate" packets per second, with "penalty_burst" burst
235 static bool sfb_rate_limit(struct sk_buff
*skb
, struct sfb_sched_data
*q
)
237 if (q
->penalty_rate
== 0 || q
->penalty_burst
== 0)
240 if (q
->tokens_avail
< 1) {
241 unsigned long age
= min(10UL * HZ
, jiffies
- q
->token_time
);
243 q
->tokens_avail
= (age
* q
->penalty_rate
) / HZ
;
244 if (q
->tokens_avail
> q
->penalty_burst
)
245 q
->tokens_avail
= q
->penalty_burst
;
246 q
->token_time
= jiffies
;
247 if (q
->tokens_avail
< 1)
255 static bool sfb_classify(struct sk_buff
*skb
, struct sfb_sched_data
*q
,
256 int *qerr
, u32
*salt
)
258 struct tcf_result res
;
261 result
= tc_classify(skb
, q
->filter_list
, &res
);
263 #ifdef CONFIG_NET_CLS_ACT
267 *qerr
= NET_XMIT_SUCCESS
| __NET_XMIT_STOLEN
;
272 *salt
= TC_H_MIN(res
.classid
);
278 static int sfb_enqueue(struct sk_buff
*skb
, struct Qdisc
*sch
)
281 struct sfb_sched_data
*q
= qdisc_priv(sch
);
282 struct Qdisc
*child
= q
->qdisc
;
286 u32 r
, slot
, salt
, sfbhash
;
287 int ret
= NET_XMIT_SUCCESS
| __NET_XMIT_BYPASS
;
289 if (q
->rehash_interval
> 0) {
290 unsigned long limit
= q
->rehash_time
+ q
->rehash_interval
;
292 if (unlikely(time_after(jiffies
, limit
))) {
294 q
->rehash_time
= jiffies
;
295 } else if (unlikely(!q
->double_buffering
&& q
->warmup_time
> 0 &&
296 time_after(jiffies
, limit
- q
->warmup_time
))) {
297 q
->double_buffering
= true;
301 if (q
->filter_list
) {
302 /* If using external classifiers, get result and record it. */
303 if (!sfb_classify(skb
, q
, &ret
, &salt
))
306 salt
= skb_get_rxhash(skb
);
311 sfbhash
= jhash_1word(salt
, q
->bins
[slot
].perturbation
);
314 sfb_skb_cb(skb
)->hashes
[slot
] = sfbhash
;
316 for (i
= 0; i
< SFB_LEVELS
; i
++) {
317 u32 hash
= sfbhash
& SFB_BUCKET_MASK
;
318 struct sfb_bucket
*b
= &q
->bins
[slot
].bins
[i
][hash
];
320 sfbhash
>>= SFB_BUCKET_SHIFT
;
322 decrement_prob(b
, q
);
323 else if (b
->qlen
>= q
->bin_size
)
324 increment_prob(b
, q
);
325 if (minqlen
> b
->qlen
)
327 if (p_min
> b
->p_mark
)
332 sfb_skb_cb(skb
)->hashes
[slot
] = 0;
334 if (unlikely(minqlen
>= q
->max
|| sch
->q
.qlen
>= q
->limit
)) {
335 sch
->qstats
.overlimits
++;
336 if (minqlen
>= q
->max
)
337 q
->stats
.bucketdrop
++;
339 q
->stats
.queuedrop
++;
343 if (unlikely(p_min
>= SFB_MAX_PROB
)) {
345 if (q
->double_buffering
) {
346 sfbhash
= jhash_1word(salt
, q
->bins
[slot
].perturbation
);
349 sfb_skb_cb(skb
)->hashes
[slot
] = sfbhash
;
351 for (i
= 0; i
< SFB_LEVELS
; i
++) {
352 u32 hash
= sfbhash
& SFB_BUCKET_MASK
;
353 struct sfb_bucket
*b
= &q
->bins
[slot
].bins
[i
][hash
];
355 sfbhash
>>= SFB_BUCKET_SHIFT
;
357 decrement_prob(b
, q
);
358 else if (b
->qlen
>= q
->bin_size
)
359 increment_prob(b
, q
);
362 if (sfb_rate_limit(skb
, q
)) {
363 sch
->qstats
.overlimits
++;
364 q
->stats
.penaltydrop
++;
370 r
= net_random() & SFB_MAX_PROB
;
372 if (unlikely(r
< p_min
)) {
373 if (unlikely(p_min
> SFB_MAX_PROB
/ 2)) {
374 /* If we're marking that many packets, then either
375 * this flow is unresponsive, or we're badly congested.
376 * In either case, we want to start dropping packets.
378 if (r
< (p_min
- SFB_MAX_PROB
/ 2) * 2) {
379 q
->stats
.earlydrop
++;
383 if (INET_ECN_set_ce(skb
)) {
386 q
->stats
.earlydrop
++;
392 ret
= qdisc_enqueue(skb
, child
);
393 if (likely(ret
== NET_XMIT_SUCCESS
)) {
395 increment_qlen(skb
, q
);
396 } else if (net_xmit_drop_count(ret
)) {
397 q
->stats
.childdrop
++;
403 qdisc_drop(skb
, sch
);
406 if (ret
& __NET_XMIT_BYPASS
)
412 static struct sk_buff
*sfb_dequeue(struct Qdisc
*sch
)
414 struct sfb_sched_data
*q
= qdisc_priv(sch
);
415 struct Qdisc
*child
= q
->qdisc
;
418 skb
= child
->dequeue(q
->qdisc
);
421 qdisc_bstats_update(sch
, skb
);
423 decrement_qlen(skb
, q
);
429 static struct sk_buff
*sfb_peek(struct Qdisc
*sch
)
431 struct sfb_sched_data
*q
= qdisc_priv(sch
);
432 struct Qdisc
*child
= q
->qdisc
;
434 return child
->ops
->peek(child
);
437 /* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
439 static void sfb_reset(struct Qdisc
*sch
)
441 struct sfb_sched_data
*q
= qdisc_priv(sch
);
443 qdisc_reset(q
->qdisc
);
446 q
->double_buffering
= false;
447 sfb_zero_all_buckets(q
);
448 sfb_init_perturbation(0, q
);
451 static void sfb_destroy(struct Qdisc
*sch
)
453 struct sfb_sched_data
*q
= qdisc_priv(sch
);
455 tcf_destroy_chain(&q
->filter_list
);
456 qdisc_destroy(q
->qdisc
);
459 static const struct nla_policy sfb_policy
[TCA_SFB_MAX
+ 1] = {
460 [TCA_SFB_PARMS
] = { .len
= sizeof(struct tc_sfb_qopt
) },
463 static const struct tc_sfb_qopt sfb_default_ops
= {
464 .rehash_interval
= 600 * MSEC_PER_SEC
,
465 .warmup_time
= 60 * MSEC_PER_SEC
,
469 .increment
= (SFB_MAX_PROB
+ 500) / 1000, /* 0.1 % */
470 .decrement
= (SFB_MAX_PROB
+ 3000) / 6000,
475 static int sfb_change(struct Qdisc
*sch
, struct nlattr
*opt
)
477 struct sfb_sched_data
*q
= qdisc_priv(sch
);
479 struct nlattr
*tb
[TCA_SFB_MAX
+ 1];
480 const struct tc_sfb_qopt
*ctl
= &sfb_default_ops
;
485 err
= nla_parse_nested(tb
, TCA_SFB_MAX
, opt
, sfb_policy
);
489 if (tb
[TCA_SFB_PARMS
] == NULL
)
492 ctl
= nla_data(tb
[TCA_SFB_PARMS
]);
497 limit
= max_t(u32
, qdisc_dev(sch
)->tx_queue_len
, 1);
499 child
= fifo_create_dflt(sch
, &pfifo_qdisc_ops
, limit
);
501 return PTR_ERR(child
);
505 qdisc_tree_decrease_qlen(q
->qdisc
, q
->qdisc
->q
.qlen
);
506 qdisc_destroy(q
->qdisc
);
509 q
->rehash_interval
= msecs_to_jiffies(ctl
->rehash_interval
);
510 q
->warmup_time
= msecs_to_jiffies(ctl
->warmup_time
);
511 q
->rehash_time
= jiffies
;
513 q
->increment
= ctl
->increment
;
514 q
->decrement
= ctl
->decrement
;
516 q
->bin_size
= ctl
->bin_size
;
517 q
->penalty_rate
= ctl
->penalty_rate
;
518 q
->penalty_burst
= ctl
->penalty_burst
;
519 q
->tokens_avail
= ctl
->penalty_burst
;
520 q
->token_time
= jiffies
;
523 q
->double_buffering
= false;
524 sfb_zero_all_buckets(q
);
525 sfb_init_perturbation(0, q
);
526 sfb_init_perturbation(1, q
);
528 sch_tree_unlock(sch
);
533 static int sfb_init(struct Qdisc
*sch
, struct nlattr
*opt
)
535 struct sfb_sched_data
*q
= qdisc_priv(sch
);
537 q
->qdisc
= &noop_qdisc
;
538 return sfb_change(sch
, opt
);
541 static int sfb_dump(struct Qdisc
*sch
, struct sk_buff
*skb
)
543 struct sfb_sched_data
*q
= qdisc_priv(sch
);
545 struct tc_sfb_qopt opt
= {
546 .rehash_interval
= jiffies_to_msecs(q
->rehash_interval
),
547 .warmup_time
= jiffies_to_msecs(q
->warmup_time
),
550 .bin_size
= q
->bin_size
,
551 .increment
= q
->increment
,
552 .decrement
= q
->decrement
,
553 .penalty_rate
= q
->penalty_rate
,
554 .penalty_burst
= q
->penalty_burst
,
557 sch
->qstats
.backlog
= q
->qdisc
->qstats
.backlog
;
558 opts
= nla_nest_start(skb
, TCA_OPTIONS
);
559 NLA_PUT(skb
, TCA_SFB_PARMS
, sizeof(opt
), &opt
);
560 return nla_nest_end(skb
, opts
);
563 nla_nest_cancel(skb
, opts
);
567 static int sfb_dump_stats(struct Qdisc
*sch
, struct gnet_dump
*d
)
569 struct sfb_sched_data
*q
= qdisc_priv(sch
);
570 struct tc_sfb_xstats st
= {
571 .earlydrop
= q
->stats
.earlydrop
,
572 .penaltydrop
= q
->stats
.penaltydrop
,
573 .bucketdrop
= q
->stats
.bucketdrop
,
574 .queuedrop
= q
->stats
.queuedrop
,
575 .childdrop
= q
->stats
.childdrop
,
576 .marked
= q
->stats
.marked
,
579 st
.maxqlen
= sfb_compute_qlen(&st
.maxprob
, &st
.avgprob
, q
);
581 return gnet_stats_copy_app(d
, &st
, sizeof(st
));
584 static int sfb_dump_class(struct Qdisc
*sch
, unsigned long cl
,
585 struct sk_buff
*skb
, struct tcmsg
*tcm
)
590 static int sfb_graft(struct Qdisc
*sch
, unsigned long arg
, struct Qdisc
*new,
593 struct sfb_sched_data
*q
= qdisc_priv(sch
);
601 qdisc_tree_decrease_qlen(*old
, (*old
)->q
.qlen
);
603 sch_tree_unlock(sch
);
607 static struct Qdisc
*sfb_leaf(struct Qdisc
*sch
, unsigned long arg
)
609 struct sfb_sched_data
*q
= qdisc_priv(sch
);
614 static unsigned long sfb_get(struct Qdisc
*sch
, u32 classid
)
619 static void sfb_put(struct Qdisc
*sch
, unsigned long arg
)
623 static int sfb_change_class(struct Qdisc
*sch
, u32 classid
, u32 parentid
,
624 struct nlattr
**tca
, unsigned long *arg
)
629 static int sfb_delete(struct Qdisc
*sch
, unsigned long cl
)
634 static void sfb_walk(struct Qdisc
*sch
, struct qdisc_walker
*walker
)
637 if (walker
->count
>= walker
->skip
)
638 if (walker
->fn(sch
, 1, walker
) < 0) {
646 static struct tcf_proto
**sfb_find_tcf(struct Qdisc
*sch
, unsigned long cl
)
648 struct sfb_sched_data
*q
= qdisc_priv(sch
);
652 return &q
->filter_list
;
655 static unsigned long sfb_bind(struct Qdisc
*sch
, unsigned long parent
,
662 static const struct Qdisc_class_ops sfb_class_ops
= {
667 .change
= sfb_change_class
,
668 .delete = sfb_delete
,
670 .tcf_chain
= sfb_find_tcf
,
671 .bind_tcf
= sfb_bind
,
672 .unbind_tcf
= sfb_put
,
673 .dump
= sfb_dump_class
,
676 static struct Qdisc_ops sfb_qdisc_ops __read_mostly
= {
678 .priv_size
= sizeof(struct sfb_sched_data
),
679 .cl_ops
= &sfb_class_ops
,
680 .enqueue
= sfb_enqueue
,
681 .dequeue
= sfb_dequeue
,
685 .destroy
= sfb_destroy
,
686 .change
= sfb_change
,
688 .dump_stats
= sfb_dump_stats
,
689 .owner
= THIS_MODULE
,
692 static int __init
sfb_module_init(void)
694 return register_qdisc(&sfb_qdisc_ops
);
697 static void __exit
sfb_module_exit(void)
699 unregister_qdisc(&sfb_qdisc_ops
);
702 module_init(sfb_module_init
)
703 module_exit(sfb_module_exit
)
705 MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
706 MODULE_AUTHOR("Juliusz Chroboczek");
707 MODULE_AUTHOR("Eric Dumazet");
708 MODULE_LICENSE("GPL");