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>
29 #include <net/flow_keys.h>
32 * SFB uses two B[l][n] : L x N arrays of bins (L levels, N bins per level)
33 * This implementation uses L = 8 and N = 16
34 * This permits us to split one 32bit hash (provided per packet by rxhash or
35 * external classifier) into 8 subhashes of 4 bits.
37 #define SFB_BUCKET_SHIFT 4
38 #define SFB_NUMBUCKETS (1 << SFB_BUCKET_SHIFT) /* N bins per Level */
39 #define SFB_BUCKET_MASK (SFB_NUMBUCKETS - 1)
40 #define SFB_LEVELS (32 / SFB_BUCKET_SHIFT) /* L */
42 /* SFB algo uses a virtual queue, named "bin" */
44 u16 qlen
; /* length of virtual queue */
45 u16 p_mark
; /* marking probability */
48 /* We use a double buffering right before hash change
49 * (Section 4.4 of SFB reference : moving hash functions)
52 u32 perturbation
; /* jhash perturbation */
53 struct sfb_bucket bins
[SFB_LEVELS
][SFB_NUMBUCKETS
];
56 struct sfb_sched_data
{
58 struct tcf_proto __rcu
*filter_list
;
59 unsigned long rehash_interval
;
60 unsigned long warmup_time
; /* double buffering warmup time in jiffies */
62 u32 bin_size
; /* maximum queue length per bin */
63 u32 increment
; /* d1 */
64 u32 decrement
; /* d2 */
65 u32 limit
; /* HARD maximal queue length */
69 unsigned long rehash_time
;
70 unsigned long token_time
;
72 u8 slot
; /* current active bins (0 or 1) */
73 bool double_buffering
;
74 struct sfb_bins bins
[2];
81 u32 childdrop
; /* drops in child qdisc */
82 u32 marked
; /* ECN mark */
87 * Each queued skb might be hashed on one or two bins
88 * We store in skb_cb the two hash values.
89 * (A zero value means double buffering was not used)
95 static inline struct sfb_skb_cb
*sfb_skb_cb(const struct sk_buff
*skb
)
97 qdisc_cb_private_validate(skb
, sizeof(struct sfb_skb_cb
));
98 return (struct sfb_skb_cb
*)qdisc_skb_cb(skb
)->data
;
102 * If using 'internal' SFB flow classifier, hash comes from skb rxhash
103 * If using external classifier, hash comes from the classid.
105 static u32
sfb_hash(const struct sk_buff
*skb
, u32 slot
)
107 return sfb_skb_cb(skb
)->hashes
[slot
];
110 /* Probabilities are coded as Q0.16 fixed-point values,
111 * with 0xFFFF representing 65535/65536 (almost 1.0)
112 * Addition and subtraction are saturating in [0, 65535]
114 static u32
prob_plus(u32 p1
, u32 p2
)
118 return min_t(u32
, res
, SFB_MAX_PROB
);
121 static u32
prob_minus(u32 p1
, u32 p2
)
123 return p1
> p2
? p1
- p2
: 0;
126 static void increment_one_qlen(u32 sfbhash
, u32 slot
, struct sfb_sched_data
*q
)
129 struct sfb_bucket
*b
= &q
->bins
[slot
].bins
[0][0];
131 for (i
= 0; i
< SFB_LEVELS
; i
++) {
132 u32 hash
= sfbhash
& SFB_BUCKET_MASK
;
134 sfbhash
>>= SFB_BUCKET_SHIFT
;
135 if (b
[hash
].qlen
< 0xFFFF)
137 b
+= SFB_NUMBUCKETS
; /* next level */
141 static void increment_qlen(const struct sk_buff
*skb
, struct sfb_sched_data
*q
)
145 sfbhash
= sfb_hash(skb
, 0);
147 increment_one_qlen(sfbhash
, 0, q
);
149 sfbhash
= sfb_hash(skb
, 1);
151 increment_one_qlen(sfbhash
, 1, q
);
154 static void decrement_one_qlen(u32 sfbhash
, u32 slot
,
155 struct sfb_sched_data
*q
)
158 struct sfb_bucket
*b
= &q
->bins
[slot
].bins
[0][0];
160 for (i
= 0; i
< SFB_LEVELS
; i
++) {
161 u32 hash
= sfbhash
& SFB_BUCKET_MASK
;
163 sfbhash
>>= SFB_BUCKET_SHIFT
;
164 if (b
[hash
].qlen
> 0)
166 b
+= SFB_NUMBUCKETS
; /* next level */
170 static void decrement_qlen(const struct sk_buff
*skb
, struct sfb_sched_data
*q
)
174 sfbhash
= sfb_hash(skb
, 0);
176 decrement_one_qlen(sfbhash
, 0, q
);
178 sfbhash
= sfb_hash(skb
, 1);
180 decrement_one_qlen(sfbhash
, 1, q
);
183 static void decrement_prob(struct sfb_bucket
*b
, struct sfb_sched_data
*q
)
185 b
->p_mark
= prob_minus(b
->p_mark
, q
->decrement
);
188 static void increment_prob(struct sfb_bucket
*b
, struct sfb_sched_data
*q
)
190 b
->p_mark
= prob_plus(b
->p_mark
, q
->increment
);
193 static void sfb_zero_all_buckets(struct sfb_sched_data
*q
)
195 memset(&q
->bins
, 0, sizeof(q
->bins
));
199 * compute max qlen, max p_mark, and avg p_mark
201 static u32
sfb_compute_qlen(u32
*prob_r
, u32
*avgpm_r
, const struct sfb_sched_data
*q
)
204 u32 qlen
= 0, prob
= 0, totalpm
= 0;
205 const struct sfb_bucket
*b
= &q
->bins
[q
->slot
].bins
[0][0];
207 for (i
= 0; i
< SFB_LEVELS
* SFB_NUMBUCKETS
; i
++) {
210 totalpm
+= b
->p_mark
;
211 if (prob
< b
->p_mark
)
216 *avgpm_r
= totalpm
/ (SFB_LEVELS
* SFB_NUMBUCKETS
);
221 static void sfb_init_perturbation(u32 slot
, struct sfb_sched_data
*q
)
223 q
->bins
[slot
].perturbation
= prandom_u32();
226 static void sfb_swap_slot(struct sfb_sched_data
*q
)
228 sfb_init_perturbation(q
->slot
, q
);
230 q
->double_buffering
= false;
233 /* Non elastic flows are allowed to use part of the bandwidth, expressed
234 * in "penalty_rate" packets per second, with "penalty_burst" burst
236 static bool sfb_rate_limit(struct sk_buff
*skb
, struct sfb_sched_data
*q
)
238 if (q
->penalty_rate
== 0 || q
->penalty_burst
== 0)
241 if (q
->tokens_avail
< 1) {
242 unsigned long age
= min(10UL * HZ
, jiffies
- q
->token_time
);
244 q
->tokens_avail
= (age
* q
->penalty_rate
) / HZ
;
245 if (q
->tokens_avail
> q
->penalty_burst
)
246 q
->tokens_avail
= q
->penalty_burst
;
247 q
->token_time
= jiffies
;
248 if (q
->tokens_avail
< 1)
256 static bool sfb_classify(struct sk_buff
*skb
, struct tcf_proto
*fl
,
257 int *qerr
, u32
*salt
)
259 struct tcf_result res
;
262 result
= tc_classify(skb
, fl
, &res
);
264 #ifdef CONFIG_NET_CLS_ACT
268 *qerr
= NET_XMIT_SUCCESS
| __NET_XMIT_STOLEN
;
273 *salt
= TC_H_MIN(res
.classid
);
279 static int sfb_enqueue(struct sk_buff
*skb
, struct Qdisc
*sch
)
282 struct sfb_sched_data
*q
= qdisc_priv(sch
);
283 struct Qdisc
*child
= q
->qdisc
;
284 struct tcf_proto
*fl
;
288 u32 r
, slot
, salt
, sfbhash
;
289 int ret
= NET_XMIT_SUCCESS
| __NET_XMIT_BYPASS
;
290 struct flow_keys keys
;
292 if (unlikely(sch
->q
.qlen
>= q
->limit
)) {
293 qdisc_qstats_overlimit(sch
);
294 q
->stats
.queuedrop
++;
298 if (q
->rehash_interval
> 0) {
299 unsigned long limit
= q
->rehash_time
+ q
->rehash_interval
;
301 if (unlikely(time_after(jiffies
, limit
))) {
303 q
->rehash_time
= jiffies
;
304 } else if (unlikely(!q
->double_buffering
&& q
->warmup_time
> 0 &&
305 time_after(jiffies
, limit
- q
->warmup_time
))) {
306 q
->double_buffering
= true;
310 fl
= rcu_dereference_bh(q
->filter_list
);
312 /* If using external classifiers, get result and record it. */
313 if (!sfb_classify(skb
, fl
, &ret
, &salt
))
319 skb_flow_dissect(skb
, &keys
);
324 sfbhash
= jhash_3words((__force u32
)keys
.dst
,
325 (__force u32
)keys
.src
,
326 (__force u32
)keys
.ports
,
327 q
->bins
[slot
].perturbation
);
330 sfb_skb_cb(skb
)->hashes
[slot
] = sfbhash
;
332 for (i
= 0; i
< SFB_LEVELS
; i
++) {
333 u32 hash
= sfbhash
& SFB_BUCKET_MASK
;
334 struct sfb_bucket
*b
= &q
->bins
[slot
].bins
[i
][hash
];
336 sfbhash
>>= SFB_BUCKET_SHIFT
;
338 decrement_prob(b
, q
);
339 else if (b
->qlen
>= q
->bin_size
)
340 increment_prob(b
, q
);
341 if (minqlen
> b
->qlen
)
343 if (p_min
> b
->p_mark
)
348 sfb_skb_cb(skb
)->hashes
[slot
] = 0;
350 if (unlikely(minqlen
>= q
->max
)) {
351 qdisc_qstats_overlimit(sch
);
352 q
->stats
.bucketdrop
++;
356 if (unlikely(p_min
>= SFB_MAX_PROB
)) {
358 if (q
->double_buffering
) {
359 sfbhash
= jhash_3words((__force u32
)keys
.dst
,
360 (__force u32
)keys
.src
,
361 (__force u32
)keys
.ports
,
362 q
->bins
[slot
].perturbation
);
365 sfb_skb_cb(skb
)->hashes
[slot
] = sfbhash
;
367 for (i
= 0; i
< SFB_LEVELS
; i
++) {
368 u32 hash
= sfbhash
& SFB_BUCKET_MASK
;
369 struct sfb_bucket
*b
= &q
->bins
[slot
].bins
[i
][hash
];
371 sfbhash
>>= SFB_BUCKET_SHIFT
;
373 decrement_prob(b
, q
);
374 else if (b
->qlen
>= q
->bin_size
)
375 increment_prob(b
, q
);
378 if (sfb_rate_limit(skb
, q
)) {
379 qdisc_qstats_overlimit(sch
);
380 q
->stats
.penaltydrop
++;
386 r
= prandom_u32() & SFB_MAX_PROB
;
388 if (unlikely(r
< p_min
)) {
389 if (unlikely(p_min
> SFB_MAX_PROB
/ 2)) {
390 /* If we're marking that many packets, then either
391 * this flow is unresponsive, or we're badly congested.
392 * In either case, we want to start dropping packets.
394 if (r
< (p_min
- SFB_MAX_PROB
/ 2) * 2) {
395 q
->stats
.earlydrop
++;
399 if (INET_ECN_set_ce(skb
)) {
402 q
->stats
.earlydrop
++;
408 ret
= qdisc_enqueue(skb
, child
);
409 if (likely(ret
== NET_XMIT_SUCCESS
)) {
411 increment_qlen(skb
, q
);
412 } else if (net_xmit_drop_count(ret
)) {
413 q
->stats
.childdrop
++;
414 qdisc_qstats_drop(sch
);
419 qdisc_drop(skb
, sch
);
422 if (ret
& __NET_XMIT_BYPASS
)
423 qdisc_qstats_drop(sch
);
428 static struct sk_buff
*sfb_dequeue(struct Qdisc
*sch
)
430 struct sfb_sched_data
*q
= qdisc_priv(sch
);
431 struct Qdisc
*child
= q
->qdisc
;
434 skb
= child
->dequeue(q
->qdisc
);
437 qdisc_bstats_update(sch
, skb
);
439 decrement_qlen(skb
, q
);
445 static struct sk_buff
*sfb_peek(struct Qdisc
*sch
)
447 struct sfb_sched_data
*q
= qdisc_priv(sch
);
448 struct Qdisc
*child
= q
->qdisc
;
450 return child
->ops
->peek(child
);
453 /* No sfb_drop -- impossible since the child doesn't return the dropped skb. */
455 static void sfb_reset(struct Qdisc
*sch
)
457 struct sfb_sched_data
*q
= qdisc_priv(sch
);
459 qdisc_reset(q
->qdisc
);
462 q
->double_buffering
= false;
463 sfb_zero_all_buckets(q
);
464 sfb_init_perturbation(0, q
);
467 static void sfb_destroy(struct Qdisc
*sch
)
469 struct sfb_sched_data
*q
= qdisc_priv(sch
);
471 tcf_destroy_chain(&q
->filter_list
);
472 qdisc_destroy(q
->qdisc
);
475 static const struct nla_policy sfb_policy
[TCA_SFB_MAX
+ 1] = {
476 [TCA_SFB_PARMS
] = { .len
= sizeof(struct tc_sfb_qopt
) },
479 static const struct tc_sfb_qopt sfb_default_ops
= {
480 .rehash_interval
= 600 * MSEC_PER_SEC
,
481 .warmup_time
= 60 * MSEC_PER_SEC
,
485 .increment
= (SFB_MAX_PROB
+ 500) / 1000, /* 0.1 % */
486 .decrement
= (SFB_MAX_PROB
+ 3000) / 6000,
491 static int sfb_change(struct Qdisc
*sch
, struct nlattr
*opt
)
493 struct sfb_sched_data
*q
= qdisc_priv(sch
);
495 struct nlattr
*tb
[TCA_SFB_MAX
+ 1];
496 const struct tc_sfb_qopt
*ctl
= &sfb_default_ops
;
501 err
= nla_parse_nested(tb
, TCA_SFB_MAX
, opt
, sfb_policy
);
505 if (tb
[TCA_SFB_PARMS
] == NULL
)
508 ctl
= nla_data(tb
[TCA_SFB_PARMS
]);
513 limit
= max_t(u32
, qdisc_dev(sch
)->tx_queue_len
, 1);
515 child
= fifo_create_dflt(sch
, &pfifo_qdisc_ops
, limit
);
517 return PTR_ERR(child
);
521 qdisc_tree_decrease_qlen(q
->qdisc
, q
->qdisc
->q
.qlen
);
522 qdisc_destroy(q
->qdisc
);
525 q
->rehash_interval
= msecs_to_jiffies(ctl
->rehash_interval
);
526 q
->warmup_time
= msecs_to_jiffies(ctl
->warmup_time
);
527 q
->rehash_time
= jiffies
;
529 q
->increment
= ctl
->increment
;
530 q
->decrement
= ctl
->decrement
;
532 q
->bin_size
= ctl
->bin_size
;
533 q
->penalty_rate
= ctl
->penalty_rate
;
534 q
->penalty_burst
= ctl
->penalty_burst
;
535 q
->tokens_avail
= ctl
->penalty_burst
;
536 q
->token_time
= jiffies
;
539 q
->double_buffering
= false;
540 sfb_zero_all_buckets(q
);
541 sfb_init_perturbation(0, q
);
542 sfb_init_perturbation(1, q
);
544 sch_tree_unlock(sch
);
549 static int sfb_init(struct Qdisc
*sch
, struct nlattr
*opt
)
551 struct sfb_sched_data
*q
= qdisc_priv(sch
);
553 q
->qdisc
= &noop_qdisc
;
554 return sfb_change(sch
, opt
);
557 static int sfb_dump(struct Qdisc
*sch
, struct sk_buff
*skb
)
559 struct sfb_sched_data
*q
= qdisc_priv(sch
);
561 struct tc_sfb_qopt opt
= {
562 .rehash_interval
= jiffies_to_msecs(q
->rehash_interval
),
563 .warmup_time
= jiffies_to_msecs(q
->warmup_time
),
566 .bin_size
= q
->bin_size
,
567 .increment
= q
->increment
,
568 .decrement
= q
->decrement
,
569 .penalty_rate
= q
->penalty_rate
,
570 .penalty_burst
= q
->penalty_burst
,
573 sch
->qstats
.backlog
= q
->qdisc
->qstats
.backlog
;
574 opts
= nla_nest_start(skb
, TCA_OPTIONS
);
576 goto nla_put_failure
;
577 if (nla_put(skb
, TCA_SFB_PARMS
, sizeof(opt
), &opt
))
578 goto nla_put_failure
;
579 return nla_nest_end(skb
, opts
);
582 nla_nest_cancel(skb
, opts
);
586 static int sfb_dump_stats(struct Qdisc
*sch
, struct gnet_dump
*d
)
588 struct sfb_sched_data
*q
= qdisc_priv(sch
);
589 struct tc_sfb_xstats st
= {
590 .earlydrop
= q
->stats
.earlydrop
,
591 .penaltydrop
= q
->stats
.penaltydrop
,
592 .bucketdrop
= q
->stats
.bucketdrop
,
593 .queuedrop
= q
->stats
.queuedrop
,
594 .childdrop
= q
->stats
.childdrop
,
595 .marked
= q
->stats
.marked
,
598 st
.maxqlen
= sfb_compute_qlen(&st
.maxprob
, &st
.avgprob
, q
);
600 return gnet_stats_copy_app(d
, &st
, sizeof(st
));
603 static int sfb_dump_class(struct Qdisc
*sch
, unsigned long cl
,
604 struct sk_buff
*skb
, struct tcmsg
*tcm
)
609 static int sfb_graft(struct Qdisc
*sch
, unsigned long arg
, struct Qdisc
*new,
612 struct sfb_sched_data
*q
= qdisc_priv(sch
);
620 qdisc_tree_decrease_qlen(*old
, (*old
)->q
.qlen
);
622 sch_tree_unlock(sch
);
626 static struct Qdisc
*sfb_leaf(struct Qdisc
*sch
, unsigned long arg
)
628 struct sfb_sched_data
*q
= qdisc_priv(sch
);
633 static unsigned long sfb_get(struct Qdisc
*sch
, u32 classid
)
638 static void sfb_put(struct Qdisc
*sch
, unsigned long arg
)
642 static int sfb_change_class(struct Qdisc
*sch
, u32 classid
, u32 parentid
,
643 struct nlattr
**tca
, unsigned long *arg
)
648 static int sfb_delete(struct Qdisc
*sch
, unsigned long cl
)
653 static void sfb_walk(struct Qdisc
*sch
, struct qdisc_walker
*walker
)
656 if (walker
->count
>= walker
->skip
)
657 if (walker
->fn(sch
, 1, walker
) < 0) {
665 static struct tcf_proto __rcu
**sfb_find_tcf(struct Qdisc
*sch
,
668 struct sfb_sched_data
*q
= qdisc_priv(sch
);
672 return &q
->filter_list
;
675 static unsigned long sfb_bind(struct Qdisc
*sch
, unsigned long parent
,
682 static const struct Qdisc_class_ops sfb_class_ops
= {
687 .change
= sfb_change_class
,
688 .delete = sfb_delete
,
690 .tcf_chain
= sfb_find_tcf
,
691 .bind_tcf
= sfb_bind
,
692 .unbind_tcf
= sfb_put
,
693 .dump
= sfb_dump_class
,
696 static struct Qdisc_ops sfb_qdisc_ops __read_mostly
= {
698 .priv_size
= sizeof(struct sfb_sched_data
),
699 .cl_ops
= &sfb_class_ops
,
700 .enqueue
= sfb_enqueue
,
701 .dequeue
= sfb_dequeue
,
705 .destroy
= sfb_destroy
,
706 .change
= sfb_change
,
708 .dump_stats
= sfb_dump_stats
,
709 .owner
= THIS_MODULE
,
712 static int __init
sfb_module_init(void)
714 return register_qdisc(&sfb_qdisc_ops
);
717 static void __exit
sfb_module_exit(void)
719 unregister_qdisc(&sfb_qdisc_ops
);
722 module_init(sfb_module_init
)
723 module_exit(sfb_module_exit
)
725 MODULE_DESCRIPTION("Stochastic Fair Blue queue discipline");
726 MODULE_AUTHOR("Juliusz Chroboczek");
727 MODULE_AUTHOR("Eric Dumazet");
728 MODULE_LICENSE("GPL");