2 * net/sched/sch_choke.c CHOKE scheduler
4 * Copyright (c) 2011 Stephen Hemminger <shemminger@vyatta.com>
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.
13 #include <linux/module.h>
14 #include <linux/types.h>
15 #include <linux/kernel.h>
16 #include <linux/skbuff.h>
17 #include <linux/vmalloc.h>
18 #include <net/pkt_sched.h>
19 #include <net/pkt_cls.h>
20 #include <net/inet_ecn.h>
22 #include <net/flow_dissector.h>
25 CHOKe stateless AQM for fair bandwidth allocation
26 =================================================
28 CHOKe (CHOose and Keep for responsive flows, CHOose and Kill for
29 unresponsive flows) is a variant of RED that penalizes misbehaving flows but
30 maintains no flow state. The difference from RED is an additional step
31 during the enqueuing process. If average queue size is over the
32 low threshold (qmin), a packet is chosen at random from the queue.
33 If both the new and chosen packet are from the same flow, both
34 are dropped. Unlike RED, CHOKe is not really a "classful" qdisc because it
35 needs to access packets in queue randomly. It has a minimal class
36 interface to allow overriding the builtin flow classifier with
40 R. Pan, B. Prabhakar, and K. Psounis, "CHOKe, A Stateless
41 Active Queue Management Scheme for Approximating Fair Bandwidth Allocation",
44 A. Tang, J. Wang, S. Low, "Understanding CHOKe: Throughput and Spatial
45 Characteristics", IEEE/ACM Transactions on Networking, 2004
49 /* Upper bound on size of sk_buff table (packets) */
50 #define CHOKE_MAX_QUEUE (128*1024 - 1)
52 struct choke_sched_data
{
57 struct red_parms parms
;
62 u32 prob_drop
; /* Early probability drops */
63 u32 prob_mark
; /* Early probability marks */
64 u32 forced_drop
; /* Forced drops, qavg > max_thresh */
65 u32 forced_mark
; /* Forced marks, qavg > max_thresh */
66 u32 pdrop
; /* Drops due to queue limits */
67 u32 other
; /* Drops due to drop() calls */
68 u32 matched
; /* Drops to flow match */
74 unsigned int tab_mask
; /* size - 1 */
79 /* number of elements in queue including holes */
80 static unsigned int choke_len(const struct choke_sched_data
*q
)
82 return (q
->tail
- q
->head
) & q
->tab_mask
;
85 /* Is ECN parameter configured */
86 static int use_ecn(const struct choke_sched_data
*q
)
88 return q
->flags
& TC_RED_ECN
;
91 /* Should packets over max just be dropped (versus marked) */
92 static int use_harddrop(const struct choke_sched_data
*q
)
94 return q
->flags
& TC_RED_HARDDROP
;
97 /* Move head pointer forward to skip over holes */
98 static void choke_zap_head_holes(struct choke_sched_data
*q
)
101 q
->head
= (q
->head
+ 1) & q
->tab_mask
;
102 if (q
->head
== q
->tail
)
104 } while (q
->tab
[q
->head
] == NULL
);
107 /* Move tail pointer backwards to reuse holes */
108 static void choke_zap_tail_holes(struct choke_sched_data
*q
)
111 q
->tail
= (q
->tail
- 1) & q
->tab_mask
;
112 if (q
->head
== q
->tail
)
114 } while (q
->tab
[q
->tail
] == NULL
);
117 /* Drop packet from queue array by creating a "hole" */
118 static void choke_drop_by_idx(struct Qdisc
*sch
, unsigned int idx
,
119 struct sk_buff
**to_free
)
121 struct choke_sched_data
*q
= qdisc_priv(sch
);
122 struct sk_buff
*skb
= q
->tab
[idx
];
127 choke_zap_head_holes(q
);
129 choke_zap_tail_holes(q
);
131 qdisc_qstats_backlog_dec(sch
, skb
);
132 qdisc_tree_reduce_backlog(sch
, 1, qdisc_pkt_len(skb
));
133 qdisc_drop(skb
, sch
, to_free
);
137 struct choke_skb_cb
{
140 struct flow_keys_digest keys
;
143 static inline struct choke_skb_cb
*choke_skb_cb(const struct sk_buff
*skb
)
145 qdisc_cb_private_validate(skb
, sizeof(struct choke_skb_cb
));
146 return (struct choke_skb_cb
*)qdisc_skb_cb(skb
)->data
;
149 static inline void choke_set_classid(struct sk_buff
*skb
, u16 classid
)
151 choke_skb_cb(skb
)->classid
= classid
;
155 * Compare flow of two packets
156 * Returns true only if source and destination address and port match.
157 * false for special cases
159 static bool choke_match_flow(struct sk_buff
*skb1
,
160 struct sk_buff
*skb2
)
162 struct flow_keys temp
;
164 if (skb1
->protocol
!= skb2
->protocol
)
167 if (!choke_skb_cb(skb1
)->keys_valid
) {
168 choke_skb_cb(skb1
)->keys_valid
= 1;
169 skb_flow_dissect_flow_keys(skb1
, &temp
, 0);
170 make_flow_keys_digest(&choke_skb_cb(skb1
)->keys
, &temp
);
173 if (!choke_skb_cb(skb2
)->keys_valid
) {
174 choke_skb_cb(skb2
)->keys_valid
= 1;
175 skb_flow_dissect_flow_keys(skb2
, &temp
, 0);
176 make_flow_keys_digest(&choke_skb_cb(skb2
)->keys
, &temp
);
179 return !memcmp(&choke_skb_cb(skb1
)->keys
,
180 &choke_skb_cb(skb2
)->keys
,
181 sizeof(choke_skb_cb(skb1
)->keys
));
185 * Select a packet at random from queue
186 * HACK: since queue can have holes from previous deletion; retry several
187 * times to find a random skb but then just give up and return the head
188 * Will return NULL if queue is empty (q->head == q->tail)
190 static struct sk_buff
*choke_peek_random(const struct choke_sched_data
*q
,
197 *pidx
= (q
->head
+ prandom_u32_max(choke_len(q
))) & q
->tab_mask
;
201 } while (--retrys
> 0);
203 return q
->tab
[*pidx
= q
->head
];
207 * Compare new packet with random packet in queue
208 * returns true if matched and sets *pidx
210 static bool choke_match_random(const struct choke_sched_data
*q
,
211 struct sk_buff
*nskb
,
214 struct sk_buff
*oskb
;
216 if (q
->head
== q
->tail
)
219 oskb
= choke_peek_random(q
, pidx
);
220 return choke_match_flow(oskb
, nskb
);
223 static int choke_enqueue(struct sk_buff
*skb
, struct Qdisc
*sch
,
224 struct sk_buff
**to_free
)
226 struct choke_sched_data
*q
= qdisc_priv(sch
);
227 const struct red_parms
*p
= &q
->parms
;
229 choke_skb_cb(skb
)->keys_valid
= 0;
230 /* Compute average queue usage (see RED) */
231 q
->vars
.qavg
= red_calc_qavg(p
, &q
->vars
, sch
->q
.qlen
);
232 if (red_is_idling(&q
->vars
))
233 red_end_of_idle_period(&q
->vars
);
235 /* Is queue small? */
236 if (q
->vars
.qavg
<= p
->qth_min
)
241 /* Draw a packet at random from queue and compare flow */
242 if (choke_match_random(q
, skb
, &idx
)) {
244 choke_drop_by_idx(sch
, idx
, to_free
);
245 goto congestion_drop
;
248 /* Queue is large, always mark/drop */
249 if (q
->vars
.qavg
> p
->qth_max
) {
252 qdisc_qstats_overlimit(sch
);
253 if (use_harddrop(q
) || !use_ecn(q
) ||
254 !INET_ECN_set_ce(skb
)) {
255 q
->stats
.forced_drop
++;
256 goto congestion_drop
;
259 q
->stats
.forced_mark
++;
260 } else if (++q
->vars
.qcount
) {
261 if (red_mark_probability(p
, &q
->vars
, q
->vars
.qavg
)) {
263 q
->vars
.qR
= red_random(p
);
265 qdisc_qstats_overlimit(sch
);
266 if (!use_ecn(q
) || !INET_ECN_set_ce(skb
)) {
267 q
->stats
.prob_drop
++;
268 goto congestion_drop
;
271 q
->stats
.prob_mark
++;
274 q
->vars
.qR
= red_random(p
);
277 /* Admit new packet */
278 if (sch
->q
.qlen
< q
->limit
) {
279 q
->tab
[q
->tail
] = skb
;
280 q
->tail
= (q
->tail
+ 1) & q
->tab_mask
;
282 qdisc_qstats_backlog_inc(sch
, skb
);
283 return NET_XMIT_SUCCESS
;
287 return qdisc_drop(skb
, sch
, to_free
);
290 qdisc_drop(skb
, sch
, to_free
);
294 static struct sk_buff
*choke_dequeue(struct Qdisc
*sch
)
296 struct choke_sched_data
*q
= qdisc_priv(sch
);
299 if (q
->head
== q
->tail
) {
300 if (!red_is_idling(&q
->vars
))
301 red_start_of_idle_period(&q
->vars
);
305 skb
= q
->tab
[q
->head
];
306 q
->tab
[q
->head
] = NULL
;
307 choke_zap_head_holes(q
);
309 qdisc_qstats_backlog_dec(sch
, skb
);
310 qdisc_bstats_update(sch
, skb
);
315 static void choke_reset(struct Qdisc
*sch
)
317 struct choke_sched_data
*q
= qdisc_priv(sch
);
319 while (q
->head
!= q
->tail
) {
320 struct sk_buff
*skb
= q
->tab
[q
->head
];
322 q
->head
= (q
->head
+ 1) & q
->tab_mask
;
325 rtnl_qdisc_drop(skb
, sch
);
329 sch
->qstats
.backlog
= 0;
330 memset(q
->tab
, 0, (q
->tab_mask
+ 1) * sizeof(struct sk_buff
*));
331 q
->head
= q
->tail
= 0;
332 red_restart(&q
->vars
);
335 static const struct nla_policy choke_policy
[TCA_CHOKE_MAX
+ 1] = {
336 [TCA_CHOKE_PARMS
] = { .len
= sizeof(struct tc_red_qopt
) },
337 [TCA_CHOKE_STAB
] = { .len
= RED_STAB_SIZE
},
338 [TCA_CHOKE_MAX_P
] = { .type
= NLA_U32
},
342 static void choke_free(void *addr
)
347 static int choke_change(struct Qdisc
*sch
, struct nlattr
*opt
)
349 struct choke_sched_data
*q
= qdisc_priv(sch
);
350 struct nlattr
*tb
[TCA_CHOKE_MAX
+ 1];
351 const struct tc_red_qopt
*ctl
;
353 struct sk_buff
**old
= NULL
;
360 err
= nla_parse_nested(tb
, TCA_CHOKE_MAX
, opt
, choke_policy
);
364 if (tb
[TCA_CHOKE_PARMS
] == NULL
||
365 tb
[TCA_CHOKE_STAB
] == NULL
)
368 max_P
= tb
[TCA_CHOKE_MAX_P
] ? nla_get_u32(tb
[TCA_CHOKE_MAX_P
]) : 0;
370 ctl
= nla_data(tb
[TCA_CHOKE_PARMS
]);
372 if (ctl
->limit
> CHOKE_MAX_QUEUE
)
375 mask
= roundup_pow_of_two(ctl
->limit
+ 1) - 1;
376 if (mask
!= q
->tab_mask
) {
377 struct sk_buff
**ntab
;
379 ntab
= kcalloc(mask
+ 1, sizeof(struct sk_buff
*),
380 GFP_KERNEL
| __GFP_NOWARN
);
382 ntab
= vzalloc((mask
+ 1) * sizeof(struct sk_buff
*));
389 unsigned int oqlen
= sch
->q
.qlen
, tail
= 0;
390 unsigned dropped
= 0;
392 while (q
->head
!= q
->tail
) {
393 struct sk_buff
*skb
= q
->tab
[q
->head
];
395 q
->head
= (q
->head
+ 1) & q
->tab_mask
;
402 dropped
+= qdisc_pkt_len(skb
);
403 qdisc_qstats_backlog_dec(sch
, skb
);
405 rtnl_qdisc_drop(skb
, sch
);
407 qdisc_tree_reduce_backlog(sch
, oqlen
- sch
->q
.qlen
, dropped
);
417 q
->flags
= ctl
->flags
;
418 q
->limit
= ctl
->limit
;
420 red_set_parms(&q
->parms
, ctl
->qth_min
, ctl
->qth_max
, ctl
->Wlog
,
421 ctl
->Plog
, ctl
->Scell_log
,
422 nla_data(tb
[TCA_CHOKE_STAB
]),
424 red_set_vars(&q
->vars
);
426 if (q
->head
== q
->tail
)
427 red_end_of_idle_period(&q
->vars
);
429 sch_tree_unlock(sch
);
434 static int choke_init(struct Qdisc
*sch
, struct nlattr
*opt
)
436 return choke_change(sch
, opt
);
439 static int choke_dump(struct Qdisc
*sch
, struct sk_buff
*skb
)
441 struct choke_sched_data
*q
= qdisc_priv(sch
);
442 struct nlattr
*opts
= NULL
;
443 struct tc_red_qopt opt
= {
446 .qth_min
= q
->parms
.qth_min
>> q
->parms
.Wlog
,
447 .qth_max
= q
->parms
.qth_max
>> q
->parms
.Wlog
,
448 .Wlog
= q
->parms
.Wlog
,
449 .Plog
= q
->parms
.Plog
,
450 .Scell_log
= q
->parms
.Scell_log
,
453 opts
= nla_nest_start(skb
, TCA_OPTIONS
);
455 goto nla_put_failure
;
457 if (nla_put(skb
, TCA_CHOKE_PARMS
, sizeof(opt
), &opt
) ||
458 nla_put_u32(skb
, TCA_CHOKE_MAX_P
, q
->parms
.max_P
))
459 goto nla_put_failure
;
460 return nla_nest_end(skb
, opts
);
463 nla_nest_cancel(skb
, opts
);
467 static int choke_dump_stats(struct Qdisc
*sch
, struct gnet_dump
*d
)
469 struct choke_sched_data
*q
= qdisc_priv(sch
);
470 struct tc_choke_xstats st
= {
471 .early
= q
->stats
.prob_drop
+ q
->stats
.forced_drop
,
472 .marked
= q
->stats
.prob_mark
+ q
->stats
.forced_mark
,
473 .pdrop
= q
->stats
.pdrop
,
474 .other
= q
->stats
.other
,
475 .matched
= q
->stats
.matched
,
478 return gnet_stats_copy_app(d
, &st
, sizeof(st
));
481 static void choke_destroy(struct Qdisc
*sch
)
483 struct choke_sched_data
*q
= qdisc_priv(sch
);
488 static struct sk_buff
*choke_peek_head(struct Qdisc
*sch
)
490 struct choke_sched_data
*q
= qdisc_priv(sch
);
492 return (q
->head
!= q
->tail
) ? q
->tab
[q
->head
] : NULL
;
495 static struct Qdisc_ops choke_qdisc_ops __read_mostly
= {
497 .priv_size
= sizeof(struct choke_sched_data
),
499 .enqueue
= choke_enqueue
,
500 .dequeue
= choke_dequeue
,
501 .peek
= choke_peek_head
,
503 .destroy
= choke_destroy
,
504 .reset
= choke_reset
,
505 .change
= choke_change
,
507 .dump_stats
= choke_dump_stats
,
508 .owner
= THIS_MODULE
,
511 static int __init
choke_module_init(void)
513 return register_qdisc(&choke_qdisc_ops
);
516 static void __exit
choke_module_exit(void)
518 unregister_qdisc(&choke_qdisc_ops
);
521 module_init(choke_module_init
)
522 module_exit(choke_module_exit
)
524 MODULE_LICENSE("GPL");