1 // SPDX-License-Identifier: GPL-2.0-only
3 * net/sched/sch_choke.c CHOKE scheduler
5 * Copyright (c) 2011 Stephen Hemminger <shemminger@vyatta.com>
6 * Copyright (c) 2011 Eric Dumazet <eric.dumazet@gmail.com>
9 #include <linux/module.h>
10 #include <linux/types.h>
11 #include <linux/kernel.h>
12 #include <linux/skbuff.h>
13 #include <linux/vmalloc.h>
14 #include <net/pkt_sched.h>
15 #include <net/pkt_cls.h>
16 #include <net/inet_ecn.h>
18 #include <net/flow_dissector.h>
21 CHOKe stateless AQM for fair bandwidth allocation
22 =================================================
24 CHOKe (CHOose and Keep for responsive flows, CHOose and Kill for
25 unresponsive flows) is a variant of RED that penalizes misbehaving flows but
26 maintains no flow state. The difference from RED is an additional step
27 during the enqueuing process. If average queue size is over the
28 low threshold (qmin), a packet is chosen at random from the queue.
29 If both the new and chosen packet are from the same flow, both
30 are dropped. Unlike RED, CHOKe is not really a "classful" qdisc because it
31 needs to access packets in queue randomly. It has a minimal class
32 interface to allow overriding the builtin flow classifier with
36 R. Pan, B. Prabhakar, and K. Psounis, "CHOKe, A Stateless
37 Active Queue Management Scheme for Approximating Fair Bandwidth Allocation",
40 A. Tang, J. Wang, S. Low, "Understanding CHOKe: Throughput and Spatial
41 Characteristics", IEEE/ACM Transactions on Networking, 2004
45 /* Upper bound on size of sk_buff table (packets) */
46 #define CHOKE_MAX_QUEUE (128*1024 - 1)
48 struct choke_sched_data {
53 struct red_parms parms;
58 u32 prob_drop; /* Early probability drops */
59 u32 prob_mark; /* Early probability marks */
60 u32 forced_drop; /* Forced drops, qavg > max_thresh */
61 u32 forced_mark; /* Forced marks, qavg > max_thresh */
62 u32 pdrop; /* Drops due to queue limits */
63 u32 matched; /* Drops to flow match */
69 unsigned int tab_mask; /* size - 1 */
74 /* number of elements in queue including holes */
75 static unsigned int choke_len(const struct choke_sched_data *q)
77 return (q->tail - q->head) & q->tab_mask;
80 /* Is ECN parameter configured */
81 static int use_ecn(const struct choke_sched_data *q)
83 return q->flags & TC_RED_ECN;
86 /* Should packets over max just be dropped (versus marked) */
87 static int use_harddrop(const struct choke_sched_data *q)
89 return q->flags & TC_RED_HARDDROP;
92 /* Move head pointer forward to skip over holes */
93 static void choke_zap_head_holes(struct choke_sched_data *q)
96 q->head = (q->head + 1) & q->tab_mask;
97 if (q->head == q->tail)
99 } while (q->tab[q->head] == NULL);
102 /* Move tail pointer backwards to reuse holes */
103 static void choke_zap_tail_holes(struct choke_sched_data *q)
106 q->tail = (q->tail - 1) & q->tab_mask;
107 if (q->head == q->tail)
109 } while (q->tab[q->tail] == NULL);
112 /* Drop packet from queue array by creating a "hole" */
113 static void choke_drop_by_idx(struct Qdisc *sch, unsigned int idx,
114 struct sk_buff **to_free)
116 struct choke_sched_data *q = qdisc_priv(sch);
117 struct sk_buff *skb = q->tab[idx];
122 choke_zap_head_holes(q);
124 choke_zap_tail_holes(q);
126 qdisc_qstats_backlog_dec(sch, skb);
127 qdisc_tree_reduce_backlog(sch, 1, qdisc_pkt_len(skb));
128 qdisc_drop(skb, sch, to_free);
132 struct choke_skb_cb {
134 struct flow_keys_digest keys;
137 static inline struct choke_skb_cb *choke_skb_cb(const struct sk_buff *skb)
139 qdisc_cb_private_validate(skb, sizeof(struct choke_skb_cb));
140 return (struct choke_skb_cb *)qdisc_skb_cb(skb)->data;
144 * Compare flow of two packets
145 * Returns true only if source and destination address and port match.
146 * false for special cases
148 static bool choke_match_flow(struct sk_buff *skb1,
149 struct sk_buff *skb2)
151 struct flow_keys temp;
153 if (skb1->protocol != skb2->protocol)
156 if (!choke_skb_cb(skb1)->keys_valid) {
157 choke_skb_cb(skb1)->keys_valid = 1;
158 skb_flow_dissect_flow_keys(skb1, &temp, 0);
159 make_flow_keys_digest(&choke_skb_cb(skb1)->keys, &temp);
162 if (!choke_skb_cb(skb2)->keys_valid) {
163 choke_skb_cb(skb2)->keys_valid = 1;
164 skb_flow_dissect_flow_keys(skb2, &temp, 0);
165 make_flow_keys_digest(&choke_skb_cb(skb2)->keys, &temp);
168 return !memcmp(&choke_skb_cb(skb1)->keys,
169 &choke_skb_cb(skb2)->keys,
170 sizeof(choke_skb_cb(skb1)->keys));
174 * Select a packet at random from queue
175 * HACK: since queue can have holes from previous deletion; retry several
176 * times to find a random skb but then just give up and return the head
177 * Will return NULL if queue is empty (q->head == q->tail)
179 static struct sk_buff *choke_peek_random(const struct choke_sched_data *q,
186 *pidx = (q->head + get_random_u32_below(choke_len(q))) & q->tab_mask;
190 } while (--retrys > 0);
192 return q->tab[*pidx = q->head];
196 * Compare new packet with random packet in queue
197 * returns true if matched and sets *pidx
199 static bool choke_match_random(const struct choke_sched_data *q,
200 struct sk_buff *nskb,
203 struct sk_buff *oskb;
205 if (q->head == q->tail)
208 oskb = choke_peek_random(q, pidx);
209 return choke_match_flow(oskb, nskb);
212 static int choke_enqueue(struct sk_buff *skb, struct Qdisc *sch,
213 struct sk_buff **to_free)
215 struct choke_sched_data *q = qdisc_priv(sch);
216 const struct red_parms *p = &q->parms;
218 choke_skb_cb(skb)->keys_valid = 0;
219 /* Compute average queue usage (see RED) */
220 q->vars.qavg = red_calc_qavg(p, &q->vars, sch->q.qlen);
221 if (red_is_idling(&q->vars))
222 red_end_of_idle_period(&q->vars);
224 /* Is queue small? */
225 if (q->vars.qavg <= p->qth_min)
230 /* Draw a packet at random from queue and compare flow */
231 if (choke_match_random(q, skb, &idx)) {
233 choke_drop_by_idx(sch, idx, to_free);
234 goto congestion_drop;
237 /* Queue is large, always mark/drop */
238 if (q->vars.qavg > p->qth_max) {
241 qdisc_qstats_overlimit(sch);
242 if (use_harddrop(q) || !use_ecn(q) ||
243 !INET_ECN_set_ce(skb)) {
244 q->stats.forced_drop++;
245 goto congestion_drop;
248 q->stats.forced_mark++;
249 } else if (++q->vars.qcount) {
250 if (red_mark_probability(p, &q->vars, q->vars.qavg)) {
252 q->vars.qR = red_random(p);
254 qdisc_qstats_overlimit(sch);
255 if (!use_ecn(q) || !INET_ECN_set_ce(skb)) {
256 q->stats.prob_drop++;
257 goto congestion_drop;
260 q->stats.prob_mark++;
263 q->vars.qR = red_random(p);
266 /* Admit new packet */
267 if (sch->q.qlen < q->limit) {
268 q->tab[q->tail] = skb;
269 q->tail = (q->tail + 1) & q->tab_mask;
271 qdisc_qstats_backlog_inc(sch, skb);
272 return NET_XMIT_SUCCESS;
276 return qdisc_drop(skb, sch, to_free);
279 qdisc_drop(skb, sch, to_free);
283 static struct sk_buff *choke_dequeue(struct Qdisc *sch)
285 struct choke_sched_data *q = qdisc_priv(sch);
288 if (q->head == q->tail) {
289 if (!red_is_idling(&q->vars))
290 red_start_of_idle_period(&q->vars);
294 skb = q->tab[q->head];
295 q->tab[q->head] = NULL;
296 choke_zap_head_holes(q);
298 qdisc_qstats_backlog_dec(sch, skb);
299 qdisc_bstats_update(sch, skb);
304 static void choke_reset(struct Qdisc *sch)
306 struct choke_sched_data *q = qdisc_priv(sch);
308 while (q->head != q->tail) {
309 struct sk_buff *skb = q->tab[q->head];
311 q->head = (q->head + 1) & q->tab_mask;
314 rtnl_qdisc_drop(skb, sch);
318 memset(q->tab, 0, (q->tab_mask + 1) * sizeof(struct sk_buff *));
319 q->head = q->tail = 0;
320 red_restart(&q->vars);
323 static const struct nla_policy choke_policy[TCA_CHOKE_MAX + 1] = {
324 [TCA_CHOKE_PARMS] = { .len = sizeof(struct tc_red_qopt) },
325 [TCA_CHOKE_STAB] = { .len = RED_STAB_SIZE },
326 [TCA_CHOKE_MAX_P] = { .type = NLA_U32 },
330 static void choke_free(void *addr)
335 static int choke_change(struct Qdisc *sch, struct nlattr *opt,
336 struct netlink_ext_ack *extack)
338 struct choke_sched_data *q = qdisc_priv(sch);
339 struct nlattr *tb[TCA_CHOKE_MAX + 1];
340 const struct tc_red_qopt *ctl;
342 struct sk_buff **old = NULL;
350 err = nla_parse_nested_deprecated(tb, TCA_CHOKE_MAX, opt,
355 if (tb[TCA_CHOKE_PARMS] == NULL ||
356 tb[TCA_CHOKE_STAB] == NULL)
359 max_P = tb[TCA_CHOKE_MAX_P] ? nla_get_u32(tb[TCA_CHOKE_MAX_P]) : 0;
361 ctl = nla_data(tb[TCA_CHOKE_PARMS]);
362 stab = nla_data(tb[TCA_CHOKE_STAB]);
363 if (!red_check_params(ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Scell_log, stab))
366 if (ctl->limit > CHOKE_MAX_QUEUE)
369 mask = roundup_pow_of_two(ctl->limit + 1) - 1;
370 if (mask != q->tab_mask) {
371 struct sk_buff **ntab;
373 ntab = kvcalloc(mask + 1, sizeof(struct sk_buff *), GFP_KERNEL);
380 unsigned int oqlen = sch->q.qlen, tail = 0;
381 unsigned dropped = 0;
383 while (q->head != q->tail) {
384 struct sk_buff *skb = q->tab[q->head];
386 q->head = (q->head + 1) & q->tab_mask;
393 dropped += qdisc_pkt_len(skb);
394 qdisc_qstats_backlog_dec(sch, skb);
396 rtnl_qdisc_drop(skb, sch);
398 qdisc_tree_reduce_backlog(sch, oqlen - sch->q.qlen, dropped);
408 q->flags = ctl->flags;
409 q->limit = ctl->limit;
411 red_set_parms(&q->parms, ctl->qth_min, ctl->qth_max, ctl->Wlog,
412 ctl->Plog, ctl->Scell_log,
415 red_set_vars(&q->vars);
417 if (q->head == q->tail)
418 red_end_of_idle_period(&q->vars);
420 sch_tree_unlock(sch);
425 static int choke_init(struct Qdisc *sch, struct nlattr *opt,
426 struct netlink_ext_ack *extack)
428 return choke_change(sch, opt, extack);
431 static int choke_dump(struct Qdisc *sch, struct sk_buff *skb)
433 struct choke_sched_data *q = qdisc_priv(sch);
434 struct nlattr *opts = NULL;
435 struct tc_red_qopt opt = {
438 .qth_min = q->parms.qth_min >> q->parms.Wlog,
439 .qth_max = q->parms.qth_max >> q->parms.Wlog,
440 .Wlog = q->parms.Wlog,
441 .Plog = q->parms.Plog,
442 .Scell_log = q->parms.Scell_log,
445 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
447 goto nla_put_failure;
449 if (nla_put(skb, TCA_CHOKE_PARMS, sizeof(opt), &opt) ||
450 nla_put_u32(skb, TCA_CHOKE_MAX_P, q->parms.max_P))
451 goto nla_put_failure;
452 return nla_nest_end(skb, opts);
455 nla_nest_cancel(skb, opts);
459 static int choke_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
461 struct choke_sched_data *q = qdisc_priv(sch);
462 struct tc_choke_xstats st = {
463 .early = q->stats.prob_drop + q->stats.forced_drop,
464 .marked = q->stats.prob_mark + q->stats.forced_mark,
465 .pdrop = q->stats.pdrop,
466 .matched = q->stats.matched,
469 return gnet_stats_copy_app(d, &st, sizeof(st));
472 static void choke_destroy(struct Qdisc *sch)
474 struct choke_sched_data *q = qdisc_priv(sch);
479 static struct sk_buff *choke_peek_head(struct Qdisc *sch)
481 struct choke_sched_data *q = qdisc_priv(sch);
483 return (q->head != q->tail) ? q->tab[q->head] : NULL;
486 static struct Qdisc_ops choke_qdisc_ops __read_mostly = {
488 .priv_size = sizeof(struct choke_sched_data),
490 .enqueue = choke_enqueue,
491 .dequeue = choke_dequeue,
492 .peek = choke_peek_head,
494 .destroy = choke_destroy,
495 .reset = choke_reset,
496 .change = choke_change,
498 .dump_stats = choke_dump_stats,
499 .owner = THIS_MODULE,
502 static int __init choke_module_init(void)
504 return register_qdisc(&choke_qdisc_ops);
507 static void __exit choke_module_exit(void)
509 unregister_qdisc(&choke_qdisc_ops);
512 module_init(choke_module_init)
513 module_exit(choke_module_exit)
515 MODULE_LICENSE("GPL");