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/reciprocal_div.h>
18 #include <linux/vmalloc.h>
19 #include <net/pkt_sched.h>
20 #include <net/inet_ecn.h>
24 #include <linux/ipv6.h>
28 CHOKe stateless AQM for fair bandwidth allocation
29 =================================================
31 CHOKe (CHOose and Keep for responsive flows, CHOose and Kill for
32 unresponsive flows) is a variant of RED that penalizes misbehaving flows but
33 maintains no flow state. The difference from RED is an additional step
34 during the enqueuing process. If average queue size is over the
35 low threshold (qmin), a packet is chosen at random from the queue.
36 If both the new and chosen packet are from the same flow, both
37 are dropped. Unlike RED, CHOKe is not really a "classful" qdisc because it
38 needs to access packets in queue randomly. It has a minimal class
39 interface to allow overriding the builtin flow classifier with
43 R. Pan, B. Prabhakar, and K. Psounis, "CHOKe, A Stateless
44 Active Queue Management Scheme for Approximating Fair Bandwidth Allocation",
47 A. Tang, J. Wang, S. Low, "Understanding CHOKe: Throughput and Spatial
48 Characteristics", IEEE/ACM Transactions on Networking, 2004
52 /* Upper bound on size of sk_buff table (packets) */
53 #define CHOKE_MAX_QUEUE (128*1024 - 1)
55 struct choke_sched_data {
60 struct red_parms parms;
63 struct tcf_proto *filter_list;
65 u32 prob_drop; /* Early probability drops */
66 u32 prob_mark; /* Early probability marks */
67 u32 forced_drop; /* Forced drops, qavg > max_thresh */
68 u32 forced_mark; /* Forced marks, qavg > max_thresh */
69 u32 pdrop; /* Drops due to queue limits */
70 u32 other; /* Drops due to drop() calls */
71 u32 matched; /* Drops to flow match */
77 unsigned int tab_mask; /* size - 1 */
82 /* deliver a random number between 0 and N - 1 */
83 static u32 random_N(unsigned int N)
85 return reciprocal_divide(random32(), N);
88 /* number of elements in queue including holes */
89 static unsigned int choke_len(const struct choke_sched_data *q)
91 return (q->tail - q->head) & q->tab_mask;
94 /* Is ECN parameter configured */
95 static int use_ecn(const struct choke_sched_data *q)
97 return q->flags & TC_RED_ECN;
100 /* Should packets over max just be dropped (versus marked) */
101 static int use_harddrop(const struct choke_sched_data *q)
103 return q->flags & TC_RED_HARDDROP;
106 /* Move head pointer forward to skip over holes */
107 static void choke_zap_head_holes(struct choke_sched_data *q)
110 q->head = (q->head + 1) & q->tab_mask;
111 if (q->head == q->tail)
113 } while (q->tab[q->head] == NULL);
116 /* Move tail pointer backwards to reuse holes */
117 static void choke_zap_tail_holes(struct choke_sched_data *q)
120 q->tail = (q->tail - 1) & q->tab_mask;
121 if (q->head == q->tail)
123 } while (q->tab[q->tail] == NULL);
126 /* Drop packet from queue array by creating a "hole" */
127 static void choke_drop_by_idx(struct Qdisc *sch, unsigned int idx)
129 struct choke_sched_data *q = qdisc_priv(sch);
130 struct sk_buff *skb = q->tab[idx];
135 choke_zap_head_holes(q);
137 choke_zap_tail_holes(q);
139 sch->qstats.backlog -= qdisc_pkt_len(skb);
140 qdisc_drop(skb, sch);
141 qdisc_tree_decrease_qlen(sch, 1);
146 * Compare flow of two packets
147 * Returns true only if source and destination address and port match.
148 * false for special cases
150 static bool choke_match_flow(struct sk_buff *skb1,
151 struct sk_buff *skb2)
153 int off1, off2, poff;
154 const u32 *ports1, *ports2;
158 if (skb1->protocol != skb2->protocol)
161 /* Use hash value as quick check
162 * Assumes that __skb_get_rxhash makes IP header and ports linear
164 hash1 = skb_get_rxhash(skb1);
165 if (!hash1 || hash1 != skb_get_rxhash(skb2))
168 /* Probably match, but be sure to avoid hash collisions */
169 off1 = skb_network_offset(skb1);
170 off2 = skb_network_offset(skb2);
172 switch (skb1->protocol) {
173 case __constant_htons(ETH_P_IP): {
174 const struct iphdr *ip1, *ip2;
176 ip1 = (const struct iphdr *) (skb1->data + off1);
177 ip2 = (const struct iphdr *) (skb2->data + off2);
179 ip_proto = ip1->protocol;
180 if (ip_proto != ip2->protocol ||
181 ip1->saddr != ip2->saddr || ip1->daddr != ip2->daddr)
184 if (ip_is_fragment(ip1) | ip_is_fragment(ip2))
186 off1 += ip1->ihl * 4;
187 off2 += ip2->ihl * 4;
191 case __constant_htons(ETH_P_IPV6): {
192 const struct ipv6hdr *ip1, *ip2;
194 ip1 = (const struct ipv6hdr *) (skb1->data + off1);
195 ip2 = (const struct ipv6hdr *) (skb2->data + off2);
197 ip_proto = ip1->nexthdr;
198 if (ip_proto != ip2->nexthdr ||
199 ipv6_addr_cmp(&ip1->saddr, &ip2->saddr) ||
200 ipv6_addr_cmp(&ip1->daddr, &ip2->daddr))
206 default: /* Maybe compare MAC header here? */
210 poff = proto_ports_offset(ip_proto);
217 ports1 = (__force u32 *)(skb1->data + off1);
218 ports2 = (__force u32 *)(skb2->data + off2);
219 return *ports1 == *ports2;
222 struct choke_skb_cb {
226 static inline struct choke_skb_cb *choke_skb_cb(const struct sk_buff *skb)
228 BUILD_BUG_ON(sizeof(skb->cb) <
229 sizeof(struct qdisc_skb_cb) + sizeof(struct choke_skb_cb));
230 return (struct choke_skb_cb *)qdisc_skb_cb(skb)->data;
233 static inline void choke_set_classid(struct sk_buff *skb, u16 classid)
235 choke_skb_cb(skb)->classid = classid;
238 static u16 choke_get_classid(const struct sk_buff *skb)
240 return choke_skb_cb(skb)->classid;
244 * Classify flow using either:
245 * 1. pre-existing classification result in skb
246 * 2. fast internal classification
247 * 3. use TC filter based classification
249 static bool choke_classify(struct sk_buff *skb,
250 struct Qdisc *sch, int *qerr)
253 struct choke_sched_data *q = qdisc_priv(sch);
254 struct tcf_result res;
257 result = tc_classify(skb, q->filter_list, &res);
259 #ifdef CONFIG_NET_CLS_ACT
263 *qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
268 choke_set_classid(skb, TC_H_MIN(res.classid));
276 * Select a packet at random from queue
277 * HACK: since queue can have holes from previous deletion; retry several
278 * times to find a random skb but then just give up and return the head
279 * Will return NULL if queue is empty (q->head == q->tail)
281 static struct sk_buff *choke_peek_random(const struct choke_sched_data *q,
288 *pidx = (q->head + random_N(choke_len(q))) & q->tab_mask;
292 } while (--retrys > 0);
294 return q->tab[*pidx = q->head];
298 * Compare new packet with random packet in queue
299 * returns true if matched and sets *pidx
301 static bool choke_match_random(const struct choke_sched_data *q,
302 struct sk_buff *nskb,
305 struct sk_buff *oskb;
307 if (q->head == q->tail)
310 oskb = choke_peek_random(q, pidx);
312 return choke_get_classid(nskb) == choke_get_classid(oskb);
314 return choke_match_flow(oskb, nskb);
317 static int choke_enqueue(struct sk_buff *skb, struct Qdisc *sch)
319 struct choke_sched_data *q = qdisc_priv(sch);
320 struct red_parms *p = &q->parms;
321 int ret = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
323 if (q->filter_list) {
324 /* If using external classifiers, get result and record it. */
325 if (!choke_classify(skb, sch, &ret))
326 goto other_drop; /* Packet was eaten by filter */
329 /* Compute average queue usage (see RED) */
330 p->qavg = red_calc_qavg(p, sch->q.qlen);
331 if (red_is_idling(p))
332 red_end_of_idle_period(p);
334 /* Is queue small? */
335 if (p->qavg <= p->qth_min)
340 /* Draw a packet at random from queue and compare flow */
341 if (choke_match_random(q, skb, &idx)) {
343 choke_drop_by_idx(sch, idx);
344 goto congestion_drop;
347 /* Queue is large, always mark/drop */
348 if (p->qavg > p->qth_max) {
351 sch->qstats.overlimits++;
352 if (use_harddrop(q) || !use_ecn(q) ||
353 !INET_ECN_set_ce(skb)) {
354 q->stats.forced_drop++;
355 goto congestion_drop;
358 q->stats.forced_mark++;
359 } else if (++p->qcount) {
360 if (red_mark_probability(p, p->qavg)) {
362 p->qR = red_random(p);
364 sch->qstats.overlimits++;
365 if (!use_ecn(q) || !INET_ECN_set_ce(skb)) {
366 q->stats.prob_drop++;
367 goto congestion_drop;
370 q->stats.prob_mark++;
373 p->qR = red_random(p);
376 /* Admit new packet */
377 if (sch->q.qlen < q->limit) {
378 q->tab[q->tail] = skb;
379 q->tail = (q->tail + 1) & q->tab_mask;
381 sch->qstats.backlog += qdisc_pkt_len(skb);
382 return NET_XMIT_SUCCESS;
388 return NET_XMIT_DROP;
391 qdisc_drop(skb, sch);
395 if (ret & __NET_XMIT_BYPASS)
401 static struct sk_buff *choke_dequeue(struct Qdisc *sch)
403 struct choke_sched_data *q = qdisc_priv(sch);
406 if (q->head == q->tail) {
407 if (!red_is_idling(&q->parms))
408 red_start_of_idle_period(&q->parms);
412 skb = q->tab[q->head];
413 q->tab[q->head] = NULL;
414 choke_zap_head_holes(q);
416 sch->qstats.backlog -= qdisc_pkt_len(skb);
417 qdisc_bstats_update(sch, skb);
422 static unsigned int choke_drop(struct Qdisc *sch)
424 struct choke_sched_data *q = qdisc_priv(sch);
427 len = qdisc_queue_drop(sch);
431 if (!red_is_idling(&q->parms))
432 red_start_of_idle_period(&q->parms);
438 static void choke_reset(struct Qdisc *sch)
440 struct choke_sched_data *q = qdisc_priv(sch);
442 red_restart(&q->parms);
445 static const struct nla_policy choke_policy[TCA_CHOKE_MAX + 1] = {
446 [TCA_CHOKE_PARMS] = { .len = sizeof(struct tc_red_qopt) },
447 [TCA_CHOKE_STAB] = { .len = RED_STAB_SIZE },
451 static void choke_free(void *addr)
454 if (is_vmalloc_addr(addr))
461 static int choke_change(struct Qdisc *sch, struct nlattr *opt)
463 struct choke_sched_data *q = qdisc_priv(sch);
464 struct nlattr *tb[TCA_CHOKE_MAX + 1];
465 const struct tc_red_qopt *ctl;
467 struct sk_buff **old = NULL;
473 err = nla_parse_nested(tb, TCA_CHOKE_MAX, opt, choke_policy);
477 if (tb[TCA_CHOKE_PARMS] == NULL ||
478 tb[TCA_CHOKE_STAB] == NULL)
481 ctl = nla_data(tb[TCA_CHOKE_PARMS]);
483 if (ctl->limit > CHOKE_MAX_QUEUE)
486 mask = roundup_pow_of_two(ctl->limit + 1) - 1;
487 if (mask != q->tab_mask) {
488 struct sk_buff **ntab;
490 ntab = kcalloc(mask + 1, sizeof(struct sk_buff *), GFP_KERNEL);
492 ntab = vzalloc((mask + 1) * sizeof(struct sk_buff *));
499 unsigned int oqlen = sch->q.qlen, tail = 0;
501 while (q->head != q->tail) {
502 struct sk_buff *skb = q->tab[q->head];
504 q->head = (q->head + 1) & q->tab_mask;
511 sch->qstats.backlog -= qdisc_pkt_len(skb);
513 qdisc_drop(skb, sch);
515 qdisc_tree_decrease_qlen(sch, oqlen - sch->q.qlen);
525 q->flags = ctl->flags;
526 q->limit = ctl->limit;
528 red_set_parms(&q->parms, ctl->qth_min, ctl->qth_max, ctl->Wlog,
529 ctl->Plog, ctl->Scell_log,
530 nla_data(tb[TCA_CHOKE_STAB]));
532 if (q->head == q->tail)
533 red_end_of_idle_period(&q->parms);
535 sch_tree_unlock(sch);
540 static int choke_init(struct Qdisc *sch, struct nlattr *opt)
542 return choke_change(sch, opt);
545 static int choke_dump(struct Qdisc *sch, struct sk_buff *skb)
547 struct choke_sched_data *q = qdisc_priv(sch);
548 struct nlattr *opts = NULL;
549 struct tc_red_qopt opt = {
552 .qth_min = q->parms.qth_min >> q->parms.Wlog,
553 .qth_max = q->parms.qth_max >> q->parms.Wlog,
554 .Wlog = q->parms.Wlog,
555 .Plog = q->parms.Plog,
556 .Scell_log = q->parms.Scell_log,
559 opts = nla_nest_start(skb, TCA_OPTIONS);
561 goto nla_put_failure;
563 NLA_PUT(skb, TCA_CHOKE_PARMS, sizeof(opt), &opt);
564 return nla_nest_end(skb, opts);
567 nla_nest_cancel(skb, opts);
571 static int choke_dump_stats(struct Qdisc *sch, struct gnet_dump *d)
573 struct choke_sched_data *q = qdisc_priv(sch);
574 struct tc_choke_xstats st = {
575 .early = q->stats.prob_drop + q->stats.forced_drop,
576 .marked = q->stats.prob_mark + q->stats.forced_mark,
577 .pdrop = q->stats.pdrop,
578 .other = q->stats.other,
579 .matched = q->stats.matched,
582 return gnet_stats_copy_app(d, &st, sizeof(st));
585 static void choke_destroy(struct Qdisc *sch)
587 struct choke_sched_data *q = qdisc_priv(sch);
589 tcf_destroy_chain(&q->filter_list);
593 static struct Qdisc *choke_leaf(struct Qdisc *sch, unsigned long arg)
598 static unsigned long choke_get(struct Qdisc *sch, u32 classid)
603 static void choke_put(struct Qdisc *q, unsigned long cl)
607 static unsigned long choke_bind(struct Qdisc *sch, unsigned long parent,
613 static struct tcf_proto **choke_find_tcf(struct Qdisc *sch, unsigned long cl)
615 struct choke_sched_data *q = qdisc_priv(sch);
619 return &q->filter_list;
622 static int choke_dump_class(struct Qdisc *sch, unsigned long cl,
623 struct sk_buff *skb, struct tcmsg *tcm)
625 tcm->tcm_handle |= TC_H_MIN(cl);
629 static void choke_walk(struct Qdisc *sch, struct qdisc_walker *arg)
632 if (arg->fn(sch, 1, arg) < 0) {
640 static const struct Qdisc_class_ops choke_class_ops = {
644 .tcf_chain = choke_find_tcf,
645 .bind_tcf = choke_bind,
646 .unbind_tcf = choke_put,
647 .dump = choke_dump_class,
651 static struct sk_buff *choke_peek_head(struct Qdisc *sch)
653 struct choke_sched_data *q = qdisc_priv(sch);
655 return (q->head != q->tail) ? q->tab[q->head] : NULL;
658 static struct Qdisc_ops choke_qdisc_ops __read_mostly = {
660 .priv_size = sizeof(struct choke_sched_data),
662 .enqueue = choke_enqueue,
663 .dequeue = choke_dequeue,
664 .peek = choke_peek_head,
667 .destroy = choke_destroy,
668 .reset = choke_reset,
669 .change = choke_change,
671 .dump_stats = choke_dump_stats,
672 .owner = THIS_MODULE,
675 static int __init choke_module_init(void)
677 return register_qdisc(&choke_qdisc_ops);
680 static void __exit choke_module_exit(void)
682 unregister_qdisc(&choke_qdisc_ops);
685 module_init(choke_module_init)
686 module_exit(choke_module_exit)
688 MODULE_LICENSE("GPL");