2 * net/sched/sch_gred.c Generic Random Early Detection queue.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License
7 * as published by the Free Software Foundation; either version
8 * 2 of the License, or (at your option) any later version.
10 * Authors: J Hadi Salim (hadi@cyberus.ca) 1998-2002
12 * 991129: - Bug fix with grio mode
13 * - a better sing. AvgQ mode with Grio(WRED)
14 * - A finer grained VQ dequeue based on sugestion
18 * For all the glorious comments look at include/net/red.h
21 #include <linux/slab.h>
22 #include <linux/module.h>
23 #include <linux/types.h>
24 #include <linux/kernel.h>
25 #include <linux/skbuff.h>
26 #include <net/pkt_sched.h>
29 #define GRED_DEF_PRIO (MAX_DPs / 2)
30 #define GRED_VQ_MASK (MAX_DPs - 1)
32 struct gred_sched_data;
35 struct gred_sched_data {
36 u32 limit; /* HARD maximal queue length */
37 u32 DP; /* the drop pramaters */
38 u32 bytesin; /* bytes seen on virtualQ so far*/
39 u32 packetsin; /* packets seen on virtualQ so far*/
40 u32 backlog; /* bytes on the virtualQ */
41 u8 prio; /* the prio of this vq */
43 struct red_parms parms;
44 struct red_stats stats;
53 struct gred_sched_data *tab[MAX_DPs];
58 struct red_parms wred_set;
61 static inline int gred_wred_mode(struct gred_sched *table)
63 return test_bit(GRED_WRED_MODE, &table->flags);
66 static inline void gred_enable_wred_mode(struct gred_sched *table)
68 __set_bit(GRED_WRED_MODE, &table->flags);
71 static inline void gred_disable_wred_mode(struct gred_sched *table)
73 __clear_bit(GRED_WRED_MODE, &table->flags);
76 static inline int gred_rio_mode(struct gred_sched *table)
78 return test_bit(GRED_RIO_MODE, &table->flags);
81 static inline void gred_enable_rio_mode(struct gred_sched *table)
83 __set_bit(GRED_RIO_MODE, &table->flags);
86 static inline void gred_disable_rio_mode(struct gred_sched *table)
88 __clear_bit(GRED_RIO_MODE, &table->flags);
91 static inline int gred_wred_mode_check(struct Qdisc *sch)
93 struct gred_sched *table = qdisc_priv(sch);
96 /* Really ugly O(n^2) but shouldn't be necessary too frequent. */
97 for (i = 0; i < table->DPs; i++) {
98 struct gred_sched_data *q = table->tab[i];
104 for (n = 0; n < table->DPs; n++)
105 if (table->tab[n] && table->tab[n] != q &&
106 table->tab[n]->prio == q->prio)
113 static inline unsigned int gred_backlog(struct gred_sched *table,
114 struct gred_sched_data *q,
117 if (gred_wred_mode(table))
118 return sch->qstats.backlog;
123 static inline u16 tc_index_to_dp(struct sk_buff *skb)
125 return skb->tc_index & GRED_VQ_MASK;
128 static inline void gred_load_wred_set(struct gred_sched *table,
129 struct gred_sched_data *q)
131 q->parms.qavg = table->wred_set.qavg;
132 q->parms.qidlestart = table->wred_set.qidlestart;
135 static inline void gred_store_wred_set(struct gred_sched *table,
136 struct gred_sched_data *q)
138 table->wred_set.qavg = q->parms.qavg;
141 static inline int gred_use_ecn(struct gred_sched *t)
143 return t->red_flags & TC_RED_ECN;
146 static inline int gred_use_harddrop(struct gred_sched *t)
148 return t->red_flags & TC_RED_HARDDROP;
151 static int gred_enqueue(struct sk_buff *skb, struct Qdisc *sch)
153 struct gred_sched_data *q = NULL;
154 struct gred_sched *t = qdisc_priv(sch);
155 unsigned long qavg = 0;
156 u16 dp = tc_index_to_dp(skb);
158 if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
163 /* Pass through packets not assigned to a DP
164 * if no default DP has been configured. This
165 * allows for DP flows to be left untouched.
167 if (skb_queue_len(&sch->q) < qdisc_dev(sch)->tx_queue_len)
168 return qdisc_enqueue_tail(skb, sch);
173 /* fix tc_index? --could be controvesial but needed for
175 skb->tc_index = (skb->tc_index & ~GRED_VQ_MASK) | dp;
178 /* sum up all the qaves of prios <= to ours to get the new qave */
179 if (!gred_wred_mode(t) && gred_rio_mode(t)) {
182 for (i = 0; i < t->DPs; i++) {
183 if (t->tab[i] && t->tab[i]->prio < q->prio &&
184 !red_is_idling(&t->tab[i]->parms))
185 qavg += t->tab[i]->parms.qavg;
191 q->bytesin += qdisc_pkt_len(skb);
193 if (gred_wred_mode(t))
194 gred_load_wred_set(t, q);
196 q->parms.qavg = red_calc_qavg(&q->parms, gred_backlog(t, q, sch));
198 if (red_is_idling(&q->parms))
199 red_end_of_idle_period(&q->parms);
201 if (gred_wred_mode(t))
202 gred_store_wred_set(t, q);
204 switch (red_action(&q->parms, q->parms.qavg + qavg)) {
209 sch->qstats.overlimits++;
210 if (!gred_use_ecn(t) || !INET_ECN_set_ce(skb)) {
211 q->stats.prob_drop++;
212 goto congestion_drop;
215 q->stats.prob_mark++;
219 sch->qstats.overlimits++;
220 if (gred_use_harddrop(t) || !gred_use_ecn(t) ||
221 !INET_ECN_set_ce(skb)) {
222 q->stats.forced_drop++;
223 goto congestion_drop;
225 q->stats.forced_mark++;
229 if (q->backlog + qdisc_pkt_len(skb) <= q->limit) {
230 q->backlog += qdisc_pkt_len(skb);
231 return qdisc_enqueue_tail(skb, sch);
236 return qdisc_drop(skb, sch);
239 qdisc_drop(skb, sch);
243 static struct sk_buff *gred_dequeue(struct Qdisc *sch)
246 struct gred_sched *t = qdisc_priv(sch);
248 skb = qdisc_dequeue_head(sch);
251 struct gred_sched_data *q;
252 u16 dp = tc_index_to_dp(skb);
254 if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
256 pr_warning("GRED: Unable to relocate VQ 0x%x "
257 "after dequeue, screwing up "
258 "backlog.\n", tc_index_to_dp(skb));
260 q->backlog -= qdisc_pkt_len(skb);
262 if (!q->backlog && !gred_wred_mode(t))
263 red_start_of_idle_period(&q->parms);
269 if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
270 red_start_of_idle_period(&t->wred_set);
275 static unsigned int gred_drop(struct Qdisc *sch)
278 struct gred_sched *t = qdisc_priv(sch);
280 skb = qdisc_dequeue_tail(sch);
282 unsigned int len = qdisc_pkt_len(skb);
283 struct gred_sched_data *q;
284 u16 dp = tc_index_to_dp(skb);
286 if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
288 pr_warning("GRED: Unable to relocate VQ 0x%x "
289 "while dropping, screwing up "
290 "backlog.\n", tc_index_to_dp(skb));
295 if (!q->backlog && !gred_wred_mode(t))
296 red_start_of_idle_period(&q->parms);
299 qdisc_drop(skb, sch);
303 if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
304 red_start_of_idle_period(&t->wred_set);
310 static void gred_reset(struct Qdisc *sch)
313 struct gred_sched *t = qdisc_priv(sch);
315 qdisc_reset_queue(sch);
317 for (i = 0; i < t->DPs; i++) {
318 struct gred_sched_data *q = t->tab[i];
323 red_restart(&q->parms);
328 static inline void gred_destroy_vq(struct gred_sched_data *q)
333 static inline int gred_change_table_def(struct Qdisc *sch, struct nlattr *dps)
335 struct gred_sched *table = qdisc_priv(sch);
336 struct tc_gred_sopt *sopt;
342 sopt = nla_data(dps);
344 if (sopt->DPs > MAX_DPs || sopt->DPs == 0 || sopt->def_DP >= sopt->DPs)
348 table->DPs = sopt->DPs;
349 table->def = sopt->def_DP;
350 table->red_flags = sopt->flags;
353 * Every entry point to GRED is synchronized with the above code
354 * and the DP is checked against DPs, i.e. shadowed VQs can no
355 * longer be found so we can unlock right here.
357 sch_tree_unlock(sch);
360 gred_enable_rio_mode(table);
361 gred_disable_wred_mode(table);
362 if (gred_wred_mode_check(sch))
363 gred_enable_wred_mode(table);
365 gred_disable_rio_mode(table);
366 gred_disable_wred_mode(table);
369 for (i = table->DPs; i < MAX_DPs; i++) {
371 pr_warning("GRED: Warning: Destroying "
372 "shadowed VQ 0x%x\n", i);
373 gred_destroy_vq(table->tab[i]);
374 table->tab[i] = NULL;
381 static inline int gred_change_vq(struct Qdisc *sch, int dp,
382 struct tc_gred_qopt *ctl, int prio, u8 *stab)
384 struct gred_sched *table = qdisc_priv(sch);
385 struct gred_sched_data *q;
387 if (!red_check_params(ctl->qth_min, ctl->qth_max, ctl->Wlog))
390 if (table->tab[dp] == NULL) {
391 table->tab[dp] = kzalloc(sizeof(*q), GFP_ATOMIC);
392 if (table->tab[dp] == NULL)
399 q->limit = ctl->limit;
402 red_end_of_idle_period(&q->parms);
404 red_set_parms(&q->parms,
405 ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Plog,
406 ctl->Scell_log, stab);
411 static const struct nla_policy gred_policy[TCA_GRED_MAX + 1] = {
412 [TCA_GRED_PARMS] = { .len = sizeof(struct tc_gred_qopt) },
413 [TCA_GRED_STAB] = { .len = 256 },
414 [TCA_GRED_DPS] = { .len = sizeof(struct tc_gred_sopt) },
417 static int gred_change(struct Qdisc *sch, struct nlattr *opt)
419 struct gred_sched *table = qdisc_priv(sch);
420 struct tc_gred_qopt *ctl;
421 struct nlattr *tb[TCA_GRED_MAX + 1];
422 int err, prio = GRED_DEF_PRIO;
428 err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
432 if (tb[TCA_GRED_PARMS] == NULL && tb[TCA_GRED_STAB] == NULL)
433 return gred_change_table_def(sch, opt);
435 if (tb[TCA_GRED_PARMS] == NULL ||
436 tb[TCA_GRED_STAB] == NULL)
440 ctl = nla_data(tb[TCA_GRED_PARMS]);
441 stab = nla_data(tb[TCA_GRED_STAB]);
443 if (ctl->DP >= table->DPs)
446 if (gred_rio_mode(table)) {
447 if (ctl->prio == 0) {
448 int def_prio = GRED_DEF_PRIO;
450 if (table->tab[table->def])
451 def_prio = table->tab[table->def]->prio;
453 printk(KERN_DEBUG "GRED: DP %u does not have a prio "
454 "setting default to %d\n", ctl->DP, def_prio);
463 err = gred_change_vq(sch, ctl->DP, ctl, prio, stab);
467 if (gred_rio_mode(table)) {
468 gred_disable_wred_mode(table);
469 if (gred_wred_mode_check(sch))
470 gred_enable_wred_mode(table);
476 sch_tree_unlock(sch);
481 static int gred_init(struct Qdisc *sch, struct nlattr *opt)
483 struct nlattr *tb[TCA_GRED_MAX + 1];
489 err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
493 if (tb[TCA_GRED_PARMS] || tb[TCA_GRED_STAB])
496 return gred_change_table_def(sch, tb[TCA_GRED_DPS]);
499 static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
501 struct gred_sched *table = qdisc_priv(sch);
502 struct nlattr *parms, *opts = NULL;
504 struct tc_gred_sopt sopt = {
506 .def_DP = table->def,
507 .grio = gred_rio_mode(table),
508 .flags = table->red_flags,
511 opts = nla_nest_start(skb, TCA_OPTIONS);
513 goto nla_put_failure;
514 NLA_PUT(skb, TCA_GRED_DPS, sizeof(sopt), &sopt);
515 parms = nla_nest_start(skb, TCA_GRED_PARMS);
517 goto nla_put_failure;
519 for (i = 0; i < MAX_DPs; i++) {
520 struct gred_sched_data *q = table->tab[i];
521 struct tc_gred_qopt opt;
523 memset(&opt, 0, sizeof(opt));
526 /* hack -- fix at some point with proper message
527 This is how we indicate to tc that there is no VQ
530 opt.DP = MAX_DPs + i;
534 opt.limit = q->limit;
536 opt.backlog = q->backlog;
538 opt.qth_min = q->parms.qth_min >> q->parms.Wlog;
539 opt.qth_max = q->parms.qth_max >> q->parms.Wlog;
540 opt.Wlog = q->parms.Wlog;
541 opt.Plog = q->parms.Plog;
542 opt.Scell_log = q->parms.Scell_log;
543 opt.other = q->stats.other;
544 opt.early = q->stats.prob_drop;
545 opt.forced = q->stats.forced_drop;
546 opt.pdrop = q->stats.pdrop;
547 opt.packets = q->packetsin;
548 opt.bytesin = q->bytesin;
550 if (gred_wred_mode(table))
551 gred_load_wred_set(table, q);
553 opt.qave = red_calc_qavg(&q->parms, q->parms.qavg);
556 if (nla_append(skb, sizeof(opt), &opt) < 0)
557 goto nla_put_failure;
560 nla_nest_end(skb, parms);
562 return nla_nest_end(skb, opts);
565 nla_nest_cancel(skb, opts);
569 static void gred_destroy(struct Qdisc *sch)
571 struct gred_sched *table = qdisc_priv(sch);
574 for (i = 0; i < table->DPs; i++) {
576 gred_destroy_vq(table->tab[i]);
580 static struct Qdisc_ops gred_qdisc_ops __read_mostly = {
582 .priv_size = sizeof(struct gred_sched),
583 .enqueue = gred_enqueue,
584 .dequeue = gred_dequeue,
585 .peek = qdisc_peek_head,
589 .destroy = gred_destroy,
590 .change = gred_change,
592 .owner = THIS_MODULE,
595 static int __init gred_module_init(void)
597 return register_qdisc(&gred_qdisc_ops);
600 static void __exit gred_module_exit(void)
602 unregister_qdisc(&gred_qdisc_ops);
605 module_init(gred_module_init)
606 module_exit(gred_module_exit)
608 MODULE_LICENSE("GPL");