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
20 * For all the glorious comments look at Alexey's sch_red.c
23 #include <linux/config.h>
24 #include <linux/module.h>
25 #include <asm/uaccess.h>
26 #include <asm/system.h>
27 #include <linux/bitops.h>
28 #include <linux/types.h>
29 #include <linux/kernel.h>
30 #include <linux/sched.h>
31 #include <linux/string.h>
33 #include <linux/socket.h>
34 #include <linux/sockios.h>
36 #include <linux/errno.h>
37 #include <linux/interrupt.h>
38 #include <linux/if_ether.h>
39 #include <linux/inet.h>
40 #include <linux/netdevice.h>
41 #include <linux/etherdevice.h>
42 #include <linux/notifier.h>
44 #include <net/route.h>
45 #include <linux/skbuff.h>
47 #include <net/pkt_sched.h>
51 #define DPRINTK(format,args...) printk(KERN_DEBUG format,##args)
53 #define DPRINTK(format,args...)
57 #define D2PRINTK(format,args...) printk(KERN_DEBUG format,##args)
59 #define D2PRINTK(format,args...)
62 #define GRED_DEF_PRIO (MAX_DPs / 2)
64 struct gred_sched_data;
67 struct gred_sched_data
69 u32 limit; /* HARD maximal queue length */
70 u32 DP; /* the drop pramaters */
71 u32 bytesin; /* bytes seen on virtualQ so far*/
72 u32 packetsin; /* packets seen on virtualQ so far*/
73 u32 backlog; /* bytes on the virtualQ */
74 u8 prio; /* the prio of this vq */
76 struct red_parms parms;
77 struct red_stats stats;
87 struct gred_sched_data *tab[MAX_DPs];
94 static inline int gred_wred_mode(struct gred_sched *table)
96 return test_bit(GRED_WRED_MODE, &table->flags);
99 static inline void gred_enable_wred_mode(struct gred_sched *table)
101 __set_bit(GRED_WRED_MODE, &table->flags);
104 static inline void gred_disable_wred_mode(struct gred_sched *table)
106 __clear_bit(GRED_WRED_MODE, &table->flags);
109 static inline int gred_rio_mode(struct gred_sched *table)
111 return test_bit(GRED_RIO_MODE, &table->flags);
114 static inline void gred_enable_rio_mode(struct gred_sched *table)
116 __set_bit(GRED_RIO_MODE, &table->flags);
119 static inline void gred_disable_rio_mode(struct gred_sched *table)
121 __clear_bit(GRED_RIO_MODE, &table->flags);
124 static inline int gred_wred_mode_check(struct Qdisc *sch)
126 struct gred_sched *table = qdisc_priv(sch);
129 /* Really ugly O(n^2) but shouldn't be necessary too frequent. */
130 for (i = 0; i < table->DPs; i++) {
131 struct gred_sched_data *q = table->tab[i];
137 for (n = 0; n < table->DPs; n++)
138 if (table->tab[n] && table->tab[n] != q &&
139 table->tab[n]->prio == q->prio)
146 static inline unsigned int gred_backlog(struct gred_sched *table,
147 struct gred_sched_data *q,
150 if (gred_wred_mode(table))
151 return sch->qstats.backlog;
157 gred_enqueue(struct sk_buff *skb, struct Qdisc* sch)
159 struct gred_sched_data *q=NULL;
160 struct gred_sched *t= qdisc_priv(sch);
161 unsigned long qavg = 0;
164 if (!t->initd && skb_queue_len(&sch->q) < (sch->dev->tx_queue_len ? : 1)) {
165 D2PRINTK("NO GRED Queues setup yet! Enqueued anyway\n");
170 if ( ((skb->tc_index&0xf) > (t->DPs -1)) || !(q=t->tab[skb->tc_index&0xf])) {
171 printk("GRED: setting to default (%d)\n ",t->def);
172 if (!(q=t->tab[t->def])) {
173 DPRINTK("GRED: setting to default FAILED! dropping!! "
177 /* fix tc_index? --could be controvesial but needed for
179 skb->tc_index=(skb->tc_index&0xfffffff0) | t->def;
182 D2PRINTK("gred_enqueue virtualQ 0x%x classid %x backlog %d "
183 "general backlog %d\n",skb->tc_index&0xf,sch->handle,q->backlog,
184 sch->qstats.backlog);
185 /* sum up all the qaves of prios <= to ours to get the new qave*/
186 if (!gred_wred_mode(t) && gred_rio_mode(t)) {
187 for (i=0;i<t->DPs;i++) {
188 if ((!t->tab[i]) || (i==q->DP))
191 if (t->tab[i]->prio < q->prio &&
192 !red_is_idling(&t->tab[i]->parms))
193 qavg +=t->tab[i]->parms.qavg;
199 q->bytesin+=skb->len;
201 if (gred_wred_mode(t)) {
203 q->parms.qavg = t->tab[t->def]->parms.qavg;
204 q->parms.qidlestart = t->tab[t->def]->parms.qidlestart;
207 q->parms.qavg = red_calc_qavg(&q->parms, gred_backlog(t, q, sch));
209 if (red_is_idling(&q->parms))
210 red_end_of_idle_period(&q->parms);
212 if (gred_wred_mode(t))
213 t->tab[t->def]->parms.qavg = q->parms.qavg;
215 switch (red_action(&q->parms, q->parms.qavg + qavg)) {
220 sch->qstats.overlimits++;
221 q->stats.prob_drop++;
222 goto congestion_drop;
225 sch->qstats.overlimits++;
226 q->stats.forced_drop++;
227 goto congestion_drop;
230 if (q->backlog + skb->len <= q->limit) {
231 q->backlog += skb->len;
233 return qdisc_enqueue_tail(skb, sch);
238 return qdisc_drop(skb, sch);
241 qdisc_drop(skb, sch);
246 gred_requeue(struct sk_buff *skb, struct Qdisc* sch)
248 struct gred_sched_data *q;
249 struct gred_sched *t= qdisc_priv(sch);
250 q= t->tab[(skb->tc_index&0xf)];
251 /* error checking here -- probably unnecessary */
253 if (red_is_idling(&q->parms))
254 red_end_of_idle_period(&q->parms);
256 q->backlog += skb->len;
257 return qdisc_requeue(skb, sch);
260 static struct sk_buff *
261 gred_dequeue(struct Qdisc* sch)
264 struct gred_sched_data *q;
265 struct gred_sched *t= qdisc_priv(sch);
267 skb = qdisc_dequeue_head(sch);
270 q= t->tab[(skb->tc_index&0xf)];
272 q->backlog -= skb->len;
273 if (!q->backlog && !gred_wred_mode(t))
274 red_start_of_idle_period(&q->parms);
276 D2PRINTK("gred_dequeue: skb has bad tcindex %x\n",skb->tc_index&0xf);
281 if (gred_wred_mode(t)) {
284 D2PRINTK("no default VQ set: Results will be "
287 red_start_of_idle_period(&q->parms);
293 static unsigned int gred_drop(struct Qdisc* sch)
297 struct gred_sched_data *q;
298 struct gred_sched *t= qdisc_priv(sch);
300 skb = qdisc_dequeue_tail(sch);
302 unsigned int len = skb->len;
303 q= t->tab[(skb->tc_index&0xf)];
307 if (!q->backlog && !gred_wred_mode(t))
308 red_start_of_idle_period(&q->parms);
310 D2PRINTK("gred_dequeue: skb has bad tcindex %x\n",skb->tc_index&0xf);
313 qdisc_drop(skb, sch);
319 D2PRINTK("no default VQ set: Results might be screwed up\n");
323 red_start_of_idle_period(&q->parms);
328 static void gred_reset(struct Qdisc* sch)
331 struct gred_sched_data *q;
332 struct gred_sched *t= qdisc_priv(sch);
334 qdisc_reset_queue(sch);
336 for (i=0;i<t->DPs;i++) {
340 red_restart(&q->parms);
345 static inline void gred_destroy_vq(struct gred_sched_data *q)
350 static inline int gred_change_table_def(struct Qdisc *sch, struct rtattr *dps)
352 struct gred_sched *table = qdisc_priv(sch);
353 struct tc_gred_sopt *sopt;
356 if (dps == NULL || RTA_PAYLOAD(dps) < sizeof(*sopt))
359 sopt = RTA_DATA(dps);
361 if (sopt->DPs > MAX_DPs || sopt->DPs == 0 || sopt->def_DP >= sopt->DPs)
365 table->DPs = sopt->DPs;
366 table->def = sopt->def_DP;
369 * Every entry point to GRED is synchronized with the above code
370 * and the DP is checked against DPs, i.e. shadowed VQs can no
371 * longer be found so we can unlock right here.
373 sch_tree_unlock(sch);
376 gred_enable_rio_mode(table);
377 gred_disable_wred_mode(table);
378 if (gred_wred_mode_check(sch))
379 gred_enable_wred_mode(table);
381 gred_disable_rio_mode(table);
382 gred_disable_wred_mode(table);
385 for (i = table->DPs; i < MAX_DPs; i++) {
387 printk(KERN_WARNING "GRED: Warning: Destroying "
388 "shadowed VQ 0x%x\n", i);
389 gred_destroy_vq(table->tab[i]);
390 table->tab[i] = NULL;
399 static inline int gred_change_vq(struct Qdisc *sch, int dp,
400 struct tc_gred_qopt *ctl, int prio, u8 *stab)
402 struct gred_sched *table = qdisc_priv(sch);
403 struct gred_sched_data *q;
405 if (table->tab[dp] == NULL) {
406 table->tab[dp] = kmalloc(sizeof(*q), GFP_KERNEL);
407 if (table->tab[dp] == NULL)
409 memset(table->tab[dp], 0, sizeof(*q));
415 q->limit = ctl->limit;
418 red_end_of_idle_period(&q->parms);
420 red_set_parms(&q->parms,
421 ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Plog,
422 ctl->Scell_log, stab);
427 static int gred_change(struct Qdisc *sch, struct rtattr *opt)
429 struct gred_sched *table = qdisc_priv(sch);
430 struct tc_gred_qopt *ctl;
431 struct rtattr *tb[TCA_GRED_MAX];
432 int err = -EINVAL, prio = GRED_DEF_PRIO;
435 if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_MAX, opt))
438 if (tb[TCA_GRED_PARMS-1] == NULL && tb[TCA_GRED_STAB-1] == NULL)
439 return gred_change_table_def(sch, opt);
441 if (tb[TCA_GRED_PARMS-1] == NULL ||
442 RTA_PAYLOAD(tb[TCA_GRED_PARMS-1]) < sizeof(*ctl) ||
443 tb[TCA_GRED_STAB-1] == NULL ||
444 RTA_PAYLOAD(tb[TCA_GRED_STAB-1]) < 256)
447 ctl = RTA_DATA(tb[TCA_GRED_PARMS-1]);
448 stab = RTA_DATA(tb[TCA_GRED_STAB-1]);
450 if (ctl->DP >= table->DPs)
453 if (gred_rio_mode(table)) {
454 if (ctl->prio == 0) {
455 int def_prio = GRED_DEF_PRIO;
457 if (table->tab[table->def])
458 def_prio = table->tab[table->def]->prio;
460 printk(KERN_DEBUG "GRED: DP %u does not have a prio "
461 "setting default to %d\n", ctl->DP, def_prio);
470 err = gred_change_vq(sch, ctl->DP, ctl, prio, stab);
474 if (table->tab[table->def] == NULL) {
475 if (gred_rio_mode(table))
476 prio = table->tab[ctl->DP]->prio;
478 err = gred_change_vq(sch, table->def, ctl, prio, stab);
485 if (gred_rio_mode(table)) {
486 gred_disable_wred_mode(table);
487 if (gred_wred_mode_check(sch))
488 gred_enable_wred_mode(table);
494 sch_tree_unlock(sch);
499 static int gred_init(struct Qdisc *sch, struct rtattr *opt)
501 struct rtattr *tb[TCA_GRED_MAX];
503 if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_MAX, opt))
506 if (tb[TCA_GRED_PARMS-1] || tb[TCA_GRED_STAB-1])
509 return gred_change_table_def(sch, tb[TCA_GRED_DPS-1]);
512 static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
514 struct gred_sched *table = qdisc_priv(sch);
515 struct rtattr *parms, *opts = NULL;
517 struct tc_gred_sopt sopt = {
519 .def_DP = table->def,
520 .grio = gred_rio_mode(table),
523 opts = RTA_NEST(skb, TCA_OPTIONS);
524 RTA_PUT(skb, TCA_GRED_DPS, sizeof(sopt), &sopt);
525 parms = RTA_NEST(skb, TCA_GRED_PARMS);
527 for (i = 0; i < MAX_DPs; i++) {
528 struct gred_sched_data *q = table->tab[i];
529 struct tc_gred_qopt opt;
531 memset(&opt, 0, sizeof(opt));
534 /* hack -- fix at some point with proper message
535 This is how we indicate to tc that there is no VQ
538 opt.DP = MAX_DPs + i;
542 opt.limit = q->limit;
544 opt.backlog = q->backlog;
546 opt.qth_min = q->parms.qth_min >> q->parms.Wlog;
547 opt.qth_max = q->parms.qth_max >> q->parms.Wlog;
548 opt.Wlog = q->parms.Wlog;
549 opt.Plog = q->parms.Plog;
550 opt.Scell_log = q->parms.Scell_log;
551 opt.other = q->stats.other;
552 opt.early = q->stats.prob_drop;
553 opt.forced = q->stats.forced_drop;
554 opt.pdrop = q->stats.pdrop;
555 opt.packets = q->packetsin;
556 opt.bytesin = q->bytesin;
558 if (gred_wred_mode(table)) {
559 q->parms.qidlestart =
560 table->tab[table->def]->parms.qidlestart;
561 q->parms.qavg = table->tab[table->def]->parms.qavg;
564 opt.qave = red_calc_qavg(&q->parms, q->parms.qavg);
567 RTA_APPEND(skb, sizeof(opt), &opt);
570 RTA_NEST_END(skb, parms);
572 return RTA_NEST_END(skb, opts);
575 return RTA_NEST_CANCEL(skb, opts);
578 static void gred_destroy(struct Qdisc *sch)
580 struct gred_sched *table = qdisc_priv(sch);
583 for (i = 0;i < table->DPs; i++) {
585 gred_destroy_vq(table->tab[i]);
589 static struct Qdisc_ops gred_qdisc_ops = {
593 .priv_size = sizeof(struct gred_sched),
594 .enqueue = gred_enqueue,
595 .dequeue = gred_dequeue,
596 .requeue = gred_requeue,
600 .destroy = gred_destroy,
601 .change = gred_change,
603 .owner = THIS_MODULE,
606 static int __init gred_module_init(void)
608 return register_qdisc(&gred_qdisc_ops);
610 static void __exit gred_module_exit(void)
612 unregister_qdisc(&gred_qdisc_ops);
614 module_init(gred_module_init)
615 module_exit(gred_module_exit)
616 MODULE_LICENSE("GPL");