sch_gred.c revision 1fe37b106b039d9358fd1211c39b1fa199e547a8 
1/*
2 * net/sched/sch_gred.c	Generic Random Early Detection queue.
3 *
4 *
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.
9 *
10 * Authors:    J Hadi Salim (hadi@cyberus.ca) 1998-2002
11 *
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
15 *		         from Ren Liu
16 *		       - More error checks
17 *
18 *  For all the glorious comments look at include/net/red.h
19 */
20
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>
27#include <net/red.h>
28
29#define GRED_DEF_PRIO (MAX_DPs / 2)
30#define GRED_VQ_MASK (MAX_DPs - 1)
31
32struct gred_sched_data;
33struct gred_sched;
34
35struct gred_sched_data {
36	u32		limit;		/* HARD maximal queue length	*/
37	u32		DP;		/* the drop parameters */
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 */
42
43	struct red_parms parms;
44	struct red_vars  vars;
45	struct red_stats stats;
46};
47
48enum {
49	GRED_WRED_MODE = 1,
50	GRED_RIO_MODE,
51};
52
53struct gred_sched {
54	struct gred_sched_data *tab[MAX_DPs];
55	unsigned long	flags;
56	u32		red_flags;
57	u32 		DPs;
58	u32 		def;
59	struct red_vars wred_set;
60};
61
62static inline int gred_wred_mode(struct gred_sched *table)
63{
64	return test_bit(GRED_WRED_MODE, &table->flags);
65}
66
67static inline void gred_enable_wred_mode(struct gred_sched *table)
68{
69	__set_bit(GRED_WRED_MODE, &table->flags);
70}
71
72static inline void gred_disable_wred_mode(struct gred_sched *table)
73{
74	__clear_bit(GRED_WRED_MODE, &table->flags);
75}
76
77static inline int gred_rio_mode(struct gred_sched *table)
78{
79	return test_bit(GRED_RIO_MODE, &table->flags);
80}
81
82static inline void gred_enable_rio_mode(struct gred_sched *table)
83{
84	__set_bit(GRED_RIO_MODE, &table->flags);
85}
86
87static inline void gred_disable_rio_mode(struct gred_sched *table)
88{
89	__clear_bit(GRED_RIO_MODE, &table->flags);
90}
91
92static inline int gred_wred_mode_check(struct Qdisc *sch)
93{
94	struct gred_sched *table = qdisc_priv(sch);
95	int i;
96
97	/* Really ugly O(n^2) but shouldn't be necessary too frequent. */
98	for (i = 0; i < table->DPs; i++) {
99		struct gred_sched_data *q = table->tab[i];
100		int n;
101
102		if (q == NULL)
103			continue;
104
105		for (n = i + 1; n < table->DPs; n++)
106			if (table->tab[n] && table->tab[n]->prio == q->prio)
107				return 1;
108	}
109
110	return 0;
111}
112
113static inline unsigned int gred_backlog(struct gred_sched *table,
114					struct gred_sched_data *q,
115					struct Qdisc *sch)
116{
117	if (gred_wred_mode(table))
118		return sch->qstats.backlog;
119	else
120		return q->backlog;
121}
122
123static inline u16 tc_index_to_dp(struct sk_buff *skb)
124{
125	return skb->tc_index & GRED_VQ_MASK;
126}
127
128static inline void gred_load_wred_set(const struct gred_sched *table,
129				      struct gred_sched_data *q)
130{
131	q->vars.qavg = table->wred_set.qavg;
132	q->vars.qidlestart = table->wred_set.qidlestart;
133}
134
135static inline void gred_store_wred_set(struct gred_sched *table,
136				       struct gred_sched_data *q)
137{
138	table->wred_set.qavg = q->vars.qavg;
139}
140
141static inline int gred_use_ecn(struct gred_sched *t)
142{
143	return t->red_flags & TC_RED_ECN;
144}
145
146static inline int gred_use_harddrop(struct gred_sched *t)
147{
148	return t->red_flags & TC_RED_HARDDROP;
149}
150
151static int gred_enqueue(struct sk_buff *skb, struct Qdisc *sch)
152{
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);
157
158	if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
159		dp = t->def;
160
161		q = t->tab[dp];
162		if (!q) {
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.
166			 */
167			if (skb_queue_len(&sch->q) < qdisc_dev(sch)->tx_queue_len)
168				return qdisc_enqueue_tail(skb, sch);
169			else
170				goto drop;
171		}
172
173		/* fix tc_index? --could be controversial but needed for
174		   requeueing */
175		skb->tc_index = (skb->tc_index & ~GRED_VQ_MASK) | dp;
176	}
177
178	/* sum up all the qaves of prios < ours to get the new qave */
179	if (!gred_wred_mode(t) && gred_rio_mode(t)) {
180		int i;
181
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]->vars))
185				qavg += t->tab[i]->vars.qavg;
186		}
187
188	}
189
190	q->packetsin++;
191	q->bytesin += qdisc_pkt_len(skb);
192
193	if (gred_wred_mode(t))
194		gred_load_wred_set(t, q);
195
196	q->vars.qavg = red_calc_qavg(&q->parms,
197				     &q->vars,
198				     gred_backlog(t, q, sch));
199
200	if (red_is_idling(&q->vars))
201		red_end_of_idle_period(&q->vars);
202
203	if (gred_wred_mode(t))
204		gred_store_wred_set(t, q);
205
206	switch (red_action(&q->parms, &q->vars, q->vars.qavg + qavg)) {
207	case RED_DONT_MARK:
208		break;
209
210	case RED_PROB_MARK:
211		sch->qstats.overlimits++;
212		if (!gred_use_ecn(t) || !INET_ECN_set_ce(skb)) {
213			q->stats.prob_drop++;
214			goto congestion_drop;
215		}
216
217		q->stats.prob_mark++;
218		break;
219
220	case RED_HARD_MARK:
221		sch->qstats.overlimits++;
222		if (gred_use_harddrop(t) || !gred_use_ecn(t) ||
223		    !INET_ECN_set_ce(skb)) {
224			q->stats.forced_drop++;
225			goto congestion_drop;
226		}
227		q->stats.forced_mark++;
228		break;
229	}
230
231	if (q->backlog + qdisc_pkt_len(skb) <= q->limit) {
232		q->backlog += qdisc_pkt_len(skb);
233		return qdisc_enqueue_tail(skb, sch);
234	}
235
236	q->stats.pdrop++;
237drop:
238	return qdisc_drop(skb, sch);
239
240congestion_drop:
241	qdisc_drop(skb, sch);
242	return NET_XMIT_CN;
243}
244
245static struct sk_buff *gred_dequeue(struct Qdisc *sch)
246{
247	struct sk_buff *skb;
248	struct gred_sched *t = qdisc_priv(sch);
249
250	skb = qdisc_dequeue_head(sch);
251
252	if (skb) {
253		struct gred_sched_data *q;
254		u16 dp = tc_index_to_dp(skb);
255
256		if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
257			net_warn_ratelimited("GRED: Unable to relocate VQ 0x%x after dequeue, screwing up backlog\n",
258					     tc_index_to_dp(skb));
259		} else {
260			q->backlog -= qdisc_pkt_len(skb);
261
262			if (!q->backlog && !gred_wred_mode(t))
263				red_start_of_idle_period(&q->vars);
264		}
265
266		return skb;
267	}
268
269	if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
270		red_start_of_idle_period(&t->wred_set);
271
272	return NULL;
273}
274
275static unsigned int gred_drop(struct Qdisc *sch)
276{
277	struct sk_buff *skb;
278	struct gred_sched *t = qdisc_priv(sch);
279
280	skb = qdisc_dequeue_tail(sch);
281	if (skb) {
282		unsigned int len = qdisc_pkt_len(skb);
283		struct gred_sched_data *q;
284		u16 dp = tc_index_to_dp(skb);
285
286		if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
287			net_warn_ratelimited("GRED: Unable to relocate VQ 0x%x while dropping, screwing up backlog\n",
288					     tc_index_to_dp(skb));
289		} else {
290			q->backlog -= len;
291			q->stats.other++;
292
293			if (!q->backlog && !gred_wred_mode(t))
294				red_start_of_idle_period(&q->vars);
295		}
296
297		qdisc_drop(skb, sch);
298		return len;
299	}
300
301	if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
302		red_start_of_idle_period(&t->wred_set);
303
304	return 0;
305
306}
307
308static void gred_reset(struct Qdisc *sch)
309{
310	int i;
311	struct gred_sched *t = qdisc_priv(sch);
312
313	qdisc_reset_queue(sch);
314
315	for (i = 0; i < t->DPs; i++) {
316		struct gred_sched_data *q = t->tab[i];
317
318		if (!q)
319			continue;
320
321		red_restart(&q->vars);
322		q->backlog = 0;
323	}
324}
325
326static inline void gred_destroy_vq(struct gred_sched_data *q)
327{
328	kfree(q);
329}
330
331static inline int gred_change_table_def(struct Qdisc *sch, struct nlattr *dps)
332{
333	struct gred_sched *table = qdisc_priv(sch);
334	struct tc_gred_sopt *sopt;
335	int i;
336
337	if (dps == NULL)
338		return -EINVAL;
339
340	sopt = nla_data(dps);
341
342	if (sopt->DPs > MAX_DPs || sopt->DPs == 0 || sopt->def_DP >= sopt->DPs)
343		return -EINVAL;
344
345	sch_tree_lock(sch);
346	table->DPs = sopt->DPs;
347	table->def = sopt->def_DP;
348	table->red_flags = sopt->flags;
349
350	/*
351	 * Every entry point to GRED is synchronized with the above code
352	 * and the DP is checked against DPs, i.e. shadowed VQs can no
353	 * longer be found so we can unlock right here.
354	 */
355	sch_tree_unlock(sch);
356
357	if (sopt->grio) {
358		gred_enable_rio_mode(table);
359		gred_disable_wred_mode(table);
360		if (gred_wred_mode_check(sch))
361			gred_enable_wred_mode(table);
362	} else {
363		gred_disable_rio_mode(table);
364		gred_disable_wred_mode(table);
365	}
366
367	for (i = table->DPs; i < MAX_DPs; i++) {
368		if (table->tab[i]) {
369			pr_warning("GRED: Warning: Destroying "
370				   "shadowed VQ 0x%x\n", i);
371			gred_destroy_vq(table->tab[i]);
372			table->tab[i] = NULL;
373		}
374	}
375
376	return 0;
377}
378
379static inline int gred_change_vq(struct Qdisc *sch, int dp,
380				 struct tc_gred_qopt *ctl, int prio,
381				 u8 *stab, u32 max_P,
382				 struct gred_sched_data **prealloc)
383{
384	struct gred_sched *table = qdisc_priv(sch);
385	struct gred_sched_data *q = table->tab[dp];
386
387	if (!q) {
388		table->tab[dp] = q = *prealloc;
389		*prealloc = NULL;
390		if (!q)
391			return -ENOMEM;
392	}
393
394	q->DP = dp;
395	q->prio = prio;
396	q->limit = ctl->limit;
397
398	if (q->backlog == 0)
399		red_end_of_idle_period(&q->vars);
400
401	red_set_parms(&q->parms,
402		      ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Plog,
403		      ctl->Scell_log, stab, max_P);
404	red_set_vars(&q->vars);
405	return 0;
406}
407
408static const struct nla_policy gred_policy[TCA_GRED_MAX + 1] = {
409	[TCA_GRED_PARMS]	= { .len = sizeof(struct tc_gred_qopt) },
410	[TCA_GRED_STAB]		= { .len = 256 },
411	[TCA_GRED_DPS]		= { .len = sizeof(struct tc_gred_sopt) },
412	[TCA_GRED_MAX_P]	= { .type = NLA_U32 },
413};
414
415static int gred_change(struct Qdisc *sch, struct nlattr *opt)
416{
417	struct gred_sched *table = qdisc_priv(sch);
418	struct tc_gred_qopt *ctl;
419	struct nlattr *tb[TCA_GRED_MAX + 1];
420	int err, prio = GRED_DEF_PRIO;
421	u8 *stab;
422	u32 max_P;
423	struct gred_sched_data *prealloc;
424
425	if (opt == NULL)
426		return -EINVAL;
427
428	err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
429	if (err < 0)
430		return err;
431
432	if (tb[TCA_GRED_PARMS] == NULL && tb[TCA_GRED_STAB] == NULL)
433		return gred_change_table_def(sch, opt);
434
435	if (tb[TCA_GRED_PARMS] == NULL ||
436	    tb[TCA_GRED_STAB] == NULL)
437		return -EINVAL;
438
439	max_P = tb[TCA_GRED_MAX_P] ? nla_get_u32(tb[TCA_GRED_MAX_P]) : 0;
440
441	err = -EINVAL;
442	ctl = nla_data(tb[TCA_GRED_PARMS]);
443	stab = nla_data(tb[TCA_GRED_STAB]);
444
445	if (ctl->DP >= table->DPs)
446		goto errout;
447
448	if (gred_rio_mode(table)) {
449		if (ctl->prio == 0) {
450			int def_prio = GRED_DEF_PRIO;
451
452			if (table->tab[table->def])
453				def_prio = table->tab[table->def]->prio;
454
455			printk(KERN_DEBUG "GRED: DP %u does not have a prio "
456			       "setting default to %d\n", ctl->DP, def_prio);
457
458			prio = def_prio;
459		} else
460			prio = ctl->prio;
461	}
462
463	prealloc = kzalloc(sizeof(*prealloc), GFP_KERNEL);
464	sch_tree_lock(sch);
465
466	err = gred_change_vq(sch, ctl->DP, ctl, prio, stab, max_P, &prealloc);
467	if (err < 0)
468		goto errout_locked;
469
470	if (gred_rio_mode(table)) {
471		gred_disable_wred_mode(table);
472		if (gred_wred_mode_check(sch))
473			gred_enable_wred_mode(table);
474	}
475
476	err = 0;
477
478errout_locked:
479	sch_tree_unlock(sch);
480	kfree(prealloc);
481errout:
482	return err;
483}
484
485static int gred_init(struct Qdisc *sch, struct nlattr *opt)
486{
487	struct nlattr *tb[TCA_GRED_MAX + 1];
488	int err;
489
490	if (opt == NULL)
491		return -EINVAL;
492
493	err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
494	if (err < 0)
495		return err;
496
497	if (tb[TCA_GRED_PARMS] || tb[TCA_GRED_STAB])
498		return -EINVAL;
499
500	return gred_change_table_def(sch, tb[TCA_GRED_DPS]);
501}
502
503static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
504{
505	struct gred_sched *table = qdisc_priv(sch);
506	struct nlattr *parms, *opts = NULL;
507	int i;
508	u32 max_p[MAX_DPs];
509	struct tc_gred_sopt sopt = {
510		.DPs	= table->DPs,
511		.def_DP	= table->def,
512		.grio	= gred_rio_mode(table),
513		.flags	= table->red_flags,
514	};
515
516	opts = nla_nest_start(skb, TCA_OPTIONS);
517	if (opts == NULL)
518		goto nla_put_failure;
519	if (nla_put(skb, TCA_GRED_DPS, sizeof(sopt), &sopt))
520		goto nla_put_failure;
521
522	for (i = 0; i < MAX_DPs; i++) {
523		struct gred_sched_data *q = table->tab[i];
524
525		max_p[i] = q ? q->parms.max_P : 0;
526	}
527	if (nla_put(skb, TCA_GRED_MAX_P, sizeof(max_p), max_p))
528		goto nla_put_failure;
529
530	parms = nla_nest_start(skb, TCA_GRED_PARMS);
531	if (parms == NULL)
532		goto nla_put_failure;
533
534	for (i = 0; i < MAX_DPs; i++) {
535		struct gred_sched_data *q = table->tab[i];
536		struct tc_gred_qopt opt;
537		unsigned long qavg;
538
539		memset(&opt, 0, sizeof(opt));
540
541		if (!q) {
542			/* hack -- fix at some point with proper message
543			   This is how we indicate to tc that there is no VQ
544			   at this DP */
545
546			opt.DP = MAX_DPs + i;
547			goto append_opt;
548		}
549
550		opt.limit	= q->limit;
551		opt.DP		= q->DP;
552		opt.backlog	= q->backlog;
553		opt.prio	= q->prio;
554		opt.qth_min	= q->parms.qth_min >> q->parms.Wlog;
555		opt.qth_max	= q->parms.qth_max >> q->parms.Wlog;
556		opt.Wlog	= q->parms.Wlog;
557		opt.Plog	= q->parms.Plog;
558		opt.Scell_log	= q->parms.Scell_log;
559		opt.other	= q->stats.other;
560		opt.early	= q->stats.prob_drop;
561		opt.forced	= q->stats.forced_drop;
562		opt.pdrop	= q->stats.pdrop;
563		opt.packets	= q->packetsin;
564		opt.bytesin	= q->bytesin;
565
566		if (gred_wred_mode(table))
567			gred_load_wred_set(table, q);
568
569		qavg = red_calc_qavg(&q->parms, &q->vars,
570				     q->vars.qavg >> q->parms.Wlog);
571		opt.qave = qavg >> q->parms.Wlog;
572
573append_opt:
574		if (nla_append(skb, sizeof(opt), &opt) < 0)
575			goto nla_put_failure;
576	}
577
578	nla_nest_end(skb, parms);
579
580	return nla_nest_end(skb, opts);
581
582nla_put_failure:
583	nla_nest_cancel(skb, opts);
584	return -EMSGSIZE;
585}
586
587static void gred_destroy(struct Qdisc *sch)
588{
589	struct gred_sched *table = qdisc_priv(sch);
590	int i;
591
592	for (i = 0; i < table->DPs; i++) {
593		if (table->tab[i])
594			gred_destroy_vq(table->tab[i]);
595	}
596}
597
598static struct Qdisc_ops gred_qdisc_ops __read_mostly = {
599	.id		=	"gred",
600	.priv_size	=	sizeof(struct gred_sched),
601	.enqueue	=	gred_enqueue,
602	.dequeue	=	gred_dequeue,
603	.peek		=	qdisc_peek_head,
604	.drop		=	gred_drop,
605	.init		=	gred_init,
606	.reset		=	gred_reset,
607	.destroy	=	gred_destroy,
608	.change		=	gred_change,
609	.dump		=	gred_dump,
610	.owner		=	THIS_MODULE,
611};
612
613static int __init gred_module_init(void)
614{
615	return register_qdisc(&gred_qdisc_ops);
616}
617
618static void __exit gred_module_exit(void)
619{
620	unregister_qdisc(&gred_qdisc_ops);
621}
622
623module_init(gred_module_init)
624module_exit(gred_module_exit)
625
626MODULE_LICENSE("GPL");
627