sch_gred.c revision 1b34ec43c9b3de44a5420841ab293d1b2035a94c 
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 = 0; n < table->DPs; n++)
106			if (table->tab[n] && table->tab[n] != q &&
107			    table->tab[n]->prio == q->prio)
108				return 1;
109	}
110
111	return 0;
112}
113
114static inline unsigned int gred_backlog(struct gred_sched *table,
115					struct gred_sched_data *q,
116					struct Qdisc *sch)
117{
118	if (gred_wred_mode(table))
119		return sch->qstats.backlog;
120	else
121		return q->backlog;
122}
123
124static inline u16 tc_index_to_dp(struct sk_buff *skb)
125{
126	return skb->tc_index & GRED_VQ_MASK;
127}
128
129static inline void gred_load_wred_set(const struct gred_sched *table,
130				      struct gred_sched_data *q)
131{
132	q->vars.qavg = table->wred_set.qavg;
133	q->vars.qidlestart = table->wred_set.qidlestart;
134}
135
136static inline void gred_store_wred_set(struct gred_sched *table,
137				       struct gred_sched_data *q)
138{
139	table->wred_set.qavg = q->vars.qavg;
140}
141
142static inline int gred_use_ecn(struct gred_sched *t)
143{
144	return t->red_flags & TC_RED_ECN;
145}
146
147static inline int gred_use_harddrop(struct gred_sched *t)
148{
149	return t->red_flags & TC_RED_HARDDROP;
150}
151
152static int gred_enqueue(struct sk_buff *skb, struct Qdisc *sch)
153{
154	struct gred_sched_data *q = NULL;
155	struct gred_sched *t = qdisc_priv(sch);
156	unsigned long qavg = 0;
157	u16 dp = tc_index_to_dp(skb);
158
159	if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
160		dp = t->def;
161
162		q = t->tab[dp];
163		if (!q) {
164			/* Pass through packets not assigned to a DP
165			 * if no default DP has been configured. This
166			 * allows for DP flows to be left untouched.
167			 */
168			if (skb_queue_len(&sch->q) < qdisc_dev(sch)->tx_queue_len)
169				return qdisc_enqueue_tail(skb, sch);
170			else
171				goto drop;
172		}
173
174		/* fix tc_index? --could be controversial but needed for
175		   requeueing */
176		skb->tc_index = (skb->tc_index & ~GRED_VQ_MASK) | dp;
177	}
178
179	/* sum up all the qaves of prios <= to ours to get the new qave */
180	if (!gred_wred_mode(t) && gred_rio_mode(t)) {
181		int i;
182
183		for (i = 0; i < t->DPs; i++) {
184			if (t->tab[i] && t->tab[i]->prio < q->prio &&
185			    !red_is_idling(&t->tab[i]->vars))
186				qavg += t->tab[i]->vars.qavg;
187		}
188
189	}
190
191	q->packetsin++;
192	q->bytesin += qdisc_pkt_len(skb);
193
194	if (gred_wred_mode(t))
195		gred_load_wred_set(t, q);
196
197	q->vars.qavg = red_calc_qavg(&q->parms,
198				     &q->vars,
199				     gred_backlog(t, q, sch));
200
201	if (red_is_idling(&q->vars))
202		red_end_of_idle_period(&q->vars);
203
204	if (gred_wred_mode(t))
205		gred_store_wred_set(t, q);
206
207	switch (red_action(&q->parms, &q->vars, q->vars.qavg + qavg)) {
208	case RED_DONT_MARK:
209		break;
210
211	case RED_PROB_MARK:
212		sch->qstats.overlimits++;
213		if (!gred_use_ecn(t) || !INET_ECN_set_ce(skb)) {
214			q->stats.prob_drop++;
215			goto congestion_drop;
216		}
217
218		q->stats.prob_mark++;
219		break;
220
221	case RED_HARD_MARK:
222		sch->qstats.overlimits++;
223		if (gred_use_harddrop(t) || !gred_use_ecn(t) ||
224		    !INET_ECN_set_ce(skb)) {
225			q->stats.forced_drop++;
226			goto congestion_drop;
227		}
228		q->stats.forced_mark++;
229		break;
230	}
231
232	if (q->backlog + qdisc_pkt_len(skb) <= q->limit) {
233		q->backlog += qdisc_pkt_len(skb);
234		return qdisc_enqueue_tail(skb, sch);
235	}
236
237	q->stats.pdrop++;
238drop:
239	return qdisc_drop(skb, sch);
240
241congestion_drop:
242	qdisc_drop(skb, sch);
243	return NET_XMIT_CN;
244}
245
246static struct sk_buff *gred_dequeue(struct Qdisc *sch)
247{
248	struct sk_buff *skb;
249	struct gred_sched *t = qdisc_priv(sch);
250
251	skb = qdisc_dequeue_head(sch);
252
253	if (skb) {
254		struct gred_sched_data *q;
255		u16 dp = tc_index_to_dp(skb);
256
257		if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
258			if (net_ratelimit())
259				pr_warning("GRED: Unable to relocate VQ 0x%x "
260					   "after dequeue, screwing up "
261					   "backlog.\n", tc_index_to_dp(skb));
262		} else {
263			q->backlog -= qdisc_pkt_len(skb);
264
265			if (!q->backlog && !gred_wred_mode(t))
266				red_start_of_idle_period(&q->vars);
267		}
268
269		return skb;
270	}
271
272	if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
273		red_start_of_idle_period(&t->wred_set);
274
275	return NULL;
276}
277
278static unsigned int gred_drop(struct Qdisc *sch)
279{
280	struct sk_buff *skb;
281	struct gred_sched *t = qdisc_priv(sch);
282
283	skb = qdisc_dequeue_tail(sch);
284	if (skb) {
285		unsigned int len = qdisc_pkt_len(skb);
286		struct gred_sched_data *q;
287		u16 dp = tc_index_to_dp(skb);
288
289		if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
290			if (net_ratelimit())
291				pr_warning("GRED: Unable to relocate VQ 0x%x "
292					   "while dropping, screwing up "
293					   "backlog.\n", tc_index_to_dp(skb));
294		} else {
295			q->backlog -= len;
296			q->stats.other++;
297
298			if (!q->backlog && !gred_wred_mode(t))
299				red_start_of_idle_period(&q->vars);
300		}
301
302		qdisc_drop(skb, sch);
303		return len;
304	}
305
306	if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
307		red_start_of_idle_period(&t->wred_set);
308
309	return 0;
310
311}
312
313static void gred_reset(struct Qdisc *sch)
314{
315	int i;
316	struct gred_sched *t = qdisc_priv(sch);
317
318	qdisc_reset_queue(sch);
319
320	for (i = 0; i < t->DPs; i++) {
321		struct gred_sched_data *q = t->tab[i];
322
323		if (!q)
324			continue;
325
326		red_restart(&q->vars);
327		q->backlog = 0;
328	}
329}
330
331static inline void gred_destroy_vq(struct gred_sched_data *q)
332{
333	kfree(q);
334}
335
336static inline int gred_change_table_def(struct Qdisc *sch, struct nlattr *dps)
337{
338	struct gred_sched *table = qdisc_priv(sch);
339	struct tc_gred_sopt *sopt;
340	int i;
341
342	if (dps == NULL)
343		return -EINVAL;
344
345	sopt = nla_data(dps);
346
347	if (sopt->DPs > MAX_DPs || sopt->DPs == 0 || sopt->def_DP >= sopt->DPs)
348		return -EINVAL;
349
350	sch_tree_lock(sch);
351	table->DPs = sopt->DPs;
352	table->def = sopt->def_DP;
353	table->red_flags = sopt->flags;
354
355	/*
356	 * Every entry point to GRED is synchronized with the above code
357	 * and the DP is checked against DPs, i.e. shadowed VQs can no
358	 * longer be found so we can unlock right here.
359	 */
360	sch_tree_unlock(sch);
361
362	if (sopt->grio) {
363		gred_enable_rio_mode(table);
364		gred_disable_wred_mode(table);
365		if (gred_wred_mode_check(sch))
366			gred_enable_wred_mode(table);
367	} else {
368		gred_disable_rio_mode(table);
369		gred_disable_wred_mode(table);
370	}
371
372	for (i = table->DPs; i < MAX_DPs; i++) {
373		if (table->tab[i]) {
374			pr_warning("GRED: Warning: Destroying "
375				   "shadowed VQ 0x%x\n", i);
376			gred_destroy_vq(table->tab[i]);
377			table->tab[i] = NULL;
378		}
379	}
380
381	return 0;
382}
383
384static inline int gred_change_vq(struct Qdisc *sch, int dp,
385				 struct tc_gred_qopt *ctl, int prio,
386				 u8 *stab, u32 max_P,
387				 struct gred_sched_data **prealloc)
388{
389	struct gred_sched *table = qdisc_priv(sch);
390	struct gred_sched_data *q = table->tab[dp];
391
392	if (!q) {
393		table->tab[dp] = q = *prealloc;
394		*prealloc = NULL;
395		if (!q)
396			return -ENOMEM;
397	}
398
399	q->DP = dp;
400	q->prio = prio;
401	q->limit = ctl->limit;
402
403	if (q->backlog == 0)
404		red_end_of_idle_period(&q->vars);
405
406	red_set_parms(&q->parms,
407		      ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Plog,
408		      ctl->Scell_log, stab, max_P);
409	red_set_vars(&q->vars);
410	return 0;
411}
412
413static const struct nla_policy gred_policy[TCA_GRED_MAX + 1] = {
414	[TCA_GRED_PARMS]	= { .len = sizeof(struct tc_gred_qopt) },
415	[TCA_GRED_STAB]		= { .len = 256 },
416	[TCA_GRED_DPS]		= { .len = sizeof(struct tc_gred_sopt) },
417	[TCA_GRED_MAX_P]	= { .type = NLA_U32 },
418};
419
420static int gred_change(struct Qdisc *sch, struct nlattr *opt)
421{
422	struct gred_sched *table = qdisc_priv(sch);
423	struct tc_gred_qopt *ctl;
424	struct nlattr *tb[TCA_GRED_MAX + 1];
425	int err, prio = GRED_DEF_PRIO;
426	u8 *stab;
427	u32 max_P;
428	struct gred_sched_data *prealloc;
429
430	if (opt == NULL)
431		return -EINVAL;
432
433	err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
434	if (err < 0)
435		return err;
436
437	if (tb[TCA_GRED_PARMS] == NULL && tb[TCA_GRED_STAB] == NULL)
438		return gred_change_table_def(sch, opt);
439
440	if (tb[TCA_GRED_PARMS] == NULL ||
441	    tb[TCA_GRED_STAB] == NULL)
442		return -EINVAL;
443
444	max_P = tb[TCA_GRED_MAX_P] ? nla_get_u32(tb[TCA_GRED_MAX_P]) : 0;
445
446	err = -EINVAL;
447	ctl = nla_data(tb[TCA_GRED_PARMS]);
448	stab = nla_data(tb[TCA_GRED_STAB]);
449
450	if (ctl->DP >= table->DPs)
451		goto errout;
452
453	if (gred_rio_mode(table)) {
454		if (ctl->prio == 0) {
455			int def_prio = GRED_DEF_PRIO;
456
457			if (table->tab[table->def])
458				def_prio = table->tab[table->def]->prio;
459
460			printk(KERN_DEBUG "GRED: DP %u does not have a prio "
461			       "setting default to %d\n", ctl->DP, def_prio);
462
463			prio = def_prio;
464		} else
465			prio = ctl->prio;
466	}
467
468	prealloc = kzalloc(sizeof(*prealloc), GFP_KERNEL);
469	sch_tree_lock(sch);
470
471	err = gred_change_vq(sch, ctl->DP, ctl, prio, stab, max_P, &prealloc);
472	if (err < 0)
473		goto errout_locked;
474
475	if (gred_rio_mode(table)) {
476		gred_disable_wred_mode(table);
477		if (gred_wred_mode_check(sch))
478			gred_enable_wred_mode(table);
479	}
480
481	err = 0;
482
483errout_locked:
484	sch_tree_unlock(sch);
485	kfree(prealloc);
486errout:
487	return err;
488}
489
490static int gred_init(struct Qdisc *sch, struct nlattr *opt)
491{
492	struct nlattr *tb[TCA_GRED_MAX + 1];
493	int err;
494
495	if (opt == NULL)
496		return -EINVAL;
497
498	err = nla_parse_nested(tb, TCA_GRED_MAX, opt, gred_policy);
499	if (err < 0)
500		return err;
501
502	if (tb[TCA_GRED_PARMS] || tb[TCA_GRED_STAB])
503		return -EINVAL;
504
505	return gred_change_table_def(sch, tb[TCA_GRED_DPS]);
506}
507
508static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
509{
510	struct gred_sched *table = qdisc_priv(sch);
511	struct nlattr *parms, *opts = NULL;
512	int i;
513	u32 max_p[MAX_DPs];
514	struct tc_gred_sopt sopt = {
515		.DPs	= table->DPs,
516		.def_DP	= table->def,
517		.grio	= gred_rio_mode(table),
518		.flags	= table->red_flags,
519	};
520
521	opts = nla_nest_start(skb, TCA_OPTIONS);
522	if (opts == NULL)
523		goto nla_put_failure;
524	if (nla_put(skb, TCA_GRED_DPS, sizeof(sopt), &sopt))
525		goto nla_put_failure;
526
527	for (i = 0; i < MAX_DPs; i++) {
528		struct gred_sched_data *q = table->tab[i];
529
530		max_p[i] = q ? q->parms.max_P : 0;
531	}
532	if (nla_put(skb, TCA_GRED_MAX_P, sizeof(max_p), max_p))
533		goto nla_put_failure;
534
535	parms = nla_nest_start(skb, TCA_GRED_PARMS);
536	if (parms == NULL)
537		goto nla_put_failure;
538
539	for (i = 0; i < MAX_DPs; i++) {
540		struct gred_sched_data *q = table->tab[i];
541		struct tc_gred_qopt opt;
542
543		memset(&opt, 0, sizeof(opt));
544
545		if (!q) {
546			/* hack -- fix at some point with proper message
547			   This is how we indicate to tc that there is no VQ
548			   at this DP */
549
550			opt.DP = MAX_DPs + i;
551			goto append_opt;
552		}
553
554		opt.limit	= q->limit;
555		opt.DP		= q->DP;
556		opt.backlog	= q->backlog;
557		opt.prio	= q->prio;
558		opt.qth_min	= q->parms.qth_min >> q->parms.Wlog;
559		opt.qth_max	= q->parms.qth_max >> q->parms.Wlog;
560		opt.Wlog	= q->parms.Wlog;
561		opt.Plog	= q->parms.Plog;
562		opt.Scell_log	= q->parms.Scell_log;
563		opt.other	= q->stats.other;
564		opt.early	= q->stats.prob_drop;
565		opt.forced	= q->stats.forced_drop;
566		opt.pdrop	= q->stats.pdrop;
567		opt.packets	= q->packetsin;
568		opt.bytesin	= q->bytesin;
569
570		if (gred_wred_mode(table)) {
571			q->vars.qidlestart =
572				table->tab[table->def]->vars.qidlestart;
573			q->vars.qavg = table->tab[table->def]->vars.qavg;
574		}
575
576		opt.qave = red_calc_qavg(&q->parms, &q->vars, q->vars.qavg);
577
578append_opt:
579		if (nla_append(skb, sizeof(opt), &opt) < 0)
580			goto nla_put_failure;
581	}
582
583	nla_nest_end(skb, parms);
584
585	return nla_nest_end(skb, opts);
586
587nla_put_failure:
588	nla_nest_cancel(skb, opts);
589	return -EMSGSIZE;
590}
591
592static void gred_destroy(struct Qdisc *sch)
593{
594	struct gred_sched *table = qdisc_priv(sch);
595	int i;
596
597	for (i = 0; i < table->DPs; i++) {
598		if (table->tab[i])
599			gred_destroy_vq(table->tab[i]);
600	}
601}
602
603static struct Qdisc_ops gred_qdisc_ops __read_mostly = {
604	.id		=	"gred",
605	.priv_size	=	sizeof(struct gred_sched),
606	.enqueue	=	gred_enqueue,
607	.dequeue	=	gred_dequeue,
608	.peek		=	qdisc_peek_head,
609	.drop		=	gred_drop,
610	.init		=	gred_init,
611	.reset		=	gred_reset,
612	.destroy	=	gred_destroy,
613	.change		=	gred_change,
614	.dump		=	gred_dump,
615	.owner		=	THIS_MODULE,
616};
617
618static int __init gred_module_init(void)
619{
620	return register_qdisc(&gred_qdisc_ops);
621}
622
623static void __exit gred_module_exit(void)
624{
625	unregister_qdisc(&gred_qdisc_ops);
626}
627
628module_init(gred_module_init)
629module_exit(gred_module_exit)
630
631MODULE_LICENSE("GPL");
632