[PKT_SCHED]: GRED: Fix restart of idle period in WRED mode upon dequeue and drop
[linux-2.6.git] / net / sched / sch_gred.c
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/config.h>
22 #include <linux/module.h>
23 #include <linux/types.h>
24 #include <linux/kernel.h>
25 #include <linux/netdevice.h>
26 #include <linux/skbuff.h>
27 #include <net/pkt_sched.h>
28 #include <net/red.h>
29
30 #define GRED_DEF_PRIO (MAX_DPs / 2)
31 #define GRED_VQ_MASK (MAX_DPs - 1)
32
33 struct gred_sched_data;
34 struct gred_sched;
35
36 struct gred_sched_data
37 {
38         u32             limit;          /* HARD maximal queue length    */
39         u32             DP;             /* the drop pramaters */
40         u32             bytesin;        /* bytes seen on virtualQ so far*/
41         u32             packetsin;      /* packets seen on virtualQ so far*/
42         u32             backlog;        /* bytes on the virtualQ */
43         u8              prio;           /* the prio of this vq */
44
45         struct red_parms parms;
46         struct red_stats stats;
47 };
48
49 enum {
50         GRED_WRED_MODE = 1,
51         GRED_RIO_MODE,
52 };
53
54 struct gred_sched
55 {
56         struct gred_sched_data *tab[MAX_DPs];
57         unsigned long   flags;
58         u32             DPs;
59         u32             def;
60         struct red_parms wred_set;
61 };
62
63 static inline int gred_wred_mode(struct gred_sched *table)
64 {
65         return test_bit(GRED_WRED_MODE, &table->flags);
66 }
67
68 static inline void gred_enable_wred_mode(struct gred_sched *table)
69 {
70         __set_bit(GRED_WRED_MODE, &table->flags);
71 }
72
73 static inline void gred_disable_wred_mode(struct gred_sched *table)
74 {
75         __clear_bit(GRED_WRED_MODE, &table->flags);
76 }
77
78 static inline int gred_rio_mode(struct gred_sched *table)
79 {
80         return test_bit(GRED_RIO_MODE, &table->flags);
81 }
82
83 static inline void gred_enable_rio_mode(struct gred_sched *table)
84 {
85         __set_bit(GRED_RIO_MODE, &table->flags);
86 }
87
88 static inline void gred_disable_rio_mode(struct gred_sched *table)
89 {
90         __clear_bit(GRED_RIO_MODE, &table->flags);
91 }
92
93 static inline int gred_wred_mode_check(struct Qdisc *sch)
94 {
95         struct gred_sched *table = qdisc_priv(sch);
96         int i;
97
98         /* Really ugly O(n^2) but shouldn't be necessary too frequent. */
99         for (i = 0; i < table->DPs; i++) {
100                 struct gred_sched_data *q = table->tab[i];
101                 int n;
102
103                 if (q == NULL)
104                         continue;
105
106                 for (n = 0; n < table->DPs; n++)
107                         if (table->tab[n] && table->tab[n] != q &&
108                             table->tab[n]->prio == q->prio)
109                                 return 1;
110         }
111
112         return 0;
113 }
114
115 static inline unsigned int gred_backlog(struct gred_sched *table,
116                                         struct gred_sched_data *q,
117                                         struct Qdisc *sch)
118 {
119         if (gred_wred_mode(table))
120                 return sch->qstats.backlog;
121         else
122                 return q->backlog;
123 }
124
125 static inline u16 tc_index_to_dp(struct sk_buff *skb)
126 {
127         return skb->tc_index & GRED_VQ_MASK;
128 }
129
130 static inline void gred_load_wred_set(struct gred_sched *table,
131                                       struct gred_sched_data *q)
132 {
133         q->parms.qavg = table->wred_set.qavg;
134         q->parms.qidlestart = table->wred_set.qidlestart;
135 }
136
137 static inline void gred_store_wred_set(struct gred_sched *table,
138                                        struct gred_sched_data *q)
139 {
140         table->wred_set.qavg = q->parms.qavg;
141 }
142
143 static int gred_enqueue(struct sk_buff *skb, struct Qdisc* sch)
144 {
145         struct gred_sched_data *q=NULL;
146         struct gred_sched *t= qdisc_priv(sch);
147         unsigned long qavg = 0;
148         u16 dp = tc_index_to_dp(skb);
149
150         if (dp >= t->DPs  || (q = t->tab[dp]) == NULL) {
151                 dp = t->def;
152
153                 if ((q = t->tab[dp]) == NULL) {
154                         /* Pass through packets not assigned to a DP
155                          * if no default DP has been configured. This
156                          * allows for DP flows to be left untouched.
157                          */
158                         if (skb_queue_len(&sch->q) < sch->dev->tx_queue_len)
159                                 return qdisc_enqueue_tail(skb, sch);
160                         else
161                                 goto drop;
162                 }
163
164                 /* fix tc_index? --could be controvesial but needed for
165                    requeueing */
166                 skb->tc_index = (skb->tc_index & ~GRED_VQ_MASK) | dp;
167         }
168
169         /* sum up all the qaves of prios <= to ours to get the new qave */
170         if (!gred_wred_mode(t) && gred_rio_mode(t)) {
171                 int i;
172
173                 for (i = 0; i < t->DPs; i++) {
174                         if (t->tab[i] && t->tab[i]->prio < q->prio &&
175                             !red_is_idling(&t->tab[i]->parms))
176                                 qavg +=t->tab[i]->parms.qavg;
177                 }
178
179         }
180
181         q->packetsin++;
182         q->bytesin += skb->len;
183
184         if (gred_wred_mode(t))
185                 gred_load_wred_set(t, q);
186
187         q->parms.qavg = red_calc_qavg(&q->parms, gred_backlog(t, q, sch));
188
189         if (red_is_idling(&q->parms))
190                 red_end_of_idle_period(&q->parms);
191
192         if (gred_wred_mode(t))
193                 gred_store_wred_set(t, q);
194
195         switch (red_action(&q->parms, q->parms.qavg + qavg)) {
196                 case RED_DONT_MARK:
197                         break;
198
199                 case RED_PROB_MARK:
200                         sch->qstats.overlimits++;
201                         q->stats.prob_drop++;
202                         goto congestion_drop;
203
204                 case RED_HARD_MARK:
205                         sch->qstats.overlimits++;
206                         q->stats.forced_drop++;
207                         goto congestion_drop;
208         }
209
210         if (q->backlog + skb->len <= q->limit) {
211                 q->backlog += skb->len;
212                 return qdisc_enqueue_tail(skb, sch);
213         }
214
215         q->stats.pdrop++;
216 drop:
217         return qdisc_drop(skb, sch);
218
219 congestion_drop:
220         qdisc_drop(skb, sch);
221         return NET_XMIT_CN;
222 }
223
224 static int gred_requeue(struct sk_buff *skb, struct Qdisc* sch)
225 {
226         struct gred_sched *t = qdisc_priv(sch);
227         struct gred_sched_data *q;
228         u16 dp = tc_index_to_dp(skb);
229
230         if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
231                 if (net_ratelimit())
232                         printk(KERN_WARNING "GRED: Unable to relocate VQ 0x%x "
233                                "for requeue, screwing up backlog.\n",
234                                tc_index_to_dp(skb));
235         } else {
236                 if (red_is_idling(&q->parms))
237                         red_end_of_idle_period(&q->parms);
238                 q->backlog += skb->len;
239         }
240
241         return qdisc_requeue(skb, sch);
242 }
243
244 static struct sk_buff *gred_dequeue(struct Qdisc* sch)
245 {
246         struct sk_buff *skb;
247         struct gred_sched *t = qdisc_priv(sch);
248
249         skb = qdisc_dequeue_head(sch);
250
251         if (skb) {
252                 struct gred_sched_data *q;
253                 u16 dp = tc_index_to_dp(skb);
254
255                 if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
256                         if (net_ratelimit())
257                                 printk(KERN_WARNING "GRED: Unable to relocate "
258                                        "VQ 0x%x after dequeue, screwing up "
259                                        "backlog.\n", tc_index_to_dp(skb));
260                 } else {
261                         q->backlog -= skb->len;
262
263                         if (!q->backlog && !gred_wred_mode(t))
264                                 red_start_of_idle_period(&q->parms);
265                 }
266
267                 return skb;
268         }
269
270         if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
271                 red_start_of_idle_period(&t->wred_set);
272
273         return NULL;
274 }
275
276 static unsigned int gred_drop(struct Qdisc* sch)
277 {
278         struct sk_buff *skb;
279         struct gred_sched *t = qdisc_priv(sch);
280
281         skb = qdisc_dequeue_tail(sch);
282         if (skb) {
283                 unsigned int len = skb->len;
284                 struct gred_sched_data *q;
285                 u16 dp = tc_index_to_dp(skb);
286
287                 if (dp >= t->DPs || (q = t->tab[dp]) == NULL) {
288                         if (net_ratelimit())
289                                 printk(KERN_WARNING "GRED: Unable to relocate "
290                                        "VQ 0x%x while dropping, screwing up "
291                                        "backlog.\n", tc_index_to_dp(skb));
292                 } else {
293                         q->backlog -= len;
294                         q->stats.other++;
295
296                         if (!q->backlog && !gred_wred_mode(t))
297                                 red_start_of_idle_period(&q->parms);
298                 }
299
300                 qdisc_drop(skb, sch);
301                 return len;
302         }
303
304         if (gred_wred_mode(t) && !red_is_idling(&t->wred_set))
305                 red_start_of_idle_period(&t->wred_set);
306
307         return 0;
308
309 }
310
311 static void gred_reset(struct Qdisc* sch)
312 {
313         int i;
314         struct gred_sched *t = qdisc_priv(sch);
315
316         qdisc_reset_queue(sch);
317
318         for (i = 0; i < t->DPs; i++) {
319                 struct gred_sched_data *q = t->tab[i];
320
321                 if (!q)
322                         continue;
323
324                 red_restart(&q->parms);
325                 q->backlog = 0;
326         }
327 }
328
329 static inline void gred_destroy_vq(struct gred_sched_data *q)
330 {
331         kfree(q);
332 }
333
334 static inline int gred_change_table_def(struct Qdisc *sch, struct rtattr *dps)
335 {
336         struct gred_sched *table = qdisc_priv(sch);
337         struct tc_gred_sopt *sopt;
338         int i;
339
340         if (dps == NULL || RTA_PAYLOAD(dps) < sizeof(*sopt))
341                 return -EINVAL;
342
343         sopt = RTA_DATA(dps);
344
345         if (sopt->DPs > MAX_DPs || sopt->DPs == 0 || sopt->def_DP >= sopt->DPs)
346                 return -EINVAL;
347
348         sch_tree_lock(sch);
349         table->DPs = sopt->DPs;
350         table->def = sopt->def_DP;
351
352         /*
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.
356          */
357         sch_tree_unlock(sch);
358
359         if (sopt->grio) {
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);
364         } else {
365                 gred_disable_rio_mode(table);
366                 gred_disable_wred_mode(table);
367         }
368
369         for (i = table->DPs; i < MAX_DPs; i++) {
370                 if (table->tab[i]) {
371                         printk(KERN_WARNING "GRED: Warning: Destroying "
372                                "shadowed VQ 0x%x\n", i);
373                         gred_destroy_vq(table->tab[i]);
374                         table->tab[i] = NULL;
375                 }
376         }
377
378         return 0;
379 }
380
381 static inline int gred_change_vq(struct Qdisc *sch, int dp,
382                                  struct tc_gred_qopt *ctl, int prio, u8 *stab)
383 {
384         struct gred_sched *table = qdisc_priv(sch);
385         struct gred_sched_data *q;
386
387         if (table->tab[dp] == NULL) {
388                 table->tab[dp] = kmalloc(sizeof(*q), GFP_KERNEL);
389                 if (table->tab[dp] == NULL)
390                         return -ENOMEM;
391                 memset(table->tab[dp], 0, sizeof(*q));
392         }
393
394         q = table->tab[dp];
395         q->DP = dp;
396         q->prio = prio;
397         q->limit = ctl->limit;
398
399         if (q->backlog == 0)
400                 red_end_of_idle_period(&q->parms);
401
402         red_set_parms(&q->parms,
403                       ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Plog,
404                       ctl->Scell_log, stab);
405
406         return 0;
407 }
408
409 static int gred_change(struct Qdisc *sch, struct rtattr *opt)
410 {
411         struct gred_sched *table = qdisc_priv(sch);
412         struct tc_gred_qopt *ctl;
413         struct rtattr *tb[TCA_GRED_MAX];
414         int err = -EINVAL, prio = GRED_DEF_PRIO;
415         u8 *stab;
416
417         if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_MAX, opt))
418                 return -EINVAL;
419
420         if (tb[TCA_GRED_PARMS-1] == NULL && tb[TCA_GRED_STAB-1] == NULL)
421                 return gred_change_table_def(sch, opt);
422
423         if (tb[TCA_GRED_PARMS-1] == NULL ||
424             RTA_PAYLOAD(tb[TCA_GRED_PARMS-1]) < sizeof(*ctl) ||
425             tb[TCA_GRED_STAB-1] == NULL ||
426             RTA_PAYLOAD(tb[TCA_GRED_STAB-1]) < 256)
427                 return -EINVAL;
428
429         ctl = RTA_DATA(tb[TCA_GRED_PARMS-1]);
430         stab = RTA_DATA(tb[TCA_GRED_STAB-1]);
431
432         if (ctl->DP >= table->DPs)
433                 goto errout;
434
435         if (gred_rio_mode(table)) {
436                 if (ctl->prio == 0) {
437                         int def_prio = GRED_DEF_PRIO;
438
439                         if (table->tab[table->def])
440                                 def_prio = table->tab[table->def]->prio;
441
442                         printk(KERN_DEBUG "GRED: DP %u does not have a prio "
443                                "setting default to %d\n", ctl->DP, def_prio);
444
445                         prio = def_prio;
446                 } else
447                         prio = ctl->prio;
448         }
449
450         sch_tree_lock(sch);
451
452         err = gred_change_vq(sch, ctl->DP, ctl, prio, stab);
453         if (err < 0)
454                 goto errout_locked;
455
456         if (gred_rio_mode(table)) {
457                 gred_disable_wred_mode(table);
458                 if (gred_wred_mode_check(sch))
459                         gred_enable_wred_mode(table);
460         }
461
462         err = 0;
463
464 errout_locked:
465         sch_tree_unlock(sch);
466 errout:
467         return err;
468 }
469
470 static int gred_init(struct Qdisc *sch, struct rtattr *opt)
471 {
472         struct rtattr *tb[TCA_GRED_MAX];
473
474         if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_MAX, opt))
475                 return -EINVAL;
476
477         if (tb[TCA_GRED_PARMS-1] || tb[TCA_GRED_STAB-1])
478                 return -EINVAL;
479
480         return gred_change_table_def(sch, tb[TCA_GRED_DPS-1]);
481 }
482
483 static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
484 {
485         struct gred_sched *table = qdisc_priv(sch);
486         struct rtattr *parms, *opts = NULL;
487         int i;
488         struct tc_gred_sopt sopt = {
489                 .DPs    = table->DPs,
490                 .def_DP = table->def,
491                 .grio   = gred_rio_mode(table),
492         };
493
494         opts = RTA_NEST(skb, TCA_OPTIONS);
495         RTA_PUT(skb, TCA_GRED_DPS, sizeof(sopt), &sopt);
496         parms = RTA_NEST(skb, TCA_GRED_PARMS);
497
498         for (i = 0; i < MAX_DPs; i++) {
499                 struct gred_sched_data *q = table->tab[i];
500                 struct tc_gred_qopt opt;
501
502                 memset(&opt, 0, sizeof(opt));
503
504                 if (!q) {
505                         /* hack -- fix at some point with proper message
506                            This is how we indicate to tc that there is no VQ
507                            at this DP */
508
509                         opt.DP = MAX_DPs + i;
510                         goto append_opt;
511                 }
512
513                 opt.limit       = q->limit;
514                 opt.DP          = q->DP;
515                 opt.backlog     = q->backlog;
516                 opt.prio        = q->prio;
517                 opt.qth_min     = q->parms.qth_min >> q->parms.Wlog;
518                 opt.qth_max     = q->parms.qth_max >> q->parms.Wlog;
519                 opt.Wlog        = q->parms.Wlog;
520                 opt.Plog        = q->parms.Plog;
521                 opt.Scell_log   = q->parms.Scell_log;
522                 opt.other       = q->stats.other;
523                 opt.early       = q->stats.prob_drop;
524                 opt.forced      = q->stats.forced_drop;
525                 opt.pdrop       = q->stats.pdrop;
526                 opt.packets     = q->packetsin;
527                 opt.bytesin     = q->bytesin;
528
529                 if (gred_wred_mode(table)) {
530                         q->parms.qidlestart =
531                                 table->tab[table->def]->parms.qidlestart;
532                         q->parms.qavg = table->tab[table->def]->parms.qavg;
533                 }
534
535                 opt.qave = red_calc_qavg(&q->parms, q->parms.qavg);
536
537 append_opt:
538                 RTA_APPEND(skb, sizeof(opt), &opt);
539         }
540
541         RTA_NEST_END(skb, parms);
542
543         return RTA_NEST_END(skb, opts);
544
545 rtattr_failure:
546         return RTA_NEST_CANCEL(skb, opts);
547 }
548
549 static void gred_destroy(struct Qdisc *sch)
550 {
551         struct gred_sched *table = qdisc_priv(sch);
552         int i;
553
554         for (i = 0; i < table->DPs; i++) {
555                 if (table->tab[i])
556                         gred_destroy_vq(table->tab[i]);
557         }
558 }
559
560 static struct Qdisc_ops gred_qdisc_ops = {
561         .id             =       "gred",
562         .priv_size      =       sizeof(struct gred_sched),
563         .enqueue        =       gred_enqueue,
564         .dequeue        =       gred_dequeue,
565         .requeue        =       gred_requeue,
566         .drop           =       gred_drop,
567         .init           =       gred_init,
568         .reset          =       gred_reset,
569         .destroy        =       gred_destroy,
570         .change         =       gred_change,
571         .dump           =       gred_dump,
572         .owner          =       THIS_MODULE,
573 };
574
575 static int __init gred_module_init(void)
576 {
577         return register_qdisc(&gred_qdisc_ops);
578 }
579
580 static void __exit gred_module_exit(void)
581 {
582         unregister_qdisc(&gred_qdisc_ops);
583 }
584
585 module_init(gred_module_init)
586 module_exit(gred_module_exit)
587
588 MODULE_LICENSE("GPL");