[PKT_SCHED]: GRED: Introduce tc_index_to_dp()
[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  *
19  *
20  *  For all the glorious comments look at Alexey's sch_red.c
21  */
22
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>
32 #include <linux/mm.h>
33 #include <linux/socket.h>
34 #include <linux/sockios.h>
35 #include <linux/in.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>
43 #include <net/ip.h>
44 #include <net/route.h>
45 #include <linux/skbuff.h>
46 #include <net/sock.h>
47 #include <net/pkt_sched.h>
48 #include <net/red.h>
49
50 #if 1 /* control */
51 #define DPRINTK(format,args...) printk(KERN_DEBUG format,##args)
52 #else
53 #define DPRINTK(format,args...)
54 #endif
55
56 #if 0 /* data */
57 #define D2PRINTK(format,args...) printk(KERN_DEBUG format,##args)
58 #else
59 #define D2PRINTK(format,args...)
60 #endif
61
62 #define GRED_DEF_PRIO (MAX_DPs / 2)
63 #define GRED_VQ_MASK (MAX_DPs - 1)
64
65 struct gred_sched_data;
66 struct gred_sched;
67
68 struct gred_sched_data
69 {
70         u32             limit;          /* HARD maximal queue length    */
71         u32             DP;             /* the drop pramaters */
72         u32             bytesin;        /* bytes seen on virtualQ so far*/
73         u32             packetsin;      /* packets seen on virtualQ so far*/
74         u32             backlog;        /* bytes on the virtualQ */
75         u8              prio;        /* the prio of this vq */
76
77         struct red_parms parms;
78         struct red_stats stats;
79 };
80
81 enum {
82         GRED_WRED_MODE = 1,
83         GRED_RIO_MODE,
84 };
85
86 struct gred_sched
87 {
88         struct gred_sched_data *tab[MAX_DPs];
89         unsigned long   flags;
90         u32             DPs;   
91         u32             def; 
92         u8              initd; 
93 };
94
95 static inline int gred_wred_mode(struct gred_sched *table)
96 {
97         return test_bit(GRED_WRED_MODE, &table->flags);
98 }
99
100 static inline void gred_enable_wred_mode(struct gred_sched *table)
101 {
102         __set_bit(GRED_WRED_MODE, &table->flags);
103 }
104
105 static inline void gred_disable_wred_mode(struct gred_sched *table)
106 {
107         __clear_bit(GRED_WRED_MODE, &table->flags);
108 }
109
110 static inline int gred_rio_mode(struct gred_sched *table)
111 {
112         return test_bit(GRED_RIO_MODE, &table->flags);
113 }
114
115 static inline void gred_enable_rio_mode(struct gred_sched *table)
116 {
117         __set_bit(GRED_RIO_MODE, &table->flags);
118 }
119
120 static inline void gred_disable_rio_mode(struct gred_sched *table)
121 {
122         __clear_bit(GRED_RIO_MODE, &table->flags);
123 }
124
125 static inline int gred_wred_mode_check(struct Qdisc *sch)
126 {
127         struct gred_sched *table = qdisc_priv(sch);
128         int i;
129
130         /* Really ugly O(n^2) but shouldn't be necessary too frequent. */
131         for (i = 0; i < table->DPs; i++) {
132                 struct gred_sched_data *q = table->tab[i];
133                 int n;
134
135                 if (q == NULL)
136                         continue;
137
138                 for (n = 0; n < table->DPs; n++)
139                         if (table->tab[n] && table->tab[n] != q &&
140                             table->tab[n]->prio == q->prio)
141                                 return 1;
142         }
143
144         return 0;
145 }
146
147 static inline unsigned int gred_backlog(struct gred_sched *table,
148                                         struct gred_sched_data *q,
149                                         struct Qdisc *sch)
150 {
151         if (gred_wred_mode(table))
152                 return sch->qstats.backlog;
153         else
154                 return q->backlog;
155 }
156
157 static inline u16 tc_index_to_dp(struct sk_buff *skb)
158 {
159         return skb->tc_index & GRED_VQ_MASK;
160 }
161
162 static int
163 gred_enqueue(struct sk_buff *skb, struct Qdisc* sch)
164 {
165         struct gred_sched_data *q=NULL;
166         struct gred_sched *t= qdisc_priv(sch);
167         unsigned long qavg = 0;
168         int i=0;
169         u16 dp;
170
171         if (!t->initd && skb_queue_len(&sch->q) < (sch->dev->tx_queue_len ? : 1)) {
172                 D2PRINTK("NO GRED Queues setup yet! Enqueued anyway\n");
173                 goto do_enqueue;
174         }
175
176         dp = tc_index_to_dp(skb);
177
178         if (dp >= t->DPs  || (q = t->tab[dp]) == NULL) {
179                 printk("GRED: setting to default (%d)\n ",t->def);
180                 if (!(q=t->tab[t->def])) {
181                         DPRINTK("GRED: setting to default FAILED! dropping!! "
182                             "(%d)\n ", t->def);
183                         goto drop;
184                 }
185                 /* fix tc_index? --could be controvesial but needed for
186                    requeueing */
187                 skb->tc_index=(skb->tc_index & ~GRED_VQ_MASK) | t->def;
188         }
189
190         D2PRINTK("gred_enqueue virtualQ 0x%x classid %x backlog %d "
191             "general backlog %d\n",skb->tc_index&0xf,sch->handle,q->backlog,
192             sch->qstats.backlog);
193         /* sum up all the qaves of prios <= to ours to get the new qave*/
194         if (!gred_wred_mode(t) && gred_rio_mode(t)) {
195                 for (i=0;i<t->DPs;i++) {
196                         if ((!t->tab[i]) || (i==q->DP)) 
197                                 continue; 
198                                 
199                         if (t->tab[i]->prio < q->prio &&
200                             !red_is_idling(&t->tab[i]->parms))
201                                 qavg +=t->tab[i]->parms.qavg;
202                 }
203                         
204         }
205
206         q->packetsin++;
207         q->bytesin+=skb->len;
208
209         if (gred_wred_mode(t)) {
210                 qavg = 0;
211                 q->parms.qavg = t->tab[t->def]->parms.qavg;
212                 q->parms.qidlestart = t->tab[t->def]->parms.qidlestart;
213         }
214
215         q->parms.qavg = red_calc_qavg(&q->parms, gred_backlog(t, q, sch));
216
217         if (red_is_idling(&q->parms))
218                 red_end_of_idle_period(&q->parms);
219
220         if (gred_wred_mode(t))
221                 t->tab[t->def]->parms.qavg = q->parms.qavg;
222
223         switch (red_action(&q->parms, q->parms.qavg + qavg)) {
224                 case RED_DONT_MARK:
225                         break;
226
227                 case RED_PROB_MARK:
228                         sch->qstats.overlimits++;
229                         q->stats.prob_drop++;
230                         goto congestion_drop;
231
232                 case RED_HARD_MARK:
233                         sch->qstats.overlimits++;
234                         q->stats.forced_drop++;
235                         goto congestion_drop;
236         }
237
238         if (q->backlog + skb->len <= q->limit) {
239                 q->backlog += skb->len;
240 do_enqueue:
241                 return qdisc_enqueue_tail(skb, sch);
242         }
243
244         q->stats.pdrop++;
245 drop:
246         return qdisc_drop(skb, sch);
247
248 congestion_drop:
249         qdisc_drop(skb, sch);
250         return NET_XMIT_CN;
251 }
252
253 static int
254 gred_requeue(struct sk_buff *skb, struct Qdisc* sch)
255 {
256         struct gred_sched *t = qdisc_priv(sch);
257         struct gred_sched_data *q = t->tab[tc_index_to_dp(skb)];
258 /* error checking here -- probably unnecessary */
259
260         if (red_is_idling(&q->parms))
261                 red_end_of_idle_period(&q->parms);
262
263         q->backlog += skb->len;
264         return qdisc_requeue(skb, sch);
265 }
266
267 static struct sk_buff *
268 gred_dequeue(struct Qdisc* sch)
269 {
270         struct sk_buff *skb;
271         struct gred_sched_data *q;
272         struct gred_sched *t= qdisc_priv(sch);
273
274         skb = qdisc_dequeue_head(sch);
275
276         if (skb) {
277                 q = t->tab[tc_index_to_dp(skb)];
278                 if (q) {
279                         q->backlog -= skb->len;
280                         if (!q->backlog && !gred_wred_mode(t))
281                                 red_start_of_idle_period(&q->parms);
282                 } else {
283                         D2PRINTK("gred_dequeue: skb has bad tcindex %x\n",
284                                  tc_index_to_dp(skb));
285                 }
286                 return skb;
287         }
288
289         if (gred_wred_mode(t)) {
290                         q= t->tab[t->def];
291                         if (!q) 
292                                 D2PRINTK("no default VQ set: Results will be "
293                                        "screwed up\n");
294                         else
295                                 red_start_of_idle_period(&q->parms);
296         }
297
298         return NULL;
299 }
300
301 static unsigned int gred_drop(struct Qdisc* sch)
302 {
303         struct sk_buff *skb;
304
305         struct gred_sched_data *q;
306         struct gred_sched *t= qdisc_priv(sch);
307
308         skb = qdisc_dequeue_tail(sch);
309         if (skb) {
310                 unsigned int len = skb->len;
311                 q = t->tab[tc_index_to_dp(skb)];
312                 if (q) {
313                         q->backlog -= len;
314                         q->stats.other++;
315                         if (!q->backlog && !gred_wred_mode(t))
316                                 red_start_of_idle_period(&q->parms);
317                 } else {
318                         D2PRINTK("gred_dequeue: skb has bad tcindex %x\n",
319                                  tc_index_to_dp(skb));
320                 }
321
322                 qdisc_drop(skb, sch);
323                 return len;
324         }
325
326         q=t->tab[t->def];
327         if (!q) {
328                 D2PRINTK("no default VQ set: Results might be screwed up\n");
329                 return 0;
330         }
331
332         red_start_of_idle_period(&q->parms);
333         return 0;
334
335 }
336
337 static void gred_reset(struct Qdisc* sch)
338 {
339         int i;
340         struct gred_sched_data *q;
341         struct gred_sched *t= qdisc_priv(sch);
342
343         qdisc_reset_queue(sch);
344
345         for (i=0;i<t->DPs;i++) {
346                 q= t->tab[i];
347                 if (!q) 
348                         continue; 
349                 red_restart(&q->parms);
350                 q->backlog = 0;
351         }
352 }
353
354 static inline void gred_destroy_vq(struct gred_sched_data *q)
355 {
356         kfree(q);
357 }
358
359 static inline int gred_change_table_def(struct Qdisc *sch, struct rtattr *dps)
360 {
361         struct gred_sched *table = qdisc_priv(sch);
362         struct tc_gred_sopt *sopt;
363         int i;
364
365         if (dps == NULL || RTA_PAYLOAD(dps) < sizeof(*sopt))
366                 return -EINVAL;
367
368         sopt = RTA_DATA(dps);
369
370         if (sopt->DPs > MAX_DPs || sopt->DPs == 0 || sopt->def_DP >= sopt->DPs)
371                 return -EINVAL;
372
373         sch_tree_lock(sch);
374         table->DPs = sopt->DPs;
375         table->def = sopt->def_DP;
376
377         /*
378          * Every entry point to GRED is synchronized with the above code
379          * and the DP is checked against DPs, i.e. shadowed VQs can no
380          * longer be found so we can unlock right here.
381          */
382         sch_tree_unlock(sch);
383
384         if (sopt->grio) {
385                 gred_enable_rio_mode(table);
386                 gred_disable_wred_mode(table);
387                 if (gred_wred_mode_check(sch))
388                         gred_enable_wred_mode(table);
389         } else {
390                 gred_disable_rio_mode(table);
391                 gred_disable_wred_mode(table);
392         }
393
394         for (i = table->DPs; i < MAX_DPs; i++) {
395                 if (table->tab[i]) {
396                         printk(KERN_WARNING "GRED: Warning: Destroying "
397                                "shadowed VQ 0x%x\n", i);
398                         gred_destroy_vq(table->tab[i]);
399                         table->tab[i] = NULL;
400                 }
401         }
402
403         table->initd = 0;
404
405         return 0;
406 }
407
408 static inline int gred_change_vq(struct Qdisc *sch, int dp,
409                                  struct tc_gred_qopt *ctl, int prio, u8 *stab)
410 {
411         struct gred_sched *table = qdisc_priv(sch);
412         struct gred_sched_data *q;
413
414         if (table->tab[dp] == NULL) {
415                 table->tab[dp] = kmalloc(sizeof(*q), GFP_KERNEL);
416                 if (table->tab[dp] == NULL)
417                         return -ENOMEM;
418                 memset(table->tab[dp], 0, sizeof(*q));
419         }
420
421         q = table->tab[dp];
422         q->DP = dp;
423         q->prio = prio;
424         q->limit = ctl->limit;
425
426         if (q->backlog == 0)
427                 red_end_of_idle_period(&q->parms);
428
429         red_set_parms(&q->parms,
430                       ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Plog,
431                       ctl->Scell_log, stab);
432
433         return 0;
434 }
435
436 static int gred_change(struct Qdisc *sch, struct rtattr *opt)
437 {
438         struct gred_sched *table = qdisc_priv(sch);
439         struct tc_gred_qopt *ctl;
440         struct rtattr *tb[TCA_GRED_MAX];
441         int err = -EINVAL, prio = GRED_DEF_PRIO;
442         u8 *stab;
443
444         if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_MAX, opt))
445                 return -EINVAL;
446
447         if (tb[TCA_GRED_PARMS-1] == NULL && tb[TCA_GRED_STAB-1] == NULL)
448                 return gred_change_table_def(sch, opt);
449
450         if (tb[TCA_GRED_PARMS-1] == NULL ||
451             RTA_PAYLOAD(tb[TCA_GRED_PARMS-1]) < sizeof(*ctl) ||
452             tb[TCA_GRED_STAB-1] == NULL ||
453             RTA_PAYLOAD(tb[TCA_GRED_STAB-1]) < 256)
454                 return -EINVAL;
455
456         ctl = RTA_DATA(tb[TCA_GRED_PARMS-1]);
457         stab = RTA_DATA(tb[TCA_GRED_STAB-1]);
458
459         if (ctl->DP >= table->DPs)
460                 goto errout;
461
462         if (gred_rio_mode(table)) {
463                 if (ctl->prio == 0) {
464                         int def_prio = GRED_DEF_PRIO;
465
466                         if (table->tab[table->def])
467                                 def_prio = table->tab[table->def]->prio;
468
469                         printk(KERN_DEBUG "GRED: DP %u does not have a prio "
470                                "setting default to %d\n", ctl->DP, def_prio);
471
472                         prio = def_prio;
473                 } else
474                         prio = ctl->prio;
475         }
476
477         sch_tree_lock(sch);
478
479         err = gred_change_vq(sch, ctl->DP, ctl, prio, stab);
480         if (err < 0)
481                 goto errout_locked;
482
483         if (table->tab[table->def] == NULL) {
484                 if (gred_rio_mode(table))
485                         prio = table->tab[ctl->DP]->prio;
486
487                 err = gred_change_vq(sch, table->def, ctl, prio, stab);
488                 if (err < 0)
489                         goto errout_locked;
490         }
491
492         table->initd = 1;
493
494         if (gred_rio_mode(table)) {
495                 gred_disable_wred_mode(table);
496                 if (gred_wred_mode_check(sch))
497                         gred_enable_wred_mode(table);
498         }
499
500         err = 0;
501
502 errout_locked:
503         sch_tree_unlock(sch);
504 errout:
505         return err;
506 }
507
508 static int gred_init(struct Qdisc *sch, struct rtattr *opt)
509 {
510         struct rtattr *tb[TCA_GRED_MAX];
511
512         if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_MAX, opt))
513                 return -EINVAL;
514
515         if (tb[TCA_GRED_PARMS-1] || tb[TCA_GRED_STAB-1])
516                 return -EINVAL;
517
518         return gred_change_table_def(sch, tb[TCA_GRED_DPS-1]);
519 }
520
521 static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
522 {
523         struct gred_sched *table = qdisc_priv(sch);
524         struct rtattr *parms, *opts = NULL;
525         int i;
526         struct tc_gred_sopt sopt = {
527                 .DPs    = table->DPs,
528                 .def_DP = table->def,
529                 .grio   = gred_rio_mode(table),
530         };
531
532         opts = RTA_NEST(skb, TCA_OPTIONS);
533         RTA_PUT(skb, TCA_GRED_DPS, sizeof(sopt), &sopt);
534         parms = RTA_NEST(skb, TCA_GRED_PARMS);
535
536         for (i = 0; i < MAX_DPs; i++) {
537                 struct gred_sched_data *q = table->tab[i];
538                 struct tc_gred_qopt opt;
539
540                 memset(&opt, 0, sizeof(opt));
541
542                 if (!q) {
543                         /* hack -- fix at some point with proper message
544                            This is how we indicate to tc that there is no VQ
545                            at this DP */
546
547                         opt.DP = MAX_DPs + i;
548                         goto append_opt;
549                 }
550
551                 opt.limit       = q->limit;
552                 opt.DP          = q->DP;
553                 opt.backlog     = q->backlog;
554                 opt.prio        = q->prio;
555                 opt.qth_min     = q->parms.qth_min >> q->parms.Wlog;
556                 opt.qth_max     = q->parms.qth_max >> q->parms.Wlog;
557                 opt.Wlog        = q->parms.Wlog;
558                 opt.Plog        = q->parms.Plog;
559                 opt.Scell_log   = q->parms.Scell_log;
560                 opt.other       = q->stats.other;
561                 opt.early       = q->stats.prob_drop;
562                 opt.forced      = q->stats.forced_drop;
563                 opt.pdrop       = q->stats.pdrop;
564                 opt.packets     = q->packetsin;
565                 opt.bytesin     = q->bytesin;
566
567                 if (gred_wred_mode(table)) {
568                         q->parms.qidlestart =
569                                 table->tab[table->def]->parms.qidlestart;
570                         q->parms.qavg = table->tab[table->def]->parms.qavg;
571                 }
572
573                 opt.qave = red_calc_qavg(&q->parms, q->parms.qavg);
574
575 append_opt:
576                 RTA_APPEND(skb, sizeof(opt), &opt);
577         }
578
579         RTA_NEST_END(skb, parms);
580
581         return RTA_NEST_END(skb, opts);
582
583 rtattr_failure:
584         return RTA_NEST_CANCEL(skb, opts);
585 }
586
587 static 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
598 static struct Qdisc_ops gred_qdisc_ops = {
599         .next           =       NULL,
600         .cl_ops         =       NULL,
601         .id             =       "gred",
602         .priv_size      =       sizeof(struct gred_sched),
603         .enqueue        =       gred_enqueue,
604         .dequeue        =       gred_dequeue,
605         .requeue        =       gred_requeue,
606         .drop           =       gred_drop,
607         .init           =       gred_init,
608         .reset          =       gred_reset,
609         .destroy        =       gred_destroy,
610         .change         =       gred_change,
611         .dump           =       gred_dump,
612         .owner          =       THIS_MODULE,
613 };
614
615 static int __init gred_module_init(void)
616 {
617         return register_qdisc(&gred_qdisc_ops);
618 }
619 static void __exit gred_module_exit(void) 
620 {
621         unregister_qdisc(&gred_qdisc_ops);
622 }
623 module_init(gred_module_init)
624 module_exit(gred_module_exit)
625 MODULE_LICENSE("GPL");