[PKT_SCHED]: GRED: Report congestion related drops as NET_XMIT_CN
[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
64 struct gred_sched_data;
65 struct gred_sched;
66
67 struct gred_sched_data
68 {
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 */
75
76         struct red_parms parms;
77         struct red_stats stats;
78 };
79
80 enum {
81         GRED_WRED_MODE = 1,
82         GRED_RIO_MODE,
83 };
84
85 struct gred_sched
86 {
87         struct gred_sched_data *tab[MAX_DPs];
88         unsigned long   flags;
89         u32             DPs;   
90         u32             def; 
91         u8              initd; 
92 };
93
94 static inline int gred_wred_mode(struct gred_sched *table)
95 {
96         return test_bit(GRED_WRED_MODE, &table->flags);
97 }
98
99 static inline void gred_enable_wred_mode(struct gred_sched *table)
100 {
101         __set_bit(GRED_WRED_MODE, &table->flags);
102 }
103
104 static inline void gred_disable_wred_mode(struct gred_sched *table)
105 {
106         __clear_bit(GRED_WRED_MODE, &table->flags);
107 }
108
109 static inline int gred_rio_mode(struct gred_sched *table)
110 {
111         return test_bit(GRED_RIO_MODE, &table->flags);
112 }
113
114 static inline void gred_enable_rio_mode(struct gred_sched *table)
115 {
116         __set_bit(GRED_RIO_MODE, &table->flags);
117 }
118
119 static inline void gred_disable_rio_mode(struct gred_sched *table)
120 {
121         __clear_bit(GRED_RIO_MODE, &table->flags);
122 }
123
124 static inline int gred_wred_mode_check(struct Qdisc *sch)
125 {
126         struct gred_sched *table = qdisc_priv(sch);
127         int i;
128
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];
132                 int n;
133
134                 if (q == NULL)
135                         continue;
136
137                 for (n = 0; n < table->DPs; n++)
138                         if (table->tab[n] && table->tab[n] != q &&
139                             table->tab[n]->prio == q->prio)
140                                 return 1;
141         }
142
143         return 0;
144 }
145
146 static inline unsigned int gred_backlog(struct gred_sched *table,
147                                         struct gred_sched_data *q,
148                                         struct Qdisc *sch)
149 {
150         if (gred_wred_mode(table))
151                 return sch->qstats.backlog;
152         else
153                 return q->backlog;
154 }
155
156 static int
157 gred_enqueue(struct sk_buff *skb, struct Qdisc* sch)
158 {
159         struct gred_sched_data *q=NULL;
160         struct gred_sched *t= qdisc_priv(sch);
161         unsigned long qavg = 0;
162         int i=0;
163
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");
166                 goto do_enqueue;
167         }
168
169
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!! "
174                             "(%d)\n ", t->def);
175                         goto drop;
176                 }
177                 /* fix tc_index? --could be controvesial but needed for
178                    requeueing */
179                 skb->tc_index=(skb->tc_index&0xfffffff0) | t->def;
180         }
181
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)) 
189                                 continue; 
190                                 
191                         if (t->tab[i]->prio < q->prio &&
192                             !red_is_idling(&t->tab[i]->parms))
193                                 qavg +=t->tab[i]->parms.qavg;
194                 }
195                         
196         }
197
198         q->packetsin++;
199         q->bytesin+=skb->len;
200
201         if (gred_wred_mode(t)) {
202                 qavg = 0;
203                 q->parms.qavg = t->tab[t->def]->parms.qavg;
204                 q->parms.qidlestart = t->tab[t->def]->parms.qidlestart;
205         }
206
207         q->parms.qavg = red_calc_qavg(&q->parms, gred_backlog(t, q, sch));
208
209         if (red_is_idling(&q->parms))
210                 red_end_of_idle_period(&q->parms);
211
212         if (gred_wred_mode(t))
213                 t->tab[t->def]->parms.qavg = q->parms.qavg;
214
215         switch (red_action(&q->parms, q->parms.qavg + qavg)) {
216                 case RED_DONT_MARK:
217                         break;
218
219                 case RED_PROB_MARK:
220                         sch->qstats.overlimits++;
221                         q->stats.prob_drop++;
222                         goto congestion_drop;
223
224                 case RED_HARD_MARK:
225                         sch->qstats.overlimits++;
226                         q->stats.forced_drop++;
227                         goto congestion_drop;
228         }
229
230         if (q->backlog + skb->len <= q->limit) {
231                 q->backlog += skb->len;
232 do_enqueue:
233                 __skb_queue_tail(&sch->q, skb);
234                 sch->qstats.backlog += skb->len;
235                 sch->bstats.bytes += skb->len;
236                 sch->bstats.packets++;
237                 return 0;
238         }
239
240         q->stats.pdrop++;
241 drop:
242         kfree_skb(skb);
243         sch->qstats.drops++;
244         return NET_XMIT_DROP;
245
246 congestion_drop:
247         kfree_skb(skb);
248         sch->qstats.drops++;
249         return NET_XMIT_CN;
250 }
251
252 static int
253 gred_requeue(struct sk_buff *skb, struct Qdisc* sch)
254 {
255         struct gred_sched_data *q;
256         struct gred_sched *t= qdisc_priv(sch);
257         q= t->tab[(skb->tc_index&0xf)];
258 /* error checking here -- probably unnecessary */
259
260         if (red_is_idling(&q->parms))
261                 red_end_of_idle_period(&q->parms);
262
263         __skb_queue_head(&sch->q, skb);
264         sch->qstats.backlog += skb->len;
265         sch->qstats.requeues++;
266         q->backlog += skb->len;
267         return 0;
268 }
269
270 static struct sk_buff *
271 gred_dequeue(struct Qdisc* sch)
272 {
273         struct sk_buff *skb;
274         struct gred_sched_data *q;
275         struct gred_sched *t= qdisc_priv(sch);
276
277         skb = __skb_dequeue(&sch->q);
278         if (skb) {
279                 sch->qstats.backlog -= skb->len;
280                 q= t->tab[(skb->tc_index&0xf)];
281                 if (q) {
282                         q->backlog -= skb->len;
283                         if (!q->backlog && !gred_wred_mode(t))
284                                 red_start_of_idle_period(&q->parms);
285                 } else {
286                         D2PRINTK("gred_dequeue: skb has bad tcindex %x\n",skb->tc_index&0xf); 
287                 }
288                 return skb;
289         }
290
291         if (gred_wred_mode(t)) {
292                         q= t->tab[t->def];
293                         if (!q) 
294                                 D2PRINTK("no default VQ set: Results will be "
295                                        "screwed up\n");
296                         else
297                                 red_start_of_idle_period(&q->parms);
298         }
299
300         return NULL;
301 }
302
303 static unsigned int gred_drop(struct Qdisc* sch)
304 {
305         struct sk_buff *skb;
306
307         struct gred_sched_data *q;
308         struct gred_sched *t= qdisc_priv(sch);
309
310         skb = __skb_dequeue_tail(&sch->q);
311         if (skb) {
312                 unsigned int len = skb->len;
313                 sch->qstats.backlog -= len;
314                 sch->qstats.drops++;
315                 q= t->tab[(skb->tc_index&0xf)];
316                 if (q) {
317                         q->backlog -= len;
318                         q->stats.other++;
319                         if (!q->backlog && !gred_wred_mode(t))
320                                 red_start_of_idle_period(&q->parms);
321                 } else {
322                         D2PRINTK("gred_dequeue: skb has bad tcindex %x\n",skb->tc_index&0xf); 
323                 }
324
325                 kfree_skb(skb);
326                 return len;
327         }
328
329         q=t->tab[t->def];
330         if (!q) {
331                 D2PRINTK("no default VQ set: Results might be screwed up\n");
332                 return 0;
333         }
334
335         red_start_of_idle_period(&q->parms);
336         return 0;
337
338 }
339
340 static void gred_reset(struct Qdisc* sch)
341 {
342         int i;
343         struct gred_sched_data *q;
344         struct gred_sched *t= qdisc_priv(sch);
345
346         __skb_queue_purge(&sch->q);
347
348         sch->qstats.backlog = 0;
349
350         for (i=0;i<t->DPs;i++) {
351                 q= t->tab[i];
352                 if (!q) 
353                         continue; 
354                 red_restart(&q->parms);
355                 q->backlog = 0;
356         }
357 }
358
359 static inline void gred_destroy_vq(struct gred_sched_data *q)
360 {
361         kfree(q);
362 }
363
364 static inline int gred_change_table_def(struct Qdisc *sch, struct rtattr *dps)
365 {
366         struct gred_sched *table = qdisc_priv(sch);
367         struct tc_gred_sopt *sopt;
368         int i;
369
370         if (dps == NULL || RTA_PAYLOAD(dps) < sizeof(*sopt))
371                 return -EINVAL;
372
373         sopt = RTA_DATA(dps);
374
375         if (sopt->DPs > MAX_DPs || sopt->DPs == 0 || sopt->def_DP >= sopt->DPs)
376                 return -EINVAL;
377
378         sch_tree_lock(sch);
379         table->DPs = sopt->DPs;
380         table->def = sopt->def_DP;
381
382         /*
383          * Every entry point to GRED is synchronized with the above code
384          * and the DP is checked against DPs, i.e. shadowed VQs can no
385          * longer be found so we can unlock right here.
386          */
387         sch_tree_unlock(sch);
388
389         if (sopt->grio) {
390                 gred_enable_rio_mode(table);
391                 gred_disable_wred_mode(table);
392                 if (gred_wred_mode_check(sch))
393                         gred_enable_wred_mode(table);
394         } else {
395                 gred_disable_rio_mode(table);
396                 gred_disable_wred_mode(table);
397         }
398
399         for (i = table->DPs; i < MAX_DPs; i++) {
400                 if (table->tab[i]) {
401                         printk(KERN_WARNING "GRED: Warning: Destroying "
402                                "shadowed VQ 0x%x\n", i);
403                         gred_destroy_vq(table->tab[i]);
404                         table->tab[i] = NULL;
405                 }
406         }
407
408         table->initd = 0;
409
410         return 0;
411 }
412
413 static inline int gred_change_vq(struct Qdisc *sch, int dp,
414                                  struct tc_gred_qopt *ctl, int prio, u8 *stab)
415 {
416         struct gred_sched *table = qdisc_priv(sch);
417         struct gred_sched_data *q;
418
419         if (table->tab[dp] == NULL) {
420                 table->tab[dp] = kmalloc(sizeof(*q), GFP_KERNEL);
421                 if (table->tab[dp] == NULL)
422                         return -ENOMEM;
423                 memset(table->tab[dp], 0, sizeof(*q));
424         }
425
426         q = table->tab[dp];
427         q->DP = dp;
428         q->prio = prio;
429         q->limit = ctl->limit;
430
431         if (q->backlog == 0)
432                 red_end_of_idle_period(&q->parms);
433
434         red_set_parms(&q->parms,
435                       ctl->qth_min, ctl->qth_max, ctl->Wlog, ctl->Plog,
436                       ctl->Scell_log, stab);
437
438         return 0;
439 }
440
441 static int gred_change(struct Qdisc *sch, struct rtattr *opt)
442 {
443         struct gred_sched *table = qdisc_priv(sch);
444         struct tc_gred_qopt *ctl;
445         struct rtattr *tb[TCA_GRED_MAX];
446         int err = -EINVAL, prio = GRED_DEF_PRIO;
447         u8 *stab;
448
449         if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_MAX, opt))
450                 return -EINVAL;
451
452         if (tb[TCA_GRED_PARMS-1] == NULL && tb[TCA_GRED_STAB-1] == NULL)
453                 return gred_change_table_def(sch, opt);
454
455         if (tb[TCA_GRED_PARMS-1] == NULL ||
456             RTA_PAYLOAD(tb[TCA_GRED_PARMS-1]) < sizeof(*ctl) ||
457             tb[TCA_GRED_STAB-1] == NULL ||
458             RTA_PAYLOAD(tb[TCA_GRED_STAB-1]) < 256)
459                 return -EINVAL;
460
461         ctl = RTA_DATA(tb[TCA_GRED_PARMS-1]);
462         stab = RTA_DATA(tb[TCA_GRED_STAB-1]);
463
464         if (ctl->DP >= table->DPs)
465                 goto errout;
466
467         if (gred_rio_mode(table)) {
468                 if (ctl->prio == 0) {
469                         int def_prio = GRED_DEF_PRIO;
470
471                         if (table->tab[table->def])
472                                 def_prio = table->tab[table->def]->prio;
473
474                         printk(KERN_DEBUG "GRED: DP %u does not have a prio "
475                                "setting default to %d\n", ctl->DP, def_prio);
476
477                         prio = def_prio;
478                 } else
479                         prio = ctl->prio;
480         }
481
482         sch_tree_lock(sch);
483
484         err = gred_change_vq(sch, ctl->DP, ctl, prio, stab);
485         if (err < 0)
486                 goto errout_locked;
487
488         if (table->tab[table->def] == NULL) {
489                 if (gred_rio_mode(table))
490                         prio = table->tab[ctl->DP]->prio;
491
492                 err = gred_change_vq(sch, table->def, ctl, prio, stab);
493                 if (err < 0)
494                         goto errout_locked;
495         }
496
497         table->initd = 1;
498
499         if (gred_rio_mode(table)) {
500                 gred_disable_wred_mode(table);
501                 if (gred_wred_mode_check(sch))
502                         gred_enable_wred_mode(table);
503         }
504
505         err = 0;
506
507 errout_locked:
508         sch_tree_unlock(sch);
509 errout:
510         return err;
511 }
512
513 static int gred_init(struct Qdisc *sch, struct rtattr *opt)
514 {
515         struct rtattr *tb[TCA_GRED_MAX];
516
517         if (opt == NULL || rtattr_parse_nested(tb, TCA_GRED_MAX, opt))
518                 return -EINVAL;
519
520         if (tb[TCA_GRED_PARMS-1] || tb[TCA_GRED_STAB-1])
521                 return -EINVAL;
522
523         return gred_change_table_def(sch, tb[TCA_GRED_DPS-1]);
524 }
525
526 static int gred_dump(struct Qdisc *sch, struct sk_buff *skb)
527 {
528         struct gred_sched *table = qdisc_priv(sch);
529         struct rtattr *parms, *opts = NULL;
530         int i;
531         struct tc_gred_sopt sopt = {
532                 .DPs    = table->DPs,
533                 .def_DP = table->def,
534                 .grio   = gred_rio_mode(table),
535         };
536
537         opts = RTA_NEST(skb, TCA_OPTIONS);
538         RTA_PUT(skb, TCA_GRED_DPS, sizeof(sopt), &sopt);
539         parms = RTA_NEST(skb, TCA_GRED_PARMS);
540
541         for (i = 0; i < MAX_DPs; i++) {
542                 struct gred_sched_data *q = table->tab[i];
543                 struct tc_gred_qopt opt;
544
545                 memset(&opt, 0, sizeof(opt));
546
547                 if (!q) {
548                         /* hack -- fix at some point with proper message
549                            This is how we indicate to tc that there is no VQ
550                            at this DP */
551
552                         opt.DP = MAX_DPs + i;
553                         goto append_opt;
554                 }
555
556                 opt.limit       = q->limit;
557                 opt.DP          = q->DP;
558                 opt.backlog     = q->backlog;
559                 opt.prio        = q->prio;
560                 opt.qth_min     = q->parms.qth_min >> q->parms.Wlog;
561                 opt.qth_max     = q->parms.qth_max >> q->parms.Wlog;
562                 opt.Wlog        = q->parms.Wlog;
563                 opt.Plog        = q->parms.Plog;
564                 opt.Scell_log   = q->parms.Scell_log;
565                 opt.other       = q->stats.other;
566                 opt.early       = q->stats.prob_drop;
567                 opt.forced      = q->stats.forced_drop;
568                 opt.pdrop       = q->stats.pdrop;
569                 opt.packets     = q->packetsin;
570                 opt.bytesin     = q->bytesin;
571
572                 if (gred_wred_mode(table)) {
573                         q->parms.qidlestart =
574                                 table->tab[table->def]->parms.qidlestart;
575                         q->parms.qavg = table->tab[table->def]->parms.qavg;
576                 }
577
578                 opt.qave = red_calc_qavg(&q->parms, q->parms.qavg);
579
580 append_opt:
581                 RTA_APPEND(skb, sizeof(opt), &opt);
582         }
583
584         RTA_NEST_END(skb, parms);
585
586         return RTA_NEST_END(skb, opts);
587
588 rtattr_failure:
589         return RTA_NEST_CANCEL(skb, opts);
590 }
591
592 static 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
603 static struct Qdisc_ops gred_qdisc_ops = {
604         .next           =       NULL,
605         .cl_ops         =       NULL,
606         .id             =       "gred",
607         .priv_size      =       sizeof(struct gred_sched),
608         .enqueue        =       gred_enqueue,
609         .dequeue        =       gred_dequeue,
610         .requeue        =       gred_requeue,
611         .drop           =       gred_drop,
612         .init           =       gred_init,
613         .reset          =       gred_reset,
614         .destroy        =       gred_destroy,
615         .change         =       gred_change,
616         .dump           =       gred_dump,
617         .owner          =       THIS_MODULE,
618 };
619
620 static int __init gred_module_init(void)
621 {
622         return register_qdisc(&gred_qdisc_ops);
623 }
624 static void __exit gred_module_exit(void) 
625 {
626         unregister_qdisc(&gred_qdisc_ops);
627 }
628 module_init(gred_module_init)
629 module_exit(gred_module_exit)
630 MODULE_LICENSE("GPL");