net_sched: factorize qdisc stats handling
[linux-2.6.git] / net / sched / sch_dsmark.c
1 /* net/sched/sch_dsmark.c - Differentiated Services field marker */
2
3 /* Written 1998-2000 by Werner Almesberger, EPFL ICA */
4
5
6 #include <linux/module.h>
7 #include <linux/init.h>
8 #include <linux/slab.h>
9 #include <linux/types.h>
10 #include <linux/string.h>
11 #include <linux/errno.h>
12 #include <linux/skbuff.h>
13 #include <linux/rtnetlink.h>
14 #include <linux/bitops.h>
15 #include <net/pkt_sched.h>
16 #include <net/dsfield.h>
17 #include <net/inet_ecn.h>
18 #include <asm/byteorder.h>
19
20 /*
21  * classid      class           marking
22  * -------      -----           -------
23  *   n/a          0             n/a
24  *   x:0          1             use entry [0]
25  *   ...         ...            ...
26  *   x:y y>0     y+1            use entry [y]
27  *   ...         ...            ...
28  * x:indices-1  indices         use entry [indices-1]
29  *   ...         ...            ...
30  *   x:y         y+1            use entry [y & (indices-1)]
31  *   ...         ...            ...
32  * 0xffff       0x10000         use entry [indices-1]
33  */
34
35
36 #define NO_DEFAULT_INDEX        (1 << 16)
37
38 struct dsmark_qdisc_data {
39         struct Qdisc            *q;
40         struct tcf_proto        *filter_list;
41         u8                      *mask;  /* "owns" the array */
42         u8                      *value;
43         u16                     indices;
44         u32                     default_index;  /* index range is 0...0xffff */
45         int                     set_tc_index;
46 };
47
48 static inline int dsmark_valid_index(struct dsmark_qdisc_data *p, u16 index)
49 {
50         return (index <= p->indices && index > 0);
51 }
52
53 /* ------------------------- Class/flow operations ------------------------- */
54
55 static int dsmark_graft(struct Qdisc *sch, unsigned long arg,
56                         struct Qdisc *new, struct Qdisc **old)
57 {
58         struct dsmark_qdisc_data *p = qdisc_priv(sch);
59
60         pr_debug("dsmark_graft(sch %p,[qdisc %p],new %p,old %p)\n",
61                 sch, p, new, old);
62
63         if (new == NULL) {
64                 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
65                                         sch->handle);
66                 if (new == NULL)
67                         new = &noop_qdisc;
68         }
69
70         sch_tree_lock(sch);
71         *old = p->q;
72         p->q = new;
73         qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
74         qdisc_reset(*old);
75         sch_tree_unlock(sch);
76
77         return 0;
78 }
79
80 static struct Qdisc *dsmark_leaf(struct Qdisc *sch, unsigned long arg)
81 {
82         struct dsmark_qdisc_data *p = qdisc_priv(sch);
83         return p->q;
84 }
85
86 static unsigned long dsmark_get(struct Qdisc *sch, u32 classid)
87 {
88         pr_debug("dsmark_get(sch %p,[qdisc %p],classid %x)\n",
89                 sch, qdisc_priv(sch), classid);
90
91         return TC_H_MIN(classid) + 1;
92 }
93
94 static unsigned long dsmark_bind_filter(struct Qdisc *sch,
95                                         unsigned long parent, u32 classid)
96 {
97         return dsmark_get(sch, classid);
98 }
99
100 static void dsmark_put(struct Qdisc *sch, unsigned long cl)
101 {
102 }
103
104 static const struct nla_policy dsmark_policy[TCA_DSMARK_MAX + 1] = {
105         [TCA_DSMARK_INDICES]            = { .type = NLA_U16 },
106         [TCA_DSMARK_DEFAULT_INDEX]      = { .type = NLA_U16 },
107         [TCA_DSMARK_SET_TC_INDEX]       = { .type = NLA_FLAG },
108         [TCA_DSMARK_MASK]               = { .type = NLA_U8 },
109         [TCA_DSMARK_VALUE]              = { .type = NLA_U8 },
110 };
111
112 static int dsmark_change(struct Qdisc *sch, u32 classid, u32 parent,
113                          struct nlattr **tca, unsigned long *arg)
114 {
115         struct dsmark_qdisc_data *p = qdisc_priv(sch);
116         struct nlattr *opt = tca[TCA_OPTIONS];
117         struct nlattr *tb[TCA_DSMARK_MAX + 1];
118         int err = -EINVAL;
119         u8 mask = 0;
120
121         pr_debug("dsmark_change(sch %p,[qdisc %p],classid %x,parent %x),"
122                 "arg 0x%lx\n", sch, p, classid, parent, *arg);
123
124         if (!dsmark_valid_index(p, *arg)) {
125                 err = -ENOENT;
126                 goto errout;
127         }
128
129         if (!opt)
130                 goto errout;
131
132         err = nla_parse_nested(tb, TCA_DSMARK_MAX, opt, dsmark_policy);
133         if (err < 0)
134                 goto errout;
135
136         if (tb[TCA_DSMARK_MASK])
137                 mask = nla_get_u8(tb[TCA_DSMARK_MASK]);
138
139         if (tb[TCA_DSMARK_VALUE])
140                 p->value[*arg-1] = nla_get_u8(tb[TCA_DSMARK_VALUE]);
141
142         if (tb[TCA_DSMARK_MASK])
143                 p->mask[*arg-1] = mask;
144
145         err = 0;
146
147 errout:
148         return err;
149 }
150
151 static int dsmark_delete(struct Qdisc *sch, unsigned long arg)
152 {
153         struct dsmark_qdisc_data *p = qdisc_priv(sch);
154
155         if (!dsmark_valid_index(p, arg))
156                 return -EINVAL;
157
158         p->mask[arg-1] = 0xff;
159         p->value[arg-1] = 0;
160
161         return 0;
162 }
163
164 static void dsmark_walk(struct Qdisc *sch, struct qdisc_walker *walker)
165 {
166         struct dsmark_qdisc_data *p = qdisc_priv(sch);
167         int i;
168
169         pr_debug("dsmark_walk(sch %p,[qdisc %p],walker %p)\n", sch, p, walker);
170
171         if (walker->stop)
172                 return;
173
174         for (i = 0; i < p->indices; i++) {
175                 if (p->mask[i] == 0xff && !p->value[i])
176                         goto ignore;
177                 if (walker->count >= walker->skip) {
178                         if (walker->fn(sch, i+1, walker) < 0) {
179                                 walker->stop = 1;
180                                 break;
181                         }
182                 }
183 ignore:
184                 walker->count++;
185         }
186 }
187
188 static inline struct tcf_proto **dsmark_find_tcf(struct Qdisc *sch,
189                                                  unsigned long cl)
190 {
191         struct dsmark_qdisc_data *p = qdisc_priv(sch);
192         return &p->filter_list;
193 }
194
195 /* --------------------------- Qdisc operations ---------------------------- */
196
197 static int dsmark_enqueue(struct sk_buff *skb, struct Qdisc *sch)
198 {
199         struct dsmark_qdisc_data *p = qdisc_priv(sch);
200         int err;
201
202         pr_debug("dsmark_enqueue(skb %p,sch %p,[qdisc %p])\n", skb, sch, p);
203
204         if (p->set_tc_index) {
205                 switch (skb->protocol) {
206                 case htons(ETH_P_IP):
207                         if (skb_cow_head(skb, sizeof(struct iphdr)))
208                                 goto drop;
209
210                         skb->tc_index = ipv4_get_dsfield(ip_hdr(skb))
211                                 & ~INET_ECN_MASK;
212                         break;
213
214                 case htons(ETH_P_IPV6):
215                         if (skb_cow_head(skb, sizeof(struct ipv6hdr)))
216                                 goto drop;
217
218                         skb->tc_index = ipv6_get_dsfield(ipv6_hdr(skb))
219                                 & ~INET_ECN_MASK;
220                         break;
221                 default:
222                         skb->tc_index = 0;
223                         break;
224                 }
225         }
226
227         if (TC_H_MAJ(skb->priority) == sch->handle)
228                 skb->tc_index = TC_H_MIN(skb->priority);
229         else {
230                 struct tcf_result res;
231                 int result = tc_classify(skb, p->filter_list, &res);
232
233                 pr_debug("result %d class 0x%04x\n", result, res.classid);
234
235                 switch (result) {
236 #ifdef CONFIG_NET_CLS_ACT
237                 case TC_ACT_QUEUED:
238                 case TC_ACT_STOLEN:
239                         kfree_skb(skb);
240                         return NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
241
242                 case TC_ACT_SHOT:
243                         goto drop;
244 #endif
245                 case TC_ACT_OK:
246                         skb->tc_index = TC_H_MIN(res.classid);
247                         break;
248
249                 default:
250                         if (p->default_index != NO_DEFAULT_INDEX)
251                                 skb->tc_index = p->default_index;
252                         break;
253                 }
254         }
255
256         err = qdisc_enqueue(skb, p->q);
257         if (err != NET_XMIT_SUCCESS) {
258                 if (net_xmit_drop_count(err))
259                         sch->qstats.drops++;
260                 return err;
261         }
262
263         qdisc_bstats_update(sch, skb);
264         sch->q.qlen++;
265
266         return NET_XMIT_SUCCESS;
267
268 drop:
269         kfree_skb(skb);
270         sch->qstats.drops++;
271         return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
272 }
273
274 static struct sk_buff *dsmark_dequeue(struct Qdisc *sch)
275 {
276         struct dsmark_qdisc_data *p = qdisc_priv(sch);
277         struct sk_buff *skb;
278         u32 index;
279
280         pr_debug("dsmark_dequeue(sch %p,[qdisc %p])\n", sch, p);
281
282         skb = p->q->ops->dequeue(p->q);
283         if (skb == NULL)
284                 return NULL;
285
286         sch->q.qlen--;
287
288         index = skb->tc_index & (p->indices - 1);
289         pr_debug("index %d->%d\n", skb->tc_index, index);
290
291         switch (skb->protocol) {
292         case htons(ETH_P_IP):
293                 ipv4_change_dsfield(ip_hdr(skb), p->mask[index],
294                                     p->value[index]);
295                         break;
296         case htons(ETH_P_IPV6):
297                 ipv6_change_dsfield(ipv6_hdr(skb), p->mask[index],
298                                     p->value[index]);
299                         break;
300         default:
301                 /*
302                  * Only complain if a change was actually attempted.
303                  * This way, we can send non-IP traffic through dsmark
304                  * and don't need yet another qdisc as a bypass.
305                  */
306                 if (p->mask[index] != 0xff || p->value[index])
307                         printk(KERN_WARNING
308                                "dsmark_dequeue: unsupported protocol %d\n",
309                                ntohs(skb->protocol));
310                 break;
311         }
312
313         return skb;
314 }
315
316 static struct sk_buff *dsmark_peek(struct Qdisc *sch)
317 {
318         struct dsmark_qdisc_data *p = qdisc_priv(sch);
319
320         pr_debug("dsmark_peek(sch %p,[qdisc %p])\n", sch, p);
321
322         return p->q->ops->peek(p->q);
323 }
324
325 static unsigned int dsmark_drop(struct Qdisc *sch)
326 {
327         struct dsmark_qdisc_data *p = qdisc_priv(sch);
328         unsigned int len;
329
330         pr_debug("dsmark_reset(sch %p,[qdisc %p])\n", sch, p);
331
332         if (p->q->ops->drop == NULL)
333                 return 0;
334
335         len = p->q->ops->drop(p->q);
336         if (len)
337                 sch->q.qlen--;
338
339         return len;
340 }
341
342 static int dsmark_init(struct Qdisc *sch, struct nlattr *opt)
343 {
344         struct dsmark_qdisc_data *p = qdisc_priv(sch);
345         struct nlattr *tb[TCA_DSMARK_MAX + 1];
346         int err = -EINVAL;
347         u32 default_index = NO_DEFAULT_INDEX;
348         u16 indices;
349         u8 *mask;
350
351         pr_debug("dsmark_init(sch %p,[qdisc %p],opt %p)\n", sch, p, opt);
352
353         if (!opt)
354                 goto errout;
355
356         err = nla_parse_nested(tb, TCA_DSMARK_MAX, opt, dsmark_policy);
357         if (err < 0)
358                 goto errout;
359
360         err = -EINVAL;
361         indices = nla_get_u16(tb[TCA_DSMARK_INDICES]);
362
363         if (hweight32(indices) != 1)
364                 goto errout;
365
366         if (tb[TCA_DSMARK_DEFAULT_INDEX])
367                 default_index = nla_get_u16(tb[TCA_DSMARK_DEFAULT_INDEX]);
368
369         mask = kmalloc(indices * 2, GFP_KERNEL);
370         if (mask == NULL) {
371                 err = -ENOMEM;
372                 goto errout;
373         }
374
375         p->mask = mask;
376         memset(p->mask, 0xff, indices);
377
378         p->value = p->mask + indices;
379         memset(p->value, 0, indices);
380
381         p->indices = indices;
382         p->default_index = default_index;
383         p->set_tc_index = nla_get_flag(tb[TCA_DSMARK_SET_TC_INDEX]);
384
385         p->q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, sch->handle);
386         if (p->q == NULL)
387                 p->q = &noop_qdisc;
388
389         pr_debug("dsmark_init: qdisc %p\n", p->q);
390
391         err = 0;
392 errout:
393         return err;
394 }
395
396 static void dsmark_reset(struct Qdisc *sch)
397 {
398         struct dsmark_qdisc_data *p = qdisc_priv(sch);
399
400         pr_debug("dsmark_reset(sch %p,[qdisc %p])\n", sch, p);
401         qdisc_reset(p->q);
402         sch->q.qlen = 0;
403 }
404
405 static void dsmark_destroy(struct Qdisc *sch)
406 {
407         struct dsmark_qdisc_data *p = qdisc_priv(sch);
408
409         pr_debug("dsmark_destroy(sch %p,[qdisc %p])\n", sch, p);
410
411         tcf_destroy_chain(&p->filter_list);
412         qdisc_destroy(p->q);
413         kfree(p->mask);
414 }
415
416 static int dsmark_dump_class(struct Qdisc *sch, unsigned long cl,
417                              struct sk_buff *skb, struct tcmsg *tcm)
418 {
419         struct dsmark_qdisc_data *p = qdisc_priv(sch);
420         struct nlattr *opts = NULL;
421
422         pr_debug("dsmark_dump_class(sch %p,[qdisc %p],class %ld\n", sch, p, cl);
423
424         if (!dsmark_valid_index(p, cl))
425                 return -EINVAL;
426
427         tcm->tcm_handle = TC_H_MAKE(TC_H_MAJ(sch->handle), cl-1);
428         tcm->tcm_info = p->q->handle;
429
430         opts = nla_nest_start(skb, TCA_OPTIONS);
431         if (opts == NULL)
432                 goto nla_put_failure;
433         NLA_PUT_U8(skb, TCA_DSMARK_MASK, p->mask[cl-1]);
434         NLA_PUT_U8(skb, TCA_DSMARK_VALUE, p->value[cl-1]);
435
436         return nla_nest_end(skb, opts);
437
438 nla_put_failure:
439         nla_nest_cancel(skb, opts);
440         return -EMSGSIZE;
441 }
442
443 static int dsmark_dump(struct Qdisc *sch, struct sk_buff *skb)
444 {
445         struct dsmark_qdisc_data *p = qdisc_priv(sch);
446         struct nlattr *opts = NULL;
447
448         opts = nla_nest_start(skb, TCA_OPTIONS);
449         if (opts == NULL)
450                 goto nla_put_failure;
451         NLA_PUT_U16(skb, TCA_DSMARK_INDICES, p->indices);
452
453         if (p->default_index != NO_DEFAULT_INDEX)
454                 NLA_PUT_U16(skb, TCA_DSMARK_DEFAULT_INDEX, p->default_index);
455
456         if (p->set_tc_index)
457                 NLA_PUT_FLAG(skb, TCA_DSMARK_SET_TC_INDEX);
458
459         return nla_nest_end(skb, opts);
460
461 nla_put_failure:
462         nla_nest_cancel(skb, opts);
463         return -EMSGSIZE;
464 }
465
466 static const struct Qdisc_class_ops dsmark_class_ops = {
467         .graft          =       dsmark_graft,
468         .leaf           =       dsmark_leaf,
469         .get            =       dsmark_get,
470         .put            =       dsmark_put,
471         .change         =       dsmark_change,
472         .delete         =       dsmark_delete,
473         .walk           =       dsmark_walk,
474         .tcf_chain      =       dsmark_find_tcf,
475         .bind_tcf       =       dsmark_bind_filter,
476         .unbind_tcf     =       dsmark_put,
477         .dump           =       dsmark_dump_class,
478 };
479
480 static struct Qdisc_ops dsmark_qdisc_ops __read_mostly = {
481         .next           =       NULL,
482         .cl_ops         =       &dsmark_class_ops,
483         .id             =       "dsmark",
484         .priv_size      =       sizeof(struct dsmark_qdisc_data),
485         .enqueue        =       dsmark_enqueue,
486         .dequeue        =       dsmark_dequeue,
487         .peek           =       dsmark_peek,
488         .drop           =       dsmark_drop,
489         .init           =       dsmark_init,
490         .reset          =       dsmark_reset,
491         .destroy        =       dsmark_destroy,
492         .change         =       NULL,
493         .dump           =       dsmark_dump,
494         .owner          =       THIS_MODULE,
495 };
496
497 static int __init dsmark_module_init(void)
498 {
499         return register_qdisc(&dsmark_qdisc_ops);
500 }
501
502 static void __exit dsmark_module_exit(void)
503 {
504         unregister_qdisc(&dsmark_qdisc_ops);
505 }
506
507 module_init(dsmark_module_init)
508 module_exit(dsmark_module_exit)
509
510 MODULE_LICENSE("GPL");