pkt_sched: Remove qdisc->ops->requeue() etc.
[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/types.h>
9 #include <linux/string.h>
10 #include <linux/errno.h>
11 #include <linux/skbuff.h>
12 #include <linux/rtnetlink.h>
13 #include <linux/bitops.h>
14 #include <net/pkt_sched.h>
15 #include <net/dsfield.h>
16 #include <net/inet_ecn.h>
17 #include <asm/byteorder.h>
18
19 /*
20  * classid      class           marking
21  * -------      -----           -------
22  *   n/a          0             n/a
23  *   x:0          1             use entry [0]
24  *   ...         ...            ...
25  *   x:y y>0     y+1            use entry [y]
26  *   ...         ...            ...
27  * x:indices-1  indices         use entry [indices-1]
28  *   ...         ...            ...
29  *   x:y         y+1            use entry [y & (indices-1)]
30  *   ...         ...            ...
31  * 0xffff       0x10000         use entry [indices-1]
32  */
33
34
35 #define NO_DEFAULT_INDEX        (1 << 16)
36
37 struct dsmark_qdisc_data {
38         struct Qdisc            *q;
39         struct tcf_proto        *filter_list;
40         u8                      *mask;  /* "owns" the array */
41         u8                      *value;
42         u16                     indices;
43         u32                     default_index;  /* index range is 0...0xffff */
44         int                     set_tc_index;
45 };
46
47 static inline int dsmark_valid_index(struct dsmark_qdisc_data *p, u16 index)
48 {
49         return (index <= p->indices && index > 0);
50 }
51
52 /* ------------------------- Class/flow operations ------------------------- */
53
54 static int dsmark_graft(struct Qdisc *sch, unsigned long arg,
55                         struct Qdisc *new, struct Qdisc **old)
56 {
57         struct dsmark_qdisc_data *p = qdisc_priv(sch);
58
59         pr_debug("dsmark_graft(sch %p,[qdisc %p],new %p,old %p)\n",
60                 sch, p, new, old);
61
62         if (new == NULL) {
63                 new = qdisc_create_dflt(qdisc_dev(sch), sch->dev_queue,
64                                         &pfifo_qdisc_ops,
65                                         sch->handle);
66                 if (new == NULL)
67                         new = &noop_qdisc;
68         }
69
70         sch_tree_lock(sch);
71         *old = xchg(&p->q, new);
72         qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
73         qdisc_reset(*old);
74         sch_tree_unlock(sch);
75
76         return 0;
77 }
78
79 static struct Qdisc *dsmark_leaf(struct Qdisc *sch, unsigned long arg)
80 {
81         struct dsmark_qdisc_data *p = qdisc_priv(sch);
82         return p->q;
83 }
84
85 static unsigned long dsmark_get(struct Qdisc *sch, u32 classid)
86 {
87         pr_debug("dsmark_get(sch %p,[qdisc %p],classid %x)\n",
88                 sch, qdisc_priv(sch), classid);
89
90         return TC_H_MIN(classid) + 1;
91 }
92
93 static unsigned long dsmark_bind_filter(struct Qdisc *sch,
94                                         unsigned long parent, u32 classid)
95 {
96         return dsmark_get(sch, classid);
97 }
98
99 static void dsmark_put(struct Qdisc *sch, unsigned long cl)
100 {
101 }
102
103 static const struct nla_policy dsmark_policy[TCA_DSMARK_MAX + 1] = {
104         [TCA_DSMARK_INDICES]            = { .type = NLA_U16 },
105         [TCA_DSMARK_DEFAULT_INDEX]      = { .type = NLA_U16 },
106         [TCA_DSMARK_SET_TC_INDEX]       = { .type = NLA_FLAG },
107         [TCA_DSMARK_MASK]               = { .type = NLA_U8 },
108         [TCA_DSMARK_VALUE]              = { .type = NLA_U8 },
109 };
110
111 static int dsmark_change(struct Qdisc *sch, u32 classid, u32 parent,
112                          struct nlattr **tca, unsigned long *arg)
113 {
114         struct dsmark_qdisc_data *p = qdisc_priv(sch);
115         struct nlattr *opt = tca[TCA_OPTIONS];
116         struct nlattr *tb[TCA_DSMARK_MAX + 1];
117         int err = -EINVAL;
118         u8 mask = 0;
119
120         pr_debug("dsmark_change(sch %p,[qdisc %p],classid %x,parent %x),"
121                 "arg 0x%lx\n", sch, p, classid, parent, *arg);
122
123         if (!dsmark_valid_index(p, *arg)) {
124                 err = -ENOENT;
125                 goto errout;
126         }
127
128         if (!opt)
129                 goto errout;
130
131         err = nla_parse_nested(tb, TCA_DSMARK_MAX, opt, dsmark_policy);
132         if (err < 0)
133                 goto errout;
134
135         if (tb[TCA_DSMARK_MASK])
136                 mask = nla_get_u8(tb[TCA_DSMARK_MASK]);
137
138         if (tb[TCA_DSMARK_VALUE])
139                 p->value[*arg-1] = nla_get_u8(tb[TCA_DSMARK_VALUE]);
140
141         if (tb[TCA_DSMARK_MASK])
142                 p->mask[*arg-1] = mask;
143
144         err = 0;
145
146 errout:
147         return err;
148 }
149
150 static int dsmark_delete(struct Qdisc *sch, unsigned long arg)
151 {
152         struct dsmark_qdisc_data *p = qdisc_priv(sch);
153
154         if (!dsmark_valid_index(p, arg))
155                 return -EINVAL;
156
157         p->mask[arg-1] = 0xff;
158         p->value[arg-1] = 0;
159
160         return 0;
161 }
162
163 static void dsmark_walk(struct Qdisc *sch, struct qdisc_walker *walker)
164 {
165         struct dsmark_qdisc_data *p = qdisc_priv(sch);
166         int i;
167
168         pr_debug("dsmark_walk(sch %p,[qdisc %p],walker %p)\n", sch, p, walker);
169
170         if (walker->stop)
171                 return;
172
173         for (i = 0; i < p->indices; i++) {
174                 if (p->mask[i] == 0xff && !p->value[i])
175                         goto ignore;
176                 if (walker->count >= walker->skip) {
177                         if (walker->fn(sch, i+1, walker) < 0) {
178                                 walker->stop = 1;
179                                 break;
180                         }
181                 }
182 ignore:
183                 walker->count++;
184         }
185 }
186
187 static inline struct tcf_proto **dsmark_find_tcf(struct Qdisc *sch,
188                                                  unsigned long cl)
189 {
190         struct dsmark_qdisc_data *p = qdisc_priv(sch);
191         return &p->filter_list;
192 }
193
194 /* --------------------------- Qdisc operations ---------------------------- */
195
196 static int dsmark_enqueue(struct sk_buff *skb, struct Qdisc *sch)
197 {
198         struct dsmark_qdisc_data *p = qdisc_priv(sch);
199         int err;
200
201         pr_debug("dsmark_enqueue(skb %p,sch %p,[qdisc %p])\n", skb, sch, p);
202
203         if (p->set_tc_index) {
204                 switch (skb->protocol) {
205                 case htons(ETH_P_IP):
206                         if (skb_cow_head(skb, sizeof(struct iphdr)))
207                                 goto drop;
208
209                         skb->tc_index = ipv4_get_dsfield(ip_hdr(skb))
210                                 & ~INET_ECN_MASK;
211                         break;
212
213                 case htons(ETH_P_IPV6):
214                         if (skb_cow_head(skb, sizeof(struct ipv6hdr)))
215                                 goto drop;
216
217                         skb->tc_index = ipv6_get_dsfield(ipv6_hdr(skb))
218                                 & ~INET_ECN_MASK;
219                         break;
220                 default:
221                         skb->tc_index = 0;
222                         break;
223                 }
224         }
225
226         if (TC_H_MAJ(skb->priority) == sch->handle)
227                 skb->tc_index = TC_H_MIN(skb->priority);
228         else {
229                 struct tcf_result res;
230                 int result = tc_classify(skb, p->filter_list, &res);
231
232                 pr_debug("result %d class 0x%04x\n", result, res.classid);
233
234                 switch (result) {
235 #ifdef CONFIG_NET_CLS_ACT
236                 case TC_ACT_QUEUED:
237                 case TC_ACT_STOLEN:
238                         kfree_skb(skb);
239                         return NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
240
241                 case TC_ACT_SHOT:
242                         goto drop;
243 #endif
244                 case TC_ACT_OK:
245                         skb->tc_index = TC_H_MIN(res.classid);
246                         break;
247
248                 default:
249                         if (p->default_index != NO_DEFAULT_INDEX)
250                                 skb->tc_index = p->default_index;
251                         break;
252                 }
253         }
254
255         err = qdisc_enqueue(skb, p->q);
256         if (err != NET_XMIT_SUCCESS) {
257                 if (net_xmit_drop_count(err))
258                         sch->qstats.drops++;
259                 return err;
260         }
261
262         sch->bstats.bytes += qdisc_pkt_len(skb);
263         sch->bstats.packets++;
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(qdisc_dev(sch), sch->dev_queue,
386                                  &pfifo_qdisc_ops, sch->handle);
387         if (p->q == NULL)
388                 p->q = &noop_qdisc;
389
390         pr_debug("dsmark_init: qdisc %p\n", p->q);
391
392         err = 0;
393 errout:
394         return err;
395 }
396
397 static void dsmark_reset(struct Qdisc *sch)
398 {
399         struct dsmark_qdisc_data *p = qdisc_priv(sch);
400
401         pr_debug("dsmark_reset(sch %p,[qdisc %p])\n", sch, p);
402         qdisc_reset(p->q);
403         sch->q.qlen = 0;
404 }
405
406 static void dsmark_destroy(struct Qdisc *sch)
407 {
408         struct dsmark_qdisc_data *p = qdisc_priv(sch);
409
410         pr_debug("dsmark_destroy(sch %p,[qdisc %p])\n", sch, p);
411
412         tcf_destroy_chain(&p->filter_list);
413         qdisc_destroy(p->q);
414         kfree(p->mask);
415 }
416
417 static int dsmark_dump_class(struct Qdisc *sch, unsigned long cl,
418                              struct sk_buff *skb, struct tcmsg *tcm)
419 {
420         struct dsmark_qdisc_data *p = qdisc_priv(sch);
421         struct nlattr *opts = NULL;
422
423         pr_debug("dsmark_dump_class(sch %p,[qdisc %p],class %ld\n", sch, p, cl);
424
425         if (!dsmark_valid_index(p, cl))
426                 return -EINVAL;
427
428         tcm->tcm_handle = TC_H_MAKE(TC_H_MAJ(sch->handle), cl-1);
429         tcm->tcm_info = p->q->handle;
430
431         opts = nla_nest_start(skb, TCA_OPTIONS);
432         if (opts == NULL)
433                 goto nla_put_failure;
434         NLA_PUT_U8(skb, TCA_DSMARK_MASK, p->mask[cl-1]);
435         NLA_PUT_U8(skb, TCA_DSMARK_VALUE, p->value[cl-1]);
436
437         return nla_nest_end(skb, opts);
438
439 nla_put_failure:
440         nla_nest_cancel(skb, opts);
441         return -EMSGSIZE;
442 }
443
444 static int dsmark_dump(struct Qdisc *sch, struct sk_buff *skb)
445 {
446         struct dsmark_qdisc_data *p = qdisc_priv(sch);
447         struct nlattr *opts = NULL;
448
449         opts = nla_nest_start(skb, TCA_OPTIONS);
450         if (opts == NULL)
451                 goto nla_put_failure;
452         NLA_PUT_U16(skb, TCA_DSMARK_INDICES, p->indices);
453
454         if (p->default_index != NO_DEFAULT_INDEX)
455                 NLA_PUT_U16(skb, TCA_DSMARK_DEFAULT_INDEX, p->default_index);
456
457         if (p->set_tc_index)
458                 NLA_PUT_FLAG(skb, TCA_DSMARK_SET_TC_INDEX);
459
460         return nla_nest_end(skb, opts);
461
462 nla_put_failure:
463         nla_nest_cancel(skb, opts);
464         return -EMSGSIZE;
465 }
466
467 static const struct Qdisc_class_ops dsmark_class_ops = {
468         .graft          =       dsmark_graft,
469         .leaf           =       dsmark_leaf,
470         .get            =       dsmark_get,
471         .put            =       dsmark_put,
472         .change         =       dsmark_change,
473         .delete         =       dsmark_delete,
474         .walk           =       dsmark_walk,
475         .tcf_chain      =       dsmark_find_tcf,
476         .bind_tcf       =       dsmark_bind_filter,
477         .unbind_tcf     =       dsmark_put,
478         .dump           =       dsmark_dump_class,
479 };
480
481 static struct Qdisc_ops dsmark_qdisc_ops __read_mostly = {
482         .next           =       NULL,
483         .cl_ops         =       &dsmark_class_ops,
484         .id             =       "dsmark",
485         .priv_size      =       sizeof(struct dsmark_qdisc_data),
486         .enqueue        =       dsmark_enqueue,
487         .dequeue        =       dsmark_dequeue,
488         .peek           =       dsmark_peek,
489         .drop           =       dsmark_drop,
490         .init           =       dsmark_init,
491         .reset          =       dsmark_reset,
492         .destroy        =       dsmark_destroy,
493         .change         =       NULL,
494         .dump           =       dsmark_dump,
495         .owner          =       THIS_MODULE,
496 };
497
498 static int __init dsmark_module_init(void)
499 {
500         return register_qdisc(&dsmark_qdisc_ops);
501 }
502
503 static void __exit dsmark_module_exit(void)
504 {
505         unregister_qdisc(&dsmark_qdisc_ops);
506 }
507
508 module_init(dsmark_module_init)
509 module_exit(dsmark_module_exit)
510
511 MODULE_LICENSE("GPL");