e38e0d00d1e6204be6e1fcd7e24eff74529d921d
[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/netdevice.h> /* for pkt_sched */
13 #include <linux/rtnetlink.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 #if 0 /* control */
21 #define DPRINTK(format,args...) printk(KERN_DEBUG format,##args)
22 #else
23 #define DPRINTK(format,args...)
24 #endif
25
26 #if 0 /* data */
27 #define D2PRINTK(format,args...) printk(KERN_DEBUG format,##args)
28 #else
29 #define D2PRINTK(format,args...)
30 #endif
31
32
33 #define PRIV(sch) ((struct dsmark_qdisc_data *) qdisc_priv(sch))
34
35
36 /*
37  * classid      class           marking
38  * -------      -----           -------
39  *   n/a          0             n/a
40  *   x:0          1             use entry [0]
41  *   ...         ...            ...
42  *   x:y y>0     y+1            use entry [y]
43  *   ...         ...            ...
44  * x:indices-1  indices         use entry [indices-1]
45  *   ...         ...            ...
46  *   x:y         y+1            use entry [y & (indices-1)]
47  *   ...         ...            ...
48  * 0xffff       0x10000         use entry [indices-1]
49  */
50
51
52 #define NO_DEFAULT_INDEX        (1 << 16)
53
54 struct dsmark_qdisc_data {
55         struct Qdisc            *q;
56         struct tcf_proto        *filter_list;
57         u8                      *mask;  /* "owns" the array */
58         u8                      *value;
59         u16                     indices;
60         u32                     default_index;  /* index range is 0...0xffff */
61         int                     set_tc_index;
62 };
63
64 static inline int dsmark_valid_indices(u16 indices)
65 {
66         while (indices != 1) {
67                 if (indices & 1)
68                         return 0;
69                 indices >>= 1;
70         }
71
72         return 1;
73 }
74
75 static inline int dsmark_valid_index(struct dsmark_qdisc_data *p, u16 index)
76 {
77         return (index <= p->indices && index > 0);
78 }
79
80 /* ------------------------- Class/flow operations ------------------------- */
81
82 static int dsmark_graft(struct Qdisc *sch, unsigned long arg,
83                         struct Qdisc *new, struct Qdisc **old)
84 {
85         struct dsmark_qdisc_data *p = PRIV(sch);
86
87         DPRINTK("dsmark_graft(sch %p,[qdisc %p],new %p,old %p)\n",
88                 sch, p, new, old);
89
90         if (new == NULL) {
91                 new = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops,
92                                         sch->handle);
93                 if (new == NULL)
94                         new = &noop_qdisc;
95         }
96
97         sch_tree_lock(sch);
98         *old = xchg(&p->q, new);
99         qdisc_tree_decrease_qlen(*old, (*old)->q.qlen);
100         qdisc_reset(*old);
101         sch_tree_unlock(sch);
102
103         return 0;
104 }
105
106 static struct Qdisc *dsmark_leaf(struct Qdisc *sch, unsigned long arg)
107 {
108         return PRIV(sch)->q;
109 }
110
111 static unsigned long dsmark_get(struct Qdisc *sch, u32 classid)
112 {
113         DPRINTK("dsmark_get(sch %p,[qdisc %p],classid %x)\n",
114                 sch, PRIV(sch), classid);
115
116         return TC_H_MIN(classid) + 1;
117 }
118
119 static unsigned long dsmark_bind_filter(struct Qdisc *sch,
120                                         unsigned long parent, u32 classid)
121 {
122         return dsmark_get(sch, classid);
123 }
124
125 static void dsmark_put(struct Qdisc *sch, unsigned long cl)
126 {
127 }
128
129 static int dsmark_change(struct Qdisc *sch, u32 classid, u32 parent,
130                          struct rtattr **tca, unsigned long *arg)
131 {
132         struct dsmark_qdisc_data *p = PRIV(sch);
133         struct rtattr *opt = tca[TCA_OPTIONS-1];
134         struct rtattr *tb[TCA_DSMARK_MAX];
135         int err = -EINVAL;
136         u8 mask = 0;
137
138         DPRINTK("dsmark_change(sch %p,[qdisc %p],classid %x,parent %x),"
139                 "arg 0x%lx\n", sch, p, classid, parent, *arg);
140
141         if (!dsmark_valid_index(p, *arg)) {
142                 err = -ENOENT;
143                 goto rtattr_failure;
144         }
145
146         if (!opt || rtattr_parse_nested(tb, TCA_DSMARK_MAX, opt))
147                 goto rtattr_failure;
148
149         if (tb[TCA_DSMARK_MASK-1])
150                 mask = RTA_GET_U8(tb[TCA_DSMARK_MASK-1]);
151
152         if (tb[TCA_DSMARK_VALUE-1])
153                 p->value[*arg-1] = RTA_GET_U8(tb[TCA_DSMARK_VALUE-1]);
154
155         if (tb[TCA_DSMARK_MASK-1])
156                 p->mask[*arg-1] = mask;
157
158         err = 0;
159
160 rtattr_failure:
161         return err;
162 }
163
164 static int dsmark_delete(struct Qdisc *sch, unsigned long arg)
165 {
166         struct dsmark_qdisc_data *p = PRIV(sch);
167
168         if (!dsmark_valid_index(p, arg))
169                 return -EINVAL;
170
171         p->mask[arg-1] = 0xff;
172         p->value[arg-1] = 0;
173
174         return 0;
175 }
176
177 static void dsmark_walk(struct Qdisc *sch,struct qdisc_walker *walker)
178 {
179         struct dsmark_qdisc_data *p = PRIV(sch);
180         int i;
181
182         DPRINTK("dsmark_walk(sch %p,[qdisc %p],walker %p)\n", sch, p, walker);
183
184         if (walker->stop)
185                 return;
186
187         for (i = 0; i < p->indices; i++) {
188                 if (p->mask[i] == 0xff && !p->value[i])
189                         goto ignore;
190                 if (walker->count >= walker->skip) {
191                         if (walker->fn(sch, i+1, walker) < 0) {
192                                 walker->stop = 1;
193                                 break;
194                         }
195                 }
196 ignore:
197                 walker->count++;
198         }
199 }
200
201 static struct tcf_proto **dsmark_find_tcf(struct Qdisc *sch,unsigned long cl)
202 {
203         return &PRIV(sch)->filter_list;
204 }
205
206 /* --------------------------- Qdisc operations ---------------------------- */
207
208 static int dsmark_enqueue(struct sk_buff *skb,struct Qdisc *sch)
209 {
210         struct dsmark_qdisc_data *p = PRIV(sch);
211         int err;
212
213         D2PRINTK("dsmark_enqueue(skb %p,sch %p,[qdisc %p])\n", skb, sch, p);
214
215         if (p->set_tc_index) {
216                 /* FIXME: Safe with non-linear skbs? --RR */
217                 switch (skb->protocol) {
218                         case __constant_htons(ETH_P_IP):
219                                 skb->tc_index = ipv4_get_dsfield(ip_hdr(skb))
220                                         & ~INET_ECN_MASK;
221                                 break;
222                         case __constant_htons(ETH_P_IPV6):
223                                 skb->tc_index = ipv6_get_dsfield(ipv6_hdr(skb))
224                                         & ~INET_ECN_MASK;
225                                 break;
226                         default:
227                                 skb->tc_index = 0;
228                                 break;
229                 };
230         }
231
232         if (TC_H_MAJ(skb->priority) == sch->handle)
233                 skb->tc_index = TC_H_MIN(skb->priority);
234         else {
235                 struct tcf_result res;
236                 int result = tc_classify(skb, p->filter_list, &res);
237
238                 D2PRINTK("result %d class 0x%04x\n", result, res.classid);
239
240                 switch (result) {
241 #ifdef CONFIG_NET_CLS_POLICE
242                         case TC_POLICE_SHOT:
243                                 kfree_skb(skb);
244                                 sch->qstats.drops++;
245                                 return NET_XMIT_POLICED;
246 #if 0
247                         case TC_POLICE_RECLASSIFY:
248                                 /* FIXME: what to do here ??? */
249 #endif
250 #endif
251                         case TC_POLICE_OK:
252                                 skb->tc_index = TC_H_MIN(res.classid);
253                                 break;
254                         case TC_POLICE_UNSPEC:
255                                 /* fall through */
256                         default:
257                                 if (p->default_index != NO_DEFAULT_INDEX)
258                                         skb->tc_index = p->default_index;
259                                 break;
260                 };
261         }
262
263         err = p->q->enqueue(skb,p->q);
264         if (err != NET_XMIT_SUCCESS) {
265                 sch->qstats.drops++;
266                 return err;
267         }
268
269         sch->bstats.bytes += skb->len;
270         sch->bstats.packets++;
271         sch->q.qlen++;
272
273         return NET_XMIT_SUCCESS;
274 }
275
276 static struct sk_buff *dsmark_dequeue(struct Qdisc *sch)
277 {
278         struct dsmark_qdisc_data *p = PRIV(sch);
279         struct sk_buff *skb;
280         u32 index;
281
282         D2PRINTK("dsmark_dequeue(sch %p,[qdisc %p])\n", sch, p);
283
284         skb = p->q->ops->dequeue(p->q);
285         if (skb == NULL)
286                 return NULL;
287
288         sch->q.qlen--;
289
290         index = skb->tc_index & (p->indices - 1);
291         D2PRINTK("index %d->%d\n", skb->tc_index, index);
292
293         switch (skb->protocol) {
294                 case __constant_htons(ETH_P_IP):
295                         ipv4_change_dsfield(ip_hdr(skb), p->mask[index],
296                                             p->value[index]);
297                         break;
298                 case __constant_htons(ETH_P_IPV6):
299                         ipv6_change_dsfield(ipv6_hdr(skb), p->mask[index],
300                                             p->value[index]);
301                         break;
302                 default:
303                         /*
304                          * Only complain if a change was actually attempted.
305                          * This way, we can send non-IP traffic through dsmark
306                          * and don't need yet another qdisc as a bypass.
307                          */
308                         if (p->mask[index] != 0xff || p->value[index])
309                                 printk(KERN_WARNING "dsmark_dequeue: "
310                                        "unsupported protocol %d\n",
311                                        ntohs(skb->protocol));
312                         break;
313         };
314
315         return skb;
316 }
317
318 static int dsmark_requeue(struct sk_buff *skb,struct Qdisc *sch)
319 {
320         struct dsmark_qdisc_data *p = PRIV(sch);
321         int err;
322
323         D2PRINTK("dsmark_requeue(skb %p,sch %p,[qdisc %p])\n", skb, sch, p);
324
325         err = p->q->ops->requeue(skb, p->q);
326         if (err != NET_XMIT_SUCCESS) {
327                 sch->qstats.drops++;
328                 return err;
329         }
330
331         sch->q.qlen++;
332         sch->qstats.requeues++;
333
334         return NET_XMIT_SUCCESS;
335 }
336
337 static unsigned int dsmark_drop(struct Qdisc *sch)
338 {
339         struct dsmark_qdisc_data *p = PRIV(sch);
340         unsigned int len;
341
342         DPRINTK("dsmark_reset(sch %p,[qdisc %p])\n", sch, p);
343
344         if (p->q->ops->drop == NULL)
345                 return 0;
346
347         len = p->q->ops->drop(p->q);
348         if (len)
349                 sch->q.qlen--;
350
351         return len;
352 }
353
354 static int dsmark_init(struct Qdisc *sch, struct rtattr *opt)
355 {
356         struct dsmark_qdisc_data *p = PRIV(sch);
357         struct rtattr *tb[TCA_DSMARK_MAX];
358         int err = -EINVAL;
359         u32 default_index = NO_DEFAULT_INDEX;
360         u16 indices;
361         u8 *mask;
362
363         DPRINTK("dsmark_init(sch %p,[qdisc %p],opt %p)\n", sch, p, opt);
364
365         if (!opt || rtattr_parse_nested(tb, TCA_DSMARK_MAX, opt) < 0)
366                 goto errout;
367
368         indices = RTA_GET_U16(tb[TCA_DSMARK_INDICES-1]);
369         if (!indices || !dsmark_valid_indices(indices))
370                 goto errout;
371
372         if (tb[TCA_DSMARK_DEFAULT_INDEX-1])
373                 default_index = RTA_GET_U16(tb[TCA_DSMARK_DEFAULT_INDEX-1]);
374
375         mask = kmalloc(indices * 2, GFP_KERNEL);
376         if (mask == NULL) {
377                 err = -ENOMEM;
378                 goto errout;
379         }
380
381         p->mask = mask;
382         memset(p->mask, 0xff, indices);
383
384         p->value = p->mask + indices;
385         memset(p->value, 0, indices);
386
387         p->indices = indices;
388         p->default_index = default_index;
389         p->set_tc_index = RTA_GET_FLAG(tb[TCA_DSMARK_SET_TC_INDEX-1]);
390
391         p->q = qdisc_create_dflt(sch->dev, &pfifo_qdisc_ops, sch->handle);
392         if (p->q == NULL)
393                 p->q = &noop_qdisc;
394
395         DPRINTK("dsmark_init: qdisc %p\n", p->q);
396
397         err = 0;
398 errout:
399 rtattr_failure:
400         return err;
401 }
402
403 static void dsmark_reset(struct Qdisc *sch)
404 {
405         struct dsmark_qdisc_data *p = PRIV(sch);
406
407         DPRINTK("dsmark_reset(sch %p,[qdisc %p])\n", sch, p);
408         qdisc_reset(p->q);
409         sch->q.qlen = 0;
410 }
411
412 static void dsmark_destroy(struct Qdisc *sch)
413 {
414         struct dsmark_qdisc_data *p = PRIV(sch);
415
416         DPRINTK("dsmark_destroy(sch %p,[qdisc %p])\n", sch, p);
417
418         tcf_destroy_chain(p->filter_list);
419         qdisc_destroy(p->q);
420         kfree(p->mask);
421 }
422
423 static int dsmark_dump_class(struct Qdisc *sch, unsigned long cl,
424                              struct sk_buff *skb, struct tcmsg *tcm)
425 {
426         struct dsmark_qdisc_data *p = PRIV(sch);
427         struct rtattr *opts = NULL;
428
429         DPRINTK("dsmark_dump_class(sch %p,[qdisc %p],class %ld\n", sch, p, cl);
430
431         if (!dsmark_valid_index(p, cl))
432                 return -EINVAL;
433
434         tcm->tcm_handle = TC_H_MAKE(TC_H_MAJ(sch->handle), cl-1);
435         tcm->tcm_info = p->q->handle;
436
437         opts = RTA_NEST(skb, TCA_OPTIONS);
438         RTA_PUT_U8(skb,TCA_DSMARK_MASK, p->mask[cl-1]);
439         RTA_PUT_U8(skb,TCA_DSMARK_VALUE, p->value[cl-1]);
440
441         return RTA_NEST_END(skb, opts);
442
443 rtattr_failure:
444         return RTA_NEST_CANCEL(skb, opts);
445 }
446
447 static int dsmark_dump(struct Qdisc *sch, struct sk_buff *skb)
448 {
449         struct dsmark_qdisc_data *p = PRIV(sch);
450         struct rtattr *opts = NULL;
451
452         opts = RTA_NEST(skb, TCA_OPTIONS);
453         RTA_PUT_U16(skb, TCA_DSMARK_INDICES, p->indices);
454
455         if (p->default_index != NO_DEFAULT_INDEX)
456                 RTA_PUT_U16(skb, TCA_DSMARK_DEFAULT_INDEX, p->default_index);
457
458         if (p->set_tc_index)
459                 RTA_PUT_FLAG(skb, TCA_DSMARK_SET_TC_INDEX);
460
461         return RTA_NEST_END(skb, opts);
462
463 rtattr_failure:
464         return RTA_NEST_CANCEL(skb, opts);
465 }
466
467 static 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 = {
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         .requeue        =       dsmark_requeue,
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");