1 // SPDX-License-Identifier: GPL-2.0-only
2 /* net/sched/sch_dsmark.c - Differentiated Services field marker */
3
4 /* Written 1998-2000 by Werner Almesberger, EPFL ICA */
5
6
7 #include <linux/module.h>
8 #include <linux/init.h>
9 #include <linux/slab.h>
10 #include <linux/types.h>
11 #include <linux/string.h>
12 #include <linux/errno.h>
13 #include <linux/skbuff.h>
14 #include <linux/rtnetlink.h>
15 #include <linux/bitops.h>
16 #include <net/pkt_sched.h>
17 #include <net/pkt_cls.h>
18 #include <net/dsfield.h>
19 #include <net/inet_ecn.h>
20 #include <asm/byteorder.h>
21
22 /*
23 * classid class marking
24 * ------- ----- -------
25 * n/a 0 n/a
26 * x:0 1 use entry [0]
27 * ... ... ...
28 * x:y y>0 y+1 use entry [y]
29 * ... ... ...
30 * x:indices-1 indices use entry [indices-1]
31 * ... ... ...
32 * x:y y+1 use entry [y & (indices-1)]
33 * ... ... ...
34 * 0xffff 0x10000 use entry [indices-1]
35 */
36
37
38 #define NO_DEFAULT_INDEX (1 << 16)
39
40 struct mask_value {
41 u8 mask;
42 u8 value;
43 };
44
45 struct dsmark_qdisc_data {
46 struct Qdisc *q;
47 struct tcf_proto __rcu *filter_list;
48 struct tcf_block *block;
49 struct mask_value *mv;
50 u16 indices;
51 u8 set_tc_index;
52 u32 default_index; /* index range is 0...0xffff */
53 #define DSMARK_EMBEDDED_SZ 16
54 struct mask_value embedded[DSMARK_EMBEDDED_SZ];
55 };
56
57 static inline int dsmark_valid_index(struct dsmark_qdisc_data *p, u16 index)
58 {
59 return index <= p->indices && index > 0;
60 }
61
62 /* ------------------------- Class/flow operations ------------------------- */
63
64 static int dsmark_graft(struct Qdisc *sch, unsigned long arg,
65 struct Qdisc *new, struct Qdisc **old,
66 struct netlink_ext_ack *extack)
67 {
68 struct dsmark_qdisc_data *p = qdisc_priv(sch);
69
70 pr_debug("%s(sch %p,[qdisc %p],new %p,old %p)\n",
71 __func__, sch, p, new, old);
72
73 if (new == NULL) {
74 new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
75 sch->handle, NULL);
76 if (new == NULL)
77 new = &noop_qdisc;
78 }
79
80 *old = qdisc_replace(sch, new, &p->q);
81 return 0;
82 }
83
84 static struct Qdisc *dsmark_leaf(struct Qdisc *sch, unsigned long arg)
85 {
86 struct dsmark_qdisc_data *p = qdisc_priv(sch);
87 return p->q;
88 }
89
90 static unsigned long dsmark_find(struct Qdisc *sch, u32 classid)
91 {
92 return TC_H_MIN(classid) + 1;
93 }
94
95 static unsigned long dsmark_bind_filter(struct Qdisc *sch,
96 unsigned long parent, u32 classid)
97 {
98 pr_debug("%s(sch %p,[qdisc %p],classid %x)\n",
99 __func__, sch, qdisc_priv(sch), classid);
100
101 return dsmark_find(sch, classid);
102 }
103
104 static void dsmark_unbind_filter(struct Qdisc *sch, unsigned long cl)
105 {
106 }
107
108 static const struct nla_policy dsmark_policy[TCA_DSMARK_MAX + 1] = {
109 [TCA_DSMARK_INDICES] = { .type = NLA_U16 },
110 [TCA_DSMARK_DEFAULT_INDEX] = { .type = NLA_U16 },
111 [TCA_DSMARK_SET_TC_INDEX] = { .type = NLA_FLAG },
112 [TCA_DSMARK_MASK] = { .type = NLA_U8 },
113 [TCA_DSMARK_VALUE] = { .type = NLA_U8 },
114 };
115
116 static int dsmark_change(struct Qdisc *sch, u32 classid, u32 parent,
117 struct nlattr **tca, unsigned long *arg,
118 struct netlink_ext_ack *extack)
119 {
120 struct dsmark_qdisc_data *p = qdisc_priv(sch);
121 struct nlattr *opt = tca[TCA_OPTIONS];
122 struct nlattr *tb[TCA_DSMARK_MAX + 1];
123 int err = -EINVAL;
124
125 pr_debug("%s(sch %p,[qdisc %p],classid %x,parent %x), arg 0x%lx\n",
126 __func__, sch, p, classid, parent, *arg);
127
128 if (!dsmark_valid_index(p, *arg)) {
129 err = -ENOENT;
130 goto errout;
131 }
132
133 if (!opt)
134 goto errout;
135
136 err = nla_parse_nested_deprecated(tb, TCA_DSMARK_MAX, opt,
137 dsmark_policy, NULL);
138 if (err < 0)
139 goto errout;
140
141 if (tb[TCA_DSMARK_VALUE])
142 p->mv[*arg - 1].value = nla_get_u8(tb[TCA_DSMARK_VALUE]);
143
144 if (tb[TCA_DSMARK_MASK])
145 p->mv[*arg - 1].mask = nla_get_u8(tb[TCA_DSMARK_MASK]);
146
147 err = 0;
148
149 errout:
150 return err;
151 }
152
153 static int dsmark_delete(struct Qdisc *sch, unsigned long arg)
154 {
155 struct dsmark_qdisc_data *p = qdisc_priv(sch);
156
157 if (!dsmark_valid_index(p, arg))
158 return -EINVAL;
159
160 p->mv[arg - 1].mask = 0xff;
161 p->mv[arg - 1].value = 0;
162
163 return 0;
164 }
165
166 static void dsmark_walk(struct Qdisc *sch, struct qdisc_walker *walker)
167 {
168 struct dsmark_qdisc_data *p = qdisc_priv(sch);
169 int i;
170
171 pr_debug("%s(sch %p,[qdisc %p],walker %p)\n",
172 __func__, sch, p, walker);
173
174 if (walker->stop)
175 return;
176
177 for (i = 0; i < p->indices; i++) {
178 if (p->mv[i].mask == 0xff && !p->mv[i].value)
179 goto ignore;
180 if (walker->count >= walker->skip) {
181 if (walker->fn(sch, i + 1, walker) < 0) {
182 walker->stop = 1;
183 break;
184 }
185 }
186 ignore:
187 walker->count++;
188 }
189 }
190
191 static struct tcf_block *dsmark_tcf_block(struct Qdisc *sch, unsigned long cl,
192 struct netlink_ext_ack *extack)
193 {
194 struct dsmark_qdisc_data *p = qdisc_priv(sch);
195
196 return p->block;
197 }
198
199 /* --------------------------- Qdisc operations ---------------------------- */
200
201 static int dsmark_enqueue(struct sk_buff *skb, struct Qdisc *sch,
202 struct sk_buff **to_free)
203 {
204 unsigned int len = qdisc_pkt_len(skb);
205 struct dsmark_qdisc_data *p = qdisc_priv(sch);
206 int err;
207
208 pr_debug("%s(skb %p,sch %p,[qdisc %p])\n", __func__, skb, sch, p);
209
210 if (p->set_tc_index) {
211 int wlen = skb_network_offset(skb);
212
213 switch (tc_skb_protocol(skb)) {
214 case htons(ETH_P_IP):
215 wlen += sizeof(struct iphdr);
216 if (!pskb_may_pull(skb, wlen) ||
217 skb_try_make_writable(skb, wlen))
218 goto drop;
219
220 skb->tc_index = ipv4_get_dsfield(ip_hdr(skb))
221 & ~INET_ECN_MASK;
222 break;
223
224 case htons(ETH_P_IPV6):
225 wlen += sizeof(struct ipv6hdr);
226 if (!pskb_may_pull(skb, wlen) ||
227 skb_try_make_writable(skb, wlen))
228 goto drop;
229
230 skb->tc_index = ipv6_get_dsfield(ipv6_hdr(skb))
231 & ~INET_ECN_MASK;
232 break;
233 default:
234 skb->tc_index = 0;
235 break;
236 }
237 }
238
239 if (TC_H_MAJ(skb->priority) == sch->handle)
240 skb->tc_index = TC_H_MIN(skb->priority);
241 else {
242 struct tcf_result res;
243 struct tcf_proto *fl = rcu_dereference_bh(p->filter_list);
244 int result = tcf_classify(skb, fl, &res, false);
245
246 pr_debug("result %d class 0x%04x\n", result, res.classid);
247
248 switch (result) {
249 #ifdef CONFIG_NET_CLS_ACT
250 case TC_ACT_QUEUED:
251 case TC_ACT_STOLEN:
252 case TC_ACT_TRAP:
253 __qdisc_drop(skb, to_free);
254 return NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
255
256 case TC_ACT_SHOT:
257 goto drop;
258 #endif
259 case TC_ACT_OK:
260 skb->tc_index = TC_H_MIN(res.classid);
261 break;
262
263 default:
264 if (p->default_index != NO_DEFAULT_INDEX)
265 skb->tc_index = p->default_index;
266 break;
267 }
268 }
269
270 err = qdisc_enqueue(skb, p->q, to_free);
271 if (err != NET_XMIT_SUCCESS) {
272 if (net_xmit_drop_count(err))
273 qdisc_qstats_drop(sch);
274 return err;
275 }
276
277 sch->qstats.backlog += len;
278 sch->q.qlen++;
279
280 return NET_XMIT_SUCCESS;
281
282 drop:
283 qdisc_drop(skb, sch, to_free);
284 return NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
285 }
286
287 static struct sk_buff *dsmark_dequeue(struct Qdisc *sch)
288 {
289 struct dsmark_qdisc_data *p = qdisc_priv(sch);
290 struct sk_buff *skb;
291 u32 index;
292
293 pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
294
295 skb = qdisc_dequeue_peeked(p->q);
296 if (skb == NULL)
297 return NULL;
298
299 qdisc_bstats_update(sch, skb);
300 qdisc_qstats_backlog_dec(sch, skb);
301 sch->q.qlen--;
302
303 index = skb->tc_index & (p->indices - 1);
304 pr_debug("index %d->%d\n", skb->tc_index, index);
305
306 switch (tc_skb_protocol(skb)) {
307 case htons(ETH_P_IP):
308 ipv4_change_dsfield(ip_hdr(skb), p->mv[index].mask,
309 p->mv[index].value);
310 break;
311 case htons(ETH_P_IPV6):
312 ipv6_change_dsfield(ipv6_hdr(skb), p->mv[index].mask,
313 p->mv[index].value);
314 break;
315 default:
316 /*
317 * Only complain if a change was actually attempted.
318 * This way, we can send non-IP traffic through dsmark
319 * and don't need yet another qdisc as a bypass.
320 */
321 if (p->mv[index].mask != 0xff || p->mv[index].value)
322 pr_warn("%s: unsupported protocol %d\n",
323 __func__, ntohs(tc_skb_protocol(skb)));
324 break;
325 }
326
327 return skb;
328 }
329
330 static struct sk_buff *dsmark_peek(struct Qdisc *sch)
331 {
332 struct dsmark_qdisc_data *p = qdisc_priv(sch);
333
334 pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
335
336 return p->q->ops->peek(p->q);
337 }
338
339 static int dsmark_init(struct Qdisc *sch, struct nlattr *opt,
340 struct netlink_ext_ack *extack)
341 {
342 struct dsmark_qdisc_data *p = qdisc_priv(sch);
343 struct nlattr *tb[TCA_DSMARK_MAX + 1];
344 int err = -EINVAL;
345 u32 default_index = NO_DEFAULT_INDEX;
346 u16 indices;
347 int i;
348
349 pr_debug("%s(sch %p,[qdisc %p],opt %p)\n", __func__, sch, p, opt);
350
351 if (!opt)
352 goto errout;
353
354 err = tcf_block_get(&p->block, &p->filter_list, sch, extack);
355 if (err)
356 return err;
357
358 err = nla_parse_nested_deprecated(tb, TCA_DSMARK_MAX, opt,
359 dsmark_policy, NULL);
360 if (err < 0)
361 goto errout;
362
363 err = -EINVAL;
364 if (!tb[TCA_DSMARK_INDICES])
365 goto errout;
366 indices = nla_get_u16(tb[TCA_DSMARK_INDICES]);
367
368 if (hweight32(indices) != 1)
369 goto errout;
370
371 if (tb[TCA_DSMARK_DEFAULT_INDEX])
372 default_index = nla_get_u16(tb[TCA_DSMARK_DEFAULT_INDEX]);
373
374 if (indices <= DSMARK_EMBEDDED_SZ)
375 p->mv = p->embedded;
376 else
377 p->mv = kmalloc_array(indices, sizeof(*p->mv), GFP_KERNEL);
378 if (!p->mv) {
379 err = -ENOMEM;
380 goto errout;
381 }
382 for (i = 0; i < indices; i++) {
383 p->mv[i].mask = 0xff;
384 p->mv[i].value = 0;
385 }
386 p->indices = indices;
387 p->default_index = default_index;
388 p->set_tc_index = nla_get_flag(tb[TCA_DSMARK_SET_TC_INDEX]);
389
390 p->q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops, sch->handle,
391 NULL);
392 if (p->q == NULL)
393 p->q = &noop_qdisc;
394 else
395 qdisc_hash_add(p->q, true);
396
397 pr_debug("%s: qdisc %p\n", __func__, p->q);
398
399 err = 0;
400 errout:
401 return err;
402 }
403
404 static void dsmark_reset(struct Qdisc *sch)
405 {
406 struct dsmark_qdisc_data *p = qdisc_priv(sch);
407
408 pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
409 qdisc_reset(p->q);
410 sch->qstats.backlog = 0;
411 sch->q.qlen = 0;
412 }
413
414 static void dsmark_destroy(struct Qdisc *sch)
415 {
416 struct dsmark_qdisc_data *p = qdisc_priv(sch);
417
418 pr_debug("%s(sch %p,[qdisc %p])\n", __func__, sch, p);
419
420 tcf_block_put(p->block);
421 qdisc_put(p->q);
422 if (p->mv != p->embedded)
423 kfree(p->mv);
424 }
425
426 static int dsmark_dump_class(struct Qdisc *sch, unsigned long cl,
427 struct sk_buff *skb, struct tcmsg *tcm)
428 {
429 struct dsmark_qdisc_data *p = qdisc_priv(sch);
430 struct nlattr *opts = NULL;
431
432 pr_debug("%s(sch %p,[qdisc %p],class %ld\n", __func__, sch, p, cl);
433
434 if (!dsmark_valid_index(p, cl))
435 return -EINVAL;
436
437 tcm->tcm_handle = TC_H_MAKE(TC_H_MAJ(sch->handle), cl - 1);
438 tcm->tcm_info = p->q->handle;
439
440 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
441 if (opts == NULL)
442 goto nla_put_failure;
443 if (nla_put_u8(skb, TCA_DSMARK_MASK, p->mv[cl - 1].mask) ||
444 nla_put_u8(skb, TCA_DSMARK_VALUE, p->mv[cl - 1].value))
445 goto nla_put_failure;
446
447 return nla_nest_end(skb, opts);
448
449 nla_put_failure:
450 nla_nest_cancel(skb, opts);
451 return -EMSGSIZE;
452 }
453
454 static int dsmark_dump(struct Qdisc *sch, struct sk_buff *skb)
455 {
456 struct dsmark_qdisc_data *p = qdisc_priv(sch);
457 struct nlattr *opts = NULL;
458
459 opts = nla_nest_start_noflag(skb, TCA_OPTIONS);
460 if (opts == NULL)
461 goto nla_put_failure;
462 if (nla_put_u16(skb, TCA_DSMARK_INDICES, p->indices))
463 goto nla_put_failure;
464
465 if (p->default_index != NO_DEFAULT_INDEX &&
466 nla_put_u16(skb, TCA_DSMARK_DEFAULT_INDEX, p->default_index))
467 goto nla_put_failure;
468
469 if (p->set_tc_index &&
470 nla_put_flag(skb, TCA_DSMARK_SET_TC_INDEX))
471 goto nla_put_failure;
472
473 return nla_nest_end(skb, opts);
474
475 nla_put_failure:
476 nla_nest_cancel(skb, opts);
477 return -EMSGSIZE;
478 }
479
480 static const struct Qdisc_class_ops dsmark_class_ops = {
481 .graft = dsmark_graft,
482 .leaf = dsmark_leaf,
483 .find = dsmark_find,
484 .change = dsmark_change,
485 .delete = dsmark_delete,
486 .walk = dsmark_walk,
487 .tcf_block = dsmark_tcf_block,
488 .bind_tcf = dsmark_bind_filter,
489 .unbind_tcf = dsmark_unbind_filter,
490 .dump = dsmark_dump_class,
491 };
492
493 static struct Qdisc_ops dsmark_qdisc_ops __read_mostly = {
494 .next = NULL,
495 .cl_ops = &dsmark_class_ops,
496 .id = "dsmark",
497 .priv_size = sizeof(struct dsmark_qdisc_data),
498 .enqueue = dsmark_enqueue,
499 .dequeue = dsmark_dequeue,
500 .peek = dsmark_peek,
501 .init = dsmark_init,
502 .reset = dsmark_reset,
503 .destroy = dsmark_destroy,
504 .change = NULL,
505 .dump = dsmark_dump,
506 .owner = THIS_MODULE,
507 };
508
509 static int __init dsmark_module_init(void)
510 {
511 return register_qdisc(&dsmark_qdisc_ops);
512 }
513
514 static void __exit dsmark_module_exit(void)
515 {
516 unregister_qdisc(&dsmark_qdisc_ops);
517 }
518
519 module_init(dsmark_module_init)
520 module_exit(dsmark_module_exit)
521
522 MODULE_LICENSE("GPL");
523
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.