xref: /linux/net/sched/em_meta.c (revision b6ebbac51bedf9e98e837688bc838f400196da5e)
1 /*
2  * net/sched/em_meta.c	Metadata ematch
3  *
4  *		This program is free software; you can redistribute it and/or
5  *		modify it under the terms of the GNU General Public License
6  *		as published by the Free Software Foundation; either version
7  *		2 of the License, or (at your option) any later version.
8  *
9  * Authors:	Thomas Graf <tgraf@suug.ch>
10  *
11  * ==========================================================================
12  *
13  * 	The metadata ematch compares two meta objects where each object
14  * 	represents either a meta value stored in the kernel or a static
15  * 	value provided by userspace. The objects are not provided by
16  * 	userspace itself but rather a definition providing the information
17  * 	to build them. Every object is of a certain type which must be
18  * 	equal to the object it is being compared to.
19  *
20  * 	The definition of a objects conists of the type (meta type), a
21  * 	identifier (meta id) and additional type specific information.
22  * 	The meta id is either TCF_META_TYPE_VALUE for values provided by
23  * 	userspace or a index to the meta operations table consisting of
24  * 	function pointers to type specific meta data collectors returning
25  * 	the value of the requested meta value.
26  *
27  * 	         lvalue                                   rvalue
28  * 	      +-----------+                           +-----------+
29  * 	      | type: INT |                           | type: INT |
30  * 	 def  | id: DEV   |                           | id: VALUE |
31  * 	      | data:     |                           | data: 3   |
32  * 	      +-----------+                           +-----------+
33  * 	            |                                       |
34  * 	            ---> meta_ops[INT][DEV](...)            |
35  *	                      |                             |
36  * 	            -----------                             |
37  * 	            V                                       V
38  * 	      +-----------+                           +-----------+
39  * 	      | type: INT |                           | type: INT |
40  * 	 obj  | id: DEV |                             | id: VALUE |
41  * 	      | data: 2   |<--data got filled out     | data: 3   |
42  * 	      +-----------+                           +-----------+
43  * 	            |                                         |
44  * 	            --------------> 2  equals 3 <--------------
45  *
46  * 	This is a simplified schema, the complexity varies depending
47  * 	on the meta type. Obviously, the length of the data must also
48  * 	be provided for non-numeric types.
49  *
50  * 	Additionally, type dependent modifiers such as shift operators
51  * 	or mask may be applied to extend the functionaliy. As of now,
52  * 	the variable length type supports shifting the byte string to
53  * 	the right, eating up any number of octets and thus supporting
54  * 	wildcard interface name comparisons such as "ppp%" matching
55  * 	ppp0..9.
56  *
57  * 	NOTE: Certain meta values depend on other subsystems and are
58  * 	      only available if that subsystem is enabled in the kernel.
59  */
60 
61 #include <linux/slab.h>
62 #include <linux/module.h>
63 #include <linux/types.h>
64 #include <linux/kernel.h>
65 #include <linux/sched.h>
66 #include <linux/string.h>
67 #include <linux/skbuff.h>
68 #include <linux/random.h>
69 #include <linux/if_vlan.h>
70 #include <linux/tc_ematch/tc_em_meta.h>
71 #include <net/dst.h>
72 #include <net/route.h>
73 #include <net/pkt_cls.h>
74 #include <net/sock.h>
75 
76 struct meta_obj {
77 	unsigned long		value;
78 	unsigned int		len;
79 };
80 
81 struct meta_value {
82 	struct tcf_meta_val	hdr;
83 	unsigned long		val;
84 	unsigned int		len;
85 };
86 
87 struct meta_match {
88 	struct meta_value	lvalue;
89 	struct meta_value	rvalue;
90 };
91 
92 static inline int meta_id(struct meta_value *v)
93 {
94 	return TCF_META_ID(v->hdr.kind);
95 }
96 
97 static inline int meta_type(struct meta_value *v)
98 {
99 	return TCF_META_TYPE(v->hdr.kind);
100 }
101 
102 #define META_COLLECTOR(FUNC) static void meta_##FUNC(struct sk_buff *skb, \
103 	struct tcf_pkt_info *info, struct meta_value *v, \
104 	struct meta_obj *dst, int *err)
105 
106 /**************************************************************************
107  * System status & misc
108  **************************************************************************/
109 
110 META_COLLECTOR(int_random)
111 {
112 	get_random_bytes(&dst->value, sizeof(dst->value));
113 }
114 
115 static inline unsigned long fixed_loadavg(int load)
116 {
117 	int rnd_load = load + (FIXED_1/200);
118 	int rnd_frac = ((rnd_load & (FIXED_1-1)) * 100) >> FSHIFT;
119 
120 	return ((rnd_load >> FSHIFT) * 100) + rnd_frac;
121 }
122 
123 META_COLLECTOR(int_loadavg_0)
124 {
125 	dst->value = fixed_loadavg(avenrun[0]);
126 }
127 
128 META_COLLECTOR(int_loadavg_1)
129 {
130 	dst->value = fixed_loadavg(avenrun[1]);
131 }
132 
133 META_COLLECTOR(int_loadavg_2)
134 {
135 	dst->value = fixed_loadavg(avenrun[2]);
136 }
137 
138 /**************************************************************************
139  * Device names & indices
140  **************************************************************************/
141 
142 static inline int int_dev(struct net_device *dev, struct meta_obj *dst)
143 {
144 	if (unlikely(dev == NULL))
145 		return -1;
146 
147 	dst->value = dev->ifindex;
148 	return 0;
149 }
150 
151 static inline int var_dev(struct net_device *dev, struct meta_obj *dst)
152 {
153 	if (unlikely(dev == NULL))
154 		return -1;
155 
156 	dst->value = (unsigned long) dev->name;
157 	dst->len = strlen(dev->name);
158 	return 0;
159 }
160 
161 META_COLLECTOR(int_dev)
162 {
163 	*err = int_dev(skb->dev, dst);
164 }
165 
166 META_COLLECTOR(var_dev)
167 {
168 	*err = var_dev(skb->dev, dst);
169 }
170 
171 /**************************************************************************
172  * vlan tag
173  **************************************************************************/
174 
175 META_COLLECTOR(int_vlan_tag)
176 {
177 	unsigned short tag;
178 
179 	tag = skb_vlan_tag_get(skb);
180 	if (!tag && __vlan_get_tag(skb, &tag))
181 		*err = -1;
182 	else
183 		dst->value = tag;
184 }
185 
186 
187 
188 /**************************************************************************
189  * skb attributes
190  **************************************************************************/
191 
192 META_COLLECTOR(int_priority)
193 {
194 	dst->value = skb->priority;
195 }
196 
197 META_COLLECTOR(int_protocol)
198 {
199 	/* Let userspace take care of the byte ordering */
200 	dst->value = tc_skb_protocol(skb);
201 }
202 
203 META_COLLECTOR(int_pkttype)
204 {
205 	dst->value = skb->pkt_type;
206 }
207 
208 META_COLLECTOR(int_pktlen)
209 {
210 	dst->value = skb->len;
211 }
212 
213 META_COLLECTOR(int_datalen)
214 {
215 	dst->value = skb->data_len;
216 }
217 
218 META_COLLECTOR(int_maclen)
219 {
220 	dst->value = skb->mac_len;
221 }
222 
223 META_COLLECTOR(int_rxhash)
224 {
225 	dst->value = skb_get_hash(skb);
226 }
227 
228 /**************************************************************************
229  * Netfilter
230  **************************************************************************/
231 
232 META_COLLECTOR(int_mark)
233 {
234 	dst->value = skb->mark;
235 }
236 
237 /**************************************************************************
238  * Traffic Control
239  **************************************************************************/
240 
241 META_COLLECTOR(int_tcindex)
242 {
243 	dst->value = skb->tc_index;
244 }
245 
246 /**************************************************************************
247  * Routing
248  **************************************************************************/
249 
250 META_COLLECTOR(int_rtclassid)
251 {
252 	if (unlikely(skb_dst(skb) == NULL))
253 		*err = -1;
254 	else
255 #ifdef CONFIG_IP_ROUTE_CLASSID
256 		dst->value = skb_dst(skb)->tclassid;
257 #else
258 		dst->value = 0;
259 #endif
260 }
261 
262 META_COLLECTOR(int_rtiif)
263 {
264 	if (unlikely(skb_rtable(skb) == NULL))
265 		*err = -1;
266 	else
267 		dst->value = inet_iif(skb);
268 }
269 
270 /**************************************************************************
271  * Socket Attributes
272  **************************************************************************/
273 
274 #define skip_nonlocal(skb) \
275 	(unlikely(skb->sk == NULL))
276 
277 META_COLLECTOR(int_sk_family)
278 {
279 	if (skip_nonlocal(skb)) {
280 		*err = -1;
281 		return;
282 	}
283 	dst->value = skb->sk->sk_family;
284 }
285 
286 META_COLLECTOR(int_sk_state)
287 {
288 	if (skip_nonlocal(skb)) {
289 		*err = -1;
290 		return;
291 	}
292 	dst->value = skb->sk->sk_state;
293 }
294 
295 META_COLLECTOR(int_sk_reuse)
296 {
297 	if (skip_nonlocal(skb)) {
298 		*err = -1;
299 		return;
300 	}
301 	dst->value = skb->sk->sk_reuse;
302 }
303 
304 META_COLLECTOR(int_sk_bound_if)
305 {
306 	if (skip_nonlocal(skb)) {
307 		*err = -1;
308 		return;
309 	}
310 	/* No error if bound_dev_if is 0, legal userspace check */
311 	dst->value = skb->sk->sk_bound_dev_if;
312 }
313 
314 META_COLLECTOR(var_sk_bound_if)
315 {
316 	if (skip_nonlocal(skb)) {
317 		*err = -1;
318 		return;
319 	}
320 
321 	if (skb->sk->sk_bound_dev_if == 0) {
322 		dst->value = (unsigned long) "any";
323 		dst->len = 3;
324 	} else {
325 		struct net_device *dev;
326 
327 		rcu_read_lock();
328 		dev = dev_get_by_index_rcu(sock_net(skb->sk),
329 					   skb->sk->sk_bound_dev_if);
330 		*err = var_dev(dev, dst);
331 		rcu_read_unlock();
332 	}
333 }
334 
335 META_COLLECTOR(int_sk_refcnt)
336 {
337 	if (skip_nonlocal(skb)) {
338 		*err = -1;
339 		return;
340 	}
341 	dst->value = atomic_read(&skb->sk->sk_refcnt);
342 }
343 
344 META_COLLECTOR(int_sk_rcvbuf)
345 {
346 	const struct sock *sk = skb_to_full_sk(skb);
347 
348 	if (!sk) {
349 		*err = -1;
350 		return;
351 	}
352 	dst->value = sk->sk_rcvbuf;
353 }
354 
355 META_COLLECTOR(int_sk_shutdown)
356 {
357 	const struct sock *sk = skb_to_full_sk(skb);
358 
359 	if (!sk) {
360 		*err = -1;
361 		return;
362 	}
363 	dst->value = sk->sk_shutdown;
364 }
365 
366 META_COLLECTOR(int_sk_proto)
367 {
368 	const struct sock *sk = skb_to_full_sk(skb);
369 
370 	if (!sk) {
371 		*err = -1;
372 		return;
373 	}
374 	dst->value = sk->sk_protocol;
375 }
376 
377 META_COLLECTOR(int_sk_type)
378 {
379 	const struct sock *sk = skb_to_full_sk(skb);
380 
381 	if (!sk) {
382 		*err = -1;
383 		return;
384 	}
385 	dst->value = sk->sk_type;
386 }
387 
388 META_COLLECTOR(int_sk_rmem_alloc)
389 {
390 	const struct sock *sk = skb_to_full_sk(skb);
391 
392 	if (!sk) {
393 		*err = -1;
394 		return;
395 	}
396 	dst->value = sk_rmem_alloc_get(sk);
397 }
398 
399 META_COLLECTOR(int_sk_wmem_alloc)
400 {
401 	const struct sock *sk = skb_to_full_sk(skb);
402 
403 	if (!sk) {
404 		*err = -1;
405 		return;
406 	}
407 	dst->value = sk_wmem_alloc_get(sk);
408 }
409 
410 META_COLLECTOR(int_sk_omem_alloc)
411 {
412 	const struct sock *sk = skb_to_full_sk(skb);
413 
414 	if (!sk) {
415 		*err = -1;
416 		return;
417 	}
418 	dst->value = atomic_read(&sk->sk_omem_alloc);
419 }
420 
421 META_COLLECTOR(int_sk_rcv_qlen)
422 {
423 	const struct sock *sk = skb_to_full_sk(skb);
424 
425 	if (!sk) {
426 		*err = -1;
427 		return;
428 	}
429 	dst->value = sk->sk_receive_queue.qlen;
430 }
431 
432 META_COLLECTOR(int_sk_snd_qlen)
433 {
434 	const struct sock *sk = skb_to_full_sk(skb);
435 
436 	if (!sk) {
437 		*err = -1;
438 		return;
439 	}
440 	dst->value = sk->sk_write_queue.qlen;
441 }
442 
443 META_COLLECTOR(int_sk_wmem_queued)
444 {
445 	const struct sock *sk = skb_to_full_sk(skb);
446 
447 	if (!sk) {
448 		*err = -1;
449 		return;
450 	}
451 	dst->value = sk->sk_wmem_queued;
452 }
453 
454 META_COLLECTOR(int_sk_fwd_alloc)
455 {
456 	const struct sock *sk = skb_to_full_sk(skb);
457 
458 	if (!sk) {
459 		*err = -1;
460 		return;
461 	}
462 	dst->value = sk->sk_forward_alloc;
463 }
464 
465 META_COLLECTOR(int_sk_sndbuf)
466 {
467 	const struct sock *sk = skb_to_full_sk(skb);
468 
469 	if (!sk) {
470 		*err = -1;
471 		return;
472 	}
473 	dst->value = sk->sk_sndbuf;
474 }
475 
476 META_COLLECTOR(int_sk_alloc)
477 {
478 	const struct sock *sk = skb_to_full_sk(skb);
479 
480 	if (!sk) {
481 		*err = -1;
482 		return;
483 	}
484 	dst->value = (__force int) sk->sk_allocation;
485 }
486 
487 META_COLLECTOR(int_sk_hash)
488 {
489 	if (skip_nonlocal(skb)) {
490 		*err = -1;
491 		return;
492 	}
493 	dst->value = skb->sk->sk_hash;
494 }
495 
496 META_COLLECTOR(int_sk_lingertime)
497 {
498 	const struct sock *sk = skb_to_full_sk(skb);
499 
500 	if (!sk) {
501 		*err = -1;
502 		return;
503 	}
504 	dst->value = sk->sk_lingertime / HZ;
505 }
506 
507 META_COLLECTOR(int_sk_err_qlen)
508 {
509 	const struct sock *sk = skb_to_full_sk(skb);
510 
511 	if (!sk) {
512 		*err = -1;
513 		return;
514 	}
515 	dst->value = sk->sk_error_queue.qlen;
516 }
517 
518 META_COLLECTOR(int_sk_ack_bl)
519 {
520 	const struct sock *sk = skb_to_full_sk(skb);
521 
522 	if (!sk) {
523 		*err = -1;
524 		return;
525 	}
526 	dst->value = sk->sk_ack_backlog;
527 }
528 
529 META_COLLECTOR(int_sk_max_ack_bl)
530 {
531 	const struct sock *sk = skb_to_full_sk(skb);
532 
533 	if (!sk) {
534 		*err = -1;
535 		return;
536 	}
537 	dst->value = sk->sk_max_ack_backlog;
538 }
539 
540 META_COLLECTOR(int_sk_prio)
541 {
542 	const struct sock *sk = skb_to_full_sk(skb);
543 
544 	if (!sk) {
545 		*err = -1;
546 		return;
547 	}
548 	dst->value = sk->sk_priority;
549 }
550 
551 META_COLLECTOR(int_sk_rcvlowat)
552 {
553 	const struct sock *sk = skb_to_full_sk(skb);
554 
555 	if (!sk) {
556 		*err = -1;
557 		return;
558 	}
559 	dst->value = sk->sk_rcvlowat;
560 }
561 
562 META_COLLECTOR(int_sk_rcvtimeo)
563 {
564 	const struct sock *sk = skb_to_full_sk(skb);
565 
566 	if (!sk) {
567 		*err = -1;
568 		return;
569 	}
570 	dst->value = sk->sk_rcvtimeo / HZ;
571 }
572 
573 META_COLLECTOR(int_sk_sndtimeo)
574 {
575 	const struct sock *sk = skb_to_full_sk(skb);
576 
577 	if (!sk) {
578 		*err = -1;
579 		return;
580 	}
581 	dst->value = sk->sk_sndtimeo / HZ;
582 }
583 
584 META_COLLECTOR(int_sk_sendmsg_off)
585 {
586 	const struct sock *sk = skb_to_full_sk(skb);
587 
588 	if (!sk) {
589 		*err = -1;
590 		return;
591 	}
592 	dst->value = sk->sk_frag.offset;
593 }
594 
595 META_COLLECTOR(int_sk_write_pend)
596 {
597 	const struct sock *sk = skb_to_full_sk(skb);
598 
599 	if (!sk) {
600 		*err = -1;
601 		return;
602 	}
603 	dst->value = sk->sk_write_pending;
604 }
605 
606 /**************************************************************************
607  * Meta value collectors assignment table
608  **************************************************************************/
609 
610 struct meta_ops {
611 	void		(*get)(struct sk_buff *, struct tcf_pkt_info *,
612 			       struct meta_value *, struct meta_obj *, int *);
613 };
614 
615 #define META_ID(name) TCF_META_ID_##name
616 #define META_FUNC(name) { .get = meta_##name }
617 
618 /* Meta value operations table listing all meta value collectors and
619  * assigns them to a type and meta id. */
620 static struct meta_ops __meta_ops[TCF_META_TYPE_MAX + 1][TCF_META_ID_MAX + 1] = {
621 	[TCF_META_TYPE_VAR] = {
622 		[META_ID(DEV)]			= META_FUNC(var_dev),
623 		[META_ID(SK_BOUND_IF)] 		= META_FUNC(var_sk_bound_if),
624 	},
625 	[TCF_META_TYPE_INT] = {
626 		[META_ID(RANDOM)]		= META_FUNC(int_random),
627 		[META_ID(LOADAVG_0)]		= META_FUNC(int_loadavg_0),
628 		[META_ID(LOADAVG_1)]		= META_FUNC(int_loadavg_1),
629 		[META_ID(LOADAVG_2)]		= META_FUNC(int_loadavg_2),
630 		[META_ID(DEV)]			= META_FUNC(int_dev),
631 		[META_ID(PRIORITY)]		= META_FUNC(int_priority),
632 		[META_ID(PROTOCOL)]		= META_FUNC(int_protocol),
633 		[META_ID(PKTTYPE)]		= META_FUNC(int_pkttype),
634 		[META_ID(PKTLEN)]		= META_FUNC(int_pktlen),
635 		[META_ID(DATALEN)]		= META_FUNC(int_datalen),
636 		[META_ID(MACLEN)]		= META_FUNC(int_maclen),
637 		[META_ID(NFMARK)]		= META_FUNC(int_mark),
638 		[META_ID(TCINDEX)]		= META_FUNC(int_tcindex),
639 		[META_ID(RTCLASSID)]		= META_FUNC(int_rtclassid),
640 		[META_ID(RTIIF)]		= META_FUNC(int_rtiif),
641 		[META_ID(SK_FAMILY)]		= META_FUNC(int_sk_family),
642 		[META_ID(SK_STATE)]		= META_FUNC(int_sk_state),
643 		[META_ID(SK_REUSE)]		= META_FUNC(int_sk_reuse),
644 		[META_ID(SK_BOUND_IF)]		= META_FUNC(int_sk_bound_if),
645 		[META_ID(SK_REFCNT)]		= META_FUNC(int_sk_refcnt),
646 		[META_ID(SK_RCVBUF)]		= META_FUNC(int_sk_rcvbuf),
647 		[META_ID(SK_SNDBUF)]		= META_FUNC(int_sk_sndbuf),
648 		[META_ID(SK_SHUTDOWN)]		= META_FUNC(int_sk_shutdown),
649 		[META_ID(SK_PROTO)]		= META_FUNC(int_sk_proto),
650 		[META_ID(SK_TYPE)]		= META_FUNC(int_sk_type),
651 		[META_ID(SK_RMEM_ALLOC)]	= META_FUNC(int_sk_rmem_alloc),
652 		[META_ID(SK_WMEM_ALLOC)]	= META_FUNC(int_sk_wmem_alloc),
653 		[META_ID(SK_OMEM_ALLOC)]	= META_FUNC(int_sk_omem_alloc),
654 		[META_ID(SK_WMEM_QUEUED)]	= META_FUNC(int_sk_wmem_queued),
655 		[META_ID(SK_RCV_QLEN)]		= META_FUNC(int_sk_rcv_qlen),
656 		[META_ID(SK_SND_QLEN)]		= META_FUNC(int_sk_snd_qlen),
657 		[META_ID(SK_ERR_QLEN)]		= META_FUNC(int_sk_err_qlen),
658 		[META_ID(SK_FORWARD_ALLOCS)]	= META_FUNC(int_sk_fwd_alloc),
659 		[META_ID(SK_ALLOCS)]		= META_FUNC(int_sk_alloc),
660 		[META_ID(SK_HASH)]		= META_FUNC(int_sk_hash),
661 		[META_ID(SK_LINGERTIME)]	= META_FUNC(int_sk_lingertime),
662 		[META_ID(SK_ACK_BACKLOG)]	= META_FUNC(int_sk_ack_bl),
663 		[META_ID(SK_MAX_ACK_BACKLOG)]	= META_FUNC(int_sk_max_ack_bl),
664 		[META_ID(SK_PRIO)]		= META_FUNC(int_sk_prio),
665 		[META_ID(SK_RCVLOWAT)]		= META_FUNC(int_sk_rcvlowat),
666 		[META_ID(SK_RCVTIMEO)]		= META_FUNC(int_sk_rcvtimeo),
667 		[META_ID(SK_SNDTIMEO)]		= META_FUNC(int_sk_sndtimeo),
668 		[META_ID(SK_SENDMSG_OFF)]	= META_FUNC(int_sk_sendmsg_off),
669 		[META_ID(SK_WRITE_PENDING)]	= META_FUNC(int_sk_write_pend),
670 		[META_ID(VLAN_TAG)]		= META_FUNC(int_vlan_tag),
671 		[META_ID(RXHASH)]		= META_FUNC(int_rxhash),
672 	}
673 };
674 
675 static inline struct meta_ops *meta_ops(struct meta_value *val)
676 {
677 	return &__meta_ops[meta_type(val)][meta_id(val)];
678 }
679 
680 /**************************************************************************
681  * Type specific operations for TCF_META_TYPE_VAR
682  **************************************************************************/
683 
684 static int meta_var_compare(struct meta_obj *a, struct meta_obj *b)
685 {
686 	int r = a->len - b->len;
687 
688 	if (r == 0)
689 		r = memcmp((void *) a->value, (void *) b->value, a->len);
690 
691 	return r;
692 }
693 
694 static int meta_var_change(struct meta_value *dst, struct nlattr *nla)
695 {
696 	int len = nla_len(nla);
697 
698 	dst->val = (unsigned long)kmemdup(nla_data(nla), len, GFP_KERNEL);
699 	if (dst->val == 0UL)
700 		return -ENOMEM;
701 	dst->len = len;
702 	return 0;
703 }
704 
705 static void meta_var_destroy(struct meta_value *v)
706 {
707 	kfree((void *) v->val);
708 }
709 
710 static void meta_var_apply_extras(struct meta_value *v,
711 				  struct meta_obj *dst)
712 {
713 	int shift = v->hdr.shift;
714 
715 	if (shift && shift < dst->len)
716 		dst->len -= shift;
717 }
718 
719 static int meta_var_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
720 {
721 	if (v->val && v->len &&
722 	    nla_put(skb, tlv, v->len, (void *) v->val))
723 		goto nla_put_failure;
724 	return 0;
725 
726 nla_put_failure:
727 	return -1;
728 }
729 
730 /**************************************************************************
731  * Type specific operations for TCF_META_TYPE_INT
732  **************************************************************************/
733 
734 static int meta_int_compare(struct meta_obj *a, struct meta_obj *b)
735 {
736 	/* Let gcc optimize it, the unlikely is not really based on
737 	 * some numbers but jump free code for mismatches seems
738 	 * more logical. */
739 	if (unlikely(a->value == b->value))
740 		return 0;
741 	else if (a->value < b->value)
742 		return -1;
743 	else
744 		return 1;
745 }
746 
747 static int meta_int_change(struct meta_value *dst, struct nlattr *nla)
748 {
749 	if (nla_len(nla) >= sizeof(unsigned long)) {
750 		dst->val = *(unsigned long *) nla_data(nla);
751 		dst->len = sizeof(unsigned long);
752 	} else if (nla_len(nla) == sizeof(u32)) {
753 		dst->val = nla_get_u32(nla);
754 		dst->len = sizeof(u32);
755 	} else
756 		return -EINVAL;
757 
758 	return 0;
759 }
760 
761 static void meta_int_apply_extras(struct meta_value *v,
762 				  struct meta_obj *dst)
763 {
764 	if (v->hdr.shift)
765 		dst->value >>= v->hdr.shift;
766 
767 	if (v->val)
768 		dst->value &= v->val;
769 }
770 
771 static int meta_int_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
772 {
773 	if (v->len == sizeof(unsigned long)) {
774 		if (nla_put(skb, tlv, sizeof(unsigned long), &v->val))
775 			goto nla_put_failure;
776 	} else if (v->len == sizeof(u32)) {
777 		if (nla_put_u32(skb, tlv, v->val))
778 			goto nla_put_failure;
779 	}
780 
781 	return 0;
782 
783 nla_put_failure:
784 	return -1;
785 }
786 
787 /**************************************************************************
788  * Type specific operations table
789  **************************************************************************/
790 
791 struct meta_type_ops {
792 	void	(*destroy)(struct meta_value *);
793 	int	(*compare)(struct meta_obj *, struct meta_obj *);
794 	int	(*change)(struct meta_value *, struct nlattr *);
795 	void	(*apply_extras)(struct meta_value *, struct meta_obj *);
796 	int	(*dump)(struct sk_buff *, struct meta_value *, int);
797 };
798 
799 static const struct meta_type_ops __meta_type_ops[TCF_META_TYPE_MAX + 1] = {
800 	[TCF_META_TYPE_VAR] = {
801 		.destroy = meta_var_destroy,
802 		.compare = meta_var_compare,
803 		.change = meta_var_change,
804 		.apply_extras = meta_var_apply_extras,
805 		.dump = meta_var_dump
806 	},
807 	[TCF_META_TYPE_INT] = {
808 		.compare = meta_int_compare,
809 		.change = meta_int_change,
810 		.apply_extras = meta_int_apply_extras,
811 		.dump = meta_int_dump
812 	}
813 };
814 
815 static inline const struct meta_type_ops *meta_type_ops(struct meta_value *v)
816 {
817 	return &__meta_type_ops[meta_type(v)];
818 }
819 
820 /**************************************************************************
821  * Core
822  **************************************************************************/
823 
824 static int meta_get(struct sk_buff *skb, struct tcf_pkt_info *info,
825 		    struct meta_value *v, struct meta_obj *dst)
826 {
827 	int err = 0;
828 
829 	if (meta_id(v) == TCF_META_ID_VALUE) {
830 		dst->value = v->val;
831 		dst->len = v->len;
832 		return 0;
833 	}
834 
835 	meta_ops(v)->get(skb, info, v, dst, &err);
836 	if (err < 0)
837 		return err;
838 
839 	if (meta_type_ops(v)->apply_extras)
840 		meta_type_ops(v)->apply_extras(v, dst);
841 
842 	return 0;
843 }
844 
845 static int em_meta_match(struct sk_buff *skb, struct tcf_ematch *m,
846 			 struct tcf_pkt_info *info)
847 {
848 	int r;
849 	struct meta_match *meta = (struct meta_match *) m->data;
850 	struct meta_obj l_value, r_value;
851 
852 	if (meta_get(skb, info, &meta->lvalue, &l_value) < 0 ||
853 	    meta_get(skb, info, &meta->rvalue, &r_value) < 0)
854 		return 0;
855 
856 	r = meta_type_ops(&meta->lvalue)->compare(&l_value, &r_value);
857 
858 	switch (meta->lvalue.hdr.op) {
859 	case TCF_EM_OPND_EQ:
860 		return !r;
861 	case TCF_EM_OPND_LT:
862 		return r < 0;
863 	case TCF_EM_OPND_GT:
864 		return r > 0;
865 	}
866 
867 	return 0;
868 }
869 
870 static void meta_delete(struct meta_match *meta)
871 {
872 	if (meta) {
873 		const struct meta_type_ops *ops = meta_type_ops(&meta->lvalue);
874 
875 		if (ops && ops->destroy) {
876 			ops->destroy(&meta->lvalue);
877 			ops->destroy(&meta->rvalue);
878 		}
879 	}
880 
881 	kfree(meta);
882 }
883 
884 static inline int meta_change_data(struct meta_value *dst, struct nlattr *nla)
885 {
886 	if (nla) {
887 		if (nla_len(nla) == 0)
888 			return -EINVAL;
889 
890 		return meta_type_ops(dst)->change(dst, nla);
891 	}
892 
893 	return 0;
894 }
895 
896 static inline int meta_is_supported(struct meta_value *val)
897 {
898 	return !meta_id(val) || meta_ops(val)->get;
899 }
900 
901 static const struct nla_policy meta_policy[TCA_EM_META_MAX + 1] = {
902 	[TCA_EM_META_HDR]	= { .len = sizeof(struct tcf_meta_hdr) },
903 };
904 
905 static int em_meta_change(struct net *net, void *data, int len,
906 			  struct tcf_ematch *m)
907 {
908 	int err;
909 	struct nlattr *tb[TCA_EM_META_MAX + 1];
910 	struct tcf_meta_hdr *hdr;
911 	struct meta_match *meta = NULL;
912 
913 	err = nla_parse(tb, TCA_EM_META_MAX, data, len, meta_policy);
914 	if (err < 0)
915 		goto errout;
916 
917 	err = -EINVAL;
918 	if (tb[TCA_EM_META_HDR] == NULL)
919 		goto errout;
920 	hdr = nla_data(tb[TCA_EM_META_HDR]);
921 
922 	if (TCF_META_TYPE(hdr->left.kind) != TCF_META_TYPE(hdr->right.kind) ||
923 	    TCF_META_TYPE(hdr->left.kind) > TCF_META_TYPE_MAX ||
924 	    TCF_META_ID(hdr->left.kind) > TCF_META_ID_MAX ||
925 	    TCF_META_ID(hdr->right.kind) > TCF_META_ID_MAX)
926 		goto errout;
927 
928 	meta = kzalloc(sizeof(*meta), GFP_KERNEL);
929 	if (meta == NULL) {
930 		err = -ENOMEM;
931 		goto errout;
932 	}
933 
934 	memcpy(&meta->lvalue.hdr, &hdr->left, sizeof(hdr->left));
935 	memcpy(&meta->rvalue.hdr, &hdr->right, sizeof(hdr->right));
936 
937 	if (!meta_is_supported(&meta->lvalue) ||
938 	    !meta_is_supported(&meta->rvalue)) {
939 		err = -EOPNOTSUPP;
940 		goto errout;
941 	}
942 
943 	if (meta_change_data(&meta->lvalue, tb[TCA_EM_META_LVALUE]) < 0 ||
944 	    meta_change_data(&meta->rvalue, tb[TCA_EM_META_RVALUE]) < 0)
945 		goto errout;
946 
947 	m->datalen = sizeof(*meta);
948 	m->data = (unsigned long) meta;
949 
950 	err = 0;
951 errout:
952 	if (err && meta)
953 		meta_delete(meta);
954 	return err;
955 }
956 
957 static void em_meta_destroy(struct tcf_ematch *m)
958 {
959 	if (m)
960 		meta_delete((struct meta_match *) m->data);
961 }
962 
963 static int em_meta_dump(struct sk_buff *skb, struct tcf_ematch *em)
964 {
965 	struct meta_match *meta = (struct meta_match *) em->data;
966 	struct tcf_meta_hdr hdr;
967 	const struct meta_type_ops *ops;
968 
969 	memset(&hdr, 0, sizeof(hdr));
970 	memcpy(&hdr.left, &meta->lvalue.hdr, sizeof(hdr.left));
971 	memcpy(&hdr.right, &meta->rvalue.hdr, sizeof(hdr.right));
972 
973 	if (nla_put(skb, TCA_EM_META_HDR, sizeof(hdr), &hdr))
974 		goto nla_put_failure;
975 
976 	ops = meta_type_ops(&meta->lvalue);
977 	if (ops->dump(skb, &meta->lvalue, TCA_EM_META_LVALUE) < 0 ||
978 	    ops->dump(skb, &meta->rvalue, TCA_EM_META_RVALUE) < 0)
979 		goto nla_put_failure;
980 
981 	return 0;
982 
983 nla_put_failure:
984 	return -1;
985 }
986 
987 static struct tcf_ematch_ops em_meta_ops = {
988 	.kind	  = TCF_EM_META,
989 	.change	  = em_meta_change,
990 	.match	  = em_meta_match,
991 	.destroy  = em_meta_destroy,
992 	.dump	  = em_meta_dump,
993 	.owner	  = THIS_MODULE,
994 	.link	  = LIST_HEAD_INIT(em_meta_ops.link)
995 };
996 
997 static int __init init_em_meta(void)
998 {
999 	return tcf_em_register(&em_meta_ops);
1000 }
1001 
1002 static void __exit exit_em_meta(void)
1003 {
1004 	tcf_em_unregister(&em_meta_ops);
1005 }
1006 
1007 MODULE_LICENSE("GPL");
1008 
1009 module_init(init_em_meta);
1010 module_exit(exit_em_meta);
1011 
1012 MODULE_ALIAS_TCF_EMATCH(TCF_EM_META);
1013