xref: /linux/net/sched/em_meta.c (revision b595076a180a56d1bb170e6eceda6eb9d76f4cd3)
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  * 	Additionaly, type dependant 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 {
78 	unsigned long		value;
79 	unsigned int		len;
80 };
81 
82 struct meta_value
83 {
84 	struct tcf_meta_val	hdr;
85 	unsigned long		val;
86 	unsigned int		len;
87 };
88 
89 struct meta_match
90 {
91 	struct meta_value	lvalue;
92 	struct meta_value	rvalue;
93 };
94 
95 static inline int meta_id(struct meta_value *v)
96 {
97 	return TCF_META_ID(v->hdr.kind);
98 }
99 
100 static inline int meta_type(struct meta_value *v)
101 {
102 	return TCF_META_TYPE(v->hdr.kind);
103 }
104 
105 #define META_COLLECTOR(FUNC) static void meta_##FUNC(struct sk_buff *skb, \
106 	struct tcf_pkt_info *info, struct meta_value *v, \
107 	struct meta_obj *dst, int *err)
108 
109 /**************************************************************************
110  * System status & misc
111  **************************************************************************/
112 
113 META_COLLECTOR(int_random)
114 {
115 	get_random_bytes(&dst->value, sizeof(dst->value));
116 }
117 
118 static inline unsigned long fixed_loadavg(int load)
119 {
120 	int rnd_load = load + (FIXED_1/200);
121 	int rnd_frac = ((rnd_load & (FIXED_1-1)) * 100) >> FSHIFT;
122 
123 	return ((rnd_load >> FSHIFT) * 100) + rnd_frac;
124 }
125 
126 META_COLLECTOR(int_loadavg_0)
127 {
128 	dst->value = fixed_loadavg(avenrun[0]);
129 }
130 
131 META_COLLECTOR(int_loadavg_1)
132 {
133 	dst->value = fixed_loadavg(avenrun[1]);
134 }
135 
136 META_COLLECTOR(int_loadavg_2)
137 {
138 	dst->value = fixed_loadavg(avenrun[2]);
139 }
140 
141 /**************************************************************************
142  * Device names & indices
143  **************************************************************************/
144 
145 static inline int int_dev(struct net_device *dev, struct meta_obj *dst)
146 {
147 	if (unlikely(dev == NULL))
148 		return -1;
149 
150 	dst->value = dev->ifindex;
151 	return 0;
152 }
153 
154 static inline int var_dev(struct net_device *dev, struct meta_obj *dst)
155 {
156 	if (unlikely(dev == NULL))
157 		return -1;
158 
159 	dst->value = (unsigned long) dev->name;
160 	dst->len = strlen(dev->name);
161 	return 0;
162 }
163 
164 META_COLLECTOR(int_dev)
165 {
166 	*err = int_dev(skb->dev, dst);
167 }
168 
169 META_COLLECTOR(var_dev)
170 {
171 	*err = var_dev(skb->dev, dst);
172 }
173 
174 /**************************************************************************
175  * vlan tag
176  **************************************************************************/
177 
178 META_COLLECTOR(int_vlan_tag)
179 {
180 	unsigned short tag;
181 
182 	tag = vlan_tx_tag_get(skb);
183 	if (!tag && __vlan_get_tag(skb, &tag))
184 		*err = -1;
185 	else
186 		dst->value = tag;
187 }
188 
189 
190 
191 /**************************************************************************
192  * skb attributes
193  **************************************************************************/
194 
195 META_COLLECTOR(int_priority)
196 {
197 	dst->value = skb->priority;
198 }
199 
200 META_COLLECTOR(int_protocol)
201 {
202 	/* Let userspace take care of the byte ordering */
203 	dst->value = skb->protocol;
204 }
205 
206 META_COLLECTOR(int_pkttype)
207 {
208 	dst->value = skb->pkt_type;
209 }
210 
211 META_COLLECTOR(int_pktlen)
212 {
213 	dst->value = skb->len;
214 }
215 
216 META_COLLECTOR(int_datalen)
217 {
218 	dst->value = skb->data_len;
219 }
220 
221 META_COLLECTOR(int_maclen)
222 {
223 	dst->value = skb->mac_len;
224 }
225 
226 META_COLLECTOR(int_rxhash)
227 {
228 	dst->value = skb_get_rxhash(skb);
229 }
230 
231 /**************************************************************************
232  * Netfilter
233  **************************************************************************/
234 
235 META_COLLECTOR(int_mark)
236 {
237 	dst->value = skb->mark;
238 }
239 
240 /**************************************************************************
241  * Traffic Control
242  **************************************************************************/
243 
244 META_COLLECTOR(int_tcindex)
245 {
246 	dst->value = skb->tc_index;
247 }
248 
249 /**************************************************************************
250  * Routing
251  **************************************************************************/
252 
253 META_COLLECTOR(int_rtclassid)
254 {
255 	if (unlikely(skb_dst(skb) == NULL))
256 		*err = -1;
257 	else
258 #ifdef CONFIG_NET_CLS_ROUTE
259 		dst->value = skb_dst(skb)->tclassid;
260 #else
261 		dst->value = 0;
262 #endif
263 }
264 
265 META_COLLECTOR(int_rtiif)
266 {
267 	if (unlikely(skb_rtable(skb) == NULL))
268 		*err = -1;
269 	else
270 		dst->value = skb_rtable(skb)->fl.iif;
271 }
272 
273 /**************************************************************************
274  * Socket Attributes
275  **************************************************************************/
276 
277 #define SKIP_NONLOCAL(skb)			\
278 	if (unlikely(skb->sk == NULL)) {	\
279 		*err = -1;			\
280 		return;				\
281 	}
282 
283 META_COLLECTOR(int_sk_family)
284 {
285 	SKIP_NONLOCAL(skb);
286 	dst->value = skb->sk->sk_family;
287 }
288 
289 META_COLLECTOR(int_sk_state)
290 {
291 	SKIP_NONLOCAL(skb);
292 	dst->value = skb->sk->sk_state;
293 }
294 
295 META_COLLECTOR(int_sk_reuse)
296 {
297 	SKIP_NONLOCAL(skb);
298 	dst->value = skb->sk->sk_reuse;
299 }
300 
301 META_COLLECTOR(int_sk_bound_if)
302 {
303 	SKIP_NONLOCAL(skb);
304 	/* No error if bound_dev_if is 0, legal userspace check */
305 	dst->value = skb->sk->sk_bound_dev_if;
306 }
307 
308 META_COLLECTOR(var_sk_bound_if)
309 {
310 	SKIP_NONLOCAL(skb);
311 
312 	if (skb->sk->sk_bound_dev_if == 0) {
313 		dst->value = (unsigned long) "any";
314 		dst->len = 3;
315 	} else {
316 		struct net_device *dev;
317 
318 		rcu_read_lock();
319 		dev = dev_get_by_index_rcu(sock_net(skb->sk),
320 					   skb->sk->sk_bound_dev_if);
321 		*err = var_dev(dev, dst);
322 		rcu_read_unlock();
323 	}
324 }
325 
326 META_COLLECTOR(int_sk_refcnt)
327 {
328 	SKIP_NONLOCAL(skb);
329 	dst->value = atomic_read(&skb->sk->sk_refcnt);
330 }
331 
332 META_COLLECTOR(int_sk_rcvbuf)
333 {
334 	SKIP_NONLOCAL(skb);
335 	dst->value = skb->sk->sk_rcvbuf;
336 }
337 
338 META_COLLECTOR(int_sk_shutdown)
339 {
340 	SKIP_NONLOCAL(skb);
341 	dst->value = skb->sk->sk_shutdown;
342 }
343 
344 META_COLLECTOR(int_sk_proto)
345 {
346 	SKIP_NONLOCAL(skb);
347 	dst->value = skb->sk->sk_protocol;
348 }
349 
350 META_COLLECTOR(int_sk_type)
351 {
352 	SKIP_NONLOCAL(skb);
353 	dst->value = skb->sk->sk_type;
354 }
355 
356 META_COLLECTOR(int_sk_rmem_alloc)
357 {
358 	SKIP_NONLOCAL(skb);
359 	dst->value = sk_rmem_alloc_get(skb->sk);
360 }
361 
362 META_COLLECTOR(int_sk_wmem_alloc)
363 {
364 	SKIP_NONLOCAL(skb);
365 	dst->value = sk_wmem_alloc_get(skb->sk);
366 }
367 
368 META_COLLECTOR(int_sk_omem_alloc)
369 {
370 	SKIP_NONLOCAL(skb);
371 	dst->value = atomic_read(&skb->sk->sk_omem_alloc);
372 }
373 
374 META_COLLECTOR(int_sk_rcv_qlen)
375 {
376 	SKIP_NONLOCAL(skb);
377 	dst->value = skb->sk->sk_receive_queue.qlen;
378 }
379 
380 META_COLLECTOR(int_sk_snd_qlen)
381 {
382 	SKIP_NONLOCAL(skb);
383 	dst->value = skb->sk->sk_write_queue.qlen;
384 }
385 
386 META_COLLECTOR(int_sk_wmem_queued)
387 {
388 	SKIP_NONLOCAL(skb);
389 	dst->value = skb->sk->sk_wmem_queued;
390 }
391 
392 META_COLLECTOR(int_sk_fwd_alloc)
393 {
394 	SKIP_NONLOCAL(skb);
395 	dst->value = skb->sk->sk_forward_alloc;
396 }
397 
398 META_COLLECTOR(int_sk_sndbuf)
399 {
400 	SKIP_NONLOCAL(skb);
401 	dst->value = skb->sk->sk_sndbuf;
402 }
403 
404 META_COLLECTOR(int_sk_alloc)
405 {
406 	SKIP_NONLOCAL(skb);
407 	dst->value = skb->sk->sk_allocation;
408 }
409 
410 META_COLLECTOR(int_sk_route_caps)
411 {
412 	SKIP_NONLOCAL(skb);
413 	dst->value = skb->sk->sk_route_caps;
414 }
415 
416 META_COLLECTOR(int_sk_hash)
417 {
418 	SKIP_NONLOCAL(skb);
419 	dst->value = skb->sk->sk_hash;
420 }
421 
422 META_COLLECTOR(int_sk_lingertime)
423 {
424 	SKIP_NONLOCAL(skb);
425 	dst->value = skb->sk->sk_lingertime / HZ;
426 }
427 
428 META_COLLECTOR(int_sk_err_qlen)
429 {
430 	SKIP_NONLOCAL(skb);
431 	dst->value = skb->sk->sk_error_queue.qlen;
432 }
433 
434 META_COLLECTOR(int_sk_ack_bl)
435 {
436 	SKIP_NONLOCAL(skb);
437 	dst->value = skb->sk->sk_ack_backlog;
438 }
439 
440 META_COLLECTOR(int_sk_max_ack_bl)
441 {
442 	SKIP_NONLOCAL(skb);
443 	dst->value = skb->sk->sk_max_ack_backlog;
444 }
445 
446 META_COLLECTOR(int_sk_prio)
447 {
448 	SKIP_NONLOCAL(skb);
449 	dst->value = skb->sk->sk_priority;
450 }
451 
452 META_COLLECTOR(int_sk_rcvlowat)
453 {
454 	SKIP_NONLOCAL(skb);
455 	dst->value = skb->sk->sk_rcvlowat;
456 }
457 
458 META_COLLECTOR(int_sk_rcvtimeo)
459 {
460 	SKIP_NONLOCAL(skb);
461 	dst->value = skb->sk->sk_rcvtimeo / HZ;
462 }
463 
464 META_COLLECTOR(int_sk_sndtimeo)
465 {
466 	SKIP_NONLOCAL(skb);
467 	dst->value = skb->sk->sk_sndtimeo / HZ;
468 }
469 
470 META_COLLECTOR(int_sk_sendmsg_off)
471 {
472 	SKIP_NONLOCAL(skb);
473 	dst->value = skb->sk->sk_sndmsg_off;
474 }
475 
476 META_COLLECTOR(int_sk_write_pend)
477 {
478 	SKIP_NONLOCAL(skb);
479 	dst->value = skb->sk->sk_write_pending;
480 }
481 
482 /**************************************************************************
483  * Meta value collectors assignment table
484  **************************************************************************/
485 
486 struct meta_ops
487 {
488 	void		(*get)(struct sk_buff *, struct tcf_pkt_info *,
489 			       struct meta_value *, struct meta_obj *, int *);
490 };
491 
492 #define META_ID(name) TCF_META_ID_##name
493 #define META_FUNC(name) { .get = meta_##name }
494 
495 /* Meta value operations table listing all meta value collectors and
496  * assigns them to a type and meta id. */
497 static struct meta_ops __meta_ops[TCF_META_TYPE_MAX+1][TCF_META_ID_MAX+1] = {
498 	[TCF_META_TYPE_VAR] = {
499 		[META_ID(DEV)]			= META_FUNC(var_dev),
500 		[META_ID(SK_BOUND_IF)] 		= META_FUNC(var_sk_bound_if),
501 	},
502 	[TCF_META_TYPE_INT] = {
503 		[META_ID(RANDOM)]		= META_FUNC(int_random),
504 		[META_ID(LOADAVG_0)]		= META_FUNC(int_loadavg_0),
505 		[META_ID(LOADAVG_1)]		= META_FUNC(int_loadavg_1),
506 		[META_ID(LOADAVG_2)]		= META_FUNC(int_loadavg_2),
507 		[META_ID(DEV)]			= META_FUNC(int_dev),
508 		[META_ID(PRIORITY)]		= META_FUNC(int_priority),
509 		[META_ID(PROTOCOL)]		= META_FUNC(int_protocol),
510 		[META_ID(PKTTYPE)]		= META_FUNC(int_pkttype),
511 		[META_ID(PKTLEN)]		= META_FUNC(int_pktlen),
512 		[META_ID(DATALEN)]		= META_FUNC(int_datalen),
513 		[META_ID(MACLEN)]		= META_FUNC(int_maclen),
514 		[META_ID(NFMARK)]		= META_FUNC(int_mark),
515 		[META_ID(TCINDEX)]		= META_FUNC(int_tcindex),
516 		[META_ID(RTCLASSID)]		= META_FUNC(int_rtclassid),
517 		[META_ID(RTIIF)]		= META_FUNC(int_rtiif),
518 		[META_ID(SK_FAMILY)]		= META_FUNC(int_sk_family),
519 		[META_ID(SK_STATE)]		= META_FUNC(int_sk_state),
520 		[META_ID(SK_REUSE)]		= META_FUNC(int_sk_reuse),
521 		[META_ID(SK_BOUND_IF)]		= META_FUNC(int_sk_bound_if),
522 		[META_ID(SK_REFCNT)]		= META_FUNC(int_sk_refcnt),
523 		[META_ID(SK_RCVBUF)]		= META_FUNC(int_sk_rcvbuf),
524 		[META_ID(SK_SNDBUF)]		= META_FUNC(int_sk_sndbuf),
525 		[META_ID(SK_SHUTDOWN)]		= META_FUNC(int_sk_shutdown),
526 		[META_ID(SK_PROTO)]		= META_FUNC(int_sk_proto),
527 		[META_ID(SK_TYPE)]		= META_FUNC(int_sk_type),
528 		[META_ID(SK_RMEM_ALLOC)]	= META_FUNC(int_sk_rmem_alloc),
529 		[META_ID(SK_WMEM_ALLOC)]	= META_FUNC(int_sk_wmem_alloc),
530 		[META_ID(SK_OMEM_ALLOC)]	= META_FUNC(int_sk_omem_alloc),
531 		[META_ID(SK_WMEM_QUEUED)]	= META_FUNC(int_sk_wmem_queued),
532 		[META_ID(SK_RCV_QLEN)]		= META_FUNC(int_sk_rcv_qlen),
533 		[META_ID(SK_SND_QLEN)]		= META_FUNC(int_sk_snd_qlen),
534 		[META_ID(SK_ERR_QLEN)]		= META_FUNC(int_sk_err_qlen),
535 		[META_ID(SK_FORWARD_ALLOCS)]	= META_FUNC(int_sk_fwd_alloc),
536 		[META_ID(SK_ALLOCS)]		= META_FUNC(int_sk_alloc),
537 		[META_ID(SK_ROUTE_CAPS)]	= META_FUNC(int_sk_route_caps),
538 		[META_ID(SK_HASH)]		= META_FUNC(int_sk_hash),
539 		[META_ID(SK_LINGERTIME)]	= META_FUNC(int_sk_lingertime),
540 		[META_ID(SK_ACK_BACKLOG)]	= META_FUNC(int_sk_ack_bl),
541 		[META_ID(SK_MAX_ACK_BACKLOG)]	= META_FUNC(int_sk_max_ack_bl),
542 		[META_ID(SK_PRIO)]		= META_FUNC(int_sk_prio),
543 		[META_ID(SK_RCVLOWAT)]		= META_FUNC(int_sk_rcvlowat),
544 		[META_ID(SK_RCVTIMEO)]		= META_FUNC(int_sk_rcvtimeo),
545 		[META_ID(SK_SNDTIMEO)]		= META_FUNC(int_sk_sndtimeo),
546 		[META_ID(SK_SENDMSG_OFF)]	= META_FUNC(int_sk_sendmsg_off),
547 		[META_ID(SK_WRITE_PENDING)]	= META_FUNC(int_sk_write_pend),
548 		[META_ID(VLAN_TAG)]		= META_FUNC(int_vlan_tag),
549 		[META_ID(RXHASH)]		= META_FUNC(int_rxhash),
550 	}
551 };
552 
553 static inline struct meta_ops * meta_ops(struct meta_value *val)
554 {
555 	return &__meta_ops[meta_type(val)][meta_id(val)];
556 }
557 
558 /**************************************************************************
559  * Type specific operations for TCF_META_TYPE_VAR
560  **************************************************************************/
561 
562 static int meta_var_compare(struct meta_obj *a, struct meta_obj *b)
563 {
564 	int r = a->len - b->len;
565 
566 	if (r == 0)
567 		r = memcmp((void *) a->value, (void *) b->value, a->len);
568 
569 	return r;
570 }
571 
572 static int meta_var_change(struct meta_value *dst, struct nlattr *nla)
573 {
574 	int len = nla_len(nla);
575 
576 	dst->val = (unsigned long)kmemdup(nla_data(nla), len, GFP_KERNEL);
577 	if (dst->val == 0UL)
578 		return -ENOMEM;
579 	dst->len = len;
580 	return 0;
581 }
582 
583 static void meta_var_destroy(struct meta_value *v)
584 {
585 	kfree((void *) v->val);
586 }
587 
588 static void meta_var_apply_extras(struct meta_value *v,
589 				  struct meta_obj *dst)
590 {
591 	int shift = v->hdr.shift;
592 
593 	if (shift && shift < dst->len)
594 		dst->len -= shift;
595 }
596 
597 static int meta_var_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
598 {
599 	if (v->val && v->len)
600 		NLA_PUT(skb, tlv, v->len, (void *) v->val);
601 	return 0;
602 
603 nla_put_failure:
604 	return -1;
605 }
606 
607 /**************************************************************************
608  * Type specific operations for TCF_META_TYPE_INT
609  **************************************************************************/
610 
611 static int meta_int_compare(struct meta_obj *a, struct meta_obj *b)
612 {
613 	/* Let gcc optimize it, the unlikely is not really based on
614 	 * some numbers but jump free code for mismatches seems
615 	 * more logical. */
616 	if (unlikely(a->value == b->value))
617 		return 0;
618 	else if (a->value < b->value)
619 		return -1;
620 	else
621 		return 1;
622 }
623 
624 static int meta_int_change(struct meta_value *dst, struct nlattr *nla)
625 {
626 	if (nla_len(nla) >= sizeof(unsigned long)) {
627 		dst->val = *(unsigned long *) nla_data(nla);
628 		dst->len = sizeof(unsigned long);
629 	} else if (nla_len(nla) == sizeof(u32)) {
630 		dst->val = nla_get_u32(nla);
631 		dst->len = sizeof(u32);
632 	} else
633 		return -EINVAL;
634 
635 	return 0;
636 }
637 
638 static void meta_int_apply_extras(struct meta_value *v,
639 				  struct meta_obj *dst)
640 {
641 	if (v->hdr.shift)
642 		dst->value >>= v->hdr.shift;
643 
644 	if (v->val)
645 		dst->value &= v->val;
646 }
647 
648 static int meta_int_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
649 {
650 	if (v->len == sizeof(unsigned long))
651 		NLA_PUT(skb, tlv, sizeof(unsigned long), &v->val);
652 	else if (v->len == sizeof(u32)) {
653 		NLA_PUT_U32(skb, tlv, v->val);
654 	}
655 
656 	return 0;
657 
658 nla_put_failure:
659 	return -1;
660 }
661 
662 /**************************************************************************
663  * Type specific operations table
664  **************************************************************************/
665 
666 struct meta_type_ops
667 {
668 	void	(*destroy)(struct meta_value *);
669 	int	(*compare)(struct meta_obj *, struct meta_obj *);
670 	int	(*change)(struct meta_value *, struct nlattr *);
671 	void	(*apply_extras)(struct meta_value *, struct meta_obj *);
672 	int	(*dump)(struct sk_buff *, struct meta_value *, int);
673 };
674 
675 static struct meta_type_ops __meta_type_ops[TCF_META_TYPE_MAX+1] = {
676 	[TCF_META_TYPE_VAR] = {
677 		.destroy = meta_var_destroy,
678 		.compare = meta_var_compare,
679 		.change = meta_var_change,
680 		.apply_extras = meta_var_apply_extras,
681 		.dump = meta_var_dump
682 	},
683 	[TCF_META_TYPE_INT] = {
684 		.compare = meta_int_compare,
685 		.change = meta_int_change,
686 		.apply_extras = meta_int_apply_extras,
687 		.dump = meta_int_dump
688 	}
689 };
690 
691 static inline struct meta_type_ops * meta_type_ops(struct meta_value *v)
692 {
693 	return &__meta_type_ops[meta_type(v)];
694 }
695 
696 /**************************************************************************
697  * Core
698  **************************************************************************/
699 
700 static int meta_get(struct sk_buff *skb, struct tcf_pkt_info *info,
701 		    struct meta_value *v, struct meta_obj *dst)
702 {
703 	int err = 0;
704 
705 	if (meta_id(v) == TCF_META_ID_VALUE) {
706 		dst->value = v->val;
707 		dst->len = v->len;
708 		return 0;
709 	}
710 
711 	meta_ops(v)->get(skb, info, v, dst, &err);
712 	if (err < 0)
713 		return err;
714 
715 	if (meta_type_ops(v)->apply_extras)
716 	    meta_type_ops(v)->apply_extras(v, dst);
717 
718 	return 0;
719 }
720 
721 static int em_meta_match(struct sk_buff *skb, struct tcf_ematch *m,
722 			 struct tcf_pkt_info *info)
723 {
724 	int r;
725 	struct meta_match *meta = (struct meta_match *) m->data;
726 	struct meta_obj l_value, r_value;
727 
728 	if (meta_get(skb, info, &meta->lvalue, &l_value) < 0 ||
729 	    meta_get(skb, info, &meta->rvalue, &r_value) < 0)
730 		return 0;
731 
732 	r = meta_type_ops(&meta->lvalue)->compare(&l_value, &r_value);
733 
734 	switch (meta->lvalue.hdr.op) {
735 		case TCF_EM_OPND_EQ:
736 			return !r;
737 		case TCF_EM_OPND_LT:
738 			return r < 0;
739 		case TCF_EM_OPND_GT:
740 			return r > 0;
741 	}
742 
743 	return 0;
744 }
745 
746 static void meta_delete(struct meta_match *meta)
747 {
748 	if (meta) {
749 		struct meta_type_ops *ops = meta_type_ops(&meta->lvalue);
750 
751 		if (ops && ops->destroy) {
752 			ops->destroy(&meta->lvalue);
753 			ops->destroy(&meta->rvalue);
754 		}
755 	}
756 
757 	kfree(meta);
758 }
759 
760 static inline int meta_change_data(struct meta_value *dst, struct nlattr *nla)
761 {
762 	if (nla) {
763 		if (nla_len(nla) == 0)
764 			return -EINVAL;
765 
766 		return meta_type_ops(dst)->change(dst, nla);
767 	}
768 
769 	return 0;
770 }
771 
772 static inline int meta_is_supported(struct meta_value *val)
773 {
774 	return (!meta_id(val) || meta_ops(val)->get);
775 }
776 
777 static const struct nla_policy meta_policy[TCA_EM_META_MAX + 1] = {
778 	[TCA_EM_META_HDR]	= { .len = sizeof(struct tcf_meta_hdr) },
779 };
780 
781 static int em_meta_change(struct tcf_proto *tp, void *data, int len,
782 			  struct tcf_ematch *m)
783 {
784 	int err;
785 	struct nlattr *tb[TCA_EM_META_MAX + 1];
786 	struct tcf_meta_hdr *hdr;
787 	struct meta_match *meta = NULL;
788 
789 	err = nla_parse(tb, TCA_EM_META_MAX, data, len, meta_policy);
790 	if (err < 0)
791 		goto errout;
792 
793 	err = -EINVAL;
794 	if (tb[TCA_EM_META_HDR] == NULL)
795 		goto errout;
796 	hdr = nla_data(tb[TCA_EM_META_HDR]);
797 
798 	if (TCF_META_TYPE(hdr->left.kind) != TCF_META_TYPE(hdr->right.kind) ||
799 	    TCF_META_TYPE(hdr->left.kind) > TCF_META_TYPE_MAX ||
800 	    TCF_META_ID(hdr->left.kind) > TCF_META_ID_MAX ||
801 	    TCF_META_ID(hdr->right.kind) > TCF_META_ID_MAX)
802 		goto errout;
803 
804 	meta = kzalloc(sizeof(*meta), GFP_KERNEL);
805 	if (meta == NULL)
806 		goto errout;
807 
808 	memcpy(&meta->lvalue.hdr, &hdr->left, sizeof(hdr->left));
809 	memcpy(&meta->rvalue.hdr, &hdr->right, sizeof(hdr->right));
810 
811 	if (!meta_is_supported(&meta->lvalue) ||
812 	    !meta_is_supported(&meta->rvalue)) {
813 		err = -EOPNOTSUPP;
814 		goto errout;
815 	}
816 
817 	if (meta_change_data(&meta->lvalue, tb[TCA_EM_META_LVALUE]) < 0 ||
818 	    meta_change_data(&meta->rvalue, tb[TCA_EM_META_RVALUE]) < 0)
819 		goto errout;
820 
821 	m->datalen = sizeof(*meta);
822 	m->data = (unsigned long) meta;
823 
824 	err = 0;
825 errout:
826 	if (err && meta)
827 		meta_delete(meta);
828 	return err;
829 }
830 
831 static void em_meta_destroy(struct tcf_proto *tp, struct tcf_ematch *m)
832 {
833 	if (m)
834 		meta_delete((struct meta_match *) m->data);
835 }
836 
837 static int em_meta_dump(struct sk_buff *skb, struct tcf_ematch *em)
838 {
839 	struct meta_match *meta = (struct meta_match *) em->data;
840 	struct tcf_meta_hdr hdr;
841 	struct meta_type_ops *ops;
842 
843 	memset(&hdr, 0, sizeof(hdr));
844 	memcpy(&hdr.left, &meta->lvalue.hdr, sizeof(hdr.left));
845 	memcpy(&hdr.right, &meta->rvalue.hdr, sizeof(hdr.right));
846 
847 	NLA_PUT(skb, TCA_EM_META_HDR, sizeof(hdr), &hdr);
848 
849 	ops = meta_type_ops(&meta->lvalue);
850 	if (ops->dump(skb, &meta->lvalue, TCA_EM_META_LVALUE) < 0 ||
851 	    ops->dump(skb, &meta->rvalue, TCA_EM_META_RVALUE) < 0)
852 		goto nla_put_failure;
853 
854 	return 0;
855 
856 nla_put_failure:
857 	return -1;
858 }
859 
860 static struct tcf_ematch_ops em_meta_ops = {
861 	.kind	  = TCF_EM_META,
862 	.change	  = em_meta_change,
863 	.match	  = em_meta_match,
864 	.destroy  = em_meta_destroy,
865 	.dump	  = em_meta_dump,
866 	.owner	  = THIS_MODULE,
867 	.link	  = LIST_HEAD_INIT(em_meta_ops.link)
868 };
869 
870 static int __init init_em_meta(void)
871 {
872 	return tcf_em_register(&em_meta_ops);
873 }
874 
875 static void __exit exit_em_meta(void)
876 {
877 	tcf_em_unregister(&em_meta_ops);
878 }
879 
880 MODULE_LICENSE("GPL");
881 
882 module_init(init_em_meta);
883 module_exit(exit_em_meta);
884 
885 MODULE_ALIAS_TCF_EMATCH(TCF_EM_META);
886