xref: /openbmc/linux/net/sched/em_meta.c (revision d78c317f) 
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 = vlan_tx_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 = skb->protocol;
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_rxhash(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 = skb_rtable(skb)->rt_iif;
268 }
269 
270 /**************************************************************************
271  * Socket Attributes
272  **************************************************************************/
273 
274 #define SKIP_NONLOCAL(skb)			\
275 	if (unlikely(skb->sk == NULL)) {	\
276 		*err = -1;			\
277 		return;				\
278 	}
279 
280 META_COLLECTOR(int_sk_family)
281 {
282 	SKIP_NONLOCAL(skb);
283 	dst->value = skb->sk->sk_family;
284 }
285 
286 META_COLLECTOR(int_sk_state)
287 {
288 	SKIP_NONLOCAL(skb);
289 	dst->value = skb->sk->sk_state;
290 }
291 
292 META_COLLECTOR(int_sk_reuse)
293 {
294 	SKIP_NONLOCAL(skb);
295 	dst->value = skb->sk->sk_reuse;
296 }
297 
298 META_COLLECTOR(int_sk_bound_if)
299 {
300 	SKIP_NONLOCAL(skb);
301 	/* No error if bound_dev_if is 0, legal userspace check */
302 	dst->value = skb->sk->sk_bound_dev_if;
303 }
304 
305 META_COLLECTOR(var_sk_bound_if)
306 {
307 	SKIP_NONLOCAL(skb);
308 
309 	if (skb->sk->sk_bound_dev_if == 0) {
310 		dst->value = (unsigned long) "any";
311 		dst->len = 3;
312 	} else {
313 		struct net_device *dev;
314 
315 		rcu_read_lock();
316 		dev = dev_get_by_index_rcu(sock_net(skb->sk),
317 					   skb->sk->sk_bound_dev_if);
318 		*err = var_dev(dev, dst);
319 		rcu_read_unlock();
320 	}
321 }
322 
323 META_COLLECTOR(int_sk_refcnt)
324 {
325 	SKIP_NONLOCAL(skb);
326 	dst->value = atomic_read(&skb->sk->sk_refcnt);
327 }
328 
329 META_COLLECTOR(int_sk_rcvbuf)
330 {
331 	SKIP_NONLOCAL(skb);
332 	dst->value = skb->sk->sk_rcvbuf;
333 }
334 
335 META_COLLECTOR(int_sk_shutdown)
336 {
337 	SKIP_NONLOCAL(skb);
338 	dst->value = skb->sk->sk_shutdown;
339 }
340 
341 META_COLLECTOR(int_sk_proto)
342 {
343 	SKIP_NONLOCAL(skb);
344 	dst->value = skb->sk->sk_protocol;
345 }
346 
347 META_COLLECTOR(int_sk_type)
348 {
349 	SKIP_NONLOCAL(skb);
350 	dst->value = skb->sk->sk_type;
351 }
352 
353 META_COLLECTOR(int_sk_rmem_alloc)
354 {
355 	SKIP_NONLOCAL(skb);
356 	dst->value = sk_rmem_alloc_get(skb->sk);
357 }
358 
359 META_COLLECTOR(int_sk_wmem_alloc)
360 {
361 	SKIP_NONLOCAL(skb);
362 	dst->value = sk_wmem_alloc_get(skb->sk);
363 }
364 
365 META_COLLECTOR(int_sk_omem_alloc)
366 {
367 	SKIP_NONLOCAL(skb);
368 	dst->value = atomic_read(&skb->sk->sk_omem_alloc);
369 }
370 
371 META_COLLECTOR(int_sk_rcv_qlen)
372 {
373 	SKIP_NONLOCAL(skb);
374 	dst->value = skb->sk->sk_receive_queue.qlen;
375 }
376 
377 META_COLLECTOR(int_sk_snd_qlen)
378 {
379 	SKIP_NONLOCAL(skb);
380 	dst->value = skb->sk->sk_write_queue.qlen;
381 }
382 
383 META_COLLECTOR(int_sk_wmem_queued)
384 {
385 	SKIP_NONLOCAL(skb);
386 	dst->value = skb->sk->sk_wmem_queued;
387 }
388 
389 META_COLLECTOR(int_sk_fwd_alloc)
390 {
391 	SKIP_NONLOCAL(skb);
392 	dst->value = skb->sk->sk_forward_alloc;
393 }
394 
395 META_COLLECTOR(int_sk_sndbuf)
396 {
397 	SKIP_NONLOCAL(skb);
398 	dst->value = skb->sk->sk_sndbuf;
399 }
400 
401 META_COLLECTOR(int_sk_alloc)
402 {
403 	SKIP_NONLOCAL(skb);
404 	dst->value = (__force int) skb->sk->sk_allocation;
405 }
406 
407 META_COLLECTOR(int_sk_hash)
408 {
409 	SKIP_NONLOCAL(skb);
410 	dst->value = skb->sk->sk_hash;
411 }
412 
413 META_COLLECTOR(int_sk_lingertime)
414 {
415 	SKIP_NONLOCAL(skb);
416 	dst->value = skb->sk->sk_lingertime / HZ;
417 }
418 
419 META_COLLECTOR(int_sk_err_qlen)
420 {
421 	SKIP_NONLOCAL(skb);
422 	dst->value = skb->sk->sk_error_queue.qlen;
423 }
424 
425 META_COLLECTOR(int_sk_ack_bl)
426 {
427 	SKIP_NONLOCAL(skb);
428 	dst->value = skb->sk->sk_ack_backlog;
429 }
430 
431 META_COLLECTOR(int_sk_max_ack_bl)
432 {
433 	SKIP_NONLOCAL(skb);
434 	dst->value = skb->sk->sk_max_ack_backlog;
435 }
436 
437 META_COLLECTOR(int_sk_prio)
438 {
439 	SKIP_NONLOCAL(skb);
440 	dst->value = skb->sk->sk_priority;
441 }
442 
443 META_COLLECTOR(int_sk_rcvlowat)
444 {
445 	SKIP_NONLOCAL(skb);
446 	dst->value = skb->sk->sk_rcvlowat;
447 }
448 
449 META_COLLECTOR(int_sk_rcvtimeo)
450 {
451 	SKIP_NONLOCAL(skb);
452 	dst->value = skb->sk->sk_rcvtimeo / HZ;
453 }
454 
455 META_COLLECTOR(int_sk_sndtimeo)
456 {
457 	SKIP_NONLOCAL(skb);
458 	dst->value = skb->sk->sk_sndtimeo / HZ;
459 }
460 
461 META_COLLECTOR(int_sk_sendmsg_off)
462 {
463 	SKIP_NONLOCAL(skb);
464 	dst->value = skb->sk->sk_sndmsg_off;
465 }
466 
467 META_COLLECTOR(int_sk_write_pend)
468 {
469 	SKIP_NONLOCAL(skb);
470 	dst->value = skb->sk->sk_write_pending;
471 }
472 
473 /**************************************************************************
474  * Meta value collectors assignment table
475  **************************************************************************/
476 
477 struct meta_ops {
478 	void		(*get)(struct sk_buff *, struct tcf_pkt_info *,
479 			       struct meta_value *, struct meta_obj *, int *);
480 };
481 
482 #define META_ID(name) TCF_META_ID_##name
483 #define META_FUNC(name) { .get = meta_##name }
484 
485 /* Meta value operations table listing all meta value collectors and
486  * assigns them to a type and meta id. */
487 static struct meta_ops __meta_ops[TCF_META_TYPE_MAX + 1][TCF_META_ID_MAX + 1] = {
488 	[TCF_META_TYPE_VAR] = {
489 		[META_ID(DEV)]			= META_FUNC(var_dev),
490 		[META_ID(SK_BOUND_IF)] 		= META_FUNC(var_sk_bound_if),
491 	},
492 	[TCF_META_TYPE_INT] = {
493 		[META_ID(RANDOM)]		= META_FUNC(int_random),
494 		[META_ID(LOADAVG_0)]		= META_FUNC(int_loadavg_0),
495 		[META_ID(LOADAVG_1)]		= META_FUNC(int_loadavg_1),
496 		[META_ID(LOADAVG_2)]		= META_FUNC(int_loadavg_2),
497 		[META_ID(DEV)]			= META_FUNC(int_dev),
498 		[META_ID(PRIORITY)]		= META_FUNC(int_priority),
499 		[META_ID(PROTOCOL)]		= META_FUNC(int_protocol),
500 		[META_ID(PKTTYPE)]		= META_FUNC(int_pkttype),
501 		[META_ID(PKTLEN)]		= META_FUNC(int_pktlen),
502 		[META_ID(DATALEN)]		= META_FUNC(int_datalen),
503 		[META_ID(MACLEN)]		= META_FUNC(int_maclen),
504 		[META_ID(NFMARK)]		= META_FUNC(int_mark),
505 		[META_ID(TCINDEX)]		= META_FUNC(int_tcindex),
506 		[META_ID(RTCLASSID)]		= META_FUNC(int_rtclassid),
507 		[META_ID(RTIIF)]		= META_FUNC(int_rtiif),
508 		[META_ID(SK_FAMILY)]		= META_FUNC(int_sk_family),
509 		[META_ID(SK_STATE)]		= META_FUNC(int_sk_state),
510 		[META_ID(SK_REUSE)]		= META_FUNC(int_sk_reuse),
511 		[META_ID(SK_BOUND_IF)]		= META_FUNC(int_sk_bound_if),
512 		[META_ID(SK_REFCNT)]		= META_FUNC(int_sk_refcnt),
513 		[META_ID(SK_RCVBUF)]		= META_FUNC(int_sk_rcvbuf),
514 		[META_ID(SK_SNDBUF)]		= META_FUNC(int_sk_sndbuf),
515 		[META_ID(SK_SHUTDOWN)]		= META_FUNC(int_sk_shutdown),
516 		[META_ID(SK_PROTO)]		= META_FUNC(int_sk_proto),
517 		[META_ID(SK_TYPE)]		= META_FUNC(int_sk_type),
518 		[META_ID(SK_RMEM_ALLOC)]	= META_FUNC(int_sk_rmem_alloc),
519 		[META_ID(SK_WMEM_ALLOC)]	= META_FUNC(int_sk_wmem_alloc),
520 		[META_ID(SK_OMEM_ALLOC)]	= META_FUNC(int_sk_omem_alloc),
521 		[META_ID(SK_WMEM_QUEUED)]	= META_FUNC(int_sk_wmem_queued),
522 		[META_ID(SK_RCV_QLEN)]		= META_FUNC(int_sk_rcv_qlen),
523 		[META_ID(SK_SND_QLEN)]		= META_FUNC(int_sk_snd_qlen),
524 		[META_ID(SK_ERR_QLEN)]		= META_FUNC(int_sk_err_qlen),
525 		[META_ID(SK_FORWARD_ALLOCS)]	= META_FUNC(int_sk_fwd_alloc),
526 		[META_ID(SK_ALLOCS)]		= META_FUNC(int_sk_alloc),
527 		[META_ID(SK_HASH)]		= META_FUNC(int_sk_hash),
528 		[META_ID(SK_LINGERTIME)]	= META_FUNC(int_sk_lingertime),
529 		[META_ID(SK_ACK_BACKLOG)]	= META_FUNC(int_sk_ack_bl),
530 		[META_ID(SK_MAX_ACK_BACKLOG)]	= META_FUNC(int_sk_max_ack_bl),
531 		[META_ID(SK_PRIO)]		= META_FUNC(int_sk_prio),
532 		[META_ID(SK_RCVLOWAT)]		= META_FUNC(int_sk_rcvlowat),
533 		[META_ID(SK_RCVTIMEO)]		= META_FUNC(int_sk_rcvtimeo),
534 		[META_ID(SK_SNDTIMEO)]		= META_FUNC(int_sk_sndtimeo),
535 		[META_ID(SK_SENDMSG_OFF)]	= META_FUNC(int_sk_sendmsg_off),
536 		[META_ID(SK_WRITE_PENDING)]	= META_FUNC(int_sk_write_pend),
537 		[META_ID(VLAN_TAG)]		= META_FUNC(int_vlan_tag),
538 		[META_ID(RXHASH)]		= META_FUNC(int_rxhash),
539 	}
540 };
541 
542 static inline struct meta_ops *meta_ops(struct meta_value *val)
543 {
544 	return &__meta_ops[meta_type(val)][meta_id(val)];
545 }
546 
547 /**************************************************************************
548  * Type specific operations for TCF_META_TYPE_VAR
549  **************************************************************************/
550 
551 static int meta_var_compare(struct meta_obj *a, struct meta_obj *b)
552 {
553 	int r = a->len - b->len;
554 
555 	if (r == 0)
556 		r = memcmp((void *) a->value, (void *) b->value, a->len);
557 
558 	return r;
559 }
560 
561 static int meta_var_change(struct meta_value *dst, struct nlattr *nla)
562 {
563 	int len = nla_len(nla);
564 
565 	dst->val = (unsigned long)kmemdup(nla_data(nla), len, GFP_KERNEL);
566 	if (dst->val == 0UL)
567 		return -ENOMEM;
568 	dst->len = len;
569 	return 0;
570 }
571 
572 static void meta_var_destroy(struct meta_value *v)
573 {
574 	kfree((void *) v->val);
575 }
576 
577 static void meta_var_apply_extras(struct meta_value *v,
578 				  struct meta_obj *dst)
579 {
580 	int shift = v->hdr.shift;
581 
582 	if (shift && shift < dst->len)
583 		dst->len -= shift;
584 }
585 
586 static int meta_var_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
587 {
588 	if (v->val && v->len)
589 		NLA_PUT(skb, tlv, v->len, (void *) v->val);
590 	return 0;
591 
592 nla_put_failure:
593 	return -1;
594 }
595 
596 /**************************************************************************
597  * Type specific operations for TCF_META_TYPE_INT
598  **************************************************************************/
599 
600 static int meta_int_compare(struct meta_obj *a, struct meta_obj *b)
601 {
602 	/* Let gcc optimize it, the unlikely is not really based on
603 	 * some numbers but jump free code for mismatches seems
604 	 * more logical. */
605 	if (unlikely(a->value == b->value))
606 		return 0;
607 	else if (a->value < b->value)
608 		return -1;
609 	else
610 		return 1;
611 }
612 
613 static int meta_int_change(struct meta_value *dst, struct nlattr *nla)
614 {
615 	if (nla_len(nla) >= sizeof(unsigned long)) {
616 		dst->val = *(unsigned long *) nla_data(nla);
617 		dst->len = sizeof(unsigned long);
618 	} else if (nla_len(nla) == sizeof(u32)) {
619 		dst->val = nla_get_u32(nla);
620 		dst->len = sizeof(u32);
621 	} else
622 		return -EINVAL;
623 
624 	return 0;
625 }
626 
627 static void meta_int_apply_extras(struct meta_value *v,
628 				  struct meta_obj *dst)
629 {
630 	if (v->hdr.shift)
631 		dst->value >>= v->hdr.shift;
632 
633 	if (v->val)
634 		dst->value &= v->val;
635 }
636 
637 static int meta_int_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
638 {
639 	if (v->len == sizeof(unsigned long))
640 		NLA_PUT(skb, tlv, sizeof(unsigned long), &v->val);
641 	else if (v->len == sizeof(u32))
642 		NLA_PUT_U32(skb, tlv, v->val);
643 
644 	return 0;
645 
646 nla_put_failure:
647 	return -1;
648 }
649 
650 /**************************************************************************
651  * Type specific operations table
652  **************************************************************************/
653 
654 struct meta_type_ops {
655 	void	(*destroy)(struct meta_value *);
656 	int	(*compare)(struct meta_obj *, struct meta_obj *);
657 	int	(*change)(struct meta_value *, struct nlattr *);
658 	void	(*apply_extras)(struct meta_value *, struct meta_obj *);
659 	int	(*dump)(struct sk_buff *, struct meta_value *, int);
660 };
661 
662 static struct meta_type_ops __meta_type_ops[TCF_META_TYPE_MAX + 1] = {
663 	[TCF_META_TYPE_VAR] = {
664 		.destroy = meta_var_destroy,
665 		.compare = meta_var_compare,
666 		.change = meta_var_change,
667 		.apply_extras = meta_var_apply_extras,
668 		.dump = meta_var_dump
669 	},
670 	[TCF_META_TYPE_INT] = {
671 		.compare = meta_int_compare,
672 		.change = meta_int_change,
673 		.apply_extras = meta_int_apply_extras,
674 		.dump = meta_int_dump
675 	}
676 };
677 
678 static inline struct meta_type_ops *meta_type_ops(struct meta_value *v)
679 {
680 	return &__meta_type_ops[meta_type(v)];
681 }
682 
683 /**************************************************************************
684  * Core
685  **************************************************************************/
686 
687 static int meta_get(struct sk_buff *skb, struct tcf_pkt_info *info,
688 		    struct meta_value *v, struct meta_obj *dst)
689 {
690 	int err = 0;
691 
692 	if (meta_id(v) == TCF_META_ID_VALUE) {
693 		dst->value = v->val;
694 		dst->len = v->len;
695 		return 0;
696 	}
697 
698 	meta_ops(v)->get(skb, info, v, dst, &err);
699 	if (err < 0)
700 		return err;
701 
702 	if (meta_type_ops(v)->apply_extras)
703 		meta_type_ops(v)->apply_extras(v, dst);
704 
705 	return 0;
706 }
707 
708 static int em_meta_match(struct sk_buff *skb, struct tcf_ematch *m,
709 			 struct tcf_pkt_info *info)
710 {
711 	int r;
712 	struct meta_match *meta = (struct meta_match *) m->data;
713 	struct meta_obj l_value, r_value;
714 
715 	if (meta_get(skb, info, &meta->lvalue, &l_value) < 0 ||
716 	    meta_get(skb, info, &meta->rvalue, &r_value) < 0)
717 		return 0;
718 
719 	r = meta_type_ops(&meta->lvalue)->compare(&l_value, &r_value);
720 
721 	switch (meta->lvalue.hdr.op) {
722 	case TCF_EM_OPND_EQ:
723 		return !r;
724 	case TCF_EM_OPND_LT:
725 		return r < 0;
726 	case TCF_EM_OPND_GT:
727 		return r > 0;
728 	}
729 
730 	return 0;
731 }
732 
733 static void meta_delete(struct meta_match *meta)
734 {
735 	if (meta) {
736 		struct meta_type_ops *ops = meta_type_ops(&meta->lvalue);
737 
738 		if (ops && ops->destroy) {
739 			ops->destroy(&meta->lvalue);
740 			ops->destroy(&meta->rvalue);
741 		}
742 	}
743 
744 	kfree(meta);
745 }
746 
747 static inline int meta_change_data(struct meta_value *dst, struct nlattr *nla)
748 {
749 	if (nla) {
750 		if (nla_len(nla) == 0)
751 			return -EINVAL;
752 
753 		return meta_type_ops(dst)->change(dst, nla);
754 	}
755 
756 	return 0;
757 }
758 
759 static inline int meta_is_supported(struct meta_value *val)
760 {
761 	return !meta_id(val) || meta_ops(val)->get;
762 }
763 
764 static const struct nla_policy meta_policy[TCA_EM_META_MAX + 1] = {
765 	[TCA_EM_META_HDR]	= { .len = sizeof(struct tcf_meta_hdr) },
766 };
767 
768 static int em_meta_change(struct tcf_proto *tp, void *data, int len,
769 			  struct tcf_ematch *m)
770 {
771 	int err;
772 	struct nlattr *tb[TCA_EM_META_MAX + 1];
773 	struct tcf_meta_hdr *hdr;
774 	struct meta_match *meta = NULL;
775 
776 	err = nla_parse(tb, TCA_EM_META_MAX, data, len, meta_policy);
777 	if (err < 0)
778 		goto errout;
779 
780 	err = -EINVAL;
781 	if (tb[TCA_EM_META_HDR] == NULL)
782 		goto errout;
783 	hdr = nla_data(tb[TCA_EM_META_HDR]);
784 
785 	if (TCF_META_TYPE(hdr->left.kind) != TCF_META_TYPE(hdr->right.kind) ||
786 	    TCF_META_TYPE(hdr->left.kind) > TCF_META_TYPE_MAX ||
787 	    TCF_META_ID(hdr->left.kind) > TCF_META_ID_MAX ||
788 	    TCF_META_ID(hdr->right.kind) > TCF_META_ID_MAX)
789 		goto errout;
790 
791 	meta = kzalloc(sizeof(*meta), GFP_KERNEL);
792 	if (meta == NULL)
793 		goto errout;
794 
795 	memcpy(&meta->lvalue.hdr, &hdr->left, sizeof(hdr->left));
796 	memcpy(&meta->rvalue.hdr, &hdr->right, sizeof(hdr->right));
797 
798 	if (!meta_is_supported(&meta->lvalue) ||
799 	    !meta_is_supported(&meta->rvalue)) {
800 		err = -EOPNOTSUPP;
801 		goto errout;
802 	}
803 
804 	if (meta_change_data(&meta->lvalue, tb[TCA_EM_META_LVALUE]) < 0 ||
805 	    meta_change_data(&meta->rvalue, tb[TCA_EM_META_RVALUE]) < 0)
806 		goto errout;
807 
808 	m->datalen = sizeof(*meta);
809 	m->data = (unsigned long) meta;
810 
811 	err = 0;
812 errout:
813 	if (err && meta)
814 		meta_delete(meta);
815 	return err;
816 }
817 
818 static void em_meta_destroy(struct tcf_proto *tp, struct tcf_ematch *m)
819 {
820 	if (m)
821 		meta_delete((struct meta_match *) m->data);
822 }
823 
824 static int em_meta_dump(struct sk_buff *skb, struct tcf_ematch *em)
825 {
826 	struct meta_match *meta = (struct meta_match *) em->data;
827 	struct tcf_meta_hdr hdr;
828 	struct meta_type_ops *ops;
829 
830 	memset(&hdr, 0, sizeof(hdr));
831 	memcpy(&hdr.left, &meta->lvalue.hdr, sizeof(hdr.left));
832 	memcpy(&hdr.right, &meta->rvalue.hdr, sizeof(hdr.right));
833 
834 	NLA_PUT(skb, TCA_EM_META_HDR, sizeof(hdr), &hdr);
835 
836 	ops = meta_type_ops(&meta->lvalue);
837 	if (ops->dump(skb, &meta->lvalue, TCA_EM_META_LVALUE) < 0 ||
838 	    ops->dump(skb, &meta->rvalue, TCA_EM_META_RVALUE) < 0)
839 		goto nla_put_failure;
840 
841 	return 0;
842 
843 nla_put_failure:
844 	return -1;
845 }
846 
847 static struct tcf_ematch_ops em_meta_ops = {
848 	.kind	  = TCF_EM_META,
849 	.change	  = em_meta_change,
850 	.match	  = em_meta_match,
851 	.destroy  = em_meta_destroy,
852 	.dump	  = em_meta_dump,
853 	.owner	  = THIS_MODULE,
854 	.link	  = LIST_HEAD_INIT(em_meta_ops.link)
855 };
856 
857 static int __init init_em_meta(void)
858 {
859 	return tcf_em_register(&em_meta_ops);
860 }
861 
862 static void __exit exit_em_meta(void)
863 {
864 	tcf_em_unregister(&em_meta_ops);
865 }
866 
867 MODULE_LICENSE("GPL");
868 
869 module_init(init_em_meta);
870 module_exit(exit_em_meta);
871 
872 MODULE_ALIAS_TCF_EMATCH(TCF_EM_META);
873