xref: /openbmc/linux/net/sched/em_meta.c (revision 643d1f7f) 
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/module.h>
62 #include <linux/types.h>
63 #include <linux/kernel.h>
64 #include <linux/sched.h>
65 #include <linux/string.h>
66 #include <linux/skbuff.h>
67 #include <linux/random.h>
68 #include <linux/tc_ematch/tc_em_meta.h>
69 #include <net/dst.h>
70 #include <net/route.h>
71 #include <net/pkt_cls.h>
72 #include <net/sock.h>
73 
74 struct meta_obj
75 {
76 	unsigned long		value;
77 	unsigned int		len;
78 };
79 
80 struct meta_value
81 {
82 	struct tcf_meta_val	hdr;
83 	unsigned long		val;
84 	unsigned int		len;
85 };
86 
87 struct meta_match
88 {
89 	struct meta_value	lvalue;
90 	struct meta_value	rvalue;
91 };
92 
93 static inline int meta_id(struct meta_value *v)
94 {
95 	return TCF_META_ID(v->hdr.kind);
96 }
97 
98 static inline int meta_type(struct meta_value *v)
99 {
100 	return TCF_META_TYPE(v->hdr.kind);
101 }
102 
103 #define META_COLLECTOR(FUNC) static void meta_##FUNC(struct sk_buff *skb, \
104 	struct tcf_pkt_info *info, struct meta_value *v, \
105 	struct meta_obj *dst, int *err)
106 
107 /**************************************************************************
108  * System status & misc
109  **************************************************************************/
110 
111 META_COLLECTOR(int_random)
112 {
113 	get_random_bytes(&dst->value, sizeof(dst->value));
114 }
115 
116 static inline unsigned long fixed_loadavg(int load)
117 {
118 	int rnd_load = load + (FIXED_1/200);
119 	int rnd_frac = ((rnd_load & (FIXED_1-1)) * 100) >> FSHIFT;
120 
121 	return ((rnd_load >> FSHIFT) * 100) + rnd_frac;
122 }
123 
124 META_COLLECTOR(int_loadavg_0)
125 {
126 	dst->value = fixed_loadavg(avenrun[0]);
127 }
128 
129 META_COLLECTOR(int_loadavg_1)
130 {
131 	dst->value = fixed_loadavg(avenrun[1]);
132 }
133 
134 META_COLLECTOR(int_loadavg_2)
135 {
136 	dst->value = fixed_loadavg(avenrun[2]);
137 }
138 
139 /**************************************************************************
140  * Device names & indices
141  **************************************************************************/
142 
143 static inline int int_dev(struct net_device *dev, struct meta_obj *dst)
144 {
145 	if (unlikely(dev == NULL))
146 		return -1;
147 
148 	dst->value = dev->ifindex;
149 	return 0;
150 }
151 
152 static inline int var_dev(struct net_device *dev, struct meta_obj *dst)
153 {
154 	if (unlikely(dev == NULL))
155 		return -1;
156 
157 	dst->value = (unsigned long) dev->name;
158 	dst->len = strlen(dev->name);
159 	return 0;
160 }
161 
162 META_COLLECTOR(int_dev)
163 {
164 	*err = int_dev(skb->dev, dst);
165 }
166 
167 META_COLLECTOR(var_dev)
168 {
169 	*err = var_dev(skb->dev, dst);
170 }
171 
172 /**************************************************************************
173  * skb attributes
174  **************************************************************************/
175 
176 META_COLLECTOR(int_priority)
177 {
178 	dst->value = skb->priority;
179 }
180 
181 META_COLLECTOR(int_protocol)
182 {
183 	/* Let userspace take care of the byte ordering */
184 	dst->value = skb->protocol;
185 }
186 
187 META_COLLECTOR(int_pkttype)
188 {
189 	dst->value = skb->pkt_type;
190 }
191 
192 META_COLLECTOR(int_pktlen)
193 {
194 	dst->value = skb->len;
195 }
196 
197 META_COLLECTOR(int_datalen)
198 {
199 	dst->value = skb->data_len;
200 }
201 
202 META_COLLECTOR(int_maclen)
203 {
204 	dst->value = skb->mac_len;
205 }
206 
207 /**************************************************************************
208  * Netfilter
209  **************************************************************************/
210 
211 META_COLLECTOR(int_mark)
212 {
213 	dst->value = skb->mark;
214 }
215 
216 /**************************************************************************
217  * Traffic Control
218  **************************************************************************/
219 
220 META_COLLECTOR(int_tcindex)
221 {
222 	dst->value = skb->tc_index;
223 }
224 
225 /**************************************************************************
226  * Routing
227  **************************************************************************/
228 
229 META_COLLECTOR(int_rtclassid)
230 {
231 	if (unlikely(skb->dst == NULL))
232 		*err = -1;
233 	else
234 #ifdef CONFIG_NET_CLS_ROUTE
235 		dst->value = skb->dst->tclassid;
236 #else
237 		dst->value = 0;
238 #endif
239 }
240 
241 META_COLLECTOR(int_rtiif)
242 {
243 	if (unlikely(skb->dst == NULL))
244 		*err = -1;
245 	else
246 		dst->value = ((struct rtable*) skb->dst)->fl.iif;
247 }
248 
249 /**************************************************************************
250  * Socket Attributes
251  **************************************************************************/
252 
253 #define SKIP_NONLOCAL(skb)			\
254 	if (unlikely(skb->sk == NULL)) {	\
255 		*err = -1;			\
256 		return;				\
257 	}
258 
259 META_COLLECTOR(int_sk_family)
260 {
261 	SKIP_NONLOCAL(skb);
262 	dst->value = skb->sk->sk_family;
263 }
264 
265 META_COLLECTOR(int_sk_state)
266 {
267 	SKIP_NONLOCAL(skb);
268 	dst->value = skb->sk->sk_state;
269 }
270 
271 META_COLLECTOR(int_sk_reuse)
272 {
273 	SKIP_NONLOCAL(skb);
274 	dst->value = skb->sk->sk_reuse;
275 }
276 
277 META_COLLECTOR(int_sk_bound_if)
278 {
279 	SKIP_NONLOCAL(skb);
280 	/* No error if bound_dev_if is 0, legal userspace check */
281 	dst->value = skb->sk->sk_bound_dev_if;
282 }
283 
284 META_COLLECTOR(var_sk_bound_if)
285 {
286 	SKIP_NONLOCAL(skb);
287 
288 	 if (skb->sk->sk_bound_dev_if == 0) {
289 		dst->value = (unsigned long) "any";
290 		dst->len = 3;
291 	 } else  {
292 		struct net_device *dev;
293 
294 		dev = dev_get_by_index(&init_net, skb->sk->sk_bound_dev_if);
295 		*err = var_dev(dev, dst);
296 		if (dev)
297 			dev_put(dev);
298 	 }
299 }
300 
301 META_COLLECTOR(int_sk_refcnt)
302 {
303 	SKIP_NONLOCAL(skb);
304 	dst->value = atomic_read(&skb->sk->sk_refcnt);
305 }
306 
307 META_COLLECTOR(int_sk_rcvbuf)
308 {
309 	SKIP_NONLOCAL(skb);
310 	dst->value = skb->sk->sk_rcvbuf;
311 }
312 
313 META_COLLECTOR(int_sk_shutdown)
314 {
315 	SKIP_NONLOCAL(skb);
316 	dst->value = skb->sk->sk_shutdown;
317 }
318 
319 META_COLLECTOR(int_sk_proto)
320 {
321 	SKIP_NONLOCAL(skb);
322 	dst->value = skb->sk->sk_protocol;
323 }
324 
325 META_COLLECTOR(int_sk_type)
326 {
327 	SKIP_NONLOCAL(skb);
328 	dst->value = skb->sk->sk_type;
329 }
330 
331 META_COLLECTOR(int_sk_rmem_alloc)
332 {
333 	SKIP_NONLOCAL(skb);
334 	dst->value = atomic_read(&skb->sk->sk_rmem_alloc);
335 }
336 
337 META_COLLECTOR(int_sk_wmem_alloc)
338 {
339 	SKIP_NONLOCAL(skb);
340 	dst->value = atomic_read(&skb->sk->sk_wmem_alloc);
341 }
342 
343 META_COLLECTOR(int_sk_omem_alloc)
344 {
345 	SKIP_NONLOCAL(skb);
346 	dst->value = atomic_read(&skb->sk->sk_omem_alloc);
347 }
348 
349 META_COLLECTOR(int_sk_rcv_qlen)
350 {
351 	SKIP_NONLOCAL(skb);
352 	dst->value = skb->sk->sk_receive_queue.qlen;
353 }
354 
355 META_COLLECTOR(int_sk_snd_qlen)
356 {
357 	SKIP_NONLOCAL(skb);
358 	dst->value = skb->sk->sk_write_queue.qlen;
359 }
360 
361 META_COLLECTOR(int_sk_wmem_queued)
362 {
363 	SKIP_NONLOCAL(skb);
364 	dst->value = skb->sk->sk_wmem_queued;
365 }
366 
367 META_COLLECTOR(int_sk_fwd_alloc)
368 {
369 	SKIP_NONLOCAL(skb);
370 	dst->value = skb->sk->sk_forward_alloc;
371 }
372 
373 META_COLLECTOR(int_sk_sndbuf)
374 {
375 	SKIP_NONLOCAL(skb);
376 	dst->value = skb->sk->sk_sndbuf;
377 }
378 
379 META_COLLECTOR(int_sk_alloc)
380 {
381 	SKIP_NONLOCAL(skb);
382 	dst->value = skb->sk->sk_allocation;
383 }
384 
385 META_COLLECTOR(int_sk_route_caps)
386 {
387 	SKIP_NONLOCAL(skb);
388 	dst->value = skb->sk->sk_route_caps;
389 }
390 
391 META_COLLECTOR(int_sk_hash)
392 {
393 	SKIP_NONLOCAL(skb);
394 	dst->value = skb->sk->sk_hash;
395 }
396 
397 META_COLLECTOR(int_sk_lingertime)
398 {
399 	SKIP_NONLOCAL(skb);
400 	dst->value = skb->sk->sk_lingertime / HZ;
401 }
402 
403 META_COLLECTOR(int_sk_err_qlen)
404 {
405 	SKIP_NONLOCAL(skb);
406 	dst->value = skb->sk->sk_error_queue.qlen;
407 }
408 
409 META_COLLECTOR(int_sk_ack_bl)
410 {
411 	SKIP_NONLOCAL(skb);
412 	dst->value = skb->sk->sk_ack_backlog;
413 }
414 
415 META_COLLECTOR(int_sk_max_ack_bl)
416 {
417 	SKIP_NONLOCAL(skb);
418 	dst->value = skb->sk->sk_max_ack_backlog;
419 }
420 
421 META_COLLECTOR(int_sk_prio)
422 {
423 	SKIP_NONLOCAL(skb);
424 	dst->value = skb->sk->sk_priority;
425 }
426 
427 META_COLLECTOR(int_sk_rcvlowat)
428 {
429 	SKIP_NONLOCAL(skb);
430 	dst->value = skb->sk->sk_rcvlowat;
431 }
432 
433 META_COLLECTOR(int_sk_rcvtimeo)
434 {
435 	SKIP_NONLOCAL(skb);
436 	dst->value = skb->sk->sk_rcvtimeo / HZ;
437 }
438 
439 META_COLLECTOR(int_sk_sndtimeo)
440 {
441 	SKIP_NONLOCAL(skb);
442 	dst->value = skb->sk->sk_sndtimeo / HZ;
443 }
444 
445 META_COLLECTOR(int_sk_sendmsg_off)
446 {
447 	SKIP_NONLOCAL(skb);
448 	dst->value = skb->sk->sk_sndmsg_off;
449 }
450 
451 META_COLLECTOR(int_sk_write_pend)
452 {
453 	SKIP_NONLOCAL(skb);
454 	dst->value = skb->sk->sk_write_pending;
455 }
456 
457 /**************************************************************************
458  * Meta value collectors assignment table
459  **************************************************************************/
460 
461 struct meta_ops
462 {
463 	void		(*get)(struct sk_buff *, struct tcf_pkt_info *,
464 			       struct meta_value *, struct meta_obj *, int *);
465 };
466 
467 #define META_ID(name) TCF_META_ID_##name
468 #define META_FUNC(name) { .get = meta_##name }
469 
470 /* Meta value operations table listing all meta value collectors and
471  * assigns them to a type and meta id. */
472 static struct meta_ops __meta_ops[TCF_META_TYPE_MAX+1][TCF_META_ID_MAX+1] = {
473 	[TCF_META_TYPE_VAR] = {
474 		[META_ID(DEV)]			= META_FUNC(var_dev),
475 		[META_ID(SK_BOUND_IF)] 		= META_FUNC(var_sk_bound_if),
476 	},
477 	[TCF_META_TYPE_INT] = {
478 		[META_ID(RANDOM)]		= META_FUNC(int_random),
479 		[META_ID(LOADAVG_0)]		= META_FUNC(int_loadavg_0),
480 		[META_ID(LOADAVG_1)]		= META_FUNC(int_loadavg_1),
481 		[META_ID(LOADAVG_2)]		= META_FUNC(int_loadavg_2),
482 		[META_ID(DEV)]			= META_FUNC(int_dev),
483 		[META_ID(PRIORITY)]		= META_FUNC(int_priority),
484 		[META_ID(PROTOCOL)]		= META_FUNC(int_protocol),
485 		[META_ID(PKTTYPE)]		= META_FUNC(int_pkttype),
486 		[META_ID(PKTLEN)]		= META_FUNC(int_pktlen),
487 		[META_ID(DATALEN)]		= META_FUNC(int_datalen),
488 		[META_ID(MACLEN)]		= META_FUNC(int_maclen),
489 		[META_ID(NFMARK)]		= META_FUNC(int_mark),
490 		[META_ID(TCINDEX)]		= META_FUNC(int_tcindex),
491 		[META_ID(RTCLASSID)]		= META_FUNC(int_rtclassid),
492 		[META_ID(RTIIF)]		= META_FUNC(int_rtiif),
493 		[META_ID(SK_FAMILY)]		= META_FUNC(int_sk_family),
494 		[META_ID(SK_STATE)]		= META_FUNC(int_sk_state),
495 		[META_ID(SK_REUSE)]		= META_FUNC(int_sk_reuse),
496 		[META_ID(SK_BOUND_IF)]		= META_FUNC(int_sk_bound_if),
497 		[META_ID(SK_REFCNT)]		= META_FUNC(int_sk_refcnt),
498 		[META_ID(SK_RCVBUF)]		= META_FUNC(int_sk_rcvbuf),
499 		[META_ID(SK_SNDBUF)]		= META_FUNC(int_sk_sndbuf),
500 		[META_ID(SK_SHUTDOWN)]		= META_FUNC(int_sk_shutdown),
501 		[META_ID(SK_PROTO)]		= META_FUNC(int_sk_proto),
502 		[META_ID(SK_TYPE)]		= META_FUNC(int_sk_type),
503 		[META_ID(SK_RMEM_ALLOC)]	= META_FUNC(int_sk_rmem_alloc),
504 		[META_ID(SK_WMEM_ALLOC)]	= META_FUNC(int_sk_wmem_alloc),
505 		[META_ID(SK_OMEM_ALLOC)]	= META_FUNC(int_sk_omem_alloc),
506 		[META_ID(SK_WMEM_QUEUED)]	= META_FUNC(int_sk_wmem_queued),
507 		[META_ID(SK_RCV_QLEN)]		= META_FUNC(int_sk_rcv_qlen),
508 		[META_ID(SK_SND_QLEN)]		= META_FUNC(int_sk_snd_qlen),
509 		[META_ID(SK_ERR_QLEN)]		= META_FUNC(int_sk_err_qlen),
510 		[META_ID(SK_FORWARD_ALLOCS)]	= META_FUNC(int_sk_fwd_alloc),
511 		[META_ID(SK_ALLOCS)]		= META_FUNC(int_sk_alloc),
512 		[META_ID(SK_ROUTE_CAPS)]	= META_FUNC(int_sk_route_caps),
513 		[META_ID(SK_HASH)]		= META_FUNC(int_sk_hash),
514 		[META_ID(SK_LINGERTIME)]	= META_FUNC(int_sk_lingertime),
515 		[META_ID(SK_ACK_BACKLOG)]	= META_FUNC(int_sk_ack_bl),
516 		[META_ID(SK_MAX_ACK_BACKLOG)]	= META_FUNC(int_sk_max_ack_bl),
517 		[META_ID(SK_PRIO)]		= META_FUNC(int_sk_prio),
518 		[META_ID(SK_RCVLOWAT)]		= META_FUNC(int_sk_rcvlowat),
519 		[META_ID(SK_RCVTIMEO)]		= META_FUNC(int_sk_rcvtimeo),
520 		[META_ID(SK_SNDTIMEO)]		= META_FUNC(int_sk_sndtimeo),
521 		[META_ID(SK_SENDMSG_OFF)]	= META_FUNC(int_sk_sendmsg_off),
522 		[META_ID(SK_WRITE_PENDING)]	= META_FUNC(int_sk_write_pend),
523 	}
524 };
525 
526 static inline struct meta_ops * meta_ops(struct meta_value *val)
527 {
528 	return &__meta_ops[meta_type(val)][meta_id(val)];
529 }
530 
531 /**************************************************************************
532  * Type specific operations for TCF_META_TYPE_VAR
533  **************************************************************************/
534 
535 static int meta_var_compare(struct meta_obj *a, struct meta_obj *b)
536 {
537 	int r = a->len - b->len;
538 
539 	if (r == 0)
540 		r = memcmp((void *) a->value, (void *) b->value, a->len);
541 
542 	return r;
543 }
544 
545 static int meta_var_change(struct meta_value *dst, struct nlattr *nla)
546 {
547 	int len = nla_len(nla);
548 
549 	dst->val = (unsigned long)kmemdup(nla_data(nla), len, GFP_KERNEL);
550 	if (dst->val == 0UL)
551 		return -ENOMEM;
552 	dst->len = len;
553 	return 0;
554 }
555 
556 static void meta_var_destroy(struct meta_value *v)
557 {
558 	kfree((void *) v->val);
559 }
560 
561 static void meta_var_apply_extras(struct meta_value *v,
562 				  struct meta_obj *dst)
563 {
564 	int shift = v->hdr.shift;
565 
566 	if (shift && shift < dst->len)
567 		dst->len -= shift;
568 }
569 
570 static int meta_var_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
571 {
572 	if (v->val && v->len)
573 		NLA_PUT(skb, tlv, v->len, (void *) v->val);
574 	return 0;
575 
576 nla_put_failure:
577 	return -1;
578 }
579 
580 /**************************************************************************
581  * Type specific operations for TCF_META_TYPE_INT
582  **************************************************************************/
583 
584 static int meta_int_compare(struct meta_obj *a, struct meta_obj *b)
585 {
586 	/* Let gcc optimize it, the unlikely is not really based on
587 	 * some numbers but jump free code for mismatches seems
588 	 * more logical. */
589 	if (unlikely(a->value == b->value))
590 		return 0;
591 	else if (a->value < b->value)
592 		return -1;
593 	else
594 		return 1;
595 }
596 
597 static int meta_int_change(struct meta_value *dst, struct nlattr *nla)
598 {
599 	if (nla_len(nla) >= sizeof(unsigned long)) {
600 		dst->val = *(unsigned long *) nla_data(nla);
601 		dst->len = sizeof(unsigned long);
602 	} else if (nla_len(nla) == sizeof(u32)) {
603 		dst->val = nla_get_u32(nla);
604 		dst->len = sizeof(u32);
605 	} else
606 		return -EINVAL;
607 
608 	return 0;
609 }
610 
611 static void meta_int_apply_extras(struct meta_value *v,
612 				  struct meta_obj *dst)
613 {
614 	if (v->hdr.shift)
615 		dst->value >>= v->hdr.shift;
616 
617 	if (v->val)
618 		dst->value &= v->val;
619 }
620 
621 static int meta_int_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
622 {
623 	if (v->len == sizeof(unsigned long))
624 		NLA_PUT(skb, tlv, sizeof(unsigned long), &v->val);
625 	else if (v->len == sizeof(u32)) {
626 		NLA_PUT_U32(skb, tlv, v->val);
627 	}
628 
629 	return 0;
630 
631 nla_put_failure:
632 	return -1;
633 }
634 
635 /**************************************************************************
636  * Type specific operations table
637  **************************************************************************/
638 
639 struct meta_type_ops
640 {
641 	void	(*destroy)(struct meta_value *);
642 	int	(*compare)(struct meta_obj *, struct meta_obj *);
643 	int	(*change)(struct meta_value *, struct nlattr *);
644 	void	(*apply_extras)(struct meta_value *, struct meta_obj *);
645 	int	(*dump)(struct sk_buff *, struct meta_value *, int);
646 };
647 
648 static struct meta_type_ops __meta_type_ops[TCF_META_TYPE_MAX+1] = {
649 	[TCF_META_TYPE_VAR] = {
650 		.destroy = meta_var_destroy,
651 		.compare = meta_var_compare,
652 		.change = meta_var_change,
653 		.apply_extras = meta_var_apply_extras,
654 		.dump = meta_var_dump
655 	},
656 	[TCF_META_TYPE_INT] = {
657 		.compare = meta_int_compare,
658 		.change = meta_int_change,
659 		.apply_extras = meta_int_apply_extras,
660 		.dump = meta_int_dump
661 	}
662 };
663 
664 static inline struct meta_type_ops * meta_type_ops(struct meta_value *v)
665 {
666 	return &__meta_type_ops[meta_type(v)];
667 }
668 
669 /**************************************************************************
670  * Core
671  **************************************************************************/
672 
673 static inline int meta_get(struct sk_buff *skb, struct tcf_pkt_info *info,
674 			   struct meta_value *v, struct meta_obj *dst)
675 {
676 	int err = 0;
677 
678 	if (meta_id(v) == TCF_META_ID_VALUE) {
679 		dst->value = v->val;
680 		dst->len = v->len;
681 		return 0;
682 	}
683 
684 	meta_ops(v)->get(skb, info, v, dst, &err);
685 	if (err < 0)
686 		return err;
687 
688 	if (meta_type_ops(v)->apply_extras)
689 	    meta_type_ops(v)->apply_extras(v, dst);
690 
691 	return 0;
692 }
693 
694 static int em_meta_match(struct sk_buff *skb, struct tcf_ematch *m,
695 			 struct tcf_pkt_info *info)
696 {
697 	int r;
698 	struct meta_match *meta = (struct meta_match *) m->data;
699 	struct meta_obj l_value, r_value;
700 
701 	if (meta_get(skb, info, &meta->lvalue, &l_value) < 0 ||
702 	    meta_get(skb, info, &meta->rvalue, &r_value) < 0)
703 		return 0;
704 
705 	r = meta_type_ops(&meta->lvalue)->compare(&l_value, &r_value);
706 
707 	switch (meta->lvalue.hdr.op) {
708 		case TCF_EM_OPND_EQ:
709 			return !r;
710 		case TCF_EM_OPND_LT:
711 			return r < 0;
712 		case TCF_EM_OPND_GT:
713 			return r > 0;
714 	}
715 
716 	return 0;
717 }
718 
719 static inline void meta_delete(struct meta_match *meta)
720 {
721 	struct meta_type_ops *ops = meta_type_ops(&meta->lvalue);
722 
723 	if (ops && ops->destroy) {
724 		ops->destroy(&meta->lvalue);
725 		ops->destroy(&meta->rvalue);
726 	}
727 
728 	kfree(meta);
729 }
730 
731 static inline int meta_change_data(struct meta_value *dst, struct nlattr *nla)
732 {
733 	if (nla) {
734 		if (nla_len(nla) == 0)
735 			return -EINVAL;
736 
737 		return meta_type_ops(dst)->change(dst, nla);
738 	}
739 
740 	return 0;
741 }
742 
743 static inline int meta_is_supported(struct meta_value *val)
744 {
745 	return (!meta_id(val) || meta_ops(val)->get);
746 }
747 
748 static const struct nla_policy meta_policy[TCA_EM_META_MAX + 1] = {
749 	[TCA_EM_META_HDR]	= { .len = sizeof(struct tcf_meta_hdr) },
750 };
751 
752 static int em_meta_change(struct tcf_proto *tp, void *data, int len,
753 			  struct tcf_ematch *m)
754 {
755 	int err;
756 	struct nlattr *tb[TCA_EM_META_MAX + 1];
757 	struct tcf_meta_hdr *hdr;
758 	struct meta_match *meta = NULL;
759 
760 	err = nla_parse(tb, TCA_EM_META_MAX, data, len, meta_policy);
761 	if (err < 0)
762 		goto errout;
763 
764 	err = -EINVAL;
765 	if (tb[TCA_EM_META_HDR] == NULL)
766 		goto errout;
767 	hdr = nla_data(tb[TCA_EM_META_HDR]);
768 
769 	if (TCF_META_TYPE(hdr->left.kind) != TCF_META_TYPE(hdr->right.kind) ||
770 	    TCF_META_TYPE(hdr->left.kind) > TCF_META_TYPE_MAX ||
771 	    TCF_META_ID(hdr->left.kind) > TCF_META_ID_MAX ||
772 	    TCF_META_ID(hdr->right.kind) > TCF_META_ID_MAX)
773 		goto errout;
774 
775 	meta = kzalloc(sizeof(*meta), GFP_KERNEL);
776 	if (meta == NULL)
777 		goto errout;
778 
779 	memcpy(&meta->lvalue.hdr, &hdr->left, sizeof(hdr->left));
780 	memcpy(&meta->rvalue.hdr, &hdr->right, sizeof(hdr->right));
781 
782 	if (!meta_is_supported(&meta->lvalue) ||
783 	    !meta_is_supported(&meta->rvalue)) {
784 		err = -EOPNOTSUPP;
785 		goto errout;
786 	}
787 
788 	if (meta_change_data(&meta->lvalue, tb[TCA_EM_META_LVALUE]) < 0 ||
789 	    meta_change_data(&meta->rvalue, tb[TCA_EM_META_RVALUE]) < 0)
790 		goto errout;
791 
792 	m->datalen = sizeof(*meta);
793 	m->data = (unsigned long) meta;
794 
795 	err = 0;
796 errout:
797 	if (err && meta)
798 		meta_delete(meta);
799 	return err;
800 }
801 
802 static void em_meta_destroy(struct tcf_proto *tp, struct tcf_ematch *m)
803 {
804 	if (m)
805 		meta_delete((struct meta_match *) m->data);
806 }
807 
808 static int em_meta_dump(struct sk_buff *skb, struct tcf_ematch *em)
809 {
810 	struct meta_match *meta = (struct meta_match *) em->data;
811 	struct tcf_meta_hdr hdr;
812 	struct meta_type_ops *ops;
813 
814 	memset(&hdr, 0, sizeof(hdr));
815 	memcpy(&hdr.left, &meta->lvalue.hdr, sizeof(hdr.left));
816 	memcpy(&hdr.right, &meta->rvalue.hdr, sizeof(hdr.right));
817 
818 	NLA_PUT(skb, TCA_EM_META_HDR, sizeof(hdr), &hdr);
819 
820 	ops = meta_type_ops(&meta->lvalue);
821 	if (ops->dump(skb, &meta->lvalue, TCA_EM_META_LVALUE) < 0 ||
822 	    ops->dump(skb, &meta->rvalue, TCA_EM_META_RVALUE) < 0)
823 		goto nla_put_failure;
824 
825 	return 0;
826 
827 nla_put_failure:
828 	return -1;
829 }
830 
831 static struct tcf_ematch_ops em_meta_ops = {
832 	.kind	  = TCF_EM_META,
833 	.change	  = em_meta_change,
834 	.match	  = em_meta_match,
835 	.destroy  = em_meta_destroy,
836 	.dump	  = em_meta_dump,
837 	.owner	  = THIS_MODULE,
838 	.link	  = LIST_HEAD_INIT(em_meta_ops.link)
839 };
840 
841 static int __init init_em_meta(void)
842 {
843 	return tcf_em_register(&em_meta_ops);
844 }
845 
846 static void __exit exit_em_meta(void)
847 {
848 	tcf_em_unregister(&em_meta_ops);
849 }
850 
851 MODULE_LICENSE("GPL");
852 
853 module_init(init_em_meta);
854 module_exit(exit_em_meta);
855 
856 MODULE_ALIAS_TCF_EMATCH(TCF_EM_META);
857