xref: /openbmc/linux/net/sched/em_meta.c (revision e8e0929d) 
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/if_vlan.h>
69 #include <linux/tc_ematch/tc_em_meta.h>
70 #include <net/dst.h>
71 #include <net/route.h>
72 #include <net/pkt_cls.h>
73 #include <net/sock.h>
74 
75 struct meta_obj
76 {
77 	unsigned long		value;
78 	unsigned int		len;
79 };
80 
81 struct meta_value
82 {
83 	struct tcf_meta_val	hdr;
84 	unsigned long		val;
85 	unsigned int		len;
86 };
87 
88 struct meta_match
89 {
90 	struct meta_value	lvalue;
91 	struct meta_value	rvalue;
92 };
93 
94 static inline int meta_id(struct meta_value *v)
95 {
96 	return TCF_META_ID(v->hdr.kind);
97 }
98 
99 static inline int meta_type(struct meta_value *v)
100 {
101 	return TCF_META_TYPE(v->hdr.kind);
102 }
103 
104 #define META_COLLECTOR(FUNC) static void meta_##FUNC(struct sk_buff *skb, \
105 	struct tcf_pkt_info *info, struct meta_value *v, \
106 	struct meta_obj *dst, int *err)
107 
108 /**************************************************************************
109  * System status & misc
110  **************************************************************************/
111 
112 META_COLLECTOR(int_random)
113 {
114 	get_random_bytes(&dst->value, sizeof(dst->value));
115 }
116 
117 static inline unsigned long fixed_loadavg(int load)
118 {
119 	int rnd_load = load + (FIXED_1/200);
120 	int rnd_frac = ((rnd_load & (FIXED_1-1)) * 100) >> FSHIFT;
121 
122 	return ((rnd_load >> FSHIFT) * 100) + rnd_frac;
123 }
124 
125 META_COLLECTOR(int_loadavg_0)
126 {
127 	dst->value = fixed_loadavg(avenrun[0]);
128 }
129 
130 META_COLLECTOR(int_loadavg_1)
131 {
132 	dst->value = fixed_loadavg(avenrun[1]);
133 }
134 
135 META_COLLECTOR(int_loadavg_2)
136 {
137 	dst->value = fixed_loadavg(avenrun[2]);
138 }
139 
140 /**************************************************************************
141  * Device names & indices
142  **************************************************************************/
143 
144 static inline int int_dev(struct net_device *dev, struct meta_obj *dst)
145 {
146 	if (unlikely(dev == NULL))
147 		return -1;
148 
149 	dst->value = dev->ifindex;
150 	return 0;
151 }
152 
153 static inline int var_dev(struct net_device *dev, struct meta_obj *dst)
154 {
155 	if (unlikely(dev == NULL))
156 		return -1;
157 
158 	dst->value = (unsigned long) dev->name;
159 	dst->len = strlen(dev->name);
160 	return 0;
161 }
162 
163 META_COLLECTOR(int_dev)
164 {
165 	*err = int_dev(skb->dev, dst);
166 }
167 
168 META_COLLECTOR(var_dev)
169 {
170 	*err = var_dev(skb->dev, dst);
171 }
172 
173 /**************************************************************************
174  * vlan tag
175  **************************************************************************/
176 
177 META_COLLECTOR(int_vlan_tag)
178 {
179 	unsigned short tag;
180 
181 	tag = vlan_tx_tag_get(skb);
182 	if (!tag && __vlan_get_tag(skb, &tag))
183 		*err = -1;
184 	else
185 		dst->value = tag;
186 }
187 
188 
189 
190 /**************************************************************************
191  * skb attributes
192  **************************************************************************/
193 
194 META_COLLECTOR(int_priority)
195 {
196 	dst->value = skb->priority;
197 }
198 
199 META_COLLECTOR(int_protocol)
200 {
201 	/* Let userspace take care of the byte ordering */
202 	dst->value = skb->protocol;
203 }
204 
205 META_COLLECTOR(int_pkttype)
206 {
207 	dst->value = skb->pkt_type;
208 }
209 
210 META_COLLECTOR(int_pktlen)
211 {
212 	dst->value = skb->len;
213 }
214 
215 META_COLLECTOR(int_datalen)
216 {
217 	dst->value = skb->data_len;
218 }
219 
220 META_COLLECTOR(int_maclen)
221 {
222 	dst->value = skb->mac_len;
223 }
224 
225 /**************************************************************************
226  * Netfilter
227  **************************************************************************/
228 
229 META_COLLECTOR(int_mark)
230 {
231 	dst->value = skb->mark;
232 }
233 
234 /**************************************************************************
235  * Traffic Control
236  **************************************************************************/
237 
238 META_COLLECTOR(int_tcindex)
239 {
240 	dst->value = skb->tc_index;
241 }
242 
243 /**************************************************************************
244  * Routing
245  **************************************************************************/
246 
247 META_COLLECTOR(int_rtclassid)
248 {
249 	if (unlikely(skb_dst(skb) == NULL))
250 		*err = -1;
251 	else
252 #ifdef CONFIG_NET_CLS_ROUTE
253 		dst->value = skb_dst(skb)->tclassid;
254 #else
255 		dst->value = 0;
256 #endif
257 }
258 
259 META_COLLECTOR(int_rtiif)
260 {
261 	if (unlikely(skb_rtable(skb) == NULL))
262 		*err = -1;
263 	else
264 		dst->value = skb_rtable(skb)->fl.iif;
265 }
266 
267 /**************************************************************************
268  * Socket Attributes
269  **************************************************************************/
270 
271 #define SKIP_NONLOCAL(skb)			\
272 	if (unlikely(skb->sk == NULL)) {	\
273 		*err = -1;			\
274 		return;				\
275 	}
276 
277 META_COLLECTOR(int_sk_family)
278 {
279 	SKIP_NONLOCAL(skb);
280 	dst->value = skb->sk->sk_family;
281 }
282 
283 META_COLLECTOR(int_sk_state)
284 {
285 	SKIP_NONLOCAL(skb);
286 	dst->value = skb->sk->sk_state;
287 }
288 
289 META_COLLECTOR(int_sk_reuse)
290 {
291 	SKIP_NONLOCAL(skb);
292 	dst->value = skb->sk->sk_reuse;
293 }
294 
295 META_COLLECTOR(int_sk_bound_if)
296 {
297 	SKIP_NONLOCAL(skb);
298 	/* No error if bound_dev_if is 0, legal userspace check */
299 	dst->value = skb->sk->sk_bound_dev_if;
300 }
301 
302 META_COLLECTOR(var_sk_bound_if)
303 {
304 	SKIP_NONLOCAL(skb);
305 
306 	 if (skb->sk->sk_bound_dev_if == 0) {
307 		dst->value = (unsigned long) "any";
308 		dst->len = 3;
309 	 } else  {
310 		struct net_device *dev;
311 
312 		dev = dev_get_by_index(&init_net, skb->sk->sk_bound_dev_if);
313 		*err = var_dev(dev, dst);
314 		if (dev)
315 			dev_put(dev);
316 	 }
317 }
318 
319 META_COLLECTOR(int_sk_refcnt)
320 {
321 	SKIP_NONLOCAL(skb);
322 	dst->value = atomic_read(&skb->sk->sk_refcnt);
323 }
324 
325 META_COLLECTOR(int_sk_rcvbuf)
326 {
327 	SKIP_NONLOCAL(skb);
328 	dst->value = skb->sk->sk_rcvbuf;
329 }
330 
331 META_COLLECTOR(int_sk_shutdown)
332 {
333 	SKIP_NONLOCAL(skb);
334 	dst->value = skb->sk->sk_shutdown;
335 }
336 
337 META_COLLECTOR(int_sk_proto)
338 {
339 	SKIP_NONLOCAL(skb);
340 	dst->value = skb->sk->sk_protocol;
341 }
342 
343 META_COLLECTOR(int_sk_type)
344 {
345 	SKIP_NONLOCAL(skb);
346 	dst->value = skb->sk->sk_type;
347 }
348 
349 META_COLLECTOR(int_sk_rmem_alloc)
350 {
351 	SKIP_NONLOCAL(skb);
352 	dst->value = sk_rmem_alloc_get(skb->sk);
353 }
354 
355 META_COLLECTOR(int_sk_wmem_alloc)
356 {
357 	SKIP_NONLOCAL(skb);
358 	dst->value = sk_wmem_alloc_get(skb->sk);
359 }
360 
361 META_COLLECTOR(int_sk_omem_alloc)
362 {
363 	SKIP_NONLOCAL(skb);
364 	dst->value = atomic_read(&skb->sk->sk_omem_alloc);
365 }
366 
367 META_COLLECTOR(int_sk_rcv_qlen)
368 {
369 	SKIP_NONLOCAL(skb);
370 	dst->value = skb->sk->sk_receive_queue.qlen;
371 }
372 
373 META_COLLECTOR(int_sk_snd_qlen)
374 {
375 	SKIP_NONLOCAL(skb);
376 	dst->value = skb->sk->sk_write_queue.qlen;
377 }
378 
379 META_COLLECTOR(int_sk_wmem_queued)
380 {
381 	SKIP_NONLOCAL(skb);
382 	dst->value = skb->sk->sk_wmem_queued;
383 }
384 
385 META_COLLECTOR(int_sk_fwd_alloc)
386 {
387 	SKIP_NONLOCAL(skb);
388 	dst->value = skb->sk->sk_forward_alloc;
389 }
390 
391 META_COLLECTOR(int_sk_sndbuf)
392 {
393 	SKIP_NONLOCAL(skb);
394 	dst->value = skb->sk->sk_sndbuf;
395 }
396 
397 META_COLLECTOR(int_sk_alloc)
398 {
399 	SKIP_NONLOCAL(skb);
400 	dst->value = skb->sk->sk_allocation;
401 }
402 
403 META_COLLECTOR(int_sk_route_caps)
404 {
405 	SKIP_NONLOCAL(skb);
406 	dst->value = skb->sk->sk_route_caps;
407 }
408 
409 META_COLLECTOR(int_sk_hash)
410 {
411 	SKIP_NONLOCAL(skb);
412 	dst->value = skb->sk->sk_hash;
413 }
414 
415 META_COLLECTOR(int_sk_lingertime)
416 {
417 	SKIP_NONLOCAL(skb);
418 	dst->value = skb->sk->sk_lingertime / HZ;
419 }
420 
421 META_COLLECTOR(int_sk_err_qlen)
422 {
423 	SKIP_NONLOCAL(skb);
424 	dst->value = skb->sk->sk_error_queue.qlen;
425 }
426 
427 META_COLLECTOR(int_sk_ack_bl)
428 {
429 	SKIP_NONLOCAL(skb);
430 	dst->value = skb->sk->sk_ack_backlog;
431 }
432 
433 META_COLLECTOR(int_sk_max_ack_bl)
434 {
435 	SKIP_NONLOCAL(skb);
436 	dst->value = skb->sk->sk_max_ack_backlog;
437 }
438 
439 META_COLLECTOR(int_sk_prio)
440 {
441 	SKIP_NONLOCAL(skb);
442 	dst->value = skb->sk->sk_priority;
443 }
444 
445 META_COLLECTOR(int_sk_rcvlowat)
446 {
447 	SKIP_NONLOCAL(skb);
448 	dst->value = skb->sk->sk_rcvlowat;
449 }
450 
451 META_COLLECTOR(int_sk_rcvtimeo)
452 {
453 	SKIP_NONLOCAL(skb);
454 	dst->value = skb->sk->sk_rcvtimeo / HZ;
455 }
456 
457 META_COLLECTOR(int_sk_sndtimeo)
458 {
459 	SKIP_NONLOCAL(skb);
460 	dst->value = skb->sk->sk_sndtimeo / HZ;
461 }
462 
463 META_COLLECTOR(int_sk_sendmsg_off)
464 {
465 	SKIP_NONLOCAL(skb);
466 	dst->value = skb->sk->sk_sndmsg_off;
467 }
468 
469 META_COLLECTOR(int_sk_write_pend)
470 {
471 	SKIP_NONLOCAL(skb);
472 	dst->value = skb->sk->sk_write_pending;
473 }
474 
475 /**************************************************************************
476  * Meta value collectors assignment table
477  **************************************************************************/
478 
479 struct meta_ops
480 {
481 	void		(*get)(struct sk_buff *, struct tcf_pkt_info *,
482 			       struct meta_value *, struct meta_obj *, int *);
483 };
484 
485 #define META_ID(name) TCF_META_ID_##name
486 #define META_FUNC(name) { .get = meta_##name }
487 
488 /* Meta value operations table listing all meta value collectors and
489  * assigns them to a type and meta id. */
490 static struct meta_ops __meta_ops[TCF_META_TYPE_MAX+1][TCF_META_ID_MAX+1] = {
491 	[TCF_META_TYPE_VAR] = {
492 		[META_ID(DEV)]			= META_FUNC(var_dev),
493 		[META_ID(SK_BOUND_IF)] 		= META_FUNC(var_sk_bound_if),
494 	},
495 	[TCF_META_TYPE_INT] = {
496 		[META_ID(RANDOM)]		= META_FUNC(int_random),
497 		[META_ID(LOADAVG_0)]		= META_FUNC(int_loadavg_0),
498 		[META_ID(LOADAVG_1)]		= META_FUNC(int_loadavg_1),
499 		[META_ID(LOADAVG_2)]		= META_FUNC(int_loadavg_2),
500 		[META_ID(DEV)]			= META_FUNC(int_dev),
501 		[META_ID(PRIORITY)]		= META_FUNC(int_priority),
502 		[META_ID(PROTOCOL)]		= META_FUNC(int_protocol),
503 		[META_ID(PKTTYPE)]		= META_FUNC(int_pkttype),
504 		[META_ID(PKTLEN)]		= META_FUNC(int_pktlen),
505 		[META_ID(DATALEN)]		= META_FUNC(int_datalen),
506 		[META_ID(MACLEN)]		= META_FUNC(int_maclen),
507 		[META_ID(NFMARK)]		= META_FUNC(int_mark),
508 		[META_ID(TCINDEX)]		= META_FUNC(int_tcindex),
509 		[META_ID(RTCLASSID)]		= META_FUNC(int_rtclassid),
510 		[META_ID(RTIIF)]		= META_FUNC(int_rtiif),
511 		[META_ID(SK_FAMILY)]		= META_FUNC(int_sk_family),
512 		[META_ID(SK_STATE)]		= META_FUNC(int_sk_state),
513 		[META_ID(SK_REUSE)]		= META_FUNC(int_sk_reuse),
514 		[META_ID(SK_BOUND_IF)]		= META_FUNC(int_sk_bound_if),
515 		[META_ID(SK_REFCNT)]		= META_FUNC(int_sk_refcnt),
516 		[META_ID(SK_RCVBUF)]		= META_FUNC(int_sk_rcvbuf),
517 		[META_ID(SK_SNDBUF)]		= META_FUNC(int_sk_sndbuf),
518 		[META_ID(SK_SHUTDOWN)]		= META_FUNC(int_sk_shutdown),
519 		[META_ID(SK_PROTO)]		= META_FUNC(int_sk_proto),
520 		[META_ID(SK_TYPE)]		= META_FUNC(int_sk_type),
521 		[META_ID(SK_RMEM_ALLOC)]	= META_FUNC(int_sk_rmem_alloc),
522 		[META_ID(SK_WMEM_ALLOC)]	= META_FUNC(int_sk_wmem_alloc),
523 		[META_ID(SK_OMEM_ALLOC)]	= META_FUNC(int_sk_omem_alloc),
524 		[META_ID(SK_WMEM_QUEUED)]	= META_FUNC(int_sk_wmem_queued),
525 		[META_ID(SK_RCV_QLEN)]		= META_FUNC(int_sk_rcv_qlen),
526 		[META_ID(SK_SND_QLEN)]		= META_FUNC(int_sk_snd_qlen),
527 		[META_ID(SK_ERR_QLEN)]		= META_FUNC(int_sk_err_qlen),
528 		[META_ID(SK_FORWARD_ALLOCS)]	= META_FUNC(int_sk_fwd_alloc),
529 		[META_ID(SK_ALLOCS)]		= META_FUNC(int_sk_alloc),
530 		[META_ID(SK_ROUTE_CAPS)]	= META_FUNC(int_sk_route_caps),
531 		[META_ID(SK_HASH)]		= META_FUNC(int_sk_hash),
532 		[META_ID(SK_LINGERTIME)]	= META_FUNC(int_sk_lingertime),
533 		[META_ID(SK_ACK_BACKLOG)]	= META_FUNC(int_sk_ack_bl),
534 		[META_ID(SK_MAX_ACK_BACKLOG)]	= META_FUNC(int_sk_max_ack_bl),
535 		[META_ID(SK_PRIO)]		= META_FUNC(int_sk_prio),
536 		[META_ID(SK_RCVLOWAT)]		= META_FUNC(int_sk_rcvlowat),
537 		[META_ID(SK_RCVTIMEO)]		= META_FUNC(int_sk_rcvtimeo),
538 		[META_ID(SK_SNDTIMEO)]		= META_FUNC(int_sk_sndtimeo),
539 		[META_ID(SK_SENDMSG_OFF)]	= META_FUNC(int_sk_sendmsg_off),
540 		[META_ID(SK_WRITE_PENDING)]	= META_FUNC(int_sk_write_pend),
541 		[META_ID(VLAN_TAG)]		= META_FUNC(int_vlan_tag),
542 	}
543 };
544 
545 static inline struct meta_ops * meta_ops(struct meta_value *val)
546 {
547 	return &__meta_ops[meta_type(val)][meta_id(val)];
548 }
549 
550 /**************************************************************************
551  * Type specific operations for TCF_META_TYPE_VAR
552  **************************************************************************/
553 
554 static int meta_var_compare(struct meta_obj *a, struct meta_obj *b)
555 {
556 	int r = a->len - b->len;
557 
558 	if (r == 0)
559 		r = memcmp((void *) a->value, (void *) b->value, a->len);
560 
561 	return r;
562 }
563 
564 static int meta_var_change(struct meta_value *dst, struct nlattr *nla)
565 {
566 	int len = nla_len(nla);
567 
568 	dst->val = (unsigned long)kmemdup(nla_data(nla), len, GFP_KERNEL);
569 	if (dst->val == 0UL)
570 		return -ENOMEM;
571 	dst->len = len;
572 	return 0;
573 }
574 
575 static void meta_var_destroy(struct meta_value *v)
576 {
577 	kfree((void *) v->val);
578 }
579 
580 static void meta_var_apply_extras(struct meta_value *v,
581 				  struct meta_obj *dst)
582 {
583 	int shift = v->hdr.shift;
584 
585 	if (shift && shift < dst->len)
586 		dst->len -= shift;
587 }
588 
589 static int meta_var_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
590 {
591 	if (v->val && v->len)
592 		NLA_PUT(skb, tlv, v->len, (void *) v->val);
593 	return 0;
594 
595 nla_put_failure:
596 	return -1;
597 }
598 
599 /**************************************************************************
600  * Type specific operations for TCF_META_TYPE_INT
601  **************************************************************************/
602 
603 static int meta_int_compare(struct meta_obj *a, struct meta_obj *b)
604 {
605 	/* Let gcc optimize it, the unlikely is not really based on
606 	 * some numbers but jump free code for mismatches seems
607 	 * more logical. */
608 	if (unlikely(a->value == b->value))
609 		return 0;
610 	else if (a->value < b->value)
611 		return -1;
612 	else
613 		return 1;
614 }
615 
616 static int meta_int_change(struct meta_value *dst, struct nlattr *nla)
617 {
618 	if (nla_len(nla) >= sizeof(unsigned long)) {
619 		dst->val = *(unsigned long *) nla_data(nla);
620 		dst->len = sizeof(unsigned long);
621 	} else if (nla_len(nla) == sizeof(u32)) {
622 		dst->val = nla_get_u32(nla);
623 		dst->len = sizeof(u32);
624 	} else
625 		return -EINVAL;
626 
627 	return 0;
628 }
629 
630 static void meta_int_apply_extras(struct meta_value *v,
631 				  struct meta_obj *dst)
632 {
633 	if (v->hdr.shift)
634 		dst->value >>= v->hdr.shift;
635 
636 	if (v->val)
637 		dst->value &= v->val;
638 }
639 
640 static int meta_int_dump(struct sk_buff *skb, struct meta_value *v, int tlv)
641 {
642 	if (v->len == sizeof(unsigned long))
643 		NLA_PUT(skb, tlv, sizeof(unsigned long), &v->val);
644 	else if (v->len == sizeof(u32)) {
645 		NLA_PUT_U32(skb, tlv, v->val);
646 	}
647 
648 	return 0;
649 
650 nla_put_failure:
651 	return -1;
652 }
653 
654 /**************************************************************************
655  * Type specific operations table
656  **************************************************************************/
657 
658 struct meta_type_ops
659 {
660 	void	(*destroy)(struct meta_value *);
661 	int	(*compare)(struct meta_obj *, struct meta_obj *);
662 	int	(*change)(struct meta_value *, struct nlattr *);
663 	void	(*apply_extras)(struct meta_value *, struct meta_obj *);
664 	int	(*dump)(struct sk_buff *, struct meta_value *, int);
665 };
666 
667 static struct meta_type_ops __meta_type_ops[TCF_META_TYPE_MAX+1] = {
668 	[TCF_META_TYPE_VAR] = {
669 		.destroy = meta_var_destroy,
670 		.compare = meta_var_compare,
671 		.change = meta_var_change,
672 		.apply_extras = meta_var_apply_extras,
673 		.dump = meta_var_dump
674 	},
675 	[TCF_META_TYPE_INT] = {
676 		.compare = meta_int_compare,
677 		.change = meta_int_change,
678 		.apply_extras = meta_int_apply_extras,
679 		.dump = meta_int_dump
680 	}
681 };
682 
683 static inline struct meta_type_ops * meta_type_ops(struct meta_value *v)
684 {
685 	return &__meta_type_ops[meta_type(v)];
686 }
687 
688 /**************************************************************************
689  * Core
690  **************************************************************************/
691 
692 static int meta_get(struct sk_buff *skb, struct tcf_pkt_info *info,
693 		    struct meta_value *v, struct meta_obj *dst)
694 {
695 	int err = 0;
696 
697 	if (meta_id(v) == TCF_META_ID_VALUE) {
698 		dst->value = v->val;
699 		dst->len = v->len;
700 		return 0;
701 	}
702 
703 	meta_ops(v)->get(skb, info, v, dst, &err);
704 	if (err < 0)
705 		return err;
706 
707 	if (meta_type_ops(v)->apply_extras)
708 	    meta_type_ops(v)->apply_extras(v, dst);
709 
710 	return 0;
711 }
712 
713 static int em_meta_match(struct sk_buff *skb, struct tcf_ematch *m,
714 			 struct tcf_pkt_info *info)
715 {
716 	int r;
717 	struct meta_match *meta = (struct meta_match *) m->data;
718 	struct meta_obj l_value, r_value;
719 
720 	if (meta_get(skb, info, &meta->lvalue, &l_value) < 0 ||
721 	    meta_get(skb, info, &meta->rvalue, &r_value) < 0)
722 		return 0;
723 
724 	r = meta_type_ops(&meta->lvalue)->compare(&l_value, &r_value);
725 
726 	switch (meta->lvalue.hdr.op) {
727 		case TCF_EM_OPND_EQ:
728 			return !r;
729 		case TCF_EM_OPND_LT:
730 			return r < 0;
731 		case TCF_EM_OPND_GT:
732 			return r > 0;
733 	}
734 
735 	return 0;
736 }
737 
738 static void meta_delete(struct meta_match *meta)
739 {
740 	if (meta) {
741 		struct meta_type_ops *ops = meta_type_ops(&meta->lvalue);
742 
743 		if (ops && ops->destroy) {
744 			ops->destroy(&meta->lvalue);
745 			ops->destroy(&meta->rvalue);
746 		}
747 	}
748 
749 	kfree(meta);
750 }
751 
752 static inline int meta_change_data(struct meta_value *dst, struct nlattr *nla)
753 {
754 	if (nla) {
755 		if (nla_len(nla) == 0)
756 			return -EINVAL;
757 
758 		return meta_type_ops(dst)->change(dst, nla);
759 	}
760 
761 	return 0;
762 }
763 
764 static inline int meta_is_supported(struct meta_value *val)
765 {
766 	return (!meta_id(val) || meta_ops(val)->get);
767 }
768 
769 static const struct nla_policy meta_policy[TCA_EM_META_MAX + 1] = {
770 	[TCA_EM_META_HDR]	= { .len = sizeof(struct tcf_meta_hdr) },
771 };
772 
773 static int em_meta_change(struct tcf_proto *tp, void *data, int len,
774 			  struct tcf_ematch *m)
775 {
776 	int err;
777 	struct nlattr *tb[TCA_EM_META_MAX + 1];
778 	struct tcf_meta_hdr *hdr;
779 	struct meta_match *meta = NULL;
780 
781 	err = nla_parse(tb, TCA_EM_META_MAX, data, len, meta_policy);
782 	if (err < 0)
783 		goto errout;
784 
785 	err = -EINVAL;
786 	if (tb[TCA_EM_META_HDR] == NULL)
787 		goto errout;
788 	hdr = nla_data(tb[TCA_EM_META_HDR]);
789 
790 	if (TCF_META_TYPE(hdr->left.kind) != TCF_META_TYPE(hdr->right.kind) ||
791 	    TCF_META_TYPE(hdr->left.kind) > TCF_META_TYPE_MAX ||
792 	    TCF_META_ID(hdr->left.kind) > TCF_META_ID_MAX ||
793 	    TCF_META_ID(hdr->right.kind) > TCF_META_ID_MAX)
794 		goto errout;
795 
796 	meta = kzalloc(sizeof(*meta), GFP_KERNEL);
797 	if (meta == NULL)
798 		goto errout;
799 
800 	memcpy(&meta->lvalue.hdr, &hdr->left, sizeof(hdr->left));
801 	memcpy(&meta->rvalue.hdr, &hdr->right, sizeof(hdr->right));
802 
803 	if (!meta_is_supported(&meta->lvalue) ||
804 	    !meta_is_supported(&meta->rvalue)) {
805 		err = -EOPNOTSUPP;
806 		goto errout;
807 	}
808 
809 	if (meta_change_data(&meta->lvalue, tb[TCA_EM_META_LVALUE]) < 0 ||
810 	    meta_change_data(&meta->rvalue, tb[TCA_EM_META_RVALUE]) < 0)
811 		goto errout;
812 
813 	m->datalen = sizeof(*meta);
814 	m->data = (unsigned long) meta;
815 
816 	err = 0;
817 errout:
818 	if (err && meta)
819 		meta_delete(meta);
820 	return err;
821 }
822 
823 static void em_meta_destroy(struct tcf_proto *tp, struct tcf_ematch *m)
824 {
825 	if (m)
826 		meta_delete((struct meta_match *) m->data);
827 }
828 
829 static int em_meta_dump(struct sk_buff *skb, struct tcf_ematch *em)
830 {
831 	struct meta_match *meta = (struct meta_match *) em->data;
832 	struct tcf_meta_hdr hdr;
833 	struct meta_type_ops *ops;
834 
835 	memset(&hdr, 0, sizeof(hdr));
836 	memcpy(&hdr.left, &meta->lvalue.hdr, sizeof(hdr.left));
837 	memcpy(&hdr.right, &meta->rvalue.hdr, sizeof(hdr.right));
838 
839 	NLA_PUT(skb, TCA_EM_META_HDR, sizeof(hdr), &hdr);
840 
841 	ops = meta_type_ops(&meta->lvalue);
842 	if (ops->dump(skb, &meta->lvalue, TCA_EM_META_LVALUE) < 0 ||
843 	    ops->dump(skb, &meta->rvalue, TCA_EM_META_RVALUE) < 0)
844 		goto nla_put_failure;
845 
846 	return 0;
847 
848 nla_put_failure:
849 	return -1;
850 }
851 
852 static struct tcf_ematch_ops em_meta_ops = {
853 	.kind	  = TCF_EM_META,
854 	.change	  = em_meta_change,
855 	.match	  = em_meta_match,
856 	.destroy  = em_meta_destroy,
857 	.dump	  = em_meta_dump,
858 	.owner	  = THIS_MODULE,
859 	.link	  = LIST_HEAD_INIT(em_meta_ops.link)
860 };
861 
862 static int __init init_em_meta(void)
863 {
864 	return tcf_em_register(&em_meta_ops);
865 }
866 
867 static void __exit exit_em_meta(void)
868 {
869 	tcf_em_unregister(&em_meta_ops);
870 }
871 
872 MODULE_LICENSE("GPL");
873 
874 module_init(init_em_meta);
875 module_exit(exit_em_meta);
876 
877 MODULE_ALIAS_TCF_EMATCH(TCF_EM_META);
878