1/*
2 * Berkeley Packet Filter based traffic classifier
3 *
4 * Might be used to classify traffic through flexible, user-defined and
5 * possibly JIT-ed BPF filters for traffic control as an alternative to
6 * ematches.
7 *
8 * (C) 2013 Daniel Borkmann <dborkman@redhat.com>
9 *
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
13 */
14
15#include <linux/module.h>
16#include <linux/types.h>
17#include <linux/skbuff.h>
18#include <linux/filter.h>
19#include <linux/bpf.h>
20
21#include <net/rtnetlink.h>
22#include <net/pkt_cls.h>
23#include <net/sock.h>
24
25MODULE_LICENSE("GPL");
26MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
27MODULE_DESCRIPTION("TC BPF based classifier");
28
29#define CLS_BPF_NAME_LEN	256
30
31struct cls_bpf_head {
32	struct list_head plist;
33	u32 hgen;
34	struct rcu_head rcu;
35};
36
37struct cls_bpf_prog {
38	struct bpf_prog *filter;
39	struct list_head link;
40	struct tcf_result res;
41	bool exts_integrated;
42	struct tcf_exts exts;
43	u32 handle;
44	union {
45		u32 bpf_fd;
46		u16 bpf_num_ops;
47	};
48	struct sock_filter *bpf_ops;
49	const char *bpf_name;
50	struct tcf_proto *tp;
51	struct rcu_head rcu;
52};
53
54static const struct nla_policy bpf_policy[TCA_BPF_MAX + 1] = {
55	[TCA_BPF_CLASSID]	= { .type = NLA_U32 },
56	[TCA_BPF_FLAGS]		= { .type = NLA_U32 },
57	[TCA_BPF_FD]		= { .type = NLA_U32 },
58	[TCA_BPF_NAME]		= { .type = NLA_NUL_STRING, .len = CLS_BPF_NAME_LEN },
59	[TCA_BPF_OPS_LEN]	= { .type = NLA_U16 },
60	[TCA_BPF_OPS]		= { .type = NLA_BINARY,
61				    .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
62};
63
64static int cls_bpf_exec_opcode(int code)
65{
66	switch (code) {
67	case TC_ACT_OK:
68	case TC_ACT_SHOT:
69	case TC_ACT_STOLEN:
70	case TC_ACT_REDIRECT:
71	case TC_ACT_UNSPEC:
72		return code;
73	default:
74		return TC_ACT_UNSPEC;
75	}
76}
77
78static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
79			    struct tcf_result *res)
80{
81	struct cls_bpf_head *head = rcu_dereference_bh(tp->root);
82	struct cls_bpf_prog *prog;
83#ifdef CONFIG_NET_CLS_ACT
84	bool at_ingress = G_TC_AT(skb->tc_verd) & AT_INGRESS;
85#else
86	bool at_ingress = false;
87#endif
88	int ret = -1;
89
90	if (unlikely(!skb_mac_header_was_set(skb)))
91		return -1;
92
93	/* Needed here for accessing maps. */
94	rcu_read_lock();
95	list_for_each_entry_rcu(prog, &head->plist, link) {
96		int filter_res;
97
98		qdisc_skb_cb(skb)->tc_classid = prog->res.classid;
99
100		if (at_ingress) {
101			/* It is safe to push/pull even if skb_shared() */
102			__skb_push(skb, skb->mac_len);
103			filter_res = BPF_PROG_RUN(prog->filter, skb);
104			__skb_pull(skb, skb->mac_len);
105		} else {
106			filter_res = BPF_PROG_RUN(prog->filter, skb);
107		}
108
109		if (prog->exts_integrated) {
110			res->class = prog->res.class;
111			res->classid = qdisc_skb_cb(skb)->tc_classid;
112
113			ret = cls_bpf_exec_opcode(filter_res);
114			if (ret == TC_ACT_UNSPEC)
115				continue;
116			break;
117		}
118
119		if (filter_res == 0)
120			continue;
121
122		*res = prog->res;
123		if (filter_res != -1)
124			res->classid = filter_res;
125
126		ret = tcf_exts_exec(skb, &prog->exts, res);
127		if (ret < 0)
128			continue;
129
130		break;
131	}
132	rcu_read_unlock();
133
134	return ret;
135}
136
137static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog)
138{
139	return !prog->bpf_ops;
140}
141
142static int cls_bpf_init(struct tcf_proto *tp)
143{
144	struct cls_bpf_head *head;
145
146	head = kzalloc(sizeof(*head), GFP_KERNEL);
147	if (head == NULL)
148		return -ENOBUFS;
149
150	INIT_LIST_HEAD_RCU(&head->plist);
151	rcu_assign_pointer(tp->root, head);
152
153	return 0;
154}
155
156static void cls_bpf_delete_prog(struct tcf_proto *tp, struct cls_bpf_prog *prog)
157{
158	tcf_exts_destroy(&prog->exts);
159
160	if (cls_bpf_is_ebpf(prog))
161		bpf_prog_put(prog->filter);
162	else
163		bpf_prog_destroy(prog->filter);
164
165	kfree(prog->bpf_name);
166	kfree(prog->bpf_ops);
167	kfree(prog);
168}
169
170static void __cls_bpf_delete_prog(struct rcu_head *rcu)
171{
172	struct cls_bpf_prog *prog = container_of(rcu, struct cls_bpf_prog, rcu);
173
174	cls_bpf_delete_prog(prog->tp, prog);
175}
176
177static int cls_bpf_delete(struct tcf_proto *tp, unsigned long arg)
178{
179	struct cls_bpf_prog *prog = (struct cls_bpf_prog *) arg;
180
181	list_del_rcu(&prog->link);
182	tcf_unbind_filter(tp, &prog->res);
183	call_rcu(&prog->rcu, __cls_bpf_delete_prog);
184
185	return 0;
186}
187
188static bool cls_bpf_destroy(struct tcf_proto *tp, bool force)
189{
190	struct cls_bpf_head *head = rtnl_dereference(tp->root);
191	struct cls_bpf_prog *prog, *tmp;
192
193	if (!force && !list_empty(&head->plist))
194		return false;
195
196	list_for_each_entry_safe(prog, tmp, &head->plist, link) {
197		list_del_rcu(&prog->link);
198		tcf_unbind_filter(tp, &prog->res);
199		call_rcu(&prog->rcu, __cls_bpf_delete_prog);
200	}
201
202	RCU_INIT_POINTER(tp->root, NULL);
203	kfree_rcu(head, rcu);
204	return true;
205}
206
207static unsigned long cls_bpf_get(struct tcf_proto *tp, u32 handle)
208{
209	struct cls_bpf_head *head = rtnl_dereference(tp->root);
210	struct cls_bpf_prog *prog;
211	unsigned long ret = 0UL;
212
213	if (head == NULL)
214		return 0UL;
215
216	list_for_each_entry(prog, &head->plist, link) {
217		if (prog->handle == handle) {
218			ret = (unsigned long) prog;
219			break;
220		}
221	}
222
223	return ret;
224}
225
226static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog)
227{
228	struct sock_filter *bpf_ops;
229	struct sock_fprog_kern fprog_tmp;
230	struct bpf_prog *fp;
231	u16 bpf_size, bpf_num_ops;
232	int ret;
233
234	bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]);
235	if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
236		return -EINVAL;
237
238	bpf_size = bpf_num_ops * sizeof(*bpf_ops);
239	if (bpf_size != nla_len(tb[TCA_BPF_OPS]))
240		return -EINVAL;
241
242	bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
243	if (bpf_ops == NULL)
244		return -ENOMEM;
245
246	memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size);
247
248	fprog_tmp.len = bpf_num_ops;
249	fprog_tmp.filter = bpf_ops;
250
251	ret = bpf_prog_create(&fp, &fprog_tmp);
252	if (ret < 0) {
253		kfree(bpf_ops);
254		return ret;
255	}
256
257	prog->bpf_ops = bpf_ops;
258	prog->bpf_num_ops = bpf_num_ops;
259	prog->bpf_name = NULL;
260	prog->filter = fp;
261
262	return 0;
263}
264
265static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog,
266				 const struct tcf_proto *tp)
267{
268	struct bpf_prog *fp;
269	char *name = NULL;
270	u32 bpf_fd;
271
272	bpf_fd = nla_get_u32(tb[TCA_BPF_FD]);
273
274	fp = bpf_prog_get(bpf_fd);
275	if (IS_ERR(fp))
276		return PTR_ERR(fp);
277
278	if (fp->type != BPF_PROG_TYPE_SCHED_CLS) {
279		bpf_prog_put(fp);
280		return -EINVAL;
281	}
282
283	if (tb[TCA_BPF_NAME]) {
284		name = kmemdup(nla_data(tb[TCA_BPF_NAME]),
285			       nla_len(tb[TCA_BPF_NAME]),
286			       GFP_KERNEL);
287		if (!name) {
288			bpf_prog_put(fp);
289			return -ENOMEM;
290		}
291	}
292
293	prog->bpf_ops = NULL;
294	prog->bpf_fd = bpf_fd;
295	prog->bpf_name = name;
296	prog->filter = fp;
297
298	if (fp->dst_needed)
299		netif_keep_dst(qdisc_dev(tp->q));
300
301	return 0;
302}
303
304static int cls_bpf_modify_existing(struct net *net, struct tcf_proto *tp,
305				   struct cls_bpf_prog *prog,
306				   unsigned long base, struct nlattr **tb,
307				   struct nlattr *est, bool ovr)
308{
309	bool is_bpf, is_ebpf, have_exts = false;
310	struct tcf_exts exts;
311	int ret;
312
313	is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS];
314	is_ebpf = tb[TCA_BPF_FD];
315	if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf))
316		return -EINVAL;
317
318	tcf_exts_init(&exts, TCA_BPF_ACT, TCA_BPF_POLICE);
319	ret = tcf_exts_validate(net, tp, tb, est, &exts, ovr);
320	if (ret < 0)
321		return ret;
322
323	if (tb[TCA_BPF_FLAGS]) {
324		u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]);
325
326		if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT) {
327			tcf_exts_destroy(&exts);
328			return -EINVAL;
329		}
330
331		have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT;
332	}
333
334	prog->exts_integrated = have_exts;
335
336	ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) :
337		       cls_bpf_prog_from_efd(tb, prog, tp);
338	if (ret < 0) {
339		tcf_exts_destroy(&exts);
340		return ret;
341	}
342
343	if (tb[TCA_BPF_CLASSID]) {
344		prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]);
345		tcf_bind_filter(tp, &prog->res, base);
346	}
347
348	tcf_exts_change(tp, &prog->exts, &exts);
349	return 0;
350}
351
352static u32 cls_bpf_grab_new_handle(struct tcf_proto *tp,
353				   struct cls_bpf_head *head)
354{
355	unsigned int i = 0x80000000;
356	u32 handle;
357
358	do {
359		if (++head->hgen == 0x7FFFFFFF)
360			head->hgen = 1;
361	} while (--i > 0 && cls_bpf_get(tp, head->hgen));
362
363	if (unlikely(i == 0)) {
364		pr_err("Insufficient number of handles\n");
365		handle = 0;
366	} else {
367		handle = head->hgen;
368	}
369
370	return handle;
371}
372
373static int cls_bpf_change(struct net *net, struct sk_buff *in_skb,
374			  struct tcf_proto *tp, unsigned long base,
375			  u32 handle, struct nlattr **tca,
376			  unsigned long *arg, bool ovr)
377{
378	struct cls_bpf_head *head = rtnl_dereference(tp->root);
379	struct cls_bpf_prog *oldprog = (struct cls_bpf_prog *) *arg;
380	struct nlattr *tb[TCA_BPF_MAX + 1];
381	struct cls_bpf_prog *prog;
382	int ret;
383
384	if (tca[TCA_OPTIONS] == NULL)
385		return -EINVAL;
386
387	ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy);
388	if (ret < 0)
389		return ret;
390
391	prog = kzalloc(sizeof(*prog), GFP_KERNEL);
392	if (!prog)
393		return -ENOBUFS;
394
395	tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE);
396
397	if (oldprog) {
398		if (handle && oldprog->handle != handle) {
399			ret = -EINVAL;
400			goto errout;
401		}
402	}
403
404	if (handle == 0)
405		prog->handle = cls_bpf_grab_new_handle(tp, head);
406	else
407		prog->handle = handle;
408	if (prog->handle == 0) {
409		ret = -EINVAL;
410		goto errout;
411	}
412
413	ret = cls_bpf_modify_existing(net, tp, prog, base, tb, tca[TCA_RATE], ovr);
414	if (ret < 0)
415		goto errout;
416
417	if (oldprog) {
418		list_replace_rcu(&oldprog->link, &prog->link);
419		tcf_unbind_filter(tp, &oldprog->res);
420		call_rcu(&oldprog->rcu, __cls_bpf_delete_prog);
421	} else {
422		list_add_rcu(&prog->link, &head->plist);
423	}
424
425	*arg = (unsigned long) prog;
426	return 0;
427errout:
428	kfree(prog);
429
430	return ret;
431}
432
433static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog,
434				 struct sk_buff *skb)
435{
436	struct nlattr *nla;
437
438	if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops))
439		return -EMSGSIZE;
440
441	nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops *
442			  sizeof(struct sock_filter));
443	if (nla == NULL)
444		return -EMSGSIZE;
445
446	memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
447
448	return 0;
449}
450
451static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog,
452				  struct sk_buff *skb)
453{
454	if (nla_put_u32(skb, TCA_BPF_FD, prog->bpf_fd))
455		return -EMSGSIZE;
456
457	if (prog->bpf_name &&
458	    nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name))
459		return -EMSGSIZE;
460
461	return 0;
462}
463
464static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
465			struct sk_buff *skb, struct tcmsg *tm)
466{
467	struct cls_bpf_prog *prog = (struct cls_bpf_prog *) fh;
468	struct nlattr *nest;
469	u32 bpf_flags = 0;
470	int ret;
471
472	if (prog == NULL)
473		return skb->len;
474
475	tm->tcm_handle = prog->handle;
476
477	nest = nla_nest_start(skb, TCA_OPTIONS);
478	if (nest == NULL)
479		goto nla_put_failure;
480
481	if (prog->res.classid &&
482	    nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid))
483		goto nla_put_failure;
484
485	if (cls_bpf_is_ebpf(prog))
486		ret = cls_bpf_dump_ebpf_info(prog, skb);
487	else
488		ret = cls_bpf_dump_bpf_info(prog, skb);
489	if (ret)
490		goto nla_put_failure;
491
492	if (tcf_exts_dump(skb, &prog->exts) < 0)
493		goto nla_put_failure;
494
495	if (prog->exts_integrated)
496		bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT;
497	if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags))
498		goto nla_put_failure;
499
500	nla_nest_end(skb, nest);
501
502	if (tcf_exts_dump_stats(skb, &prog->exts) < 0)
503		goto nla_put_failure;
504
505	return skb->len;
506
507nla_put_failure:
508	nla_nest_cancel(skb, nest);
509	return -1;
510}
511
512static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg)
513{
514	struct cls_bpf_head *head = rtnl_dereference(tp->root);
515	struct cls_bpf_prog *prog;
516
517	list_for_each_entry(prog, &head->plist, link) {
518		if (arg->count < arg->skip)
519			goto skip;
520		if (arg->fn(tp, (unsigned long) prog, arg) < 0) {
521			arg->stop = 1;
522			break;
523		}
524skip:
525		arg->count++;
526	}
527}
528
529static struct tcf_proto_ops cls_bpf_ops __read_mostly = {
530	.kind		=	"bpf",
531	.owner		=	THIS_MODULE,
532	.classify	=	cls_bpf_classify,
533	.init		=	cls_bpf_init,
534	.destroy	=	cls_bpf_destroy,
535	.get		=	cls_bpf_get,
536	.change		=	cls_bpf_change,
537	.delete		=	cls_bpf_delete,
538	.walk		=	cls_bpf_walk,
539	.dump		=	cls_bpf_dump,
540};
541
542static int __init cls_bpf_init_mod(void)
543{
544	return register_tcf_proto_ops(&cls_bpf_ops);
545}
546
547static void __exit cls_bpf_exit_mod(void)
548{
549	unregister_tcf_proto_ops(&cls_bpf_ops);
550}
551
552module_init(cls_bpf_init_mod);
553module_exit(cls_bpf_exit_mod);
554