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	struct tcf_exts exts;
42	u32 handle;
43	union {
44		u32 bpf_fd;
45		u16 bpf_num_ops;
46	};
47	struct sock_filter *bpf_ops;
48	const char *bpf_name;
49	struct tcf_proto *tp;
50	struct rcu_head rcu;
51};
52
53static const struct nla_policy bpf_policy[TCA_BPF_MAX + 1] = {
54	[TCA_BPF_CLASSID]	= { .type = NLA_U32 },
55	[TCA_BPF_FD]		= { .type = NLA_U32 },
56	[TCA_BPF_NAME]		= { .type = NLA_NUL_STRING, .len = CLS_BPF_NAME_LEN },
57	[TCA_BPF_OPS_LEN]	= { .type = NLA_U16 },
58	[TCA_BPF_OPS]		= { .type = NLA_BINARY,
59				    .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
60};
61
62static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
63			    struct tcf_result *res)
64{
65	struct cls_bpf_head *head = rcu_dereference_bh(tp->root);
66	struct cls_bpf_prog *prog;
67	int ret = -1;
68
69	if (unlikely(!skb_mac_header_was_set(skb)))
70		return -1;
71
72	/* Needed here for accessing maps. */
73	rcu_read_lock();
74	list_for_each_entry_rcu(prog, &head->plist, link) {
75		int filter_res = BPF_PROG_RUN(prog->filter, skb);
76
77		if (filter_res == 0)
78			continue;
79
80		*res = prog->res;
81		if (filter_res != -1)
82			res->classid = filter_res;
83
84		ret = tcf_exts_exec(skb, &prog->exts, res);
85		if (ret < 0)
86			continue;
87
88		break;
89	}
90	rcu_read_unlock();
91
92	return ret;
93}
94
95static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog)
96{
97	return !prog->bpf_ops;
98}
99
100static int cls_bpf_init(struct tcf_proto *tp)
101{
102	struct cls_bpf_head *head;
103
104	head = kzalloc(sizeof(*head), GFP_KERNEL);
105	if (head == NULL)
106		return -ENOBUFS;
107
108	INIT_LIST_HEAD_RCU(&head->plist);
109	rcu_assign_pointer(tp->root, head);
110
111	return 0;
112}
113
114static void cls_bpf_delete_prog(struct tcf_proto *tp, struct cls_bpf_prog *prog)
115{
116	tcf_exts_destroy(&prog->exts);
117
118	if (cls_bpf_is_ebpf(prog))
119		bpf_prog_put(prog->filter);
120	else
121		bpf_prog_destroy(prog->filter);
122
123	kfree(prog->bpf_name);
124	kfree(prog->bpf_ops);
125	kfree(prog);
126}
127
128static void __cls_bpf_delete_prog(struct rcu_head *rcu)
129{
130	struct cls_bpf_prog *prog = container_of(rcu, struct cls_bpf_prog, rcu);
131
132	cls_bpf_delete_prog(prog->tp, prog);
133}
134
135static int cls_bpf_delete(struct tcf_proto *tp, unsigned long arg)
136{
137	struct cls_bpf_prog *prog = (struct cls_bpf_prog *) arg;
138
139	list_del_rcu(&prog->link);
140	tcf_unbind_filter(tp, &prog->res);
141	call_rcu(&prog->rcu, __cls_bpf_delete_prog);
142
143	return 0;
144}
145
146static bool cls_bpf_destroy(struct tcf_proto *tp, bool force)
147{
148	struct cls_bpf_head *head = rtnl_dereference(tp->root);
149	struct cls_bpf_prog *prog, *tmp;
150
151	if (!force && !list_empty(&head->plist))
152		return false;
153
154	list_for_each_entry_safe(prog, tmp, &head->plist, link) {
155		list_del_rcu(&prog->link);
156		tcf_unbind_filter(tp, &prog->res);
157		call_rcu(&prog->rcu, __cls_bpf_delete_prog);
158	}
159
160	RCU_INIT_POINTER(tp->root, NULL);
161	kfree_rcu(head, rcu);
162	return true;
163}
164
165static unsigned long cls_bpf_get(struct tcf_proto *tp, u32 handle)
166{
167	struct cls_bpf_head *head = rtnl_dereference(tp->root);
168	struct cls_bpf_prog *prog;
169	unsigned long ret = 0UL;
170
171	if (head == NULL)
172		return 0UL;
173
174	list_for_each_entry(prog, &head->plist, link) {
175		if (prog->handle == handle) {
176			ret = (unsigned long) prog;
177			break;
178		}
179	}
180
181	return ret;
182}
183
184static int cls_bpf_prog_from_ops(struct nlattr **tb,
185				 struct cls_bpf_prog *prog, u32 classid)
186{
187	struct sock_filter *bpf_ops;
188	struct sock_fprog_kern fprog_tmp;
189	struct bpf_prog *fp;
190	u16 bpf_size, bpf_num_ops;
191	int ret;
192
193	bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]);
194	if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
195		return -EINVAL;
196
197	bpf_size = bpf_num_ops * sizeof(*bpf_ops);
198	if (bpf_size != nla_len(tb[TCA_BPF_OPS]))
199		return -EINVAL;
200
201	bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
202	if (bpf_ops == NULL)
203		return -ENOMEM;
204
205	memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size);
206
207	fprog_tmp.len = bpf_num_ops;
208	fprog_tmp.filter = bpf_ops;
209
210	ret = bpf_prog_create(&fp, &fprog_tmp);
211	if (ret < 0) {
212		kfree(bpf_ops);
213		return ret;
214	}
215
216	prog->bpf_ops = bpf_ops;
217	prog->bpf_num_ops = bpf_num_ops;
218	prog->bpf_name = NULL;
219
220	prog->filter = fp;
221	prog->res.classid = classid;
222
223	return 0;
224}
225
226static int cls_bpf_prog_from_efd(struct nlattr **tb,
227				 struct cls_bpf_prog *prog, u32 classid)
228{
229	struct bpf_prog *fp;
230	char *name = NULL;
231	u32 bpf_fd;
232
233	bpf_fd = nla_get_u32(tb[TCA_BPF_FD]);
234
235	fp = bpf_prog_get(bpf_fd);
236	if (IS_ERR(fp))
237		return PTR_ERR(fp);
238
239	if (fp->type != BPF_PROG_TYPE_SCHED_CLS) {
240		bpf_prog_put(fp);
241		return -EINVAL;
242	}
243
244	if (tb[TCA_BPF_NAME]) {
245		name = kmemdup(nla_data(tb[TCA_BPF_NAME]),
246			       nla_len(tb[TCA_BPF_NAME]),
247			       GFP_KERNEL);
248		if (!name) {
249			bpf_prog_put(fp);
250			return -ENOMEM;
251		}
252	}
253
254	prog->bpf_ops = NULL;
255	prog->bpf_fd = bpf_fd;
256	prog->bpf_name = name;
257
258	prog->filter = fp;
259	prog->res.classid = classid;
260
261	return 0;
262}
263
264static int cls_bpf_modify_existing(struct net *net, struct tcf_proto *tp,
265				   struct cls_bpf_prog *prog,
266				   unsigned long base, struct nlattr **tb,
267				   struct nlattr *est, bool ovr)
268{
269	struct tcf_exts exts;
270	bool is_bpf, is_ebpf;
271	u32 classid;
272	int ret;
273
274	is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS];
275	is_ebpf = tb[TCA_BPF_FD];
276
277	if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf) ||
278	    !tb[TCA_BPF_CLASSID])
279		return -EINVAL;
280
281	tcf_exts_init(&exts, TCA_BPF_ACT, TCA_BPF_POLICE);
282	ret = tcf_exts_validate(net, tp, tb, est, &exts, ovr);
283	if (ret < 0)
284		return ret;
285
286	classid = nla_get_u32(tb[TCA_BPF_CLASSID]);
287
288	ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog, classid) :
289		       cls_bpf_prog_from_efd(tb, prog, classid);
290	if (ret < 0) {
291		tcf_exts_destroy(&exts);
292		return ret;
293	}
294
295	tcf_bind_filter(tp, &prog->res, base);
296	tcf_exts_change(tp, &prog->exts, &exts);
297
298	return 0;
299}
300
301static u32 cls_bpf_grab_new_handle(struct tcf_proto *tp,
302				   struct cls_bpf_head *head)
303{
304	unsigned int i = 0x80000000;
305	u32 handle;
306
307	do {
308		if (++head->hgen == 0x7FFFFFFF)
309			head->hgen = 1;
310	} while (--i > 0 && cls_bpf_get(tp, head->hgen));
311
312	if (unlikely(i == 0)) {
313		pr_err("Insufficient number of handles\n");
314		handle = 0;
315	} else {
316		handle = head->hgen;
317	}
318
319	return handle;
320}
321
322static int cls_bpf_change(struct net *net, struct sk_buff *in_skb,
323			  struct tcf_proto *tp, unsigned long base,
324			  u32 handle, struct nlattr **tca,
325			  unsigned long *arg, bool ovr)
326{
327	struct cls_bpf_head *head = rtnl_dereference(tp->root);
328	struct cls_bpf_prog *oldprog = (struct cls_bpf_prog *) *arg;
329	struct nlattr *tb[TCA_BPF_MAX + 1];
330	struct cls_bpf_prog *prog;
331	int ret;
332
333	if (tca[TCA_OPTIONS] == NULL)
334		return -EINVAL;
335
336	ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy);
337	if (ret < 0)
338		return ret;
339
340	prog = kzalloc(sizeof(*prog), GFP_KERNEL);
341	if (!prog)
342		return -ENOBUFS;
343
344	tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE);
345
346	if (oldprog) {
347		if (handle && oldprog->handle != handle) {
348			ret = -EINVAL;
349			goto errout;
350		}
351	}
352
353	if (handle == 0)
354		prog->handle = cls_bpf_grab_new_handle(tp, head);
355	else
356		prog->handle = handle;
357	if (prog->handle == 0) {
358		ret = -EINVAL;
359		goto errout;
360	}
361
362	ret = cls_bpf_modify_existing(net, tp, prog, base, tb, tca[TCA_RATE], ovr);
363	if (ret < 0)
364		goto errout;
365
366	if (oldprog) {
367		list_replace_rcu(&oldprog->link, &prog->link);
368		tcf_unbind_filter(tp, &oldprog->res);
369		call_rcu(&oldprog->rcu, __cls_bpf_delete_prog);
370	} else {
371		list_add_rcu(&prog->link, &head->plist);
372	}
373
374	*arg = (unsigned long) prog;
375	return 0;
376errout:
377	kfree(prog);
378
379	return ret;
380}
381
382static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog,
383				 struct sk_buff *skb)
384{
385	struct nlattr *nla;
386
387	if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops))
388		return -EMSGSIZE;
389
390	nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops *
391			  sizeof(struct sock_filter));
392	if (nla == NULL)
393		return -EMSGSIZE;
394
395	memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
396
397	return 0;
398}
399
400static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog,
401				  struct sk_buff *skb)
402{
403	if (nla_put_u32(skb, TCA_BPF_FD, prog->bpf_fd))
404		return -EMSGSIZE;
405
406	if (prog->bpf_name &&
407	    nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name))
408		return -EMSGSIZE;
409
410	return 0;
411}
412
413static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
414			struct sk_buff *skb, struct tcmsg *tm)
415{
416	struct cls_bpf_prog *prog = (struct cls_bpf_prog *) fh;
417	struct nlattr *nest;
418	int ret;
419
420	if (prog == NULL)
421		return skb->len;
422
423	tm->tcm_handle = prog->handle;
424
425	nest = nla_nest_start(skb, TCA_OPTIONS);
426	if (nest == NULL)
427		goto nla_put_failure;
428
429	if (nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid))
430		goto nla_put_failure;
431
432	if (cls_bpf_is_ebpf(prog))
433		ret = cls_bpf_dump_ebpf_info(prog, skb);
434	else
435		ret = cls_bpf_dump_bpf_info(prog, skb);
436	if (ret)
437		goto nla_put_failure;
438
439	if (tcf_exts_dump(skb, &prog->exts) < 0)
440		goto nla_put_failure;
441
442	nla_nest_end(skb, nest);
443
444	if (tcf_exts_dump_stats(skb, &prog->exts) < 0)
445		goto nla_put_failure;
446
447	return skb->len;
448
449nla_put_failure:
450	nla_nest_cancel(skb, nest);
451	return -1;
452}
453
454static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg)
455{
456	struct cls_bpf_head *head = rtnl_dereference(tp->root);
457	struct cls_bpf_prog *prog;
458
459	list_for_each_entry(prog, &head->plist, link) {
460		if (arg->count < arg->skip)
461			goto skip;
462		if (arg->fn(tp, (unsigned long) prog, arg) < 0) {
463			arg->stop = 1;
464			break;
465		}
466skip:
467		arg->count++;
468	}
469}
470
471static struct tcf_proto_ops cls_bpf_ops __read_mostly = {
472	.kind		=	"bpf",
473	.owner		=	THIS_MODULE,
474	.classify	=	cls_bpf_classify,
475	.init		=	cls_bpf_init,
476	.destroy	=	cls_bpf_destroy,
477	.get		=	cls_bpf_get,
478	.change		=	cls_bpf_change,
479	.delete		=	cls_bpf_delete,
480	.walk		=	cls_bpf_walk,
481	.dump		=	cls_bpf_dump,
482};
483
484static int __init cls_bpf_init_mod(void)
485{
486	return register_tcf_proto_ops(&cls_bpf_ops);
487}
488
489static void __exit cls_bpf_exit_mod(void)
490{
491	unregister_tcf_proto_ops(&cls_bpf_ops);
492}
493
494module_init(cls_bpf_init_mod);
495module_exit(cls_bpf_exit_mod);
496