]>
Commit | Line | Data |
---|---|---|
d23b8ad8 JP |
1 | /* |
2 | * Copyright (c) 2015 Jiri Pirko <jiri@resnulli.us> | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or modify | |
5 | * it under the terms of the GNU General Public License as published by | |
6 | * the Free Software Foundation; either version 2 of the License, or | |
7 | * (at your option) any later version. | |
8 | */ | |
9 | ||
10 | #include <linux/module.h> | |
11 | #include <linux/init.h> | |
12 | #include <linux/kernel.h> | |
13 | #include <linux/skbuff.h> | |
14 | #include <linux/rtnetlink.h> | |
15 | #include <linux/filter.h> | |
a8cb5f55 DB |
16 | #include <linux/bpf.h> |
17 | ||
d23b8ad8 JP |
18 | #include <net/netlink.h> |
19 | #include <net/pkt_sched.h> | |
20 | ||
21 | #include <linux/tc_act/tc_bpf.h> | |
22 | #include <net/tc_act/tc_bpf.h> | |
23 | ||
a8cb5f55 DB |
24 | #define BPF_TAB_MASK 15 |
25 | #define ACT_BPF_NAME_LEN 256 | |
26 | ||
27 | struct tcf_bpf_cfg { | |
28 | struct bpf_prog *filter; | |
29 | struct sock_filter *bpf_ops; | |
f4eaed28 | 30 | const char *bpf_name; |
a8cb5f55 | 31 | u16 bpf_num_ops; |
f4eaed28 | 32 | bool is_ebpf; |
a8cb5f55 | 33 | }; |
d23b8ad8 | 34 | |
c7d03a00 | 35 | static unsigned int bpf_net_id; |
a85a970a | 36 | static struct tc_action_ops act_bpf_ops; |
ddf97ccd | 37 | |
a8cb5f55 | 38 | static int tcf_bpf(struct sk_buff *skb, const struct tc_action *act, |
d23b8ad8 JP |
39 | struct tcf_result *res) |
40 | { | |
f53d8c7b | 41 | bool at_ingress = skb_at_tc_ingress(skb); |
a85a970a | 42 | struct tcf_bpf *prog = to_bpf(act); |
cff82457 | 43 | struct bpf_prog *filter; |
ced585c8 | 44 | int action, filter_res; |
d23b8ad8 | 45 | |
cff82457 AS |
46 | tcf_lastuse_update(&prog->tcf_tm); |
47 | bstats_cpu_update(this_cpu_ptr(prog->common.cpu_bstats), skb); | |
d23b8ad8 | 48 | |
a8cb5f55 | 49 | rcu_read_lock(); |
cff82457 | 50 | filter = rcu_dereference(prog->filter); |
3431205e AS |
51 | if (at_ingress) { |
52 | __skb_push(skb, skb->mac_len); | |
db58ba45 | 53 | bpf_compute_data_end(skb); |
cff82457 | 54 | filter_res = BPF_PROG_RUN(filter, skb); |
3431205e AS |
55 | __skb_pull(skb, skb->mac_len); |
56 | } else { | |
db58ba45 | 57 | bpf_compute_data_end(skb); |
cff82457 | 58 | filter_res = BPF_PROG_RUN(filter, skb); |
3431205e | 59 | } |
a8cb5f55 | 60 | rcu_read_unlock(); |
ced585c8 DB |
61 | |
62 | /* A BPF program may overwrite the default action opcode. | |
63 | * Similarly as in cls_bpf, if filter_res == -1 we use the | |
64 | * default action specified from tc. | |
65 | * | |
66 | * In case a different well-known TC_ACT opcode has been | |
67 | * returned, it will overwrite the default one. | |
68 | * | |
69 | * For everything else that is unkown, TC_ACT_UNSPEC is | |
70 | * returned. | |
71 | */ | |
72 | switch (filter_res) { | |
73 | case TC_ACT_PIPE: | |
74 | case TC_ACT_RECLASSIFY: | |
75 | case TC_ACT_OK: | |
27b29f63 | 76 | case TC_ACT_REDIRECT: |
ced585c8 DB |
77 | action = filter_res; |
78 | break; | |
79 | case TC_ACT_SHOT: | |
80 | action = filter_res; | |
cff82457 | 81 | qstats_drop_inc(this_cpu_ptr(prog->common.cpu_qstats)); |
ced585c8 DB |
82 | break; |
83 | case TC_ACT_UNSPEC: | |
a8cb5f55 | 84 | action = prog->tcf_action; |
ced585c8 DB |
85 | break; |
86 | default: | |
87 | action = TC_ACT_UNSPEC; | |
88 | break; | |
d23b8ad8 JP |
89 | } |
90 | ||
d23b8ad8 JP |
91 | return action; |
92 | } | |
93 | ||
a8cb5f55 DB |
94 | static bool tcf_bpf_is_ebpf(const struct tcf_bpf *prog) |
95 | { | |
96 | return !prog->bpf_ops; | |
97 | } | |
98 | ||
99 | static int tcf_bpf_dump_bpf_info(const struct tcf_bpf *prog, | |
100 | struct sk_buff *skb) | |
101 | { | |
102 | struct nlattr *nla; | |
103 | ||
104 | if (nla_put_u16(skb, TCA_ACT_BPF_OPS_LEN, prog->bpf_num_ops)) | |
105 | return -EMSGSIZE; | |
106 | ||
107 | nla = nla_reserve(skb, TCA_ACT_BPF_OPS, prog->bpf_num_ops * | |
108 | sizeof(struct sock_filter)); | |
109 | if (nla == NULL) | |
110 | return -EMSGSIZE; | |
111 | ||
112 | memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla)); | |
113 | ||
114 | return 0; | |
115 | } | |
116 | ||
117 | static int tcf_bpf_dump_ebpf_info(const struct tcf_bpf *prog, | |
118 | struct sk_buff *skb) | |
119 | { | |
7bd509e3 DB |
120 | struct nlattr *nla; |
121 | ||
a8cb5f55 DB |
122 | if (prog->bpf_name && |
123 | nla_put_string(skb, TCA_ACT_BPF_NAME, prog->bpf_name)) | |
124 | return -EMSGSIZE; | |
125 | ||
e8628307 DB |
126 | if (nla_put_u32(skb, TCA_ACT_BPF_ID, prog->filter->aux->id)) |
127 | return -EMSGSIZE; | |
128 | ||
f1f7714e | 129 | nla = nla_reserve(skb, TCA_ACT_BPF_TAG, sizeof(prog->filter->tag)); |
7bd509e3 DB |
130 | if (nla == NULL) |
131 | return -EMSGSIZE; | |
132 | ||
f1f7714e | 133 | memcpy(nla_data(nla), prog->filter->tag, nla_len(nla)); |
7bd509e3 | 134 | |
a8cb5f55 DB |
135 | return 0; |
136 | } | |
137 | ||
138 | static int tcf_bpf_dump(struct sk_buff *skb, struct tc_action *act, | |
d23b8ad8 JP |
139 | int bind, int ref) |
140 | { | |
141 | unsigned char *tp = skb_tail_pointer(skb); | |
a85a970a | 142 | struct tcf_bpf *prog = to_bpf(act); |
d23b8ad8 | 143 | struct tc_act_bpf opt = { |
a8cb5f55 DB |
144 | .index = prog->tcf_index, |
145 | .refcnt = prog->tcf_refcnt - ref, | |
146 | .bindcnt = prog->tcf_bindcnt - bind, | |
147 | .action = prog->tcf_action, | |
d23b8ad8 | 148 | }; |
a8cb5f55 DB |
149 | struct tcf_t tm; |
150 | int ret; | |
d23b8ad8 JP |
151 | |
152 | if (nla_put(skb, TCA_ACT_BPF_PARMS, sizeof(opt), &opt)) | |
153 | goto nla_put_failure; | |
154 | ||
a8cb5f55 DB |
155 | if (tcf_bpf_is_ebpf(prog)) |
156 | ret = tcf_bpf_dump_ebpf_info(prog, skb); | |
157 | else | |
158 | ret = tcf_bpf_dump_bpf_info(prog, skb); | |
159 | if (ret) | |
d23b8ad8 JP |
160 | goto nla_put_failure; |
161 | ||
48d8ee16 | 162 | tcf_tm_dump(&tm, &prog->tcf_tm); |
9854518e ND |
163 | if (nla_put_64bit(skb, TCA_ACT_BPF_TM, sizeof(tm), &tm, |
164 | TCA_ACT_BPF_PAD)) | |
d23b8ad8 | 165 | goto nla_put_failure; |
a8cb5f55 | 166 | |
d23b8ad8 JP |
167 | return skb->len; |
168 | ||
169 | nla_put_failure: | |
170 | nlmsg_trim(skb, tp); | |
171 | return -1; | |
172 | } | |
173 | ||
174 | static const struct nla_policy act_bpf_policy[TCA_ACT_BPF_MAX + 1] = { | |
175 | [TCA_ACT_BPF_PARMS] = { .len = sizeof(struct tc_act_bpf) }, | |
a8cb5f55 | 176 | [TCA_ACT_BPF_FD] = { .type = NLA_U32 }, |
0b0f43fe JHS |
177 | [TCA_ACT_BPF_NAME] = { .type = NLA_NUL_STRING, |
178 | .len = ACT_BPF_NAME_LEN }, | |
d23b8ad8 JP |
179 | [TCA_ACT_BPF_OPS_LEN] = { .type = NLA_U16 }, |
180 | [TCA_ACT_BPF_OPS] = { .type = NLA_BINARY, | |
181 | .len = sizeof(struct sock_filter) * BPF_MAXINSNS }, | |
182 | }; | |
183 | ||
a8cb5f55 | 184 | static int tcf_bpf_init_from_ops(struct nlattr **tb, struct tcf_bpf_cfg *cfg) |
d23b8ad8 | 185 | { |
d23b8ad8 | 186 | struct sock_filter *bpf_ops; |
a8cb5f55 | 187 | struct sock_fprog_kern fprog_tmp; |
d23b8ad8 | 188 | struct bpf_prog *fp; |
a8cb5f55 | 189 | u16 bpf_size, bpf_num_ops; |
d23b8ad8 JP |
190 | int ret; |
191 | ||
d23b8ad8 JP |
192 | bpf_num_ops = nla_get_u16(tb[TCA_ACT_BPF_OPS_LEN]); |
193 | if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0) | |
194 | return -EINVAL; | |
195 | ||
196 | bpf_size = bpf_num_ops * sizeof(*bpf_ops); | |
fd3e646c DB |
197 | if (bpf_size != nla_len(tb[TCA_ACT_BPF_OPS])) |
198 | return -EINVAL; | |
199 | ||
d23b8ad8 | 200 | bpf_ops = kzalloc(bpf_size, GFP_KERNEL); |
a8cb5f55 | 201 | if (bpf_ops == NULL) |
d23b8ad8 JP |
202 | return -ENOMEM; |
203 | ||
204 | memcpy(bpf_ops, nla_data(tb[TCA_ACT_BPF_OPS]), bpf_size); | |
205 | ||
a8cb5f55 DB |
206 | fprog_tmp.len = bpf_num_ops; |
207 | fprog_tmp.filter = bpf_ops; | |
d23b8ad8 | 208 | |
a8cb5f55 DB |
209 | ret = bpf_prog_create(&fp, &fprog_tmp); |
210 | if (ret < 0) { | |
211 | kfree(bpf_ops); | |
212 | return ret; | |
213 | } | |
d23b8ad8 | 214 | |
a8cb5f55 DB |
215 | cfg->bpf_ops = bpf_ops; |
216 | cfg->bpf_num_ops = bpf_num_ops; | |
217 | cfg->filter = fp; | |
f4eaed28 | 218 | cfg->is_ebpf = false; |
a8cb5f55 DB |
219 | |
220 | return 0; | |
221 | } | |
222 | ||
223 | static int tcf_bpf_init_from_efd(struct nlattr **tb, struct tcf_bpf_cfg *cfg) | |
224 | { | |
225 | struct bpf_prog *fp; | |
226 | char *name = NULL; | |
227 | u32 bpf_fd; | |
228 | ||
229 | bpf_fd = nla_get_u32(tb[TCA_ACT_BPF_FD]); | |
230 | ||
113214be | 231 | fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_ACT); |
a8cb5f55 DB |
232 | if (IS_ERR(fp)) |
233 | return PTR_ERR(fp); | |
234 | ||
a8cb5f55 | 235 | if (tb[TCA_ACT_BPF_NAME]) { |
b15ca182 | 236 | name = nla_memdup(tb[TCA_ACT_BPF_NAME], GFP_KERNEL); |
a8cb5f55 DB |
237 | if (!name) { |
238 | bpf_prog_put(fp); | |
239 | return -ENOMEM; | |
240 | } | |
241 | } | |
242 | ||
a8cb5f55 DB |
243 | cfg->bpf_name = name; |
244 | cfg->filter = fp; | |
f4eaed28 | 245 | cfg->is_ebpf = true; |
a8cb5f55 DB |
246 | |
247 | return 0; | |
248 | } | |
249 | ||
f4eaed28 DB |
250 | static void tcf_bpf_cfg_cleanup(const struct tcf_bpf_cfg *cfg) |
251 | { | |
252 | if (cfg->is_ebpf) | |
253 | bpf_prog_put(cfg->filter); | |
254 | else | |
255 | bpf_prog_destroy(cfg->filter); | |
256 | ||
257 | kfree(cfg->bpf_ops); | |
258 | kfree(cfg->bpf_name); | |
259 | } | |
260 | ||
261 | static void tcf_bpf_prog_fill_cfg(const struct tcf_bpf *prog, | |
262 | struct tcf_bpf_cfg *cfg) | |
263 | { | |
264 | cfg->is_ebpf = tcf_bpf_is_ebpf(prog); | |
cff82457 AS |
265 | /* updates to prog->filter are prevented, since it's called either |
266 | * with rtnl lock or during final cleanup in rcu callback | |
267 | */ | |
268 | cfg->filter = rcu_dereference_protected(prog->filter, 1); | |
f4eaed28 DB |
269 | |
270 | cfg->bpf_ops = prog->bpf_ops; | |
271 | cfg->bpf_name = prog->bpf_name; | |
272 | } | |
273 | ||
a8cb5f55 | 274 | static int tcf_bpf_init(struct net *net, struct nlattr *nla, |
a85a970a | 275 | struct nlattr *est, struct tc_action **act, |
a8cb5f55 DB |
276 | int replace, int bind) |
277 | { | |
ddf97ccd | 278 | struct tc_action_net *tn = net_generic(net, bpf_net_id); |
a8cb5f55 | 279 | struct nlattr *tb[TCA_ACT_BPF_MAX + 1]; |
f4eaed28 | 280 | struct tcf_bpf_cfg cfg, old; |
a8cb5f55 DB |
281 | struct tc_act_bpf *parm; |
282 | struct tcf_bpf *prog; | |
a8cb5f55 | 283 | bool is_bpf, is_ebpf; |
a5c90b29 | 284 | int ret, res = 0; |
a8cb5f55 DB |
285 | |
286 | if (!nla) | |
287 | return -EINVAL; | |
288 | ||
fceb6435 | 289 | ret = nla_parse_nested(tb, TCA_ACT_BPF_MAX, nla, act_bpf_policy, NULL); |
a8cb5f55 DB |
290 | if (ret < 0) |
291 | return ret; | |
292 | ||
a5c90b29 | 293 | if (!tb[TCA_ACT_BPF_PARMS]) |
a8cb5f55 DB |
294 | return -EINVAL; |
295 | ||
296 | parm = nla_data(tb[TCA_ACT_BPF_PARMS]); | |
297 | ||
ddf97ccd WC |
298 | if (!tcf_hash_check(tn, parm->index, act, bind)) { |
299 | ret = tcf_hash_create(tn, parm->index, est, act, | |
a85a970a | 300 | &act_bpf_ops, bind, true); |
a8cb5f55 | 301 | if (ret < 0) |
a5c90b29 | 302 | return ret; |
d23b8ad8 | 303 | |
a5c90b29 | 304 | res = ACT_P_CREATED; |
d23b8ad8 | 305 | } else { |
a8cb5f55 | 306 | /* Don't override defaults. */ |
d23b8ad8 | 307 | if (bind) |
a5c90b29 | 308 | return 0; |
a8cb5f55 | 309 | |
a85a970a | 310 | tcf_hash_release(*act, bind); |
a5c90b29 DB |
311 | if (!replace) |
312 | return -EEXIST; | |
d23b8ad8 JP |
313 | } |
314 | ||
a5c90b29 DB |
315 | is_bpf = tb[TCA_ACT_BPF_OPS_LEN] && tb[TCA_ACT_BPF_OPS]; |
316 | is_ebpf = tb[TCA_ACT_BPF_FD]; | |
317 | ||
318 | if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf)) { | |
319 | ret = -EINVAL; | |
320 | goto out; | |
321 | } | |
322 | ||
323 | memset(&cfg, 0, sizeof(cfg)); | |
324 | ||
325 | ret = is_bpf ? tcf_bpf_init_from_ops(tb, &cfg) : | |
326 | tcf_bpf_init_from_efd(tb, &cfg); | |
327 | if (ret < 0) | |
328 | goto out; | |
329 | ||
a85a970a | 330 | prog = to_bpf(*act); |
cff82457 | 331 | ASSERT_RTNL(); |
a8cb5f55 | 332 | |
faa54be4 | 333 | if (res != ACT_P_CREATED) |
f4eaed28 DB |
334 | tcf_bpf_prog_fill_cfg(prog, &old); |
335 | ||
a8cb5f55 DB |
336 | prog->bpf_ops = cfg.bpf_ops; |
337 | prog->bpf_name = cfg.bpf_name; | |
338 | ||
339 | if (cfg.bpf_num_ops) | |
340 | prog->bpf_num_ops = cfg.bpf_num_ops; | |
a8cb5f55 DB |
341 | |
342 | prog->tcf_action = parm->action; | |
cff82457 | 343 | rcu_assign_pointer(prog->filter, cfg.filter); |
d23b8ad8 | 344 | |
cff82457 | 345 | if (res == ACT_P_CREATED) { |
a85a970a | 346 | tcf_hash_insert(tn, *act); |
cff82457 AS |
347 | } else { |
348 | /* make sure the program being replaced is no longer executing */ | |
349 | synchronize_rcu(); | |
f4eaed28 | 350 | tcf_bpf_cfg_cleanup(&old); |
cff82457 | 351 | } |
a8cb5f55 | 352 | |
a5c90b29 DB |
353 | return res; |
354 | out: | |
355 | if (res == ACT_P_CREATED) | |
a85a970a | 356 | tcf_hash_cleanup(*act, est); |
d23b8ad8 | 357 | |
d23b8ad8 JP |
358 | return ret; |
359 | } | |
360 | ||
a8cb5f55 | 361 | static void tcf_bpf_cleanup(struct tc_action *act, int bind) |
d23b8ad8 | 362 | { |
f4eaed28 | 363 | struct tcf_bpf_cfg tmp; |
ddf06c1e | 364 | |
a85a970a | 365 | tcf_bpf_prog_fill_cfg(to_bpf(act), &tmp); |
f4eaed28 | 366 | tcf_bpf_cfg_cleanup(&tmp); |
d23b8ad8 JP |
367 | } |
368 | ||
ddf97ccd WC |
369 | static int tcf_bpf_walker(struct net *net, struct sk_buff *skb, |
370 | struct netlink_callback *cb, int type, | |
a85a970a | 371 | const struct tc_action_ops *ops) |
ddf97ccd WC |
372 | { |
373 | struct tc_action_net *tn = net_generic(net, bpf_net_id); | |
374 | ||
a85a970a | 375 | return tcf_generic_walker(tn, skb, cb, type, ops); |
ddf97ccd WC |
376 | } |
377 | ||
a85a970a | 378 | static int tcf_bpf_search(struct net *net, struct tc_action **a, u32 index) |
ddf97ccd WC |
379 | { |
380 | struct tc_action_net *tn = net_generic(net, bpf_net_id); | |
381 | ||
382 | return tcf_hash_search(tn, a, index); | |
383 | } | |
384 | ||
a8cb5f55 DB |
385 | static struct tc_action_ops act_bpf_ops __read_mostly = { |
386 | .kind = "bpf", | |
387 | .type = TCA_ACT_BPF, | |
388 | .owner = THIS_MODULE, | |
389 | .act = tcf_bpf, | |
390 | .dump = tcf_bpf_dump, | |
391 | .cleanup = tcf_bpf_cleanup, | |
392 | .init = tcf_bpf_init, | |
ddf97ccd WC |
393 | .walk = tcf_bpf_walker, |
394 | .lookup = tcf_bpf_search, | |
a85a970a | 395 | .size = sizeof(struct tcf_bpf), |
ddf97ccd WC |
396 | }; |
397 | ||
398 | static __net_init int bpf_init_net(struct net *net) | |
399 | { | |
400 | struct tc_action_net *tn = net_generic(net, bpf_net_id); | |
401 | ||
402 | return tc_action_net_init(tn, &act_bpf_ops, BPF_TAB_MASK); | |
403 | } | |
404 | ||
405 | static void __net_exit bpf_exit_net(struct net *net) | |
406 | { | |
407 | struct tc_action_net *tn = net_generic(net, bpf_net_id); | |
408 | ||
409 | tc_action_net_exit(tn); | |
410 | } | |
411 | ||
412 | static struct pernet_operations bpf_net_ops = { | |
413 | .init = bpf_init_net, | |
414 | .exit = bpf_exit_net, | |
415 | .id = &bpf_net_id, | |
416 | .size = sizeof(struct tc_action_net), | |
d23b8ad8 JP |
417 | }; |
418 | ||
419 | static int __init bpf_init_module(void) | |
420 | { | |
ddf97ccd | 421 | return tcf_register_action(&act_bpf_ops, &bpf_net_ops); |
d23b8ad8 JP |
422 | } |
423 | ||
424 | static void __exit bpf_cleanup_module(void) | |
425 | { | |
ddf97ccd | 426 | tcf_unregister_action(&act_bpf_ops, &bpf_net_ops); |
d23b8ad8 JP |
427 | } |
428 | ||
429 | module_init(bpf_init_module); | |
430 | module_exit(bpf_cleanup_module); | |
431 | ||
432 | MODULE_AUTHOR("Jiri Pirko <jiri@resnulli.us>"); | |
433 | MODULE_DESCRIPTION("TC BPF based action"); | |
434 | MODULE_LICENSE("GPL v2"); |