1
2
3
4
5
6
7
8
9
10
11
12
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#define CLS_BPF_SUPPORTED_GEN_FLAGS \
31 (TCA_CLS_FLAGS_SKIP_HW | TCA_CLS_FLAGS_SKIP_SW)
32
33struct cls_bpf_head {
34 struct list_head plist;
35 u32 hgen;
36 struct rcu_head rcu;
37};
38
39struct cls_bpf_prog {
40 struct bpf_prog *filter;
41 struct list_head link;
42 struct tcf_result res;
43 bool exts_integrated;
44 bool offloaded;
45 u32 gen_flags;
46 struct tcf_exts exts;
47 u32 handle;
48 union {
49 u32 bpf_fd;
50 u16 bpf_num_ops;
51 };
52 struct sock_filter *bpf_ops;
53 const char *bpf_name;
54 struct tcf_proto *tp;
55 struct rcu_head rcu;
56};
57
58static const struct nla_policy bpf_policy[TCA_BPF_MAX + 1] = {
59 [TCA_BPF_CLASSID] = { .type = NLA_U32 },
60 [TCA_BPF_FLAGS] = { .type = NLA_U32 },
61 [TCA_BPF_FLAGS_GEN] = { .type = NLA_U32 },
62 [TCA_BPF_FD] = { .type = NLA_U32 },
63 [TCA_BPF_NAME] = { .type = NLA_NUL_STRING,
64 .len = CLS_BPF_NAME_LEN },
65 [TCA_BPF_OPS_LEN] = { .type = NLA_U16 },
66 [TCA_BPF_OPS] = { .type = NLA_BINARY,
67 .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
68};
69
70static int cls_bpf_exec_opcode(int code)
71{
72 switch (code) {
73 case TC_ACT_OK:
74 case TC_ACT_SHOT:
75 case TC_ACT_STOLEN:
76 case TC_ACT_REDIRECT:
77 case TC_ACT_UNSPEC:
78 return code;
79 default:
80 return TC_ACT_UNSPEC;
81 }
82}
83
84static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
85 struct tcf_result *res)
86{
87 struct cls_bpf_head *head = rcu_dereference_bh(tp->root);
88 bool at_ingress = skb_at_tc_ingress(skb);
89 struct cls_bpf_prog *prog;
90 int ret = -1;
91
92
93 rcu_read_lock();
94 list_for_each_entry_rcu(prog, &head->plist, link) {
95 int filter_res;
96
97 qdisc_skb_cb(skb)->tc_classid = prog->res.classid;
98
99 if (tc_skip_sw(prog->gen_flags)) {
100 filter_res = prog->exts_integrated ? TC_ACT_UNSPEC : 0;
101 } else if (at_ingress) {
102
103 __skb_push(skb, skb->mac_len);
104 bpf_compute_data_end(skb);
105 filter_res = BPF_PROG_RUN(prog->filter, skb);
106 __skb_pull(skb, skb->mac_len);
107 } else {
108 bpf_compute_data_end(skb);
109 filter_res = BPF_PROG_RUN(prog->filter, skb);
110 }
111
112 if (prog->exts_integrated) {
113 res->class = 0;
114 res->classid = TC_H_MAJ(prog->res.classid) |
115 qdisc_skb_cb(skb)->tc_classid;
116
117 ret = cls_bpf_exec_opcode(filter_res);
118 if (ret == TC_ACT_UNSPEC)
119 continue;
120 break;
121 }
122
123 if (filter_res == 0)
124 continue;
125 if (filter_res != -1) {
126 res->class = 0;
127 res->classid = filter_res;
128 } else {
129 *res = prog->res;
130 }
131
132 ret = tcf_exts_exec(skb, &prog->exts, res);
133 if (ret < 0)
134 continue;
135
136 break;
137 }
138 rcu_read_unlock();
139
140 return ret;
141}
142
143static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog)
144{
145 return !prog->bpf_ops;
146}
147
148static int cls_bpf_offload_cmd(struct tcf_proto *tp, struct cls_bpf_prog *prog,
149 enum tc_clsbpf_command cmd)
150{
151 struct net_device *dev = tp->q->dev_queue->dev;
152 struct tc_cls_bpf_offload bpf_offload = {};
153 struct tc_to_netdev offload;
154
155 offload.type = TC_SETUP_CLSBPF;
156 offload.cls_bpf = &bpf_offload;
157
158 bpf_offload.command = cmd;
159 bpf_offload.exts = &prog->exts;
160 bpf_offload.prog = prog->filter;
161 bpf_offload.name = prog->bpf_name;
162 bpf_offload.exts_integrated = prog->exts_integrated;
163 bpf_offload.gen_flags = prog->gen_flags;
164
165 return dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle,
166 tp->protocol, &offload);
167}
168
169static int cls_bpf_offload(struct tcf_proto *tp, struct cls_bpf_prog *prog,
170 struct cls_bpf_prog *oldprog)
171{
172 struct net_device *dev = tp->q->dev_queue->dev;
173 struct cls_bpf_prog *obj = prog;
174 enum tc_clsbpf_command cmd;
175 bool skip_sw;
176 int ret;
177
178 skip_sw = tc_skip_sw(prog->gen_flags) ||
179 (oldprog && tc_skip_sw(oldprog->gen_flags));
180
181 if (oldprog && oldprog->offloaded) {
182 if (tc_should_offload(dev, tp, prog->gen_flags)) {
183 cmd = TC_CLSBPF_REPLACE;
184 } else if (!tc_skip_sw(prog->gen_flags)) {
185 obj = oldprog;
186 cmd = TC_CLSBPF_DESTROY;
187 } else {
188 return -EINVAL;
189 }
190 } else {
191 if (!tc_should_offload(dev, tp, prog->gen_flags))
192 return skip_sw ? -EINVAL : 0;
193 cmd = TC_CLSBPF_ADD;
194 }
195
196 ret = cls_bpf_offload_cmd(tp, obj, cmd);
197 if (ret)
198 return skip_sw ? ret : 0;
199
200 obj->offloaded = true;
201 if (oldprog)
202 oldprog->offloaded = false;
203
204 return 0;
205}
206
207static void cls_bpf_stop_offload(struct tcf_proto *tp,
208 struct cls_bpf_prog *prog)
209{
210 int err;
211
212 if (!prog->offloaded)
213 return;
214
215 err = cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_DESTROY);
216 if (err) {
217 pr_err("Stopping hardware offload failed: %d\n", err);
218 return;
219 }
220
221 prog->offloaded = false;
222}
223
224static void cls_bpf_offload_update_stats(struct tcf_proto *tp,
225 struct cls_bpf_prog *prog)
226{
227 if (!prog->offloaded)
228 return;
229
230 cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_STATS);
231}
232
233static int cls_bpf_init(struct tcf_proto *tp)
234{
235 struct cls_bpf_head *head;
236
237 head = kzalloc(sizeof(*head), GFP_KERNEL);
238 if (head == NULL)
239 return -ENOBUFS;
240
241 INIT_LIST_HEAD_RCU(&head->plist);
242 rcu_assign_pointer(tp->root, head);
243
244 return 0;
245}
246
247static void cls_bpf_delete_prog(struct tcf_proto *tp, struct cls_bpf_prog *prog)
248{
249 tcf_exts_destroy(&prog->exts);
250
251 if (cls_bpf_is_ebpf(prog))
252 bpf_prog_put(prog->filter);
253 else
254 bpf_prog_destroy(prog->filter);
255
256 kfree(prog->bpf_name);
257 kfree(prog->bpf_ops);
258 kfree(prog);
259}
260
261static void __cls_bpf_delete_prog(struct rcu_head *rcu)
262{
263 struct cls_bpf_prog *prog = container_of(rcu, struct cls_bpf_prog, rcu);
264
265 cls_bpf_delete_prog(prog->tp, prog);
266}
267
268static int cls_bpf_delete(struct tcf_proto *tp, unsigned long arg)
269{
270 struct cls_bpf_prog *prog = (struct cls_bpf_prog *) arg;
271
272 cls_bpf_stop_offload(tp, prog);
273 list_del_rcu(&prog->link);
274 tcf_unbind_filter(tp, &prog->res);
275 call_rcu(&prog->rcu, __cls_bpf_delete_prog);
276
277 return 0;
278}
279
280static bool cls_bpf_destroy(struct tcf_proto *tp, bool force)
281{
282 struct cls_bpf_head *head = rtnl_dereference(tp->root);
283 struct cls_bpf_prog *prog, *tmp;
284
285 if (!force && !list_empty(&head->plist))
286 return false;
287
288 list_for_each_entry_safe(prog, tmp, &head->plist, link) {
289 cls_bpf_stop_offload(tp, prog);
290 list_del_rcu(&prog->link);
291 tcf_unbind_filter(tp, &prog->res);
292 call_rcu(&prog->rcu, __cls_bpf_delete_prog);
293 }
294
295 kfree_rcu(head, rcu);
296 return true;
297}
298
299static unsigned long cls_bpf_get(struct tcf_proto *tp, u32 handle)
300{
301 struct cls_bpf_head *head = rtnl_dereference(tp->root);
302 struct cls_bpf_prog *prog;
303 unsigned long ret = 0UL;
304
305 list_for_each_entry(prog, &head->plist, link) {
306 if (prog->handle == handle) {
307 ret = (unsigned long) prog;
308 break;
309 }
310 }
311
312 return ret;
313}
314
315static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog)
316{
317 struct sock_filter *bpf_ops;
318 struct sock_fprog_kern fprog_tmp;
319 struct bpf_prog *fp;
320 u16 bpf_size, bpf_num_ops;
321 int ret;
322
323 bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]);
324 if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
325 return -EINVAL;
326
327 bpf_size = bpf_num_ops * sizeof(*bpf_ops);
328 if (bpf_size != nla_len(tb[TCA_BPF_OPS]))
329 return -EINVAL;
330
331 bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
332 if (bpf_ops == NULL)
333 return -ENOMEM;
334
335 memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size);
336
337 fprog_tmp.len = bpf_num_ops;
338 fprog_tmp.filter = bpf_ops;
339
340 ret = bpf_prog_create(&fp, &fprog_tmp);
341 if (ret < 0) {
342 kfree(bpf_ops);
343 return ret;
344 }
345
346 prog->bpf_ops = bpf_ops;
347 prog->bpf_num_ops = bpf_num_ops;
348 prog->bpf_name = NULL;
349 prog->filter = fp;
350
351 return 0;
352}
353
354static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog,
355 const struct tcf_proto *tp)
356{
357 struct bpf_prog *fp;
358 char *name = NULL;
359 u32 bpf_fd;
360
361 bpf_fd = nla_get_u32(tb[TCA_BPF_FD]);
362
363 fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_CLS);
364 if (IS_ERR(fp))
365 return PTR_ERR(fp);
366
367 if (tb[TCA_BPF_NAME]) {
368 name = kmemdup(nla_data(tb[TCA_BPF_NAME]),
369 nla_len(tb[TCA_BPF_NAME]),
370 GFP_KERNEL);
371 if (!name) {
372 bpf_prog_put(fp);
373 return -ENOMEM;
374 }
375 }
376
377 prog->bpf_ops = NULL;
378 prog->bpf_fd = bpf_fd;
379 prog->bpf_name = name;
380 prog->filter = fp;
381
382 if (fp->dst_needed && !(tp->q->flags & TCQ_F_INGRESS))
383 netif_keep_dst(qdisc_dev(tp->q));
384
385 return 0;
386}
387
388static int cls_bpf_modify_existing(struct net *net, struct tcf_proto *tp,
389 struct cls_bpf_prog *prog,
390 unsigned long base, struct nlattr **tb,
391 struct nlattr *est, bool ovr)
392{
393 bool is_bpf, is_ebpf, have_exts = false;
394 struct tcf_exts exts;
395 u32 gen_flags = 0;
396 int ret;
397
398 is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS];
399 is_ebpf = tb[TCA_BPF_FD];
400 if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf))
401 return -EINVAL;
402
403 ret = tcf_exts_init(&exts, TCA_BPF_ACT, TCA_BPF_POLICE);
404 if (ret < 0)
405 return ret;
406 ret = tcf_exts_validate(net, tp, tb, est, &exts, ovr);
407 if (ret < 0)
408 goto errout;
409
410 if (tb[TCA_BPF_FLAGS]) {
411 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]);
412
413 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT) {
414 ret = -EINVAL;
415 goto errout;
416 }
417
418 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT;
419 }
420 if (tb[TCA_BPF_FLAGS_GEN]) {
421 gen_flags = nla_get_u32(tb[TCA_BPF_FLAGS_GEN]);
422 if (gen_flags & ~CLS_BPF_SUPPORTED_GEN_FLAGS ||
423 !tc_flags_valid(gen_flags)) {
424 ret = -EINVAL;
425 goto errout;
426 }
427 }
428
429 prog->exts_integrated = have_exts;
430 prog->gen_flags = gen_flags;
431
432 ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) :
433 cls_bpf_prog_from_efd(tb, prog, tp);
434 if (ret < 0)
435 goto errout;
436
437 if (tb[TCA_BPF_CLASSID]) {
438 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]);
439 tcf_bind_filter(tp, &prog->res, base);
440 }
441
442 tcf_exts_change(tp, &prog->exts, &exts);
443 return 0;
444
445errout:
446 tcf_exts_destroy(&exts);
447 return ret;
448}
449
450static u32 cls_bpf_grab_new_handle(struct tcf_proto *tp,
451 struct cls_bpf_head *head)
452{
453 unsigned int i = 0x80000000;
454 u32 handle;
455
456 do {
457 if (++head->hgen == 0x7FFFFFFF)
458 head->hgen = 1;
459 } while (--i > 0 && cls_bpf_get(tp, head->hgen));
460
461 if (unlikely(i == 0)) {
462 pr_err("Insufficient number of handles\n");
463 handle = 0;
464 } else {
465 handle = head->hgen;
466 }
467
468 return handle;
469}
470
471static int cls_bpf_change(struct net *net, struct sk_buff *in_skb,
472 struct tcf_proto *tp, unsigned long base,
473 u32 handle, struct nlattr **tca,
474 unsigned long *arg, bool ovr)
475{
476 struct cls_bpf_head *head = rtnl_dereference(tp->root);
477 struct cls_bpf_prog *oldprog = (struct cls_bpf_prog *) *arg;
478 struct nlattr *tb[TCA_BPF_MAX + 1];
479 struct cls_bpf_prog *prog;
480 int ret;
481
482 if (tca[TCA_OPTIONS] == NULL)
483 return -EINVAL;
484
485 ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy);
486 if (ret < 0)
487 return ret;
488
489 prog = kzalloc(sizeof(*prog), GFP_KERNEL);
490 if (!prog)
491 return -ENOBUFS;
492
493 ret = tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE);
494 if (ret < 0)
495 goto errout;
496
497 if (oldprog) {
498 if (handle && oldprog->handle != handle) {
499 ret = -EINVAL;
500 goto errout;
501 }
502 }
503
504 if (handle == 0)
505 prog->handle = cls_bpf_grab_new_handle(tp, head);
506 else
507 prog->handle = handle;
508 if (prog->handle == 0) {
509 ret = -EINVAL;
510 goto errout;
511 }
512
513 ret = cls_bpf_modify_existing(net, tp, prog, base, tb, tca[TCA_RATE],
514 ovr);
515 if (ret < 0)
516 goto errout;
517
518 ret = cls_bpf_offload(tp, prog, oldprog);
519 if (ret) {
520 cls_bpf_delete_prog(tp, prog);
521 return ret;
522 }
523
524 if (oldprog) {
525 list_replace_rcu(&oldprog->link, &prog->link);
526 tcf_unbind_filter(tp, &oldprog->res);
527 call_rcu(&oldprog->rcu, __cls_bpf_delete_prog);
528 } else {
529 list_add_rcu(&prog->link, &head->plist);
530 }
531
532 *arg = (unsigned long) prog;
533 return 0;
534
535errout:
536 tcf_exts_destroy(&prog->exts);
537 kfree(prog);
538 return ret;
539}
540
541static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog,
542 struct sk_buff *skb)
543{
544 struct nlattr *nla;
545
546 if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops))
547 return -EMSGSIZE;
548
549 nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops *
550 sizeof(struct sock_filter));
551 if (nla == NULL)
552 return -EMSGSIZE;
553
554 memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
555
556 return 0;
557}
558
559static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog,
560 struct sk_buff *skb)
561{
562 if (nla_put_u32(skb, TCA_BPF_FD, prog->bpf_fd))
563 return -EMSGSIZE;
564
565 if (prog->bpf_name &&
566 nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name))
567 return -EMSGSIZE;
568
569 return 0;
570}
571
572static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
573 struct sk_buff *skb, struct tcmsg *tm)
574{
575 struct cls_bpf_prog *prog = (struct cls_bpf_prog *) fh;
576 struct nlattr *nest;
577 u32 bpf_flags = 0;
578 int ret;
579
580 if (prog == NULL)
581 return skb->len;
582
583 tm->tcm_handle = prog->handle;
584
585 cls_bpf_offload_update_stats(tp, prog);
586
587 nest = nla_nest_start(skb, TCA_OPTIONS);
588 if (nest == NULL)
589 goto nla_put_failure;
590
591 if (prog->res.classid &&
592 nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid))
593 goto nla_put_failure;
594
595 if (cls_bpf_is_ebpf(prog))
596 ret = cls_bpf_dump_ebpf_info(prog, skb);
597 else
598 ret = cls_bpf_dump_bpf_info(prog, skb);
599 if (ret)
600 goto nla_put_failure;
601
602 if (tcf_exts_dump(skb, &prog->exts) < 0)
603 goto nla_put_failure;
604
605 if (prog->exts_integrated)
606 bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT;
607 if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags))
608 goto nla_put_failure;
609 if (prog->gen_flags &&
610 nla_put_u32(skb, TCA_BPF_FLAGS_GEN, prog->gen_flags))
611 goto nla_put_failure;
612
613 nla_nest_end(skb, nest);
614
615 if (tcf_exts_dump_stats(skb, &prog->exts) < 0)
616 goto nla_put_failure;
617
618 return skb->len;
619
620nla_put_failure:
621 nla_nest_cancel(skb, nest);
622 return -1;
623}
624
625static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg)
626{
627 struct cls_bpf_head *head = rtnl_dereference(tp->root);
628 struct cls_bpf_prog *prog;
629
630 list_for_each_entry(prog, &head->plist, link) {
631 if (arg->count < arg->skip)
632 goto skip;
633 if (arg->fn(tp, (unsigned long) prog, arg) < 0) {
634 arg->stop = 1;
635 break;
636 }
637skip:
638 arg->count++;
639 }
640}
641
642static struct tcf_proto_ops cls_bpf_ops __read_mostly = {
643 .kind = "bpf",
644 .owner = THIS_MODULE,
645 .classify = cls_bpf_classify,
646 .init = cls_bpf_init,
647 .destroy = cls_bpf_destroy,
648 .get = cls_bpf_get,
649 .change = cls_bpf_change,
650 .delete = cls_bpf_delete,
651 .walk = cls_bpf_walk,
652 .dump = cls_bpf_dump,
653};
654
655static int __init cls_bpf_init_mod(void)
656{
657 return register_tcf_proto_ops(&cls_bpf_ops);
658}
659
660static void __exit cls_bpf_exit_mod(void)
661{
662 unregister_tcf_proto_ops(&cls_bpf_ops);
663}
664
665module_init(cls_bpf_init_mod);
666module_exit(cls_bpf_exit_mod);
667
