2 * Berkeley Packet Filter based traffic classifier
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
8 * (C) 2013 Daniel Borkmann <dborkman@redhat.com>
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.
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>
21 #include <net/rtnetlink.h>
22 #include <net/pkt_cls.h>
25 MODULE_LICENSE("GPL");
26 MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
27 MODULE_DESCRIPTION("TC BPF based classifier");
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)
34 struct list_head plist;
40 struct bpf_prog *filter;
41 struct list_head link;
42 struct tcf_result res;
49 struct sock_filter *bpf_ops;
55 static const struct nla_policy bpf_policy[TCA_BPF_MAX + 1] = {
56 [TCA_BPF_CLASSID] = { .type = NLA_U32 },
57 [TCA_BPF_FLAGS] = { .type = NLA_U32 },
58 [TCA_BPF_FLAGS_GEN] = { .type = NLA_U32 },
59 [TCA_BPF_FD] = { .type = NLA_U32 },
60 [TCA_BPF_NAME] = { .type = NLA_NUL_STRING,
61 .len = CLS_BPF_NAME_LEN },
62 [TCA_BPF_OPS_LEN] = { .type = NLA_U16 },
63 [TCA_BPF_OPS] = { .type = NLA_BINARY,
64 .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
67 static int cls_bpf_exec_opcode(int code)
81 static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
82 struct tcf_result *res)
84 struct cls_bpf_head *head = rcu_dereference_bh(tp->root);
85 bool at_ingress = skb_at_tc_ingress(skb);
86 struct cls_bpf_prog *prog;
89 /* Needed here for accessing maps. */
91 list_for_each_entry_rcu(prog, &head->plist, link) {
94 qdisc_skb_cb(skb)->tc_classid = prog->res.classid;
96 if (tc_skip_sw(prog->gen_flags)) {
97 filter_res = prog->exts_integrated ? TC_ACT_UNSPEC : 0;
98 } else if (at_ingress) {
99 /* It is safe to push/pull even if skb_shared() */
100 __skb_push(skb, skb->mac_len);
101 bpf_compute_data_end(skb);
102 filter_res = BPF_PROG_RUN(prog->filter, skb);
103 __skb_pull(skb, skb->mac_len);
105 bpf_compute_data_end(skb);
106 filter_res = BPF_PROG_RUN(prog->filter, skb);
109 if (prog->exts_integrated) {
111 res->classid = TC_H_MAJ(prog->res.classid) |
112 qdisc_skb_cb(skb)->tc_classid;
114 ret = cls_bpf_exec_opcode(filter_res);
115 if (ret == TC_ACT_UNSPEC)
122 if (filter_res != -1) {
124 res->classid = filter_res;
129 ret = tcf_exts_exec(skb, &prog->exts, res);
140 static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog)
142 return !prog->bpf_ops;
145 static int cls_bpf_offload_cmd(struct tcf_proto *tp, struct cls_bpf_prog *prog,
146 enum tc_clsbpf_command cmd)
148 struct net_device *dev = tp->q->dev_queue->dev;
149 struct tc_cls_bpf_offload bpf_offload = {};
150 struct tc_to_netdev offload;
153 offload.type = TC_SETUP_CLSBPF;
154 offload.cls_bpf = &bpf_offload;
156 bpf_offload.command = cmd;
157 bpf_offload.exts = &prog->exts;
158 bpf_offload.prog = prog->filter;
159 bpf_offload.name = prog->bpf_name;
160 bpf_offload.exts_integrated = prog->exts_integrated;
161 bpf_offload.gen_flags = prog->gen_flags;
163 err = dev->netdev_ops->ndo_setup_tc(dev, tp->q->handle,
164 tp->protocol, &offload);
166 if (!err && (cmd == TC_CLSBPF_ADD || cmd == TC_CLSBPF_REPLACE))
167 prog->gen_flags |= TCA_CLS_FLAGS_IN_HW;
172 static int cls_bpf_offload(struct tcf_proto *tp, struct cls_bpf_prog *prog,
173 struct cls_bpf_prog *oldprog)
175 struct net_device *dev = tp->q->dev_queue->dev;
176 struct cls_bpf_prog *obj = prog;
177 enum tc_clsbpf_command cmd;
181 skip_sw = tc_skip_sw(prog->gen_flags) ||
182 (oldprog && tc_skip_sw(oldprog->gen_flags));
184 if (oldprog && oldprog->offloaded) {
185 if (tc_should_offload(dev, tp, prog->gen_flags)) {
186 cmd = TC_CLSBPF_REPLACE;
187 } else if (!tc_skip_sw(prog->gen_flags)) {
189 cmd = TC_CLSBPF_DESTROY;
194 if (!tc_should_offload(dev, tp, prog->gen_flags))
195 return skip_sw ? -EINVAL : 0;
199 ret = cls_bpf_offload_cmd(tp, obj, cmd);
201 return skip_sw ? ret : 0;
203 obj->offloaded = true;
205 oldprog->offloaded = false;
210 static void cls_bpf_stop_offload(struct tcf_proto *tp,
211 struct cls_bpf_prog *prog)
215 if (!prog->offloaded)
218 err = cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_DESTROY);
220 pr_err("Stopping hardware offload failed: %d\n", err);
224 prog->offloaded = false;
227 static void cls_bpf_offload_update_stats(struct tcf_proto *tp,
228 struct cls_bpf_prog *prog)
230 if (!prog->offloaded)
233 cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_STATS);
236 static int cls_bpf_init(struct tcf_proto *tp)
238 struct cls_bpf_head *head;
240 head = kzalloc(sizeof(*head), GFP_KERNEL);
244 INIT_LIST_HEAD_RCU(&head->plist);
245 rcu_assign_pointer(tp->root, head);
250 static void __cls_bpf_delete_prog(struct cls_bpf_prog *prog)
252 tcf_exts_destroy(&prog->exts);
254 if (cls_bpf_is_ebpf(prog))
255 bpf_prog_put(prog->filter);
257 bpf_prog_destroy(prog->filter);
259 kfree(prog->bpf_name);
260 kfree(prog->bpf_ops);
264 static void cls_bpf_delete_prog_rcu(struct rcu_head *rcu)
266 __cls_bpf_delete_prog(container_of(rcu, struct cls_bpf_prog, rcu));
269 static void __cls_bpf_delete(struct tcf_proto *tp, struct cls_bpf_prog *prog)
271 cls_bpf_stop_offload(tp, prog);
272 list_del_rcu(&prog->link);
273 tcf_unbind_filter(tp, &prog->res);
274 call_rcu(&prog->rcu, cls_bpf_delete_prog_rcu);
277 static int cls_bpf_delete(struct tcf_proto *tp, unsigned long arg, bool *last)
279 struct cls_bpf_head *head = rtnl_dereference(tp->root);
281 __cls_bpf_delete(tp, (struct cls_bpf_prog *) arg);
282 *last = list_empty(&head->plist);
286 static void cls_bpf_destroy(struct tcf_proto *tp)
288 struct cls_bpf_head *head = rtnl_dereference(tp->root);
289 struct cls_bpf_prog *prog, *tmp;
291 list_for_each_entry_safe(prog, tmp, &head->plist, link)
292 __cls_bpf_delete(tp, prog);
294 kfree_rcu(head, rcu);
297 static unsigned long cls_bpf_get(struct tcf_proto *tp, u32 handle)
299 struct cls_bpf_head *head = rtnl_dereference(tp->root);
300 struct cls_bpf_prog *prog;
301 unsigned long ret = 0UL;
303 list_for_each_entry(prog, &head->plist, link) {
304 if (prog->handle == handle) {
305 ret = (unsigned long) prog;
313 static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog)
315 struct sock_filter *bpf_ops;
316 struct sock_fprog_kern fprog_tmp;
318 u16 bpf_size, bpf_num_ops;
321 bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]);
322 if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
325 bpf_size = bpf_num_ops * sizeof(*bpf_ops);
326 if (bpf_size != nla_len(tb[TCA_BPF_OPS]))
329 bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
333 memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size);
335 fprog_tmp.len = bpf_num_ops;
336 fprog_tmp.filter = bpf_ops;
338 ret = bpf_prog_create(&fp, &fprog_tmp);
344 prog->bpf_ops = bpf_ops;
345 prog->bpf_num_ops = bpf_num_ops;
346 prog->bpf_name = NULL;
352 static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog,
353 const struct tcf_proto *tp)
359 bpf_fd = nla_get_u32(tb[TCA_BPF_FD]);
361 fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_CLS);
365 if (tb[TCA_BPF_NAME]) {
366 name = nla_memdup(tb[TCA_BPF_NAME], GFP_KERNEL);
373 prog->bpf_ops = NULL;
374 prog->bpf_name = name;
377 if (fp->dst_needed && !(tp->q->flags & TCQ_F_INGRESS))
378 netif_keep_dst(qdisc_dev(tp->q));
383 static int cls_bpf_modify_existing(struct net *net, struct tcf_proto *tp,
384 struct cls_bpf_prog *prog,
385 unsigned long base, struct nlattr **tb,
386 struct nlattr *est, bool ovr)
388 bool is_bpf, is_ebpf, have_exts = false;
389 struct tcf_exts exts;
393 is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS];
394 is_ebpf = tb[TCA_BPF_FD];
395 if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf))
398 ret = tcf_exts_init(&exts, TCA_BPF_ACT, TCA_BPF_POLICE);
401 ret = tcf_exts_validate(net, tp, tb, est, &exts, ovr);
405 if (tb[TCA_BPF_FLAGS]) {
406 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]);
408 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT) {
413 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT;
415 if (tb[TCA_BPF_FLAGS_GEN]) {
416 gen_flags = nla_get_u32(tb[TCA_BPF_FLAGS_GEN]);
417 if (gen_flags & ~CLS_BPF_SUPPORTED_GEN_FLAGS ||
418 !tc_flags_valid(gen_flags)) {
424 prog->exts_integrated = have_exts;
425 prog->gen_flags = gen_flags;
427 ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) :
428 cls_bpf_prog_from_efd(tb, prog, tp);
432 if (tb[TCA_BPF_CLASSID]) {
433 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]);
434 tcf_bind_filter(tp, &prog->res, base);
437 tcf_exts_change(tp, &prog->exts, &exts);
441 tcf_exts_destroy(&exts);
445 static u32 cls_bpf_grab_new_handle(struct tcf_proto *tp,
446 struct cls_bpf_head *head)
448 unsigned int i = 0x80000000;
452 if (++head->hgen == 0x7FFFFFFF)
454 } while (--i > 0 && cls_bpf_get(tp, head->hgen));
456 if (unlikely(i == 0)) {
457 pr_err("Insufficient number of handles\n");
466 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb,
467 struct tcf_proto *tp, unsigned long base,
468 u32 handle, struct nlattr **tca,
469 unsigned long *arg, bool ovr)
471 struct cls_bpf_head *head = rtnl_dereference(tp->root);
472 struct cls_bpf_prog *oldprog = (struct cls_bpf_prog *) *arg;
473 struct nlattr *tb[TCA_BPF_MAX + 1];
474 struct cls_bpf_prog *prog;
477 if (tca[TCA_OPTIONS] == NULL)
480 ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy,
485 prog = kzalloc(sizeof(*prog), GFP_KERNEL);
489 ret = tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE);
494 if (handle && oldprog->handle != handle) {
501 prog->handle = cls_bpf_grab_new_handle(tp, head);
503 prog->handle = handle;
504 if (prog->handle == 0) {
509 ret = cls_bpf_modify_existing(net, tp, prog, base, tb, tca[TCA_RATE],
514 ret = cls_bpf_offload(tp, prog, oldprog);
516 __cls_bpf_delete_prog(prog);
520 if (!tc_in_hw(prog->gen_flags))
521 prog->gen_flags |= TCA_CLS_FLAGS_NOT_IN_HW;
524 list_replace_rcu(&oldprog->link, &prog->link);
525 tcf_unbind_filter(tp, &oldprog->res);
526 call_rcu(&oldprog->rcu, cls_bpf_delete_prog_rcu);
528 list_add_rcu(&prog->link, &head->plist);
531 *arg = (unsigned long) prog;
535 tcf_exts_destroy(&prog->exts);
540 static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog,
545 if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops))
548 nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops *
549 sizeof(struct sock_filter));
553 memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
558 static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog,
563 if (prog->bpf_name &&
564 nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name))
567 nla = nla_reserve(skb, TCA_BPF_TAG, sizeof(prog->filter->tag));
571 memcpy(nla_data(nla), prog->filter->tag, nla_len(nla));
576 static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, unsigned long fh,
577 struct sk_buff *skb, struct tcmsg *tm)
579 struct cls_bpf_prog *prog = (struct cls_bpf_prog *) fh;
587 tm->tcm_handle = prog->handle;
589 cls_bpf_offload_update_stats(tp, prog);
591 nest = nla_nest_start(skb, TCA_OPTIONS);
593 goto nla_put_failure;
595 if (prog->res.classid &&
596 nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid))
597 goto nla_put_failure;
599 if (cls_bpf_is_ebpf(prog))
600 ret = cls_bpf_dump_ebpf_info(prog, skb);
602 ret = cls_bpf_dump_bpf_info(prog, skb);
604 goto nla_put_failure;
606 if (tcf_exts_dump(skb, &prog->exts) < 0)
607 goto nla_put_failure;
609 if (prog->exts_integrated)
610 bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT;
611 if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags))
612 goto nla_put_failure;
613 if (prog->gen_flags &&
614 nla_put_u32(skb, TCA_BPF_FLAGS_GEN, prog->gen_flags))
615 goto nla_put_failure;
617 nla_nest_end(skb, nest);
619 if (tcf_exts_dump_stats(skb, &prog->exts) < 0)
620 goto nla_put_failure;
625 nla_nest_cancel(skb, nest);
629 static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg)
631 struct cls_bpf_head *head = rtnl_dereference(tp->root);
632 struct cls_bpf_prog *prog;
634 list_for_each_entry(prog, &head->plist, link) {
635 if (arg->count < arg->skip)
637 if (arg->fn(tp, (unsigned long) prog, arg) < 0) {
646 static struct tcf_proto_ops cls_bpf_ops __read_mostly = {
648 .owner = THIS_MODULE,
649 .classify = cls_bpf_classify,
650 .init = cls_bpf_init,
651 .destroy = cls_bpf_destroy,
653 .change = cls_bpf_change,
654 .delete = cls_bpf_delete,
655 .walk = cls_bpf_walk,
656 .dump = cls_bpf_dump,
659 static int __init cls_bpf_init_mod(void)
661 return register_tcf_proto_ops(&cls_bpf_ops);
664 static void __exit cls_bpf_exit_mod(void)
666 unregister_tcf_proto_ops(&cls_bpf_ops);
669 module_init(cls_bpf_init_mod);
670 module_exit(cls_bpf_exit_mod);