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)
82 static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
83 struct tcf_result *res)
85 struct cls_bpf_head *head = rcu_dereference_bh(tp->root);
86 bool at_ingress = skb_at_tc_ingress(skb);
87 struct cls_bpf_prog *prog;
90 /* Needed here for accessing maps. */
92 list_for_each_entry_rcu(prog, &head->plist, link) {
95 qdisc_skb_cb(skb)->tc_classid = prog->res.classid;
97 if (tc_skip_sw(prog->gen_flags)) {
98 filter_res = prog->exts_integrated ? TC_ACT_UNSPEC : 0;
99 } else if (at_ingress) {
100 /* It is safe to push/pull even if skb_shared() */
101 __skb_push(skb, skb->mac_len);
102 bpf_compute_data_end(skb);
103 filter_res = BPF_PROG_RUN(prog->filter, skb);
104 __skb_pull(skb, skb->mac_len);
106 bpf_compute_data_end(skb);
107 filter_res = BPF_PROG_RUN(prog->filter, skb);
110 if (prog->exts_integrated) {
112 res->classid = TC_H_MAJ(prog->res.classid) |
113 qdisc_skb_cb(skb)->tc_classid;
115 ret = cls_bpf_exec_opcode(filter_res);
116 if (ret == TC_ACT_UNSPEC)
123 if (filter_res != -1) {
125 res->classid = filter_res;
130 ret = tcf_exts_exec(skb, &prog->exts, res);
141 static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog)
143 return !prog->bpf_ops;
146 static int cls_bpf_offload_cmd(struct tcf_proto *tp, struct cls_bpf_prog *prog,
147 enum tc_clsbpf_command cmd)
149 struct net_device *dev = tp->q->dev_queue->dev;
150 struct tc_cls_bpf_offload cls_bpf = {};
153 tc_cls_common_offload_init(&cls_bpf.common, tp);
154 cls_bpf.command = cmd;
155 cls_bpf.exts = &prog->exts;
156 cls_bpf.prog = prog->filter;
157 cls_bpf.name = prog->bpf_name;
158 cls_bpf.exts_integrated = prog->exts_integrated;
159 cls_bpf.gen_flags = prog->gen_flags;
161 err = dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_CLSBPF, &cls_bpf);
162 if (!err && (cmd == TC_CLSBPF_ADD || cmd == TC_CLSBPF_REPLACE))
163 prog->gen_flags |= TCA_CLS_FLAGS_IN_HW;
168 static int cls_bpf_offload(struct tcf_proto *tp, struct cls_bpf_prog *prog,
169 struct cls_bpf_prog *oldprog)
171 struct net_device *dev = tp->q->dev_queue->dev;
172 struct cls_bpf_prog *obj = prog;
173 enum tc_clsbpf_command cmd;
177 skip_sw = tc_skip_sw(prog->gen_flags) ||
178 (oldprog && tc_skip_sw(oldprog->gen_flags));
180 if (oldprog && oldprog->offloaded) {
181 if (tc_should_offload(dev, prog->gen_flags)) {
182 cmd = TC_CLSBPF_REPLACE;
183 } else if (!tc_skip_sw(prog->gen_flags)) {
185 cmd = TC_CLSBPF_DESTROY;
190 if (!tc_should_offload(dev, prog->gen_flags))
191 return skip_sw ? -EINVAL : 0;
195 ret = cls_bpf_offload_cmd(tp, obj, cmd);
197 return skip_sw ? ret : 0;
199 obj->offloaded = true;
201 oldprog->offloaded = false;
206 static void cls_bpf_stop_offload(struct tcf_proto *tp,
207 struct cls_bpf_prog *prog)
211 if (!prog->offloaded)
214 err = cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_DESTROY);
216 pr_err("Stopping hardware offload failed: %d\n", err);
220 prog->offloaded = false;
223 static void cls_bpf_offload_update_stats(struct tcf_proto *tp,
224 struct cls_bpf_prog *prog)
226 if (!prog->offloaded)
229 cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_STATS);
232 static int cls_bpf_init(struct tcf_proto *tp)
234 struct cls_bpf_head *head;
236 head = kzalloc(sizeof(*head), GFP_KERNEL);
240 INIT_LIST_HEAD_RCU(&head->plist);
241 rcu_assign_pointer(tp->root, head);
246 static void __cls_bpf_delete_prog(struct cls_bpf_prog *prog)
248 tcf_exts_destroy(&prog->exts);
250 if (cls_bpf_is_ebpf(prog))
251 bpf_prog_put(prog->filter);
253 bpf_prog_destroy(prog->filter);
255 kfree(prog->bpf_name);
256 kfree(prog->bpf_ops);
260 static void cls_bpf_delete_prog_rcu(struct rcu_head *rcu)
262 __cls_bpf_delete_prog(container_of(rcu, struct cls_bpf_prog, rcu));
265 static void __cls_bpf_delete(struct tcf_proto *tp, struct cls_bpf_prog *prog)
267 cls_bpf_stop_offload(tp, prog);
268 list_del_rcu(&prog->link);
269 tcf_unbind_filter(tp, &prog->res);
270 call_rcu(&prog->rcu, cls_bpf_delete_prog_rcu);
273 static int cls_bpf_delete(struct tcf_proto *tp, void *arg, bool *last)
275 struct cls_bpf_head *head = rtnl_dereference(tp->root);
277 __cls_bpf_delete(tp, arg);
278 *last = list_empty(&head->plist);
282 static void cls_bpf_destroy(struct tcf_proto *tp)
284 struct cls_bpf_head *head = rtnl_dereference(tp->root);
285 struct cls_bpf_prog *prog, *tmp;
287 list_for_each_entry_safe(prog, tmp, &head->plist, link)
288 __cls_bpf_delete(tp, prog);
290 kfree_rcu(head, rcu);
293 static void *cls_bpf_get(struct tcf_proto *tp, u32 handle)
295 struct cls_bpf_head *head = rtnl_dereference(tp->root);
296 struct cls_bpf_prog *prog;
298 list_for_each_entry(prog, &head->plist, link) {
299 if (prog->handle == handle)
306 static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog)
308 struct sock_filter *bpf_ops;
309 struct sock_fprog_kern fprog_tmp;
311 u16 bpf_size, bpf_num_ops;
314 bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]);
315 if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
318 bpf_size = bpf_num_ops * sizeof(*bpf_ops);
319 if (bpf_size != nla_len(tb[TCA_BPF_OPS]))
322 bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
326 memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size);
328 fprog_tmp.len = bpf_num_ops;
329 fprog_tmp.filter = bpf_ops;
331 ret = bpf_prog_create(&fp, &fprog_tmp);
337 prog->bpf_ops = bpf_ops;
338 prog->bpf_num_ops = bpf_num_ops;
339 prog->bpf_name = NULL;
345 static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog,
346 const struct tcf_proto *tp)
352 bpf_fd = nla_get_u32(tb[TCA_BPF_FD]);
354 fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_CLS);
358 if (tb[TCA_BPF_NAME]) {
359 name = nla_memdup(tb[TCA_BPF_NAME], GFP_KERNEL);
366 prog->bpf_ops = NULL;
367 prog->bpf_name = name;
370 if (fp->dst_needed && !(tp->q->flags & TCQ_F_INGRESS))
371 netif_keep_dst(qdisc_dev(tp->q));
376 static int cls_bpf_set_parms(struct net *net, struct tcf_proto *tp,
377 struct cls_bpf_prog *prog, unsigned long base,
378 struct nlattr **tb, struct nlattr *est, bool ovr)
380 bool is_bpf, is_ebpf, have_exts = false;
384 is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS];
385 is_ebpf = tb[TCA_BPF_FD];
386 if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf))
389 ret = tcf_exts_validate(net, tp, tb, est, &prog->exts, ovr);
393 if (tb[TCA_BPF_FLAGS]) {
394 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]);
396 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT)
399 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT;
401 if (tb[TCA_BPF_FLAGS_GEN]) {
402 gen_flags = nla_get_u32(tb[TCA_BPF_FLAGS_GEN]);
403 if (gen_flags & ~CLS_BPF_SUPPORTED_GEN_FLAGS ||
404 !tc_flags_valid(gen_flags))
408 prog->exts_integrated = have_exts;
409 prog->gen_flags = gen_flags;
411 ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) :
412 cls_bpf_prog_from_efd(tb, prog, tp);
416 if (tb[TCA_BPF_CLASSID]) {
417 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]);
418 tcf_bind_filter(tp, &prog->res, base);
424 static u32 cls_bpf_grab_new_handle(struct tcf_proto *tp,
425 struct cls_bpf_head *head)
427 unsigned int i = 0x80000000;
431 if (++head->hgen == 0x7FFFFFFF)
433 } while (--i > 0 && cls_bpf_get(tp, head->hgen));
435 if (unlikely(i == 0)) {
436 pr_err("Insufficient number of handles\n");
445 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb,
446 struct tcf_proto *tp, unsigned long base,
447 u32 handle, struct nlattr **tca,
448 void **arg, bool ovr)
450 struct cls_bpf_head *head = rtnl_dereference(tp->root);
451 struct cls_bpf_prog *oldprog = *arg;
452 struct nlattr *tb[TCA_BPF_MAX + 1];
453 struct cls_bpf_prog *prog;
456 if (tca[TCA_OPTIONS] == NULL)
459 ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy,
464 prog = kzalloc(sizeof(*prog), GFP_KERNEL);
468 ret = tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE);
473 if (handle && oldprog->handle != handle) {
480 prog->handle = cls_bpf_grab_new_handle(tp, head);
482 prog->handle = handle;
483 if (prog->handle == 0) {
488 ret = cls_bpf_set_parms(net, tp, prog, base, tb, tca[TCA_RATE], ovr);
492 ret = cls_bpf_offload(tp, prog, oldprog);
494 __cls_bpf_delete_prog(prog);
498 if (!tc_in_hw(prog->gen_flags))
499 prog->gen_flags |= TCA_CLS_FLAGS_NOT_IN_HW;
502 list_replace_rcu(&oldprog->link, &prog->link);
503 tcf_unbind_filter(tp, &oldprog->res);
504 call_rcu(&oldprog->rcu, cls_bpf_delete_prog_rcu);
506 list_add_rcu(&prog->link, &head->plist);
513 tcf_exts_destroy(&prog->exts);
518 static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog,
523 if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops))
526 nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops *
527 sizeof(struct sock_filter));
531 memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
536 static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog,
541 if (prog->bpf_name &&
542 nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name))
545 if (nla_put_u32(skb, TCA_BPF_ID, prog->filter->aux->id))
548 nla = nla_reserve(skb, TCA_BPF_TAG, sizeof(prog->filter->tag));
552 memcpy(nla_data(nla), prog->filter->tag, nla_len(nla));
557 static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, void *fh,
558 struct sk_buff *skb, struct tcmsg *tm)
560 struct cls_bpf_prog *prog = fh;
568 tm->tcm_handle = prog->handle;
570 cls_bpf_offload_update_stats(tp, prog);
572 nest = nla_nest_start(skb, TCA_OPTIONS);
574 goto nla_put_failure;
576 if (prog->res.classid &&
577 nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid))
578 goto nla_put_failure;
580 if (cls_bpf_is_ebpf(prog))
581 ret = cls_bpf_dump_ebpf_info(prog, skb);
583 ret = cls_bpf_dump_bpf_info(prog, skb);
585 goto nla_put_failure;
587 if (tcf_exts_dump(skb, &prog->exts) < 0)
588 goto nla_put_failure;
590 if (prog->exts_integrated)
591 bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT;
592 if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags))
593 goto nla_put_failure;
594 if (prog->gen_flags &&
595 nla_put_u32(skb, TCA_BPF_FLAGS_GEN, prog->gen_flags))
596 goto nla_put_failure;
598 nla_nest_end(skb, nest);
600 if (tcf_exts_dump_stats(skb, &prog->exts) < 0)
601 goto nla_put_failure;
606 nla_nest_cancel(skb, nest);
610 static void cls_bpf_bind_class(void *fh, u32 classid, unsigned long cl)
612 struct cls_bpf_prog *prog = fh;
614 if (prog && prog->res.classid == classid)
615 prog->res.class = cl;
618 static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg)
620 struct cls_bpf_head *head = rtnl_dereference(tp->root);
621 struct cls_bpf_prog *prog;
623 list_for_each_entry(prog, &head->plist, link) {
624 if (arg->count < arg->skip)
626 if (arg->fn(tp, prog, arg) < 0) {
635 static struct tcf_proto_ops cls_bpf_ops __read_mostly = {
637 .owner = THIS_MODULE,
638 .classify = cls_bpf_classify,
639 .init = cls_bpf_init,
640 .destroy = cls_bpf_destroy,
642 .change = cls_bpf_change,
643 .delete = cls_bpf_delete,
644 .walk = cls_bpf_walk,
645 .dump = cls_bpf_dump,
646 .bind_class = cls_bpf_bind_class,
649 static int __init cls_bpf_init_mod(void)
651 return register_tcf_proto_ops(&cls_bpf_ops);
654 static void __exit cls_bpf_exit_mod(void)
656 unregister_tcf_proto_ops(&cls_bpf_ops);
659 module_init(cls_bpf_init_mod);
660 module_exit(cls_bpf_exit_mod);