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>
20 #include <linux/idr.h>
22 #include <net/rtnetlink.h>
23 #include <net/pkt_cls.h>
26 MODULE_LICENSE("GPL");
27 MODULE_AUTHOR("Daniel Borkmann <dborkman@redhat.com>");
28 MODULE_DESCRIPTION("TC BPF based classifier");
30 #define CLS_BPF_NAME_LEN 256
31 #define CLS_BPF_SUPPORTED_GEN_FLAGS \
32 (TCA_CLS_FLAGS_SKIP_HW | TCA_CLS_FLAGS_SKIP_SW)
35 struct list_head plist;
36 struct idr handle_idr;
41 struct bpf_prog *filter;
42 struct list_head link;
43 struct tcf_result res;
50 struct sock_filter *bpf_ops;
54 struct work_struct work;
59 static const struct nla_policy bpf_policy[TCA_BPF_MAX + 1] = {
60 [TCA_BPF_CLASSID] = { .type = NLA_U32 },
61 [TCA_BPF_FLAGS] = { .type = NLA_U32 },
62 [TCA_BPF_FLAGS_GEN] = { .type = NLA_U32 },
63 [TCA_BPF_FD] = { .type = NLA_U32 },
64 [TCA_BPF_NAME] = { .type = NLA_NUL_STRING,
65 .len = CLS_BPF_NAME_LEN },
66 [TCA_BPF_OPS_LEN] = { .type = NLA_U16 },
67 [TCA_BPF_OPS] = { .type = NLA_BINARY,
68 .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
71 static int cls_bpf_exec_opcode(int code)
86 static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
87 struct tcf_result *res)
89 struct cls_bpf_head *head = rcu_dereference_bh(tp->root);
90 bool at_ingress = skb_at_tc_ingress(skb);
91 struct cls_bpf_prog *prog;
94 /* Needed here for accessing maps. */
96 list_for_each_entry_rcu(prog, &head->plist, link) {
99 qdisc_skb_cb(skb)->tc_classid = prog->res.classid;
101 if (tc_skip_sw(prog->gen_flags)) {
102 filter_res = prog->exts_integrated ? TC_ACT_UNSPEC : 0;
103 } else if (at_ingress) {
104 /* It is safe to push/pull even if skb_shared() */
105 __skb_push(skb, skb->mac_len);
106 bpf_compute_data_pointers(skb);
107 filter_res = BPF_PROG_RUN(prog->filter, skb);
108 __skb_pull(skb, skb->mac_len);
110 bpf_compute_data_pointers(skb);
111 filter_res = BPF_PROG_RUN(prog->filter, skb);
114 if (prog->exts_integrated) {
116 res->classid = TC_H_MAJ(prog->res.classid) |
117 qdisc_skb_cb(skb)->tc_classid;
119 ret = cls_bpf_exec_opcode(filter_res);
120 if (ret == TC_ACT_UNSPEC)
127 if (filter_res != -1) {
129 res->classid = filter_res;
134 ret = tcf_exts_exec(skb, &prog->exts, res);
145 static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog)
147 return !prog->bpf_ops;
150 static int cls_bpf_offload_cmd(struct tcf_proto *tp, struct cls_bpf_prog *prog,
151 enum tc_clsbpf_command cmd)
153 bool addorrep = cmd == TC_CLSBPF_ADD || cmd == TC_CLSBPF_REPLACE;
154 struct tcf_block *block = tp->chain->block;
155 bool skip_sw = tc_skip_sw(prog->gen_flags);
156 struct tc_cls_bpf_offload cls_bpf = {};
159 tc_cls_common_offload_init(&cls_bpf.common, tp);
160 cls_bpf.command = cmd;
161 cls_bpf.exts = &prog->exts;
162 cls_bpf.prog = prog->filter;
163 cls_bpf.name = prog->bpf_name;
164 cls_bpf.exts_integrated = prog->exts_integrated;
165 cls_bpf.gen_flags = prog->gen_flags;
167 err = tc_setup_cb_call(block, NULL, TC_SETUP_CLSBPF, &cls_bpf, skip_sw);
170 cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_DESTROY);
172 } else if (err > 0) {
173 prog->gen_flags |= TCA_CLS_FLAGS_IN_HW;
177 if (addorrep && skip_sw && !(prog->gen_flags & TCA_CLS_FLAGS_IN_HW))
183 static int cls_bpf_offload(struct tcf_proto *tp, struct cls_bpf_prog *prog,
184 struct cls_bpf_prog *oldprog)
186 struct cls_bpf_prog *obj = prog;
187 enum tc_clsbpf_command cmd;
191 skip_sw = tc_skip_sw(prog->gen_flags) ||
192 (oldprog && tc_skip_sw(oldprog->gen_flags));
194 if (oldprog && oldprog->offloaded) {
195 if (!tc_skip_hw(prog->gen_flags)) {
196 cmd = TC_CLSBPF_REPLACE;
197 } else if (!tc_skip_sw(prog->gen_flags)) {
199 cmd = TC_CLSBPF_DESTROY;
204 if (tc_skip_hw(prog->gen_flags))
205 return skip_sw ? -EINVAL : 0;
209 ret = cls_bpf_offload_cmd(tp, obj, cmd);
213 obj->offloaded = true;
215 oldprog->offloaded = false;
220 static void cls_bpf_stop_offload(struct tcf_proto *tp,
221 struct cls_bpf_prog *prog)
225 if (!prog->offloaded)
228 err = cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_DESTROY);
230 pr_err("Stopping hardware offload failed: %d\n", err);
234 prog->offloaded = false;
237 static void cls_bpf_offload_update_stats(struct tcf_proto *tp,
238 struct cls_bpf_prog *prog)
240 if (!prog->offloaded)
243 cls_bpf_offload_cmd(tp, prog, TC_CLSBPF_STATS);
246 static int cls_bpf_init(struct tcf_proto *tp)
248 struct cls_bpf_head *head;
250 head = kzalloc(sizeof(*head), GFP_KERNEL);
254 INIT_LIST_HEAD_RCU(&head->plist);
255 idr_init(&head->handle_idr);
256 rcu_assign_pointer(tp->root, head);
261 static void __cls_bpf_delete_prog(struct cls_bpf_prog *prog)
263 tcf_exts_destroy(&prog->exts);
264 tcf_exts_put_net(&prog->exts);
266 if (cls_bpf_is_ebpf(prog))
267 bpf_prog_put(prog->filter);
269 bpf_prog_destroy(prog->filter);
271 kfree(prog->bpf_name);
272 kfree(prog->bpf_ops);
276 static void cls_bpf_delete_prog_work(struct work_struct *work)
278 struct cls_bpf_prog *prog = container_of(work, struct cls_bpf_prog, work);
281 __cls_bpf_delete_prog(prog);
285 static void cls_bpf_delete_prog_rcu(struct rcu_head *rcu)
287 struct cls_bpf_prog *prog = container_of(rcu, struct cls_bpf_prog, rcu);
289 INIT_WORK(&prog->work, cls_bpf_delete_prog_work);
290 tcf_queue_work(&prog->work);
293 static void __cls_bpf_delete(struct tcf_proto *tp, struct cls_bpf_prog *prog)
295 struct cls_bpf_head *head = rtnl_dereference(tp->root);
297 idr_remove_ext(&head->handle_idr, prog->handle);
298 cls_bpf_stop_offload(tp, prog);
299 list_del_rcu(&prog->link);
300 tcf_unbind_filter(tp, &prog->res);
301 if (tcf_exts_get_net(&prog->exts))
302 call_rcu(&prog->rcu, cls_bpf_delete_prog_rcu);
304 __cls_bpf_delete_prog(prog);
307 static int cls_bpf_delete(struct tcf_proto *tp, void *arg, bool *last)
309 struct cls_bpf_head *head = rtnl_dereference(tp->root);
311 __cls_bpf_delete(tp, arg);
312 *last = list_empty(&head->plist);
316 static void cls_bpf_destroy(struct tcf_proto *tp)
318 struct cls_bpf_head *head = rtnl_dereference(tp->root);
319 struct cls_bpf_prog *prog, *tmp;
321 list_for_each_entry_safe(prog, tmp, &head->plist, link)
322 __cls_bpf_delete(tp, prog);
324 idr_destroy(&head->handle_idr);
325 kfree_rcu(head, rcu);
328 static void *cls_bpf_get(struct tcf_proto *tp, u32 handle)
330 struct cls_bpf_head *head = rtnl_dereference(tp->root);
331 struct cls_bpf_prog *prog;
333 list_for_each_entry(prog, &head->plist, link) {
334 if (prog->handle == handle)
341 static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog)
343 struct sock_filter *bpf_ops;
344 struct sock_fprog_kern fprog_tmp;
346 u16 bpf_size, bpf_num_ops;
349 bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]);
350 if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
353 bpf_size = bpf_num_ops * sizeof(*bpf_ops);
354 if (bpf_size != nla_len(tb[TCA_BPF_OPS]))
357 bpf_ops = kzalloc(bpf_size, GFP_KERNEL);
361 memcpy(bpf_ops, nla_data(tb[TCA_BPF_OPS]), bpf_size);
363 fprog_tmp.len = bpf_num_ops;
364 fprog_tmp.filter = bpf_ops;
366 ret = bpf_prog_create(&fp, &fprog_tmp);
372 prog->bpf_ops = bpf_ops;
373 prog->bpf_num_ops = bpf_num_ops;
374 prog->bpf_name = NULL;
380 static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog,
381 u32 gen_flags, const struct tcf_proto *tp)
387 bpf_fd = nla_get_u32(tb[TCA_BPF_FD]);
389 if (gen_flags & TCA_CLS_FLAGS_SKIP_SW)
390 fp = bpf_prog_get_type_dev(bpf_fd, BPF_PROG_TYPE_SCHED_CLS,
393 fp = bpf_prog_get_type(bpf_fd, BPF_PROG_TYPE_SCHED_CLS);
397 if (tb[TCA_BPF_NAME]) {
398 name = nla_memdup(tb[TCA_BPF_NAME], GFP_KERNEL);
405 prog->bpf_ops = NULL;
406 prog->bpf_name = name;
409 if (fp->dst_needed && !(tp->q->flags & TCQ_F_INGRESS))
410 netif_keep_dst(qdisc_dev(tp->q));
415 static int cls_bpf_set_parms(struct net *net, struct tcf_proto *tp,
416 struct cls_bpf_prog *prog, unsigned long base,
417 struct nlattr **tb, struct nlattr *est, bool ovr)
419 bool is_bpf, is_ebpf, have_exts = false;
423 is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS];
424 is_ebpf = tb[TCA_BPF_FD];
425 if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf))
428 ret = tcf_exts_validate(net, tp, tb, est, &prog->exts, ovr);
432 if (tb[TCA_BPF_FLAGS]) {
433 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]);
435 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT)
438 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT;
440 if (tb[TCA_BPF_FLAGS_GEN]) {
441 gen_flags = nla_get_u32(tb[TCA_BPF_FLAGS_GEN]);
442 if (gen_flags & ~CLS_BPF_SUPPORTED_GEN_FLAGS ||
443 !tc_flags_valid(gen_flags))
447 prog->exts_integrated = have_exts;
448 prog->gen_flags = gen_flags;
450 ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) :
451 cls_bpf_prog_from_efd(tb, prog, gen_flags, tp);
455 if (tb[TCA_BPF_CLASSID]) {
456 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]);
457 tcf_bind_filter(tp, &prog->res, base);
463 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb,
464 struct tcf_proto *tp, unsigned long base,
465 u32 handle, struct nlattr **tca,
466 void **arg, bool ovr)
468 struct cls_bpf_head *head = rtnl_dereference(tp->root);
469 struct cls_bpf_prog *oldprog = *arg;
470 struct nlattr *tb[TCA_BPF_MAX + 1];
471 struct cls_bpf_prog *prog;
472 unsigned long idr_index;
475 if (tca[TCA_OPTIONS] == NULL)
478 ret = nla_parse_nested(tb, TCA_BPF_MAX, tca[TCA_OPTIONS], bpf_policy,
483 prog = kzalloc(sizeof(*prog), GFP_KERNEL);
487 ret = tcf_exts_init(&prog->exts, TCA_BPF_ACT, TCA_BPF_POLICE);
492 if (handle && oldprog->handle != handle) {
499 ret = idr_alloc_ext(&head->handle_idr, prog, &idr_index,
500 1, 0x7FFFFFFF, GFP_KERNEL);
503 prog->handle = idr_index;
506 ret = idr_alloc_ext(&head->handle_idr, prog, &idr_index,
507 handle, handle + 1, GFP_KERNEL);
511 prog->handle = handle;
514 ret = cls_bpf_set_parms(net, tp, prog, base, tb, tca[TCA_RATE], ovr);
518 ret = cls_bpf_offload(tp, prog, oldprog);
521 idr_remove_ext(&head->handle_idr, prog->handle);
522 __cls_bpf_delete_prog(prog);
526 if (!tc_in_hw(prog->gen_flags))
527 prog->gen_flags |= TCA_CLS_FLAGS_NOT_IN_HW;
530 idr_replace_ext(&head->handle_idr, prog, handle);
531 list_replace_rcu(&oldprog->link, &prog->link);
532 tcf_unbind_filter(tp, &oldprog->res);
533 tcf_exts_get_net(&oldprog->exts);
534 call_rcu(&oldprog->rcu, cls_bpf_delete_prog_rcu);
536 list_add_rcu(&prog->link, &head->plist);
544 idr_remove_ext(&head->handle_idr, prog->handle);
546 tcf_exts_destroy(&prog->exts);
551 static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog,
556 if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops))
559 nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops *
560 sizeof(struct sock_filter));
564 memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
569 static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog,
574 if (prog->bpf_name &&
575 nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name))
578 if (nla_put_u32(skb, TCA_BPF_ID, prog->filter->aux->id))
581 nla = nla_reserve(skb, TCA_BPF_TAG, sizeof(prog->filter->tag));
585 memcpy(nla_data(nla), prog->filter->tag, nla_len(nla));
590 static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, void *fh,
591 struct sk_buff *skb, struct tcmsg *tm)
593 struct cls_bpf_prog *prog = fh;
601 tm->tcm_handle = prog->handle;
603 cls_bpf_offload_update_stats(tp, prog);
605 nest = nla_nest_start(skb, TCA_OPTIONS);
607 goto nla_put_failure;
609 if (prog->res.classid &&
610 nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid))
611 goto nla_put_failure;
613 if (cls_bpf_is_ebpf(prog))
614 ret = cls_bpf_dump_ebpf_info(prog, skb);
616 ret = cls_bpf_dump_bpf_info(prog, skb);
618 goto nla_put_failure;
620 if (tcf_exts_dump(skb, &prog->exts) < 0)
621 goto nla_put_failure;
623 if (prog->exts_integrated)
624 bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT;
625 if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags))
626 goto nla_put_failure;
627 if (prog->gen_flags &&
628 nla_put_u32(skb, TCA_BPF_FLAGS_GEN, prog->gen_flags))
629 goto nla_put_failure;
631 nla_nest_end(skb, nest);
633 if (tcf_exts_dump_stats(skb, &prog->exts) < 0)
634 goto nla_put_failure;
639 nla_nest_cancel(skb, nest);
643 static void cls_bpf_bind_class(void *fh, u32 classid, unsigned long cl)
645 struct cls_bpf_prog *prog = fh;
647 if (prog && prog->res.classid == classid)
648 prog->res.class = cl;
651 static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg)
653 struct cls_bpf_head *head = rtnl_dereference(tp->root);
654 struct cls_bpf_prog *prog;
656 list_for_each_entry(prog, &head->plist, link) {
657 if (arg->count < arg->skip)
659 if (arg->fn(tp, prog, arg) < 0) {
668 static struct tcf_proto_ops cls_bpf_ops __read_mostly = {
670 .owner = THIS_MODULE,
671 .classify = cls_bpf_classify,
672 .init = cls_bpf_init,
673 .destroy = cls_bpf_destroy,
675 .change = cls_bpf_change,
676 .delete = cls_bpf_delete,
677 .walk = cls_bpf_walk,
678 .dump = cls_bpf_dump,
679 .bind_class = cls_bpf_bind_class,
682 static int __init cls_bpf_init_mod(void)
684 return register_tcf_proto_ops(&cls_bpf_ops);
687 static void __exit cls_bpf_exit_mod(void)
689 unregister_tcf_proto_ops(&cls_bpf_ops);
692 module_init(cls_bpf_init_mod);
693 module_exit(cls_bpf_exit_mod);