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;
46 unsigned int in_hw_count;
50 struct sock_filter *bpf_ops;
53 struct rcu_work rwork;
56 static const struct nla_policy bpf_policy[TCA_BPF_MAX + 1] = {
57 [TCA_BPF_CLASSID] = { .type = NLA_U32 },
58 [TCA_BPF_FLAGS] = { .type = NLA_U32 },
59 [TCA_BPF_FLAGS_GEN] = { .type = NLA_U32 },
60 [TCA_BPF_FD] = { .type = NLA_U32 },
61 [TCA_BPF_NAME] = { .type = NLA_NUL_STRING,
62 .len = CLS_BPF_NAME_LEN },
63 [TCA_BPF_OPS_LEN] = { .type = NLA_U16 },
64 [TCA_BPF_OPS] = { .type = NLA_BINARY,
65 .len = sizeof(struct sock_filter) * BPF_MAXINSNS },
68 static int cls_bpf_exec_opcode(int code)
83 static int cls_bpf_classify(struct sk_buff *skb, const struct tcf_proto *tp,
84 struct tcf_result *res)
86 struct cls_bpf_head *head = rcu_dereference_bh(tp->root);
87 bool at_ingress = skb_at_tc_ingress(skb);
88 struct cls_bpf_prog *prog;
91 /* Needed here for accessing maps. */
93 list_for_each_entry_rcu(prog, &head->plist, link) {
96 qdisc_skb_cb(skb)->tc_classid = prog->res.classid;
98 if (tc_skip_sw(prog->gen_flags)) {
99 filter_res = prog->exts_integrated ? TC_ACT_UNSPEC : 0;
100 } else if (at_ingress) {
101 /* It is safe to push/pull even if skb_shared() */
102 __skb_push(skb, skb->mac_len);
103 bpf_compute_data_pointers(skb);
104 filter_res = BPF_PROG_RUN(prog->filter, skb);
105 __skb_pull(skb, skb->mac_len);
107 bpf_compute_data_pointers(skb);
108 filter_res = BPF_PROG_RUN(prog->filter, skb);
111 if (prog->exts_integrated) {
113 res->classid = TC_H_MAJ(prog->res.classid) |
114 qdisc_skb_cb(skb)->tc_classid;
116 ret = cls_bpf_exec_opcode(filter_res);
117 if (ret == TC_ACT_UNSPEC)
124 if (filter_res != -1) {
126 res->classid = filter_res;
131 ret = tcf_exts_exec(skb, &prog->exts, res);
142 static bool cls_bpf_is_ebpf(const struct cls_bpf_prog *prog)
144 return !prog->bpf_ops;
147 static int cls_bpf_offload_cmd(struct tcf_proto *tp, struct cls_bpf_prog *prog,
148 struct cls_bpf_prog *oldprog,
149 struct netlink_ext_ack *extack)
151 struct tcf_block *block = tp->chain->block;
152 struct tc_cls_bpf_offload cls_bpf = {};
153 struct cls_bpf_prog *obj;
157 skip_sw = prog && tc_skip_sw(prog->gen_flags);
158 obj = prog ?: oldprog;
160 tc_cls_common_offload_init(&cls_bpf.common, tp, obj->gen_flags, extack);
161 cls_bpf.command = TC_CLSBPF_OFFLOAD;
162 cls_bpf.exts = &obj->exts;
163 cls_bpf.prog = prog ? prog->filter : NULL;
164 cls_bpf.oldprog = oldprog ? oldprog->filter : NULL;
165 cls_bpf.name = obj->bpf_name;
166 cls_bpf.exts_integrated = obj->exts_integrated;
169 tcf_block_offload_dec(block, &oldprog->gen_flags);
171 err = tc_setup_cb_call(block, TC_SETUP_CLSBPF, &cls_bpf, skip_sw);
174 cls_bpf_offload_cmd(tp, oldprog, prog, extack);
176 } else if (err > 0) {
177 prog->in_hw_count = err;
178 tcf_block_offload_inc(block, &prog->gen_flags);
182 if (prog && skip_sw && !(prog->gen_flags & TCA_CLS_FLAGS_IN_HW))
188 static u32 cls_bpf_flags(u32 flags)
190 return flags & CLS_BPF_SUPPORTED_GEN_FLAGS;
193 static int cls_bpf_offload(struct tcf_proto *tp, struct cls_bpf_prog *prog,
194 struct cls_bpf_prog *oldprog,
195 struct netlink_ext_ack *extack)
197 if (prog && oldprog &&
198 cls_bpf_flags(prog->gen_flags) !=
199 cls_bpf_flags(oldprog->gen_flags))
202 if (prog && tc_skip_hw(prog->gen_flags))
204 if (oldprog && tc_skip_hw(oldprog->gen_flags))
206 if (!prog && !oldprog)
209 return cls_bpf_offload_cmd(tp, prog, oldprog, extack);
212 static void cls_bpf_stop_offload(struct tcf_proto *tp,
213 struct cls_bpf_prog *prog,
214 struct netlink_ext_ack *extack)
218 err = cls_bpf_offload_cmd(tp, NULL, prog, extack);
220 pr_err("Stopping hardware offload failed: %d\n", err);
223 static void cls_bpf_offload_update_stats(struct tcf_proto *tp,
224 struct cls_bpf_prog *prog)
226 struct tcf_block *block = tp->chain->block;
227 struct tc_cls_bpf_offload cls_bpf = {};
229 tc_cls_common_offload_init(&cls_bpf.common, tp, prog->gen_flags, NULL);
230 cls_bpf.command = TC_CLSBPF_STATS;
231 cls_bpf.exts = &prog->exts;
232 cls_bpf.prog = prog->filter;
233 cls_bpf.name = prog->bpf_name;
234 cls_bpf.exts_integrated = prog->exts_integrated;
236 tc_setup_cb_call(block, TC_SETUP_CLSBPF, &cls_bpf, false);
239 static int cls_bpf_init(struct tcf_proto *tp)
241 struct cls_bpf_head *head;
243 head = kzalloc(sizeof(*head), GFP_KERNEL);
247 INIT_LIST_HEAD_RCU(&head->plist);
248 idr_init(&head->handle_idr);
249 rcu_assign_pointer(tp->root, head);
254 static void cls_bpf_free_parms(struct cls_bpf_prog *prog)
256 if (cls_bpf_is_ebpf(prog))
257 bpf_prog_put(prog->filter);
259 bpf_prog_destroy(prog->filter);
261 kfree(prog->bpf_name);
262 kfree(prog->bpf_ops);
265 static void __cls_bpf_delete_prog(struct cls_bpf_prog *prog)
267 tcf_exts_destroy(&prog->exts);
268 tcf_exts_put_net(&prog->exts);
270 cls_bpf_free_parms(prog);
274 static void cls_bpf_delete_prog_work(struct work_struct *work)
276 struct cls_bpf_prog *prog = container_of(to_rcu_work(work),
280 __cls_bpf_delete_prog(prog);
284 static void __cls_bpf_delete(struct tcf_proto *tp, struct cls_bpf_prog *prog,
285 struct netlink_ext_ack *extack)
287 struct cls_bpf_head *head = rtnl_dereference(tp->root);
289 idr_remove(&head->handle_idr, prog->handle);
290 cls_bpf_stop_offload(tp, prog, extack);
291 list_del_rcu(&prog->link);
292 tcf_unbind_filter(tp, &prog->res);
293 if (tcf_exts_get_net(&prog->exts))
294 tcf_queue_work(&prog->rwork, cls_bpf_delete_prog_work);
296 __cls_bpf_delete_prog(prog);
299 static int cls_bpf_delete(struct tcf_proto *tp, void *arg, bool *last,
300 bool rtnl_held, struct netlink_ext_ack *extack)
302 struct cls_bpf_head *head = rtnl_dereference(tp->root);
304 __cls_bpf_delete(tp, arg, extack);
305 *last = list_empty(&head->plist);
309 static void cls_bpf_destroy(struct tcf_proto *tp, bool rtnl_held,
310 struct netlink_ext_ack *extack)
312 struct cls_bpf_head *head = rtnl_dereference(tp->root);
313 struct cls_bpf_prog *prog, *tmp;
315 list_for_each_entry_safe(prog, tmp, &head->plist, link)
316 __cls_bpf_delete(tp, prog, extack);
318 idr_destroy(&head->handle_idr);
319 kfree_rcu(head, rcu);
322 static void *cls_bpf_get(struct tcf_proto *tp, u32 handle)
324 struct cls_bpf_head *head = rtnl_dereference(tp->root);
325 struct cls_bpf_prog *prog;
327 list_for_each_entry(prog, &head->plist, link) {
328 if (prog->handle == handle)
335 static int cls_bpf_prog_from_ops(struct nlattr **tb, struct cls_bpf_prog *prog)
337 struct sock_filter *bpf_ops;
338 struct sock_fprog_kern fprog_tmp;
340 u16 bpf_size, bpf_num_ops;
343 bpf_num_ops = nla_get_u16(tb[TCA_BPF_OPS_LEN]);
344 if (bpf_num_ops > BPF_MAXINSNS || bpf_num_ops == 0)
347 bpf_size = bpf_num_ops * sizeof(*bpf_ops);
348 if (bpf_size != nla_len(tb[TCA_BPF_OPS]))
351 bpf_ops = kmemdup(nla_data(tb[TCA_BPF_OPS]), bpf_size, GFP_KERNEL);
355 fprog_tmp.len = bpf_num_ops;
356 fprog_tmp.filter = bpf_ops;
358 ret = bpf_prog_create(&fp, &fprog_tmp);
364 prog->bpf_ops = bpf_ops;
365 prog->bpf_num_ops = bpf_num_ops;
366 prog->bpf_name = NULL;
372 static int cls_bpf_prog_from_efd(struct nlattr **tb, struct cls_bpf_prog *prog,
373 u32 gen_flags, const struct tcf_proto *tp)
380 bpf_fd = nla_get_u32(tb[TCA_BPF_FD]);
381 skip_sw = gen_flags & TCA_CLS_FLAGS_SKIP_SW;
383 fp = bpf_prog_get_type_dev(bpf_fd, BPF_PROG_TYPE_SCHED_CLS, skip_sw);
387 if (tb[TCA_BPF_NAME]) {
388 name = nla_memdup(tb[TCA_BPF_NAME], GFP_KERNEL);
395 prog->bpf_ops = NULL;
396 prog->bpf_name = name;
400 tcf_block_netif_keep_dst(tp->chain->block);
405 static int cls_bpf_set_parms(struct net *net, struct tcf_proto *tp,
406 struct cls_bpf_prog *prog, unsigned long base,
407 struct nlattr **tb, struct nlattr *est, bool ovr,
408 struct netlink_ext_ack *extack)
410 bool is_bpf, is_ebpf, have_exts = false;
414 is_bpf = tb[TCA_BPF_OPS_LEN] && tb[TCA_BPF_OPS];
415 is_ebpf = tb[TCA_BPF_FD];
416 if ((!is_bpf && !is_ebpf) || (is_bpf && is_ebpf))
419 ret = tcf_exts_validate(net, tp, tb, est, &prog->exts, ovr, true,
424 if (tb[TCA_BPF_FLAGS]) {
425 u32 bpf_flags = nla_get_u32(tb[TCA_BPF_FLAGS]);
427 if (bpf_flags & ~TCA_BPF_FLAG_ACT_DIRECT)
430 have_exts = bpf_flags & TCA_BPF_FLAG_ACT_DIRECT;
432 if (tb[TCA_BPF_FLAGS_GEN]) {
433 gen_flags = nla_get_u32(tb[TCA_BPF_FLAGS_GEN]);
434 if (gen_flags & ~CLS_BPF_SUPPORTED_GEN_FLAGS ||
435 !tc_flags_valid(gen_flags))
439 prog->exts_integrated = have_exts;
440 prog->gen_flags = gen_flags;
442 ret = is_bpf ? cls_bpf_prog_from_ops(tb, prog) :
443 cls_bpf_prog_from_efd(tb, prog, gen_flags, tp);
447 if (tb[TCA_BPF_CLASSID]) {
448 prog->res.classid = nla_get_u32(tb[TCA_BPF_CLASSID]);
449 tcf_bind_filter(tp, &prog->res, base);
455 static int cls_bpf_change(struct net *net, struct sk_buff *in_skb,
456 struct tcf_proto *tp, unsigned long base,
457 u32 handle, struct nlattr **tca,
458 void **arg, bool ovr, bool rtnl_held,
459 struct netlink_ext_ack *extack)
461 struct cls_bpf_head *head = rtnl_dereference(tp->root);
462 struct cls_bpf_prog *oldprog = *arg;
463 struct nlattr *tb[TCA_BPF_MAX + 1];
464 struct cls_bpf_prog *prog;
467 if (tca[TCA_OPTIONS] == NULL)
470 ret = nla_parse_nested_deprecated(tb, TCA_BPF_MAX, tca[TCA_OPTIONS],
475 prog = kzalloc(sizeof(*prog), GFP_KERNEL);
479 ret = tcf_exts_init(&prog->exts, net, TCA_BPF_ACT, TCA_BPF_POLICE);
484 if (handle && oldprog->handle != handle) {
492 ret = idr_alloc_u32(&head->handle_idr, prog, &handle,
493 INT_MAX, GFP_KERNEL);
494 } else if (!oldprog) {
495 ret = idr_alloc_u32(&head->handle_idr, prog, &handle,
501 prog->handle = handle;
503 ret = cls_bpf_set_parms(net, tp, prog, base, tb, tca[TCA_RATE], ovr,
508 ret = cls_bpf_offload(tp, prog, oldprog, extack);
512 if (!tc_in_hw(prog->gen_flags))
513 prog->gen_flags |= TCA_CLS_FLAGS_NOT_IN_HW;
516 idr_replace(&head->handle_idr, prog, handle);
517 list_replace_rcu(&oldprog->link, &prog->link);
518 tcf_unbind_filter(tp, &oldprog->res);
519 tcf_exts_get_net(&oldprog->exts);
520 tcf_queue_work(&oldprog->rwork, cls_bpf_delete_prog_work);
522 list_add_rcu(&prog->link, &head->plist);
529 cls_bpf_free_parms(prog);
532 idr_remove(&head->handle_idr, prog->handle);
534 tcf_exts_destroy(&prog->exts);
539 static int cls_bpf_dump_bpf_info(const struct cls_bpf_prog *prog,
544 if (nla_put_u16(skb, TCA_BPF_OPS_LEN, prog->bpf_num_ops))
547 nla = nla_reserve(skb, TCA_BPF_OPS, prog->bpf_num_ops *
548 sizeof(struct sock_filter));
552 memcpy(nla_data(nla), prog->bpf_ops, nla_len(nla));
557 static int cls_bpf_dump_ebpf_info(const struct cls_bpf_prog *prog,
562 if (prog->bpf_name &&
563 nla_put_string(skb, TCA_BPF_NAME, prog->bpf_name))
566 if (nla_put_u32(skb, TCA_BPF_ID, prog->filter->aux->id))
569 nla = nla_reserve(skb, TCA_BPF_TAG, sizeof(prog->filter->tag));
573 memcpy(nla_data(nla), prog->filter->tag, nla_len(nla));
578 static int cls_bpf_dump(struct net *net, struct tcf_proto *tp, void *fh,
579 struct sk_buff *skb, struct tcmsg *tm, bool rtnl_held)
581 struct cls_bpf_prog *prog = fh;
589 tm->tcm_handle = prog->handle;
591 cls_bpf_offload_update_stats(tp, prog);
593 nest = nla_nest_start_noflag(skb, TCA_OPTIONS);
595 goto nla_put_failure;
597 if (prog->res.classid &&
598 nla_put_u32(skb, TCA_BPF_CLASSID, prog->res.classid))
599 goto nla_put_failure;
601 if (cls_bpf_is_ebpf(prog))
602 ret = cls_bpf_dump_ebpf_info(prog, skb);
604 ret = cls_bpf_dump_bpf_info(prog, skb);
606 goto nla_put_failure;
608 if (tcf_exts_dump(skb, &prog->exts) < 0)
609 goto nla_put_failure;
611 if (prog->exts_integrated)
612 bpf_flags |= TCA_BPF_FLAG_ACT_DIRECT;
613 if (bpf_flags && nla_put_u32(skb, TCA_BPF_FLAGS, bpf_flags))
614 goto nla_put_failure;
615 if (prog->gen_flags &&
616 nla_put_u32(skb, TCA_BPF_FLAGS_GEN, prog->gen_flags))
617 goto nla_put_failure;
619 nla_nest_end(skb, nest);
621 if (tcf_exts_dump_stats(skb, &prog->exts) < 0)
622 goto nla_put_failure;
627 nla_nest_cancel(skb, nest);
631 static void cls_bpf_bind_class(void *fh, u32 classid, unsigned long cl)
633 struct cls_bpf_prog *prog = fh;
635 if (prog && prog->res.classid == classid)
636 prog->res.class = cl;
639 static void cls_bpf_walk(struct tcf_proto *tp, struct tcf_walker *arg,
642 struct cls_bpf_head *head = rtnl_dereference(tp->root);
643 struct cls_bpf_prog *prog;
645 list_for_each_entry(prog, &head->plist, link) {
646 if (arg->count < arg->skip)
648 if (arg->fn(tp, prog, arg) < 0) {
657 static int cls_bpf_reoffload(struct tcf_proto *tp, bool add, tc_setup_cb_t *cb,
658 void *cb_priv, struct netlink_ext_ack *extack)
660 struct cls_bpf_head *head = rtnl_dereference(tp->root);
661 struct tcf_block *block = tp->chain->block;
662 struct tc_cls_bpf_offload cls_bpf = {};
663 struct cls_bpf_prog *prog;
666 list_for_each_entry(prog, &head->plist, link) {
667 if (tc_skip_hw(prog->gen_flags))
670 tc_cls_common_offload_init(&cls_bpf.common, tp, prog->gen_flags,
672 cls_bpf.command = TC_CLSBPF_OFFLOAD;
673 cls_bpf.exts = &prog->exts;
674 cls_bpf.prog = add ? prog->filter : NULL;
675 cls_bpf.oldprog = add ? NULL : prog->filter;
676 cls_bpf.name = prog->bpf_name;
677 cls_bpf.exts_integrated = prog->exts_integrated;
679 err = cb(TC_SETUP_CLSBPF, &cls_bpf, cb_priv);
681 if (add && tc_skip_sw(prog->gen_flags))
686 tc_cls_offload_cnt_update(block, &prog->in_hw_count,
687 &prog->gen_flags, add);
693 static struct tcf_proto_ops cls_bpf_ops __read_mostly = {
695 .owner = THIS_MODULE,
696 .classify = cls_bpf_classify,
697 .init = cls_bpf_init,
698 .destroy = cls_bpf_destroy,
700 .change = cls_bpf_change,
701 .delete = cls_bpf_delete,
702 .walk = cls_bpf_walk,
703 .reoffload = cls_bpf_reoffload,
704 .dump = cls_bpf_dump,
705 .bind_class = cls_bpf_bind_class,
708 static int __init cls_bpf_init_mod(void)
710 return register_tcf_proto_ops(&cls_bpf_ops);
713 static void __exit cls_bpf_exit_mod(void)
715 unregister_tcf_proto_ops(&cls_bpf_ops);
718 module_init(cls_bpf_init_mod);
719 module_exit(cls_bpf_exit_mod);