1// SPDX-License-Identifier: GPL-2.0-only
  2/* Copyright (c) 2016 Thomas Graf <tgraf@tgraf.ch>
  3 */
  4
  5#include <linux/kernel.h>
  6#include <linux/module.h>
  7#include <linux/skbuff.h>
  8#include <linux/types.h>
  9#include <linux/bpf.h>
 10#include <net/lwtunnel.h>
 11#include <net/gre.h>
 12#include <net/ip6_route.h>
 13#include <net/ipv6_stubs.h>
 14
 15struct bpf_lwt_prog {
 16	struct bpf_prog *prog;
 17	char *name;
 18};
 19
 20struct bpf_lwt {
 21	struct bpf_lwt_prog in;
 22	struct bpf_lwt_prog out;
 23	struct bpf_lwt_prog xmit;
 24	int family;
 25};
 26
 27#define MAX_PROG_NAME 256
 28
 29static inline struct bpf_lwt *bpf_lwt_lwtunnel(struct lwtunnel_state *lwt)
 30{
 31	return (struct bpf_lwt *)lwt->data;
 32}
 33
 34#define NO_REDIRECT false
 35#define CAN_REDIRECT true
 36
 37static int run_lwt_bpf(struct sk_buff *skb, struct bpf_lwt_prog *lwt,
 38		       struct dst_entry *dst, bool can_redirect)
 39{
 40	int ret;
 41
 42	/* Preempt disable is needed to protect per-cpu redirect_info between
 43	 * BPF prog and skb_do_redirect(). The call_rcu in bpf_prog_put() and
 44	 * access to maps strictly require a rcu_read_lock() for protection,
 45	 * mixing with BH RCU lock doesn't work.
 46	 */
 47	preempt_disable();
 48	bpf_compute_data_pointers(skb);
 49	ret = bpf_prog_run_save_cb(lwt->prog, skb);
 50
 51	switch (ret) {
 52	case BPF_OK:
 53	case BPF_LWT_REROUTE:
 54		break;
 55
 56	case BPF_REDIRECT:
 57		if (unlikely(!can_redirect)) {
 58			pr_warn_once("Illegal redirect return code in prog %s\n",
 59				     lwt->name ? : "<unknown>");
 60			ret = BPF_OK;
 61		} else {
 62			skb_reset_mac_header(skb);
 63			ret = skb_do_redirect(skb);
 64			if (ret == 0)
 65				ret = BPF_REDIRECT;
 66		}
 67		break;
 68
 69	case BPF_DROP:
 70		kfree_skb(skb);
 71		ret = -EPERM;
 72		break;
 73
 74	default:
 75		pr_warn_once("bpf-lwt: Illegal return value %u, expect packet loss\n", ret);
 76		kfree_skb(skb);
 77		ret = -EINVAL;
 78		break;
 79	}
 80
 81	preempt_enable();
 82
 83	return ret;
 84}
 85
 86static int bpf_lwt_input_reroute(struct sk_buff *skb)
 87{
 88	int err = -EINVAL;
 89
 90	if (skb->protocol == htons(ETH_P_IP)) {
 91		struct net_device *dev = skb_dst(skb)->dev;
 92		struct iphdr *iph = ip_hdr(skb);
 93
 94		dev_hold(dev);
 95		skb_dst_drop(skb);
 96		err = ip_route_input_noref(skb, iph->daddr, iph->saddr,
 97					   iph->tos, dev);
 98		dev_put(dev);
 99	} else if (skb->protocol == htons(ETH_P_IPV6)) {
100		skb_dst_drop(skb);
101		err = ipv6_stub->ipv6_route_input(skb);
102	} else {
103		err = -EAFNOSUPPORT;
104	}
105
106	if (err)
107		goto err;
108	return dst_input(skb);
109
110err:
111	kfree_skb(skb);
112	return err;
113}
114
115static int bpf_input(struct sk_buff *skb)
116{
117	struct dst_entry *dst = skb_dst(skb);
118	struct bpf_lwt *bpf;
119	int ret;
120
121	bpf = bpf_lwt_lwtunnel(dst->lwtstate);
122	if (bpf->in.prog) {
123		ret = run_lwt_bpf(skb, &bpf->in, dst, NO_REDIRECT);
124		if (ret < 0)
125			return ret;
126		if (ret == BPF_LWT_REROUTE)
127			return bpf_lwt_input_reroute(skb);
128	}
129
130	if (unlikely(!dst->lwtstate->orig_input)) {
131		kfree_skb(skb);
132		return -EINVAL;
133	}
134
135	return dst->lwtstate->orig_input(skb);
136}
137
138static int bpf_output(struct net *net, struct sock *sk, struct sk_buff *skb)
139{
140	struct dst_entry *dst = skb_dst(skb);
141	struct bpf_lwt *bpf;
142	int ret;
143
144	bpf = bpf_lwt_lwtunnel(dst->lwtstate);
145	if (bpf->out.prog) {
146		ret = run_lwt_bpf(skb, &bpf->out, dst, NO_REDIRECT);
147		if (ret < 0)
148			return ret;
149	}
150
151	if (unlikely(!dst->lwtstate->orig_output)) {
152		pr_warn_once("orig_output not set on dst for prog %s\n",
153			     bpf->out.name);
154		kfree_skb(skb);
155		return -EINVAL;
156	}
157
158	return dst->lwtstate->orig_output(net, sk, skb);
159}
160
161static int xmit_check_hhlen(struct sk_buff *skb)
162{
163	int hh_len = skb_dst(skb)->dev->hard_header_len;
164
165	if (skb_headroom(skb) < hh_len) {
166		int nhead = HH_DATA_ALIGN(hh_len - skb_headroom(skb));
167
168		if (pskb_expand_head(skb, nhead, 0, GFP_ATOMIC))
169			return -ENOMEM;
170	}
171
172	return 0;
173}
174
175static int bpf_lwt_xmit_reroute(struct sk_buff *skb)
176{
177	struct net_device *l3mdev = l3mdev_master_dev_rcu(skb_dst(skb)->dev);
178	int oif = l3mdev ? l3mdev->ifindex : 0;
179	struct dst_entry *dst = NULL;
180	int err = -EAFNOSUPPORT;
181	struct sock *sk;
182	struct net *net;
183	bool ipv4;
184
185	if (skb->protocol == htons(ETH_P_IP))
186		ipv4 = true;
187	else if (skb->protocol == htons(ETH_P_IPV6))
188		ipv4 = false;
189	else
190		goto err;
191
192	sk = sk_to_full_sk(skb->sk);
193	if (sk) {
194		if (sk->sk_bound_dev_if)
195			oif = sk->sk_bound_dev_if;
196		net = sock_net(sk);
197	} else {
198		net = dev_net(skb_dst(skb)->dev);
199	}
200
201	if (ipv4) {
202		struct iphdr *iph = ip_hdr(skb);
203		struct flowi4 fl4 = {};
204		struct rtable *rt;
205
206		fl4.flowi4_oif = oif;
207		fl4.flowi4_mark = skb->mark;
208		fl4.flowi4_uid = sock_net_uid(net, sk);
209		fl4.flowi4_tos = RT_TOS(iph->tos);
210		fl4.flowi4_flags = FLOWI_FLAG_ANYSRC;
211		fl4.flowi4_proto = iph->protocol;
212		fl4.daddr = iph->daddr;
213		fl4.saddr = iph->saddr;
214
215		rt = ip_route_output_key(net, &fl4);
216		if (IS_ERR(rt)) {
217			err = PTR_ERR(rt);
218			goto err;
219		}
220		dst = &rt->dst;
221	} else {
222		struct ipv6hdr *iph6 = ipv6_hdr(skb);
223		struct flowi6 fl6 = {};
224
225		fl6.flowi6_oif = oif;
226		fl6.flowi6_mark = skb->mark;
227		fl6.flowi6_uid = sock_net_uid(net, sk);
228		fl6.flowlabel = ip6_flowinfo(iph6);
229		fl6.flowi6_proto = iph6->nexthdr;
230		fl6.daddr = iph6->daddr;
231		fl6.saddr = iph6->saddr;
232
233		err = ipv6_stub->ipv6_dst_lookup(net, skb->sk, &dst, &fl6);
234		if (unlikely(err))
235			goto err;
236		if (IS_ERR(dst)) {
237			err = PTR_ERR(dst);
238			goto err;
239		}
240	}
241	if (unlikely(dst->error)) {
242		err = dst->error;
243		dst_release(dst);
244		goto err;
245	}
246
247	/* Although skb header was reserved in bpf_lwt_push_ip_encap(), it
248	 * was done for the previous dst, so we are doing it here again, in
249	 * case the new dst needs much more space. The call below is a noop
250	 * if there is enough header space in skb.
251	 */
252	err = skb_cow_head(skb, LL_RESERVED_SPACE(dst->dev));
253	if (unlikely(err))
254		goto err;
255
256	skb_dst_drop(skb);
257	skb_dst_set(skb, dst);
258
259	err = dst_output(dev_net(skb_dst(skb)->dev), skb->sk, skb);
260	if (unlikely(err))
261		return err;
262
263	/* ip[6]_finish_output2 understand LWTUNNEL_XMIT_DONE */
264	return LWTUNNEL_XMIT_DONE;
265
266err:
267	kfree_skb(skb);
268	return err;
269}
270
271static int bpf_xmit(struct sk_buff *skb)
272{
273	struct dst_entry *dst = skb_dst(skb);
274	struct bpf_lwt *bpf;
275
276	bpf = bpf_lwt_lwtunnel(dst->lwtstate);
277	if (bpf->xmit.prog) {
278		__be16 proto = skb->protocol;
279		int ret;
280
281		ret = run_lwt_bpf(skb, &bpf->xmit, dst, CAN_REDIRECT);
282		switch (ret) {
283		case BPF_OK:
284			/* If the header changed, e.g. via bpf_lwt_push_encap,
285			 * BPF_LWT_REROUTE below should have been used if the
286			 * protocol was also changed.
287			 */
288			if (skb->protocol != proto) {
289				kfree_skb(skb);
290				return -EINVAL;
291			}
292			/* If the header was expanded, headroom might be too
293			 * small for L2 header to come, expand as needed.
294			 */
295			ret = xmit_check_hhlen(skb);
296			if (unlikely(ret))
297				return ret;
298
299			return LWTUNNEL_XMIT_CONTINUE;
300		case BPF_REDIRECT:
301			return LWTUNNEL_XMIT_DONE;
302		case BPF_LWT_REROUTE:
303			return bpf_lwt_xmit_reroute(skb);
304		default:
305			return ret;
306		}
307	}
308
309	return LWTUNNEL_XMIT_CONTINUE;
310}
311
312static void bpf_lwt_prog_destroy(struct bpf_lwt_prog *prog)
313{
314	if (prog->prog)
315		bpf_prog_put(prog->prog);
316
317	kfree(prog->name);
318}
319
320static void bpf_destroy_state(struct lwtunnel_state *lwt)
321{
322	struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt);
323
324	bpf_lwt_prog_destroy(&bpf->in);
325	bpf_lwt_prog_destroy(&bpf->out);
326	bpf_lwt_prog_destroy(&bpf->xmit);
327}
328
329static const struct nla_policy bpf_prog_policy[LWT_BPF_PROG_MAX + 1] = {
330	[LWT_BPF_PROG_FD]   = { .type = NLA_U32, },
331	[LWT_BPF_PROG_NAME] = { .type = NLA_NUL_STRING,
332				.len = MAX_PROG_NAME },
333};
334
335static int bpf_parse_prog(struct nlattr *attr, struct bpf_lwt_prog *prog,
336			  enum bpf_prog_type type)
337{
338	struct nlattr *tb[LWT_BPF_PROG_MAX + 1];
339	struct bpf_prog *p;
340	int ret;
341	u32 fd;
342
343	ret = nla_parse_nested_deprecated(tb, LWT_BPF_PROG_MAX, attr,
344					  bpf_prog_policy, NULL);
345	if (ret < 0)
346		return ret;
347
348	if (!tb[LWT_BPF_PROG_FD] || !tb[LWT_BPF_PROG_NAME])
349		return -EINVAL;
350
351	prog->name = nla_memdup(tb[LWT_BPF_PROG_NAME], GFP_ATOMIC);
352	if (!prog->name)
353		return -ENOMEM;
354
355	fd = nla_get_u32(tb[LWT_BPF_PROG_FD]);
356	p = bpf_prog_get_type(fd, type);
357	if (IS_ERR(p))
358		return PTR_ERR(p);
359
360	prog->prog = p;
361
362	return 0;
363}
364
365static const struct nla_policy bpf_nl_policy[LWT_BPF_MAX + 1] = {
366	[LWT_BPF_IN]		= { .type = NLA_NESTED, },
367	[LWT_BPF_OUT]		= { .type = NLA_NESTED, },
368	[LWT_BPF_XMIT]		= { .type = NLA_NESTED, },
369	[LWT_BPF_XMIT_HEADROOM]	= { .type = NLA_U32 },
370};
371
372static int bpf_build_state(struct nlattr *nla,
373			   unsigned int family, const void *cfg,
374			   struct lwtunnel_state **ts,
375			   struct netlink_ext_ack *extack)
376{
377	struct nlattr *tb[LWT_BPF_MAX + 1];
378	struct lwtunnel_state *newts;
379	struct bpf_lwt *bpf;
380	int ret;
381
382	if (family != AF_INET && family != AF_INET6)
383		return -EAFNOSUPPORT;
384
385	ret = nla_parse_nested_deprecated(tb, LWT_BPF_MAX, nla, bpf_nl_policy,
386					  extack);
387	if (ret < 0)
388		return ret;
389
390	if (!tb[LWT_BPF_IN] && !tb[LWT_BPF_OUT] && !tb[LWT_BPF_XMIT])
391		return -EINVAL;
392
393	newts = lwtunnel_state_alloc(sizeof(*bpf));
394	if (!newts)
395		return -ENOMEM;
396
397	newts->type = LWTUNNEL_ENCAP_BPF;
398	bpf = bpf_lwt_lwtunnel(newts);
399
400	if (tb[LWT_BPF_IN]) {
401		newts->flags |= LWTUNNEL_STATE_INPUT_REDIRECT;
402		ret = bpf_parse_prog(tb[LWT_BPF_IN], &bpf->in,
403				     BPF_PROG_TYPE_LWT_IN);
404		if (ret  < 0)
405			goto errout;
406	}
407
408	if (tb[LWT_BPF_OUT]) {
409		newts->flags |= LWTUNNEL_STATE_OUTPUT_REDIRECT;
410		ret = bpf_parse_prog(tb[LWT_BPF_OUT], &bpf->out,
411				     BPF_PROG_TYPE_LWT_OUT);
412		if (ret < 0)
413			goto errout;
414	}
415
416	if (tb[LWT_BPF_XMIT]) {
417		newts->flags |= LWTUNNEL_STATE_XMIT_REDIRECT;
418		ret = bpf_parse_prog(tb[LWT_BPF_XMIT], &bpf->xmit,
419				     BPF_PROG_TYPE_LWT_XMIT);
420		if (ret < 0)
421			goto errout;
422	}
423
424	if (tb[LWT_BPF_XMIT_HEADROOM]) {
425		u32 headroom = nla_get_u32(tb[LWT_BPF_XMIT_HEADROOM]);
426
427		if (headroom > LWT_BPF_MAX_HEADROOM) {
428			ret = -ERANGE;
429			goto errout;
430		}
431
432		newts->headroom = headroom;
433	}
434
435	bpf->family = family;
436	*ts = newts;
437
438	return 0;
439
440errout:
441	bpf_destroy_state(newts);
442	kfree(newts);
443	return ret;
444}
445
446static int bpf_fill_lwt_prog(struct sk_buff *skb, int attr,
447			     struct bpf_lwt_prog *prog)
448{
449	struct nlattr *nest;
450
451	if (!prog->prog)
452		return 0;
453
454	nest = nla_nest_start_noflag(skb, attr);
455	if (!nest)
456		return -EMSGSIZE;
457
458	if (prog->name &&
459	    nla_put_string(skb, LWT_BPF_PROG_NAME, prog->name))
460		return -EMSGSIZE;
461
462	return nla_nest_end(skb, nest);
463}
464
465static int bpf_fill_encap_info(struct sk_buff *skb, struct lwtunnel_state *lwt)
466{
467	struct bpf_lwt *bpf = bpf_lwt_lwtunnel(lwt);
468
469	if (bpf_fill_lwt_prog(skb, LWT_BPF_IN, &bpf->in) < 0 ||
470	    bpf_fill_lwt_prog(skb, LWT_BPF_OUT, &bpf->out) < 0 ||
471	    bpf_fill_lwt_prog(skb, LWT_BPF_XMIT, &bpf->xmit) < 0)
472		return -EMSGSIZE;
473
474	return 0;
475}
476
477static int bpf_encap_nlsize(struct lwtunnel_state *lwtstate)
478{
479	int nest_len = nla_total_size(sizeof(struct nlattr)) +
480		       nla_total_size(MAX_PROG_NAME) + /* LWT_BPF_PROG_NAME */
481		       0;
482
483	return nest_len + /* LWT_BPF_IN */
484	       nest_len + /* LWT_BPF_OUT */
485	       nest_len + /* LWT_BPF_XMIT */
486	       0;
487}
488
489static int bpf_lwt_prog_cmp(struct bpf_lwt_prog *a, struct bpf_lwt_prog *b)
490{
491	/* FIXME:
492	 * The LWT state is currently rebuilt for delete requests which
493	 * results in a new bpf_prog instance. Comparing names for now.
494	 */
495	if (!a->name && !b->name)
496		return 0;
497
498	if (!a->name || !b->name)
499		return 1;
500
501	return strcmp(a->name, b->name);
502}
503
504static int bpf_encap_cmp(struct lwtunnel_state *a, struct lwtunnel_state *b)
505{
506	struct bpf_lwt *a_bpf = bpf_lwt_lwtunnel(a);
507	struct bpf_lwt *b_bpf = bpf_lwt_lwtunnel(b);
508
509	return bpf_lwt_prog_cmp(&a_bpf->in, &b_bpf->in) ||
510	       bpf_lwt_prog_cmp(&a_bpf->out, &b_bpf->out) ||
511	       bpf_lwt_prog_cmp(&a_bpf->xmit, &b_bpf->xmit);
512}
513
514static const struct lwtunnel_encap_ops bpf_encap_ops = {
515	.build_state	= bpf_build_state,
516	.destroy_state	= bpf_destroy_state,
517	.input		= bpf_input,
518	.output		= bpf_output,
519	.xmit		= bpf_xmit,
520	.fill_encap	= bpf_fill_encap_info,
521	.get_encap_size = bpf_encap_nlsize,
522	.cmp_encap	= bpf_encap_cmp,
523	.owner		= THIS_MODULE,
524};
525
526static int handle_gso_type(struct sk_buff *skb, unsigned int gso_type,
527			   int encap_len)
528{
529	struct skb_shared_info *shinfo = skb_shinfo(skb);
530
531	gso_type |= SKB_GSO_DODGY;
532	shinfo->gso_type |= gso_type;
533	skb_decrease_gso_size(shinfo, encap_len);
534	shinfo->gso_segs = 0;
535	return 0;
536}
537
538static int handle_gso_encap(struct sk_buff *skb, bool ipv4, int encap_len)
539{
540	int next_hdr_offset;
541	void *next_hdr;
542	__u8 protocol;
543
544	/* SCTP and UDP_L4 gso need more nuanced handling than what
545	 * handle_gso_type() does above: skb_decrease_gso_size() is not enough.
546	 * So at the moment only TCP GSO packets are let through.
547	 */
548	if (!(skb_shinfo(skb)->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
549		return -ENOTSUPP;
550
551	if (ipv4) {
552		protocol = ip_hdr(skb)->protocol;
553		next_hdr_offset = sizeof(struct iphdr);
554		next_hdr = skb_network_header(skb) + next_hdr_offset;
555	} else {
556		protocol = ipv6_hdr(skb)->nexthdr;
557		next_hdr_offset = sizeof(struct ipv6hdr);
558		next_hdr = skb_network_header(skb) + next_hdr_offset;
559	}
560
561	switch (protocol) {
562	case IPPROTO_GRE:
563		next_hdr_offset += sizeof(struct gre_base_hdr);
564		if (next_hdr_offset > encap_len)
565			return -EINVAL;
566
567		if (((struct gre_base_hdr *)next_hdr)->flags & GRE_CSUM)
568			return handle_gso_type(skb, SKB_GSO_GRE_CSUM,
569					       encap_len);
570		return handle_gso_type(skb, SKB_GSO_GRE, encap_len);
571
572	case IPPROTO_UDP:
573		next_hdr_offset += sizeof(struct udphdr);
574		if (next_hdr_offset > encap_len)
575			return -EINVAL;
576
577		if (((struct udphdr *)next_hdr)->check)
578			return handle_gso_type(skb, SKB_GSO_UDP_TUNNEL_CSUM,
579					       encap_len);
580		return handle_gso_type(skb, SKB_GSO_UDP_TUNNEL, encap_len);
581
582	case IPPROTO_IP:
583	case IPPROTO_IPV6:
584		if (ipv4)
585			return handle_gso_type(skb, SKB_GSO_IPXIP4, encap_len);
586		else
587			return handle_gso_type(skb, SKB_GSO_IPXIP6, encap_len);
588
589	default:
590		return -EPROTONOSUPPORT;
591	}
592}
593
594int bpf_lwt_push_ip_encap(struct sk_buff *skb, void *hdr, u32 len, bool ingress)
595{
596	struct iphdr *iph;
597	bool ipv4;
598	int err;
599
600	if (unlikely(len < sizeof(struct iphdr) || len > LWT_BPF_MAX_HEADROOM))
601		return -EINVAL;
602
603	/* validate protocol and length */
604	iph = (struct iphdr *)hdr;
605	if (iph->version == 4) {
606		ipv4 = true;
607		if (unlikely(len < iph->ihl * 4))
608			return -EINVAL;
609	} else if (iph->version == 6) {
610		ipv4 = false;
611		if (unlikely(len < sizeof(struct ipv6hdr)))
612			return -EINVAL;
613	} else {
614		return -EINVAL;
615	}
616
617	if (ingress)
618		err = skb_cow_head(skb, len + skb->mac_len);
619	else
620		err = skb_cow_head(skb,
621				   len + LL_RESERVED_SPACE(skb_dst(skb)->dev));
622	if (unlikely(err))
623		return err;
624
625	/* push the encap headers and fix pointers */
626	skb_reset_inner_headers(skb);
627	skb_reset_inner_mac_header(skb);  /* mac header is not yet set */
628	skb_set_inner_protocol(skb, skb->protocol);
629	skb->encapsulation = 1;
630	skb_push(skb, len);
631	if (ingress)
632		skb_postpush_rcsum(skb, iph, len);
633	skb_reset_network_header(skb);
634	memcpy(skb_network_header(skb), hdr, len);
635	bpf_compute_data_pointers(skb);
636	skb_clear_hash(skb);
637
638	if (ipv4) {
639		skb->protocol = htons(ETH_P_IP);
640		iph = ip_hdr(skb);
641
642		if (!iph->check)
643			iph->check = ip_fast_csum((unsigned char *)iph,
644						  iph->ihl);
645	} else {
646		skb->protocol = htons(ETH_P_IPV6);
647	}
648
649	if (skb_is_gso(skb))
650		return handle_gso_encap(skb, ipv4, len);
651
652	return 0;
653}
654
655static int __init bpf_lwt_init(void)
656{
657	return lwtunnel_encap_add_ops(&bpf_encap_ops, LWTUNNEL_ENCAP_BPF);
658}
659
660subsys_initcall(bpf_lwt_init)