• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Linux Socket Filter - Kernel level socket filtering
3  *
4  * Based on the design of the Berkeley Packet Filter. The new
5  * internal format has been designed by PLUMgrid:
6  *
7  *	Copyright (c) 2011 - 2014 PLUMgrid, http://plumgrid.com
8  *
9  * Authors:
10  *
11  *	Jay Schulist <jschlst@samba.org>
12  *	Alexei Starovoitov <ast@plumgrid.com>
13  *	Daniel Borkmann <dborkman@redhat.com>
14  *
15  * This program is free software; you can redistribute it and/or
16  * modify it under the terms of the GNU General Public License
17  * as published by the Free Software Foundation; either version
18  * 2 of the License, or (at your option) any later version.
19  *
20  * Andi Kleen - Fix a few bad bugs and races.
21  * Kris Katterjohn - Added many additional checks in bpf_check_classic()
22  */
23 
24 #include <linux/module.h>
25 #include <linux/types.h>
26 #include <linux/mm.h>
27 #include <linux/fcntl.h>
28 #include <linux/socket.h>
29 #include <linux/sock_diag.h>
30 #include <linux/in.h>
31 #include <linux/inet.h>
32 #include <linux/netdevice.h>
33 #include <linux/if_packet.h>
34 #include <linux/gfp.h>
35 #include <net/ip.h>
36 #include <net/protocol.h>
37 #include <net/netlink.h>
38 #include <linux/skbuff.h>
39 #include <net/sock.h>
40 #include <net/flow_dissector.h>
41 #include <linux/errno.h>
42 #include <linux/timer.h>
43 #include <asm/uaccess.h>
44 #include <asm/unaligned.h>
45 #include <linux/filter.h>
46 #include <linux/ratelimit.h>
47 #include <linux/seccomp.h>
48 #include <linux/if_vlan.h>
49 #include <linux/bpf.h>
50 #include <net/sch_generic.h>
51 #include <net/cls_cgroup.h>
52 #include <net/dst_metadata.h>
53 #include <net/dst.h>
54 #include <net/sock_reuseport.h>
55 
56 /**
57  *	sk_filter_trim_cap - run a packet through a socket filter
58  *	@sk: sock associated with &sk_buff
59  *	@skb: buffer to filter
60  *	@cap: limit on how short the eBPF program may trim the packet
61  *
62  * Run the eBPF program and then cut skb->data to correct size returned by
63  * the program. If pkt_len is 0 we toss packet. If skb->len is smaller
64  * than pkt_len we keep whole skb->data. This is the socket level
65  * wrapper to BPF_PROG_RUN. It returns 0 if the packet should
66  * be accepted or -EPERM if the packet should be tossed.
67  *
68  */
sk_filter_trim_cap(struct sock * sk,struct sk_buff * skb,unsigned int cap)69 int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap)
70 {
71 	int err;
72 	struct sk_filter *filter;
73 
74 	/*
75 	 * If the skb was allocated from pfmemalloc reserves, only
76 	 * allow SOCK_MEMALLOC sockets to use it as this socket is
77 	 * helping free memory
78 	 */
79 	if (skb_pfmemalloc(skb) && !sock_flag(sk, SOCK_MEMALLOC))
80 		return -ENOMEM;
81 
82 	err = BPF_CGROUP_RUN_PROG_INET_INGRESS(sk, skb);
83 	if (err)
84 		return err;
85 
86 	err = security_sock_rcv_skb(sk, skb);
87 	if (err)
88 		return err;
89 
90 	rcu_read_lock();
91 	filter = rcu_dereference(sk->sk_filter);
92 	if (filter) {
93 		struct sock *save_sk = skb->sk;
94 		unsigned int pkt_len;
95 
96 		skb->sk = sk;
97 		pkt_len = bpf_prog_run_save_cb(filter->prog, skb);
98 		skb->sk = save_sk;
99 		err = pkt_len ? pskb_trim(skb, max(cap, pkt_len)) : -EPERM;
100 	}
101 	rcu_read_unlock();
102 
103 	return err;
104 }
105 EXPORT_SYMBOL(sk_filter_trim_cap);
106 
BPF_CALL_1(__skb_get_pay_offset,struct sk_buff *,skb)107 BPF_CALL_1(__skb_get_pay_offset, struct sk_buff *, skb)
108 {
109 	return skb_get_poff(skb);
110 }
111 
BPF_CALL_3(__skb_get_nlattr,struct sk_buff *,skb,u32,a,u32,x)112 BPF_CALL_3(__skb_get_nlattr, struct sk_buff *, skb, u32, a, u32, x)
113 {
114 	struct nlattr *nla;
115 
116 	if (skb_is_nonlinear(skb))
117 		return 0;
118 
119 	if (skb->len < sizeof(struct nlattr))
120 		return 0;
121 
122 	if (a > skb->len - sizeof(struct nlattr))
123 		return 0;
124 
125 	nla = nla_find((struct nlattr *) &skb->data[a], skb->len - a, x);
126 	if (nla)
127 		return (void *) nla - (void *) skb->data;
128 
129 	return 0;
130 }
131 
BPF_CALL_3(__skb_get_nlattr_nest,struct sk_buff *,skb,u32,a,u32,x)132 BPF_CALL_3(__skb_get_nlattr_nest, struct sk_buff *, skb, u32, a, u32, x)
133 {
134 	struct nlattr *nla;
135 
136 	if (skb_is_nonlinear(skb))
137 		return 0;
138 
139 	if (skb->len < sizeof(struct nlattr))
140 		return 0;
141 
142 	if (a > skb->len - sizeof(struct nlattr))
143 		return 0;
144 
145 	nla = (struct nlattr *) &skb->data[a];
146 	if (nla->nla_len > skb->len - a)
147 		return 0;
148 
149 	nla = nla_find_nested(nla, x);
150 	if (nla)
151 		return (void *) nla - (void *) skb->data;
152 
153 	return 0;
154 }
155 
BPF_CALL_0(__get_raw_cpu_id)156 BPF_CALL_0(__get_raw_cpu_id)
157 {
158 	return raw_smp_processor_id();
159 }
160 
161 static const struct bpf_func_proto bpf_get_raw_smp_processor_id_proto = {
162 	.func		= __get_raw_cpu_id,
163 	.gpl_only	= false,
164 	.ret_type	= RET_INTEGER,
165 };
166 
convert_skb_access(int skb_field,int dst_reg,int src_reg,struct bpf_insn * insn_buf)167 static u32 convert_skb_access(int skb_field, int dst_reg, int src_reg,
168 			      struct bpf_insn *insn_buf)
169 {
170 	struct bpf_insn *insn = insn_buf;
171 
172 	switch (skb_field) {
173 	case SKF_AD_MARK:
174 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
175 
176 		*insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
177 				      offsetof(struct sk_buff, mark));
178 		break;
179 
180 	case SKF_AD_PKTTYPE:
181 		*insn++ = BPF_LDX_MEM(BPF_B, dst_reg, src_reg, PKT_TYPE_OFFSET());
182 		*insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, PKT_TYPE_MAX);
183 #ifdef __BIG_ENDIAN_BITFIELD
184 		*insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 5);
185 #endif
186 		break;
187 
188 	case SKF_AD_QUEUE:
189 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
190 
191 		*insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
192 				      offsetof(struct sk_buff, queue_mapping));
193 		break;
194 
195 	case SKF_AD_VLAN_TAG:
196 	case SKF_AD_VLAN_TAG_PRESENT:
197 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
198 		BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
199 
200 		/* dst_reg = *(u16 *) (src_reg + offsetof(vlan_tci)) */
201 		*insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
202 				      offsetof(struct sk_buff, vlan_tci));
203 		if (skb_field == SKF_AD_VLAN_TAG) {
204 			*insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg,
205 						~VLAN_TAG_PRESENT);
206 		} else {
207 			/* dst_reg >>= 12 */
208 			*insn++ = BPF_ALU32_IMM(BPF_RSH, dst_reg, 12);
209 			/* dst_reg &= 1 */
210 			*insn++ = BPF_ALU32_IMM(BPF_AND, dst_reg, 1);
211 		}
212 		break;
213 	}
214 
215 	return insn - insn_buf;
216 }
217 
convert_bpf_extensions(struct sock_filter * fp,struct bpf_insn ** insnp)218 static bool convert_bpf_extensions(struct sock_filter *fp,
219 				   struct bpf_insn **insnp)
220 {
221 	struct bpf_insn *insn = *insnp;
222 	u32 cnt;
223 
224 	switch (fp->k) {
225 	case SKF_AD_OFF + SKF_AD_PROTOCOL:
226 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
227 
228 		/* A = *(u16 *) (CTX + offsetof(protocol)) */
229 		*insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
230 				      offsetof(struct sk_buff, protocol));
231 		/* A = ntohs(A) [emitting a nop or swap16] */
232 		*insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
233 		break;
234 
235 	case SKF_AD_OFF + SKF_AD_PKTTYPE:
236 		cnt = convert_skb_access(SKF_AD_PKTTYPE, BPF_REG_A, BPF_REG_CTX, insn);
237 		insn += cnt - 1;
238 		break;
239 
240 	case SKF_AD_OFF + SKF_AD_IFINDEX:
241 	case SKF_AD_OFF + SKF_AD_HATYPE:
242 		BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
243 		BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, type) != 2);
244 
245 		*insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
246 				      BPF_REG_TMP, BPF_REG_CTX,
247 				      offsetof(struct sk_buff, dev));
248 		/* if (tmp != 0) goto pc + 1 */
249 		*insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_TMP, 0, 1);
250 		*insn++ = BPF_EXIT_INSN();
251 		if (fp->k == SKF_AD_OFF + SKF_AD_IFINDEX)
252 			*insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_TMP,
253 					    offsetof(struct net_device, ifindex));
254 		else
255 			*insn = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_TMP,
256 					    offsetof(struct net_device, type));
257 		break;
258 
259 	case SKF_AD_OFF + SKF_AD_MARK:
260 		cnt = convert_skb_access(SKF_AD_MARK, BPF_REG_A, BPF_REG_CTX, insn);
261 		insn += cnt - 1;
262 		break;
263 
264 	case SKF_AD_OFF + SKF_AD_RXHASH:
265 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
266 
267 		*insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX,
268 				    offsetof(struct sk_buff, hash));
269 		break;
270 
271 	case SKF_AD_OFF + SKF_AD_QUEUE:
272 		cnt = convert_skb_access(SKF_AD_QUEUE, BPF_REG_A, BPF_REG_CTX, insn);
273 		insn += cnt - 1;
274 		break;
275 
276 	case SKF_AD_OFF + SKF_AD_VLAN_TAG:
277 		cnt = convert_skb_access(SKF_AD_VLAN_TAG,
278 					 BPF_REG_A, BPF_REG_CTX, insn);
279 		insn += cnt - 1;
280 		break;
281 
282 	case SKF_AD_OFF + SKF_AD_VLAN_TAG_PRESENT:
283 		cnt = convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
284 					 BPF_REG_A, BPF_REG_CTX, insn);
285 		insn += cnt - 1;
286 		break;
287 
288 	case SKF_AD_OFF + SKF_AD_VLAN_TPID:
289 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
290 
291 		/* A = *(u16 *) (CTX + offsetof(vlan_proto)) */
292 		*insn++ = BPF_LDX_MEM(BPF_H, BPF_REG_A, BPF_REG_CTX,
293 				      offsetof(struct sk_buff, vlan_proto));
294 		/* A = ntohs(A) [emitting a nop or swap16] */
295 		*insn = BPF_ENDIAN(BPF_FROM_BE, BPF_REG_A, 16);
296 		break;
297 
298 	case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
299 	case SKF_AD_OFF + SKF_AD_NLATTR:
300 	case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
301 	case SKF_AD_OFF + SKF_AD_CPU:
302 	case SKF_AD_OFF + SKF_AD_RANDOM:
303 		/* arg1 = CTX */
304 		*insn++ = BPF_MOV64_REG(BPF_REG_ARG1, BPF_REG_CTX);
305 		/* arg2 = A */
306 		*insn++ = BPF_MOV64_REG(BPF_REG_ARG2, BPF_REG_A);
307 		/* arg3 = X */
308 		*insn++ = BPF_MOV64_REG(BPF_REG_ARG3, BPF_REG_X);
309 		/* Emit call(arg1=CTX, arg2=A, arg3=X) */
310 		switch (fp->k) {
311 		case SKF_AD_OFF + SKF_AD_PAY_OFFSET:
312 			*insn = BPF_EMIT_CALL(__skb_get_pay_offset);
313 			break;
314 		case SKF_AD_OFF + SKF_AD_NLATTR:
315 			*insn = BPF_EMIT_CALL(__skb_get_nlattr);
316 			break;
317 		case SKF_AD_OFF + SKF_AD_NLATTR_NEST:
318 			*insn = BPF_EMIT_CALL(__skb_get_nlattr_nest);
319 			break;
320 		case SKF_AD_OFF + SKF_AD_CPU:
321 			*insn = BPF_EMIT_CALL(__get_raw_cpu_id);
322 			break;
323 		case SKF_AD_OFF + SKF_AD_RANDOM:
324 			*insn = BPF_EMIT_CALL(bpf_user_rnd_u32);
325 			bpf_user_rnd_init_once();
326 			break;
327 		}
328 		break;
329 
330 	case SKF_AD_OFF + SKF_AD_ALU_XOR_X:
331 		/* A ^= X */
332 		*insn = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_X);
333 		break;
334 
335 	default:
336 		/* This is just a dummy call to avoid letting the compiler
337 		 * evict __bpf_call_base() as an optimization. Placed here
338 		 * where no-one bothers.
339 		 */
340 		BUG_ON(__bpf_call_base(0, 0, 0, 0, 0) != 0);
341 		return false;
342 	}
343 
344 	*insnp = insn;
345 	return true;
346 }
347 
348 /**
349  *	bpf_convert_filter - convert filter program
350  *	@prog: the user passed filter program
351  *	@len: the length of the user passed filter program
352  *	@new_prog: buffer where converted program will be stored
353  *	@new_len: pointer to store length of converted program
354  *
355  * Remap 'sock_filter' style BPF instruction set to 'sock_filter_ext' style.
356  * Conversion workflow:
357  *
358  * 1) First pass for calculating the new program length:
359  *   bpf_convert_filter(old_prog, old_len, NULL, &new_len)
360  *
361  * 2) 2nd pass to remap in two passes: 1st pass finds new
362  *    jump offsets, 2nd pass remapping:
363  *   new_prog = kmalloc(sizeof(struct bpf_insn) * new_len);
364  *   bpf_convert_filter(old_prog, old_len, new_prog, &new_len);
365  */
bpf_convert_filter(struct sock_filter * prog,int len,struct bpf_insn * new_prog,int * new_len)366 static int bpf_convert_filter(struct sock_filter *prog, int len,
367 			      struct bpf_insn *new_prog, int *new_len)
368 {
369 	int new_flen = 0, pass = 0, target, i;
370 	struct bpf_insn *new_insn;
371 	struct sock_filter *fp;
372 	int *addrs = NULL;
373 	u8 bpf_src;
374 
375 	BUILD_BUG_ON(BPF_MEMWORDS * sizeof(u32) > MAX_BPF_STACK);
376 	BUILD_BUG_ON(BPF_REG_FP + 1 != MAX_BPF_REG);
377 
378 	if (len <= 0 || len > BPF_MAXINSNS)
379 		return -EINVAL;
380 
381 	if (new_prog) {
382 		addrs = kcalloc(len, sizeof(*addrs),
383 				GFP_KERNEL | __GFP_NOWARN);
384 		if (!addrs)
385 			return -ENOMEM;
386 	}
387 
388 do_pass:
389 	new_insn = new_prog;
390 	fp = prog;
391 
392 	/* Classic BPF related prologue emission. */
393 	if (new_insn) {
394 		/* Classic BPF expects A and X to be reset first. These need
395 		 * to be guaranteed to be the first two instructions.
396 		 */
397 		*new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_A, BPF_REG_A);
398 		*new_insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_X, BPF_REG_X);
399 
400 		/* All programs must keep CTX in callee saved BPF_REG_CTX.
401 		 * In eBPF case it's done by the compiler, here we need to
402 		 * do this ourself. Initial CTX is present in BPF_REG_ARG1.
403 		 */
404 		*new_insn++ = BPF_MOV64_REG(BPF_REG_CTX, BPF_REG_ARG1);
405 	} else {
406 		new_insn += 3;
407 	}
408 
409 	for (i = 0; i < len; fp++, i++) {
410 		struct bpf_insn tmp_insns[6] = { };
411 		struct bpf_insn *insn = tmp_insns;
412 
413 		if (addrs)
414 			addrs[i] = new_insn - new_prog;
415 
416 		switch (fp->code) {
417 		/* All arithmetic insns and skb loads map as-is. */
418 		case BPF_ALU | BPF_ADD | BPF_X:
419 		case BPF_ALU | BPF_ADD | BPF_K:
420 		case BPF_ALU | BPF_SUB | BPF_X:
421 		case BPF_ALU | BPF_SUB | BPF_K:
422 		case BPF_ALU | BPF_AND | BPF_X:
423 		case BPF_ALU | BPF_AND | BPF_K:
424 		case BPF_ALU | BPF_OR | BPF_X:
425 		case BPF_ALU | BPF_OR | BPF_K:
426 		case BPF_ALU | BPF_LSH | BPF_X:
427 		case BPF_ALU | BPF_LSH | BPF_K:
428 		case BPF_ALU | BPF_RSH | BPF_X:
429 		case BPF_ALU | BPF_RSH | BPF_K:
430 		case BPF_ALU | BPF_XOR | BPF_X:
431 		case BPF_ALU | BPF_XOR | BPF_K:
432 		case BPF_ALU | BPF_MUL | BPF_X:
433 		case BPF_ALU | BPF_MUL | BPF_K:
434 		case BPF_ALU | BPF_DIV | BPF_X:
435 		case BPF_ALU | BPF_DIV | BPF_K:
436 		case BPF_ALU | BPF_MOD | BPF_X:
437 		case BPF_ALU | BPF_MOD | BPF_K:
438 		case BPF_ALU | BPF_NEG:
439 		case BPF_LD | BPF_ABS | BPF_W:
440 		case BPF_LD | BPF_ABS | BPF_H:
441 		case BPF_LD | BPF_ABS | BPF_B:
442 		case BPF_LD | BPF_IND | BPF_W:
443 		case BPF_LD | BPF_IND | BPF_H:
444 		case BPF_LD | BPF_IND | BPF_B:
445 			/* Check for overloaded BPF extension and
446 			 * directly convert it if found, otherwise
447 			 * just move on with mapping.
448 			 */
449 			if (BPF_CLASS(fp->code) == BPF_LD &&
450 			    BPF_MODE(fp->code) == BPF_ABS &&
451 			    convert_bpf_extensions(fp, &insn))
452 				break;
453 
454 			if (fp->code == (BPF_ALU | BPF_DIV | BPF_X) ||
455 			    fp->code == (BPF_ALU | BPF_MOD | BPF_X))
456 				*insn++ = BPF_MOV32_REG(BPF_REG_X, BPF_REG_X);
457 
458 			*insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k);
459 			break;
460 
461 		/* Jump transformation cannot use BPF block macros
462 		 * everywhere as offset calculation and target updates
463 		 * require a bit more work than the rest, i.e. jump
464 		 * opcodes map as-is, but offsets need adjustment.
465 		 */
466 
467 #define BPF_EMIT_JMP							\
468 	do {								\
469 		if (target >= len || target < 0)			\
470 			goto err;					\
471 		insn->off = addrs ? addrs[target] - addrs[i] - 1 : 0;	\
472 		/* Adjust pc relative offset for 2nd or 3rd insn. */	\
473 		insn->off -= insn - tmp_insns;				\
474 	} while (0)
475 
476 		case BPF_JMP | BPF_JA:
477 			target = i + fp->k + 1;
478 			insn->code = fp->code;
479 			BPF_EMIT_JMP;
480 			break;
481 
482 		case BPF_JMP | BPF_JEQ | BPF_K:
483 		case BPF_JMP | BPF_JEQ | BPF_X:
484 		case BPF_JMP | BPF_JSET | BPF_K:
485 		case BPF_JMP | BPF_JSET | BPF_X:
486 		case BPF_JMP | BPF_JGT | BPF_K:
487 		case BPF_JMP | BPF_JGT | BPF_X:
488 		case BPF_JMP | BPF_JGE | BPF_K:
489 		case BPF_JMP | BPF_JGE | BPF_X:
490 			if (BPF_SRC(fp->code) == BPF_K && (int) fp->k < 0) {
491 				/* BPF immediates are signed, zero extend
492 				 * immediate into tmp register and use it
493 				 * in compare insn.
494 				 */
495 				*insn++ = BPF_MOV32_IMM(BPF_REG_TMP, fp->k);
496 
497 				insn->dst_reg = BPF_REG_A;
498 				insn->src_reg = BPF_REG_TMP;
499 				bpf_src = BPF_X;
500 			} else {
501 				insn->dst_reg = BPF_REG_A;
502 				insn->imm = fp->k;
503 				bpf_src = BPF_SRC(fp->code);
504 				insn->src_reg = bpf_src == BPF_X ? BPF_REG_X : 0;
505 			}
506 
507 			/* Common case where 'jump_false' is next insn. */
508 			if (fp->jf == 0) {
509 				insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
510 				target = i + fp->jt + 1;
511 				BPF_EMIT_JMP;
512 				break;
513 			}
514 
515 			/* Convert JEQ into JNE when 'jump_true' is next insn. */
516 			if (fp->jt == 0 && BPF_OP(fp->code) == BPF_JEQ) {
517 				insn->code = BPF_JMP | BPF_JNE | bpf_src;
518 				target = i + fp->jf + 1;
519 				BPF_EMIT_JMP;
520 				break;
521 			}
522 
523 			/* Other jumps are mapped into two insns: Jxx and JA. */
524 			target = i + fp->jt + 1;
525 			insn->code = BPF_JMP | BPF_OP(fp->code) | bpf_src;
526 			BPF_EMIT_JMP;
527 			insn++;
528 
529 			insn->code = BPF_JMP | BPF_JA;
530 			target = i + fp->jf + 1;
531 			BPF_EMIT_JMP;
532 			break;
533 
534 		/* ldxb 4 * ([14] & 0xf) is remaped into 6 insns. */
535 		case BPF_LDX | BPF_MSH | BPF_B:
536 			/* tmp = A */
537 			*insn++ = BPF_MOV64_REG(BPF_REG_TMP, BPF_REG_A);
538 			/* A = BPF_R0 = *(u8 *) (skb->data + K) */
539 			*insn++ = BPF_LD_ABS(BPF_B, fp->k);
540 			/* A &= 0xf */
541 			*insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_A, 0xf);
542 			/* A <<= 2 */
543 			*insn++ = BPF_ALU32_IMM(BPF_LSH, BPF_REG_A, 2);
544 			/* X = A */
545 			*insn++ = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
546 			/* A = tmp */
547 			*insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_TMP);
548 			break;
549 
550 		/* RET_K is remaped into 2 insns. RET_A case doesn't need an
551 		 * extra mov as BPF_REG_0 is already mapped into BPF_REG_A.
552 		 */
553 		case BPF_RET | BPF_A:
554 		case BPF_RET | BPF_K:
555 			if (BPF_RVAL(fp->code) == BPF_K)
556 				*insn++ = BPF_MOV32_RAW(BPF_K, BPF_REG_0,
557 							0, fp->k);
558 			*insn = BPF_EXIT_INSN();
559 			break;
560 
561 		/* Store to stack. */
562 		case BPF_ST:
563 		case BPF_STX:
564 			*insn = BPF_STX_MEM(BPF_W, BPF_REG_FP, BPF_CLASS(fp->code) ==
565 					    BPF_ST ? BPF_REG_A : BPF_REG_X,
566 					    -(BPF_MEMWORDS - fp->k) * 4);
567 			break;
568 
569 		/* Load from stack. */
570 		case BPF_LD | BPF_MEM:
571 		case BPF_LDX | BPF_MEM:
572 			*insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD  ?
573 					    BPF_REG_A : BPF_REG_X, BPF_REG_FP,
574 					    -(BPF_MEMWORDS - fp->k) * 4);
575 			break;
576 
577 		/* A = K or X = K */
578 		case BPF_LD | BPF_IMM:
579 		case BPF_LDX | BPF_IMM:
580 			*insn = BPF_MOV32_IMM(BPF_CLASS(fp->code) == BPF_LD ?
581 					      BPF_REG_A : BPF_REG_X, fp->k);
582 			break;
583 
584 		/* X = A */
585 		case BPF_MISC | BPF_TAX:
586 			*insn = BPF_MOV64_REG(BPF_REG_X, BPF_REG_A);
587 			break;
588 
589 		/* A = X */
590 		case BPF_MISC | BPF_TXA:
591 			*insn = BPF_MOV64_REG(BPF_REG_A, BPF_REG_X);
592 			break;
593 
594 		/* A = skb->len or X = skb->len */
595 		case BPF_LD | BPF_W | BPF_LEN:
596 		case BPF_LDX | BPF_W | BPF_LEN:
597 			*insn = BPF_LDX_MEM(BPF_W, BPF_CLASS(fp->code) == BPF_LD ?
598 					    BPF_REG_A : BPF_REG_X, BPF_REG_CTX,
599 					    offsetof(struct sk_buff, len));
600 			break;
601 
602 		/* Access seccomp_data fields. */
603 		case BPF_LDX | BPF_ABS | BPF_W:
604 			/* A = *(u32 *) (ctx + K) */
605 			*insn = BPF_LDX_MEM(BPF_W, BPF_REG_A, BPF_REG_CTX, fp->k);
606 			break;
607 
608 		/* Unknown instruction. */
609 		default:
610 			goto err;
611 		}
612 
613 		insn++;
614 		if (new_prog)
615 			memcpy(new_insn, tmp_insns,
616 			       sizeof(*insn) * (insn - tmp_insns));
617 		new_insn += insn - tmp_insns;
618 	}
619 
620 	if (!new_prog) {
621 		/* Only calculating new length. */
622 		*new_len = new_insn - new_prog;
623 		return 0;
624 	}
625 
626 	pass++;
627 	if (new_flen != new_insn - new_prog) {
628 		new_flen = new_insn - new_prog;
629 		if (pass > 2)
630 			goto err;
631 		goto do_pass;
632 	}
633 
634 	kfree(addrs);
635 	BUG_ON(*new_len != new_flen);
636 	return 0;
637 err:
638 	kfree(addrs);
639 	return -EINVAL;
640 }
641 
642 /* Security:
643  *
644  * As we dont want to clear mem[] array for each packet going through
645  * __bpf_prog_run(), we check that filter loaded by user never try to read
646  * a cell if not previously written, and we check all branches to be sure
647  * a malicious user doesn't try to abuse us.
648  */
check_load_and_stores(const struct sock_filter * filter,int flen)649 static int check_load_and_stores(const struct sock_filter *filter, int flen)
650 {
651 	u16 *masks, memvalid = 0; /* One bit per cell, 16 cells */
652 	int pc, ret = 0;
653 
654 	BUILD_BUG_ON(BPF_MEMWORDS > 16);
655 
656 	masks = kmalloc_array(flen, sizeof(*masks), GFP_KERNEL);
657 	if (!masks)
658 		return -ENOMEM;
659 
660 	memset(masks, 0xff, flen * sizeof(*masks));
661 
662 	for (pc = 0; pc < flen; pc++) {
663 		memvalid &= masks[pc];
664 
665 		switch (filter[pc].code) {
666 		case BPF_ST:
667 		case BPF_STX:
668 			memvalid |= (1 << filter[pc].k);
669 			break;
670 		case BPF_LD | BPF_MEM:
671 		case BPF_LDX | BPF_MEM:
672 			if (!(memvalid & (1 << filter[pc].k))) {
673 				ret = -EINVAL;
674 				goto error;
675 			}
676 			break;
677 		case BPF_JMP | BPF_JA:
678 			/* A jump must set masks on target */
679 			masks[pc + 1 + filter[pc].k] &= memvalid;
680 			memvalid = ~0;
681 			break;
682 		case BPF_JMP | BPF_JEQ | BPF_K:
683 		case BPF_JMP | BPF_JEQ | BPF_X:
684 		case BPF_JMP | BPF_JGE | BPF_K:
685 		case BPF_JMP | BPF_JGE | BPF_X:
686 		case BPF_JMP | BPF_JGT | BPF_K:
687 		case BPF_JMP | BPF_JGT | BPF_X:
688 		case BPF_JMP | BPF_JSET | BPF_K:
689 		case BPF_JMP | BPF_JSET | BPF_X:
690 			/* A jump must set masks on targets */
691 			masks[pc + 1 + filter[pc].jt] &= memvalid;
692 			masks[pc + 1 + filter[pc].jf] &= memvalid;
693 			memvalid = ~0;
694 			break;
695 		}
696 	}
697 error:
698 	kfree(masks);
699 	return ret;
700 }
701 
chk_code_allowed(u16 code_to_probe)702 static bool chk_code_allowed(u16 code_to_probe)
703 {
704 	static const bool codes[] = {
705 		/* 32 bit ALU operations */
706 		[BPF_ALU | BPF_ADD | BPF_K] = true,
707 		[BPF_ALU | BPF_ADD | BPF_X] = true,
708 		[BPF_ALU | BPF_SUB | BPF_K] = true,
709 		[BPF_ALU | BPF_SUB | BPF_X] = true,
710 		[BPF_ALU | BPF_MUL | BPF_K] = true,
711 		[BPF_ALU | BPF_MUL | BPF_X] = true,
712 		[BPF_ALU | BPF_DIV | BPF_K] = true,
713 		[BPF_ALU | BPF_DIV | BPF_X] = true,
714 		[BPF_ALU | BPF_MOD | BPF_K] = true,
715 		[BPF_ALU | BPF_MOD | BPF_X] = true,
716 		[BPF_ALU | BPF_AND | BPF_K] = true,
717 		[BPF_ALU | BPF_AND | BPF_X] = true,
718 		[BPF_ALU | BPF_OR | BPF_K] = true,
719 		[BPF_ALU | BPF_OR | BPF_X] = true,
720 		[BPF_ALU | BPF_XOR | BPF_K] = true,
721 		[BPF_ALU | BPF_XOR | BPF_X] = true,
722 		[BPF_ALU | BPF_LSH | BPF_K] = true,
723 		[BPF_ALU | BPF_LSH | BPF_X] = true,
724 		[BPF_ALU | BPF_RSH | BPF_K] = true,
725 		[BPF_ALU | BPF_RSH | BPF_X] = true,
726 		[BPF_ALU | BPF_NEG] = true,
727 		/* Load instructions */
728 		[BPF_LD | BPF_W | BPF_ABS] = true,
729 		[BPF_LD | BPF_H | BPF_ABS] = true,
730 		[BPF_LD | BPF_B | BPF_ABS] = true,
731 		[BPF_LD | BPF_W | BPF_LEN] = true,
732 		[BPF_LD | BPF_W | BPF_IND] = true,
733 		[BPF_LD | BPF_H | BPF_IND] = true,
734 		[BPF_LD | BPF_B | BPF_IND] = true,
735 		[BPF_LD | BPF_IMM] = true,
736 		[BPF_LD | BPF_MEM] = true,
737 		[BPF_LDX | BPF_W | BPF_LEN] = true,
738 		[BPF_LDX | BPF_B | BPF_MSH] = true,
739 		[BPF_LDX | BPF_IMM] = true,
740 		[BPF_LDX | BPF_MEM] = true,
741 		/* Store instructions */
742 		[BPF_ST] = true,
743 		[BPF_STX] = true,
744 		/* Misc instructions */
745 		[BPF_MISC | BPF_TAX] = true,
746 		[BPF_MISC | BPF_TXA] = true,
747 		/* Return instructions */
748 		[BPF_RET | BPF_K] = true,
749 		[BPF_RET | BPF_A] = true,
750 		/* Jump instructions */
751 		[BPF_JMP | BPF_JA] = true,
752 		[BPF_JMP | BPF_JEQ | BPF_K] = true,
753 		[BPF_JMP | BPF_JEQ | BPF_X] = true,
754 		[BPF_JMP | BPF_JGE | BPF_K] = true,
755 		[BPF_JMP | BPF_JGE | BPF_X] = true,
756 		[BPF_JMP | BPF_JGT | BPF_K] = true,
757 		[BPF_JMP | BPF_JGT | BPF_X] = true,
758 		[BPF_JMP | BPF_JSET | BPF_K] = true,
759 		[BPF_JMP | BPF_JSET | BPF_X] = true,
760 	};
761 
762 	if (code_to_probe >= ARRAY_SIZE(codes))
763 		return false;
764 
765 	return codes[code_to_probe];
766 }
767 
bpf_check_basics_ok(const struct sock_filter * filter,unsigned int flen)768 static bool bpf_check_basics_ok(const struct sock_filter *filter,
769 				unsigned int flen)
770 {
771 	if (filter == NULL)
772 		return false;
773 	if (flen == 0 || flen > BPF_MAXINSNS)
774 		return false;
775 
776 	return true;
777 }
778 
779 /**
780  *	bpf_check_classic - verify socket filter code
781  *	@filter: filter to verify
782  *	@flen: length of filter
783  *
784  * Check the user's filter code. If we let some ugly
785  * filter code slip through kaboom! The filter must contain
786  * no references or jumps that are out of range, no illegal
787  * instructions, and must end with a RET instruction.
788  *
789  * All jumps are forward as they are not signed.
790  *
791  * Returns 0 if the rule set is legal or -EINVAL if not.
792  */
bpf_check_classic(const struct sock_filter * filter,unsigned int flen)793 static int bpf_check_classic(const struct sock_filter *filter,
794 			     unsigned int flen)
795 {
796 	bool anc_found;
797 	int pc;
798 
799 	/* Check the filter code now */
800 	for (pc = 0; pc < flen; pc++) {
801 		const struct sock_filter *ftest = &filter[pc];
802 
803 		/* May we actually operate on this code? */
804 		if (!chk_code_allowed(ftest->code))
805 			return -EINVAL;
806 
807 		/* Some instructions need special checks */
808 		switch (ftest->code) {
809 		case BPF_ALU | BPF_DIV | BPF_K:
810 		case BPF_ALU | BPF_MOD | BPF_K:
811 			/* Check for division by zero */
812 			if (ftest->k == 0)
813 				return -EINVAL;
814 			break;
815 		case BPF_ALU | BPF_LSH | BPF_K:
816 		case BPF_ALU | BPF_RSH | BPF_K:
817 			if (ftest->k >= 32)
818 				return -EINVAL;
819 			break;
820 		case BPF_LD | BPF_MEM:
821 		case BPF_LDX | BPF_MEM:
822 		case BPF_ST:
823 		case BPF_STX:
824 			/* Check for invalid memory addresses */
825 			if (ftest->k >= BPF_MEMWORDS)
826 				return -EINVAL;
827 			break;
828 		case BPF_JMP | BPF_JA:
829 			/* Note, the large ftest->k might cause loops.
830 			 * Compare this with conditional jumps below,
831 			 * where offsets are limited. --ANK (981016)
832 			 */
833 			if (ftest->k >= (unsigned int)(flen - pc - 1))
834 				return -EINVAL;
835 			break;
836 		case BPF_JMP | BPF_JEQ | BPF_K:
837 		case BPF_JMP | BPF_JEQ | BPF_X:
838 		case BPF_JMP | BPF_JGE | BPF_K:
839 		case BPF_JMP | BPF_JGE | BPF_X:
840 		case BPF_JMP | BPF_JGT | BPF_K:
841 		case BPF_JMP | BPF_JGT | BPF_X:
842 		case BPF_JMP | BPF_JSET | BPF_K:
843 		case BPF_JMP | BPF_JSET | BPF_X:
844 			/* Both conditionals must be safe */
845 			if (pc + ftest->jt + 1 >= flen ||
846 			    pc + ftest->jf + 1 >= flen)
847 				return -EINVAL;
848 			break;
849 		case BPF_LD | BPF_W | BPF_ABS:
850 		case BPF_LD | BPF_H | BPF_ABS:
851 		case BPF_LD | BPF_B | BPF_ABS:
852 			anc_found = false;
853 			if (bpf_anc_helper(ftest) & BPF_ANC)
854 				anc_found = true;
855 			/* Ancillary operation unknown or unsupported */
856 			if (anc_found == false && ftest->k >= SKF_AD_OFF)
857 				return -EINVAL;
858 		}
859 	}
860 
861 	/* Last instruction must be a RET code */
862 	switch (filter[flen - 1].code) {
863 	case BPF_RET | BPF_K:
864 	case BPF_RET | BPF_A:
865 		return check_load_and_stores(filter, flen);
866 	}
867 
868 	return -EINVAL;
869 }
870 
bpf_prog_store_orig_filter(struct bpf_prog * fp,const struct sock_fprog * fprog)871 static int bpf_prog_store_orig_filter(struct bpf_prog *fp,
872 				      const struct sock_fprog *fprog)
873 {
874 	unsigned int fsize = bpf_classic_proglen(fprog);
875 	struct sock_fprog_kern *fkprog;
876 
877 	fp->orig_prog = kmalloc(sizeof(*fkprog), GFP_KERNEL);
878 	if (!fp->orig_prog)
879 		return -ENOMEM;
880 
881 	fkprog = fp->orig_prog;
882 	fkprog->len = fprog->len;
883 
884 	fkprog->filter = kmemdup(fp->insns, fsize,
885 				 GFP_KERNEL | __GFP_NOWARN);
886 	if (!fkprog->filter) {
887 		kfree(fp->orig_prog);
888 		return -ENOMEM;
889 	}
890 
891 	return 0;
892 }
893 
bpf_release_orig_filter(struct bpf_prog * fp)894 static void bpf_release_orig_filter(struct bpf_prog *fp)
895 {
896 	struct sock_fprog_kern *fprog = fp->orig_prog;
897 
898 	if (fprog) {
899 		kfree(fprog->filter);
900 		kfree(fprog);
901 	}
902 }
903 
__bpf_prog_release(struct bpf_prog * prog)904 static void __bpf_prog_release(struct bpf_prog *prog)
905 {
906 	if (prog->type == BPF_PROG_TYPE_SOCKET_FILTER) {
907 		bpf_prog_put(prog);
908 	} else {
909 		bpf_release_orig_filter(prog);
910 		bpf_prog_free(prog);
911 	}
912 }
913 
__sk_filter_release(struct sk_filter * fp)914 static void __sk_filter_release(struct sk_filter *fp)
915 {
916 	__bpf_prog_release(fp->prog);
917 	kfree(fp);
918 }
919 
920 /**
921  * 	sk_filter_release_rcu - Release a socket filter by rcu_head
922  *	@rcu: rcu_head that contains the sk_filter to free
923  */
sk_filter_release_rcu(struct rcu_head * rcu)924 static void sk_filter_release_rcu(struct rcu_head *rcu)
925 {
926 	struct sk_filter *fp = container_of(rcu, struct sk_filter, rcu);
927 
928 	__sk_filter_release(fp);
929 }
930 
931 /**
932  *	sk_filter_release - release a socket filter
933  *	@fp: filter to remove
934  *
935  *	Remove a filter from a socket and release its resources.
936  */
sk_filter_release(struct sk_filter * fp)937 static void sk_filter_release(struct sk_filter *fp)
938 {
939 	if (atomic_dec_and_test(&fp->refcnt))
940 		call_rcu(&fp->rcu, sk_filter_release_rcu);
941 }
942 
sk_filter_uncharge(struct sock * sk,struct sk_filter * fp)943 void sk_filter_uncharge(struct sock *sk, struct sk_filter *fp)
944 {
945 	u32 filter_size = bpf_prog_size(fp->prog->len);
946 
947 	atomic_sub(filter_size, &sk->sk_omem_alloc);
948 	sk_filter_release(fp);
949 }
950 
951 /* try to charge the socket memory if there is space available
952  * return true on success
953  */
sk_filter_charge(struct sock * sk,struct sk_filter * fp)954 bool sk_filter_charge(struct sock *sk, struct sk_filter *fp)
955 {
956 	u32 filter_size = bpf_prog_size(fp->prog->len);
957 
958 	/* same check as in sock_kmalloc() */
959 	if (filter_size <= sysctl_optmem_max &&
960 	    atomic_read(&sk->sk_omem_alloc) + filter_size < sysctl_optmem_max) {
961 		atomic_inc(&fp->refcnt);
962 		atomic_add(filter_size, &sk->sk_omem_alloc);
963 		return true;
964 	}
965 	return false;
966 }
967 
bpf_migrate_filter(struct bpf_prog * fp)968 static struct bpf_prog *bpf_migrate_filter(struct bpf_prog *fp)
969 {
970 	struct sock_filter *old_prog;
971 	struct bpf_prog *old_fp;
972 	int err, new_len, old_len = fp->len;
973 
974 	/* We are free to overwrite insns et al right here as it
975 	 * won't be used at this point in time anymore internally
976 	 * after the migration to the internal BPF instruction
977 	 * representation.
978 	 */
979 	BUILD_BUG_ON(sizeof(struct sock_filter) !=
980 		     sizeof(struct bpf_insn));
981 
982 	/* Conversion cannot happen on overlapping memory areas,
983 	 * so we need to keep the user BPF around until the 2nd
984 	 * pass. At this time, the user BPF is stored in fp->insns.
985 	 */
986 	old_prog = kmemdup(fp->insns, old_len * sizeof(struct sock_filter),
987 			   GFP_KERNEL | __GFP_NOWARN);
988 	if (!old_prog) {
989 		err = -ENOMEM;
990 		goto out_err;
991 	}
992 
993 	/* 1st pass: calculate the new program length. */
994 	err = bpf_convert_filter(old_prog, old_len, NULL, &new_len);
995 	if (err)
996 		goto out_err_free;
997 
998 	/* Expand fp for appending the new filter representation. */
999 	old_fp = fp;
1000 	fp = bpf_prog_realloc(old_fp, bpf_prog_size(new_len), 0);
1001 	if (!fp) {
1002 		/* The old_fp is still around in case we couldn't
1003 		 * allocate new memory, so uncharge on that one.
1004 		 */
1005 		fp = old_fp;
1006 		err = -ENOMEM;
1007 		goto out_err_free;
1008 	}
1009 
1010 	fp->len = new_len;
1011 
1012 	/* 2nd pass: remap sock_filter insns into bpf_insn insns. */
1013 	err = bpf_convert_filter(old_prog, old_len, fp->insnsi, &new_len);
1014 	if (err)
1015 		/* 2nd bpf_convert_filter() can fail only if it fails
1016 		 * to allocate memory, remapping must succeed. Note,
1017 		 * that at this time old_fp has already been released
1018 		 * by krealloc().
1019 		 */
1020 		goto out_err_free;
1021 
1022 	fp = bpf_prog_select_runtime(fp, &err);
1023 	if (err)
1024 		goto out_err_free;
1025 
1026 	kfree(old_prog);
1027 	return fp;
1028 
1029 out_err_free:
1030 	kfree(old_prog);
1031 out_err:
1032 	__bpf_prog_release(fp);
1033 	return ERR_PTR(err);
1034 }
1035 
bpf_prepare_filter(struct bpf_prog * fp,bpf_aux_classic_check_t trans)1036 static struct bpf_prog *bpf_prepare_filter(struct bpf_prog *fp,
1037 					   bpf_aux_classic_check_t trans)
1038 {
1039 	int err;
1040 
1041 	fp->bpf_func = NULL;
1042 	fp->jited = 0;
1043 
1044 	err = bpf_check_classic(fp->insns, fp->len);
1045 	if (err) {
1046 		__bpf_prog_release(fp);
1047 		return ERR_PTR(err);
1048 	}
1049 
1050 	/* There might be additional checks and transformations
1051 	 * needed on classic filters, f.e. in case of seccomp.
1052 	 */
1053 	if (trans) {
1054 		err = trans(fp->insns, fp->len);
1055 		if (err) {
1056 			__bpf_prog_release(fp);
1057 			return ERR_PTR(err);
1058 		}
1059 	}
1060 
1061 	/* Probe if we can JIT compile the filter and if so, do
1062 	 * the compilation of the filter.
1063 	 */
1064 	bpf_jit_compile(fp);
1065 
1066 	/* JIT compiler couldn't process this filter, so do the
1067 	 * internal BPF translation for the optimized interpreter.
1068 	 */
1069 	if (!fp->jited)
1070 		fp = bpf_migrate_filter(fp);
1071 
1072 	return fp;
1073 }
1074 
1075 /**
1076  *	bpf_prog_create - create an unattached filter
1077  *	@pfp: the unattached filter that is created
1078  *	@fprog: the filter program
1079  *
1080  * Create a filter independent of any socket. We first run some
1081  * sanity checks on it to make sure it does not explode on us later.
1082  * If an error occurs or there is insufficient memory for the filter
1083  * a negative errno code is returned. On success the return is zero.
1084  */
bpf_prog_create(struct bpf_prog ** pfp,struct sock_fprog_kern * fprog)1085 int bpf_prog_create(struct bpf_prog **pfp, struct sock_fprog_kern *fprog)
1086 {
1087 	unsigned int fsize = bpf_classic_proglen(fprog);
1088 	struct bpf_prog *fp;
1089 
1090 	/* Make sure new filter is there and in the right amounts. */
1091 	if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1092 		return -EINVAL;
1093 
1094 	fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1095 	if (!fp)
1096 		return -ENOMEM;
1097 
1098 	memcpy(fp->insns, fprog->filter, fsize);
1099 
1100 	fp->len = fprog->len;
1101 	/* Since unattached filters are not copied back to user
1102 	 * space through sk_get_filter(), we do not need to hold
1103 	 * a copy here, and can spare us the work.
1104 	 */
1105 	fp->orig_prog = NULL;
1106 
1107 	/* bpf_prepare_filter() already takes care of freeing
1108 	 * memory in case something goes wrong.
1109 	 */
1110 	fp = bpf_prepare_filter(fp, NULL);
1111 	if (IS_ERR(fp))
1112 		return PTR_ERR(fp);
1113 
1114 	*pfp = fp;
1115 	return 0;
1116 }
1117 EXPORT_SYMBOL_GPL(bpf_prog_create);
1118 
1119 /**
1120  *	bpf_prog_create_from_user - create an unattached filter from user buffer
1121  *	@pfp: the unattached filter that is created
1122  *	@fprog: the filter program
1123  *	@trans: post-classic verifier transformation handler
1124  *	@save_orig: save classic BPF program
1125  *
1126  * This function effectively does the same as bpf_prog_create(), only
1127  * that it builds up its insns buffer from user space provided buffer.
1128  * It also allows for passing a bpf_aux_classic_check_t handler.
1129  */
bpf_prog_create_from_user(struct bpf_prog ** pfp,struct sock_fprog * fprog,bpf_aux_classic_check_t trans,bool save_orig)1130 int bpf_prog_create_from_user(struct bpf_prog **pfp, struct sock_fprog *fprog,
1131 			      bpf_aux_classic_check_t trans, bool save_orig)
1132 {
1133 	unsigned int fsize = bpf_classic_proglen(fprog);
1134 	struct bpf_prog *fp;
1135 	int err;
1136 
1137 	/* Make sure new filter is there and in the right amounts. */
1138 	if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1139 		return -EINVAL;
1140 
1141 	fp = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1142 	if (!fp)
1143 		return -ENOMEM;
1144 
1145 	if (copy_from_user(fp->insns, fprog->filter, fsize)) {
1146 		__bpf_prog_free(fp);
1147 		return -EFAULT;
1148 	}
1149 
1150 	fp->len = fprog->len;
1151 	fp->orig_prog = NULL;
1152 
1153 	if (save_orig) {
1154 		err = bpf_prog_store_orig_filter(fp, fprog);
1155 		if (err) {
1156 			__bpf_prog_free(fp);
1157 			return -ENOMEM;
1158 		}
1159 	}
1160 
1161 	/* bpf_prepare_filter() already takes care of freeing
1162 	 * memory in case something goes wrong.
1163 	 */
1164 	fp = bpf_prepare_filter(fp, trans);
1165 	if (IS_ERR(fp))
1166 		return PTR_ERR(fp);
1167 
1168 	*pfp = fp;
1169 	return 0;
1170 }
1171 EXPORT_SYMBOL_GPL(bpf_prog_create_from_user);
1172 
bpf_prog_destroy(struct bpf_prog * fp)1173 void bpf_prog_destroy(struct bpf_prog *fp)
1174 {
1175 	__bpf_prog_release(fp);
1176 }
1177 EXPORT_SYMBOL_GPL(bpf_prog_destroy);
1178 
__sk_attach_prog(struct bpf_prog * prog,struct sock * sk)1179 static int __sk_attach_prog(struct bpf_prog *prog, struct sock *sk)
1180 {
1181 	struct sk_filter *fp, *old_fp;
1182 
1183 	fp = kmalloc(sizeof(*fp), GFP_KERNEL);
1184 	if (!fp)
1185 		return -ENOMEM;
1186 
1187 	fp->prog = prog;
1188 	atomic_set(&fp->refcnt, 0);
1189 
1190 	if (!sk_filter_charge(sk, fp)) {
1191 		kfree(fp);
1192 		return -ENOMEM;
1193 	}
1194 
1195 	old_fp = rcu_dereference_protected(sk->sk_filter,
1196 					   lockdep_sock_is_held(sk));
1197 	rcu_assign_pointer(sk->sk_filter, fp);
1198 
1199 	if (old_fp)
1200 		sk_filter_uncharge(sk, old_fp);
1201 
1202 	return 0;
1203 }
1204 
__reuseport_attach_prog(struct bpf_prog * prog,struct sock * sk)1205 static int __reuseport_attach_prog(struct bpf_prog *prog, struct sock *sk)
1206 {
1207 	struct bpf_prog *old_prog;
1208 	int err;
1209 
1210 	if (bpf_prog_size(prog->len) > sysctl_optmem_max)
1211 		return -ENOMEM;
1212 
1213 	if (sk_unhashed(sk) && sk->sk_reuseport) {
1214 		err = reuseport_alloc(sk);
1215 		if (err)
1216 			return err;
1217 	} else if (!rcu_access_pointer(sk->sk_reuseport_cb)) {
1218 		/* The socket wasn't bound with SO_REUSEPORT */
1219 		return -EINVAL;
1220 	}
1221 
1222 	old_prog = reuseport_attach_prog(sk, prog);
1223 	if (old_prog)
1224 		bpf_prog_destroy(old_prog);
1225 
1226 	return 0;
1227 }
1228 
1229 static
__get_filter(struct sock_fprog * fprog,struct sock * sk)1230 struct bpf_prog *__get_filter(struct sock_fprog *fprog, struct sock *sk)
1231 {
1232 	unsigned int fsize = bpf_classic_proglen(fprog);
1233 	struct bpf_prog *prog;
1234 	int err;
1235 
1236 	if (sock_flag(sk, SOCK_FILTER_LOCKED))
1237 		return ERR_PTR(-EPERM);
1238 
1239 	/* Make sure new filter is there and in the right amounts. */
1240 	if (!bpf_check_basics_ok(fprog->filter, fprog->len))
1241 		return ERR_PTR(-EINVAL);
1242 
1243 	prog = bpf_prog_alloc(bpf_prog_size(fprog->len), 0);
1244 	if (!prog)
1245 		return ERR_PTR(-ENOMEM);
1246 
1247 	if (copy_from_user(prog->insns, fprog->filter, fsize)) {
1248 		__bpf_prog_free(prog);
1249 		return ERR_PTR(-EFAULT);
1250 	}
1251 
1252 	prog->len = fprog->len;
1253 
1254 	err = bpf_prog_store_orig_filter(prog, fprog);
1255 	if (err) {
1256 		__bpf_prog_free(prog);
1257 		return ERR_PTR(-ENOMEM);
1258 	}
1259 
1260 	/* bpf_prepare_filter() already takes care of freeing
1261 	 * memory in case something goes wrong.
1262 	 */
1263 	return bpf_prepare_filter(prog, NULL);
1264 }
1265 
1266 /**
1267  *	sk_attach_filter - attach a socket filter
1268  *	@fprog: the filter program
1269  *	@sk: the socket to use
1270  *
1271  * Attach the user's filter code. We first run some sanity checks on
1272  * it to make sure it does not explode on us later. If an error
1273  * occurs or there is insufficient memory for the filter a negative
1274  * errno code is returned. On success the return is zero.
1275  */
sk_attach_filter(struct sock_fprog * fprog,struct sock * sk)1276 int sk_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1277 {
1278 	struct bpf_prog *prog = __get_filter(fprog, sk);
1279 	int err;
1280 
1281 	if (IS_ERR(prog))
1282 		return PTR_ERR(prog);
1283 
1284 	err = __sk_attach_prog(prog, sk);
1285 	if (err < 0) {
1286 		__bpf_prog_release(prog);
1287 		return err;
1288 	}
1289 
1290 	return 0;
1291 }
1292 EXPORT_SYMBOL_GPL(sk_attach_filter);
1293 
sk_reuseport_attach_filter(struct sock_fprog * fprog,struct sock * sk)1294 int sk_reuseport_attach_filter(struct sock_fprog *fprog, struct sock *sk)
1295 {
1296 	struct bpf_prog *prog = __get_filter(fprog, sk);
1297 	int err;
1298 
1299 	if (IS_ERR(prog))
1300 		return PTR_ERR(prog);
1301 
1302 	err = __reuseport_attach_prog(prog, sk);
1303 	if (err < 0) {
1304 		__bpf_prog_release(prog);
1305 		return err;
1306 	}
1307 
1308 	return 0;
1309 }
1310 
__get_bpf(u32 ufd,struct sock * sk)1311 static struct bpf_prog *__get_bpf(u32 ufd, struct sock *sk)
1312 {
1313 	if (sock_flag(sk, SOCK_FILTER_LOCKED))
1314 		return ERR_PTR(-EPERM);
1315 
1316 	return bpf_prog_get_type(ufd, BPF_PROG_TYPE_SOCKET_FILTER);
1317 }
1318 
sk_attach_bpf(u32 ufd,struct sock * sk)1319 int sk_attach_bpf(u32 ufd, struct sock *sk)
1320 {
1321 	struct bpf_prog *prog = __get_bpf(ufd, sk);
1322 	int err;
1323 
1324 	if (IS_ERR(prog))
1325 		return PTR_ERR(prog);
1326 
1327 	err = __sk_attach_prog(prog, sk);
1328 	if (err < 0) {
1329 		bpf_prog_put(prog);
1330 		return err;
1331 	}
1332 
1333 	return 0;
1334 }
1335 
sk_reuseport_attach_bpf(u32 ufd,struct sock * sk)1336 int sk_reuseport_attach_bpf(u32 ufd, struct sock *sk)
1337 {
1338 	struct bpf_prog *prog = __get_bpf(ufd, sk);
1339 	int err;
1340 
1341 	if (IS_ERR(prog))
1342 		return PTR_ERR(prog);
1343 
1344 	err = __reuseport_attach_prog(prog, sk);
1345 	if (err < 0) {
1346 		bpf_prog_put(prog);
1347 		return err;
1348 	}
1349 
1350 	return 0;
1351 }
1352 
1353 struct bpf_scratchpad {
1354 	union {
1355 		__be32 diff[MAX_BPF_STACK / sizeof(__be32)];
1356 		u8     buff[MAX_BPF_STACK];
1357 	};
1358 };
1359 
1360 static DEFINE_PER_CPU(struct bpf_scratchpad, bpf_sp);
1361 
__bpf_try_make_writable(struct sk_buff * skb,unsigned int write_len)1362 static inline int __bpf_try_make_writable(struct sk_buff *skb,
1363 					  unsigned int write_len)
1364 {
1365 	return skb_ensure_writable(skb, write_len);
1366 }
1367 
bpf_try_make_writable(struct sk_buff * skb,unsigned int write_len)1368 static inline int bpf_try_make_writable(struct sk_buff *skb,
1369 					unsigned int write_len)
1370 {
1371 	int err = __bpf_try_make_writable(skb, write_len);
1372 
1373 	bpf_compute_data_end(skb);
1374 	return err;
1375 }
1376 
bpf_try_make_head_writable(struct sk_buff * skb)1377 static int bpf_try_make_head_writable(struct sk_buff *skb)
1378 {
1379 	return bpf_try_make_writable(skb, skb_headlen(skb));
1380 }
1381 
bpf_push_mac_rcsum(struct sk_buff * skb)1382 static inline void bpf_push_mac_rcsum(struct sk_buff *skb)
1383 {
1384 	if (skb_at_tc_ingress(skb))
1385 		skb_postpush_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1386 }
1387 
bpf_pull_mac_rcsum(struct sk_buff * skb)1388 static inline void bpf_pull_mac_rcsum(struct sk_buff *skb)
1389 {
1390 	if (skb_at_tc_ingress(skb))
1391 		skb_postpull_rcsum(skb, skb_mac_header(skb), skb->mac_len);
1392 }
1393 
BPF_CALL_5(bpf_skb_store_bytes,struct sk_buff *,skb,u32,offset,const void *,from,u32,len,u64,flags)1394 BPF_CALL_5(bpf_skb_store_bytes, struct sk_buff *, skb, u32, offset,
1395 	   const void *, from, u32, len, u64, flags)
1396 {
1397 	void *ptr;
1398 
1399 	if (unlikely(flags & ~(BPF_F_RECOMPUTE_CSUM | BPF_F_INVALIDATE_HASH)))
1400 		return -EINVAL;
1401 	if (unlikely(offset > 0xffff))
1402 		return -EFAULT;
1403 	if (unlikely(bpf_try_make_writable(skb, offset + len)))
1404 		return -EFAULT;
1405 
1406 	ptr = skb->data + offset;
1407 	if (flags & BPF_F_RECOMPUTE_CSUM)
1408 		__skb_postpull_rcsum(skb, ptr, len, offset);
1409 
1410 	memcpy(ptr, from, len);
1411 
1412 	if (flags & BPF_F_RECOMPUTE_CSUM)
1413 		__skb_postpush_rcsum(skb, ptr, len, offset);
1414 	if (flags & BPF_F_INVALIDATE_HASH)
1415 		skb_clear_hash(skb);
1416 
1417 	return 0;
1418 }
1419 
1420 static const struct bpf_func_proto bpf_skb_store_bytes_proto = {
1421 	.func		= bpf_skb_store_bytes,
1422 	.gpl_only	= false,
1423 	.ret_type	= RET_INTEGER,
1424 	.arg1_type	= ARG_PTR_TO_CTX,
1425 	.arg2_type	= ARG_ANYTHING,
1426 	.arg3_type	= ARG_PTR_TO_STACK,
1427 	.arg4_type	= ARG_CONST_STACK_SIZE,
1428 	.arg5_type	= ARG_ANYTHING,
1429 };
1430 
BPF_CALL_4(bpf_skb_load_bytes,const struct sk_buff *,skb,u32,offset,void *,to,u32,len)1431 BPF_CALL_4(bpf_skb_load_bytes, const struct sk_buff *, skb, u32, offset,
1432 	   void *, to, u32, len)
1433 {
1434 	void *ptr;
1435 
1436 	if (unlikely(offset > 0xffff))
1437 		goto err_clear;
1438 
1439 	ptr = skb_header_pointer(skb, offset, len, to);
1440 	if (unlikely(!ptr))
1441 		goto err_clear;
1442 	if (ptr != to)
1443 		memcpy(to, ptr, len);
1444 
1445 	return 0;
1446 err_clear:
1447 	memset(to, 0, len);
1448 	return -EFAULT;
1449 }
1450 
1451 static const struct bpf_func_proto bpf_skb_load_bytes_proto = {
1452 	.func		= bpf_skb_load_bytes,
1453 	.gpl_only	= false,
1454 	.ret_type	= RET_INTEGER,
1455 	.arg1_type	= ARG_PTR_TO_CTX,
1456 	.arg2_type	= ARG_ANYTHING,
1457 	.arg3_type	= ARG_PTR_TO_RAW_STACK,
1458 	.arg4_type	= ARG_CONST_STACK_SIZE,
1459 };
1460 
BPF_CALL_2(bpf_skb_pull_data,struct sk_buff *,skb,u32,len)1461 BPF_CALL_2(bpf_skb_pull_data, struct sk_buff *, skb, u32, len)
1462 {
1463 	/* Idea is the following: should the needed direct read/write
1464 	 * test fail during runtime, we can pull in more data and redo
1465 	 * again, since implicitly, we invalidate previous checks here.
1466 	 *
1467 	 * Or, since we know how much we need to make read/writeable,
1468 	 * this can be done once at the program beginning for direct
1469 	 * access case. By this we overcome limitations of only current
1470 	 * headroom being accessible.
1471 	 */
1472 	return bpf_try_make_writable(skb, len ? : skb_headlen(skb));
1473 }
1474 
1475 static const struct bpf_func_proto bpf_skb_pull_data_proto = {
1476 	.func		= bpf_skb_pull_data,
1477 	.gpl_only	= false,
1478 	.ret_type	= RET_INTEGER,
1479 	.arg1_type	= ARG_PTR_TO_CTX,
1480 	.arg2_type	= ARG_ANYTHING,
1481 };
1482 
BPF_CALL_5(bpf_l3_csum_replace,struct sk_buff *,skb,u32,offset,u64,from,u64,to,u64,flags)1483 BPF_CALL_5(bpf_l3_csum_replace, struct sk_buff *, skb, u32, offset,
1484 	   u64, from, u64, to, u64, flags)
1485 {
1486 	__sum16 *ptr;
1487 
1488 	if (unlikely(flags & ~(BPF_F_HDR_FIELD_MASK)))
1489 		return -EINVAL;
1490 	if (unlikely(offset > 0xffff || offset & 1))
1491 		return -EFAULT;
1492 	if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1493 		return -EFAULT;
1494 
1495 	ptr = (__sum16 *)(skb->data + offset);
1496 	switch (flags & BPF_F_HDR_FIELD_MASK) {
1497 	case 0:
1498 		if (unlikely(from != 0))
1499 			return -EINVAL;
1500 
1501 		csum_replace_by_diff(ptr, to);
1502 		break;
1503 	case 2:
1504 		csum_replace2(ptr, from, to);
1505 		break;
1506 	case 4:
1507 		csum_replace4(ptr, from, to);
1508 		break;
1509 	default:
1510 		return -EINVAL;
1511 	}
1512 
1513 	return 0;
1514 }
1515 
1516 static const struct bpf_func_proto bpf_l3_csum_replace_proto = {
1517 	.func		= bpf_l3_csum_replace,
1518 	.gpl_only	= false,
1519 	.ret_type	= RET_INTEGER,
1520 	.arg1_type	= ARG_PTR_TO_CTX,
1521 	.arg2_type	= ARG_ANYTHING,
1522 	.arg3_type	= ARG_ANYTHING,
1523 	.arg4_type	= ARG_ANYTHING,
1524 	.arg5_type	= ARG_ANYTHING,
1525 };
1526 
BPF_CALL_5(bpf_l4_csum_replace,struct sk_buff *,skb,u32,offset,u64,from,u64,to,u64,flags)1527 BPF_CALL_5(bpf_l4_csum_replace, struct sk_buff *, skb, u32, offset,
1528 	   u64, from, u64, to, u64, flags)
1529 {
1530 	bool is_pseudo = flags & BPF_F_PSEUDO_HDR;
1531 	bool is_mmzero = flags & BPF_F_MARK_MANGLED_0;
1532 	__sum16 *ptr;
1533 
1534 	if (unlikely(flags & ~(BPF_F_MARK_MANGLED_0 | BPF_F_PSEUDO_HDR |
1535 			       BPF_F_HDR_FIELD_MASK)))
1536 		return -EINVAL;
1537 	if (unlikely(offset > 0xffff || offset & 1))
1538 		return -EFAULT;
1539 	if (unlikely(bpf_try_make_writable(skb, offset + sizeof(*ptr))))
1540 		return -EFAULT;
1541 
1542 	ptr = (__sum16 *)(skb->data + offset);
1543 	if (is_mmzero && !*ptr)
1544 		return 0;
1545 
1546 	switch (flags & BPF_F_HDR_FIELD_MASK) {
1547 	case 0:
1548 		if (unlikely(from != 0))
1549 			return -EINVAL;
1550 
1551 		inet_proto_csum_replace_by_diff(ptr, skb, to, is_pseudo);
1552 		break;
1553 	case 2:
1554 		inet_proto_csum_replace2(ptr, skb, from, to, is_pseudo);
1555 		break;
1556 	case 4:
1557 		inet_proto_csum_replace4(ptr, skb, from, to, is_pseudo);
1558 		break;
1559 	default:
1560 		return -EINVAL;
1561 	}
1562 
1563 	if (is_mmzero && !*ptr)
1564 		*ptr = CSUM_MANGLED_0;
1565 	return 0;
1566 }
1567 
1568 static const struct bpf_func_proto bpf_l4_csum_replace_proto = {
1569 	.func		= bpf_l4_csum_replace,
1570 	.gpl_only	= false,
1571 	.ret_type	= RET_INTEGER,
1572 	.arg1_type	= ARG_PTR_TO_CTX,
1573 	.arg2_type	= ARG_ANYTHING,
1574 	.arg3_type	= ARG_ANYTHING,
1575 	.arg4_type	= ARG_ANYTHING,
1576 	.arg5_type	= ARG_ANYTHING,
1577 };
1578 
BPF_CALL_5(bpf_csum_diff,__be32 *,from,u32,from_size,__be32 *,to,u32,to_size,__wsum,seed)1579 BPF_CALL_5(bpf_csum_diff, __be32 *, from, u32, from_size,
1580 	   __be32 *, to, u32, to_size, __wsum, seed)
1581 {
1582 	struct bpf_scratchpad *sp = this_cpu_ptr(&bpf_sp);
1583 	u32 diff_size = from_size + to_size;
1584 	int i, j = 0;
1585 
1586 	/* This is quite flexible, some examples:
1587 	 *
1588 	 * from_size == 0, to_size > 0,  seed := csum --> pushing data
1589 	 * from_size > 0,  to_size == 0, seed := csum --> pulling data
1590 	 * from_size > 0,  to_size > 0,  seed := 0    --> diffing data
1591 	 *
1592 	 * Even for diffing, from_size and to_size don't need to be equal.
1593 	 */
1594 	if (unlikely(((from_size | to_size) & (sizeof(__be32) - 1)) ||
1595 		     diff_size > sizeof(sp->diff)))
1596 		return -EINVAL;
1597 
1598 	for (i = 0; i < from_size / sizeof(__be32); i++, j++)
1599 		sp->diff[j] = ~from[i];
1600 	for (i = 0; i <   to_size / sizeof(__be32); i++, j++)
1601 		sp->diff[j] = to[i];
1602 
1603 	return csum_partial(sp->diff, diff_size, seed);
1604 }
1605 
1606 static const struct bpf_func_proto bpf_csum_diff_proto = {
1607 	.func		= bpf_csum_diff,
1608 	.gpl_only	= false,
1609 	.pkt_access	= true,
1610 	.ret_type	= RET_INTEGER,
1611 	.arg1_type	= ARG_PTR_TO_STACK,
1612 	.arg2_type	= ARG_CONST_STACK_SIZE_OR_ZERO,
1613 	.arg3_type	= ARG_PTR_TO_STACK,
1614 	.arg4_type	= ARG_CONST_STACK_SIZE_OR_ZERO,
1615 	.arg5_type	= ARG_ANYTHING,
1616 };
1617 
BPF_CALL_2(bpf_csum_update,struct sk_buff *,skb,__wsum,csum)1618 BPF_CALL_2(bpf_csum_update, struct sk_buff *, skb, __wsum, csum)
1619 {
1620 	/* The interface is to be used in combination with bpf_csum_diff()
1621 	 * for direct packet writes. csum rotation for alignment as well
1622 	 * as emulating csum_sub() can be done from the eBPF program.
1623 	 */
1624 	if (skb->ip_summed == CHECKSUM_COMPLETE)
1625 		return (skb->csum = csum_add(skb->csum, csum));
1626 
1627 	return -ENOTSUPP;
1628 }
1629 
1630 static const struct bpf_func_proto bpf_csum_update_proto = {
1631 	.func		= bpf_csum_update,
1632 	.gpl_only	= false,
1633 	.ret_type	= RET_INTEGER,
1634 	.arg1_type	= ARG_PTR_TO_CTX,
1635 	.arg2_type	= ARG_ANYTHING,
1636 };
1637 
__bpf_rx_skb(struct net_device * dev,struct sk_buff * skb)1638 static inline int __bpf_rx_skb(struct net_device *dev, struct sk_buff *skb)
1639 {
1640 	return dev_forward_skb(dev, skb);
1641 }
1642 
__bpf_rx_skb_no_mac(struct net_device * dev,struct sk_buff * skb)1643 static inline int __bpf_rx_skb_no_mac(struct net_device *dev,
1644 				      struct sk_buff *skb)
1645 {
1646 	int ret = ____dev_forward_skb(dev, skb);
1647 
1648 	if (likely(!ret)) {
1649 		skb->dev = dev;
1650 		ret = netif_rx(skb);
1651 	}
1652 
1653 	return ret;
1654 }
1655 
__bpf_tx_skb(struct net_device * dev,struct sk_buff * skb)1656 static inline int __bpf_tx_skb(struct net_device *dev, struct sk_buff *skb)
1657 {
1658 	int ret;
1659 
1660 	if (unlikely(__this_cpu_read(xmit_recursion) > XMIT_RECURSION_LIMIT)) {
1661 		net_crit_ratelimited("bpf: recursion limit reached on datapath, buggy bpf program?\n");
1662 		kfree_skb(skb);
1663 		return -ENETDOWN;
1664 	}
1665 
1666 	skb->dev = dev;
1667 
1668 	__this_cpu_inc(xmit_recursion);
1669 	ret = dev_queue_xmit(skb);
1670 	__this_cpu_dec(xmit_recursion);
1671 
1672 	return ret;
1673 }
1674 
__bpf_redirect_no_mac(struct sk_buff * skb,struct net_device * dev,u32 flags)1675 static int __bpf_redirect_no_mac(struct sk_buff *skb, struct net_device *dev,
1676 				 u32 flags)
1677 {
1678 	/* skb->mac_len is not set on normal egress */
1679 	unsigned int mlen = skb->network_header - skb->mac_header;
1680 
1681 	__skb_pull(skb, mlen);
1682 
1683 	/* At ingress, the mac header has already been pulled once.
1684 	 * At egress, skb_pospull_rcsum has to be done in case that
1685 	 * the skb is originated from ingress (i.e. a forwarded skb)
1686 	 * to ensure that rcsum starts at net header.
1687 	 */
1688 	if (!skb_at_tc_ingress(skb))
1689 		skb_postpull_rcsum(skb, skb_mac_header(skb), mlen);
1690 	skb_pop_mac_header(skb);
1691 	skb_reset_mac_len(skb);
1692 	return flags & BPF_F_INGRESS ?
1693 	       __bpf_rx_skb_no_mac(dev, skb) : __bpf_tx_skb(dev, skb);
1694 }
1695 
__bpf_redirect_common(struct sk_buff * skb,struct net_device * dev,u32 flags)1696 static int __bpf_redirect_common(struct sk_buff *skb, struct net_device *dev,
1697 				 u32 flags)
1698 {
1699 	bpf_push_mac_rcsum(skb);
1700 	return flags & BPF_F_INGRESS ?
1701 	       __bpf_rx_skb(dev, skb) : __bpf_tx_skb(dev, skb);
1702 }
1703 
__bpf_redirect(struct sk_buff * skb,struct net_device * dev,u32 flags)1704 static int __bpf_redirect(struct sk_buff *skb, struct net_device *dev,
1705 			  u32 flags)
1706 {
1707 	switch (dev->type) {
1708 	case ARPHRD_TUNNEL:
1709 	case ARPHRD_TUNNEL6:
1710 	case ARPHRD_SIT:
1711 	case ARPHRD_IPGRE:
1712 	case ARPHRD_VOID:
1713 	case ARPHRD_NONE:
1714 		return __bpf_redirect_no_mac(skb, dev, flags);
1715 	default:
1716 		return __bpf_redirect_common(skb, dev, flags);
1717 	}
1718 }
1719 
BPF_CALL_3(bpf_clone_redirect,struct sk_buff *,skb,u32,ifindex,u64,flags)1720 BPF_CALL_3(bpf_clone_redirect, struct sk_buff *, skb, u32, ifindex, u64, flags)
1721 {
1722 	struct net_device *dev;
1723 	struct sk_buff *clone;
1724 	int ret;
1725 
1726 	if (unlikely(flags & ~(BPF_F_INGRESS)))
1727 		return -EINVAL;
1728 
1729 	dev = dev_get_by_index_rcu(dev_net(skb->dev), ifindex);
1730 	if (unlikely(!dev))
1731 		return -EINVAL;
1732 
1733 	clone = skb_clone(skb, GFP_ATOMIC);
1734 	if (unlikely(!clone))
1735 		return -ENOMEM;
1736 
1737 	/* For direct write, we need to keep the invariant that the skbs
1738 	 * we're dealing with need to be uncloned. Should uncloning fail
1739 	 * here, we need to free the just generated clone to unclone once
1740 	 * again.
1741 	 */
1742 	ret = bpf_try_make_head_writable(skb);
1743 	if (unlikely(ret)) {
1744 		kfree_skb(clone);
1745 		return -ENOMEM;
1746 	}
1747 
1748 	return __bpf_redirect(clone, dev, flags);
1749 }
1750 
1751 static const struct bpf_func_proto bpf_clone_redirect_proto = {
1752 	.func           = bpf_clone_redirect,
1753 	.gpl_only       = false,
1754 	.ret_type       = RET_INTEGER,
1755 	.arg1_type      = ARG_PTR_TO_CTX,
1756 	.arg2_type      = ARG_ANYTHING,
1757 	.arg3_type      = ARG_ANYTHING,
1758 };
1759 
1760 struct redirect_info {
1761 	u32 ifindex;
1762 	u32 flags;
1763 };
1764 
1765 static DEFINE_PER_CPU(struct redirect_info, redirect_info);
1766 
BPF_CALL_2(bpf_redirect,u32,ifindex,u64,flags)1767 BPF_CALL_2(bpf_redirect, u32, ifindex, u64, flags)
1768 {
1769 	struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1770 
1771 	if (unlikely(flags & ~(BPF_F_INGRESS)))
1772 		return TC_ACT_SHOT;
1773 
1774 	ri->ifindex = ifindex;
1775 	ri->flags = flags;
1776 
1777 	return TC_ACT_REDIRECT;
1778 }
1779 
skb_do_redirect(struct sk_buff * skb)1780 int skb_do_redirect(struct sk_buff *skb)
1781 {
1782 	struct redirect_info *ri = this_cpu_ptr(&redirect_info);
1783 	struct net_device *dev;
1784 
1785 	dev = dev_get_by_index_rcu(dev_net(skb->dev), ri->ifindex);
1786 	ri->ifindex = 0;
1787 	if (unlikely(!dev)) {
1788 		kfree_skb(skb);
1789 		return -EINVAL;
1790 	}
1791 
1792 	return __bpf_redirect(skb, dev, ri->flags);
1793 }
1794 
1795 static const struct bpf_func_proto bpf_redirect_proto = {
1796 	.func           = bpf_redirect,
1797 	.gpl_only       = false,
1798 	.ret_type       = RET_INTEGER,
1799 	.arg1_type      = ARG_ANYTHING,
1800 	.arg2_type      = ARG_ANYTHING,
1801 };
1802 
BPF_CALL_1(bpf_get_cgroup_classid,const struct sk_buff *,skb)1803 BPF_CALL_1(bpf_get_cgroup_classid, const struct sk_buff *, skb)
1804 {
1805 	return task_get_classid(skb);
1806 }
1807 
1808 static const struct bpf_func_proto bpf_get_cgroup_classid_proto = {
1809 	.func           = bpf_get_cgroup_classid,
1810 	.gpl_only       = false,
1811 	.ret_type       = RET_INTEGER,
1812 	.arg1_type      = ARG_PTR_TO_CTX,
1813 };
1814 
BPF_CALL_1(bpf_get_route_realm,const struct sk_buff *,skb)1815 BPF_CALL_1(bpf_get_route_realm, const struct sk_buff *, skb)
1816 {
1817 	return dst_tclassid(skb);
1818 }
1819 
1820 static const struct bpf_func_proto bpf_get_route_realm_proto = {
1821 	.func           = bpf_get_route_realm,
1822 	.gpl_only       = false,
1823 	.ret_type       = RET_INTEGER,
1824 	.arg1_type      = ARG_PTR_TO_CTX,
1825 };
1826 
BPF_CALL_1(bpf_get_hash_recalc,struct sk_buff *,skb)1827 BPF_CALL_1(bpf_get_hash_recalc, struct sk_buff *, skb)
1828 {
1829 	/* If skb_clear_hash() was called due to mangling, we can
1830 	 * trigger SW recalculation here. Later access to hash
1831 	 * can then use the inline skb->hash via context directly
1832 	 * instead of calling this helper again.
1833 	 */
1834 	return skb_get_hash(skb);
1835 }
1836 
1837 static const struct bpf_func_proto bpf_get_hash_recalc_proto = {
1838 	.func		= bpf_get_hash_recalc,
1839 	.gpl_only	= false,
1840 	.ret_type	= RET_INTEGER,
1841 	.arg1_type	= ARG_PTR_TO_CTX,
1842 };
1843 
BPF_CALL_1(bpf_set_hash_invalid,struct sk_buff *,skb)1844 BPF_CALL_1(bpf_set_hash_invalid, struct sk_buff *, skb)
1845 {
1846 	/* After all direct packet write, this can be used once for
1847 	 * triggering a lazy recalc on next skb_get_hash() invocation.
1848 	 */
1849 	skb_clear_hash(skb);
1850 	return 0;
1851 }
1852 
1853 static const struct bpf_func_proto bpf_set_hash_invalid_proto = {
1854 	.func		= bpf_set_hash_invalid,
1855 	.gpl_only	= false,
1856 	.ret_type	= RET_INTEGER,
1857 	.arg1_type	= ARG_PTR_TO_CTX,
1858 };
1859 
BPF_CALL_3(bpf_skb_vlan_push,struct sk_buff *,skb,__be16,vlan_proto,u16,vlan_tci)1860 BPF_CALL_3(bpf_skb_vlan_push, struct sk_buff *, skb, __be16, vlan_proto,
1861 	   u16, vlan_tci)
1862 {
1863 	int ret;
1864 
1865 	if (unlikely(vlan_proto != htons(ETH_P_8021Q) &&
1866 		     vlan_proto != htons(ETH_P_8021AD)))
1867 		vlan_proto = htons(ETH_P_8021Q);
1868 
1869 	bpf_push_mac_rcsum(skb);
1870 	ret = skb_vlan_push(skb, vlan_proto, vlan_tci);
1871 	bpf_pull_mac_rcsum(skb);
1872 
1873 	bpf_compute_data_end(skb);
1874 	return ret;
1875 }
1876 
1877 const struct bpf_func_proto bpf_skb_vlan_push_proto = {
1878 	.func           = bpf_skb_vlan_push,
1879 	.gpl_only       = false,
1880 	.ret_type       = RET_INTEGER,
1881 	.arg1_type      = ARG_PTR_TO_CTX,
1882 	.arg2_type      = ARG_ANYTHING,
1883 	.arg3_type      = ARG_ANYTHING,
1884 };
1885 EXPORT_SYMBOL_GPL(bpf_skb_vlan_push_proto);
1886 
BPF_CALL_1(bpf_skb_vlan_pop,struct sk_buff *,skb)1887 BPF_CALL_1(bpf_skb_vlan_pop, struct sk_buff *, skb)
1888 {
1889 	int ret;
1890 
1891 	bpf_push_mac_rcsum(skb);
1892 	ret = skb_vlan_pop(skb);
1893 	bpf_pull_mac_rcsum(skb);
1894 
1895 	bpf_compute_data_end(skb);
1896 	return ret;
1897 }
1898 
1899 const struct bpf_func_proto bpf_skb_vlan_pop_proto = {
1900 	.func           = bpf_skb_vlan_pop,
1901 	.gpl_only       = false,
1902 	.ret_type       = RET_INTEGER,
1903 	.arg1_type      = ARG_PTR_TO_CTX,
1904 };
1905 EXPORT_SYMBOL_GPL(bpf_skb_vlan_pop_proto);
1906 
bpf_skb_generic_push(struct sk_buff * skb,u32 off,u32 len)1907 static int bpf_skb_generic_push(struct sk_buff *skb, u32 off, u32 len)
1908 {
1909 	/* Caller already did skb_cow() with len as headroom,
1910 	 * so no need to do it here.
1911 	 */
1912 	skb_push(skb, len);
1913 	memmove(skb->data, skb->data + len, off);
1914 	memset(skb->data + off, 0, len);
1915 
1916 	/* No skb_postpush_rcsum(skb, skb->data + off, len)
1917 	 * needed here as it does not change the skb->csum
1918 	 * result for checksum complete when summing over
1919 	 * zeroed blocks.
1920 	 */
1921 	return 0;
1922 }
1923 
bpf_skb_generic_pop(struct sk_buff * skb,u32 off,u32 len)1924 static int bpf_skb_generic_pop(struct sk_buff *skb, u32 off, u32 len)
1925 {
1926 	/* skb_ensure_writable() is not needed here, as we're
1927 	 * already working on an uncloned skb.
1928 	 */
1929 	if (unlikely(!pskb_may_pull(skb, off + len)))
1930 		return -ENOMEM;
1931 
1932 	skb_postpull_rcsum(skb, skb->data + off, len);
1933 	memmove(skb->data + len, skb->data, off);
1934 	__skb_pull(skb, len);
1935 
1936 	return 0;
1937 }
1938 
bpf_skb_net_hdr_push(struct sk_buff * skb,u32 off,u32 len)1939 static int bpf_skb_net_hdr_push(struct sk_buff *skb, u32 off, u32 len)
1940 {
1941 	bool trans_same = skb->transport_header == skb->network_header;
1942 	int ret;
1943 
1944 	/* There's no need for __skb_push()/__skb_pull() pair to
1945 	 * get to the start of the mac header as we're guaranteed
1946 	 * to always start from here under eBPF.
1947 	 */
1948 	ret = bpf_skb_generic_push(skb, off, len);
1949 	if (likely(!ret)) {
1950 		skb->mac_header -= len;
1951 		skb->network_header -= len;
1952 		if (trans_same)
1953 			skb->transport_header = skb->network_header;
1954 	}
1955 
1956 	return ret;
1957 }
1958 
bpf_skb_net_hdr_pop(struct sk_buff * skb,u32 off,u32 len)1959 static int bpf_skb_net_hdr_pop(struct sk_buff *skb, u32 off, u32 len)
1960 {
1961 	bool trans_same = skb->transport_header == skb->network_header;
1962 	int ret;
1963 
1964 	/* Same here, __skb_push()/__skb_pull() pair not needed. */
1965 	ret = bpf_skb_generic_pop(skb, off, len);
1966 	if (likely(!ret)) {
1967 		skb->mac_header += len;
1968 		skb->network_header += len;
1969 		if (trans_same)
1970 			skb->transport_header = skb->network_header;
1971 	}
1972 
1973 	return ret;
1974 }
1975 
bpf_skb_proto_4_to_6(struct sk_buff * skb)1976 static int bpf_skb_proto_4_to_6(struct sk_buff *skb)
1977 {
1978 	const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
1979 	u32 off = skb->network_header - skb->mac_header;
1980 	int ret;
1981 
1982 	ret = skb_cow(skb, len_diff);
1983 	if (unlikely(ret < 0))
1984 		return ret;
1985 
1986 	ret = bpf_skb_net_hdr_push(skb, off, len_diff);
1987 	if (unlikely(ret < 0))
1988 		return ret;
1989 
1990 	if (skb_is_gso(skb)) {
1991 		/* SKB_GSO_UDP stays as is. SKB_GSO_TCPV4 needs to
1992 		 * be changed into SKB_GSO_TCPV6.
1993 		 */
1994 		if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
1995 			skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV4;
1996 			skb_shinfo(skb)->gso_type |=  SKB_GSO_TCPV6;
1997 		}
1998 
1999 		/* Due to IPv6 header, MSS needs to be downgraded. */
2000 		skb_shinfo(skb)->gso_size -= len_diff;
2001 		/* Header must be checked, and gso_segs recomputed. */
2002 		skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2003 		skb_shinfo(skb)->gso_segs = 0;
2004 	}
2005 
2006 	skb->protocol = htons(ETH_P_IPV6);
2007 	skb_clear_hash(skb);
2008 
2009 	return 0;
2010 }
2011 
bpf_skb_proto_6_to_4(struct sk_buff * skb)2012 static int bpf_skb_proto_6_to_4(struct sk_buff *skb)
2013 {
2014 	const u32 len_diff = sizeof(struct ipv6hdr) - sizeof(struct iphdr);
2015 	u32 off = skb->network_header - skb->mac_header;
2016 	int ret;
2017 
2018 	ret = skb_unclone(skb, GFP_ATOMIC);
2019 	if (unlikely(ret < 0))
2020 		return ret;
2021 
2022 	ret = bpf_skb_net_hdr_pop(skb, off, len_diff);
2023 	if (unlikely(ret < 0))
2024 		return ret;
2025 
2026 	if (skb_is_gso(skb)) {
2027 		/* SKB_GSO_UDP stays as is. SKB_GSO_TCPV6 needs to
2028 		 * be changed into SKB_GSO_TCPV4.
2029 		 */
2030 		if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6) {
2031 			skb_shinfo(skb)->gso_type &= ~SKB_GSO_TCPV6;
2032 			skb_shinfo(skb)->gso_type |=  SKB_GSO_TCPV4;
2033 		}
2034 
2035 		/* Due to IPv4 header, MSS can be upgraded. */
2036 		skb_shinfo(skb)->gso_size += len_diff;
2037 		/* Header must be checked, and gso_segs recomputed. */
2038 		skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
2039 		skb_shinfo(skb)->gso_segs = 0;
2040 	}
2041 
2042 	skb->protocol = htons(ETH_P_IP);
2043 	skb_clear_hash(skb);
2044 
2045 	return 0;
2046 }
2047 
bpf_skb_proto_xlat(struct sk_buff * skb,__be16 to_proto)2048 static int bpf_skb_proto_xlat(struct sk_buff *skb, __be16 to_proto)
2049 {
2050 	__be16 from_proto = skb->protocol;
2051 
2052 	if (from_proto == htons(ETH_P_IP) &&
2053 	      to_proto == htons(ETH_P_IPV6))
2054 		return bpf_skb_proto_4_to_6(skb);
2055 
2056 	if (from_proto == htons(ETH_P_IPV6) &&
2057 	      to_proto == htons(ETH_P_IP))
2058 		return bpf_skb_proto_6_to_4(skb);
2059 
2060 	return -ENOTSUPP;
2061 }
2062 
BPF_CALL_3(bpf_skb_change_proto,struct sk_buff *,skb,__be16,proto,u64,flags)2063 BPF_CALL_3(bpf_skb_change_proto, struct sk_buff *, skb, __be16, proto,
2064 	   u64, flags)
2065 {
2066 	int ret;
2067 
2068 	if (unlikely(flags))
2069 		return -EINVAL;
2070 
2071 	/* General idea is that this helper does the basic groundwork
2072 	 * needed for changing the protocol, and eBPF program fills the
2073 	 * rest through bpf_skb_store_bytes(), bpf_lX_csum_replace()
2074 	 * and other helpers, rather than passing a raw buffer here.
2075 	 *
2076 	 * The rationale is to keep this minimal and without a need to
2077 	 * deal with raw packet data. F.e. even if we would pass buffers
2078 	 * here, the program still needs to call the bpf_lX_csum_replace()
2079 	 * helpers anyway. Plus, this way we keep also separation of
2080 	 * concerns, since f.e. bpf_skb_store_bytes() should only take
2081 	 * care of stores.
2082 	 *
2083 	 * Currently, additional options and extension header space are
2084 	 * not supported, but flags register is reserved so we can adapt
2085 	 * that. For offloads, we mark packet as dodgy, so that headers
2086 	 * need to be verified first.
2087 	 */
2088 	ret = bpf_skb_proto_xlat(skb, proto);
2089 	bpf_compute_data_end(skb);
2090 	return ret;
2091 }
2092 
2093 static const struct bpf_func_proto bpf_skb_change_proto_proto = {
2094 	.func		= bpf_skb_change_proto,
2095 	.gpl_only	= false,
2096 	.ret_type	= RET_INTEGER,
2097 	.arg1_type	= ARG_PTR_TO_CTX,
2098 	.arg2_type	= ARG_ANYTHING,
2099 	.arg3_type	= ARG_ANYTHING,
2100 };
2101 
BPF_CALL_2(bpf_skb_change_type,struct sk_buff *,skb,u32,pkt_type)2102 BPF_CALL_2(bpf_skb_change_type, struct sk_buff *, skb, u32, pkt_type)
2103 {
2104 	/* We only allow a restricted subset to be changed for now. */
2105 	if (unlikely(!skb_pkt_type_ok(skb->pkt_type) ||
2106 		     !skb_pkt_type_ok(pkt_type)))
2107 		return -EINVAL;
2108 
2109 	skb->pkt_type = pkt_type;
2110 	return 0;
2111 }
2112 
2113 static const struct bpf_func_proto bpf_skb_change_type_proto = {
2114 	.func		= bpf_skb_change_type,
2115 	.gpl_only	= false,
2116 	.ret_type	= RET_INTEGER,
2117 	.arg1_type	= ARG_PTR_TO_CTX,
2118 	.arg2_type	= ARG_ANYTHING,
2119 };
2120 
__bpf_skb_min_len(const struct sk_buff * skb)2121 static u32 __bpf_skb_min_len(const struct sk_buff *skb)
2122 {
2123 	u32 min_len = skb_network_offset(skb);
2124 
2125 	if (skb_transport_header_was_set(skb))
2126 		min_len = skb_transport_offset(skb);
2127 	if (skb->ip_summed == CHECKSUM_PARTIAL)
2128 		min_len = skb_checksum_start_offset(skb) +
2129 			  skb->csum_offset + sizeof(__sum16);
2130 	return min_len;
2131 }
2132 
__bpf_skb_max_len(const struct sk_buff * skb)2133 static u32 __bpf_skb_max_len(const struct sk_buff *skb)
2134 {
2135 	return skb->dev->mtu + skb->dev->hard_header_len;
2136 }
2137 
bpf_skb_grow_rcsum(struct sk_buff * skb,unsigned int new_len)2138 static int bpf_skb_grow_rcsum(struct sk_buff *skb, unsigned int new_len)
2139 {
2140 	unsigned int old_len = skb->len;
2141 	int ret;
2142 
2143 	ret = __skb_grow_rcsum(skb, new_len);
2144 	if (!ret)
2145 		memset(skb->data + old_len, 0, new_len - old_len);
2146 	return ret;
2147 }
2148 
bpf_skb_trim_rcsum(struct sk_buff * skb,unsigned int new_len)2149 static int bpf_skb_trim_rcsum(struct sk_buff *skb, unsigned int new_len)
2150 {
2151 	return __skb_trim_rcsum(skb, new_len);
2152 }
2153 
BPF_CALL_3(bpf_skb_change_tail,struct sk_buff *,skb,u32,new_len,u64,flags)2154 BPF_CALL_3(bpf_skb_change_tail, struct sk_buff *, skb, u32, new_len,
2155 	   u64, flags)
2156 {
2157 	u32 max_len = __bpf_skb_max_len(skb);
2158 	u32 min_len = __bpf_skb_min_len(skb);
2159 	int ret;
2160 
2161 	if (unlikely(flags || new_len > max_len || new_len < min_len))
2162 		return -EINVAL;
2163 	if (skb->encapsulation)
2164 		return -ENOTSUPP;
2165 
2166 	/* The basic idea of this helper is that it's performing the
2167 	 * needed work to either grow or trim an skb, and eBPF program
2168 	 * rewrites the rest via helpers like bpf_skb_store_bytes(),
2169 	 * bpf_lX_csum_replace() and others rather than passing a raw
2170 	 * buffer here. This one is a slow path helper and intended
2171 	 * for replies with control messages.
2172 	 *
2173 	 * Like in bpf_skb_change_proto(), we want to keep this rather
2174 	 * minimal and without protocol specifics so that we are able
2175 	 * to separate concerns as in bpf_skb_store_bytes() should only
2176 	 * be the one responsible for writing buffers.
2177 	 *
2178 	 * It's really expected to be a slow path operation here for
2179 	 * control message replies, so we're implicitly linearizing,
2180 	 * uncloning and drop offloads from the skb by this.
2181 	 */
2182 	ret = __bpf_try_make_writable(skb, skb->len);
2183 	if (!ret) {
2184 		if (new_len > skb->len)
2185 			ret = bpf_skb_grow_rcsum(skb, new_len);
2186 		else if (new_len < skb->len)
2187 			ret = bpf_skb_trim_rcsum(skb, new_len);
2188 		if (!ret && skb_is_gso(skb))
2189 			skb_gso_reset(skb);
2190 	}
2191 
2192 	bpf_compute_data_end(skb);
2193 	return ret;
2194 }
2195 
2196 static const struct bpf_func_proto bpf_skb_change_tail_proto = {
2197 	.func		= bpf_skb_change_tail,
2198 	.gpl_only	= false,
2199 	.ret_type	= RET_INTEGER,
2200 	.arg1_type	= ARG_PTR_TO_CTX,
2201 	.arg2_type	= ARG_ANYTHING,
2202 	.arg3_type	= ARG_ANYTHING,
2203 };
2204 
bpf_helper_changes_skb_data(void * func)2205 bool bpf_helper_changes_skb_data(void *func)
2206 {
2207 	if (func == bpf_skb_vlan_push ||
2208 	    func == bpf_skb_vlan_pop ||
2209 	    func == bpf_skb_store_bytes ||
2210 	    func == bpf_skb_change_proto ||
2211 	    func == bpf_skb_change_tail ||
2212 	    func == bpf_skb_pull_data ||
2213 	    func == bpf_clone_redirect ||
2214 	    func == bpf_l3_csum_replace ||
2215 	    func == bpf_l4_csum_replace)
2216 		return true;
2217 
2218 	return false;
2219 }
2220 
bpf_skb_copy(void * dst_buff,const void * skb,unsigned long off,unsigned long len)2221 static unsigned long bpf_skb_copy(void *dst_buff, const void *skb,
2222 				  unsigned long off, unsigned long len)
2223 {
2224 	void *ptr = skb_header_pointer(skb, off, len, dst_buff);
2225 
2226 	if (unlikely(!ptr))
2227 		return len;
2228 	if (ptr != dst_buff)
2229 		memcpy(dst_buff, ptr, len);
2230 
2231 	return 0;
2232 }
2233 
BPF_CALL_5(bpf_skb_event_output,struct sk_buff *,skb,struct bpf_map *,map,u64,flags,void *,meta,u64,meta_size)2234 BPF_CALL_5(bpf_skb_event_output, struct sk_buff *, skb, struct bpf_map *, map,
2235 	   u64, flags, void *, meta, u64, meta_size)
2236 {
2237 	u64 skb_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
2238 
2239 	if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
2240 		return -EINVAL;
2241 	if (unlikely(skb_size > skb->len))
2242 		return -EFAULT;
2243 
2244 	return bpf_event_output(map, flags, meta, meta_size, skb, skb_size,
2245 				bpf_skb_copy);
2246 }
2247 
2248 static const struct bpf_func_proto bpf_skb_event_output_proto = {
2249 	.func		= bpf_skb_event_output,
2250 	.gpl_only	= true,
2251 	.ret_type	= RET_INTEGER,
2252 	.arg1_type	= ARG_PTR_TO_CTX,
2253 	.arg2_type	= ARG_CONST_MAP_PTR,
2254 	.arg3_type	= ARG_ANYTHING,
2255 	.arg4_type	= ARG_PTR_TO_STACK,
2256 	.arg5_type	= ARG_CONST_STACK_SIZE,
2257 };
2258 
bpf_tunnel_key_af(u64 flags)2259 static unsigned short bpf_tunnel_key_af(u64 flags)
2260 {
2261 	return flags & BPF_F_TUNINFO_IPV6 ? AF_INET6 : AF_INET;
2262 }
2263 
BPF_CALL_4(bpf_skb_get_tunnel_key,struct sk_buff *,skb,struct bpf_tunnel_key *,to,u32,size,u64,flags)2264 BPF_CALL_4(bpf_skb_get_tunnel_key, struct sk_buff *, skb, struct bpf_tunnel_key *, to,
2265 	   u32, size, u64, flags)
2266 {
2267 	const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2268 	u8 compat[sizeof(struct bpf_tunnel_key)];
2269 	void *to_orig = to;
2270 	int err;
2271 
2272 	if (unlikely(!info || (flags & ~(BPF_F_TUNINFO_IPV6)))) {
2273 		err = -EINVAL;
2274 		goto err_clear;
2275 	}
2276 	if (ip_tunnel_info_af(info) != bpf_tunnel_key_af(flags)) {
2277 		err = -EPROTO;
2278 		goto err_clear;
2279 	}
2280 	if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2281 		err = -EINVAL;
2282 		switch (size) {
2283 		case offsetof(struct bpf_tunnel_key, tunnel_label):
2284 		case offsetof(struct bpf_tunnel_key, tunnel_ext):
2285 			goto set_compat;
2286 		case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2287 			/* Fixup deprecated structure layouts here, so we have
2288 			 * a common path later on.
2289 			 */
2290 			if (ip_tunnel_info_af(info) != AF_INET)
2291 				goto err_clear;
2292 set_compat:
2293 			to = (struct bpf_tunnel_key *)compat;
2294 			break;
2295 		default:
2296 			goto err_clear;
2297 		}
2298 	}
2299 
2300 	to->tunnel_id = be64_to_cpu(info->key.tun_id);
2301 	to->tunnel_tos = info->key.tos;
2302 	to->tunnel_ttl = info->key.ttl;
2303 
2304 	if (flags & BPF_F_TUNINFO_IPV6) {
2305 		memcpy(to->remote_ipv6, &info->key.u.ipv6.src,
2306 		       sizeof(to->remote_ipv6));
2307 		to->tunnel_label = be32_to_cpu(info->key.label);
2308 	} else {
2309 		to->remote_ipv4 = be32_to_cpu(info->key.u.ipv4.src);
2310 	}
2311 
2312 	if (unlikely(size != sizeof(struct bpf_tunnel_key)))
2313 		memcpy(to_orig, to, size);
2314 
2315 	return 0;
2316 err_clear:
2317 	memset(to_orig, 0, size);
2318 	return err;
2319 }
2320 
2321 static const struct bpf_func_proto bpf_skb_get_tunnel_key_proto = {
2322 	.func		= bpf_skb_get_tunnel_key,
2323 	.gpl_only	= false,
2324 	.ret_type	= RET_INTEGER,
2325 	.arg1_type	= ARG_PTR_TO_CTX,
2326 	.arg2_type	= ARG_PTR_TO_RAW_STACK,
2327 	.arg3_type	= ARG_CONST_STACK_SIZE,
2328 	.arg4_type	= ARG_ANYTHING,
2329 };
2330 
BPF_CALL_3(bpf_skb_get_tunnel_opt,struct sk_buff *,skb,u8 *,to,u32,size)2331 BPF_CALL_3(bpf_skb_get_tunnel_opt, struct sk_buff *, skb, u8 *, to, u32, size)
2332 {
2333 	const struct ip_tunnel_info *info = skb_tunnel_info(skb);
2334 	int err;
2335 
2336 	if (unlikely(!info ||
2337 		     !(info->key.tun_flags & TUNNEL_OPTIONS_PRESENT))) {
2338 		err = -ENOENT;
2339 		goto err_clear;
2340 	}
2341 	if (unlikely(size < info->options_len)) {
2342 		err = -ENOMEM;
2343 		goto err_clear;
2344 	}
2345 
2346 	ip_tunnel_info_opts_get(to, info);
2347 	if (size > info->options_len)
2348 		memset(to + info->options_len, 0, size - info->options_len);
2349 
2350 	return info->options_len;
2351 err_clear:
2352 	memset(to, 0, size);
2353 	return err;
2354 }
2355 
2356 static const struct bpf_func_proto bpf_skb_get_tunnel_opt_proto = {
2357 	.func		= bpf_skb_get_tunnel_opt,
2358 	.gpl_only	= false,
2359 	.ret_type	= RET_INTEGER,
2360 	.arg1_type	= ARG_PTR_TO_CTX,
2361 	.arg2_type	= ARG_PTR_TO_RAW_STACK,
2362 	.arg3_type	= ARG_CONST_STACK_SIZE,
2363 };
2364 
2365 static struct metadata_dst __percpu *md_dst;
2366 
BPF_CALL_4(bpf_skb_set_tunnel_key,struct sk_buff *,skb,const struct bpf_tunnel_key *,from,u32,size,u64,flags)2367 BPF_CALL_4(bpf_skb_set_tunnel_key, struct sk_buff *, skb,
2368 	   const struct bpf_tunnel_key *, from, u32, size, u64, flags)
2369 {
2370 	struct metadata_dst *md = this_cpu_ptr(md_dst);
2371 	u8 compat[sizeof(struct bpf_tunnel_key)];
2372 	struct ip_tunnel_info *info;
2373 
2374 	if (unlikely(flags & ~(BPF_F_TUNINFO_IPV6 | BPF_F_ZERO_CSUM_TX |
2375 			       BPF_F_DONT_FRAGMENT)))
2376 		return -EINVAL;
2377 	if (unlikely(size != sizeof(struct bpf_tunnel_key))) {
2378 		switch (size) {
2379 		case offsetof(struct bpf_tunnel_key, tunnel_label):
2380 		case offsetof(struct bpf_tunnel_key, tunnel_ext):
2381 		case offsetof(struct bpf_tunnel_key, remote_ipv6[1]):
2382 			/* Fixup deprecated structure layouts here, so we have
2383 			 * a common path later on.
2384 			 */
2385 			memcpy(compat, from, size);
2386 			memset(compat + size, 0, sizeof(compat) - size);
2387 			from = (const struct bpf_tunnel_key *) compat;
2388 			break;
2389 		default:
2390 			return -EINVAL;
2391 		}
2392 	}
2393 	if (unlikely((!(flags & BPF_F_TUNINFO_IPV6) && from->tunnel_label) ||
2394 		     from->tunnel_ext))
2395 		return -EINVAL;
2396 
2397 	skb_dst_drop(skb);
2398 	dst_hold((struct dst_entry *) md);
2399 	skb_dst_set(skb, (struct dst_entry *) md);
2400 
2401 	info = &md->u.tun_info;
2402 	info->mode = IP_TUNNEL_INFO_TX;
2403 
2404 	info->key.tun_flags = TUNNEL_KEY | TUNNEL_CSUM | TUNNEL_NOCACHE;
2405 	if (flags & BPF_F_DONT_FRAGMENT)
2406 		info->key.tun_flags |= TUNNEL_DONT_FRAGMENT;
2407 
2408 	info->key.tun_id = cpu_to_be64(from->tunnel_id);
2409 	info->key.tos = from->tunnel_tos;
2410 	info->key.ttl = from->tunnel_ttl;
2411 
2412 	if (flags & BPF_F_TUNINFO_IPV6) {
2413 		info->mode |= IP_TUNNEL_INFO_IPV6;
2414 		memcpy(&info->key.u.ipv6.dst, from->remote_ipv6,
2415 		       sizeof(from->remote_ipv6));
2416 		info->key.label = cpu_to_be32(from->tunnel_label) &
2417 				  IPV6_FLOWLABEL_MASK;
2418 	} else {
2419 		info->key.u.ipv4.dst = cpu_to_be32(from->remote_ipv4);
2420 		if (flags & BPF_F_ZERO_CSUM_TX)
2421 			info->key.tun_flags &= ~TUNNEL_CSUM;
2422 	}
2423 
2424 	return 0;
2425 }
2426 
2427 static const struct bpf_func_proto bpf_skb_set_tunnel_key_proto = {
2428 	.func		= bpf_skb_set_tunnel_key,
2429 	.gpl_only	= false,
2430 	.ret_type	= RET_INTEGER,
2431 	.arg1_type	= ARG_PTR_TO_CTX,
2432 	.arg2_type	= ARG_PTR_TO_STACK,
2433 	.arg3_type	= ARG_CONST_STACK_SIZE,
2434 	.arg4_type	= ARG_ANYTHING,
2435 };
2436 
BPF_CALL_3(bpf_skb_set_tunnel_opt,struct sk_buff *,skb,const u8 *,from,u32,size)2437 BPF_CALL_3(bpf_skb_set_tunnel_opt, struct sk_buff *, skb,
2438 	   const u8 *, from, u32, size)
2439 {
2440 	struct ip_tunnel_info *info = skb_tunnel_info(skb);
2441 	const struct metadata_dst *md = this_cpu_ptr(md_dst);
2442 
2443 	if (unlikely(info != &md->u.tun_info || (size & (sizeof(u32) - 1))))
2444 		return -EINVAL;
2445 	if (unlikely(size > IP_TUNNEL_OPTS_MAX))
2446 		return -ENOMEM;
2447 
2448 	ip_tunnel_info_opts_set(info, from, size);
2449 
2450 	return 0;
2451 }
2452 
2453 static const struct bpf_func_proto bpf_skb_set_tunnel_opt_proto = {
2454 	.func		= bpf_skb_set_tunnel_opt,
2455 	.gpl_only	= false,
2456 	.ret_type	= RET_INTEGER,
2457 	.arg1_type	= ARG_PTR_TO_CTX,
2458 	.arg2_type	= ARG_PTR_TO_STACK,
2459 	.arg3_type	= ARG_CONST_STACK_SIZE,
2460 };
2461 
2462 static const struct bpf_func_proto *
bpf_get_skb_set_tunnel_proto(enum bpf_func_id which)2463 bpf_get_skb_set_tunnel_proto(enum bpf_func_id which)
2464 {
2465 	if (!md_dst) {
2466 		/* Race is not possible, since it's called from verifier
2467 		 * that is holding verifier mutex.
2468 		 */
2469 		md_dst = metadata_dst_alloc_percpu(IP_TUNNEL_OPTS_MAX,
2470 						   GFP_KERNEL);
2471 		if (!md_dst)
2472 			return NULL;
2473 	}
2474 
2475 	switch (which) {
2476 	case BPF_FUNC_skb_set_tunnel_key:
2477 		return &bpf_skb_set_tunnel_key_proto;
2478 	case BPF_FUNC_skb_set_tunnel_opt:
2479 		return &bpf_skb_set_tunnel_opt_proto;
2480 	default:
2481 		return NULL;
2482 	}
2483 }
2484 
BPF_CALL_3(bpf_skb_under_cgroup,struct sk_buff *,skb,struct bpf_map *,map,u32,idx)2485 BPF_CALL_3(bpf_skb_under_cgroup, struct sk_buff *, skb, struct bpf_map *, map,
2486 	   u32, idx)
2487 {
2488 	struct bpf_array *array = container_of(map, struct bpf_array, map);
2489 	struct cgroup *cgrp;
2490 	struct sock *sk;
2491 
2492 	sk = skb_to_full_sk(skb);
2493 	if (!sk || !sk_fullsock(sk))
2494 		return -ENOENT;
2495 	if (unlikely(idx >= array->map.max_entries))
2496 		return -E2BIG;
2497 
2498 	cgrp = READ_ONCE(array->ptrs[idx]);
2499 	if (unlikely(!cgrp))
2500 		return -EAGAIN;
2501 
2502 	return sk_under_cgroup_hierarchy(sk, cgrp);
2503 }
2504 
2505 static const struct bpf_func_proto bpf_skb_under_cgroup_proto = {
2506 	.func		= bpf_skb_under_cgroup,
2507 	.gpl_only	= false,
2508 	.ret_type	= RET_INTEGER,
2509 	.arg1_type	= ARG_PTR_TO_CTX,
2510 	.arg2_type	= ARG_CONST_MAP_PTR,
2511 	.arg3_type	= ARG_ANYTHING,
2512 };
2513 
bpf_xdp_copy(void * dst_buff,const void * src_buff,unsigned long off,unsigned long len)2514 static unsigned long bpf_xdp_copy(void *dst_buff, const void *src_buff,
2515 				  unsigned long off, unsigned long len)
2516 {
2517 	memcpy(dst_buff, src_buff + off, len);
2518 	return 0;
2519 }
2520 
BPF_CALL_5(bpf_xdp_event_output,struct xdp_buff *,xdp,struct bpf_map *,map,u64,flags,void *,meta,u64,meta_size)2521 BPF_CALL_5(bpf_xdp_event_output, struct xdp_buff *, xdp, struct bpf_map *, map,
2522 	   u64, flags, void *, meta, u64, meta_size)
2523 {
2524 	u64 xdp_size = (flags & BPF_F_CTXLEN_MASK) >> 32;
2525 
2526 	if (unlikely(flags & ~(BPF_F_CTXLEN_MASK | BPF_F_INDEX_MASK)))
2527 		return -EINVAL;
2528 	if (unlikely(xdp_size > (unsigned long)(xdp->data_end - xdp->data)))
2529 		return -EFAULT;
2530 
2531 	return bpf_event_output(map, flags, meta, meta_size, xdp, xdp_size,
2532 				bpf_xdp_copy);
2533 }
2534 
2535 static const struct bpf_func_proto bpf_xdp_event_output_proto = {
2536 	.func		= bpf_xdp_event_output,
2537 	.gpl_only	= true,
2538 	.ret_type	= RET_INTEGER,
2539 	.arg1_type	= ARG_PTR_TO_CTX,
2540 	.arg2_type	= ARG_CONST_MAP_PTR,
2541 	.arg3_type	= ARG_ANYTHING,
2542 	.arg4_type	= ARG_PTR_TO_STACK,
2543 	.arg5_type	= ARG_CONST_STACK_SIZE,
2544 };
2545 
BPF_CALL_1(bpf_get_socket_cookie,struct sk_buff *,skb)2546 BPF_CALL_1(bpf_get_socket_cookie, struct sk_buff *, skb)
2547 {
2548 	return skb->sk ? sock_gen_cookie(skb->sk) : 0;
2549 }
2550 
2551 static const struct bpf_func_proto bpf_get_socket_cookie_proto = {
2552 	.func           = bpf_get_socket_cookie,
2553 	.gpl_only       = false,
2554 	.ret_type       = RET_INTEGER,
2555 	.arg1_type      = ARG_PTR_TO_CTX,
2556 };
2557 
BPF_CALL_1(bpf_get_socket_uid,struct sk_buff *,skb)2558 BPF_CALL_1(bpf_get_socket_uid, struct sk_buff *, skb)
2559 {
2560 	struct sock *sk = sk_to_full_sk(skb->sk);
2561 	kuid_t kuid;
2562 
2563 	if (!sk || !sk_fullsock(sk))
2564 		return overflowuid;
2565 	kuid = sock_net_uid(sock_net(sk), sk);
2566 	return from_kuid_munged(sock_net(sk)->user_ns, kuid);
2567 }
2568 
2569 static const struct bpf_func_proto bpf_get_socket_uid_proto = {
2570 	.func           = bpf_get_socket_uid,
2571 	.gpl_only       = false,
2572 	.ret_type       = RET_INTEGER,
2573 	.arg1_type      = ARG_PTR_TO_CTX,
2574 };
2575 
2576 static const struct bpf_func_proto *
sk_filter_func_proto(enum bpf_func_id func_id)2577 sk_filter_func_proto(enum bpf_func_id func_id)
2578 {
2579 	switch (func_id) {
2580 	case BPF_FUNC_map_lookup_elem:
2581 		return &bpf_map_lookup_elem_proto;
2582 	case BPF_FUNC_map_update_elem:
2583 		return &bpf_map_update_elem_proto;
2584 	case BPF_FUNC_map_delete_elem:
2585 		return &bpf_map_delete_elem_proto;
2586 	case BPF_FUNC_get_prandom_u32:
2587 		return &bpf_get_prandom_u32_proto;
2588 	case BPF_FUNC_get_smp_processor_id:
2589 		return &bpf_get_raw_smp_processor_id_proto;
2590 	case BPF_FUNC_tail_call:
2591 		return &bpf_tail_call_proto;
2592 	case BPF_FUNC_ktime_get_ns:
2593 		return &bpf_ktime_get_ns_proto;
2594 	case BPF_FUNC_trace_printk:
2595 		if (capable(CAP_SYS_ADMIN))
2596 			return bpf_get_trace_printk_proto();
2597 	case BPF_FUNC_get_socket_cookie:
2598 		return &bpf_get_socket_cookie_proto;
2599 	case BPF_FUNC_get_socket_uid:
2600 		return &bpf_get_socket_uid_proto;
2601 	default:
2602 		return NULL;
2603 	}
2604 }
2605 
2606 static const struct bpf_func_proto *
tc_cls_act_func_proto(enum bpf_func_id func_id)2607 tc_cls_act_func_proto(enum bpf_func_id func_id)
2608 {
2609 	switch (func_id) {
2610 	case BPF_FUNC_skb_store_bytes:
2611 		return &bpf_skb_store_bytes_proto;
2612 	case BPF_FUNC_skb_load_bytes:
2613 		return &bpf_skb_load_bytes_proto;
2614 	case BPF_FUNC_skb_pull_data:
2615 		return &bpf_skb_pull_data_proto;
2616 	case BPF_FUNC_csum_diff:
2617 		return &bpf_csum_diff_proto;
2618 	case BPF_FUNC_csum_update:
2619 		return &bpf_csum_update_proto;
2620 	case BPF_FUNC_l3_csum_replace:
2621 		return &bpf_l3_csum_replace_proto;
2622 	case BPF_FUNC_l4_csum_replace:
2623 		return &bpf_l4_csum_replace_proto;
2624 	case BPF_FUNC_clone_redirect:
2625 		return &bpf_clone_redirect_proto;
2626 	case BPF_FUNC_get_cgroup_classid:
2627 		return &bpf_get_cgroup_classid_proto;
2628 	case BPF_FUNC_skb_vlan_push:
2629 		return &bpf_skb_vlan_push_proto;
2630 	case BPF_FUNC_skb_vlan_pop:
2631 		return &bpf_skb_vlan_pop_proto;
2632 	case BPF_FUNC_skb_change_proto:
2633 		return &bpf_skb_change_proto_proto;
2634 	case BPF_FUNC_skb_change_type:
2635 		return &bpf_skb_change_type_proto;
2636 	case BPF_FUNC_skb_change_tail:
2637 		return &bpf_skb_change_tail_proto;
2638 	case BPF_FUNC_skb_get_tunnel_key:
2639 		return &bpf_skb_get_tunnel_key_proto;
2640 	case BPF_FUNC_skb_set_tunnel_key:
2641 		return bpf_get_skb_set_tunnel_proto(func_id);
2642 	case BPF_FUNC_skb_get_tunnel_opt:
2643 		return &bpf_skb_get_tunnel_opt_proto;
2644 	case BPF_FUNC_skb_set_tunnel_opt:
2645 		return bpf_get_skb_set_tunnel_proto(func_id);
2646 	case BPF_FUNC_redirect:
2647 		return &bpf_redirect_proto;
2648 	case BPF_FUNC_get_route_realm:
2649 		return &bpf_get_route_realm_proto;
2650 	case BPF_FUNC_get_hash_recalc:
2651 		return &bpf_get_hash_recalc_proto;
2652 	case BPF_FUNC_set_hash_invalid:
2653 		return &bpf_set_hash_invalid_proto;
2654 	case BPF_FUNC_perf_event_output:
2655 		return &bpf_skb_event_output_proto;
2656 	case BPF_FUNC_get_smp_processor_id:
2657 		return &bpf_get_smp_processor_id_proto;
2658 	case BPF_FUNC_skb_under_cgroup:
2659 		return &bpf_skb_under_cgroup_proto;
2660 	default:
2661 		return sk_filter_func_proto(func_id);
2662 	}
2663 }
2664 
2665 static const struct bpf_func_proto *
xdp_func_proto(enum bpf_func_id func_id)2666 xdp_func_proto(enum bpf_func_id func_id)
2667 {
2668 	switch (func_id) {
2669 	case BPF_FUNC_perf_event_output:
2670 		return &bpf_xdp_event_output_proto;
2671 	case BPF_FUNC_get_smp_processor_id:
2672 		return &bpf_get_smp_processor_id_proto;
2673 	default:
2674 		return sk_filter_func_proto(func_id);
2675 	}
2676 }
2677 
2678 static const struct bpf_func_proto *
cg_skb_func_proto(enum bpf_func_id func_id)2679 cg_skb_func_proto(enum bpf_func_id func_id)
2680 {
2681 	switch (func_id) {
2682 	case BPF_FUNC_skb_load_bytes:
2683 		return &bpf_skb_load_bytes_proto;
2684 	default:
2685 		return sk_filter_func_proto(func_id);
2686 	}
2687 }
2688 
__is_valid_access(int off,int size,enum bpf_access_type type)2689 static bool __is_valid_access(int off, int size, enum bpf_access_type type)
2690 {
2691 	if (off < 0 || off >= sizeof(struct __sk_buff))
2692 		return false;
2693 	/* The verifier guarantees that size > 0. */
2694 	if (off % size != 0)
2695 		return false;
2696 	if (size != sizeof(__u32))
2697 		return false;
2698 
2699 	return true;
2700 }
2701 
sk_filter_is_valid_access(int off,int size,enum bpf_access_type type,enum bpf_reg_type * reg_type)2702 static bool sk_filter_is_valid_access(int off, int size,
2703 				      enum bpf_access_type type,
2704 				      enum bpf_reg_type *reg_type)
2705 {
2706 	switch (off) {
2707 	case offsetof(struct __sk_buff, tc_classid):
2708 	case offsetof(struct __sk_buff, data):
2709 	case offsetof(struct __sk_buff, data_end):
2710 		return false;
2711 	}
2712 
2713 	if (type == BPF_WRITE) {
2714 		switch (off) {
2715 		case offsetof(struct __sk_buff, cb[0]) ...
2716 		     offsetof(struct __sk_buff, cb[4]):
2717 			break;
2718 		default:
2719 			return false;
2720 		}
2721 	}
2722 
2723 	return __is_valid_access(off, size, type);
2724 }
2725 
tc_cls_act_prologue(struct bpf_insn * insn_buf,bool direct_write,const struct bpf_prog * prog)2726 static int tc_cls_act_prologue(struct bpf_insn *insn_buf, bool direct_write,
2727 			       const struct bpf_prog *prog)
2728 {
2729 	struct bpf_insn *insn = insn_buf;
2730 
2731 	if (!direct_write)
2732 		return 0;
2733 
2734 	/* if (!skb->cloned)
2735 	 *       goto start;
2736 	 *
2737 	 * (Fast-path, otherwise approximation that we might be
2738 	 *  a clone, do the rest in helper.)
2739 	 */
2740 	*insn++ = BPF_LDX_MEM(BPF_B, BPF_REG_6, BPF_REG_1, CLONED_OFFSET());
2741 	*insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_6, CLONED_MASK);
2742 	*insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_6, 0, 7);
2743 
2744 	/* ret = bpf_skb_pull_data(skb, 0); */
2745 	*insn++ = BPF_MOV64_REG(BPF_REG_6, BPF_REG_1);
2746 	*insn++ = BPF_ALU64_REG(BPF_XOR, BPF_REG_2, BPF_REG_2);
2747 	*insn++ = BPF_RAW_INSN(BPF_JMP | BPF_CALL, 0, 0, 0,
2748 			       BPF_FUNC_skb_pull_data);
2749 	/* if (!ret)
2750 	 *      goto restore;
2751 	 * return TC_ACT_SHOT;
2752 	 */
2753 	*insn++ = BPF_JMP_IMM(BPF_JEQ, BPF_REG_0, 0, 2);
2754 	*insn++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_0, TC_ACT_SHOT);
2755 	*insn++ = BPF_EXIT_INSN();
2756 
2757 	/* restore: */
2758 	*insn++ = BPF_MOV64_REG(BPF_REG_1, BPF_REG_6);
2759 	/* start: */
2760 	*insn++ = prog->insnsi[0];
2761 
2762 	return insn - insn_buf;
2763 }
2764 
tc_cls_act_is_valid_access(int off,int size,enum bpf_access_type type,enum bpf_reg_type * reg_type)2765 static bool tc_cls_act_is_valid_access(int off, int size,
2766 				       enum bpf_access_type type,
2767 				       enum bpf_reg_type *reg_type)
2768 {
2769 	if (type == BPF_WRITE) {
2770 		switch (off) {
2771 		case offsetof(struct __sk_buff, mark):
2772 		case offsetof(struct __sk_buff, tc_index):
2773 		case offsetof(struct __sk_buff, priority):
2774 		case offsetof(struct __sk_buff, cb[0]) ...
2775 		     offsetof(struct __sk_buff, cb[4]):
2776 		case offsetof(struct __sk_buff, tc_classid):
2777 			break;
2778 		default:
2779 			return false;
2780 		}
2781 	}
2782 
2783 	switch (off) {
2784 	case offsetof(struct __sk_buff, data):
2785 		*reg_type = PTR_TO_PACKET;
2786 		break;
2787 	case offsetof(struct __sk_buff, data_end):
2788 		*reg_type = PTR_TO_PACKET_END;
2789 		break;
2790 	}
2791 
2792 	return __is_valid_access(off, size, type);
2793 }
2794 
__is_valid_xdp_access(int off,int size,enum bpf_access_type type)2795 static bool __is_valid_xdp_access(int off, int size,
2796 				  enum bpf_access_type type)
2797 {
2798 	if (off < 0 || off >= sizeof(struct xdp_md))
2799 		return false;
2800 	if (off % size != 0)
2801 		return false;
2802 	if (size != sizeof(__u32))
2803 		return false;
2804 
2805 	return true;
2806 }
2807 
xdp_is_valid_access(int off,int size,enum bpf_access_type type,enum bpf_reg_type * reg_type)2808 static bool xdp_is_valid_access(int off, int size,
2809 				enum bpf_access_type type,
2810 				enum bpf_reg_type *reg_type)
2811 {
2812 	if (type == BPF_WRITE)
2813 		return false;
2814 
2815 	switch (off) {
2816 	case offsetof(struct xdp_md, data):
2817 		*reg_type = PTR_TO_PACKET;
2818 		break;
2819 	case offsetof(struct xdp_md, data_end):
2820 		*reg_type = PTR_TO_PACKET_END;
2821 		break;
2822 	}
2823 
2824 	return __is_valid_xdp_access(off, size, type);
2825 }
2826 
bpf_warn_invalid_xdp_action(u32 act)2827 void bpf_warn_invalid_xdp_action(u32 act)
2828 {
2829 	WARN_ONCE(1, "Illegal XDP return value %u, expect packet loss\n", act);
2830 }
2831 EXPORT_SYMBOL_GPL(bpf_warn_invalid_xdp_action);
2832 
sk_filter_convert_ctx_access(enum bpf_access_type type,int dst_reg,int src_reg,int ctx_off,struct bpf_insn * insn_buf,struct bpf_prog * prog)2833 static u32 sk_filter_convert_ctx_access(enum bpf_access_type type, int dst_reg,
2834 					int src_reg, int ctx_off,
2835 					struct bpf_insn *insn_buf,
2836 					struct bpf_prog *prog)
2837 {
2838 	struct bpf_insn *insn = insn_buf;
2839 
2840 	switch (ctx_off) {
2841 	case offsetof(struct __sk_buff, len):
2842 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
2843 
2844 		*insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2845 				      offsetof(struct sk_buff, len));
2846 		break;
2847 
2848 	case offsetof(struct __sk_buff, protocol):
2849 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
2850 
2851 		*insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2852 				      offsetof(struct sk_buff, protocol));
2853 		break;
2854 
2855 	case offsetof(struct __sk_buff, vlan_proto):
2856 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_proto) != 2);
2857 
2858 		*insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2859 				      offsetof(struct sk_buff, vlan_proto));
2860 		break;
2861 
2862 	case offsetof(struct __sk_buff, priority):
2863 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, priority) != 4);
2864 
2865 		if (type == BPF_WRITE)
2866 			*insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
2867 					      offsetof(struct sk_buff, priority));
2868 		else
2869 			*insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2870 					      offsetof(struct sk_buff, priority));
2871 		break;
2872 
2873 	case offsetof(struct __sk_buff, ingress_ifindex):
2874 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, skb_iif) != 4);
2875 
2876 		*insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2877 				      offsetof(struct sk_buff, skb_iif));
2878 		break;
2879 
2880 	case offsetof(struct __sk_buff, ifindex):
2881 		BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
2882 
2883 		*insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
2884 				      dst_reg, src_reg,
2885 				      offsetof(struct sk_buff, dev));
2886 		*insn++ = BPF_JMP_IMM(BPF_JEQ, dst_reg, 0, 1);
2887 		*insn++ = BPF_LDX_MEM(BPF_W, dst_reg, dst_reg,
2888 				      offsetof(struct net_device, ifindex));
2889 		break;
2890 
2891 	case offsetof(struct __sk_buff, hash):
2892 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
2893 
2894 		*insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2895 				      offsetof(struct sk_buff, hash));
2896 		break;
2897 
2898 	case offsetof(struct __sk_buff, mark):
2899 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
2900 
2901 		if (type == BPF_WRITE)
2902 			*insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg,
2903 					      offsetof(struct sk_buff, mark));
2904 		else
2905 			*insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg,
2906 					      offsetof(struct sk_buff, mark));
2907 		break;
2908 
2909 	case offsetof(struct __sk_buff, pkt_type):
2910 		return convert_skb_access(SKF_AD_PKTTYPE, dst_reg, src_reg, insn);
2911 
2912 	case offsetof(struct __sk_buff, queue_mapping):
2913 		return convert_skb_access(SKF_AD_QUEUE, dst_reg, src_reg, insn);
2914 
2915 	case offsetof(struct __sk_buff, vlan_present):
2916 		return convert_skb_access(SKF_AD_VLAN_TAG_PRESENT,
2917 					  dst_reg, src_reg, insn);
2918 
2919 	case offsetof(struct __sk_buff, vlan_tci):
2920 		return convert_skb_access(SKF_AD_VLAN_TAG,
2921 					  dst_reg, src_reg, insn);
2922 
2923 	case offsetof(struct __sk_buff, cb[0]) ...
2924 	     offsetof(struct __sk_buff, cb[4]):
2925 		BUILD_BUG_ON(FIELD_SIZEOF(struct qdisc_skb_cb, data) < 20);
2926 
2927 		prog->cb_access = 1;
2928 		ctx_off -= offsetof(struct __sk_buff, cb[0]);
2929 		ctx_off += offsetof(struct sk_buff, cb);
2930 		ctx_off += offsetof(struct qdisc_skb_cb, data);
2931 		if (type == BPF_WRITE)
2932 			*insn++ = BPF_STX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
2933 		else
2934 			*insn++ = BPF_LDX_MEM(BPF_W, dst_reg, src_reg, ctx_off);
2935 		break;
2936 
2937 	case offsetof(struct __sk_buff, tc_classid):
2938 		ctx_off -= offsetof(struct __sk_buff, tc_classid);
2939 		ctx_off += offsetof(struct sk_buff, cb);
2940 		ctx_off += offsetof(struct qdisc_skb_cb, tc_classid);
2941 		if (type == BPF_WRITE)
2942 			*insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg, ctx_off);
2943 		else
2944 			*insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg, ctx_off);
2945 		break;
2946 
2947 	case offsetof(struct __sk_buff, data):
2948 		*insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, data),
2949 				      dst_reg, src_reg,
2950 				      offsetof(struct sk_buff, data));
2951 		break;
2952 
2953 	case offsetof(struct __sk_buff, data_end):
2954 		ctx_off -= offsetof(struct __sk_buff, data_end);
2955 		ctx_off += offsetof(struct sk_buff, cb);
2956 		ctx_off += offsetof(struct bpf_skb_data_end, data_end);
2957 		*insn++ = BPF_LDX_MEM(BPF_SIZEOF(void *), dst_reg, src_reg,
2958 				      ctx_off);
2959 		break;
2960 
2961 	case offsetof(struct __sk_buff, tc_index):
2962 #ifdef CONFIG_NET_SCHED
2963 		BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, tc_index) != 2);
2964 
2965 		if (type == BPF_WRITE)
2966 			*insn++ = BPF_STX_MEM(BPF_H, dst_reg, src_reg,
2967 					      offsetof(struct sk_buff, tc_index));
2968 		else
2969 			*insn++ = BPF_LDX_MEM(BPF_H, dst_reg, src_reg,
2970 					      offsetof(struct sk_buff, tc_index));
2971 		break;
2972 #else
2973 		if (type == BPF_WRITE)
2974 			*insn++ = BPF_MOV64_REG(dst_reg, dst_reg);
2975 		else
2976 			*insn++ = BPF_MOV64_IMM(dst_reg, 0);
2977 		break;
2978 #endif
2979 	}
2980 
2981 	return insn - insn_buf;
2982 }
2983 
tc_cls_act_convert_ctx_access(enum bpf_access_type type,int dst_reg,int src_reg,int ctx_off,struct bpf_insn * insn_buf,struct bpf_prog * prog)2984 static u32 tc_cls_act_convert_ctx_access(enum bpf_access_type type, int dst_reg,
2985 					 int src_reg, int ctx_off,
2986 					 struct bpf_insn *insn_buf,
2987 					 struct bpf_prog *prog)
2988 {
2989 	struct bpf_insn *insn = insn_buf;
2990 
2991 	switch (ctx_off) {
2992 	case offsetof(struct __sk_buff, ifindex):
2993 		BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
2994 
2995 		*insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct sk_buff, dev),
2996 				      dst_reg, src_reg,
2997 				      offsetof(struct sk_buff, dev));
2998 		*insn++ = BPF_LDX_MEM(BPF_W, dst_reg, dst_reg,
2999 				      offsetof(struct net_device, ifindex));
3000 		break;
3001 	default:
3002 		return sk_filter_convert_ctx_access(type, dst_reg, src_reg,
3003 						    ctx_off, insn_buf, prog);
3004 	}
3005 
3006 	return insn - insn_buf;
3007 }
3008 
xdp_convert_ctx_access(enum bpf_access_type type,int dst_reg,int src_reg,int ctx_off,struct bpf_insn * insn_buf,struct bpf_prog * prog)3009 static u32 xdp_convert_ctx_access(enum bpf_access_type type, int dst_reg,
3010 				  int src_reg, int ctx_off,
3011 				  struct bpf_insn *insn_buf,
3012 				  struct bpf_prog *prog)
3013 {
3014 	struct bpf_insn *insn = insn_buf;
3015 
3016 	switch (ctx_off) {
3017 	case offsetof(struct xdp_md, data):
3018 		*insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data),
3019 				      dst_reg, src_reg,
3020 				      offsetof(struct xdp_buff, data));
3021 		break;
3022 	case offsetof(struct xdp_md, data_end):
3023 		*insn++ = BPF_LDX_MEM(BPF_FIELD_SIZEOF(struct xdp_buff, data_end),
3024 				      dst_reg, src_reg,
3025 				      offsetof(struct xdp_buff, data_end));
3026 		break;
3027 	}
3028 
3029 	return insn - insn_buf;
3030 }
3031 
3032 static const struct bpf_verifier_ops sk_filter_ops = {
3033 	.get_func_proto		= sk_filter_func_proto,
3034 	.is_valid_access	= sk_filter_is_valid_access,
3035 	.convert_ctx_access	= sk_filter_convert_ctx_access,
3036 };
3037 
3038 static const struct bpf_verifier_ops tc_cls_act_ops = {
3039 	.get_func_proto		= tc_cls_act_func_proto,
3040 	.is_valid_access	= tc_cls_act_is_valid_access,
3041 	.convert_ctx_access	= tc_cls_act_convert_ctx_access,
3042 	.gen_prologue		= tc_cls_act_prologue,
3043 };
3044 
3045 static const struct bpf_verifier_ops xdp_ops = {
3046 	.get_func_proto		= xdp_func_proto,
3047 	.is_valid_access	= xdp_is_valid_access,
3048 	.convert_ctx_access	= xdp_convert_ctx_access,
3049 };
3050 
3051 static const struct bpf_verifier_ops cg_skb_ops = {
3052 	.get_func_proto		= cg_skb_func_proto,
3053 	.is_valid_access	= sk_filter_is_valid_access,
3054 	.convert_ctx_access	= sk_filter_convert_ctx_access,
3055 };
3056 
3057 static struct bpf_prog_type_list sk_filter_type __read_mostly = {
3058 	.ops	= &sk_filter_ops,
3059 	.type	= BPF_PROG_TYPE_SOCKET_FILTER,
3060 };
3061 
3062 static struct bpf_prog_type_list sched_cls_type __read_mostly = {
3063 	.ops	= &tc_cls_act_ops,
3064 	.type	= BPF_PROG_TYPE_SCHED_CLS,
3065 };
3066 
3067 static struct bpf_prog_type_list sched_act_type __read_mostly = {
3068 	.ops	= &tc_cls_act_ops,
3069 	.type	= BPF_PROG_TYPE_SCHED_ACT,
3070 };
3071 
3072 static struct bpf_prog_type_list xdp_type __read_mostly = {
3073 	.ops	= &xdp_ops,
3074 	.type	= BPF_PROG_TYPE_XDP,
3075 };
3076 
3077 static struct bpf_prog_type_list cg_skb_type __read_mostly = {
3078 	.ops	= &cg_skb_ops,
3079 	.type	= BPF_PROG_TYPE_CGROUP_SKB,
3080 };
3081 
register_sk_filter_ops(void)3082 static int __init register_sk_filter_ops(void)
3083 {
3084 	bpf_register_prog_type(&sk_filter_type);
3085 	bpf_register_prog_type(&sched_cls_type);
3086 	bpf_register_prog_type(&sched_act_type);
3087 	bpf_register_prog_type(&xdp_type);
3088 	bpf_register_prog_type(&cg_skb_type);
3089 
3090 	return 0;
3091 }
3092 late_initcall(register_sk_filter_ops);
3093 
sk_detach_filter(struct sock * sk)3094 int sk_detach_filter(struct sock *sk)
3095 {
3096 	int ret = -ENOENT;
3097 	struct sk_filter *filter;
3098 
3099 	if (sock_flag(sk, SOCK_FILTER_LOCKED))
3100 		return -EPERM;
3101 
3102 	filter = rcu_dereference_protected(sk->sk_filter,
3103 					   lockdep_sock_is_held(sk));
3104 	if (filter) {
3105 		RCU_INIT_POINTER(sk->sk_filter, NULL);
3106 		sk_filter_uncharge(sk, filter);
3107 		ret = 0;
3108 	}
3109 
3110 	return ret;
3111 }
3112 EXPORT_SYMBOL_GPL(sk_detach_filter);
3113 
sk_get_filter(struct sock * sk,struct sock_filter __user * ubuf,unsigned int len)3114 int sk_get_filter(struct sock *sk, struct sock_filter __user *ubuf,
3115 		  unsigned int len)
3116 {
3117 	struct sock_fprog_kern *fprog;
3118 	struct sk_filter *filter;
3119 	int ret = 0;
3120 
3121 	lock_sock(sk);
3122 	filter = rcu_dereference_protected(sk->sk_filter,
3123 					   lockdep_sock_is_held(sk));
3124 	if (!filter)
3125 		goto out;
3126 
3127 	/* We're copying the filter that has been originally attached,
3128 	 * so no conversion/decode needed anymore. eBPF programs that
3129 	 * have no original program cannot be dumped through this.
3130 	 */
3131 	ret = -EACCES;
3132 	fprog = filter->prog->orig_prog;
3133 	if (!fprog)
3134 		goto out;
3135 
3136 	ret = fprog->len;
3137 	if (!len)
3138 		/* User space only enquires number of filter blocks. */
3139 		goto out;
3140 
3141 	ret = -EINVAL;
3142 	if (len < fprog->len)
3143 		goto out;
3144 
3145 	ret = -EFAULT;
3146 	if (copy_to_user(ubuf, fprog->filter, bpf_classic_proglen(fprog)))
3147 		goto out;
3148 
3149 	/* Instead of bytes, the API requests to return the number
3150 	 * of filter blocks.
3151 	 */
3152 	ret = fprog->len;
3153 out:
3154 	release_sock(sk);
3155 	return ret;
3156 }
3157