xref: /linux/kernel/bpf/core.c (revision c0e297dc61f8d4453e07afbea1fa8d0e67cd4a34)
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/filter.h>
25 #include <linux/skbuff.h>
26 #include <linux/vmalloc.h>
27 #include <linux/random.h>
28 #include <linux/moduleloader.h>
29 #include <linux/bpf.h>
30 
31 #include <asm/unaligned.h>
32 
33 /* Registers */
34 #define BPF_R0	regs[BPF_REG_0]
35 #define BPF_R1	regs[BPF_REG_1]
36 #define BPF_R2	regs[BPF_REG_2]
37 #define BPF_R3	regs[BPF_REG_3]
38 #define BPF_R4	regs[BPF_REG_4]
39 #define BPF_R5	regs[BPF_REG_5]
40 #define BPF_R6	regs[BPF_REG_6]
41 #define BPF_R7	regs[BPF_REG_7]
42 #define BPF_R8	regs[BPF_REG_8]
43 #define BPF_R9	regs[BPF_REG_9]
44 #define BPF_R10	regs[BPF_REG_10]
45 
46 /* Named registers */
47 #define DST	regs[insn->dst_reg]
48 #define SRC	regs[insn->src_reg]
49 #define FP	regs[BPF_REG_FP]
50 #define ARG1	regs[BPF_REG_ARG1]
51 #define CTX	regs[BPF_REG_CTX]
52 #define IMM	insn->imm
53 
54 /* No hurry in this branch
55  *
56  * Exported for the bpf jit load helper.
57  */
58 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
59 {
60 	u8 *ptr = NULL;
61 
62 	if (k >= SKF_NET_OFF)
63 		ptr = skb_network_header(skb) + k - SKF_NET_OFF;
64 	else if (k >= SKF_LL_OFF)
65 		ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
66 
67 	if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
68 		return ptr;
69 
70 	return NULL;
71 }
72 
73 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
74 {
75 	gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
76 			  gfp_extra_flags;
77 	struct bpf_prog_aux *aux;
78 	struct bpf_prog *fp;
79 
80 	size = round_up(size, PAGE_SIZE);
81 	fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
82 	if (fp == NULL)
83 		return NULL;
84 
85 	aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
86 	if (aux == NULL) {
87 		vfree(fp);
88 		return NULL;
89 	}
90 
91 	fp->pages = size / PAGE_SIZE;
92 	fp->aux = aux;
93 
94 	return fp;
95 }
96 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
97 
98 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
99 				  gfp_t gfp_extra_flags)
100 {
101 	gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
102 			  gfp_extra_flags;
103 	struct bpf_prog *fp;
104 
105 	BUG_ON(fp_old == NULL);
106 
107 	size = round_up(size, PAGE_SIZE);
108 	if (size <= fp_old->pages * PAGE_SIZE)
109 		return fp_old;
110 
111 	fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
112 	if (fp != NULL) {
113 		memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
114 		fp->pages = size / PAGE_SIZE;
115 
116 		/* We keep fp->aux from fp_old around in the new
117 		 * reallocated structure.
118 		 */
119 		fp_old->aux = NULL;
120 		__bpf_prog_free(fp_old);
121 	}
122 
123 	return fp;
124 }
125 EXPORT_SYMBOL_GPL(bpf_prog_realloc);
126 
127 void __bpf_prog_free(struct bpf_prog *fp)
128 {
129 	kfree(fp->aux);
130 	vfree(fp);
131 }
132 EXPORT_SYMBOL_GPL(__bpf_prog_free);
133 
134 #ifdef CONFIG_BPF_JIT
135 struct bpf_binary_header *
136 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
137 		     unsigned int alignment,
138 		     bpf_jit_fill_hole_t bpf_fill_ill_insns)
139 {
140 	struct bpf_binary_header *hdr;
141 	unsigned int size, hole, start;
142 
143 	/* Most of BPF filters are really small, but if some of them
144 	 * fill a page, allow at least 128 extra bytes to insert a
145 	 * random section of illegal instructions.
146 	 */
147 	size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
148 	hdr = module_alloc(size);
149 	if (hdr == NULL)
150 		return NULL;
151 
152 	/* Fill space with illegal/arch-dep instructions. */
153 	bpf_fill_ill_insns(hdr, size);
154 
155 	hdr->pages = size / PAGE_SIZE;
156 	hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
157 		     PAGE_SIZE - sizeof(*hdr));
158 	start = (prandom_u32() % hole) & ~(alignment - 1);
159 
160 	/* Leave a random number of instructions before BPF code. */
161 	*image_ptr = &hdr->image[start];
162 
163 	return hdr;
164 }
165 
166 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
167 {
168 	module_memfree(hdr);
169 }
170 #endif /* CONFIG_BPF_JIT */
171 
172 /* Base function for offset calculation. Needs to go into .text section,
173  * therefore keeping it non-static as well; will also be used by JITs
174  * anyway later on, so do not let the compiler omit it.
175  */
176 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
177 {
178 	return 0;
179 }
180 
181 /**
182  *	__bpf_prog_run - run eBPF program on a given context
183  *	@ctx: is the data we are operating on
184  *	@insn: is the array of eBPF instructions
185  *
186  * Decode and execute eBPF instructions.
187  */
188 static unsigned int __bpf_prog_run(void *ctx, const struct bpf_insn *insn)
189 {
190 	u64 stack[MAX_BPF_STACK / sizeof(u64)];
191 	u64 regs[MAX_BPF_REG], tmp;
192 	static const void *jumptable[256] = {
193 		[0 ... 255] = &&default_label,
194 		/* Now overwrite non-defaults ... */
195 		/* 32 bit ALU operations */
196 		[BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
197 		[BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
198 		[BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
199 		[BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
200 		[BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
201 		[BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
202 		[BPF_ALU | BPF_OR | BPF_X]  = &&ALU_OR_X,
203 		[BPF_ALU | BPF_OR | BPF_K]  = &&ALU_OR_K,
204 		[BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
205 		[BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
206 		[BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
207 		[BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
208 		[BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
209 		[BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
210 		[BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
211 		[BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
212 		[BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
213 		[BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
214 		[BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
215 		[BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
216 		[BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
217 		[BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
218 		[BPF_ALU | BPF_NEG] = &&ALU_NEG,
219 		[BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
220 		[BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
221 		/* 64 bit ALU operations */
222 		[BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
223 		[BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
224 		[BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
225 		[BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
226 		[BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
227 		[BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
228 		[BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
229 		[BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
230 		[BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
231 		[BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
232 		[BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
233 		[BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
234 		[BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
235 		[BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
236 		[BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
237 		[BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
238 		[BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
239 		[BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
240 		[BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
241 		[BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
242 		[BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
243 		[BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
244 		[BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
245 		[BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
246 		[BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
247 		/* Call instruction */
248 		[BPF_JMP | BPF_CALL] = &&JMP_CALL,
249 		[BPF_JMP | BPF_CALL | BPF_X] = &&JMP_TAIL_CALL,
250 		/* Jumps */
251 		[BPF_JMP | BPF_JA] = &&JMP_JA,
252 		[BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
253 		[BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
254 		[BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
255 		[BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
256 		[BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
257 		[BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
258 		[BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
259 		[BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
260 		[BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
261 		[BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
262 		[BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
263 		[BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
264 		[BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
265 		[BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
266 		/* Program return */
267 		[BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
268 		/* Store instructions */
269 		[BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
270 		[BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
271 		[BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
272 		[BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
273 		[BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
274 		[BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
275 		[BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
276 		[BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
277 		[BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
278 		[BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
279 		/* Load instructions */
280 		[BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
281 		[BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
282 		[BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
283 		[BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
284 		[BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
285 		[BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
286 		[BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
287 		[BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
288 		[BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
289 		[BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
290 		[BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
291 	};
292 	u32 tail_call_cnt = 0;
293 	void *ptr;
294 	int off;
295 
296 #define CONT	 ({ insn++; goto select_insn; })
297 #define CONT_JMP ({ insn++; goto select_insn; })
298 
299 	FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)];
300 	ARG1 = (u64) (unsigned long) ctx;
301 
302 	/* Registers used in classic BPF programs need to be reset first. */
303 	regs[BPF_REG_A] = 0;
304 	regs[BPF_REG_X] = 0;
305 
306 select_insn:
307 	goto *jumptable[insn->code];
308 
309 	/* ALU */
310 #define ALU(OPCODE, OP)			\
311 	ALU64_##OPCODE##_X:		\
312 		DST = DST OP SRC;	\
313 		CONT;			\
314 	ALU_##OPCODE##_X:		\
315 		DST = (u32) DST OP (u32) SRC;	\
316 		CONT;			\
317 	ALU64_##OPCODE##_K:		\
318 		DST = DST OP IMM;		\
319 		CONT;			\
320 	ALU_##OPCODE##_K:		\
321 		DST = (u32) DST OP (u32) IMM;	\
322 		CONT;
323 
324 	ALU(ADD,  +)
325 	ALU(SUB,  -)
326 	ALU(AND,  &)
327 	ALU(OR,   |)
328 	ALU(LSH, <<)
329 	ALU(RSH, >>)
330 	ALU(XOR,  ^)
331 	ALU(MUL,  *)
332 #undef ALU
333 	ALU_NEG:
334 		DST = (u32) -DST;
335 		CONT;
336 	ALU64_NEG:
337 		DST = -DST;
338 		CONT;
339 	ALU_MOV_X:
340 		DST = (u32) SRC;
341 		CONT;
342 	ALU_MOV_K:
343 		DST = (u32) IMM;
344 		CONT;
345 	ALU64_MOV_X:
346 		DST = SRC;
347 		CONT;
348 	ALU64_MOV_K:
349 		DST = IMM;
350 		CONT;
351 	LD_IMM_DW:
352 		DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
353 		insn++;
354 		CONT;
355 	ALU64_ARSH_X:
356 		(*(s64 *) &DST) >>= SRC;
357 		CONT;
358 	ALU64_ARSH_K:
359 		(*(s64 *) &DST) >>= IMM;
360 		CONT;
361 	ALU64_MOD_X:
362 		if (unlikely(SRC == 0))
363 			return 0;
364 		div64_u64_rem(DST, SRC, &tmp);
365 		DST = tmp;
366 		CONT;
367 	ALU_MOD_X:
368 		if (unlikely(SRC == 0))
369 			return 0;
370 		tmp = (u32) DST;
371 		DST = do_div(tmp, (u32) SRC);
372 		CONT;
373 	ALU64_MOD_K:
374 		div64_u64_rem(DST, IMM, &tmp);
375 		DST = tmp;
376 		CONT;
377 	ALU_MOD_K:
378 		tmp = (u32) DST;
379 		DST = do_div(tmp, (u32) IMM);
380 		CONT;
381 	ALU64_DIV_X:
382 		if (unlikely(SRC == 0))
383 			return 0;
384 		DST = div64_u64(DST, SRC);
385 		CONT;
386 	ALU_DIV_X:
387 		if (unlikely(SRC == 0))
388 			return 0;
389 		tmp = (u32) DST;
390 		do_div(tmp, (u32) SRC);
391 		DST = (u32) tmp;
392 		CONT;
393 	ALU64_DIV_K:
394 		DST = div64_u64(DST, IMM);
395 		CONT;
396 	ALU_DIV_K:
397 		tmp = (u32) DST;
398 		do_div(tmp, (u32) IMM);
399 		DST = (u32) tmp;
400 		CONT;
401 	ALU_END_TO_BE:
402 		switch (IMM) {
403 		case 16:
404 			DST = (__force u16) cpu_to_be16(DST);
405 			break;
406 		case 32:
407 			DST = (__force u32) cpu_to_be32(DST);
408 			break;
409 		case 64:
410 			DST = (__force u64) cpu_to_be64(DST);
411 			break;
412 		}
413 		CONT;
414 	ALU_END_TO_LE:
415 		switch (IMM) {
416 		case 16:
417 			DST = (__force u16) cpu_to_le16(DST);
418 			break;
419 		case 32:
420 			DST = (__force u32) cpu_to_le32(DST);
421 			break;
422 		case 64:
423 			DST = (__force u64) cpu_to_le64(DST);
424 			break;
425 		}
426 		CONT;
427 
428 	/* CALL */
429 	JMP_CALL:
430 		/* Function call scratches BPF_R1-BPF_R5 registers,
431 		 * preserves BPF_R6-BPF_R9, and stores return value
432 		 * into BPF_R0.
433 		 */
434 		BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
435 						       BPF_R4, BPF_R5);
436 		CONT;
437 
438 	JMP_TAIL_CALL: {
439 		struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
440 		struct bpf_array *array = container_of(map, struct bpf_array, map);
441 		struct bpf_prog *prog;
442 		u64 index = BPF_R3;
443 
444 		if (unlikely(index >= array->map.max_entries))
445 			goto out;
446 
447 		if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
448 			goto out;
449 
450 		tail_call_cnt++;
451 
452 		prog = READ_ONCE(array->prog[index]);
453 		if (unlikely(!prog))
454 			goto out;
455 
456 		ARG1 = BPF_R1;
457 		insn = prog->insnsi;
458 		goto select_insn;
459 out:
460 		CONT;
461 	}
462 	/* JMP */
463 	JMP_JA:
464 		insn += insn->off;
465 		CONT;
466 	JMP_JEQ_X:
467 		if (DST == SRC) {
468 			insn += insn->off;
469 			CONT_JMP;
470 		}
471 		CONT;
472 	JMP_JEQ_K:
473 		if (DST == IMM) {
474 			insn += insn->off;
475 			CONT_JMP;
476 		}
477 		CONT;
478 	JMP_JNE_X:
479 		if (DST != SRC) {
480 			insn += insn->off;
481 			CONT_JMP;
482 		}
483 		CONT;
484 	JMP_JNE_K:
485 		if (DST != IMM) {
486 			insn += insn->off;
487 			CONT_JMP;
488 		}
489 		CONT;
490 	JMP_JGT_X:
491 		if (DST > SRC) {
492 			insn += insn->off;
493 			CONT_JMP;
494 		}
495 		CONT;
496 	JMP_JGT_K:
497 		if (DST > IMM) {
498 			insn += insn->off;
499 			CONT_JMP;
500 		}
501 		CONT;
502 	JMP_JGE_X:
503 		if (DST >= SRC) {
504 			insn += insn->off;
505 			CONT_JMP;
506 		}
507 		CONT;
508 	JMP_JGE_K:
509 		if (DST >= IMM) {
510 			insn += insn->off;
511 			CONT_JMP;
512 		}
513 		CONT;
514 	JMP_JSGT_X:
515 		if (((s64) DST) > ((s64) SRC)) {
516 			insn += insn->off;
517 			CONT_JMP;
518 		}
519 		CONT;
520 	JMP_JSGT_K:
521 		if (((s64) DST) > ((s64) IMM)) {
522 			insn += insn->off;
523 			CONT_JMP;
524 		}
525 		CONT;
526 	JMP_JSGE_X:
527 		if (((s64) DST) >= ((s64) SRC)) {
528 			insn += insn->off;
529 			CONT_JMP;
530 		}
531 		CONT;
532 	JMP_JSGE_K:
533 		if (((s64) DST) >= ((s64) IMM)) {
534 			insn += insn->off;
535 			CONT_JMP;
536 		}
537 		CONT;
538 	JMP_JSET_X:
539 		if (DST & SRC) {
540 			insn += insn->off;
541 			CONT_JMP;
542 		}
543 		CONT;
544 	JMP_JSET_K:
545 		if (DST & IMM) {
546 			insn += insn->off;
547 			CONT_JMP;
548 		}
549 		CONT;
550 	JMP_EXIT:
551 		return BPF_R0;
552 
553 	/* STX and ST and LDX*/
554 #define LDST(SIZEOP, SIZE)						\
555 	STX_MEM_##SIZEOP:						\
556 		*(SIZE *)(unsigned long) (DST + insn->off) = SRC;	\
557 		CONT;							\
558 	ST_MEM_##SIZEOP:						\
559 		*(SIZE *)(unsigned long) (DST + insn->off) = IMM;	\
560 		CONT;							\
561 	LDX_MEM_##SIZEOP:						\
562 		DST = *(SIZE *)(unsigned long) (SRC + insn->off);	\
563 		CONT;
564 
565 	LDST(B,   u8)
566 	LDST(H,  u16)
567 	LDST(W,  u32)
568 	LDST(DW, u64)
569 #undef LDST
570 	STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
571 		atomic_add((u32) SRC, (atomic_t *)(unsigned long)
572 			   (DST + insn->off));
573 		CONT;
574 	STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
575 		atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
576 			     (DST + insn->off));
577 		CONT;
578 	LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
579 		off = IMM;
580 load_word:
581 		/* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are
582 		 * only appearing in the programs where ctx ==
583 		 * skb. All programs keep 'ctx' in regs[BPF_REG_CTX]
584 		 * == BPF_R6, bpf_convert_filter() saves it in BPF_R6,
585 		 * internal BPF verifier will check that BPF_R6 ==
586 		 * ctx.
587 		 *
588 		 * BPF_ABS and BPF_IND are wrappers of function calls,
589 		 * so they scratch BPF_R1-BPF_R5 registers, preserve
590 		 * BPF_R6-BPF_R9, and store return value into BPF_R0.
591 		 *
592 		 * Implicit input:
593 		 *   ctx == skb == BPF_R6 == CTX
594 		 *
595 		 * Explicit input:
596 		 *   SRC == any register
597 		 *   IMM == 32-bit immediate
598 		 *
599 		 * Output:
600 		 *   BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
601 		 */
602 
603 		ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
604 		if (likely(ptr != NULL)) {
605 			BPF_R0 = get_unaligned_be32(ptr);
606 			CONT;
607 		}
608 
609 		return 0;
610 	LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
611 		off = IMM;
612 load_half:
613 		ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
614 		if (likely(ptr != NULL)) {
615 			BPF_R0 = get_unaligned_be16(ptr);
616 			CONT;
617 		}
618 
619 		return 0;
620 	LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
621 		off = IMM;
622 load_byte:
623 		ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
624 		if (likely(ptr != NULL)) {
625 			BPF_R0 = *(u8 *)ptr;
626 			CONT;
627 		}
628 
629 		return 0;
630 	LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
631 		off = IMM + SRC;
632 		goto load_word;
633 	LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
634 		off = IMM + SRC;
635 		goto load_half;
636 	LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
637 		off = IMM + SRC;
638 		goto load_byte;
639 
640 	default_label:
641 		/* If we ever reach this, we have a bug somewhere. */
642 		WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
643 		return 0;
644 }
645 
646 bool bpf_prog_array_compatible(struct bpf_array *array,
647 			       const struct bpf_prog *fp)
648 {
649 	if (!array->owner_prog_type) {
650 		/* There's no owner yet where we could check for
651 		 * compatibility.
652 		 */
653 		array->owner_prog_type = fp->type;
654 		array->owner_jited = fp->jited;
655 
656 		return true;
657 	}
658 
659 	return array->owner_prog_type == fp->type &&
660 	       array->owner_jited == fp->jited;
661 }
662 
663 static int bpf_check_tail_call(const struct bpf_prog *fp)
664 {
665 	struct bpf_prog_aux *aux = fp->aux;
666 	int i;
667 
668 	for (i = 0; i < aux->used_map_cnt; i++) {
669 		struct bpf_map *map = aux->used_maps[i];
670 		struct bpf_array *array;
671 
672 		if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
673 			continue;
674 
675 		array = container_of(map, struct bpf_array, map);
676 		if (!bpf_prog_array_compatible(array, fp))
677 			return -EINVAL;
678 	}
679 
680 	return 0;
681 }
682 
683 /**
684  *	bpf_prog_select_runtime - select exec runtime for BPF program
685  *	@fp: bpf_prog populated with internal BPF program
686  *
687  * Try to JIT eBPF program, if JIT is not available, use interpreter.
688  * The BPF program will be executed via BPF_PROG_RUN() macro.
689  */
690 int bpf_prog_select_runtime(struct bpf_prog *fp)
691 {
692 	fp->bpf_func = (void *) __bpf_prog_run;
693 
694 	bpf_int_jit_compile(fp);
695 	bpf_prog_lock_ro(fp);
696 
697 	/* The tail call compatibility check can only be done at
698 	 * this late stage as we need to determine, if we deal
699 	 * with JITed or non JITed program concatenations and not
700 	 * all eBPF JITs might immediately support all features.
701 	 */
702 	return bpf_check_tail_call(fp);
703 }
704 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
705 
706 static void bpf_prog_free_deferred(struct work_struct *work)
707 {
708 	struct bpf_prog_aux *aux;
709 
710 	aux = container_of(work, struct bpf_prog_aux, work);
711 	bpf_jit_free(aux->prog);
712 }
713 
714 /* Free internal BPF program */
715 void bpf_prog_free(struct bpf_prog *fp)
716 {
717 	struct bpf_prog_aux *aux = fp->aux;
718 
719 	INIT_WORK(&aux->work, bpf_prog_free_deferred);
720 	aux->prog = fp;
721 	schedule_work(&aux->work);
722 }
723 EXPORT_SYMBOL_GPL(bpf_prog_free);
724 
725 /* Weak definitions of helper functions in case we don't have bpf syscall. */
726 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
727 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
728 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
729 
730 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
731 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
732 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
733 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
734 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
735 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
736 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
737 {
738 	return NULL;
739 }
740 
741 /* Always built-in helper functions. */
742 const struct bpf_func_proto bpf_tail_call_proto = {
743 	.func		= NULL,
744 	.gpl_only	= false,
745 	.ret_type	= RET_VOID,
746 	.arg1_type	= ARG_PTR_TO_CTX,
747 	.arg2_type	= ARG_CONST_MAP_PTR,
748 	.arg3_type	= ARG_ANYTHING,
749 };
750 
751 /* For classic BPF JITs that don't implement bpf_int_jit_compile(). */
752 void __weak bpf_int_jit_compile(struct bpf_prog *prog)
753 {
754 }
755 
756 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
757  * skb_copy_bits(), so provide a weak definition of it for NET-less config.
758  */
759 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
760 			 int len)
761 {
762 	return -EFAULT;
763 }
764