xref: /openbmc/linux/kernel/bpf/core.c (revision 293d5b43)
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 #include <linux/frame.h>
31 
32 #include <asm/unaligned.h>
33 
34 /* Registers */
35 #define BPF_R0	regs[BPF_REG_0]
36 #define BPF_R1	regs[BPF_REG_1]
37 #define BPF_R2	regs[BPF_REG_2]
38 #define BPF_R3	regs[BPF_REG_3]
39 #define BPF_R4	regs[BPF_REG_4]
40 #define BPF_R5	regs[BPF_REG_5]
41 #define BPF_R6	regs[BPF_REG_6]
42 #define BPF_R7	regs[BPF_REG_7]
43 #define BPF_R8	regs[BPF_REG_8]
44 #define BPF_R9	regs[BPF_REG_9]
45 #define BPF_R10	regs[BPF_REG_10]
46 
47 /* Named registers */
48 #define DST	regs[insn->dst_reg]
49 #define SRC	regs[insn->src_reg]
50 #define FP	regs[BPF_REG_FP]
51 #define ARG1	regs[BPF_REG_ARG1]
52 #define CTX	regs[BPF_REG_CTX]
53 #define IMM	insn->imm
54 
55 /* No hurry in this branch
56  *
57  * Exported for the bpf jit load helper.
58  */
59 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
60 {
61 	u8 *ptr = NULL;
62 
63 	if (k >= SKF_NET_OFF)
64 		ptr = skb_network_header(skb) + k - SKF_NET_OFF;
65 	else if (k >= SKF_LL_OFF)
66 		ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
67 
68 	if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
69 		return ptr;
70 
71 	return NULL;
72 }
73 
74 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
75 {
76 	gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
77 			  gfp_extra_flags;
78 	struct bpf_prog_aux *aux;
79 	struct bpf_prog *fp;
80 
81 	size = round_up(size, PAGE_SIZE);
82 	fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
83 	if (fp == NULL)
84 		return NULL;
85 
86 	kmemcheck_annotate_bitfield(fp, meta);
87 
88 	aux = kzalloc(sizeof(*aux), GFP_KERNEL | gfp_extra_flags);
89 	if (aux == NULL) {
90 		vfree(fp);
91 		return NULL;
92 	}
93 
94 	fp->pages = size / PAGE_SIZE;
95 	fp->aux = aux;
96 	fp->aux->prog = fp;
97 
98 	return fp;
99 }
100 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
101 
102 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
103 				  gfp_t gfp_extra_flags)
104 {
105 	gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
106 			  gfp_extra_flags;
107 	struct bpf_prog *fp;
108 
109 	BUG_ON(fp_old == NULL);
110 
111 	size = round_up(size, PAGE_SIZE);
112 	if (size <= fp_old->pages * PAGE_SIZE)
113 		return fp_old;
114 
115 	fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
116 	if (fp != NULL) {
117 		kmemcheck_annotate_bitfield(fp, meta);
118 
119 		memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
120 		fp->pages = size / PAGE_SIZE;
121 		fp->aux->prog = fp;
122 
123 		/* We keep fp->aux from fp_old around in the new
124 		 * reallocated structure.
125 		 */
126 		fp_old->aux = NULL;
127 		__bpf_prog_free(fp_old);
128 	}
129 
130 	return fp;
131 }
132 
133 void __bpf_prog_free(struct bpf_prog *fp)
134 {
135 	kfree(fp->aux);
136 	vfree(fp);
137 }
138 
139 static bool bpf_is_jmp_and_has_target(const struct bpf_insn *insn)
140 {
141 	return BPF_CLASS(insn->code) == BPF_JMP  &&
142 	       /* Call and Exit are both special jumps with no
143 		* target inside the BPF instruction image.
144 		*/
145 	       BPF_OP(insn->code) != BPF_CALL &&
146 	       BPF_OP(insn->code) != BPF_EXIT;
147 }
148 
149 static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta)
150 {
151 	struct bpf_insn *insn = prog->insnsi;
152 	u32 i, insn_cnt = prog->len;
153 
154 	for (i = 0; i < insn_cnt; i++, insn++) {
155 		if (!bpf_is_jmp_and_has_target(insn))
156 			continue;
157 
158 		/* Adjust offset of jmps if we cross boundaries. */
159 		if (i < pos && i + insn->off + 1 > pos)
160 			insn->off += delta;
161 		else if (i > pos + delta && i + insn->off + 1 <= pos + delta)
162 			insn->off -= delta;
163 	}
164 }
165 
166 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
167 				       const struct bpf_insn *patch, u32 len)
168 {
169 	u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
170 	struct bpf_prog *prog_adj;
171 
172 	/* Since our patchlet doesn't expand the image, we're done. */
173 	if (insn_delta == 0) {
174 		memcpy(prog->insnsi + off, patch, sizeof(*patch));
175 		return prog;
176 	}
177 
178 	insn_adj_cnt = prog->len + insn_delta;
179 
180 	/* Several new instructions need to be inserted. Make room
181 	 * for them. Likely, there's no need for a new allocation as
182 	 * last page could have large enough tailroom.
183 	 */
184 	prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
185 				    GFP_USER);
186 	if (!prog_adj)
187 		return NULL;
188 
189 	prog_adj->len = insn_adj_cnt;
190 
191 	/* Patching happens in 3 steps:
192 	 *
193 	 * 1) Move over tail of insnsi from next instruction onwards,
194 	 *    so we can patch the single target insn with one or more
195 	 *    new ones (patching is always from 1 to n insns, n > 0).
196 	 * 2) Inject new instructions at the target location.
197 	 * 3) Adjust branch offsets if necessary.
198 	 */
199 	insn_rest = insn_adj_cnt - off - len;
200 
201 	memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
202 		sizeof(*patch) * insn_rest);
203 	memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
204 
205 	bpf_adj_branches(prog_adj, off, insn_delta);
206 
207 	return prog_adj;
208 }
209 
210 #ifdef CONFIG_BPF_JIT
211 struct bpf_binary_header *
212 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
213 		     unsigned int alignment,
214 		     bpf_jit_fill_hole_t bpf_fill_ill_insns)
215 {
216 	struct bpf_binary_header *hdr;
217 	unsigned int size, hole, start;
218 
219 	/* Most of BPF filters are really small, but if some of them
220 	 * fill a page, allow at least 128 extra bytes to insert a
221 	 * random section of illegal instructions.
222 	 */
223 	size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
224 	hdr = module_alloc(size);
225 	if (hdr == NULL)
226 		return NULL;
227 
228 	/* Fill space with illegal/arch-dep instructions. */
229 	bpf_fill_ill_insns(hdr, size);
230 
231 	hdr->pages = size / PAGE_SIZE;
232 	hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
233 		     PAGE_SIZE - sizeof(*hdr));
234 	start = (get_random_int() % hole) & ~(alignment - 1);
235 
236 	/* Leave a random number of instructions before BPF code. */
237 	*image_ptr = &hdr->image[start];
238 
239 	return hdr;
240 }
241 
242 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
243 {
244 	module_memfree(hdr);
245 }
246 
247 int bpf_jit_harden __read_mostly;
248 
249 static int bpf_jit_blind_insn(const struct bpf_insn *from,
250 			      const struct bpf_insn *aux,
251 			      struct bpf_insn *to_buff)
252 {
253 	struct bpf_insn *to = to_buff;
254 	u32 imm_rnd = get_random_int();
255 	s16 off;
256 
257 	BUILD_BUG_ON(BPF_REG_AX  + 1 != MAX_BPF_JIT_REG);
258 	BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
259 
260 	if (from->imm == 0 &&
261 	    (from->code == (BPF_ALU   | BPF_MOV | BPF_K) ||
262 	     from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
263 		*to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
264 		goto out;
265 	}
266 
267 	switch (from->code) {
268 	case BPF_ALU | BPF_ADD | BPF_K:
269 	case BPF_ALU | BPF_SUB | BPF_K:
270 	case BPF_ALU | BPF_AND | BPF_K:
271 	case BPF_ALU | BPF_OR  | BPF_K:
272 	case BPF_ALU | BPF_XOR | BPF_K:
273 	case BPF_ALU | BPF_MUL | BPF_K:
274 	case BPF_ALU | BPF_MOV | BPF_K:
275 	case BPF_ALU | BPF_DIV | BPF_K:
276 	case BPF_ALU | BPF_MOD | BPF_K:
277 		*to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
278 		*to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
279 		*to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
280 		break;
281 
282 	case BPF_ALU64 | BPF_ADD | BPF_K:
283 	case BPF_ALU64 | BPF_SUB | BPF_K:
284 	case BPF_ALU64 | BPF_AND | BPF_K:
285 	case BPF_ALU64 | BPF_OR  | BPF_K:
286 	case BPF_ALU64 | BPF_XOR | BPF_K:
287 	case BPF_ALU64 | BPF_MUL | BPF_K:
288 	case BPF_ALU64 | BPF_MOV | BPF_K:
289 	case BPF_ALU64 | BPF_DIV | BPF_K:
290 	case BPF_ALU64 | BPF_MOD | BPF_K:
291 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
292 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
293 		*to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
294 		break;
295 
296 	case BPF_JMP | BPF_JEQ  | BPF_K:
297 	case BPF_JMP | BPF_JNE  | BPF_K:
298 	case BPF_JMP | BPF_JGT  | BPF_K:
299 	case BPF_JMP | BPF_JGE  | BPF_K:
300 	case BPF_JMP | BPF_JSGT | BPF_K:
301 	case BPF_JMP | BPF_JSGE | BPF_K:
302 	case BPF_JMP | BPF_JSET | BPF_K:
303 		/* Accommodate for extra offset in case of a backjump. */
304 		off = from->off;
305 		if (off < 0)
306 			off -= 2;
307 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
308 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
309 		*to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
310 		break;
311 
312 	case BPF_LD | BPF_ABS | BPF_W:
313 	case BPF_LD | BPF_ABS | BPF_H:
314 	case BPF_LD | BPF_ABS | BPF_B:
315 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
316 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
317 		*to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
318 		break;
319 
320 	case BPF_LD | BPF_IND | BPF_W:
321 	case BPF_LD | BPF_IND | BPF_H:
322 	case BPF_LD | BPF_IND | BPF_B:
323 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
324 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
325 		*to++ = BPF_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg);
326 		*to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
327 		break;
328 
329 	case BPF_LD | BPF_IMM | BPF_DW:
330 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
331 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
332 		*to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
333 		*to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
334 		break;
335 	case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
336 		*to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
337 		*to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
338 		*to++ = BPF_ALU64_REG(BPF_OR,  aux[0].dst_reg, BPF_REG_AX);
339 		break;
340 
341 	case BPF_ST | BPF_MEM | BPF_DW:
342 	case BPF_ST | BPF_MEM | BPF_W:
343 	case BPF_ST | BPF_MEM | BPF_H:
344 	case BPF_ST | BPF_MEM | BPF_B:
345 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
346 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
347 		*to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
348 		break;
349 	}
350 out:
351 	return to - to_buff;
352 }
353 
354 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
355 					      gfp_t gfp_extra_flags)
356 {
357 	gfp_t gfp_flags = GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO |
358 			  gfp_extra_flags;
359 	struct bpf_prog *fp;
360 
361 	fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL);
362 	if (fp != NULL) {
363 		kmemcheck_annotate_bitfield(fp, meta);
364 
365 		/* aux->prog still points to the fp_other one, so
366 		 * when promoting the clone to the real program,
367 		 * this still needs to be adapted.
368 		 */
369 		memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
370 	}
371 
372 	return fp;
373 }
374 
375 static void bpf_prog_clone_free(struct bpf_prog *fp)
376 {
377 	/* aux was stolen by the other clone, so we cannot free
378 	 * it from this path! It will be freed eventually by the
379 	 * other program on release.
380 	 *
381 	 * At this point, we don't need a deferred release since
382 	 * clone is guaranteed to not be locked.
383 	 */
384 	fp->aux = NULL;
385 	__bpf_prog_free(fp);
386 }
387 
388 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
389 {
390 	/* We have to repoint aux->prog to self, as we don't
391 	 * know whether fp here is the clone or the original.
392 	 */
393 	fp->aux->prog = fp;
394 	bpf_prog_clone_free(fp_other);
395 }
396 
397 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
398 {
399 	struct bpf_insn insn_buff[16], aux[2];
400 	struct bpf_prog *clone, *tmp;
401 	int insn_delta, insn_cnt;
402 	struct bpf_insn *insn;
403 	int i, rewritten;
404 
405 	if (!bpf_jit_blinding_enabled())
406 		return prog;
407 
408 	clone = bpf_prog_clone_create(prog, GFP_USER);
409 	if (!clone)
410 		return ERR_PTR(-ENOMEM);
411 
412 	insn_cnt = clone->len;
413 	insn = clone->insnsi;
414 
415 	for (i = 0; i < insn_cnt; i++, insn++) {
416 		/* We temporarily need to hold the original ld64 insn
417 		 * so that we can still access the first part in the
418 		 * second blinding run.
419 		 */
420 		if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
421 		    insn[1].code == 0)
422 			memcpy(aux, insn, sizeof(aux));
423 
424 		rewritten = bpf_jit_blind_insn(insn, aux, insn_buff);
425 		if (!rewritten)
426 			continue;
427 
428 		tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
429 		if (!tmp) {
430 			/* Patching may have repointed aux->prog during
431 			 * realloc from the original one, so we need to
432 			 * fix it up here on error.
433 			 */
434 			bpf_jit_prog_release_other(prog, clone);
435 			return ERR_PTR(-ENOMEM);
436 		}
437 
438 		clone = tmp;
439 		insn_delta = rewritten - 1;
440 
441 		/* Walk new program and skip insns we just inserted. */
442 		insn = clone->insnsi + i + insn_delta;
443 		insn_cnt += insn_delta;
444 		i        += insn_delta;
445 	}
446 
447 	return clone;
448 }
449 #endif /* CONFIG_BPF_JIT */
450 
451 /* Base function for offset calculation. Needs to go into .text section,
452  * therefore keeping it non-static as well; will also be used by JITs
453  * anyway later on, so do not let the compiler omit it.
454  */
455 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
456 {
457 	return 0;
458 }
459 EXPORT_SYMBOL_GPL(__bpf_call_base);
460 
461 /**
462  *	__bpf_prog_run - run eBPF program on a given context
463  *	@ctx: is the data we are operating on
464  *	@insn: is the array of eBPF instructions
465  *
466  * Decode and execute eBPF instructions.
467  */
468 static unsigned int __bpf_prog_run(void *ctx, const struct bpf_insn *insn)
469 {
470 	u64 stack[MAX_BPF_STACK / sizeof(u64)];
471 	u64 regs[MAX_BPF_REG], tmp;
472 	static const void *jumptable[256] = {
473 		[0 ... 255] = &&default_label,
474 		/* Now overwrite non-defaults ... */
475 		/* 32 bit ALU operations */
476 		[BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
477 		[BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
478 		[BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
479 		[BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
480 		[BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
481 		[BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
482 		[BPF_ALU | BPF_OR | BPF_X]  = &&ALU_OR_X,
483 		[BPF_ALU | BPF_OR | BPF_K]  = &&ALU_OR_K,
484 		[BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
485 		[BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
486 		[BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
487 		[BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
488 		[BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
489 		[BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
490 		[BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
491 		[BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
492 		[BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
493 		[BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
494 		[BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
495 		[BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
496 		[BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
497 		[BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
498 		[BPF_ALU | BPF_NEG] = &&ALU_NEG,
499 		[BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
500 		[BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
501 		/* 64 bit ALU operations */
502 		[BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
503 		[BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
504 		[BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
505 		[BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
506 		[BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
507 		[BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
508 		[BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
509 		[BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
510 		[BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
511 		[BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
512 		[BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
513 		[BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
514 		[BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
515 		[BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
516 		[BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
517 		[BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
518 		[BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
519 		[BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
520 		[BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
521 		[BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
522 		[BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
523 		[BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
524 		[BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
525 		[BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
526 		[BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
527 		/* Call instruction */
528 		[BPF_JMP | BPF_CALL] = &&JMP_CALL,
529 		[BPF_JMP | BPF_CALL | BPF_X] = &&JMP_TAIL_CALL,
530 		/* Jumps */
531 		[BPF_JMP | BPF_JA] = &&JMP_JA,
532 		[BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
533 		[BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
534 		[BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
535 		[BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
536 		[BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
537 		[BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
538 		[BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
539 		[BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
540 		[BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
541 		[BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
542 		[BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
543 		[BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
544 		[BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
545 		[BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
546 		/* Program return */
547 		[BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
548 		/* Store instructions */
549 		[BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
550 		[BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
551 		[BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
552 		[BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
553 		[BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
554 		[BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
555 		[BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
556 		[BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
557 		[BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
558 		[BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
559 		/* Load instructions */
560 		[BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
561 		[BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
562 		[BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
563 		[BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
564 		[BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
565 		[BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
566 		[BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
567 		[BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
568 		[BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
569 		[BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
570 		[BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
571 	};
572 	u32 tail_call_cnt = 0;
573 	void *ptr;
574 	int off;
575 
576 #define CONT	 ({ insn++; goto select_insn; })
577 #define CONT_JMP ({ insn++; goto select_insn; })
578 
579 	FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)];
580 	ARG1 = (u64) (unsigned long) ctx;
581 
582 select_insn:
583 	goto *jumptable[insn->code];
584 
585 	/* ALU */
586 #define ALU(OPCODE, OP)			\
587 	ALU64_##OPCODE##_X:		\
588 		DST = DST OP SRC;	\
589 		CONT;			\
590 	ALU_##OPCODE##_X:		\
591 		DST = (u32) DST OP (u32) SRC;	\
592 		CONT;			\
593 	ALU64_##OPCODE##_K:		\
594 		DST = DST OP IMM;		\
595 		CONT;			\
596 	ALU_##OPCODE##_K:		\
597 		DST = (u32) DST OP (u32) IMM;	\
598 		CONT;
599 
600 	ALU(ADD,  +)
601 	ALU(SUB,  -)
602 	ALU(AND,  &)
603 	ALU(OR,   |)
604 	ALU(LSH, <<)
605 	ALU(RSH, >>)
606 	ALU(XOR,  ^)
607 	ALU(MUL,  *)
608 #undef ALU
609 	ALU_NEG:
610 		DST = (u32) -DST;
611 		CONT;
612 	ALU64_NEG:
613 		DST = -DST;
614 		CONT;
615 	ALU_MOV_X:
616 		DST = (u32) SRC;
617 		CONT;
618 	ALU_MOV_K:
619 		DST = (u32) IMM;
620 		CONT;
621 	ALU64_MOV_X:
622 		DST = SRC;
623 		CONT;
624 	ALU64_MOV_K:
625 		DST = IMM;
626 		CONT;
627 	LD_IMM_DW:
628 		DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
629 		insn++;
630 		CONT;
631 	ALU64_ARSH_X:
632 		(*(s64 *) &DST) >>= SRC;
633 		CONT;
634 	ALU64_ARSH_K:
635 		(*(s64 *) &DST) >>= IMM;
636 		CONT;
637 	ALU64_MOD_X:
638 		if (unlikely(SRC == 0))
639 			return 0;
640 		div64_u64_rem(DST, SRC, &tmp);
641 		DST = tmp;
642 		CONT;
643 	ALU_MOD_X:
644 		if (unlikely(SRC == 0))
645 			return 0;
646 		tmp = (u32) DST;
647 		DST = do_div(tmp, (u32) SRC);
648 		CONT;
649 	ALU64_MOD_K:
650 		div64_u64_rem(DST, IMM, &tmp);
651 		DST = tmp;
652 		CONT;
653 	ALU_MOD_K:
654 		tmp = (u32) DST;
655 		DST = do_div(tmp, (u32) IMM);
656 		CONT;
657 	ALU64_DIV_X:
658 		if (unlikely(SRC == 0))
659 			return 0;
660 		DST = div64_u64(DST, SRC);
661 		CONT;
662 	ALU_DIV_X:
663 		if (unlikely(SRC == 0))
664 			return 0;
665 		tmp = (u32) DST;
666 		do_div(tmp, (u32) SRC);
667 		DST = (u32) tmp;
668 		CONT;
669 	ALU64_DIV_K:
670 		DST = div64_u64(DST, IMM);
671 		CONT;
672 	ALU_DIV_K:
673 		tmp = (u32) DST;
674 		do_div(tmp, (u32) IMM);
675 		DST = (u32) tmp;
676 		CONT;
677 	ALU_END_TO_BE:
678 		switch (IMM) {
679 		case 16:
680 			DST = (__force u16) cpu_to_be16(DST);
681 			break;
682 		case 32:
683 			DST = (__force u32) cpu_to_be32(DST);
684 			break;
685 		case 64:
686 			DST = (__force u64) cpu_to_be64(DST);
687 			break;
688 		}
689 		CONT;
690 	ALU_END_TO_LE:
691 		switch (IMM) {
692 		case 16:
693 			DST = (__force u16) cpu_to_le16(DST);
694 			break;
695 		case 32:
696 			DST = (__force u32) cpu_to_le32(DST);
697 			break;
698 		case 64:
699 			DST = (__force u64) cpu_to_le64(DST);
700 			break;
701 		}
702 		CONT;
703 
704 	/* CALL */
705 	JMP_CALL:
706 		/* Function call scratches BPF_R1-BPF_R5 registers,
707 		 * preserves BPF_R6-BPF_R9, and stores return value
708 		 * into BPF_R0.
709 		 */
710 		BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
711 						       BPF_R4, BPF_R5);
712 		CONT;
713 
714 	JMP_TAIL_CALL: {
715 		struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
716 		struct bpf_array *array = container_of(map, struct bpf_array, map);
717 		struct bpf_prog *prog;
718 		u64 index = BPF_R3;
719 
720 		if (unlikely(index >= array->map.max_entries))
721 			goto out;
722 		if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
723 			goto out;
724 
725 		tail_call_cnt++;
726 
727 		prog = READ_ONCE(array->ptrs[index]);
728 		if (!prog)
729 			goto out;
730 
731 		/* ARG1 at this point is guaranteed to point to CTX from
732 		 * the verifier side due to the fact that the tail call is
733 		 * handeled like a helper, that is, bpf_tail_call_proto,
734 		 * where arg1_type is ARG_PTR_TO_CTX.
735 		 */
736 		insn = prog->insnsi;
737 		goto select_insn;
738 out:
739 		CONT;
740 	}
741 	/* JMP */
742 	JMP_JA:
743 		insn += insn->off;
744 		CONT;
745 	JMP_JEQ_X:
746 		if (DST == SRC) {
747 			insn += insn->off;
748 			CONT_JMP;
749 		}
750 		CONT;
751 	JMP_JEQ_K:
752 		if (DST == IMM) {
753 			insn += insn->off;
754 			CONT_JMP;
755 		}
756 		CONT;
757 	JMP_JNE_X:
758 		if (DST != SRC) {
759 			insn += insn->off;
760 			CONT_JMP;
761 		}
762 		CONT;
763 	JMP_JNE_K:
764 		if (DST != IMM) {
765 			insn += insn->off;
766 			CONT_JMP;
767 		}
768 		CONT;
769 	JMP_JGT_X:
770 		if (DST > SRC) {
771 			insn += insn->off;
772 			CONT_JMP;
773 		}
774 		CONT;
775 	JMP_JGT_K:
776 		if (DST > IMM) {
777 			insn += insn->off;
778 			CONT_JMP;
779 		}
780 		CONT;
781 	JMP_JGE_X:
782 		if (DST >= SRC) {
783 			insn += insn->off;
784 			CONT_JMP;
785 		}
786 		CONT;
787 	JMP_JGE_K:
788 		if (DST >= IMM) {
789 			insn += insn->off;
790 			CONT_JMP;
791 		}
792 		CONT;
793 	JMP_JSGT_X:
794 		if (((s64) DST) > ((s64) SRC)) {
795 			insn += insn->off;
796 			CONT_JMP;
797 		}
798 		CONT;
799 	JMP_JSGT_K:
800 		if (((s64) DST) > ((s64) IMM)) {
801 			insn += insn->off;
802 			CONT_JMP;
803 		}
804 		CONT;
805 	JMP_JSGE_X:
806 		if (((s64) DST) >= ((s64) SRC)) {
807 			insn += insn->off;
808 			CONT_JMP;
809 		}
810 		CONT;
811 	JMP_JSGE_K:
812 		if (((s64) DST) >= ((s64) IMM)) {
813 			insn += insn->off;
814 			CONT_JMP;
815 		}
816 		CONT;
817 	JMP_JSET_X:
818 		if (DST & SRC) {
819 			insn += insn->off;
820 			CONT_JMP;
821 		}
822 		CONT;
823 	JMP_JSET_K:
824 		if (DST & IMM) {
825 			insn += insn->off;
826 			CONT_JMP;
827 		}
828 		CONT;
829 	JMP_EXIT:
830 		return BPF_R0;
831 
832 	/* STX and ST and LDX*/
833 #define LDST(SIZEOP, SIZE)						\
834 	STX_MEM_##SIZEOP:						\
835 		*(SIZE *)(unsigned long) (DST + insn->off) = SRC;	\
836 		CONT;							\
837 	ST_MEM_##SIZEOP:						\
838 		*(SIZE *)(unsigned long) (DST + insn->off) = IMM;	\
839 		CONT;							\
840 	LDX_MEM_##SIZEOP:						\
841 		DST = *(SIZE *)(unsigned long) (SRC + insn->off);	\
842 		CONT;
843 
844 	LDST(B,   u8)
845 	LDST(H,  u16)
846 	LDST(W,  u32)
847 	LDST(DW, u64)
848 #undef LDST
849 	STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
850 		atomic_add((u32) SRC, (atomic_t *)(unsigned long)
851 			   (DST + insn->off));
852 		CONT;
853 	STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
854 		atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
855 			     (DST + insn->off));
856 		CONT;
857 	LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
858 		off = IMM;
859 load_word:
860 		/* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are
861 		 * only appearing in the programs where ctx ==
862 		 * skb. All programs keep 'ctx' in regs[BPF_REG_CTX]
863 		 * == BPF_R6, bpf_convert_filter() saves it in BPF_R6,
864 		 * internal BPF verifier will check that BPF_R6 ==
865 		 * ctx.
866 		 *
867 		 * BPF_ABS and BPF_IND are wrappers of function calls,
868 		 * so they scratch BPF_R1-BPF_R5 registers, preserve
869 		 * BPF_R6-BPF_R9, and store return value into BPF_R0.
870 		 *
871 		 * Implicit input:
872 		 *   ctx == skb == BPF_R6 == CTX
873 		 *
874 		 * Explicit input:
875 		 *   SRC == any register
876 		 *   IMM == 32-bit immediate
877 		 *
878 		 * Output:
879 		 *   BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
880 		 */
881 
882 		ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
883 		if (likely(ptr != NULL)) {
884 			BPF_R0 = get_unaligned_be32(ptr);
885 			CONT;
886 		}
887 
888 		return 0;
889 	LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
890 		off = IMM;
891 load_half:
892 		ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
893 		if (likely(ptr != NULL)) {
894 			BPF_R0 = get_unaligned_be16(ptr);
895 			CONT;
896 		}
897 
898 		return 0;
899 	LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
900 		off = IMM;
901 load_byte:
902 		ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
903 		if (likely(ptr != NULL)) {
904 			BPF_R0 = *(u8 *)ptr;
905 			CONT;
906 		}
907 
908 		return 0;
909 	LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
910 		off = IMM + SRC;
911 		goto load_word;
912 	LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
913 		off = IMM + SRC;
914 		goto load_half;
915 	LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
916 		off = IMM + SRC;
917 		goto load_byte;
918 
919 	default_label:
920 		/* If we ever reach this, we have a bug somewhere. */
921 		WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
922 		return 0;
923 }
924 STACK_FRAME_NON_STANDARD(__bpf_prog_run); /* jump table */
925 
926 bool bpf_prog_array_compatible(struct bpf_array *array,
927 			       const struct bpf_prog *fp)
928 {
929 	if (!array->owner_prog_type) {
930 		/* There's no owner yet where we could check for
931 		 * compatibility.
932 		 */
933 		array->owner_prog_type = fp->type;
934 		array->owner_jited = fp->jited;
935 
936 		return true;
937 	}
938 
939 	return array->owner_prog_type == fp->type &&
940 	       array->owner_jited == fp->jited;
941 }
942 
943 static int bpf_check_tail_call(const struct bpf_prog *fp)
944 {
945 	struct bpf_prog_aux *aux = fp->aux;
946 	int i;
947 
948 	for (i = 0; i < aux->used_map_cnt; i++) {
949 		struct bpf_map *map = aux->used_maps[i];
950 		struct bpf_array *array;
951 
952 		if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
953 			continue;
954 
955 		array = container_of(map, struct bpf_array, map);
956 		if (!bpf_prog_array_compatible(array, fp))
957 			return -EINVAL;
958 	}
959 
960 	return 0;
961 }
962 
963 /**
964  *	bpf_prog_select_runtime - select exec runtime for BPF program
965  *	@fp: bpf_prog populated with internal BPF program
966  *	@err: pointer to error variable
967  *
968  * Try to JIT eBPF program, if JIT is not available, use interpreter.
969  * The BPF program will be executed via BPF_PROG_RUN() macro.
970  */
971 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
972 {
973 	fp->bpf_func = (void *) __bpf_prog_run;
974 
975 	/* eBPF JITs can rewrite the program in case constant
976 	 * blinding is active. However, in case of error during
977 	 * blinding, bpf_int_jit_compile() must always return a
978 	 * valid program, which in this case would simply not
979 	 * be JITed, but falls back to the interpreter.
980 	 */
981 	fp = bpf_int_jit_compile(fp);
982 	bpf_prog_lock_ro(fp);
983 
984 	/* The tail call compatibility check can only be done at
985 	 * this late stage as we need to determine, if we deal
986 	 * with JITed or non JITed program concatenations and not
987 	 * all eBPF JITs might immediately support all features.
988 	 */
989 	*err = bpf_check_tail_call(fp);
990 
991 	return fp;
992 }
993 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
994 
995 static void bpf_prog_free_deferred(struct work_struct *work)
996 {
997 	struct bpf_prog_aux *aux;
998 
999 	aux = container_of(work, struct bpf_prog_aux, work);
1000 	bpf_jit_free(aux->prog);
1001 }
1002 
1003 /* Free internal BPF program */
1004 void bpf_prog_free(struct bpf_prog *fp)
1005 {
1006 	struct bpf_prog_aux *aux = fp->aux;
1007 
1008 	INIT_WORK(&aux->work, bpf_prog_free_deferred);
1009 	schedule_work(&aux->work);
1010 }
1011 EXPORT_SYMBOL_GPL(bpf_prog_free);
1012 
1013 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
1014 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
1015 
1016 void bpf_user_rnd_init_once(void)
1017 {
1018 	prandom_init_once(&bpf_user_rnd_state);
1019 }
1020 
1021 u64 bpf_user_rnd_u32(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
1022 {
1023 	/* Should someone ever have the rather unwise idea to use some
1024 	 * of the registers passed into this function, then note that
1025 	 * this function is called from native eBPF and classic-to-eBPF
1026 	 * transformations. Register assignments from both sides are
1027 	 * different, f.e. classic always sets fn(ctx, A, X) here.
1028 	 */
1029 	struct rnd_state *state;
1030 	u32 res;
1031 
1032 	state = &get_cpu_var(bpf_user_rnd_state);
1033 	res = prandom_u32_state(state);
1034 	put_cpu_var(state);
1035 
1036 	return res;
1037 }
1038 
1039 /* Weak definitions of helper functions in case we don't have bpf syscall. */
1040 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
1041 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
1042 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
1043 
1044 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
1045 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
1046 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
1047 
1048 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
1049 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
1050 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
1051 
1052 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
1053 {
1054 	return NULL;
1055 }
1056 
1057 u64 __weak
1058 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
1059 		 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
1060 {
1061 	return -ENOTSUPP;
1062 }
1063 
1064 /* Always built-in helper functions. */
1065 const struct bpf_func_proto bpf_tail_call_proto = {
1066 	.func		= NULL,
1067 	.gpl_only	= false,
1068 	.ret_type	= RET_VOID,
1069 	.arg1_type	= ARG_PTR_TO_CTX,
1070 	.arg2_type	= ARG_CONST_MAP_PTR,
1071 	.arg3_type	= ARG_ANYTHING,
1072 };
1073 
1074 /* For classic BPF JITs that don't implement bpf_int_jit_compile(). */
1075 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
1076 {
1077 	return prog;
1078 }
1079 
1080 bool __weak bpf_helper_changes_skb_data(void *func)
1081 {
1082 	return false;
1083 }
1084 
1085 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
1086  * skb_copy_bits(), so provide a weak definition of it for NET-less config.
1087  */
1088 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
1089 			 int len)
1090 {
1091 	return -EFAULT;
1092 }
1093