xref: /linux/kernel/bpf/core.c (revision 4950276672fce5c241857540f8561c440663673d)
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 #include <linux/rbtree_latch.h>
32 #include <linux/kallsyms.h>
33 #include <linux/rcupdate.h>
34 
35 #include <asm/unaligned.h>
36 
37 /* Registers */
38 #define BPF_R0	regs[BPF_REG_0]
39 #define BPF_R1	regs[BPF_REG_1]
40 #define BPF_R2	regs[BPF_REG_2]
41 #define BPF_R3	regs[BPF_REG_3]
42 #define BPF_R4	regs[BPF_REG_4]
43 #define BPF_R5	regs[BPF_REG_5]
44 #define BPF_R6	regs[BPF_REG_6]
45 #define BPF_R7	regs[BPF_REG_7]
46 #define BPF_R8	regs[BPF_REG_8]
47 #define BPF_R9	regs[BPF_REG_9]
48 #define BPF_R10	regs[BPF_REG_10]
49 
50 /* Named registers */
51 #define DST	regs[insn->dst_reg]
52 #define SRC	regs[insn->src_reg]
53 #define FP	regs[BPF_REG_FP]
54 #define ARG1	regs[BPF_REG_ARG1]
55 #define CTX	regs[BPF_REG_CTX]
56 #define IMM	insn->imm
57 
58 /* No hurry in this branch
59  *
60  * Exported for the bpf jit load helper.
61  */
62 void *bpf_internal_load_pointer_neg_helper(const struct sk_buff *skb, int k, unsigned int size)
63 {
64 	u8 *ptr = NULL;
65 
66 	if (k >= SKF_NET_OFF)
67 		ptr = skb_network_header(skb) + k - SKF_NET_OFF;
68 	else if (k >= SKF_LL_OFF)
69 		ptr = skb_mac_header(skb) + k - SKF_LL_OFF;
70 
71 	if (ptr >= skb->head && ptr + size <= skb_tail_pointer(skb))
72 		return ptr;
73 
74 	return NULL;
75 }
76 
77 struct bpf_prog *bpf_prog_alloc(unsigned int size, gfp_t gfp_extra_flags)
78 {
79 	gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
80 	struct bpf_prog_aux *aux;
81 	struct bpf_prog *fp;
82 
83 	size = round_up(size, PAGE_SIZE);
84 	fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
85 	if (fp == NULL)
86 		return NULL;
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 	INIT_LIST_HEAD_RCU(&fp->aux->ksym_lnode);
99 
100 	return fp;
101 }
102 EXPORT_SYMBOL_GPL(bpf_prog_alloc);
103 
104 struct bpf_prog *bpf_prog_realloc(struct bpf_prog *fp_old, unsigned int size,
105 				  gfp_t gfp_extra_flags)
106 {
107 	gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
108 	struct bpf_prog *fp;
109 	u32 pages, delta;
110 	int ret;
111 
112 	BUG_ON(fp_old == NULL);
113 
114 	size = round_up(size, PAGE_SIZE);
115 	pages = size / PAGE_SIZE;
116 	if (pages <= fp_old->pages)
117 		return fp_old;
118 
119 	delta = pages - fp_old->pages;
120 	ret = __bpf_prog_charge(fp_old->aux->user, delta);
121 	if (ret)
122 		return NULL;
123 
124 	fp = __vmalloc(size, gfp_flags, PAGE_KERNEL);
125 	if (fp == NULL) {
126 		__bpf_prog_uncharge(fp_old->aux->user, delta);
127 	} else {
128 		memcpy(fp, fp_old, fp_old->pages * PAGE_SIZE);
129 		fp->pages = pages;
130 		fp->aux->prog = fp;
131 
132 		/* We keep fp->aux from fp_old around in the new
133 		 * reallocated structure.
134 		 */
135 		fp_old->aux = NULL;
136 		__bpf_prog_free(fp_old);
137 	}
138 
139 	return fp;
140 }
141 
142 void __bpf_prog_free(struct bpf_prog *fp)
143 {
144 	kfree(fp->aux);
145 	vfree(fp);
146 }
147 
148 int bpf_prog_calc_tag(struct bpf_prog *fp)
149 {
150 	const u32 bits_offset = SHA_MESSAGE_BYTES - sizeof(__be64);
151 	u32 raw_size = bpf_prog_tag_scratch_size(fp);
152 	u32 digest[SHA_DIGEST_WORDS];
153 	u32 ws[SHA_WORKSPACE_WORDS];
154 	u32 i, bsize, psize, blocks;
155 	struct bpf_insn *dst;
156 	bool was_ld_map;
157 	u8 *raw, *todo;
158 	__be32 *result;
159 	__be64 *bits;
160 
161 	raw = vmalloc(raw_size);
162 	if (!raw)
163 		return -ENOMEM;
164 
165 	sha_init(digest);
166 	memset(ws, 0, sizeof(ws));
167 
168 	/* We need to take out the map fd for the digest calculation
169 	 * since they are unstable from user space side.
170 	 */
171 	dst = (void *)raw;
172 	for (i = 0, was_ld_map = false; i < fp->len; i++) {
173 		dst[i] = fp->insnsi[i];
174 		if (!was_ld_map &&
175 		    dst[i].code == (BPF_LD | BPF_IMM | BPF_DW) &&
176 		    dst[i].src_reg == BPF_PSEUDO_MAP_FD) {
177 			was_ld_map = true;
178 			dst[i].imm = 0;
179 		} else if (was_ld_map &&
180 			   dst[i].code == 0 &&
181 			   dst[i].dst_reg == 0 &&
182 			   dst[i].src_reg == 0 &&
183 			   dst[i].off == 0) {
184 			was_ld_map = false;
185 			dst[i].imm = 0;
186 		} else {
187 			was_ld_map = false;
188 		}
189 	}
190 
191 	psize = bpf_prog_insn_size(fp);
192 	memset(&raw[psize], 0, raw_size - psize);
193 	raw[psize++] = 0x80;
194 
195 	bsize  = round_up(psize, SHA_MESSAGE_BYTES);
196 	blocks = bsize / SHA_MESSAGE_BYTES;
197 	todo   = raw;
198 	if (bsize - psize >= sizeof(__be64)) {
199 		bits = (__be64 *)(todo + bsize - sizeof(__be64));
200 	} else {
201 		bits = (__be64 *)(todo + bsize + bits_offset);
202 		blocks++;
203 	}
204 	*bits = cpu_to_be64((psize - 1) << 3);
205 
206 	while (blocks--) {
207 		sha_transform(digest, todo, ws);
208 		todo += SHA_MESSAGE_BYTES;
209 	}
210 
211 	result = (__force __be32 *)digest;
212 	for (i = 0; i < SHA_DIGEST_WORDS; i++)
213 		result[i] = cpu_to_be32(digest[i]);
214 	memcpy(fp->tag, result, sizeof(fp->tag));
215 
216 	vfree(raw);
217 	return 0;
218 }
219 
220 static bool bpf_is_jmp_and_has_target(const struct bpf_insn *insn)
221 {
222 	return BPF_CLASS(insn->code) == BPF_JMP  &&
223 	       /* Call and Exit are both special jumps with no
224 		* target inside the BPF instruction image.
225 		*/
226 	       BPF_OP(insn->code) != BPF_CALL &&
227 	       BPF_OP(insn->code) != BPF_EXIT;
228 }
229 
230 static void bpf_adj_branches(struct bpf_prog *prog, u32 pos, u32 delta)
231 {
232 	struct bpf_insn *insn = prog->insnsi;
233 	u32 i, insn_cnt = prog->len;
234 
235 	for (i = 0; i < insn_cnt; i++, insn++) {
236 		if (!bpf_is_jmp_and_has_target(insn))
237 			continue;
238 
239 		/* Adjust offset of jmps if we cross boundaries. */
240 		if (i < pos && i + insn->off + 1 > pos)
241 			insn->off += delta;
242 		else if (i > pos + delta && i + insn->off + 1 <= pos + delta)
243 			insn->off -= delta;
244 	}
245 }
246 
247 struct bpf_prog *bpf_patch_insn_single(struct bpf_prog *prog, u32 off,
248 				       const struct bpf_insn *patch, u32 len)
249 {
250 	u32 insn_adj_cnt, insn_rest, insn_delta = len - 1;
251 	struct bpf_prog *prog_adj;
252 
253 	/* Since our patchlet doesn't expand the image, we're done. */
254 	if (insn_delta == 0) {
255 		memcpy(prog->insnsi + off, patch, sizeof(*patch));
256 		return prog;
257 	}
258 
259 	insn_adj_cnt = prog->len + insn_delta;
260 
261 	/* Several new instructions need to be inserted. Make room
262 	 * for them. Likely, there's no need for a new allocation as
263 	 * last page could have large enough tailroom.
264 	 */
265 	prog_adj = bpf_prog_realloc(prog, bpf_prog_size(insn_adj_cnt),
266 				    GFP_USER);
267 	if (!prog_adj)
268 		return NULL;
269 
270 	prog_adj->len = insn_adj_cnt;
271 
272 	/* Patching happens in 3 steps:
273 	 *
274 	 * 1) Move over tail of insnsi from next instruction onwards,
275 	 *    so we can patch the single target insn with one or more
276 	 *    new ones (patching is always from 1 to n insns, n > 0).
277 	 * 2) Inject new instructions at the target location.
278 	 * 3) Adjust branch offsets if necessary.
279 	 */
280 	insn_rest = insn_adj_cnt - off - len;
281 
282 	memmove(prog_adj->insnsi + off + len, prog_adj->insnsi + off + 1,
283 		sizeof(*patch) * insn_rest);
284 	memcpy(prog_adj->insnsi + off, patch, sizeof(*patch) * len);
285 
286 	bpf_adj_branches(prog_adj, off, insn_delta);
287 
288 	return prog_adj;
289 }
290 
291 #ifdef CONFIG_BPF_JIT
292 static __always_inline void
293 bpf_get_prog_addr_region(const struct bpf_prog *prog,
294 			 unsigned long *symbol_start,
295 			 unsigned long *symbol_end)
296 {
297 	const struct bpf_binary_header *hdr = bpf_jit_binary_hdr(prog);
298 	unsigned long addr = (unsigned long)hdr;
299 
300 	WARN_ON_ONCE(!bpf_prog_ebpf_jited(prog));
301 
302 	*symbol_start = addr;
303 	*symbol_end   = addr + hdr->pages * PAGE_SIZE;
304 }
305 
306 static void bpf_get_prog_name(const struct bpf_prog *prog, char *sym)
307 {
308 	BUILD_BUG_ON(sizeof("bpf_prog_") +
309 		     sizeof(prog->tag) * 2 + 1 > KSYM_NAME_LEN);
310 
311 	sym += snprintf(sym, KSYM_NAME_LEN, "bpf_prog_");
312 	sym  = bin2hex(sym, prog->tag, sizeof(prog->tag));
313 	*sym = 0;
314 }
315 
316 static __always_inline unsigned long
317 bpf_get_prog_addr_start(struct latch_tree_node *n)
318 {
319 	unsigned long symbol_start, symbol_end;
320 	const struct bpf_prog_aux *aux;
321 
322 	aux = container_of(n, struct bpf_prog_aux, ksym_tnode);
323 	bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
324 
325 	return symbol_start;
326 }
327 
328 static __always_inline bool bpf_tree_less(struct latch_tree_node *a,
329 					  struct latch_tree_node *b)
330 {
331 	return bpf_get_prog_addr_start(a) < bpf_get_prog_addr_start(b);
332 }
333 
334 static __always_inline int bpf_tree_comp(void *key, struct latch_tree_node *n)
335 {
336 	unsigned long val = (unsigned long)key;
337 	unsigned long symbol_start, symbol_end;
338 	const struct bpf_prog_aux *aux;
339 
340 	aux = container_of(n, struct bpf_prog_aux, ksym_tnode);
341 	bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
342 
343 	if (val < symbol_start)
344 		return -1;
345 	if (val >= symbol_end)
346 		return  1;
347 
348 	return 0;
349 }
350 
351 static const struct latch_tree_ops bpf_tree_ops = {
352 	.less	= bpf_tree_less,
353 	.comp	= bpf_tree_comp,
354 };
355 
356 static DEFINE_SPINLOCK(bpf_lock);
357 static LIST_HEAD(bpf_kallsyms);
358 static struct latch_tree_root bpf_tree __cacheline_aligned;
359 
360 int bpf_jit_kallsyms __read_mostly;
361 
362 static void bpf_prog_ksym_node_add(struct bpf_prog_aux *aux)
363 {
364 	WARN_ON_ONCE(!list_empty(&aux->ksym_lnode));
365 	list_add_tail_rcu(&aux->ksym_lnode, &bpf_kallsyms);
366 	latch_tree_insert(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
367 }
368 
369 static void bpf_prog_ksym_node_del(struct bpf_prog_aux *aux)
370 {
371 	if (list_empty(&aux->ksym_lnode))
372 		return;
373 
374 	latch_tree_erase(&aux->ksym_tnode, &bpf_tree, &bpf_tree_ops);
375 	list_del_rcu(&aux->ksym_lnode);
376 }
377 
378 static bool bpf_prog_kallsyms_candidate(const struct bpf_prog *fp)
379 {
380 	return fp->jited && !bpf_prog_was_classic(fp);
381 }
382 
383 static bool bpf_prog_kallsyms_verify_off(const struct bpf_prog *fp)
384 {
385 	return list_empty(&fp->aux->ksym_lnode) ||
386 	       fp->aux->ksym_lnode.prev == LIST_POISON2;
387 }
388 
389 void bpf_prog_kallsyms_add(struct bpf_prog *fp)
390 {
391 	if (!bpf_prog_kallsyms_candidate(fp) ||
392 	    !capable(CAP_SYS_ADMIN))
393 		return;
394 
395 	spin_lock_bh(&bpf_lock);
396 	bpf_prog_ksym_node_add(fp->aux);
397 	spin_unlock_bh(&bpf_lock);
398 }
399 
400 void bpf_prog_kallsyms_del(struct bpf_prog *fp)
401 {
402 	if (!bpf_prog_kallsyms_candidate(fp))
403 		return;
404 
405 	spin_lock_bh(&bpf_lock);
406 	bpf_prog_ksym_node_del(fp->aux);
407 	spin_unlock_bh(&bpf_lock);
408 }
409 
410 static struct bpf_prog *bpf_prog_kallsyms_find(unsigned long addr)
411 {
412 	struct latch_tree_node *n;
413 
414 	if (!bpf_jit_kallsyms_enabled())
415 		return NULL;
416 
417 	n = latch_tree_find((void *)addr, &bpf_tree, &bpf_tree_ops);
418 	return n ?
419 	       container_of(n, struct bpf_prog_aux, ksym_tnode)->prog :
420 	       NULL;
421 }
422 
423 const char *__bpf_address_lookup(unsigned long addr, unsigned long *size,
424 				 unsigned long *off, char *sym)
425 {
426 	unsigned long symbol_start, symbol_end;
427 	struct bpf_prog *prog;
428 	char *ret = NULL;
429 
430 	rcu_read_lock();
431 	prog = bpf_prog_kallsyms_find(addr);
432 	if (prog) {
433 		bpf_get_prog_addr_region(prog, &symbol_start, &symbol_end);
434 		bpf_get_prog_name(prog, sym);
435 
436 		ret = sym;
437 		if (size)
438 			*size = symbol_end - symbol_start;
439 		if (off)
440 			*off  = addr - symbol_start;
441 	}
442 	rcu_read_unlock();
443 
444 	return ret;
445 }
446 
447 bool is_bpf_text_address(unsigned long addr)
448 {
449 	bool ret;
450 
451 	rcu_read_lock();
452 	ret = bpf_prog_kallsyms_find(addr) != NULL;
453 	rcu_read_unlock();
454 
455 	return ret;
456 }
457 
458 int bpf_get_kallsym(unsigned int symnum, unsigned long *value, char *type,
459 		    char *sym)
460 {
461 	unsigned long symbol_start, symbol_end;
462 	struct bpf_prog_aux *aux;
463 	unsigned int it = 0;
464 	int ret = -ERANGE;
465 
466 	if (!bpf_jit_kallsyms_enabled())
467 		return ret;
468 
469 	rcu_read_lock();
470 	list_for_each_entry_rcu(aux, &bpf_kallsyms, ksym_lnode) {
471 		if (it++ != symnum)
472 			continue;
473 
474 		bpf_get_prog_addr_region(aux->prog, &symbol_start, &symbol_end);
475 		bpf_get_prog_name(aux->prog, sym);
476 
477 		*value = symbol_start;
478 		*type  = BPF_SYM_ELF_TYPE;
479 
480 		ret = 0;
481 		break;
482 	}
483 	rcu_read_unlock();
484 
485 	return ret;
486 }
487 
488 struct bpf_binary_header *
489 bpf_jit_binary_alloc(unsigned int proglen, u8 **image_ptr,
490 		     unsigned int alignment,
491 		     bpf_jit_fill_hole_t bpf_fill_ill_insns)
492 {
493 	struct bpf_binary_header *hdr;
494 	unsigned int size, hole, start;
495 
496 	/* Most of BPF filters are really small, but if some of them
497 	 * fill a page, allow at least 128 extra bytes to insert a
498 	 * random section of illegal instructions.
499 	 */
500 	size = round_up(proglen + sizeof(*hdr) + 128, PAGE_SIZE);
501 	hdr = module_alloc(size);
502 	if (hdr == NULL)
503 		return NULL;
504 
505 	/* Fill space with illegal/arch-dep instructions. */
506 	bpf_fill_ill_insns(hdr, size);
507 
508 	hdr->pages = size / PAGE_SIZE;
509 	hole = min_t(unsigned int, size - (proglen + sizeof(*hdr)),
510 		     PAGE_SIZE - sizeof(*hdr));
511 	start = (get_random_int() % hole) & ~(alignment - 1);
512 
513 	/* Leave a random number of instructions before BPF code. */
514 	*image_ptr = &hdr->image[start];
515 
516 	return hdr;
517 }
518 
519 void bpf_jit_binary_free(struct bpf_binary_header *hdr)
520 {
521 	module_memfree(hdr);
522 }
523 
524 /* This symbol is only overridden by archs that have different
525  * requirements than the usual eBPF JITs, f.e. when they only
526  * implement cBPF JIT, do not set images read-only, etc.
527  */
528 void __weak bpf_jit_free(struct bpf_prog *fp)
529 {
530 	if (fp->jited) {
531 		struct bpf_binary_header *hdr = bpf_jit_binary_hdr(fp);
532 
533 		bpf_jit_binary_unlock_ro(hdr);
534 		bpf_jit_binary_free(hdr);
535 
536 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(fp));
537 	}
538 
539 	bpf_prog_unlock_free(fp);
540 }
541 
542 int bpf_jit_harden __read_mostly;
543 
544 static int bpf_jit_blind_insn(const struct bpf_insn *from,
545 			      const struct bpf_insn *aux,
546 			      struct bpf_insn *to_buff)
547 {
548 	struct bpf_insn *to = to_buff;
549 	u32 imm_rnd = get_random_int();
550 	s16 off;
551 
552 	BUILD_BUG_ON(BPF_REG_AX  + 1 != MAX_BPF_JIT_REG);
553 	BUILD_BUG_ON(MAX_BPF_REG + 1 != MAX_BPF_JIT_REG);
554 
555 	if (from->imm == 0 &&
556 	    (from->code == (BPF_ALU   | BPF_MOV | BPF_K) ||
557 	     from->code == (BPF_ALU64 | BPF_MOV | BPF_K))) {
558 		*to++ = BPF_ALU64_REG(BPF_XOR, from->dst_reg, from->dst_reg);
559 		goto out;
560 	}
561 
562 	switch (from->code) {
563 	case BPF_ALU | BPF_ADD | BPF_K:
564 	case BPF_ALU | BPF_SUB | BPF_K:
565 	case BPF_ALU | BPF_AND | BPF_K:
566 	case BPF_ALU | BPF_OR  | BPF_K:
567 	case BPF_ALU | BPF_XOR | BPF_K:
568 	case BPF_ALU | BPF_MUL | BPF_K:
569 	case BPF_ALU | BPF_MOV | BPF_K:
570 	case BPF_ALU | BPF_DIV | BPF_K:
571 	case BPF_ALU | BPF_MOD | BPF_K:
572 		*to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
573 		*to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
574 		*to++ = BPF_ALU32_REG(from->code, from->dst_reg, BPF_REG_AX);
575 		break;
576 
577 	case BPF_ALU64 | BPF_ADD | BPF_K:
578 	case BPF_ALU64 | BPF_SUB | BPF_K:
579 	case BPF_ALU64 | BPF_AND | BPF_K:
580 	case BPF_ALU64 | BPF_OR  | BPF_K:
581 	case BPF_ALU64 | BPF_XOR | BPF_K:
582 	case BPF_ALU64 | BPF_MUL | BPF_K:
583 	case BPF_ALU64 | BPF_MOV | BPF_K:
584 	case BPF_ALU64 | BPF_DIV | BPF_K:
585 	case BPF_ALU64 | BPF_MOD | BPF_K:
586 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
587 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
588 		*to++ = BPF_ALU64_REG(from->code, from->dst_reg, BPF_REG_AX);
589 		break;
590 
591 	case BPF_JMP | BPF_JEQ  | BPF_K:
592 	case BPF_JMP | BPF_JNE  | BPF_K:
593 	case BPF_JMP | BPF_JGT  | BPF_K:
594 	case BPF_JMP | BPF_JLT  | BPF_K:
595 	case BPF_JMP | BPF_JGE  | BPF_K:
596 	case BPF_JMP | BPF_JLE  | BPF_K:
597 	case BPF_JMP | BPF_JSGT | BPF_K:
598 	case BPF_JMP | BPF_JSLT | BPF_K:
599 	case BPF_JMP | BPF_JSGE | BPF_K:
600 	case BPF_JMP | BPF_JSLE | BPF_K:
601 	case BPF_JMP | BPF_JSET | BPF_K:
602 		/* Accommodate for extra offset in case of a backjump. */
603 		off = from->off;
604 		if (off < 0)
605 			off -= 2;
606 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
607 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
608 		*to++ = BPF_JMP_REG(from->code, from->dst_reg, BPF_REG_AX, off);
609 		break;
610 
611 	case BPF_LD | BPF_ABS | BPF_W:
612 	case BPF_LD | BPF_ABS | BPF_H:
613 	case BPF_LD | BPF_ABS | BPF_B:
614 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
615 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
616 		*to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
617 		break;
618 
619 	case BPF_LD | BPF_IND | BPF_W:
620 	case BPF_LD | BPF_IND | BPF_H:
621 	case BPF_LD | BPF_IND | BPF_B:
622 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
623 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
624 		*to++ = BPF_ALU32_REG(BPF_ADD, BPF_REG_AX, from->src_reg);
625 		*to++ = BPF_LD_IND(from->code, BPF_REG_AX, 0);
626 		break;
627 
628 	case BPF_LD | BPF_IMM | BPF_DW:
629 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[1].imm);
630 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
631 		*to++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_AX, 32);
632 		*to++ = BPF_ALU64_REG(BPF_MOV, aux[0].dst_reg, BPF_REG_AX);
633 		break;
634 	case 0: /* Part 2 of BPF_LD | BPF_IMM | BPF_DW. */
635 		*to++ = BPF_ALU32_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ aux[0].imm);
636 		*to++ = BPF_ALU32_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
637 		*to++ = BPF_ALU64_REG(BPF_OR,  aux[0].dst_reg, BPF_REG_AX);
638 		break;
639 
640 	case BPF_ST | BPF_MEM | BPF_DW:
641 	case BPF_ST | BPF_MEM | BPF_W:
642 	case BPF_ST | BPF_MEM | BPF_H:
643 	case BPF_ST | BPF_MEM | BPF_B:
644 		*to++ = BPF_ALU64_IMM(BPF_MOV, BPF_REG_AX, imm_rnd ^ from->imm);
645 		*to++ = BPF_ALU64_IMM(BPF_XOR, BPF_REG_AX, imm_rnd);
646 		*to++ = BPF_STX_MEM(from->code, from->dst_reg, BPF_REG_AX, from->off);
647 		break;
648 	}
649 out:
650 	return to - to_buff;
651 }
652 
653 static struct bpf_prog *bpf_prog_clone_create(struct bpf_prog *fp_other,
654 					      gfp_t gfp_extra_flags)
655 {
656 	gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | gfp_extra_flags;
657 	struct bpf_prog *fp;
658 
659 	fp = __vmalloc(fp_other->pages * PAGE_SIZE, gfp_flags, PAGE_KERNEL);
660 	if (fp != NULL) {
661 		/* aux->prog still points to the fp_other one, so
662 		 * when promoting the clone to the real program,
663 		 * this still needs to be adapted.
664 		 */
665 		memcpy(fp, fp_other, fp_other->pages * PAGE_SIZE);
666 	}
667 
668 	return fp;
669 }
670 
671 static void bpf_prog_clone_free(struct bpf_prog *fp)
672 {
673 	/* aux was stolen by the other clone, so we cannot free
674 	 * it from this path! It will be freed eventually by the
675 	 * other program on release.
676 	 *
677 	 * At this point, we don't need a deferred release since
678 	 * clone is guaranteed to not be locked.
679 	 */
680 	fp->aux = NULL;
681 	__bpf_prog_free(fp);
682 }
683 
684 void bpf_jit_prog_release_other(struct bpf_prog *fp, struct bpf_prog *fp_other)
685 {
686 	/* We have to repoint aux->prog to self, as we don't
687 	 * know whether fp here is the clone or the original.
688 	 */
689 	fp->aux->prog = fp;
690 	bpf_prog_clone_free(fp_other);
691 }
692 
693 struct bpf_prog *bpf_jit_blind_constants(struct bpf_prog *prog)
694 {
695 	struct bpf_insn insn_buff[16], aux[2];
696 	struct bpf_prog *clone, *tmp;
697 	int insn_delta, insn_cnt;
698 	struct bpf_insn *insn;
699 	int i, rewritten;
700 
701 	if (!bpf_jit_blinding_enabled())
702 		return prog;
703 
704 	clone = bpf_prog_clone_create(prog, GFP_USER);
705 	if (!clone)
706 		return ERR_PTR(-ENOMEM);
707 
708 	insn_cnt = clone->len;
709 	insn = clone->insnsi;
710 
711 	for (i = 0; i < insn_cnt; i++, insn++) {
712 		/* We temporarily need to hold the original ld64 insn
713 		 * so that we can still access the first part in the
714 		 * second blinding run.
715 		 */
716 		if (insn[0].code == (BPF_LD | BPF_IMM | BPF_DW) &&
717 		    insn[1].code == 0)
718 			memcpy(aux, insn, sizeof(aux));
719 
720 		rewritten = bpf_jit_blind_insn(insn, aux, insn_buff);
721 		if (!rewritten)
722 			continue;
723 
724 		tmp = bpf_patch_insn_single(clone, i, insn_buff, rewritten);
725 		if (!tmp) {
726 			/* Patching may have repointed aux->prog during
727 			 * realloc from the original one, so we need to
728 			 * fix it up here on error.
729 			 */
730 			bpf_jit_prog_release_other(prog, clone);
731 			return ERR_PTR(-ENOMEM);
732 		}
733 
734 		clone = tmp;
735 		insn_delta = rewritten - 1;
736 
737 		/* Walk new program and skip insns we just inserted. */
738 		insn = clone->insnsi + i + insn_delta;
739 		insn_cnt += insn_delta;
740 		i        += insn_delta;
741 	}
742 
743 	return clone;
744 }
745 #endif /* CONFIG_BPF_JIT */
746 
747 /* Base function for offset calculation. Needs to go into .text section,
748  * therefore keeping it non-static as well; will also be used by JITs
749  * anyway later on, so do not let the compiler omit it.
750  */
751 noinline u64 __bpf_call_base(u64 r1, u64 r2, u64 r3, u64 r4, u64 r5)
752 {
753 	return 0;
754 }
755 EXPORT_SYMBOL_GPL(__bpf_call_base);
756 
757 /**
758  *	__bpf_prog_run - run eBPF program on a given context
759  *	@ctx: is the data we are operating on
760  *	@insn: is the array of eBPF instructions
761  *
762  * Decode and execute eBPF instructions.
763  */
764 static unsigned int ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn,
765 				    u64 *stack)
766 {
767 	u64 tmp;
768 	static const void *jumptable[256] = {
769 		[0 ... 255] = &&default_label,
770 		/* Now overwrite non-defaults ... */
771 		/* 32 bit ALU operations */
772 		[BPF_ALU | BPF_ADD | BPF_X] = &&ALU_ADD_X,
773 		[BPF_ALU | BPF_ADD | BPF_K] = &&ALU_ADD_K,
774 		[BPF_ALU | BPF_SUB | BPF_X] = &&ALU_SUB_X,
775 		[BPF_ALU | BPF_SUB | BPF_K] = &&ALU_SUB_K,
776 		[BPF_ALU | BPF_AND | BPF_X] = &&ALU_AND_X,
777 		[BPF_ALU | BPF_AND | BPF_K] = &&ALU_AND_K,
778 		[BPF_ALU | BPF_OR | BPF_X]  = &&ALU_OR_X,
779 		[BPF_ALU | BPF_OR | BPF_K]  = &&ALU_OR_K,
780 		[BPF_ALU | BPF_LSH | BPF_X] = &&ALU_LSH_X,
781 		[BPF_ALU | BPF_LSH | BPF_K] = &&ALU_LSH_K,
782 		[BPF_ALU | BPF_RSH | BPF_X] = &&ALU_RSH_X,
783 		[BPF_ALU | BPF_RSH | BPF_K] = &&ALU_RSH_K,
784 		[BPF_ALU | BPF_XOR | BPF_X] = &&ALU_XOR_X,
785 		[BPF_ALU | BPF_XOR | BPF_K] = &&ALU_XOR_K,
786 		[BPF_ALU | BPF_MUL | BPF_X] = &&ALU_MUL_X,
787 		[BPF_ALU | BPF_MUL | BPF_K] = &&ALU_MUL_K,
788 		[BPF_ALU | BPF_MOV | BPF_X] = &&ALU_MOV_X,
789 		[BPF_ALU | BPF_MOV | BPF_K] = &&ALU_MOV_K,
790 		[BPF_ALU | BPF_DIV | BPF_X] = &&ALU_DIV_X,
791 		[BPF_ALU | BPF_DIV | BPF_K] = &&ALU_DIV_K,
792 		[BPF_ALU | BPF_MOD | BPF_X] = &&ALU_MOD_X,
793 		[BPF_ALU | BPF_MOD | BPF_K] = &&ALU_MOD_K,
794 		[BPF_ALU | BPF_NEG] = &&ALU_NEG,
795 		[BPF_ALU | BPF_END | BPF_TO_BE] = &&ALU_END_TO_BE,
796 		[BPF_ALU | BPF_END | BPF_TO_LE] = &&ALU_END_TO_LE,
797 		/* 64 bit ALU operations */
798 		[BPF_ALU64 | BPF_ADD | BPF_X] = &&ALU64_ADD_X,
799 		[BPF_ALU64 | BPF_ADD | BPF_K] = &&ALU64_ADD_K,
800 		[BPF_ALU64 | BPF_SUB | BPF_X] = &&ALU64_SUB_X,
801 		[BPF_ALU64 | BPF_SUB | BPF_K] = &&ALU64_SUB_K,
802 		[BPF_ALU64 | BPF_AND | BPF_X] = &&ALU64_AND_X,
803 		[BPF_ALU64 | BPF_AND | BPF_K] = &&ALU64_AND_K,
804 		[BPF_ALU64 | BPF_OR | BPF_X] = &&ALU64_OR_X,
805 		[BPF_ALU64 | BPF_OR | BPF_K] = &&ALU64_OR_K,
806 		[BPF_ALU64 | BPF_LSH | BPF_X] = &&ALU64_LSH_X,
807 		[BPF_ALU64 | BPF_LSH | BPF_K] = &&ALU64_LSH_K,
808 		[BPF_ALU64 | BPF_RSH | BPF_X] = &&ALU64_RSH_X,
809 		[BPF_ALU64 | BPF_RSH | BPF_K] = &&ALU64_RSH_K,
810 		[BPF_ALU64 | BPF_XOR | BPF_X] = &&ALU64_XOR_X,
811 		[BPF_ALU64 | BPF_XOR | BPF_K] = &&ALU64_XOR_K,
812 		[BPF_ALU64 | BPF_MUL | BPF_X] = &&ALU64_MUL_X,
813 		[BPF_ALU64 | BPF_MUL | BPF_K] = &&ALU64_MUL_K,
814 		[BPF_ALU64 | BPF_MOV | BPF_X] = &&ALU64_MOV_X,
815 		[BPF_ALU64 | BPF_MOV | BPF_K] = &&ALU64_MOV_K,
816 		[BPF_ALU64 | BPF_ARSH | BPF_X] = &&ALU64_ARSH_X,
817 		[BPF_ALU64 | BPF_ARSH | BPF_K] = &&ALU64_ARSH_K,
818 		[BPF_ALU64 | BPF_DIV | BPF_X] = &&ALU64_DIV_X,
819 		[BPF_ALU64 | BPF_DIV | BPF_K] = &&ALU64_DIV_K,
820 		[BPF_ALU64 | BPF_MOD | BPF_X] = &&ALU64_MOD_X,
821 		[BPF_ALU64 | BPF_MOD | BPF_K] = &&ALU64_MOD_K,
822 		[BPF_ALU64 | BPF_NEG] = &&ALU64_NEG,
823 		/* Call instruction */
824 		[BPF_JMP | BPF_CALL] = &&JMP_CALL,
825 		[BPF_JMP | BPF_TAIL_CALL] = &&JMP_TAIL_CALL,
826 		/* Jumps */
827 		[BPF_JMP | BPF_JA] = &&JMP_JA,
828 		[BPF_JMP | BPF_JEQ | BPF_X] = &&JMP_JEQ_X,
829 		[BPF_JMP | BPF_JEQ | BPF_K] = &&JMP_JEQ_K,
830 		[BPF_JMP | BPF_JNE | BPF_X] = &&JMP_JNE_X,
831 		[BPF_JMP | BPF_JNE | BPF_K] = &&JMP_JNE_K,
832 		[BPF_JMP | BPF_JGT | BPF_X] = &&JMP_JGT_X,
833 		[BPF_JMP | BPF_JGT | BPF_K] = &&JMP_JGT_K,
834 		[BPF_JMP | BPF_JLT | BPF_X] = &&JMP_JLT_X,
835 		[BPF_JMP | BPF_JLT | BPF_K] = &&JMP_JLT_K,
836 		[BPF_JMP | BPF_JGE | BPF_X] = &&JMP_JGE_X,
837 		[BPF_JMP | BPF_JGE | BPF_K] = &&JMP_JGE_K,
838 		[BPF_JMP | BPF_JLE | BPF_X] = &&JMP_JLE_X,
839 		[BPF_JMP | BPF_JLE | BPF_K] = &&JMP_JLE_K,
840 		[BPF_JMP | BPF_JSGT | BPF_X] = &&JMP_JSGT_X,
841 		[BPF_JMP | BPF_JSGT | BPF_K] = &&JMP_JSGT_K,
842 		[BPF_JMP | BPF_JSLT | BPF_X] = &&JMP_JSLT_X,
843 		[BPF_JMP | BPF_JSLT | BPF_K] = &&JMP_JSLT_K,
844 		[BPF_JMP | BPF_JSGE | BPF_X] = &&JMP_JSGE_X,
845 		[BPF_JMP | BPF_JSGE | BPF_K] = &&JMP_JSGE_K,
846 		[BPF_JMP | BPF_JSLE | BPF_X] = &&JMP_JSLE_X,
847 		[BPF_JMP | BPF_JSLE | BPF_K] = &&JMP_JSLE_K,
848 		[BPF_JMP | BPF_JSET | BPF_X] = &&JMP_JSET_X,
849 		[BPF_JMP | BPF_JSET | BPF_K] = &&JMP_JSET_K,
850 		/* Program return */
851 		[BPF_JMP | BPF_EXIT] = &&JMP_EXIT,
852 		/* Store instructions */
853 		[BPF_STX | BPF_MEM | BPF_B] = &&STX_MEM_B,
854 		[BPF_STX | BPF_MEM | BPF_H] = &&STX_MEM_H,
855 		[BPF_STX | BPF_MEM | BPF_W] = &&STX_MEM_W,
856 		[BPF_STX | BPF_MEM | BPF_DW] = &&STX_MEM_DW,
857 		[BPF_STX | BPF_XADD | BPF_W] = &&STX_XADD_W,
858 		[BPF_STX | BPF_XADD | BPF_DW] = &&STX_XADD_DW,
859 		[BPF_ST | BPF_MEM | BPF_B] = &&ST_MEM_B,
860 		[BPF_ST | BPF_MEM | BPF_H] = &&ST_MEM_H,
861 		[BPF_ST | BPF_MEM | BPF_W] = &&ST_MEM_W,
862 		[BPF_ST | BPF_MEM | BPF_DW] = &&ST_MEM_DW,
863 		/* Load instructions */
864 		[BPF_LDX | BPF_MEM | BPF_B] = &&LDX_MEM_B,
865 		[BPF_LDX | BPF_MEM | BPF_H] = &&LDX_MEM_H,
866 		[BPF_LDX | BPF_MEM | BPF_W] = &&LDX_MEM_W,
867 		[BPF_LDX | BPF_MEM | BPF_DW] = &&LDX_MEM_DW,
868 		[BPF_LD | BPF_ABS | BPF_W] = &&LD_ABS_W,
869 		[BPF_LD | BPF_ABS | BPF_H] = &&LD_ABS_H,
870 		[BPF_LD | BPF_ABS | BPF_B] = &&LD_ABS_B,
871 		[BPF_LD | BPF_IND | BPF_W] = &&LD_IND_W,
872 		[BPF_LD | BPF_IND | BPF_H] = &&LD_IND_H,
873 		[BPF_LD | BPF_IND | BPF_B] = &&LD_IND_B,
874 		[BPF_LD | BPF_IMM | BPF_DW] = &&LD_IMM_DW,
875 	};
876 	u32 tail_call_cnt = 0;
877 	void *ptr;
878 	int off;
879 
880 #define CONT	 ({ insn++; goto select_insn; })
881 #define CONT_JMP ({ insn++; goto select_insn; })
882 
883 select_insn:
884 	goto *jumptable[insn->code];
885 
886 	/* ALU */
887 #define ALU(OPCODE, OP)			\
888 	ALU64_##OPCODE##_X:		\
889 		DST = DST OP SRC;	\
890 		CONT;			\
891 	ALU_##OPCODE##_X:		\
892 		DST = (u32) DST OP (u32) SRC;	\
893 		CONT;			\
894 	ALU64_##OPCODE##_K:		\
895 		DST = DST OP IMM;		\
896 		CONT;			\
897 	ALU_##OPCODE##_K:		\
898 		DST = (u32) DST OP (u32) IMM;	\
899 		CONT;
900 
901 	ALU(ADD,  +)
902 	ALU(SUB,  -)
903 	ALU(AND,  &)
904 	ALU(OR,   |)
905 	ALU(LSH, <<)
906 	ALU(RSH, >>)
907 	ALU(XOR,  ^)
908 	ALU(MUL,  *)
909 #undef ALU
910 	ALU_NEG:
911 		DST = (u32) -DST;
912 		CONT;
913 	ALU64_NEG:
914 		DST = -DST;
915 		CONT;
916 	ALU_MOV_X:
917 		DST = (u32) SRC;
918 		CONT;
919 	ALU_MOV_K:
920 		DST = (u32) IMM;
921 		CONT;
922 	ALU64_MOV_X:
923 		DST = SRC;
924 		CONT;
925 	ALU64_MOV_K:
926 		DST = IMM;
927 		CONT;
928 	LD_IMM_DW:
929 		DST = (u64) (u32) insn[0].imm | ((u64) (u32) insn[1].imm) << 32;
930 		insn++;
931 		CONT;
932 	ALU64_ARSH_X:
933 		(*(s64 *) &DST) >>= SRC;
934 		CONT;
935 	ALU64_ARSH_K:
936 		(*(s64 *) &DST) >>= IMM;
937 		CONT;
938 	ALU64_MOD_X:
939 		if (unlikely(SRC == 0))
940 			return 0;
941 		div64_u64_rem(DST, SRC, &tmp);
942 		DST = tmp;
943 		CONT;
944 	ALU_MOD_X:
945 		if (unlikely(SRC == 0))
946 			return 0;
947 		tmp = (u32) DST;
948 		DST = do_div(tmp, (u32) SRC);
949 		CONT;
950 	ALU64_MOD_K:
951 		div64_u64_rem(DST, IMM, &tmp);
952 		DST = tmp;
953 		CONT;
954 	ALU_MOD_K:
955 		tmp = (u32) DST;
956 		DST = do_div(tmp, (u32) IMM);
957 		CONT;
958 	ALU64_DIV_X:
959 		if (unlikely(SRC == 0))
960 			return 0;
961 		DST = div64_u64(DST, SRC);
962 		CONT;
963 	ALU_DIV_X:
964 		if (unlikely(SRC == 0))
965 			return 0;
966 		tmp = (u32) DST;
967 		do_div(tmp, (u32) SRC);
968 		DST = (u32) tmp;
969 		CONT;
970 	ALU64_DIV_K:
971 		DST = div64_u64(DST, IMM);
972 		CONT;
973 	ALU_DIV_K:
974 		tmp = (u32) DST;
975 		do_div(tmp, (u32) IMM);
976 		DST = (u32) tmp;
977 		CONT;
978 	ALU_END_TO_BE:
979 		switch (IMM) {
980 		case 16:
981 			DST = (__force u16) cpu_to_be16(DST);
982 			break;
983 		case 32:
984 			DST = (__force u32) cpu_to_be32(DST);
985 			break;
986 		case 64:
987 			DST = (__force u64) cpu_to_be64(DST);
988 			break;
989 		}
990 		CONT;
991 	ALU_END_TO_LE:
992 		switch (IMM) {
993 		case 16:
994 			DST = (__force u16) cpu_to_le16(DST);
995 			break;
996 		case 32:
997 			DST = (__force u32) cpu_to_le32(DST);
998 			break;
999 		case 64:
1000 			DST = (__force u64) cpu_to_le64(DST);
1001 			break;
1002 		}
1003 		CONT;
1004 
1005 	/* CALL */
1006 	JMP_CALL:
1007 		/* Function call scratches BPF_R1-BPF_R5 registers,
1008 		 * preserves BPF_R6-BPF_R9, and stores return value
1009 		 * into BPF_R0.
1010 		 */
1011 		BPF_R0 = (__bpf_call_base + insn->imm)(BPF_R1, BPF_R2, BPF_R3,
1012 						       BPF_R4, BPF_R5);
1013 		CONT;
1014 
1015 	JMP_TAIL_CALL: {
1016 		struct bpf_map *map = (struct bpf_map *) (unsigned long) BPF_R2;
1017 		struct bpf_array *array = container_of(map, struct bpf_array, map);
1018 		struct bpf_prog *prog;
1019 		u32 index = BPF_R3;
1020 
1021 		if (unlikely(index >= array->map.max_entries))
1022 			goto out;
1023 		if (unlikely(tail_call_cnt > MAX_TAIL_CALL_CNT))
1024 			goto out;
1025 
1026 		tail_call_cnt++;
1027 
1028 		prog = READ_ONCE(array->ptrs[index]);
1029 		if (!prog)
1030 			goto out;
1031 
1032 		/* ARG1 at this point is guaranteed to point to CTX from
1033 		 * the verifier side due to the fact that the tail call is
1034 		 * handeled like a helper, that is, bpf_tail_call_proto,
1035 		 * where arg1_type is ARG_PTR_TO_CTX.
1036 		 */
1037 		insn = prog->insnsi;
1038 		goto select_insn;
1039 out:
1040 		CONT;
1041 	}
1042 	/* JMP */
1043 	JMP_JA:
1044 		insn += insn->off;
1045 		CONT;
1046 	JMP_JEQ_X:
1047 		if (DST == SRC) {
1048 			insn += insn->off;
1049 			CONT_JMP;
1050 		}
1051 		CONT;
1052 	JMP_JEQ_K:
1053 		if (DST == IMM) {
1054 			insn += insn->off;
1055 			CONT_JMP;
1056 		}
1057 		CONT;
1058 	JMP_JNE_X:
1059 		if (DST != SRC) {
1060 			insn += insn->off;
1061 			CONT_JMP;
1062 		}
1063 		CONT;
1064 	JMP_JNE_K:
1065 		if (DST != IMM) {
1066 			insn += insn->off;
1067 			CONT_JMP;
1068 		}
1069 		CONT;
1070 	JMP_JGT_X:
1071 		if (DST > SRC) {
1072 			insn += insn->off;
1073 			CONT_JMP;
1074 		}
1075 		CONT;
1076 	JMP_JGT_K:
1077 		if (DST > IMM) {
1078 			insn += insn->off;
1079 			CONT_JMP;
1080 		}
1081 		CONT;
1082 	JMP_JLT_X:
1083 		if (DST < SRC) {
1084 			insn += insn->off;
1085 			CONT_JMP;
1086 		}
1087 		CONT;
1088 	JMP_JLT_K:
1089 		if (DST < IMM) {
1090 			insn += insn->off;
1091 			CONT_JMP;
1092 		}
1093 		CONT;
1094 	JMP_JGE_X:
1095 		if (DST >= SRC) {
1096 			insn += insn->off;
1097 			CONT_JMP;
1098 		}
1099 		CONT;
1100 	JMP_JGE_K:
1101 		if (DST >= IMM) {
1102 			insn += insn->off;
1103 			CONT_JMP;
1104 		}
1105 		CONT;
1106 	JMP_JLE_X:
1107 		if (DST <= SRC) {
1108 			insn += insn->off;
1109 			CONT_JMP;
1110 		}
1111 		CONT;
1112 	JMP_JLE_K:
1113 		if (DST <= IMM) {
1114 			insn += insn->off;
1115 			CONT_JMP;
1116 		}
1117 		CONT;
1118 	JMP_JSGT_X:
1119 		if (((s64) DST) > ((s64) SRC)) {
1120 			insn += insn->off;
1121 			CONT_JMP;
1122 		}
1123 		CONT;
1124 	JMP_JSGT_K:
1125 		if (((s64) DST) > ((s64) IMM)) {
1126 			insn += insn->off;
1127 			CONT_JMP;
1128 		}
1129 		CONT;
1130 	JMP_JSLT_X:
1131 		if (((s64) DST) < ((s64) SRC)) {
1132 			insn += insn->off;
1133 			CONT_JMP;
1134 		}
1135 		CONT;
1136 	JMP_JSLT_K:
1137 		if (((s64) DST) < ((s64) IMM)) {
1138 			insn += insn->off;
1139 			CONT_JMP;
1140 		}
1141 		CONT;
1142 	JMP_JSGE_X:
1143 		if (((s64) DST) >= ((s64) SRC)) {
1144 			insn += insn->off;
1145 			CONT_JMP;
1146 		}
1147 		CONT;
1148 	JMP_JSGE_K:
1149 		if (((s64) DST) >= ((s64) IMM)) {
1150 			insn += insn->off;
1151 			CONT_JMP;
1152 		}
1153 		CONT;
1154 	JMP_JSLE_X:
1155 		if (((s64) DST) <= ((s64) SRC)) {
1156 			insn += insn->off;
1157 			CONT_JMP;
1158 		}
1159 		CONT;
1160 	JMP_JSLE_K:
1161 		if (((s64) DST) <= ((s64) IMM)) {
1162 			insn += insn->off;
1163 			CONT_JMP;
1164 		}
1165 		CONT;
1166 	JMP_JSET_X:
1167 		if (DST & SRC) {
1168 			insn += insn->off;
1169 			CONT_JMP;
1170 		}
1171 		CONT;
1172 	JMP_JSET_K:
1173 		if (DST & IMM) {
1174 			insn += insn->off;
1175 			CONT_JMP;
1176 		}
1177 		CONT;
1178 	JMP_EXIT:
1179 		return BPF_R0;
1180 
1181 	/* STX and ST and LDX*/
1182 #define LDST(SIZEOP, SIZE)						\
1183 	STX_MEM_##SIZEOP:						\
1184 		*(SIZE *)(unsigned long) (DST + insn->off) = SRC;	\
1185 		CONT;							\
1186 	ST_MEM_##SIZEOP:						\
1187 		*(SIZE *)(unsigned long) (DST + insn->off) = IMM;	\
1188 		CONT;							\
1189 	LDX_MEM_##SIZEOP:						\
1190 		DST = *(SIZE *)(unsigned long) (SRC + insn->off);	\
1191 		CONT;
1192 
1193 	LDST(B,   u8)
1194 	LDST(H,  u16)
1195 	LDST(W,  u32)
1196 	LDST(DW, u64)
1197 #undef LDST
1198 	STX_XADD_W: /* lock xadd *(u32 *)(dst_reg + off16) += src_reg */
1199 		atomic_add((u32) SRC, (atomic_t *)(unsigned long)
1200 			   (DST + insn->off));
1201 		CONT;
1202 	STX_XADD_DW: /* lock xadd *(u64 *)(dst_reg + off16) += src_reg */
1203 		atomic64_add((u64) SRC, (atomic64_t *)(unsigned long)
1204 			     (DST + insn->off));
1205 		CONT;
1206 	LD_ABS_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + imm32)) */
1207 		off = IMM;
1208 load_word:
1209 		/* BPF_LD + BPD_ABS and BPF_LD + BPF_IND insns are only
1210 		 * appearing in the programs where ctx == skb
1211 		 * (see may_access_skb() in the verifier). All programs
1212 		 * keep 'ctx' in regs[BPF_REG_CTX] == BPF_R6,
1213 		 * bpf_convert_filter() saves it in BPF_R6, internal BPF
1214 		 * verifier will check that BPF_R6 == ctx.
1215 		 *
1216 		 * BPF_ABS and BPF_IND are wrappers of function calls,
1217 		 * so they scratch BPF_R1-BPF_R5 registers, preserve
1218 		 * BPF_R6-BPF_R9, and store return value into BPF_R0.
1219 		 *
1220 		 * Implicit input:
1221 		 *   ctx == skb == BPF_R6 == CTX
1222 		 *
1223 		 * Explicit input:
1224 		 *   SRC == any register
1225 		 *   IMM == 32-bit immediate
1226 		 *
1227 		 * Output:
1228 		 *   BPF_R0 - 8/16/32-bit skb data converted to cpu endianness
1229 		 */
1230 
1231 		ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 4, &tmp);
1232 		if (likely(ptr != NULL)) {
1233 			BPF_R0 = get_unaligned_be32(ptr);
1234 			CONT;
1235 		}
1236 
1237 		return 0;
1238 	LD_ABS_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + imm32)) */
1239 		off = IMM;
1240 load_half:
1241 		ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 2, &tmp);
1242 		if (likely(ptr != NULL)) {
1243 			BPF_R0 = get_unaligned_be16(ptr);
1244 			CONT;
1245 		}
1246 
1247 		return 0;
1248 	LD_ABS_B: /* BPF_R0 = *(u8 *) (skb->data + imm32) */
1249 		off = IMM;
1250 load_byte:
1251 		ptr = bpf_load_pointer((struct sk_buff *) (unsigned long) CTX, off, 1, &tmp);
1252 		if (likely(ptr != NULL)) {
1253 			BPF_R0 = *(u8 *)ptr;
1254 			CONT;
1255 		}
1256 
1257 		return 0;
1258 	LD_IND_W: /* BPF_R0 = ntohl(*(u32 *) (skb->data + src_reg + imm32)) */
1259 		off = IMM + SRC;
1260 		goto load_word;
1261 	LD_IND_H: /* BPF_R0 = ntohs(*(u16 *) (skb->data + src_reg + imm32)) */
1262 		off = IMM + SRC;
1263 		goto load_half;
1264 	LD_IND_B: /* BPF_R0 = *(u8 *) (skb->data + src_reg + imm32) */
1265 		off = IMM + SRC;
1266 		goto load_byte;
1267 
1268 	default_label:
1269 		/* If we ever reach this, we have a bug somewhere. */
1270 		WARN_RATELIMIT(1, "unknown opcode %02x\n", insn->code);
1271 		return 0;
1272 }
1273 STACK_FRAME_NON_STANDARD(___bpf_prog_run); /* jump table */
1274 
1275 #define PROG_NAME(stack_size) __bpf_prog_run##stack_size
1276 #define DEFINE_BPF_PROG_RUN(stack_size) \
1277 static unsigned int PROG_NAME(stack_size)(const void *ctx, const struct bpf_insn *insn) \
1278 { \
1279 	u64 stack[stack_size / sizeof(u64)]; \
1280 	u64 regs[MAX_BPF_REG]; \
1281 \
1282 	FP = (u64) (unsigned long) &stack[ARRAY_SIZE(stack)]; \
1283 	ARG1 = (u64) (unsigned long) ctx; \
1284 	return ___bpf_prog_run(regs, insn, stack); \
1285 }
1286 
1287 #define EVAL1(FN, X) FN(X)
1288 #define EVAL2(FN, X, Y...) FN(X) EVAL1(FN, Y)
1289 #define EVAL3(FN, X, Y...) FN(X) EVAL2(FN, Y)
1290 #define EVAL4(FN, X, Y...) FN(X) EVAL3(FN, Y)
1291 #define EVAL5(FN, X, Y...) FN(X) EVAL4(FN, Y)
1292 #define EVAL6(FN, X, Y...) FN(X) EVAL5(FN, Y)
1293 
1294 EVAL6(DEFINE_BPF_PROG_RUN, 32, 64, 96, 128, 160, 192);
1295 EVAL6(DEFINE_BPF_PROG_RUN, 224, 256, 288, 320, 352, 384);
1296 EVAL4(DEFINE_BPF_PROG_RUN, 416, 448, 480, 512);
1297 
1298 #define PROG_NAME_LIST(stack_size) PROG_NAME(stack_size),
1299 
1300 static unsigned int (*interpreters[])(const void *ctx,
1301 				      const struct bpf_insn *insn) = {
1302 EVAL6(PROG_NAME_LIST, 32, 64, 96, 128, 160, 192)
1303 EVAL6(PROG_NAME_LIST, 224, 256, 288, 320, 352, 384)
1304 EVAL4(PROG_NAME_LIST, 416, 448, 480, 512)
1305 };
1306 
1307 bool bpf_prog_array_compatible(struct bpf_array *array,
1308 			       const struct bpf_prog *fp)
1309 {
1310 	if (!array->owner_prog_type) {
1311 		/* There's no owner yet where we could check for
1312 		 * compatibility.
1313 		 */
1314 		array->owner_prog_type = fp->type;
1315 		array->owner_jited = fp->jited;
1316 
1317 		return true;
1318 	}
1319 
1320 	return array->owner_prog_type == fp->type &&
1321 	       array->owner_jited == fp->jited;
1322 }
1323 
1324 static int bpf_check_tail_call(const struct bpf_prog *fp)
1325 {
1326 	struct bpf_prog_aux *aux = fp->aux;
1327 	int i;
1328 
1329 	for (i = 0; i < aux->used_map_cnt; i++) {
1330 		struct bpf_map *map = aux->used_maps[i];
1331 		struct bpf_array *array;
1332 
1333 		if (map->map_type != BPF_MAP_TYPE_PROG_ARRAY)
1334 			continue;
1335 
1336 		array = container_of(map, struct bpf_array, map);
1337 		if (!bpf_prog_array_compatible(array, fp))
1338 			return -EINVAL;
1339 	}
1340 
1341 	return 0;
1342 }
1343 
1344 /**
1345  *	bpf_prog_select_runtime - select exec runtime for BPF program
1346  *	@fp: bpf_prog populated with internal BPF program
1347  *	@err: pointer to error variable
1348  *
1349  * Try to JIT eBPF program, if JIT is not available, use interpreter.
1350  * The BPF program will be executed via BPF_PROG_RUN() macro.
1351  */
1352 struct bpf_prog *bpf_prog_select_runtime(struct bpf_prog *fp, int *err)
1353 {
1354 	u32 stack_depth = max_t(u32, fp->aux->stack_depth, 1);
1355 
1356 	fp->bpf_func = interpreters[(round_up(stack_depth, 32) / 32) - 1];
1357 
1358 	/* eBPF JITs can rewrite the program in case constant
1359 	 * blinding is active. However, in case of error during
1360 	 * blinding, bpf_int_jit_compile() must always return a
1361 	 * valid program, which in this case would simply not
1362 	 * be JITed, but falls back to the interpreter.
1363 	 */
1364 	fp = bpf_int_jit_compile(fp);
1365 	bpf_prog_lock_ro(fp);
1366 
1367 	/* The tail call compatibility check can only be done at
1368 	 * this late stage as we need to determine, if we deal
1369 	 * with JITed or non JITed program concatenations and not
1370 	 * all eBPF JITs might immediately support all features.
1371 	 */
1372 	*err = bpf_check_tail_call(fp);
1373 
1374 	return fp;
1375 }
1376 EXPORT_SYMBOL_GPL(bpf_prog_select_runtime);
1377 
1378 static void bpf_prog_free_deferred(struct work_struct *work)
1379 {
1380 	struct bpf_prog_aux *aux;
1381 
1382 	aux = container_of(work, struct bpf_prog_aux, work);
1383 	bpf_jit_free(aux->prog);
1384 }
1385 
1386 /* Free internal BPF program */
1387 void bpf_prog_free(struct bpf_prog *fp)
1388 {
1389 	struct bpf_prog_aux *aux = fp->aux;
1390 
1391 	INIT_WORK(&aux->work, bpf_prog_free_deferred);
1392 	schedule_work(&aux->work);
1393 }
1394 EXPORT_SYMBOL_GPL(bpf_prog_free);
1395 
1396 /* RNG for unpriviledged user space with separated state from prandom_u32(). */
1397 static DEFINE_PER_CPU(struct rnd_state, bpf_user_rnd_state);
1398 
1399 void bpf_user_rnd_init_once(void)
1400 {
1401 	prandom_init_once(&bpf_user_rnd_state);
1402 }
1403 
1404 BPF_CALL_0(bpf_user_rnd_u32)
1405 {
1406 	/* Should someone ever have the rather unwise idea to use some
1407 	 * of the registers passed into this function, then note that
1408 	 * this function is called from native eBPF and classic-to-eBPF
1409 	 * transformations. Register assignments from both sides are
1410 	 * different, f.e. classic always sets fn(ctx, A, X) here.
1411 	 */
1412 	struct rnd_state *state;
1413 	u32 res;
1414 
1415 	state = &get_cpu_var(bpf_user_rnd_state);
1416 	res = prandom_u32_state(state);
1417 	put_cpu_var(bpf_user_rnd_state);
1418 
1419 	return res;
1420 }
1421 
1422 /* Weak definitions of helper functions in case we don't have bpf syscall. */
1423 const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;
1424 const struct bpf_func_proto bpf_map_update_elem_proto __weak;
1425 const struct bpf_func_proto bpf_map_delete_elem_proto __weak;
1426 
1427 const struct bpf_func_proto bpf_get_prandom_u32_proto __weak;
1428 const struct bpf_func_proto bpf_get_smp_processor_id_proto __weak;
1429 const struct bpf_func_proto bpf_get_numa_node_id_proto __weak;
1430 const struct bpf_func_proto bpf_ktime_get_ns_proto __weak;
1431 
1432 const struct bpf_func_proto bpf_get_current_pid_tgid_proto __weak;
1433 const struct bpf_func_proto bpf_get_current_uid_gid_proto __weak;
1434 const struct bpf_func_proto bpf_get_current_comm_proto __weak;
1435 const struct bpf_func_proto bpf_sock_map_update_proto __weak;
1436 
1437 const struct bpf_func_proto * __weak bpf_get_trace_printk_proto(void)
1438 {
1439 	return NULL;
1440 }
1441 
1442 u64 __weak
1443 bpf_event_output(struct bpf_map *map, u64 flags, void *meta, u64 meta_size,
1444 		 void *ctx, u64 ctx_size, bpf_ctx_copy_t ctx_copy)
1445 {
1446 	return -ENOTSUPP;
1447 }
1448 
1449 /* Always built-in helper functions. */
1450 const struct bpf_func_proto bpf_tail_call_proto = {
1451 	.func		= NULL,
1452 	.gpl_only	= false,
1453 	.ret_type	= RET_VOID,
1454 	.arg1_type	= ARG_PTR_TO_CTX,
1455 	.arg2_type	= ARG_CONST_MAP_PTR,
1456 	.arg3_type	= ARG_ANYTHING,
1457 };
1458 
1459 /* Stub for JITs that only support cBPF. eBPF programs are interpreted.
1460  * It is encouraged to implement bpf_int_jit_compile() instead, so that
1461  * eBPF and implicitly also cBPF can get JITed!
1462  */
1463 struct bpf_prog * __weak bpf_int_jit_compile(struct bpf_prog *prog)
1464 {
1465 	return prog;
1466 }
1467 
1468 /* Stub for JITs that support eBPF. All cBPF code gets transformed into
1469  * eBPF by the kernel and is later compiled by bpf_int_jit_compile().
1470  */
1471 void __weak bpf_jit_compile(struct bpf_prog *prog)
1472 {
1473 }
1474 
1475 bool __weak bpf_helper_changes_pkt_data(void *func)
1476 {
1477 	return false;
1478 }
1479 
1480 /* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call
1481  * skb_copy_bits(), so provide a weak definition of it for NET-less config.
1482  */
1483 int __weak skb_copy_bits(const struct sk_buff *skb, int offset, void *to,
1484 			 int len)
1485 {
1486 	return -EFAULT;
1487 }
1488 
1489 /* All definitions of tracepoints related to BPF. */
1490 #define CREATE_TRACE_POINTS
1491 #include <linux/bpf_trace.h>
1492 
1493 EXPORT_TRACEPOINT_SYMBOL_GPL(xdp_exception);
1494 
1495 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_get_type);
1496 EXPORT_TRACEPOINT_SYMBOL_GPL(bpf_prog_put_rcu);
1497