xref: /openbmc/linux/arch/x86/net/bpf_jit_comp.c (revision 3dc4b6fb)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * bpf_jit_comp.c: BPF JIT compiler
4  *
5  * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
6  * Internal BPF Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
7  */
8 #include <linux/netdevice.h>
9 #include <linux/filter.h>
10 #include <linux/if_vlan.h>
11 #include <linux/bpf.h>
12 
13 #include <asm/set_memory.h>
14 #include <asm/nospec-branch.h>
15 
16 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
17 {
18 	if (len == 1)
19 		*ptr = bytes;
20 	else if (len == 2)
21 		*(u16 *)ptr = bytes;
22 	else {
23 		*(u32 *)ptr = bytes;
24 		barrier();
25 	}
26 	return ptr + len;
27 }
28 
29 #define EMIT(bytes, len) \
30 	do { prog = emit_code(prog, bytes, len); cnt += len; } while (0)
31 
32 #define EMIT1(b1)		EMIT(b1, 1)
33 #define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
34 #define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
35 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
36 
37 #define EMIT1_off32(b1, off) \
38 	do { EMIT1(b1); EMIT(off, 4); } while (0)
39 #define EMIT2_off32(b1, b2, off) \
40 	do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
41 #define EMIT3_off32(b1, b2, b3, off) \
42 	do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
43 #define EMIT4_off32(b1, b2, b3, b4, off) \
44 	do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
45 
46 static bool is_imm8(int value)
47 {
48 	return value <= 127 && value >= -128;
49 }
50 
51 static bool is_simm32(s64 value)
52 {
53 	return value == (s64)(s32)value;
54 }
55 
56 static bool is_uimm32(u64 value)
57 {
58 	return value == (u64)(u32)value;
59 }
60 
61 /* mov dst, src */
62 #define EMIT_mov(DST, SRC)								 \
63 	do {										 \
64 		if (DST != SRC)								 \
65 			EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
66 	} while (0)
67 
68 static int bpf_size_to_x86_bytes(int bpf_size)
69 {
70 	if (bpf_size == BPF_W)
71 		return 4;
72 	else if (bpf_size == BPF_H)
73 		return 2;
74 	else if (bpf_size == BPF_B)
75 		return 1;
76 	else if (bpf_size == BPF_DW)
77 		return 4; /* imm32 */
78 	else
79 		return 0;
80 }
81 
82 /*
83  * List of x86 cond jumps opcodes (. + s8)
84  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
85  */
86 #define X86_JB  0x72
87 #define X86_JAE 0x73
88 #define X86_JE  0x74
89 #define X86_JNE 0x75
90 #define X86_JBE 0x76
91 #define X86_JA  0x77
92 #define X86_JL  0x7C
93 #define X86_JGE 0x7D
94 #define X86_JLE 0x7E
95 #define X86_JG  0x7F
96 
97 /* Pick a register outside of BPF range for JIT internal work */
98 #define AUX_REG (MAX_BPF_JIT_REG + 1)
99 
100 /*
101  * The following table maps BPF registers to x86-64 registers.
102  *
103  * x86-64 register R12 is unused, since if used as base address
104  * register in load/store instructions, it always needs an
105  * extra byte of encoding and is callee saved.
106  *
107  * Also x86-64 register R9 is unused. x86-64 register R10 is
108  * used for blinding (if enabled).
109  */
110 static const int reg2hex[] = {
111 	[BPF_REG_0] = 0,  /* RAX */
112 	[BPF_REG_1] = 7,  /* RDI */
113 	[BPF_REG_2] = 6,  /* RSI */
114 	[BPF_REG_3] = 2,  /* RDX */
115 	[BPF_REG_4] = 1,  /* RCX */
116 	[BPF_REG_5] = 0,  /* R8  */
117 	[BPF_REG_6] = 3,  /* RBX callee saved */
118 	[BPF_REG_7] = 5,  /* R13 callee saved */
119 	[BPF_REG_8] = 6,  /* R14 callee saved */
120 	[BPF_REG_9] = 7,  /* R15 callee saved */
121 	[BPF_REG_FP] = 5, /* RBP readonly */
122 	[BPF_REG_AX] = 2, /* R10 temp register */
123 	[AUX_REG] = 3,    /* R11 temp register */
124 };
125 
126 /*
127  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
128  * which need extra byte of encoding.
129  * rax,rcx,...,rbp have simpler encoding
130  */
131 static bool is_ereg(u32 reg)
132 {
133 	return (1 << reg) & (BIT(BPF_REG_5) |
134 			     BIT(AUX_REG) |
135 			     BIT(BPF_REG_7) |
136 			     BIT(BPF_REG_8) |
137 			     BIT(BPF_REG_9) |
138 			     BIT(BPF_REG_AX));
139 }
140 
141 static bool is_axreg(u32 reg)
142 {
143 	return reg == BPF_REG_0;
144 }
145 
146 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
147 static u8 add_1mod(u8 byte, u32 reg)
148 {
149 	if (is_ereg(reg))
150 		byte |= 1;
151 	return byte;
152 }
153 
154 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
155 {
156 	if (is_ereg(r1))
157 		byte |= 1;
158 	if (is_ereg(r2))
159 		byte |= 4;
160 	return byte;
161 }
162 
163 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
164 static u8 add_1reg(u8 byte, u32 dst_reg)
165 {
166 	return byte + reg2hex[dst_reg];
167 }
168 
169 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
170 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
171 {
172 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
173 }
174 
175 static void jit_fill_hole(void *area, unsigned int size)
176 {
177 	/* Fill whole space with INT3 instructions */
178 	memset(area, 0xcc, size);
179 }
180 
181 struct jit_context {
182 	int cleanup_addr; /* Epilogue code offset */
183 };
184 
185 /* Maximum number of bytes emitted while JITing one eBPF insn */
186 #define BPF_MAX_INSN_SIZE	128
187 #define BPF_INSN_SAFETY		64
188 
189 #define PROLOGUE_SIZE		20
190 
191 /*
192  * Emit x86-64 prologue code for BPF program and check its size.
193  * bpf_tail_call helper will skip it while jumping into another program
194  */
195 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf)
196 {
197 	u8 *prog = *pprog;
198 	int cnt = 0;
199 
200 	EMIT1(0x55);             /* push rbp */
201 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
202 	/* sub rsp, rounded_stack_depth */
203 	EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
204 	EMIT1(0x53);             /* push rbx */
205 	EMIT2(0x41, 0x55);       /* push r13 */
206 	EMIT2(0x41, 0x56);       /* push r14 */
207 	EMIT2(0x41, 0x57);       /* push r15 */
208 	if (!ebpf_from_cbpf) {
209 		/* zero init tail_call_cnt */
210 		EMIT2(0x6a, 0x00);
211 		BUILD_BUG_ON(cnt != PROLOGUE_SIZE);
212 	}
213 	*pprog = prog;
214 }
215 
216 /*
217  * Generate the following code:
218  *
219  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
220  *   if (index >= array->map.max_entries)
221  *     goto out;
222  *   if (++tail_call_cnt > MAX_TAIL_CALL_CNT)
223  *     goto out;
224  *   prog = array->ptrs[index];
225  *   if (prog == NULL)
226  *     goto out;
227  *   goto *(prog->bpf_func + prologue_size);
228  * out:
229  */
230 static void emit_bpf_tail_call(u8 **pprog)
231 {
232 	u8 *prog = *pprog;
233 	int label1, label2, label3;
234 	int cnt = 0;
235 
236 	/*
237 	 * rdi - pointer to ctx
238 	 * rsi - pointer to bpf_array
239 	 * rdx - index in bpf_array
240 	 */
241 
242 	/*
243 	 * if (index >= array->map.max_entries)
244 	 *	goto out;
245 	 */
246 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
247 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
248 	      offsetof(struct bpf_array, map.max_entries));
249 #define OFFSET1 (41 + RETPOLINE_RAX_BPF_JIT_SIZE) /* Number of bytes to jump */
250 	EMIT2(X86_JBE, OFFSET1);                  /* jbe out */
251 	label1 = cnt;
252 
253 	/*
254 	 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
255 	 *	goto out;
256 	 */
257 	EMIT2_off32(0x8B, 0x85, -36 - MAX_BPF_STACK); /* mov eax, dword ptr [rbp - 548] */
258 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
259 #define OFFSET2 (30 + RETPOLINE_RAX_BPF_JIT_SIZE)
260 	EMIT2(X86_JA, OFFSET2);                   /* ja out */
261 	label2 = cnt;
262 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
263 	EMIT2_off32(0x89, 0x85, -36 - MAX_BPF_STACK); /* mov dword ptr [rbp -548], eax */
264 
265 	/* prog = array->ptrs[index]; */
266 	EMIT4_off32(0x48, 0x8B, 0x84, 0xD6,       /* mov rax, [rsi + rdx * 8 + offsetof(...)] */
267 		    offsetof(struct bpf_array, ptrs));
268 
269 	/*
270 	 * if (prog == NULL)
271 	 *	goto out;
272 	 */
273 	EMIT3(0x48, 0x85, 0xC0);		  /* test rax,rax */
274 #define OFFSET3 (8 + RETPOLINE_RAX_BPF_JIT_SIZE)
275 	EMIT2(X86_JE, OFFSET3);                   /* je out */
276 	label3 = cnt;
277 
278 	/* goto *(prog->bpf_func + prologue_size); */
279 	EMIT4(0x48, 0x8B, 0x40,                   /* mov rax, qword ptr [rax + 32] */
280 	      offsetof(struct bpf_prog, bpf_func));
281 	EMIT4(0x48, 0x83, 0xC0, PROLOGUE_SIZE);   /* add rax, prologue_size */
282 
283 	/*
284 	 * Wow we're ready to jump into next BPF program
285 	 * rdi == ctx (1st arg)
286 	 * rax == prog->bpf_func + prologue_size
287 	 */
288 	RETPOLINE_RAX_BPF_JIT();
289 
290 	/* out: */
291 	BUILD_BUG_ON(cnt - label1 != OFFSET1);
292 	BUILD_BUG_ON(cnt - label2 != OFFSET2);
293 	BUILD_BUG_ON(cnt - label3 != OFFSET3);
294 	*pprog = prog;
295 }
296 
297 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
298 			   u32 dst_reg, const u32 imm32)
299 {
300 	u8 *prog = *pprog;
301 	u8 b1, b2, b3;
302 	int cnt = 0;
303 
304 	/*
305 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
306 	 * (which zero-extends imm32) to save 2 bytes.
307 	 */
308 	if (sign_propagate && (s32)imm32 < 0) {
309 		/* 'mov %rax, imm32' sign extends imm32 */
310 		b1 = add_1mod(0x48, dst_reg);
311 		b2 = 0xC7;
312 		b3 = 0xC0;
313 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
314 		goto done;
315 	}
316 
317 	/*
318 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
319 	 * to save 3 bytes.
320 	 */
321 	if (imm32 == 0) {
322 		if (is_ereg(dst_reg))
323 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
324 		b2 = 0x31; /* xor */
325 		b3 = 0xC0;
326 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
327 		goto done;
328 	}
329 
330 	/* mov %eax, imm32 */
331 	if (is_ereg(dst_reg))
332 		EMIT1(add_1mod(0x40, dst_reg));
333 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
334 done:
335 	*pprog = prog;
336 }
337 
338 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
339 			   const u32 imm32_hi, const u32 imm32_lo)
340 {
341 	u8 *prog = *pprog;
342 	int cnt = 0;
343 
344 	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
345 		/*
346 		 * For emitting plain u32, where sign bit must not be
347 		 * propagated LLVM tends to load imm64 over mov32
348 		 * directly, so save couple of bytes by just doing
349 		 * 'mov %eax, imm32' instead.
350 		 */
351 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
352 	} else {
353 		/* movabsq %rax, imm64 */
354 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
355 		EMIT(imm32_lo, 4);
356 		EMIT(imm32_hi, 4);
357 	}
358 
359 	*pprog = prog;
360 }
361 
362 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
363 {
364 	u8 *prog = *pprog;
365 	int cnt = 0;
366 
367 	if (is64) {
368 		/* mov dst, src */
369 		EMIT_mov(dst_reg, src_reg);
370 	} else {
371 		/* mov32 dst, src */
372 		if (is_ereg(dst_reg) || is_ereg(src_reg))
373 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
374 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
375 	}
376 
377 	*pprog = prog;
378 }
379 
380 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
381 		  int oldproglen, struct jit_context *ctx)
382 {
383 	struct bpf_insn *insn = bpf_prog->insnsi;
384 	int insn_cnt = bpf_prog->len;
385 	bool seen_exit = false;
386 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
387 	int i, cnt = 0;
388 	int proglen = 0;
389 	u8 *prog = temp;
390 
391 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
392 		      bpf_prog_was_classic(bpf_prog));
393 	addrs[0] = prog - temp;
394 
395 	for (i = 1; i <= insn_cnt; i++, insn++) {
396 		const s32 imm32 = insn->imm;
397 		u32 dst_reg = insn->dst_reg;
398 		u32 src_reg = insn->src_reg;
399 		u8 b2 = 0, b3 = 0;
400 		s64 jmp_offset;
401 		u8 jmp_cond;
402 		int ilen;
403 		u8 *func;
404 
405 		switch (insn->code) {
406 			/* ALU */
407 		case BPF_ALU | BPF_ADD | BPF_X:
408 		case BPF_ALU | BPF_SUB | BPF_X:
409 		case BPF_ALU | BPF_AND | BPF_X:
410 		case BPF_ALU | BPF_OR | BPF_X:
411 		case BPF_ALU | BPF_XOR | BPF_X:
412 		case BPF_ALU64 | BPF_ADD | BPF_X:
413 		case BPF_ALU64 | BPF_SUB | BPF_X:
414 		case BPF_ALU64 | BPF_AND | BPF_X:
415 		case BPF_ALU64 | BPF_OR | BPF_X:
416 		case BPF_ALU64 | BPF_XOR | BPF_X:
417 			switch (BPF_OP(insn->code)) {
418 			case BPF_ADD: b2 = 0x01; break;
419 			case BPF_SUB: b2 = 0x29; break;
420 			case BPF_AND: b2 = 0x21; break;
421 			case BPF_OR: b2 = 0x09; break;
422 			case BPF_XOR: b2 = 0x31; break;
423 			}
424 			if (BPF_CLASS(insn->code) == BPF_ALU64)
425 				EMIT1(add_2mod(0x48, dst_reg, src_reg));
426 			else if (is_ereg(dst_reg) || is_ereg(src_reg))
427 				EMIT1(add_2mod(0x40, dst_reg, src_reg));
428 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
429 			break;
430 
431 		case BPF_ALU64 | BPF_MOV | BPF_X:
432 		case BPF_ALU | BPF_MOV | BPF_X:
433 			emit_mov_reg(&prog,
434 				     BPF_CLASS(insn->code) == BPF_ALU64,
435 				     dst_reg, src_reg);
436 			break;
437 
438 			/* neg dst */
439 		case BPF_ALU | BPF_NEG:
440 		case BPF_ALU64 | BPF_NEG:
441 			if (BPF_CLASS(insn->code) == BPF_ALU64)
442 				EMIT1(add_1mod(0x48, dst_reg));
443 			else if (is_ereg(dst_reg))
444 				EMIT1(add_1mod(0x40, dst_reg));
445 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
446 			break;
447 
448 		case BPF_ALU | BPF_ADD | BPF_K:
449 		case BPF_ALU | BPF_SUB | BPF_K:
450 		case BPF_ALU | BPF_AND | BPF_K:
451 		case BPF_ALU | BPF_OR | BPF_K:
452 		case BPF_ALU | BPF_XOR | BPF_K:
453 		case BPF_ALU64 | BPF_ADD | BPF_K:
454 		case BPF_ALU64 | BPF_SUB | BPF_K:
455 		case BPF_ALU64 | BPF_AND | BPF_K:
456 		case BPF_ALU64 | BPF_OR | BPF_K:
457 		case BPF_ALU64 | BPF_XOR | BPF_K:
458 			if (BPF_CLASS(insn->code) == BPF_ALU64)
459 				EMIT1(add_1mod(0x48, dst_reg));
460 			else if (is_ereg(dst_reg))
461 				EMIT1(add_1mod(0x40, dst_reg));
462 
463 			/*
464 			 * b3 holds 'normal' opcode, b2 short form only valid
465 			 * in case dst is eax/rax.
466 			 */
467 			switch (BPF_OP(insn->code)) {
468 			case BPF_ADD:
469 				b3 = 0xC0;
470 				b2 = 0x05;
471 				break;
472 			case BPF_SUB:
473 				b3 = 0xE8;
474 				b2 = 0x2D;
475 				break;
476 			case BPF_AND:
477 				b3 = 0xE0;
478 				b2 = 0x25;
479 				break;
480 			case BPF_OR:
481 				b3 = 0xC8;
482 				b2 = 0x0D;
483 				break;
484 			case BPF_XOR:
485 				b3 = 0xF0;
486 				b2 = 0x35;
487 				break;
488 			}
489 
490 			if (is_imm8(imm32))
491 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
492 			else if (is_axreg(dst_reg))
493 				EMIT1_off32(b2, imm32);
494 			else
495 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
496 			break;
497 
498 		case BPF_ALU64 | BPF_MOV | BPF_K:
499 		case BPF_ALU | BPF_MOV | BPF_K:
500 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
501 				       dst_reg, imm32);
502 			break;
503 
504 		case BPF_LD | BPF_IMM | BPF_DW:
505 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
506 			insn++;
507 			i++;
508 			break;
509 
510 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
511 		case BPF_ALU | BPF_MOD | BPF_X:
512 		case BPF_ALU | BPF_DIV | BPF_X:
513 		case BPF_ALU | BPF_MOD | BPF_K:
514 		case BPF_ALU | BPF_DIV | BPF_K:
515 		case BPF_ALU64 | BPF_MOD | BPF_X:
516 		case BPF_ALU64 | BPF_DIV | BPF_X:
517 		case BPF_ALU64 | BPF_MOD | BPF_K:
518 		case BPF_ALU64 | BPF_DIV | BPF_K:
519 			EMIT1(0x50); /* push rax */
520 			EMIT1(0x52); /* push rdx */
521 
522 			if (BPF_SRC(insn->code) == BPF_X)
523 				/* mov r11, src_reg */
524 				EMIT_mov(AUX_REG, src_reg);
525 			else
526 				/* mov r11, imm32 */
527 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
528 
529 			/* mov rax, dst_reg */
530 			EMIT_mov(BPF_REG_0, dst_reg);
531 
532 			/*
533 			 * xor edx, edx
534 			 * equivalent to 'xor rdx, rdx', but one byte less
535 			 */
536 			EMIT2(0x31, 0xd2);
537 
538 			if (BPF_CLASS(insn->code) == BPF_ALU64)
539 				/* div r11 */
540 				EMIT3(0x49, 0xF7, 0xF3);
541 			else
542 				/* div r11d */
543 				EMIT3(0x41, 0xF7, 0xF3);
544 
545 			if (BPF_OP(insn->code) == BPF_MOD)
546 				/* mov r11, rdx */
547 				EMIT3(0x49, 0x89, 0xD3);
548 			else
549 				/* mov r11, rax */
550 				EMIT3(0x49, 0x89, 0xC3);
551 
552 			EMIT1(0x5A); /* pop rdx */
553 			EMIT1(0x58); /* pop rax */
554 
555 			/* mov dst_reg, r11 */
556 			EMIT_mov(dst_reg, AUX_REG);
557 			break;
558 
559 		case BPF_ALU | BPF_MUL | BPF_K:
560 		case BPF_ALU | BPF_MUL | BPF_X:
561 		case BPF_ALU64 | BPF_MUL | BPF_K:
562 		case BPF_ALU64 | BPF_MUL | BPF_X:
563 		{
564 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
565 
566 			if (dst_reg != BPF_REG_0)
567 				EMIT1(0x50); /* push rax */
568 			if (dst_reg != BPF_REG_3)
569 				EMIT1(0x52); /* push rdx */
570 
571 			/* mov r11, dst_reg */
572 			EMIT_mov(AUX_REG, dst_reg);
573 
574 			if (BPF_SRC(insn->code) == BPF_X)
575 				emit_mov_reg(&prog, is64, BPF_REG_0, src_reg);
576 			else
577 				emit_mov_imm32(&prog, is64, BPF_REG_0, imm32);
578 
579 			if (is64)
580 				EMIT1(add_1mod(0x48, AUX_REG));
581 			else if (is_ereg(AUX_REG))
582 				EMIT1(add_1mod(0x40, AUX_REG));
583 			/* mul(q) r11 */
584 			EMIT2(0xF7, add_1reg(0xE0, AUX_REG));
585 
586 			if (dst_reg != BPF_REG_3)
587 				EMIT1(0x5A); /* pop rdx */
588 			if (dst_reg != BPF_REG_0) {
589 				/* mov dst_reg, rax */
590 				EMIT_mov(dst_reg, BPF_REG_0);
591 				EMIT1(0x58); /* pop rax */
592 			}
593 			break;
594 		}
595 			/* Shifts */
596 		case BPF_ALU | BPF_LSH | BPF_K:
597 		case BPF_ALU | BPF_RSH | BPF_K:
598 		case BPF_ALU | BPF_ARSH | BPF_K:
599 		case BPF_ALU64 | BPF_LSH | BPF_K:
600 		case BPF_ALU64 | BPF_RSH | BPF_K:
601 		case BPF_ALU64 | BPF_ARSH | BPF_K:
602 			if (BPF_CLASS(insn->code) == BPF_ALU64)
603 				EMIT1(add_1mod(0x48, dst_reg));
604 			else if (is_ereg(dst_reg))
605 				EMIT1(add_1mod(0x40, dst_reg));
606 
607 			switch (BPF_OP(insn->code)) {
608 			case BPF_LSH: b3 = 0xE0; break;
609 			case BPF_RSH: b3 = 0xE8; break;
610 			case BPF_ARSH: b3 = 0xF8; break;
611 			}
612 
613 			if (imm32 == 1)
614 				EMIT2(0xD1, add_1reg(b3, dst_reg));
615 			else
616 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
617 			break;
618 
619 		case BPF_ALU | BPF_LSH | BPF_X:
620 		case BPF_ALU | BPF_RSH | BPF_X:
621 		case BPF_ALU | BPF_ARSH | BPF_X:
622 		case BPF_ALU64 | BPF_LSH | BPF_X:
623 		case BPF_ALU64 | BPF_RSH | BPF_X:
624 		case BPF_ALU64 | BPF_ARSH | BPF_X:
625 
626 			/* Check for bad case when dst_reg == rcx */
627 			if (dst_reg == BPF_REG_4) {
628 				/* mov r11, dst_reg */
629 				EMIT_mov(AUX_REG, dst_reg);
630 				dst_reg = AUX_REG;
631 			}
632 
633 			if (src_reg != BPF_REG_4) { /* common case */
634 				EMIT1(0x51); /* push rcx */
635 
636 				/* mov rcx, src_reg */
637 				EMIT_mov(BPF_REG_4, src_reg);
638 			}
639 
640 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
641 			if (BPF_CLASS(insn->code) == BPF_ALU64)
642 				EMIT1(add_1mod(0x48, dst_reg));
643 			else if (is_ereg(dst_reg))
644 				EMIT1(add_1mod(0x40, dst_reg));
645 
646 			switch (BPF_OP(insn->code)) {
647 			case BPF_LSH: b3 = 0xE0; break;
648 			case BPF_RSH: b3 = 0xE8; break;
649 			case BPF_ARSH: b3 = 0xF8; break;
650 			}
651 			EMIT2(0xD3, add_1reg(b3, dst_reg));
652 
653 			if (src_reg != BPF_REG_4)
654 				EMIT1(0x59); /* pop rcx */
655 
656 			if (insn->dst_reg == BPF_REG_4)
657 				/* mov dst_reg, r11 */
658 				EMIT_mov(insn->dst_reg, AUX_REG);
659 			break;
660 
661 		case BPF_ALU | BPF_END | BPF_FROM_BE:
662 			switch (imm32) {
663 			case 16:
664 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
665 				EMIT1(0x66);
666 				if (is_ereg(dst_reg))
667 					EMIT1(0x41);
668 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
669 
670 				/* Emit 'movzwl eax, ax' */
671 				if (is_ereg(dst_reg))
672 					EMIT3(0x45, 0x0F, 0xB7);
673 				else
674 					EMIT2(0x0F, 0xB7);
675 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
676 				break;
677 			case 32:
678 				/* Emit 'bswap eax' to swap lower 4 bytes */
679 				if (is_ereg(dst_reg))
680 					EMIT2(0x41, 0x0F);
681 				else
682 					EMIT1(0x0F);
683 				EMIT1(add_1reg(0xC8, dst_reg));
684 				break;
685 			case 64:
686 				/* Emit 'bswap rax' to swap 8 bytes */
687 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
688 				      add_1reg(0xC8, dst_reg));
689 				break;
690 			}
691 			break;
692 
693 		case BPF_ALU | BPF_END | BPF_FROM_LE:
694 			switch (imm32) {
695 			case 16:
696 				/*
697 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
698 				 * into 64 bit
699 				 */
700 				if (is_ereg(dst_reg))
701 					EMIT3(0x45, 0x0F, 0xB7);
702 				else
703 					EMIT2(0x0F, 0xB7);
704 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
705 				break;
706 			case 32:
707 				/* Emit 'mov eax, eax' to clear upper 32-bits */
708 				if (is_ereg(dst_reg))
709 					EMIT1(0x45);
710 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
711 				break;
712 			case 64:
713 				/* nop */
714 				break;
715 			}
716 			break;
717 
718 			/* ST: *(u8*)(dst_reg + off) = imm */
719 		case BPF_ST | BPF_MEM | BPF_B:
720 			if (is_ereg(dst_reg))
721 				EMIT2(0x41, 0xC6);
722 			else
723 				EMIT1(0xC6);
724 			goto st;
725 		case BPF_ST | BPF_MEM | BPF_H:
726 			if (is_ereg(dst_reg))
727 				EMIT3(0x66, 0x41, 0xC7);
728 			else
729 				EMIT2(0x66, 0xC7);
730 			goto st;
731 		case BPF_ST | BPF_MEM | BPF_W:
732 			if (is_ereg(dst_reg))
733 				EMIT2(0x41, 0xC7);
734 			else
735 				EMIT1(0xC7);
736 			goto st;
737 		case BPF_ST | BPF_MEM | BPF_DW:
738 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
739 
740 st:			if (is_imm8(insn->off))
741 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
742 			else
743 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
744 
745 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
746 			break;
747 
748 			/* STX: *(u8*)(dst_reg + off) = src_reg */
749 		case BPF_STX | BPF_MEM | BPF_B:
750 			/* Emit 'mov byte ptr [rax + off], al' */
751 			if (is_ereg(dst_reg) || is_ereg(src_reg) ||
752 			    /* We have to add extra byte for x86 SIL, DIL regs */
753 			    src_reg == BPF_REG_1 || src_reg == BPF_REG_2)
754 				EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
755 			else
756 				EMIT1(0x88);
757 			goto stx;
758 		case BPF_STX | BPF_MEM | BPF_H:
759 			if (is_ereg(dst_reg) || is_ereg(src_reg))
760 				EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
761 			else
762 				EMIT2(0x66, 0x89);
763 			goto stx;
764 		case BPF_STX | BPF_MEM | BPF_W:
765 			if (is_ereg(dst_reg) || is_ereg(src_reg))
766 				EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
767 			else
768 				EMIT1(0x89);
769 			goto stx;
770 		case BPF_STX | BPF_MEM | BPF_DW:
771 			EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
772 stx:			if (is_imm8(insn->off))
773 				EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off);
774 			else
775 				EMIT1_off32(add_2reg(0x80, dst_reg, src_reg),
776 					    insn->off);
777 			break;
778 
779 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
780 		case BPF_LDX | BPF_MEM | BPF_B:
781 			/* Emit 'movzx rax, byte ptr [rax + off]' */
782 			EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
783 			goto ldx;
784 		case BPF_LDX | BPF_MEM | BPF_H:
785 			/* Emit 'movzx rax, word ptr [rax + off]' */
786 			EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
787 			goto ldx;
788 		case BPF_LDX | BPF_MEM | BPF_W:
789 			/* Emit 'mov eax, dword ptr [rax+0x14]' */
790 			if (is_ereg(dst_reg) || is_ereg(src_reg))
791 				EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
792 			else
793 				EMIT1(0x8B);
794 			goto ldx;
795 		case BPF_LDX | BPF_MEM | BPF_DW:
796 			/* Emit 'mov rax, qword ptr [rax+0x14]' */
797 			EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
798 ldx:			/*
799 			 * If insn->off == 0 we can save one extra byte, but
800 			 * special case of x86 R13 which always needs an offset
801 			 * is not worth the hassle
802 			 */
803 			if (is_imm8(insn->off))
804 				EMIT2(add_2reg(0x40, src_reg, dst_reg), insn->off);
805 			else
806 				EMIT1_off32(add_2reg(0x80, src_reg, dst_reg),
807 					    insn->off);
808 			break;
809 
810 			/* STX XADD: lock *(u32*)(dst_reg + off) += src_reg */
811 		case BPF_STX | BPF_XADD | BPF_W:
812 			/* Emit 'lock add dword ptr [rax + off], eax' */
813 			if (is_ereg(dst_reg) || is_ereg(src_reg))
814 				EMIT3(0xF0, add_2mod(0x40, dst_reg, src_reg), 0x01);
815 			else
816 				EMIT2(0xF0, 0x01);
817 			goto xadd;
818 		case BPF_STX | BPF_XADD | BPF_DW:
819 			EMIT3(0xF0, add_2mod(0x48, dst_reg, src_reg), 0x01);
820 xadd:			if (is_imm8(insn->off))
821 				EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off);
822 			else
823 				EMIT1_off32(add_2reg(0x80, dst_reg, src_reg),
824 					    insn->off);
825 			break;
826 
827 			/* call */
828 		case BPF_JMP | BPF_CALL:
829 			func = (u8 *) __bpf_call_base + imm32;
830 			jmp_offset = func - (image + addrs[i]);
831 			if (!imm32 || !is_simm32(jmp_offset)) {
832 				pr_err("unsupported BPF func %d addr %p image %p\n",
833 				       imm32, func, image);
834 				return -EINVAL;
835 			}
836 			EMIT1_off32(0xE8, jmp_offset);
837 			break;
838 
839 		case BPF_JMP | BPF_TAIL_CALL:
840 			emit_bpf_tail_call(&prog);
841 			break;
842 
843 			/* cond jump */
844 		case BPF_JMP | BPF_JEQ | BPF_X:
845 		case BPF_JMP | BPF_JNE | BPF_X:
846 		case BPF_JMP | BPF_JGT | BPF_X:
847 		case BPF_JMP | BPF_JLT | BPF_X:
848 		case BPF_JMP | BPF_JGE | BPF_X:
849 		case BPF_JMP | BPF_JLE | BPF_X:
850 		case BPF_JMP | BPF_JSGT | BPF_X:
851 		case BPF_JMP | BPF_JSLT | BPF_X:
852 		case BPF_JMP | BPF_JSGE | BPF_X:
853 		case BPF_JMP | BPF_JSLE | BPF_X:
854 		case BPF_JMP32 | BPF_JEQ | BPF_X:
855 		case BPF_JMP32 | BPF_JNE | BPF_X:
856 		case BPF_JMP32 | BPF_JGT | BPF_X:
857 		case BPF_JMP32 | BPF_JLT | BPF_X:
858 		case BPF_JMP32 | BPF_JGE | BPF_X:
859 		case BPF_JMP32 | BPF_JLE | BPF_X:
860 		case BPF_JMP32 | BPF_JSGT | BPF_X:
861 		case BPF_JMP32 | BPF_JSLT | BPF_X:
862 		case BPF_JMP32 | BPF_JSGE | BPF_X:
863 		case BPF_JMP32 | BPF_JSLE | BPF_X:
864 			/* cmp dst_reg, src_reg */
865 			if (BPF_CLASS(insn->code) == BPF_JMP)
866 				EMIT1(add_2mod(0x48, dst_reg, src_reg));
867 			else if (is_ereg(dst_reg) || is_ereg(src_reg))
868 				EMIT1(add_2mod(0x40, dst_reg, src_reg));
869 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
870 			goto emit_cond_jmp;
871 
872 		case BPF_JMP | BPF_JSET | BPF_X:
873 		case BPF_JMP32 | BPF_JSET | BPF_X:
874 			/* test dst_reg, src_reg */
875 			if (BPF_CLASS(insn->code) == BPF_JMP)
876 				EMIT1(add_2mod(0x48, dst_reg, src_reg));
877 			else if (is_ereg(dst_reg) || is_ereg(src_reg))
878 				EMIT1(add_2mod(0x40, dst_reg, src_reg));
879 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
880 			goto emit_cond_jmp;
881 
882 		case BPF_JMP | BPF_JSET | BPF_K:
883 		case BPF_JMP32 | BPF_JSET | BPF_K:
884 			/* test dst_reg, imm32 */
885 			if (BPF_CLASS(insn->code) == BPF_JMP)
886 				EMIT1(add_1mod(0x48, dst_reg));
887 			else if (is_ereg(dst_reg))
888 				EMIT1(add_1mod(0x40, dst_reg));
889 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
890 			goto emit_cond_jmp;
891 
892 		case BPF_JMP | BPF_JEQ | BPF_K:
893 		case BPF_JMP | BPF_JNE | BPF_K:
894 		case BPF_JMP | BPF_JGT | BPF_K:
895 		case BPF_JMP | BPF_JLT | BPF_K:
896 		case BPF_JMP | BPF_JGE | BPF_K:
897 		case BPF_JMP | BPF_JLE | BPF_K:
898 		case BPF_JMP | BPF_JSGT | BPF_K:
899 		case BPF_JMP | BPF_JSLT | BPF_K:
900 		case BPF_JMP | BPF_JSGE | BPF_K:
901 		case BPF_JMP | BPF_JSLE | BPF_K:
902 		case BPF_JMP32 | BPF_JEQ | BPF_K:
903 		case BPF_JMP32 | BPF_JNE | BPF_K:
904 		case BPF_JMP32 | BPF_JGT | BPF_K:
905 		case BPF_JMP32 | BPF_JLT | BPF_K:
906 		case BPF_JMP32 | BPF_JGE | BPF_K:
907 		case BPF_JMP32 | BPF_JLE | BPF_K:
908 		case BPF_JMP32 | BPF_JSGT | BPF_K:
909 		case BPF_JMP32 | BPF_JSLT | BPF_K:
910 		case BPF_JMP32 | BPF_JSGE | BPF_K:
911 		case BPF_JMP32 | BPF_JSLE | BPF_K:
912 			/* cmp dst_reg, imm8/32 */
913 			if (BPF_CLASS(insn->code) == BPF_JMP)
914 				EMIT1(add_1mod(0x48, dst_reg));
915 			else if (is_ereg(dst_reg))
916 				EMIT1(add_1mod(0x40, dst_reg));
917 
918 			if (is_imm8(imm32))
919 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
920 			else
921 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
922 
923 emit_cond_jmp:		/* Convert BPF opcode to x86 */
924 			switch (BPF_OP(insn->code)) {
925 			case BPF_JEQ:
926 				jmp_cond = X86_JE;
927 				break;
928 			case BPF_JSET:
929 			case BPF_JNE:
930 				jmp_cond = X86_JNE;
931 				break;
932 			case BPF_JGT:
933 				/* GT is unsigned '>', JA in x86 */
934 				jmp_cond = X86_JA;
935 				break;
936 			case BPF_JLT:
937 				/* LT is unsigned '<', JB in x86 */
938 				jmp_cond = X86_JB;
939 				break;
940 			case BPF_JGE:
941 				/* GE is unsigned '>=', JAE in x86 */
942 				jmp_cond = X86_JAE;
943 				break;
944 			case BPF_JLE:
945 				/* LE is unsigned '<=', JBE in x86 */
946 				jmp_cond = X86_JBE;
947 				break;
948 			case BPF_JSGT:
949 				/* Signed '>', GT in x86 */
950 				jmp_cond = X86_JG;
951 				break;
952 			case BPF_JSLT:
953 				/* Signed '<', LT in x86 */
954 				jmp_cond = X86_JL;
955 				break;
956 			case BPF_JSGE:
957 				/* Signed '>=', GE in x86 */
958 				jmp_cond = X86_JGE;
959 				break;
960 			case BPF_JSLE:
961 				/* Signed '<=', LE in x86 */
962 				jmp_cond = X86_JLE;
963 				break;
964 			default: /* to silence GCC warning */
965 				return -EFAULT;
966 			}
967 			jmp_offset = addrs[i + insn->off] - addrs[i];
968 			if (is_imm8(jmp_offset)) {
969 				EMIT2(jmp_cond, jmp_offset);
970 			} else if (is_simm32(jmp_offset)) {
971 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
972 			} else {
973 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
974 				return -EFAULT;
975 			}
976 
977 			break;
978 
979 		case BPF_JMP | BPF_JA:
980 			if (insn->off == -1)
981 				/* -1 jmp instructions will always jump
982 				 * backwards two bytes. Explicitly handling
983 				 * this case avoids wasting too many passes
984 				 * when there are long sequences of replaced
985 				 * dead code.
986 				 */
987 				jmp_offset = -2;
988 			else
989 				jmp_offset = addrs[i + insn->off] - addrs[i];
990 
991 			if (!jmp_offset)
992 				/* Optimize out nop jumps */
993 				break;
994 emit_jmp:
995 			if (is_imm8(jmp_offset)) {
996 				EMIT2(0xEB, jmp_offset);
997 			} else if (is_simm32(jmp_offset)) {
998 				EMIT1_off32(0xE9, jmp_offset);
999 			} else {
1000 				pr_err("jmp gen bug %llx\n", jmp_offset);
1001 				return -EFAULT;
1002 			}
1003 			break;
1004 
1005 		case BPF_JMP | BPF_EXIT:
1006 			if (seen_exit) {
1007 				jmp_offset = ctx->cleanup_addr - addrs[i];
1008 				goto emit_jmp;
1009 			}
1010 			seen_exit = true;
1011 			/* Update cleanup_addr */
1012 			ctx->cleanup_addr = proglen;
1013 			if (!bpf_prog_was_classic(bpf_prog))
1014 				EMIT1(0x5B); /* get rid of tail_call_cnt */
1015 			EMIT2(0x41, 0x5F);   /* pop r15 */
1016 			EMIT2(0x41, 0x5E);   /* pop r14 */
1017 			EMIT2(0x41, 0x5D);   /* pop r13 */
1018 			EMIT1(0x5B);         /* pop rbx */
1019 			EMIT1(0xC9);         /* leave */
1020 			EMIT1(0xC3);         /* ret */
1021 			break;
1022 
1023 		default:
1024 			/*
1025 			 * By design x86-64 JIT should support all BPF instructions.
1026 			 * This error will be seen if new instruction was added
1027 			 * to the interpreter, but not to the JIT, or if there is
1028 			 * junk in bpf_prog.
1029 			 */
1030 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1031 			return -EINVAL;
1032 		}
1033 
1034 		ilen = prog - temp;
1035 		if (ilen > BPF_MAX_INSN_SIZE) {
1036 			pr_err("bpf_jit: fatal insn size error\n");
1037 			return -EFAULT;
1038 		}
1039 
1040 		if (image) {
1041 			if (unlikely(proglen + ilen > oldproglen)) {
1042 				pr_err("bpf_jit: fatal error\n");
1043 				return -EFAULT;
1044 			}
1045 			memcpy(image + proglen, temp, ilen);
1046 		}
1047 		proglen += ilen;
1048 		addrs[i] = proglen;
1049 		prog = temp;
1050 	}
1051 	return proglen;
1052 }
1053 
1054 struct x64_jit_data {
1055 	struct bpf_binary_header *header;
1056 	int *addrs;
1057 	u8 *image;
1058 	int proglen;
1059 	struct jit_context ctx;
1060 };
1061 
1062 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1063 {
1064 	struct bpf_binary_header *header = NULL;
1065 	struct bpf_prog *tmp, *orig_prog = prog;
1066 	struct x64_jit_data *jit_data;
1067 	int proglen, oldproglen = 0;
1068 	struct jit_context ctx = {};
1069 	bool tmp_blinded = false;
1070 	bool extra_pass = false;
1071 	u8 *image = NULL;
1072 	int *addrs;
1073 	int pass;
1074 	int i;
1075 
1076 	if (!prog->jit_requested)
1077 		return orig_prog;
1078 
1079 	tmp = bpf_jit_blind_constants(prog);
1080 	/*
1081 	 * If blinding was requested and we failed during blinding,
1082 	 * we must fall back to the interpreter.
1083 	 */
1084 	if (IS_ERR(tmp))
1085 		return orig_prog;
1086 	if (tmp != prog) {
1087 		tmp_blinded = true;
1088 		prog = tmp;
1089 	}
1090 
1091 	jit_data = prog->aux->jit_data;
1092 	if (!jit_data) {
1093 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1094 		if (!jit_data) {
1095 			prog = orig_prog;
1096 			goto out;
1097 		}
1098 		prog->aux->jit_data = jit_data;
1099 	}
1100 	addrs = jit_data->addrs;
1101 	if (addrs) {
1102 		ctx = jit_data->ctx;
1103 		oldproglen = jit_data->proglen;
1104 		image = jit_data->image;
1105 		header = jit_data->header;
1106 		extra_pass = true;
1107 		goto skip_init_addrs;
1108 	}
1109 	addrs = kmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
1110 	if (!addrs) {
1111 		prog = orig_prog;
1112 		goto out_addrs;
1113 	}
1114 
1115 	/*
1116 	 * Before first pass, make a rough estimation of addrs[]
1117 	 * each BPF instruction is translated to less than 64 bytes
1118 	 */
1119 	for (proglen = 0, i = 0; i <= prog->len; i++) {
1120 		proglen += 64;
1121 		addrs[i] = proglen;
1122 	}
1123 	ctx.cleanup_addr = proglen;
1124 skip_init_addrs:
1125 
1126 	/*
1127 	 * JITed image shrinks with every pass and the loop iterates
1128 	 * until the image stops shrinking. Very large BPF programs
1129 	 * may converge on the last pass. In such case do one more
1130 	 * pass to emit the final image.
1131 	 */
1132 	for (pass = 0; pass < 20 || image; pass++) {
1133 		proglen = do_jit(prog, addrs, image, oldproglen, &ctx);
1134 		if (proglen <= 0) {
1135 out_image:
1136 			image = NULL;
1137 			if (header)
1138 				bpf_jit_binary_free(header);
1139 			prog = orig_prog;
1140 			goto out_addrs;
1141 		}
1142 		if (image) {
1143 			if (proglen != oldproglen) {
1144 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
1145 				       proglen, oldproglen);
1146 				goto out_image;
1147 			}
1148 			break;
1149 		}
1150 		if (proglen == oldproglen) {
1151 			header = bpf_jit_binary_alloc(proglen, &image,
1152 						      1, jit_fill_hole);
1153 			if (!header) {
1154 				prog = orig_prog;
1155 				goto out_addrs;
1156 			}
1157 		}
1158 		oldproglen = proglen;
1159 		cond_resched();
1160 	}
1161 
1162 	if (bpf_jit_enable > 1)
1163 		bpf_jit_dump(prog->len, proglen, pass + 1, image);
1164 
1165 	if (image) {
1166 		if (!prog->is_func || extra_pass) {
1167 			bpf_jit_binary_lock_ro(header);
1168 		} else {
1169 			jit_data->addrs = addrs;
1170 			jit_data->ctx = ctx;
1171 			jit_data->proglen = proglen;
1172 			jit_data->image = image;
1173 			jit_data->header = header;
1174 		}
1175 		prog->bpf_func = (void *)image;
1176 		prog->jited = 1;
1177 		prog->jited_len = proglen;
1178 	} else {
1179 		prog = orig_prog;
1180 	}
1181 
1182 	if (!image || !prog->is_func || extra_pass) {
1183 		if (image)
1184 			bpf_prog_fill_jited_linfo(prog, addrs + 1);
1185 out_addrs:
1186 		kfree(addrs);
1187 		kfree(jit_data);
1188 		prog->aux->jit_data = NULL;
1189 	}
1190 out:
1191 	if (tmp_blinded)
1192 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
1193 					   tmp : orig_prog);
1194 	return prog;
1195 }
1196