xref: /linux/arch/x86/net/bpf_jit_comp.c (revision ec63e2a4897075e427c121d863bd89c44578094f)
1 /*
2  * bpf_jit_comp.c: BPF JIT compiler
3  *
4  * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
5  * Internal BPF Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
6  *
7  * This program is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU General Public License
9  * as published by the Free Software Foundation; version 2
10  * of the License.
11  */
12 #include <linux/netdevice.h>
13 #include <linux/filter.h>
14 #include <linux/if_vlan.h>
15 #include <linux/bpf.h>
16 
17 #include <asm/set_memory.h>
18 #include <asm/nospec-branch.h>
19 
20 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
21 {
22 	if (len == 1)
23 		*ptr = bytes;
24 	else if (len == 2)
25 		*(u16 *)ptr = bytes;
26 	else {
27 		*(u32 *)ptr = bytes;
28 		barrier();
29 	}
30 	return ptr + len;
31 }
32 
33 #define EMIT(bytes, len) \
34 	do { prog = emit_code(prog, bytes, len); cnt += len; } while (0)
35 
36 #define EMIT1(b1)		EMIT(b1, 1)
37 #define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
38 #define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
39 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
40 
41 #define EMIT1_off32(b1, off) \
42 	do { EMIT1(b1); EMIT(off, 4); } while (0)
43 #define EMIT2_off32(b1, b2, off) \
44 	do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
45 #define EMIT3_off32(b1, b2, b3, off) \
46 	do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
47 #define EMIT4_off32(b1, b2, b3, b4, off) \
48 	do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
49 
50 static bool is_imm8(int value)
51 {
52 	return value <= 127 && value >= -128;
53 }
54 
55 static bool is_simm32(s64 value)
56 {
57 	return value == (s64)(s32)value;
58 }
59 
60 static bool is_uimm32(u64 value)
61 {
62 	return value == (u64)(u32)value;
63 }
64 
65 /* mov dst, src */
66 #define EMIT_mov(DST, SRC)								 \
67 	do {										 \
68 		if (DST != SRC)								 \
69 			EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
70 	} while (0)
71 
72 static int bpf_size_to_x86_bytes(int bpf_size)
73 {
74 	if (bpf_size == BPF_W)
75 		return 4;
76 	else if (bpf_size == BPF_H)
77 		return 2;
78 	else if (bpf_size == BPF_B)
79 		return 1;
80 	else if (bpf_size == BPF_DW)
81 		return 4; /* imm32 */
82 	else
83 		return 0;
84 }
85 
86 /*
87  * List of x86 cond jumps opcodes (. + s8)
88  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
89  */
90 #define X86_JB  0x72
91 #define X86_JAE 0x73
92 #define X86_JE  0x74
93 #define X86_JNE 0x75
94 #define X86_JBE 0x76
95 #define X86_JA  0x77
96 #define X86_JL  0x7C
97 #define X86_JGE 0x7D
98 #define X86_JLE 0x7E
99 #define X86_JG  0x7F
100 
101 /* Pick a register outside of BPF range for JIT internal work */
102 #define AUX_REG (MAX_BPF_JIT_REG + 1)
103 
104 /*
105  * The following table maps BPF registers to x86-64 registers.
106  *
107  * x86-64 register R12 is unused, since if used as base address
108  * register in load/store instructions, it always needs an
109  * extra byte of encoding and is callee saved.
110  *
111  * Also x86-64 register R9 is unused. x86-64 register R10 is
112  * used for blinding (if enabled).
113  */
114 static const int reg2hex[] = {
115 	[BPF_REG_0] = 0,  /* RAX */
116 	[BPF_REG_1] = 7,  /* RDI */
117 	[BPF_REG_2] = 6,  /* RSI */
118 	[BPF_REG_3] = 2,  /* RDX */
119 	[BPF_REG_4] = 1,  /* RCX */
120 	[BPF_REG_5] = 0,  /* R8  */
121 	[BPF_REG_6] = 3,  /* RBX callee saved */
122 	[BPF_REG_7] = 5,  /* R13 callee saved */
123 	[BPF_REG_8] = 6,  /* R14 callee saved */
124 	[BPF_REG_9] = 7,  /* R15 callee saved */
125 	[BPF_REG_FP] = 5, /* RBP readonly */
126 	[BPF_REG_AX] = 2, /* R10 temp register */
127 	[AUX_REG] = 3,    /* R11 temp register */
128 };
129 
130 /*
131  * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
132  * which need extra byte of encoding.
133  * rax,rcx,...,rbp have simpler encoding
134  */
135 static bool is_ereg(u32 reg)
136 {
137 	return (1 << reg) & (BIT(BPF_REG_5) |
138 			     BIT(AUX_REG) |
139 			     BIT(BPF_REG_7) |
140 			     BIT(BPF_REG_8) |
141 			     BIT(BPF_REG_9) |
142 			     BIT(BPF_REG_AX));
143 }
144 
145 static bool is_axreg(u32 reg)
146 {
147 	return reg == BPF_REG_0;
148 }
149 
150 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
151 static u8 add_1mod(u8 byte, u32 reg)
152 {
153 	if (is_ereg(reg))
154 		byte |= 1;
155 	return byte;
156 }
157 
158 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
159 {
160 	if (is_ereg(r1))
161 		byte |= 1;
162 	if (is_ereg(r2))
163 		byte |= 4;
164 	return byte;
165 }
166 
167 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
168 static u8 add_1reg(u8 byte, u32 dst_reg)
169 {
170 	return byte + reg2hex[dst_reg];
171 }
172 
173 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
174 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
175 {
176 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
177 }
178 
179 static void jit_fill_hole(void *area, unsigned int size)
180 {
181 	/* Fill whole space with INT3 instructions */
182 	memset(area, 0xcc, size);
183 }
184 
185 struct jit_context {
186 	int cleanup_addr; /* Epilogue code offset */
187 };
188 
189 /* Maximum number of bytes emitted while JITing one eBPF insn */
190 #define BPF_MAX_INSN_SIZE	128
191 #define BPF_INSN_SAFETY		64
192 
193 #define AUX_STACK_SPACE		40 /* Space for RBX, R13, R14, R15, tailcnt */
194 
195 #define PROLOGUE_SIZE		37
196 
197 /*
198  * Emit x86-64 prologue code for BPF program and check its size.
199  * bpf_tail_call helper will skip it while jumping into another program
200  */
201 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf)
202 {
203 	u8 *prog = *pprog;
204 	int cnt = 0;
205 
206 	/* push rbp */
207 	EMIT1(0x55);
208 
209 	/* mov rbp,rsp */
210 	EMIT3(0x48, 0x89, 0xE5);
211 
212 	/* sub rsp, rounded_stack_depth + AUX_STACK_SPACE */
213 	EMIT3_off32(0x48, 0x81, 0xEC,
214 		    round_up(stack_depth, 8) + AUX_STACK_SPACE);
215 
216 	/* sub rbp, AUX_STACK_SPACE */
217 	EMIT4(0x48, 0x83, 0xED, AUX_STACK_SPACE);
218 
219 	/* mov qword ptr [rbp+0],rbx */
220 	EMIT4(0x48, 0x89, 0x5D, 0);
221 	/* mov qword ptr [rbp+8],r13 */
222 	EMIT4(0x4C, 0x89, 0x6D, 8);
223 	/* mov qword ptr [rbp+16],r14 */
224 	EMIT4(0x4C, 0x89, 0x75, 16);
225 	/* mov qword ptr [rbp+24],r15 */
226 	EMIT4(0x4C, 0x89, 0x7D, 24);
227 
228 	if (!ebpf_from_cbpf) {
229 		/*
230 		 * Clear the tail call counter (tail_call_cnt): for eBPF tail
231 		 * calls we need to reset the counter to 0. It's done in two
232 		 * instructions, resetting RAX register to 0, and moving it
233 		 * to the counter location.
234 		 */
235 
236 		/* xor eax, eax */
237 		EMIT2(0x31, 0xc0);
238 		/* mov qword ptr [rbp+32], rax */
239 		EMIT4(0x48, 0x89, 0x45, 32);
240 
241 		BUILD_BUG_ON(cnt != PROLOGUE_SIZE);
242 	}
243 
244 	*pprog = prog;
245 }
246 
247 /*
248  * Generate the following code:
249  *
250  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
251  *   if (index >= array->map.max_entries)
252  *     goto out;
253  *   if (++tail_call_cnt > MAX_TAIL_CALL_CNT)
254  *     goto out;
255  *   prog = array->ptrs[index];
256  *   if (prog == NULL)
257  *     goto out;
258  *   goto *(prog->bpf_func + prologue_size);
259  * out:
260  */
261 static void emit_bpf_tail_call(u8 **pprog)
262 {
263 	u8 *prog = *pprog;
264 	int label1, label2, label3;
265 	int cnt = 0;
266 
267 	/*
268 	 * rdi - pointer to ctx
269 	 * rsi - pointer to bpf_array
270 	 * rdx - index in bpf_array
271 	 */
272 
273 	/*
274 	 * if (index >= array->map.max_entries)
275 	 *	goto out;
276 	 */
277 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
278 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
279 	      offsetof(struct bpf_array, map.max_entries));
280 #define OFFSET1 (41 + RETPOLINE_RAX_BPF_JIT_SIZE) /* Number of bytes to jump */
281 	EMIT2(X86_JBE, OFFSET1);                  /* jbe out */
282 	label1 = cnt;
283 
284 	/*
285 	 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
286 	 *	goto out;
287 	 */
288 	EMIT2_off32(0x8B, 0x85, 36);              /* mov eax, dword ptr [rbp + 36] */
289 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
290 #define OFFSET2 (30 + RETPOLINE_RAX_BPF_JIT_SIZE)
291 	EMIT2(X86_JA, OFFSET2);                   /* ja out */
292 	label2 = cnt;
293 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
294 	EMIT2_off32(0x89, 0x85, 36);              /* mov dword ptr [rbp + 36], eax */
295 
296 	/* prog = array->ptrs[index]; */
297 	EMIT4_off32(0x48, 0x8B, 0x84, 0xD6,       /* mov rax, [rsi + rdx * 8 + offsetof(...)] */
298 		    offsetof(struct bpf_array, ptrs));
299 
300 	/*
301 	 * if (prog == NULL)
302 	 *	goto out;
303 	 */
304 	EMIT3(0x48, 0x85, 0xC0);		  /* test rax,rax */
305 #define OFFSET3 (8 + RETPOLINE_RAX_BPF_JIT_SIZE)
306 	EMIT2(X86_JE, OFFSET3);                   /* je out */
307 	label3 = cnt;
308 
309 	/* goto *(prog->bpf_func + prologue_size); */
310 	EMIT4(0x48, 0x8B, 0x40,                   /* mov rax, qword ptr [rax + 32] */
311 	      offsetof(struct bpf_prog, bpf_func));
312 	EMIT4(0x48, 0x83, 0xC0, PROLOGUE_SIZE);   /* add rax, prologue_size */
313 
314 	/*
315 	 * Wow we're ready to jump into next BPF program
316 	 * rdi == ctx (1st arg)
317 	 * rax == prog->bpf_func + prologue_size
318 	 */
319 	RETPOLINE_RAX_BPF_JIT();
320 
321 	/* out: */
322 	BUILD_BUG_ON(cnt - label1 != OFFSET1);
323 	BUILD_BUG_ON(cnt - label2 != OFFSET2);
324 	BUILD_BUG_ON(cnt - label3 != OFFSET3);
325 	*pprog = prog;
326 }
327 
328 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
329 			   u32 dst_reg, const u32 imm32)
330 {
331 	u8 *prog = *pprog;
332 	u8 b1, b2, b3;
333 	int cnt = 0;
334 
335 	/*
336 	 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
337 	 * (which zero-extends imm32) to save 2 bytes.
338 	 */
339 	if (sign_propagate && (s32)imm32 < 0) {
340 		/* 'mov %rax, imm32' sign extends imm32 */
341 		b1 = add_1mod(0x48, dst_reg);
342 		b2 = 0xC7;
343 		b3 = 0xC0;
344 		EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
345 		goto done;
346 	}
347 
348 	/*
349 	 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
350 	 * to save 3 bytes.
351 	 */
352 	if (imm32 == 0) {
353 		if (is_ereg(dst_reg))
354 			EMIT1(add_2mod(0x40, dst_reg, dst_reg));
355 		b2 = 0x31; /* xor */
356 		b3 = 0xC0;
357 		EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
358 		goto done;
359 	}
360 
361 	/* mov %eax, imm32 */
362 	if (is_ereg(dst_reg))
363 		EMIT1(add_1mod(0x40, dst_reg));
364 	EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
365 done:
366 	*pprog = prog;
367 }
368 
369 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
370 			   const u32 imm32_hi, const u32 imm32_lo)
371 {
372 	u8 *prog = *pprog;
373 	int cnt = 0;
374 
375 	if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
376 		/*
377 		 * For emitting plain u32, where sign bit must not be
378 		 * propagated LLVM tends to load imm64 over mov32
379 		 * directly, so save couple of bytes by just doing
380 		 * 'mov %eax, imm32' instead.
381 		 */
382 		emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
383 	} else {
384 		/* movabsq %rax, imm64 */
385 		EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
386 		EMIT(imm32_lo, 4);
387 		EMIT(imm32_hi, 4);
388 	}
389 
390 	*pprog = prog;
391 }
392 
393 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
394 {
395 	u8 *prog = *pprog;
396 	int cnt = 0;
397 
398 	if (is64) {
399 		/* mov dst, src */
400 		EMIT_mov(dst_reg, src_reg);
401 	} else {
402 		/* mov32 dst, src */
403 		if (is_ereg(dst_reg) || is_ereg(src_reg))
404 			EMIT1(add_2mod(0x40, dst_reg, src_reg));
405 		EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
406 	}
407 
408 	*pprog = prog;
409 }
410 
411 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
412 		  int oldproglen, struct jit_context *ctx)
413 {
414 	struct bpf_insn *insn = bpf_prog->insnsi;
415 	int insn_cnt = bpf_prog->len;
416 	bool seen_exit = false;
417 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
418 	int i, cnt = 0;
419 	int proglen = 0;
420 	u8 *prog = temp;
421 
422 	emit_prologue(&prog, bpf_prog->aux->stack_depth,
423 		      bpf_prog_was_classic(bpf_prog));
424 
425 	for (i = 0; i < insn_cnt; i++, insn++) {
426 		const s32 imm32 = insn->imm;
427 		u32 dst_reg = insn->dst_reg;
428 		u32 src_reg = insn->src_reg;
429 		u8 b2 = 0, b3 = 0;
430 		s64 jmp_offset;
431 		u8 jmp_cond;
432 		int ilen;
433 		u8 *func;
434 
435 		switch (insn->code) {
436 			/* ALU */
437 		case BPF_ALU | BPF_ADD | BPF_X:
438 		case BPF_ALU | BPF_SUB | BPF_X:
439 		case BPF_ALU | BPF_AND | BPF_X:
440 		case BPF_ALU | BPF_OR | BPF_X:
441 		case BPF_ALU | BPF_XOR | BPF_X:
442 		case BPF_ALU64 | BPF_ADD | BPF_X:
443 		case BPF_ALU64 | BPF_SUB | BPF_X:
444 		case BPF_ALU64 | BPF_AND | BPF_X:
445 		case BPF_ALU64 | BPF_OR | BPF_X:
446 		case BPF_ALU64 | BPF_XOR | BPF_X:
447 			switch (BPF_OP(insn->code)) {
448 			case BPF_ADD: b2 = 0x01; break;
449 			case BPF_SUB: b2 = 0x29; break;
450 			case BPF_AND: b2 = 0x21; break;
451 			case BPF_OR: b2 = 0x09; break;
452 			case BPF_XOR: b2 = 0x31; break;
453 			}
454 			if (BPF_CLASS(insn->code) == BPF_ALU64)
455 				EMIT1(add_2mod(0x48, dst_reg, src_reg));
456 			else if (is_ereg(dst_reg) || is_ereg(src_reg))
457 				EMIT1(add_2mod(0x40, dst_reg, src_reg));
458 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
459 			break;
460 
461 		case BPF_ALU64 | BPF_MOV | BPF_X:
462 		case BPF_ALU | BPF_MOV | BPF_X:
463 			emit_mov_reg(&prog,
464 				     BPF_CLASS(insn->code) == BPF_ALU64,
465 				     dst_reg, src_reg);
466 			break;
467 
468 			/* neg dst */
469 		case BPF_ALU | BPF_NEG:
470 		case BPF_ALU64 | BPF_NEG:
471 			if (BPF_CLASS(insn->code) == BPF_ALU64)
472 				EMIT1(add_1mod(0x48, dst_reg));
473 			else if (is_ereg(dst_reg))
474 				EMIT1(add_1mod(0x40, dst_reg));
475 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
476 			break;
477 
478 		case BPF_ALU | BPF_ADD | BPF_K:
479 		case BPF_ALU | BPF_SUB | BPF_K:
480 		case BPF_ALU | BPF_AND | BPF_K:
481 		case BPF_ALU | BPF_OR | BPF_K:
482 		case BPF_ALU | BPF_XOR | BPF_K:
483 		case BPF_ALU64 | BPF_ADD | BPF_K:
484 		case BPF_ALU64 | BPF_SUB | BPF_K:
485 		case BPF_ALU64 | BPF_AND | BPF_K:
486 		case BPF_ALU64 | BPF_OR | BPF_K:
487 		case BPF_ALU64 | BPF_XOR | BPF_K:
488 			if (BPF_CLASS(insn->code) == BPF_ALU64)
489 				EMIT1(add_1mod(0x48, dst_reg));
490 			else if (is_ereg(dst_reg))
491 				EMIT1(add_1mod(0x40, dst_reg));
492 
493 			/*
494 			 * b3 holds 'normal' opcode, b2 short form only valid
495 			 * in case dst is eax/rax.
496 			 */
497 			switch (BPF_OP(insn->code)) {
498 			case BPF_ADD:
499 				b3 = 0xC0;
500 				b2 = 0x05;
501 				break;
502 			case BPF_SUB:
503 				b3 = 0xE8;
504 				b2 = 0x2D;
505 				break;
506 			case BPF_AND:
507 				b3 = 0xE0;
508 				b2 = 0x25;
509 				break;
510 			case BPF_OR:
511 				b3 = 0xC8;
512 				b2 = 0x0D;
513 				break;
514 			case BPF_XOR:
515 				b3 = 0xF0;
516 				b2 = 0x35;
517 				break;
518 			}
519 
520 			if (is_imm8(imm32))
521 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
522 			else if (is_axreg(dst_reg))
523 				EMIT1_off32(b2, imm32);
524 			else
525 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
526 			break;
527 
528 		case BPF_ALU64 | BPF_MOV | BPF_K:
529 		case BPF_ALU | BPF_MOV | BPF_K:
530 			emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
531 				       dst_reg, imm32);
532 			break;
533 
534 		case BPF_LD | BPF_IMM | BPF_DW:
535 			emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
536 			insn++;
537 			i++;
538 			break;
539 
540 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
541 		case BPF_ALU | BPF_MOD | BPF_X:
542 		case BPF_ALU | BPF_DIV | BPF_X:
543 		case BPF_ALU | BPF_MOD | BPF_K:
544 		case BPF_ALU | BPF_DIV | BPF_K:
545 		case BPF_ALU64 | BPF_MOD | BPF_X:
546 		case BPF_ALU64 | BPF_DIV | BPF_X:
547 		case BPF_ALU64 | BPF_MOD | BPF_K:
548 		case BPF_ALU64 | BPF_DIV | BPF_K:
549 			EMIT1(0x50); /* push rax */
550 			EMIT1(0x52); /* push rdx */
551 
552 			if (BPF_SRC(insn->code) == BPF_X)
553 				/* mov r11, src_reg */
554 				EMIT_mov(AUX_REG, src_reg);
555 			else
556 				/* mov r11, imm32 */
557 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
558 
559 			/* mov rax, dst_reg */
560 			EMIT_mov(BPF_REG_0, dst_reg);
561 
562 			/*
563 			 * xor edx, edx
564 			 * equivalent to 'xor rdx, rdx', but one byte less
565 			 */
566 			EMIT2(0x31, 0xd2);
567 
568 			if (BPF_CLASS(insn->code) == BPF_ALU64)
569 				/* div r11 */
570 				EMIT3(0x49, 0xF7, 0xF3);
571 			else
572 				/* div r11d */
573 				EMIT3(0x41, 0xF7, 0xF3);
574 
575 			if (BPF_OP(insn->code) == BPF_MOD)
576 				/* mov r11, rdx */
577 				EMIT3(0x49, 0x89, 0xD3);
578 			else
579 				/* mov r11, rax */
580 				EMIT3(0x49, 0x89, 0xC3);
581 
582 			EMIT1(0x5A); /* pop rdx */
583 			EMIT1(0x58); /* pop rax */
584 
585 			/* mov dst_reg, r11 */
586 			EMIT_mov(dst_reg, AUX_REG);
587 			break;
588 
589 		case BPF_ALU | BPF_MUL | BPF_K:
590 		case BPF_ALU | BPF_MUL | BPF_X:
591 		case BPF_ALU64 | BPF_MUL | BPF_K:
592 		case BPF_ALU64 | BPF_MUL | BPF_X:
593 		{
594 			bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
595 
596 			if (dst_reg != BPF_REG_0)
597 				EMIT1(0x50); /* push rax */
598 			if (dst_reg != BPF_REG_3)
599 				EMIT1(0x52); /* push rdx */
600 
601 			/* mov r11, dst_reg */
602 			EMIT_mov(AUX_REG, dst_reg);
603 
604 			if (BPF_SRC(insn->code) == BPF_X)
605 				emit_mov_reg(&prog, is64, BPF_REG_0, src_reg);
606 			else
607 				emit_mov_imm32(&prog, is64, BPF_REG_0, imm32);
608 
609 			if (is64)
610 				EMIT1(add_1mod(0x48, AUX_REG));
611 			else if (is_ereg(AUX_REG))
612 				EMIT1(add_1mod(0x40, AUX_REG));
613 			/* mul(q) r11 */
614 			EMIT2(0xF7, add_1reg(0xE0, AUX_REG));
615 
616 			if (dst_reg != BPF_REG_3)
617 				EMIT1(0x5A); /* pop rdx */
618 			if (dst_reg != BPF_REG_0) {
619 				/* mov dst_reg, rax */
620 				EMIT_mov(dst_reg, BPF_REG_0);
621 				EMIT1(0x58); /* pop rax */
622 			}
623 			break;
624 		}
625 			/* Shifts */
626 		case BPF_ALU | BPF_LSH | BPF_K:
627 		case BPF_ALU | BPF_RSH | BPF_K:
628 		case BPF_ALU | BPF_ARSH | BPF_K:
629 		case BPF_ALU64 | BPF_LSH | BPF_K:
630 		case BPF_ALU64 | BPF_RSH | BPF_K:
631 		case BPF_ALU64 | BPF_ARSH | BPF_K:
632 			if (BPF_CLASS(insn->code) == BPF_ALU64)
633 				EMIT1(add_1mod(0x48, dst_reg));
634 			else if (is_ereg(dst_reg))
635 				EMIT1(add_1mod(0x40, dst_reg));
636 
637 			switch (BPF_OP(insn->code)) {
638 			case BPF_LSH: b3 = 0xE0; break;
639 			case BPF_RSH: b3 = 0xE8; break;
640 			case BPF_ARSH: b3 = 0xF8; break;
641 			}
642 
643 			if (imm32 == 1)
644 				EMIT2(0xD1, add_1reg(b3, dst_reg));
645 			else
646 				EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
647 			break;
648 
649 		case BPF_ALU | BPF_LSH | BPF_X:
650 		case BPF_ALU | BPF_RSH | BPF_X:
651 		case BPF_ALU | BPF_ARSH | BPF_X:
652 		case BPF_ALU64 | BPF_LSH | BPF_X:
653 		case BPF_ALU64 | BPF_RSH | BPF_X:
654 		case BPF_ALU64 | BPF_ARSH | BPF_X:
655 
656 			/* Check for bad case when dst_reg == rcx */
657 			if (dst_reg == BPF_REG_4) {
658 				/* mov r11, dst_reg */
659 				EMIT_mov(AUX_REG, dst_reg);
660 				dst_reg = AUX_REG;
661 			}
662 
663 			if (src_reg != BPF_REG_4) { /* common case */
664 				EMIT1(0x51); /* push rcx */
665 
666 				/* mov rcx, src_reg */
667 				EMIT_mov(BPF_REG_4, src_reg);
668 			}
669 
670 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
671 			if (BPF_CLASS(insn->code) == BPF_ALU64)
672 				EMIT1(add_1mod(0x48, dst_reg));
673 			else if (is_ereg(dst_reg))
674 				EMIT1(add_1mod(0x40, dst_reg));
675 
676 			switch (BPF_OP(insn->code)) {
677 			case BPF_LSH: b3 = 0xE0; break;
678 			case BPF_RSH: b3 = 0xE8; break;
679 			case BPF_ARSH: b3 = 0xF8; break;
680 			}
681 			EMIT2(0xD3, add_1reg(b3, dst_reg));
682 
683 			if (src_reg != BPF_REG_4)
684 				EMIT1(0x59); /* pop rcx */
685 
686 			if (insn->dst_reg == BPF_REG_4)
687 				/* mov dst_reg, r11 */
688 				EMIT_mov(insn->dst_reg, AUX_REG);
689 			break;
690 
691 		case BPF_ALU | BPF_END | BPF_FROM_BE:
692 			switch (imm32) {
693 			case 16:
694 				/* Emit 'ror %ax, 8' to swap lower 2 bytes */
695 				EMIT1(0x66);
696 				if (is_ereg(dst_reg))
697 					EMIT1(0x41);
698 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
699 
700 				/* Emit 'movzwl eax, ax' */
701 				if (is_ereg(dst_reg))
702 					EMIT3(0x45, 0x0F, 0xB7);
703 				else
704 					EMIT2(0x0F, 0xB7);
705 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
706 				break;
707 			case 32:
708 				/* Emit 'bswap eax' to swap lower 4 bytes */
709 				if (is_ereg(dst_reg))
710 					EMIT2(0x41, 0x0F);
711 				else
712 					EMIT1(0x0F);
713 				EMIT1(add_1reg(0xC8, dst_reg));
714 				break;
715 			case 64:
716 				/* Emit 'bswap rax' to swap 8 bytes */
717 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
718 				      add_1reg(0xC8, dst_reg));
719 				break;
720 			}
721 			break;
722 
723 		case BPF_ALU | BPF_END | BPF_FROM_LE:
724 			switch (imm32) {
725 			case 16:
726 				/*
727 				 * Emit 'movzwl eax, ax' to zero extend 16-bit
728 				 * into 64 bit
729 				 */
730 				if (is_ereg(dst_reg))
731 					EMIT3(0x45, 0x0F, 0xB7);
732 				else
733 					EMIT2(0x0F, 0xB7);
734 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
735 				break;
736 			case 32:
737 				/* Emit 'mov eax, eax' to clear upper 32-bits */
738 				if (is_ereg(dst_reg))
739 					EMIT1(0x45);
740 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
741 				break;
742 			case 64:
743 				/* nop */
744 				break;
745 			}
746 			break;
747 
748 			/* ST: *(u8*)(dst_reg + off) = imm */
749 		case BPF_ST | BPF_MEM | BPF_B:
750 			if (is_ereg(dst_reg))
751 				EMIT2(0x41, 0xC6);
752 			else
753 				EMIT1(0xC6);
754 			goto st;
755 		case BPF_ST | BPF_MEM | BPF_H:
756 			if (is_ereg(dst_reg))
757 				EMIT3(0x66, 0x41, 0xC7);
758 			else
759 				EMIT2(0x66, 0xC7);
760 			goto st;
761 		case BPF_ST | BPF_MEM | BPF_W:
762 			if (is_ereg(dst_reg))
763 				EMIT2(0x41, 0xC7);
764 			else
765 				EMIT1(0xC7);
766 			goto st;
767 		case BPF_ST | BPF_MEM | BPF_DW:
768 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
769 
770 st:			if (is_imm8(insn->off))
771 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
772 			else
773 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
774 
775 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
776 			break;
777 
778 			/* STX: *(u8*)(dst_reg + off) = src_reg */
779 		case BPF_STX | BPF_MEM | BPF_B:
780 			/* Emit 'mov byte ptr [rax + off], al' */
781 			if (is_ereg(dst_reg) || is_ereg(src_reg) ||
782 			    /* We have to add extra byte for x86 SIL, DIL regs */
783 			    src_reg == BPF_REG_1 || src_reg == BPF_REG_2)
784 				EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
785 			else
786 				EMIT1(0x88);
787 			goto stx;
788 		case BPF_STX | BPF_MEM | BPF_H:
789 			if (is_ereg(dst_reg) || is_ereg(src_reg))
790 				EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
791 			else
792 				EMIT2(0x66, 0x89);
793 			goto stx;
794 		case BPF_STX | BPF_MEM | BPF_W:
795 			if (is_ereg(dst_reg) || is_ereg(src_reg))
796 				EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
797 			else
798 				EMIT1(0x89);
799 			goto stx;
800 		case BPF_STX | BPF_MEM | BPF_DW:
801 			EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
802 stx:			if (is_imm8(insn->off))
803 				EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off);
804 			else
805 				EMIT1_off32(add_2reg(0x80, dst_reg, src_reg),
806 					    insn->off);
807 			break;
808 
809 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
810 		case BPF_LDX | BPF_MEM | BPF_B:
811 			/* Emit 'movzx rax, byte ptr [rax + off]' */
812 			EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
813 			goto ldx;
814 		case BPF_LDX | BPF_MEM | BPF_H:
815 			/* Emit 'movzx rax, word ptr [rax + off]' */
816 			EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
817 			goto ldx;
818 		case BPF_LDX | BPF_MEM | BPF_W:
819 			/* Emit 'mov eax, dword ptr [rax+0x14]' */
820 			if (is_ereg(dst_reg) || is_ereg(src_reg))
821 				EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
822 			else
823 				EMIT1(0x8B);
824 			goto ldx;
825 		case BPF_LDX | BPF_MEM | BPF_DW:
826 			/* Emit 'mov rax, qword ptr [rax+0x14]' */
827 			EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
828 ldx:			/*
829 			 * If insn->off == 0 we can save one extra byte, but
830 			 * special case of x86 R13 which always needs an offset
831 			 * is not worth the hassle
832 			 */
833 			if (is_imm8(insn->off))
834 				EMIT2(add_2reg(0x40, src_reg, dst_reg), insn->off);
835 			else
836 				EMIT1_off32(add_2reg(0x80, src_reg, dst_reg),
837 					    insn->off);
838 			break;
839 
840 			/* STX XADD: lock *(u32*)(dst_reg + off) += src_reg */
841 		case BPF_STX | BPF_XADD | BPF_W:
842 			/* Emit 'lock add dword ptr [rax + off], eax' */
843 			if (is_ereg(dst_reg) || is_ereg(src_reg))
844 				EMIT3(0xF0, add_2mod(0x40, dst_reg, src_reg), 0x01);
845 			else
846 				EMIT2(0xF0, 0x01);
847 			goto xadd;
848 		case BPF_STX | BPF_XADD | BPF_DW:
849 			EMIT3(0xF0, add_2mod(0x48, dst_reg, src_reg), 0x01);
850 xadd:			if (is_imm8(insn->off))
851 				EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off);
852 			else
853 				EMIT1_off32(add_2reg(0x80, dst_reg, src_reg),
854 					    insn->off);
855 			break;
856 
857 			/* call */
858 		case BPF_JMP | BPF_CALL:
859 			func = (u8 *) __bpf_call_base + imm32;
860 			jmp_offset = func - (image + addrs[i]);
861 			if (!imm32 || !is_simm32(jmp_offset)) {
862 				pr_err("unsupported BPF func %d addr %p image %p\n",
863 				       imm32, func, image);
864 				return -EINVAL;
865 			}
866 			EMIT1_off32(0xE8, jmp_offset);
867 			break;
868 
869 		case BPF_JMP | BPF_TAIL_CALL:
870 			emit_bpf_tail_call(&prog);
871 			break;
872 
873 			/* cond jump */
874 		case BPF_JMP | BPF_JEQ | BPF_X:
875 		case BPF_JMP | BPF_JNE | BPF_X:
876 		case BPF_JMP | BPF_JGT | BPF_X:
877 		case BPF_JMP | BPF_JLT | BPF_X:
878 		case BPF_JMP | BPF_JGE | BPF_X:
879 		case BPF_JMP | BPF_JLE | BPF_X:
880 		case BPF_JMP | BPF_JSGT | BPF_X:
881 		case BPF_JMP | BPF_JSLT | BPF_X:
882 		case BPF_JMP | BPF_JSGE | BPF_X:
883 		case BPF_JMP | BPF_JSLE | BPF_X:
884 		case BPF_JMP32 | BPF_JEQ | BPF_X:
885 		case BPF_JMP32 | BPF_JNE | BPF_X:
886 		case BPF_JMP32 | BPF_JGT | BPF_X:
887 		case BPF_JMP32 | BPF_JLT | BPF_X:
888 		case BPF_JMP32 | BPF_JGE | BPF_X:
889 		case BPF_JMP32 | BPF_JLE | BPF_X:
890 		case BPF_JMP32 | BPF_JSGT | BPF_X:
891 		case BPF_JMP32 | BPF_JSLT | BPF_X:
892 		case BPF_JMP32 | BPF_JSGE | BPF_X:
893 		case BPF_JMP32 | BPF_JSLE | BPF_X:
894 			/* cmp dst_reg, src_reg */
895 			if (BPF_CLASS(insn->code) == BPF_JMP)
896 				EMIT1(add_2mod(0x48, dst_reg, src_reg));
897 			else if (is_ereg(dst_reg) || is_ereg(src_reg))
898 				EMIT1(add_2mod(0x40, dst_reg, src_reg));
899 			EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
900 			goto emit_cond_jmp;
901 
902 		case BPF_JMP | BPF_JSET | BPF_X:
903 		case BPF_JMP32 | BPF_JSET | BPF_X:
904 			/* test dst_reg, src_reg */
905 			if (BPF_CLASS(insn->code) == BPF_JMP)
906 				EMIT1(add_2mod(0x48, dst_reg, src_reg));
907 			else if (is_ereg(dst_reg) || is_ereg(src_reg))
908 				EMIT1(add_2mod(0x40, dst_reg, src_reg));
909 			EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
910 			goto emit_cond_jmp;
911 
912 		case BPF_JMP | BPF_JSET | BPF_K:
913 		case BPF_JMP32 | BPF_JSET | BPF_K:
914 			/* test dst_reg, imm32 */
915 			if (BPF_CLASS(insn->code) == BPF_JMP)
916 				EMIT1(add_1mod(0x48, dst_reg));
917 			else if (is_ereg(dst_reg))
918 				EMIT1(add_1mod(0x40, dst_reg));
919 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
920 			goto emit_cond_jmp;
921 
922 		case BPF_JMP | BPF_JEQ | BPF_K:
923 		case BPF_JMP | BPF_JNE | BPF_K:
924 		case BPF_JMP | BPF_JGT | BPF_K:
925 		case BPF_JMP | BPF_JLT | BPF_K:
926 		case BPF_JMP | BPF_JGE | BPF_K:
927 		case BPF_JMP | BPF_JLE | BPF_K:
928 		case BPF_JMP | BPF_JSGT | BPF_K:
929 		case BPF_JMP | BPF_JSLT | BPF_K:
930 		case BPF_JMP | BPF_JSGE | BPF_K:
931 		case BPF_JMP | BPF_JSLE | BPF_K:
932 		case BPF_JMP32 | BPF_JEQ | BPF_K:
933 		case BPF_JMP32 | BPF_JNE | BPF_K:
934 		case BPF_JMP32 | BPF_JGT | BPF_K:
935 		case BPF_JMP32 | BPF_JLT | BPF_K:
936 		case BPF_JMP32 | BPF_JGE | BPF_K:
937 		case BPF_JMP32 | BPF_JLE | BPF_K:
938 		case BPF_JMP32 | BPF_JSGT | BPF_K:
939 		case BPF_JMP32 | BPF_JSLT | BPF_K:
940 		case BPF_JMP32 | BPF_JSGE | BPF_K:
941 		case BPF_JMP32 | BPF_JSLE | BPF_K:
942 			/* cmp dst_reg, imm8/32 */
943 			if (BPF_CLASS(insn->code) == BPF_JMP)
944 				EMIT1(add_1mod(0x48, dst_reg));
945 			else if (is_ereg(dst_reg))
946 				EMIT1(add_1mod(0x40, dst_reg));
947 
948 			if (is_imm8(imm32))
949 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
950 			else
951 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
952 
953 emit_cond_jmp:		/* Convert BPF opcode to x86 */
954 			switch (BPF_OP(insn->code)) {
955 			case BPF_JEQ:
956 				jmp_cond = X86_JE;
957 				break;
958 			case BPF_JSET:
959 			case BPF_JNE:
960 				jmp_cond = X86_JNE;
961 				break;
962 			case BPF_JGT:
963 				/* GT is unsigned '>', JA in x86 */
964 				jmp_cond = X86_JA;
965 				break;
966 			case BPF_JLT:
967 				/* LT is unsigned '<', JB in x86 */
968 				jmp_cond = X86_JB;
969 				break;
970 			case BPF_JGE:
971 				/* GE is unsigned '>=', JAE in x86 */
972 				jmp_cond = X86_JAE;
973 				break;
974 			case BPF_JLE:
975 				/* LE is unsigned '<=', JBE in x86 */
976 				jmp_cond = X86_JBE;
977 				break;
978 			case BPF_JSGT:
979 				/* Signed '>', GT in x86 */
980 				jmp_cond = X86_JG;
981 				break;
982 			case BPF_JSLT:
983 				/* Signed '<', LT in x86 */
984 				jmp_cond = X86_JL;
985 				break;
986 			case BPF_JSGE:
987 				/* Signed '>=', GE in x86 */
988 				jmp_cond = X86_JGE;
989 				break;
990 			case BPF_JSLE:
991 				/* Signed '<=', LE in x86 */
992 				jmp_cond = X86_JLE;
993 				break;
994 			default: /* to silence GCC warning */
995 				return -EFAULT;
996 			}
997 			jmp_offset = addrs[i + insn->off] - addrs[i];
998 			if (is_imm8(jmp_offset)) {
999 				EMIT2(jmp_cond, jmp_offset);
1000 			} else if (is_simm32(jmp_offset)) {
1001 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1002 			} else {
1003 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1004 				return -EFAULT;
1005 			}
1006 
1007 			break;
1008 
1009 		case BPF_JMP | BPF_JA:
1010 			if (insn->off == -1)
1011 				/* -1 jmp instructions will always jump
1012 				 * backwards two bytes. Explicitly handling
1013 				 * this case avoids wasting too many passes
1014 				 * when there are long sequences of replaced
1015 				 * dead code.
1016 				 */
1017 				jmp_offset = -2;
1018 			else
1019 				jmp_offset = addrs[i + insn->off] - addrs[i];
1020 
1021 			if (!jmp_offset)
1022 				/* Optimize out nop jumps */
1023 				break;
1024 emit_jmp:
1025 			if (is_imm8(jmp_offset)) {
1026 				EMIT2(0xEB, jmp_offset);
1027 			} else if (is_simm32(jmp_offset)) {
1028 				EMIT1_off32(0xE9, jmp_offset);
1029 			} else {
1030 				pr_err("jmp gen bug %llx\n", jmp_offset);
1031 				return -EFAULT;
1032 			}
1033 			break;
1034 
1035 		case BPF_JMP | BPF_EXIT:
1036 			if (seen_exit) {
1037 				jmp_offset = ctx->cleanup_addr - addrs[i];
1038 				goto emit_jmp;
1039 			}
1040 			seen_exit = true;
1041 			/* Update cleanup_addr */
1042 			ctx->cleanup_addr = proglen;
1043 			/* mov rbx, qword ptr [rbp+0] */
1044 			EMIT4(0x48, 0x8B, 0x5D, 0);
1045 			/* mov r13, qword ptr [rbp+8] */
1046 			EMIT4(0x4C, 0x8B, 0x6D, 8);
1047 			/* mov r14, qword ptr [rbp+16] */
1048 			EMIT4(0x4C, 0x8B, 0x75, 16);
1049 			/* mov r15, qword ptr [rbp+24] */
1050 			EMIT4(0x4C, 0x8B, 0x7D, 24);
1051 
1052 			/* add rbp, AUX_STACK_SPACE */
1053 			EMIT4(0x48, 0x83, 0xC5, AUX_STACK_SPACE);
1054 			EMIT1(0xC9); /* leave */
1055 			EMIT1(0xC3); /* ret */
1056 			break;
1057 
1058 		default:
1059 			/*
1060 			 * By design x86-64 JIT should support all BPF instructions.
1061 			 * This error will be seen if new instruction was added
1062 			 * to the interpreter, but not to the JIT, or if there is
1063 			 * junk in bpf_prog.
1064 			 */
1065 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1066 			return -EINVAL;
1067 		}
1068 
1069 		ilen = prog - temp;
1070 		if (ilen > BPF_MAX_INSN_SIZE) {
1071 			pr_err("bpf_jit: fatal insn size error\n");
1072 			return -EFAULT;
1073 		}
1074 
1075 		if (image) {
1076 			if (unlikely(proglen + ilen > oldproglen)) {
1077 				pr_err("bpf_jit: fatal error\n");
1078 				return -EFAULT;
1079 			}
1080 			memcpy(image + proglen, temp, ilen);
1081 		}
1082 		proglen += ilen;
1083 		addrs[i] = proglen;
1084 		prog = temp;
1085 	}
1086 	return proglen;
1087 }
1088 
1089 struct x64_jit_data {
1090 	struct bpf_binary_header *header;
1091 	int *addrs;
1092 	u8 *image;
1093 	int proglen;
1094 	struct jit_context ctx;
1095 };
1096 
1097 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1098 {
1099 	struct bpf_binary_header *header = NULL;
1100 	struct bpf_prog *tmp, *orig_prog = prog;
1101 	struct x64_jit_data *jit_data;
1102 	int proglen, oldproglen = 0;
1103 	struct jit_context ctx = {};
1104 	bool tmp_blinded = false;
1105 	bool extra_pass = false;
1106 	u8 *image = NULL;
1107 	int *addrs;
1108 	int pass;
1109 	int i;
1110 
1111 	if (!prog->jit_requested)
1112 		return orig_prog;
1113 
1114 	tmp = bpf_jit_blind_constants(prog);
1115 	/*
1116 	 * If blinding was requested and we failed during blinding,
1117 	 * we must fall back to the interpreter.
1118 	 */
1119 	if (IS_ERR(tmp))
1120 		return orig_prog;
1121 	if (tmp != prog) {
1122 		tmp_blinded = true;
1123 		prog = tmp;
1124 	}
1125 
1126 	jit_data = prog->aux->jit_data;
1127 	if (!jit_data) {
1128 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1129 		if (!jit_data) {
1130 			prog = orig_prog;
1131 			goto out;
1132 		}
1133 		prog->aux->jit_data = jit_data;
1134 	}
1135 	addrs = jit_data->addrs;
1136 	if (addrs) {
1137 		ctx = jit_data->ctx;
1138 		oldproglen = jit_data->proglen;
1139 		image = jit_data->image;
1140 		header = jit_data->header;
1141 		extra_pass = true;
1142 		goto skip_init_addrs;
1143 	}
1144 	addrs = kmalloc_array(prog->len, sizeof(*addrs), GFP_KERNEL);
1145 	if (!addrs) {
1146 		prog = orig_prog;
1147 		goto out_addrs;
1148 	}
1149 
1150 	/*
1151 	 * Before first pass, make a rough estimation of addrs[]
1152 	 * each BPF instruction is translated to less than 64 bytes
1153 	 */
1154 	for (proglen = 0, i = 0; i < prog->len; i++) {
1155 		proglen += 64;
1156 		addrs[i] = proglen;
1157 	}
1158 	ctx.cleanup_addr = proglen;
1159 skip_init_addrs:
1160 
1161 	/*
1162 	 * JITed image shrinks with every pass and the loop iterates
1163 	 * until the image stops shrinking. Very large BPF programs
1164 	 * may converge on the last pass. In such case do one more
1165 	 * pass to emit the final image.
1166 	 */
1167 	for (pass = 0; pass < 20 || image; pass++) {
1168 		proglen = do_jit(prog, addrs, image, oldproglen, &ctx);
1169 		if (proglen <= 0) {
1170 out_image:
1171 			image = NULL;
1172 			if (header)
1173 				bpf_jit_binary_free(header);
1174 			prog = orig_prog;
1175 			goto out_addrs;
1176 		}
1177 		if (image) {
1178 			if (proglen != oldproglen) {
1179 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
1180 				       proglen, oldproglen);
1181 				goto out_image;
1182 			}
1183 			break;
1184 		}
1185 		if (proglen == oldproglen) {
1186 			header = bpf_jit_binary_alloc(proglen, &image,
1187 						      1, jit_fill_hole);
1188 			if (!header) {
1189 				prog = orig_prog;
1190 				goto out_addrs;
1191 			}
1192 		}
1193 		oldproglen = proglen;
1194 		cond_resched();
1195 	}
1196 
1197 	if (bpf_jit_enable > 1)
1198 		bpf_jit_dump(prog->len, proglen, pass + 1, image);
1199 
1200 	if (image) {
1201 		if (!prog->is_func || extra_pass) {
1202 			bpf_jit_binary_lock_ro(header);
1203 		} else {
1204 			jit_data->addrs = addrs;
1205 			jit_data->ctx = ctx;
1206 			jit_data->proglen = proglen;
1207 			jit_data->image = image;
1208 			jit_data->header = header;
1209 		}
1210 		prog->bpf_func = (void *)image;
1211 		prog->jited = 1;
1212 		prog->jited_len = proglen;
1213 	} else {
1214 		prog = orig_prog;
1215 	}
1216 
1217 	if (!image || !prog->is_func || extra_pass) {
1218 		if (image)
1219 			bpf_prog_fill_jited_linfo(prog, addrs);
1220 out_addrs:
1221 		kfree(addrs);
1222 		kfree(jit_data);
1223 		prog->aux->jit_data = NULL;
1224 	}
1225 out:
1226 	if (tmp_blinded)
1227 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
1228 					   tmp : orig_prog);
1229 	return prog;
1230 }
1231