xref: /openbmc/linux/arch/x86/net/bpf_jit_comp.c (revision 7f2e85840871f199057e65232ebde846192ed989)
1 /* bpf_jit_comp.c : BPF JIT compiler
2  *
3  * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
4  * Internal BPF Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * as published by the Free Software Foundation; version 2
9  * of the License.
10  */
11 #include <linux/netdevice.h>
12 #include <linux/filter.h>
13 #include <linux/if_vlan.h>
14 #include <asm/cacheflush.h>
15 #include <asm/set_memory.h>
16 #include <linux/bpf.h>
17 
18 /*
19  * assembly code in arch/x86/net/bpf_jit.S
20  */
21 extern u8 sk_load_word[], sk_load_half[], sk_load_byte[];
22 extern u8 sk_load_word_positive_offset[], sk_load_half_positive_offset[];
23 extern u8 sk_load_byte_positive_offset[];
24 extern u8 sk_load_word_negative_offset[], sk_load_half_negative_offset[];
25 extern u8 sk_load_byte_negative_offset[];
26 
27 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
28 {
29 	if (len == 1)
30 		*ptr = bytes;
31 	else if (len == 2)
32 		*(u16 *)ptr = bytes;
33 	else {
34 		*(u32 *)ptr = bytes;
35 		barrier();
36 	}
37 	return ptr + len;
38 }
39 
40 #define EMIT(bytes, len) \
41 	do { prog = emit_code(prog, bytes, len); cnt += len; } while (0)
42 
43 #define EMIT1(b1)		EMIT(b1, 1)
44 #define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
45 #define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
46 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
47 #define EMIT1_off32(b1, off) \
48 	do {EMIT1(b1); EMIT(off, 4); } while (0)
49 #define EMIT2_off32(b1, b2, off) \
50 	do {EMIT2(b1, b2); EMIT(off, 4); } while (0)
51 #define EMIT3_off32(b1, b2, b3, off) \
52 	do {EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
53 #define EMIT4_off32(b1, b2, b3, b4, off) \
54 	do {EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
55 
56 static bool is_imm8(int value)
57 {
58 	return value <= 127 && value >= -128;
59 }
60 
61 static bool is_simm32(s64 value)
62 {
63 	return value == (s64) (s32) value;
64 }
65 
66 /* mov dst, src */
67 #define EMIT_mov(DST, SRC) \
68 	do {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 /* list of x86 cond jumps opcodes (. + s8)
87  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
88  */
89 #define X86_JB  0x72
90 #define X86_JAE 0x73
91 #define X86_JE  0x74
92 #define X86_JNE 0x75
93 #define X86_JBE 0x76
94 #define X86_JA  0x77
95 #define X86_JL  0x7C
96 #define X86_JGE 0x7D
97 #define X86_JLE 0x7E
98 #define X86_JG  0x7F
99 
100 static void bpf_flush_icache(void *start, void *end)
101 {
102 	mm_segment_t old_fs = get_fs();
103 
104 	set_fs(KERNEL_DS);
105 	smp_wmb();
106 	flush_icache_range((unsigned long)start, (unsigned long)end);
107 	set_fs(old_fs);
108 }
109 
110 #define CHOOSE_LOAD_FUNC(K, func) \
111 	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
112 
113 /* pick a register outside of BPF range for JIT internal work */
114 #define AUX_REG (MAX_BPF_JIT_REG + 1)
115 
116 /* The following table maps BPF registers to x64 registers.
117  *
118  * x64 register r12 is unused, since if used as base address
119  * register in load/store instructions, it always needs an
120  * extra byte of encoding and is callee saved.
121  *
122  *  r9 caches skb->len - skb->data_len
123  * r10 caches skb->data, and used for blinding (if enabled)
124  */
125 static const int reg2hex[] = {
126 	[BPF_REG_0] = 0,  /* rax */
127 	[BPF_REG_1] = 7,  /* rdi */
128 	[BPF_REG_2] = 6,  /* rsi */
129 	[BPF_REG_3] = 2,  /* rdx */
130 	[BPF_REG_4] = 1,  /* rcx */
131 	[BPF_REG_5] = 0,  /* r8 */
132 	[BPF_REG_6] = 3,  /* rbx callee saved */
133 	[BPF_REG_7] = 5,  /* r13 callee saved */
134 	[BPF_REG_8] = 6,  /* r14 callee saved */
135 	[BPF_REG_9] = 7,  /* r15 callee saved */
136 	[BPF_REG_FP] = 5, /* rbp readonly */
137 	[BPF_REG_AX] = 2, /* r10 temp register */
138 	[AUX_REG] = 3,    /* r11 temp register */
139 };
140 
141 /* is_ereg() == true if BPF register 'reg' maps to x64 r8..r15
142  * which need extra byte of encoding.
143  * rax,rcx,...,rbp have simpler encoding
144  */
145 static bool is_ereg(u32 reg)
146 {
147 	return (1 << reg) & (BIT(BPF_REG_5) |
148 			     BIT(AUX_REG) |
149 			     BIT(BPF_REG_7) |
150 			     BIT(BPF_REG_8) |
151 			     BIT(BPF_REG_9) |
152 			     BIT(BPF_REG_AX));
153 }
154 
155 static bool is_axreg(u32 reg)
156 {
157 	return reg == BPF_REG_0;
158 }
159 
160 /* add modifiers if 'reg' maps to x64 registers r8..r15 */
161 static u8 add_1mod(u8 byte, u32 reg)
162 {
163 	if (is_ereg(reg))
164 		byte |= 1;
165 	return byte;
166 }
167 
168 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
169 {
170 	if (is_ereg(r1))
171 		byte |= 1;
172 	if (is_ereg(r2))
173 		byte |= 4;
174 	return byte;
175 }
176 
177 /* encode 'dst_reg' register into x64 opcode 'byte' */
178 static u8 add_1reg(u8 byte, u32 dst_reg)
179 {
180 	return byte + reg2hex[dst_reg];
181 }
182 
183 /* encode 'dst_reg' and 'src_reg' registers into x64 opcode 'byte' */
184 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
185 {
186 	return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
187 }
188 
189 static void jit_fill_hole(void *area, unsigned int size)
190 {
191 	/* fill whole space with int3 instructions */
192 	memset(area, 0xcc, size);
193 }
194 
195 struct jit_context {
196 	int cleanup_addr; /* epilogue code offset */
197 	bool seen_ld_abs;
198 	bool seen_ax_reg;
199 };
200 
201 /* maximum number of bytes emitted while JITing one eBPF insn */
202 #define BPF_MAX_INSN_SIZE	128
203 #define BPF_INSN_SAFETY		64
204 
205 #define AUX_STACK_SPACE \
206 	(32 /* space for rbx, r13, r14, r15 */ + \
207 	 8 /* space for skb_copy_bits() buffer */)
208 
209 #define PROLOGUE_SIZE 37
210 
211 /* emit x64 prologue code for BPF program and check it's size.
212  * bpf_tail_call helper will skip it while jumping into another program
213  */
214 static void emit_prologue(u8 **pprog, u32 stack_depth)
215 {
216 	u8 *prog = *pprog;
217 	int cnt = 0;
218 
219 	EMIT1(0x55); /* push rbp */
220 	EMIT3(0x48, 0x89, 0xE5); /* mov rbp,rsp */
221 
222 	/* sub rsp, rounded_stack_depth + AUX_STACK_SPACE */
223 	EMIT3_off32(0x48, 0x81, 0xEC,
224 		    round_up(stack_depth, 8) + AUX_STACK_SPACE);
225 
226 	/* sub rbp, AUX_STACK_SPACE */
227 	EMIT4(0x48, 0x83, 0xED, AUX_STACK_SPACE);
228 
229 	/* all classic BPF filters use R6(rbx) save it */
230 
231 	/* mov qword ptr [rbp+0],rbx */
232 	EMIT4(0x48, 0x89, 0x5D, 0);
233 
234 	/* bpf_convert_filter() maps classic BPF register X to R7 and uses R8
235 	 * as temporary, so all tcpdump filters need to spill/fill R7(r13) and
236 	 * R8(r14). R9(r15) spill could be made conditional, but there is only
237 	 * one 'bpf_error' return path out of helper functions inside bpf_jit.S
238 	 * The overhead of extra spill is negligible for any filter other
239 	 * than synthetic ones. Therefore not worth adding complexity.
240 	 */
241 
242 	/* mov qword ptr [rbp+8],r13 */
243 	EMIT4(0x4C, 0x89, 0x6D, 8);
244 	/* mov qword ptr [rbp+16],r14 */
245 	EMIT4(0x4C, 0x89, 0x75, 16);
246 	/* mov qword ptr [rbp+24],r15 */
247 	EMIT4(0x4C, 0x89, 0x7D, 24);
248 
249 	/* Clear the tail call counter (tail_call_cnt): for eBPF tail calls
250 	 * we need to reset the counter to 0. It's done in two instructions,
251 	 * resetting rax register to 0 (xor on eax gets 0 extended), and
252 	 * moving it to the counter location.
253 	 */
254 
255 	/* xor eax, eax */
256 	EMIT2(0x31, 0xc0);
257 	/* mov qword ptr [rbp+32], rax */
258 	EMIT4(0x48, 0x89, 0x45, 32);
259 
260 	BUILD_BUG_ON(cnt != PROLOGUE_SIZE);
261 	*pprog = prog;
262 }
263 
264 /* generate the following code:
265  * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
266  *   if (index >= array->map.max_entries)
267  *     goto out;
268  *   if (++tail_call_cnt > MAX_TAIL_CALL_CNT)
269  *     goto out;
270  *   prog = array->ptrs[index];
271  *   if (prog == NULL)
272  *     goto out;
273  *   goto *(prog->bpf_func + prologue_size);
274  * out:
275  */
276 static void emit_bpf_tail_call(u8 **pprog)
277 {
278 	u8 *prog = *pprog;
279 	int label1, label2, label3;
280 	int cnt = 0;
281 
282 	/* rdi - pointer to ctx
283 	 * rsi - pointer to bpf_array
284 	 * rdx - index in bpf_array
285 	 */
286 
287 	/* if (index >= array->map.max_entries)
288 	 *   goto out;
289 	 */
290 	EMIT2(0x89, 0xD2);                        /* mov edx, edx */
291 	EMIT3(0x39, 0x56,                         /* cmp dword ptr [rsi + 16], edx */
292 	      offsetof(struct bpf_array, map.max_entries));
293 #define OFFSET1 43 /* number of bytes to jump */
294 	EMIT2(X86_JBE, OFFSET1);                  /* jbe out */
295 	label1 = cnt;
296 
297 	/* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
298 	 *   goto out;
299 	 */
300 	EMIT2_off32(0x8B, 0x85, 36);              /* mov eax, dword ptr [rbp + 36] */
301 	EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT);     /* cmp eax, MAX_TAIL_CALL_CNT */
302 #define OFFSET2 32
303 	EMIT2(X86_JA, OFFSET2);                   /* ja out */
304 	label2 = cnt;
305 	EMIT3(0x83, 0xC0, 0x01);                  /* add eax, 1 */
306 	EMIT2_off32(0x89, 0x85, 36);              /* mov dword ptr [rbp + 36], eax */
307 
308 	/* prog = array->ptrs[index]; */
309 	EMIT4_off32(0x48, 0x8B, 0x84, 0xD6,       /* mov rax, [rsi + rdx * 8 + offsetof(...)] */
310 		    offsetof(struct bpf_array, ptrs));
311 
312 	/* if (prog == NULL)
313 	 *   goto out;
314 	 */
315 	EMIT3(0x48, 0x85, 0xC0);		  /* test rax,rax */
316 #define OFFSET3 10
317 	EMIT2(X86_JE, OFFSET3);                   /* je out */
318 	label3 = cnt;
319 
320 	/* goto *(prog->bpf_func + prologue_size); */
321 	EMIT4(0x48, 0x8B, 0x40,                   /* mov rax, qword ptr [rax + 32] */
322 	      offsetof(struct bpf_prog, bpf_func));
323 	EMIT4(0x48, 0x83, 0xC0, PROLOGUE_SIZE);   /* add rax, prologue_size */
324 
325 	/* now we're ready to jump into next BPF program
326 	 * rdi == ctx (1st arg)
327 	 * rax == prog->bpf_func + prologue_size
328 	 */
329 	EMIT2(0xFF, 0xE0);                        /* jmp rax */
330 
331 	/* out: */
332 	BUILD_BUG_ON(cnt - label1 != OFFSET1);
333 	BUILD_BUG_ON(cnt - label2 != OFFSET2);
334 	BUILD_BUG_ON(cnt - label3 != OFFSET3);
335 	*pprog = prog;
336 }
337 
338 
339 static void emit_load_skb_data_hlen(u8 **pprog)
340 {
341 	u8 *prog = *pprog;
342 	int cnt = 0;
343 
344 	/* r9d = skb->len - skb->data_len (headlen)
345 	 * r10 = skb->data
346 	 */
347 	/* mov %r9d, off32(%rdi) */
348 	EMIT3_off32(0x44, 0x8b, 0x8f, offsetof(struct sk_buff, len));
349 
350 	/* sub %r9d, off32(%rdi) */
351 	EMIT3_off32(0x44, 0x2b, 0x8f, offsetof(struct sk_buff, data_len));
352 
353 	/* mov %r10, off32(%rdi) */
354 	EMIT3_off32(0x4c, 0x8b, 0x97, offsetof(struct sk_buff, data));
355 	*pprog = prog;
356 }
357 
358 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
359 		  int oldproglen, struct jit_context *ctx)
360 {
361 	struct bpf_insn *insn = bpf_prog->insnsi;
362 	int insn_cnt = bpf_prog->len;
363 	bool seen_ld_abs = ctx->seen_ld_abs | (oldproglen == 0);
364 	bool seen_ax_reg = ctx->seen_ax_reg | (oldproglen == 0);
365 	bool seen_exit = false;
366 	u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
367 	int i, cnt = 0;
368 	int proglen = 0;
369 	u8 *prog = temp;
370 
371 	emit_prologue(&prog, bpf_prog->aux->stack_depth);
372 
373 	if (seen_ld_abs)
374 		emit_load_skb_data_hlen(&prog);
375 
376 	for (i = 0; i < insn_cnt; i++, insn++) {
377 		const s32 imm32 = insn->imm;
378 		u32 dst_reg = insn->dst_reg;
379 		u32 src_reg = insn->src_reg;
380 		u8 b1 = 0, b2 = 0, b3 = 0;
381 		s64 jmp_offset;
382 		u8 jmp_cond;
383 		bool reload_skb_data;
384 		int ilen;
385 		u8 *func;
386 
387 		if (dst_reg == BPF_REG_AX || src_reg == BPF_REG_AX)
388 			ctx->seen_ax_reg = seen_ax_reg = true;
389 
390 		switch (insn->code) {
391 			/* ALU */
392 		case BPF_ALU | BPF_ADD | BPF_X:
393 		case BPF_ALU | BPF_SUB | BPF_X:
394 		case BPF_ALU | BPF_AND | BPF_X:
395 		case BPF_ALU | BPF_OR | BPF_X:
396 		case BPF_ALU | BPF_XOR | BPF_X:
397 		case BPF_ALU64 | BPF_ADD | BPF_X:
398 		case BPF_ALU64 | BPF_SUB | BPF_X:
399 		case BPF_ALU64 | BPF_AND | BPF_X:
400 		case BPF_ALU64 | BPF_OR | BPF_X:
401 		case BPF_ALU64 | BPF_XOR | BPF_X:
402 			switch (BPF_OP(insn->code)) {
403 			case BPF_ADD: b2 = 0x01; break;
404 			case BPF_SUB: b2 = 0x29; break;
405 			case BPF_AND: b2 = 0x21; break;
406 			case BPF_OR: b2 = 0x09; break;
407 			case BPF_XOR: b2 = 0x31; break;
408 			}
409 			if (BPF_CLASS(insn->code) == BPF_ALU64)
410 				EMIT1(add_2mod(0x48, dst_reg, src_reg));
411 			else if (is_ereg(dst_reg) || is_ereg(src_reg))
412 				EMIT1(add_2mod(0x40, dst_reg, src_reg));
413 			EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
414 			break;
415 
416 			/* mov dst, src */
417 		case BPF_ALU64 | BPF_MOV | BPF_X:
418 			EMIT_mov(dst_reg, src_reg);
419 			break;
420 
421 			/* mov32 dst, src */
422 		case BPF_ALU | BPF_MOV | BPF_X:
423 			if (is_ereg(dst_reg) || is_ereg(src_reg))
424 				EMIT1(add_2mod(0x40, dst_reg, src_reg));
425 			EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
426 			break;
427 
428 			/* neg dst */
429 		case BPF_ALU | BPF_NEG:
430 		case BPF_ALU64 | BPF_NEG:
431 			if (BPF_CLASS(insn->code) == BPF_ALU64)
432 				EMIT1(add_1mod(0x48, dst_reg));
433 			else if (is_ereg(dst_reg))
434 				EMIT1(add_1mod(0x40, dst_reg));
435 			EMIT2(0xF7, add_1reg(0xD8, dst_reg));
436 			break;
437 
438 		case BPF_ALU | BPF_ADD | BPF_K:
439 		case BPF_ALU | BPF_SUB | BPF_K:
440 		case BPF_ALU | BPF_AND | BPF_K:
441 		case BPF_ALU | BPF_OR | BPF_K:
442 		case BPF_ALU | BPF_XOR | BPF_K:
443 		case BPF_ALU64 | BPF_ADD | BPF_K:
444 		case BPF_ALU64 | BPF_SUB | BPF_K:
445 		case BPF_ALU64 | BPF_AND | BPF_K:
446 		case BPF_ALU64 | BPF_OR | BPF_K:
447 		case BPF_ALU64 | BPF_XOR | BPF_K:
448 			if (BPF_CLASS(insn->code) == BPF_ALU64)
449 				EMIT1(add_1mod(0x48, dst_reg));
450 			else if (is_ereg(dst_reg))
451 				EMIT1(add_1mod(0x40, dst_reg));
452 
453 			/* b3 holds 'normal' opcode, b2 short form only valid
454 			 * in case dst is eax/rax.
455 			 */
456 			switch (BPF_OP(insn->code)) {
457 			case BPF_ADD:
458 				b3 = 0xC0;
459 				b2 = 0x05;
460 				break;
461 			case BPF_SUB:
462 				b3 = 0xE8;
463 				b2 = 0x2D;
464 				break;
465 			case BPF_AND:
466 				b3 = 0xE0;
467 				b2 = 0x25;
468 				break;
469 			case BPF_OR:
470 				b3 = 0xC8;
471 				b2 = 0x0D;
472 				break;
473 			case BPF_XOR:
474 				b3 = 0xF0;
475 				b2 = 0x35;
476 				break;
477 			}
478 
479 			if (is_imm8(imm32))
480 				EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
481 			else if (is_axreg(dst_reg))
482 				EMIT1_off32(b2, imm32);
483 			else
484 				EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
485 			break;
486 
487 		case BPF_ALU64 | BPF_MOV | BPF_K:
488 			/* optimization: if imm32 is positive,
489 			 * use 'mov eax, imm32' (which zero-extends imm32)
490 			 * to save 2 bytes
491 			 */
492 			if (imm32 < 0) {
493 				/* 'mov rax, imm32' sign extends imm32 */
494 				b1 = add_1mod(0x48, dst_reg);
495 				b2 = 0xC7;
496 				b3 = 0xC0;
497 				EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
498 				break;
499 			}
500 
501 		case BPF_ALU | BPF_MOV | BPF_K:
502 			/* optimization: if imm32 is zero, use 'xor <dst>,<dst>'
503 			 * to save 3 bytes.
504 			 */
505 			if (imm32 == 0) {
506 				if (is_ereg(dst_reg))
507 					EMIT1(add_2mod(0x40, dst_reg, dst_reg));
508 				b2 = 0x31; /* xor */
509 				b3 = 0xC0;
510 				EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
511 				break;
512 			}
513 
514 			/* mov %eax, imm32 */
515 			if (is_ereg(dst_reg))
516 				EMIT1(add_1mod(0x40, dst_reg));
517 			EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
518 			break;
519 
520 		case BPF_LD | BPF_IMM | BPF_DW:
521 			/* optimization: if imm64 is zero, use 'xor <dst>,<dst>'
522 			 * to save 7 bytes.
523 			 */
524 			if (insn[0].imm == 0 && insn[1].imm == 0) {
525 				b1 = add_2mod(0x48, dst_reg, dst_reg);
526 				b2 = 0x31; /* xor */
527 				b3 = 0xC0;
528 				EMIT3(b1, b2, add_2reg(b3, dst_reg, dst_reg));
529 
530 				insn++;
531 				i++;
532 				break;
533 			}
534 
535 			/* movabsq %rax, imm64 */
536 			EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
537 			EMIT(insn[0].imm, 4);
538 			EMIT(insn[1].imm, 4);
539 
540 			insn++;
541 			i++;
542 			break;
543 
544 			/* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
545 		case BPF_ALU | BPF_MOD | BPF_X:
546 		case BPF_ALU | BPF_DIV | BPF_X:
547 		case BPF_ALU | BPF_MOD | BPF_K:
548 		case BPF_ALU | BPF_DIV | BPF_K:
549 		case BPF_ALU64 | BPF_MOD | BPF_X:
550 		case BPF_ALU64 | BPF_DIV | BPF_X:
551 		case BPF_ALU64 | BPF_MOD | BPF_K:
552 		case BPF_ALU64 | BPF_DIV | BPF_K:
553 			EMIT1(0x50); /* push rax */
554 			EMIT1(0x52); /* push rdx */
555 
556 			if (BPF_SRC(insn->code) == BPF_X)
557 				/* mov r11, src_reg */
558 				EMIT_mov(AUX_REG, src_reg);
559 			else
560 				/* mov r11, imm32 */
561 				EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
562 
563 			/* mov rax, dst_reg */
564 			EMIT_mov(BPF_REG_0, dst_reg);
565 
566 			/* xor edx, edx
567 			 * equivalent to 'xor rdx, rdx', but one byte less
568 			 */
569 			EMIT2(0x31, 0xd2);
570 
571 			if (BPF_CLASS(insn->code) == BPF_ALU64)
572 				/* div r11 */
573 				EMIT3(0x49, 0xF7, 0xF3);
574 			else
575 				/* div r11d */
576 				EMIT3(0x41, 0xF7, 0xF3);
577 
578 			if (BPF_OP(insn->code) == BPF_MOD)
579 				/* mov r11, rdx */
580 				EMIT3(0x49, 0x89, 0xD3);
581 			else
582 				/* mov r11, rax */
583 				EMIT3(0x49, 0x89, 0xC3);
584 
585 			EMIT1(0x5A); /* pop rdx */
586 			EMIT1(0x58); /* pop rax */
587 
588 			/* mov dst_reg, r11 */
589 			EMIT_mov(dst_reg, AUX_REG);
590 			break;
591 
592 		case BPF_ALU | BPF_MUL | BPF_K:
593 		case BPF_ALU | BPF_MUL | BPF_X:
594 		case BPF_ALU64 | BPF_MUL | BPF_K:
595 		case BPF_ALU64 | BPF_MUL | BPF_X:
596 			EMIT1(0x50); /* push rax */
597 			EMIT1(0x52); /* push rdx */
598 
599 			/* mov r11, dst_reg */
600 			EMIT_mov(AUX_REG, dst_reg);
601 
602 			if (BPF_SRC(insn->code) == BPF_X)
603 				/* mov rax, src_reg */
604 				EMIT_mov(BPF_REG_0, src_reg);
605 			else
606 				/* mov rax, imm32 */
607 				EMIT3_off32(0x48, 0xC7, 0xC0, imm32);
608 
609 			if (BPF_CLASS(insn->code) == BPF_ALU64)
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 			/* mov r11, rax */
617 			EMIT_mov(AUX_REG, BPF_REG_0);
618 
619 			EMIT1(0x5A); /* pop rdx */
620 			EMIT1(0x58); /* pop rax */
621 
622 			/* mov dst_reg, r11 */
623 			EMIT_mov(dst_reg, AUX_REG);
624 			break;
625 
626 			/* shifts */
627 		case BPF_ALU | BPF_LSH | BPF_K:
628 		case BPF_ALU | BPF_RSH | BPF_K:
629 		case BPF_ALU | BPF_ARSH | BPF_K:
630 		case BPF_ALU64 | BPF_LSH | BPF_K:
631 		case BPF_ALU64 | BPF_RSH | BPF_K:
632 		case BPF_ALU64 | BPF_ARSH | BPF_K:
633 			if (BPF_CLASS(insn->code) == BPF_ALU64)
634 				EMIT1(add_1mod(0x48, dst_reg));
635 			else if (is_ereg(dst_reg))
636 				EMIT1(add_1mod(0x40, dst_reg));
637 
638 			switch (BPF_OP(insn->code)) {
639 			case BPF_LSH: b3 = 0xE0; break;
640 			case BPF_RSH: b3 = 0xE8; break;
641 			case BPF_ARSH: b3 = 0xF8; break;
642 			}
643 			EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
644 			break;
645 
646 		case BPF_ALU | BPF_LSH | BPF_X:
647 		case BPF_ALU | BPF_RSH | BPF_X:
648 		case BPF_ALU | BPF_ARSH | BPF_X:
649 		case BPF_ALU64 | BPF_LSH | BPF_X:
650 		case BPF_ALU64 | BPF_RSH | BPF_X:
651 		case BPF_ALU64 | BPF_ARSH | BPF_X:
652 
653 			/* check for bad case when dst_reg == rcx */
654 			if (dst_reg == BPF_REG_4) {
655 				/* mov r11, dst_reg */
656 				EMIT_mov(AUX_REG, dst_reg);
657 				dst_reg = AUX_REG;
658 			}
659 
660 			if (src_reg != BPF_REG_4) { /* common case */
661 				EMIT1(0x51); /* push rcx */
662 
663 				/* mov rcx, src_reg */
664 				EMIT_mov(BPF_REG_4, src_reg);
665 			}
666 
667 			/* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
668 			if (BPF_CLASS(insn->code) == BPF_ALU64)
669 				EMIT1(add_1mod(0x48, dst_reg));
670 			else if (is_ereg(dst_reg))
671 				EMIT1(add_1mod(0x40, dst_reg));
672 
673 			switch (BPF_OP(insn->code)) {
674 			case BPF_LSH: b3 = 0xE0; break;
675 			case BPF_RSH: b3 = 0xE8; break;
676 			case BPF_ARSH: b3 = 0xF8; break;
677 			}
678 			EMIT2(0xD3, add_1reg(b3, dst_reg));
679 
680 			if (src_reg != BPF_REG_4)
681 				EMIT1(0x59); /* pop rcx */
682 
683 			if (insn->dst_reg == BPF_REG_4)
684 				/* mov dst_reg, r11 */
685 				EMIT_mov(insn->dst_reg, AUX_REG);
686 			break;
687 
688 		case BPF_ALU | BPF_END | BPF_FROM_BE:
689 			switch (imm32) {
690 			case 16:
691 				/* emit 'ror %ax, 8' to swap lower 2 bytes */
692 				EMIT1(0x66);
693 				if (is_ereg(dst_reg))
694 					EMIT1(0x41);
695 				EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
696 
697 				/* emit 'movzwl eax, ax' */
698 				if (is_ereg(dst_reg))
699 					EMIT3(0x45, 0x0F, 0xB7);
700 				else
701 					EMIT2(0x0F, 0xB7);
702 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
703 				break;
704 			case 32:
705 				/* emit 'bswap eax' to swap lower 4 bytes */
706 				if (is_ereg(dst_reg))
707 					EMIT2(0x41, 0x0F);
708 				else
709 					EMIT1(0x0F);
710 				EMIT1(add_1reg(0xC8, dst_reg));
711 				break;
712 			case 64:
713 				/* emit 'bswap rax' to swap 8 bytes */
714 				EMIT3(add_1mod(0x48, dst_reg), 0x0F,
715 				      add_1reg(0xC8, dst_reg));
716 				break;
717 			}
718 			break;
719 
720 		case BPF_ALU | BPF_END | BPF_FROM_LE:
721 			switch (imm32) {
722 			case 16:
723 				/* emit 'movzwl eax, ax' to zero extend 16-bit
724 				 * into 64 bit
725 				 */
726 				if (is_ereg(dst_reg))
727 					EMIT3(0x45, 0x0F, 0xB7);
728 				else
729 					EMIT2(0x0F, 0xB7);
730 				EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
731 				break;
732 			case 32:
733 				/* emit 'mov eax, eax' to clear upper 32-bits */
734 				if (is_ereg(dst_reg))
735 					EMIT1(0x45);
736 				EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
737 				break;
738 			case 64:
739 				/* nop */
740 				break;
741 			}
742 			break;
743 
744 			/* ST: *(u8*)(dst_reg + off) = imm */
745 		case BPF_ST | BPF_MEM | BPF_B:
746 			if (is_ereg(dst_reg))
747 				EMIT2(0x41, 0xC6);
748 			else
749 				EMIT1(0xC6);
750 			goto st;
751 		case BPF_ST | BPF_MEM | BPF_H:
752 			if (is_ereg(dst_reg))
753 				EMIT3(0x66, 0x41, 0xC7);
754 			else
755 				EMIT2(0x66, 0xC7);
756 			goto st;
757 		case BPF_ST | BPF_MEM | BPF_W:
758 			if (is_ereg(dst_reg))
759 				EMIT2(0x41, 0xC7);
760 			else
761 				EMIT1(0xC7);
762 			goto st;
763 		case BPF_ST | BPF_MEM | BPF_DW:
764 			EMIT2(add_1mod(0x48, dst_reg), 0xC7);
765 
766 st:			if (is_imm8(insn->off))
767 				EMIT2(add_1reg(0x40, dst_reg), insn->off);
768 			else
769 				EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
770 
771 			EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
772 			break;
773 
774 			/* STX: *(u8*)(dst_reg + off) = src_reg */
775 		case BPF_STX | BPF_MEM | BPF_B:
776 			/* emit 'mov byte ptr [rax + off], al' */
777 			if (is_ereg(dst_reg) || is_ereg(src_reg) ||
778 			    /* have to add extra byte for x86 SIL, DIL regs */
779 			    src_reg == BPF_REG_1 || src_reg == BPF_REG_2)
780 				EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
781 			else
782 				EMIT1(0x88);
783 			goto stx;
784 		case BPF_STX | BPF_MEM | BPF_H:
785 			if (is_ereg(dst_reg) || is_ereg(src_reg))
786 				EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
787 			else
788 				EMIT2(0x66, 0x89);
789 			goto stx;
790 		case BPF_STX | BPF_MEM | BPF_W:
791 			if (is_ereg(dst_reg) || is_ereg(src_reg))
792 				EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
793 			else
794 				EMIT1(0x89);
795 			goto stx;
796 		case BPF_STX | BPF_MEM | BPF_DW:
797 			EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
798 stx:			if (is_imm8(insn->off))
799 				EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off);
800 			else
801 				EMIT1_off32(add_2reg(0x80, dst_reg, src_reg),
802 					    insn->off);
803 			break;
804 
805 			/* LDX: dst_reg = *(u8*)(src_reg + off) */
806 		case BPF_LDX | BPF_MEM | BPF_B:
807 			/* emit 'movzx rax, byte ptr [rax + off]' */
808 			EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
809 			goto ldx;
810 		case BPF_LDX | BPF_MEM | BPF_H:
811 			/* emit 'movzx rax, word ptr [rax + off]' */
812 			EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
813 			goto ldx;
814 		case BPF_LDX | BPF_MEM | BPF_W:
815 			/* emit 'mov eax, dword ptr [rax+0x14]' */
816 			if (is_ereg(dst_reg) || is_ereg(src_reg))
817 				EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
818 			else
819 				EMIT1(0x8B);
820 			goto ldx;
821 		case BPF_LDX | BPF_MEM | BPF_DW:
822 			/* emit 'mov rax, qword ptr [rax+0x14]' */
823 			EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
824 ldx:			/* if insn->off == 0 we can save one extra byte, but
825 			 * special case of x86 r13 which always needs an offset
826 			 * is not worth the hassle
827 			 */
828 			if (is_imm8(insn->off))
829 				EMIT2(add_2reg(0x40, src_reg, dst_reg), insn->off);
830 			else
831 				EMIT1_off32(add_2reg(0x80, src_reg, dst_reg),
832 					    insn->off);
833 			break;
834 
835 			/* STX XADD: lock *(u32*)(dst_reg + off) += src_reg */
836 		case BPF_STX | BPF_XADD | BPF_W:
837 			/* emit 'lock add dword ptr [rax + off], eax' */
838 			if (is_ereg(dst_reg) || is_ereg(src_reg))
839 				EMIT3(0xF0, add_2mod(0x40, dst_reg, src_reg), 0x01);
840 			else
841 				EMIT2(0xF0, 0x01);
842 			goto xadd;
843 		case BPF_STX | BPF_XADD | BPF_DW:
844 			EMIT3(0xF0, add_2mod(0x48, dst_reg, src_reg), 0x01);
845 xadd:			if (is_imm8(insn->off))
846 				EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off);
847 			else
848 				EMIT1_off32(add_2reg(0x80, dst_reg, src_reg),
849 					    insn->off);
850 			break;
851 
852 			/* call */
853 		case BPF_JMP | BPF_CALL:
854 			func = (u8 *) __bpf_call_base + imm32;
855 			jmp_offset = func - (image + addrs[i]);
856 			if (seen_ld_abs) {
857 				reload_skb_data = bpf_helper_changes_pkt_data(func);
858 				if (reload_skb_data) {
859 					EMIT1(0x57); /* push %rdi */
860 					jmp_offset += 22; /* pop, mov, sub, mov */
861 				} else {
862 					EMIT2(0x41, 0x52); /* push %r10 */
863 					EMIT2(0x41, 0x51); /* push %r9 */
864 					/* need to adjust jmp offset, since
865 					 * pop %r9, pop %r10 take 4 bytes after call insn
866 					 */
867 					jmp_offset += 4;
868 				}
869 			}
870 			if (!imm32 || !is_simm32(jmp_offset)) {
871 				pr_err("unsupported bpf func %d addr %p image %p\n",
872 				       imm32, func, image);
873 				return -EINVAL;
874 			}
875 			EMIT1_off32(0xE8, jmp_offset);
876 			if (seen_ld_abs) {
877 				if (reload_skb_data) {
878 					EMIT1(0x5F); /* pop %rdi */
879 					emit_load_skb_data_hlen(&prog);
880 				} else {
881 					EMIT2(0x41, 0x59); /* pop %r9 */
882 					EMIT2(0x41, 0x5A); /* pop %r10 */
883 				}
884 			}
885 			break;
886 
887 		case BPF_JMP | BPF_TAIL_CALL:
888 			emit_bpf_tail_call(&prog);
889 			break;
890 
891 			/* cond jump */
892 		case BPF_JMP | BPF_JEQ | BPF_X:
893 		case BPF_JMP | BPF_JNE | BPF_X:
894 		case BPF_JMP | BPF_JGT | BPF_X:
895 		case BPF_JMP | BPF_JLT | BPF_X:
896 		case BPF_JMP | BPF_JGE | BPF_X:
897 		case BPF_JMP | BPF_JLE | BPF_X:
898 		case BPF_JMP | BPF_JSGT | BPF_X:
899 		case BPF_JMP | BPF_JSLT | BPF_X:
900 		case BPF_JMP | BPF_JSGE | BPF_X:
901 		case BPF_JMP | BPF_JSLE | BPF_X:
902 			/* cmp dst_reg, src_reg */
903 			EMIT3(add_2mod(0x48, dst_reg, src_reg), 0x39,
904 			      add_2reg(0xC0, dst_reg, src_reg));
905 			goto emit_cond_jmp;
906 
907 		case BPF_JMP | BPF_JSET | BPF_X:
908 			/* test dst_reg, src_reg */
909 			EMIT3(add_2mod(0x48, dst_reg, src_reg), 0x85,
910 			      add_2reg(0xC0, dst_reg, src_reg));
911 			goto emit_cond_jmp;
912 
913 		case BPF_JMP | BPF_JSET | BPF_K:
914 			/* test dst_reg, imm32 */
915 			EMIT1(add_1mod(0x48, dst_reg));
916 			EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
917 			goto emit_cond_jmp;
918 
919 		case BPF_JMP | BPF_JEQ | BPF_K:
920 		case BPF_JMP | BPF_JNE | BPF_K:
921 		case BPF_JMP | BPF_JGT | BPF_K:
922 		case BPF_JMP | BPF_JLT | BPF_K:
923 		case BPF_JMP | BPF_JGE | BPF_K:
924 		case BPF_JMP | BPF_JLE | BPF_K:
925 		case BPF_JMP | BPF_JSGT | BPF_K:
926 		case BPF_JMP | BPF_JSLT | BPF_K:
927 		case BPF_JMP | BPF_JSGE | BPF_K:
928 		case BPF_JMP | BPF_JSLE | BPF_K:
929 			/* cmp dst_reg, imm8/32 */
930 			EMIT1(add_1mod(0x48, dst_reg));
931 
932 			if (is_imm8(imm32))
933 				EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
934 			else
935 				EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
936 
937 emit_cond_jmp:		/* convert BPF opcode to x86 */
938 			switch (BPF_OP(insn->code)) {
939 			case BPF_JEQ:
940 				jmp_cond = X86_JE;
941 				break;
942 			case BPF_JSET:
943 			case BPF_JNE:
944 				jmp_cond = X86_JNE;
945 				break;
946 			case BPF_JGT:
947 				/* GT is unsigned '>', JA in x86 */
948 				jmp_cond = X86_JA;
949 				break;
950 			case BPF_JLT:
951 				/* LT is unsigned '<', JB in x86 */
952 				jmp_cond = X86_JB;
953 				break;
954 			case BPF_JGE:
955 				/* GE is unsigned '>=', JAE in x86 */
956 				jmp_cond = X86_JAE;
957 				break;
958 			case BPF_JLE:
959 				/* LE is unsigned '<=', JBE in x86 */
960 				jmp_cond = X86_JBE;
961 				break;
962 			case BPF_JSGT:
963 				/* signed '>', GT in x86 */
964 				jmp_cond = X86_JG;
965 				break;
966 			case BPF_JSLT:
967 				/* signed '<', LT in x86 */
968 				jmp_cond = X86_JL;
969 				break;
970 			case BPF_JSGE:
971 				/* signed '>=', GE in x86 */
972 				jmp_cond = X86_JGE;
973 				break;
974 			case BPF_JSLE:
975 				/* signed '<=', LE in x86 */
976 				jmp_cond = X86_JLE;
977 				break;
978 			default: /* to silence gcc warning */
979 				return -EFAULT;
980 			}
981 			jmp_offset = addrs[i + insn->off] - addrs[i];
982 			if (is_imm8(jmp_offset)) {
983 				EMIT2(jmp_cond, jmp_offset);
984 			} else if (is_simm32(jmp_offset)) {
985 				EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
986 			} else {
987 				pr_err("cond_jmp gen bug %llx\n", jmp_offset);
988 				return -EFAULT;
989 			}
990 
991 			break;
992 
993 		case BPF_JMP | BPF_JA:
994 			jmp_offset = addrs[i + insn->off] - addrs[i];
995 			if (!jmp_offset)
996 				/* optimize out nop jumps */
997 				break;
998 emit_jmp:
999 			if (is_imm8(jmp_offset)) {
1000 				EMIT2(0xEB, jmp_offset);
1001 			} else if (is_simm32(jmp_offset)) {
1002 				EMIT1_off32(0xE9, jmp_offset);
1003 			} else {
1004 				pr_err("jmp gen bug %llx\n", jmp_offset);
1005 				return -EFAULT;
1006 			}
1007 			break;
1008 
1009 		case BPF_LD | BPF_IND | BPF_W:
1010 			func = sk_load_word;
1011 			goto common_load;
1012 		case BPF_LD | BPF_ABS | BPF_W:
1013 			func = CHOOSE_LOAD_FUNC(imm32, sk_load_word);
1014 common_load:
1015 			ctx->seen_ld_abs = seen_ld_abs = true;
1016 			jmp_offset = func - (image + addrs[i]);
1017 			if (!func || !is_simm32(jmp_offset)) {
1018 				pr_err("unsupported bpf func %d addr %p image %p\n",
1019 				       imm32, func, image);
1020 				return -EINVAL;
1021 			}
1022 			if (BPF_MODE(insn->code) == BPF_ABS) {
1023 				/* mov %esi, imm32 */
1024 				EMIT1_off32(0xBE, imm32);
1025 			} else {
1026 				/* mov %rsi, src_reg */
1027 				EMIT_mov(BPF_REG_2, src_reg);
1028 				if (imm32) {
1029 					if (is_imm8(imm32))
1030 						/* add %esi, imm8 */
1031 						EMIT3(0x83, 0xC6, imm32);
1032 					else
1033 						/* add %esi, imm32 */
1034 						EMIT2_off32(0x81, 0xC6, imm32);
1035 				}
1036 			}
1037 			/* skb pointer is in R6 (%rbx), it will be copied into
1038 			 * %rdi if skb_copy_bits() call is necessary.
1039 			 * sk_load_* helpers also use %r10 and %r9d.
1040 			 * See bpf_jit.S
1041 			 */
1042 			if (seen_ax_reg)
1043 				/* r10 = skb->data, mov %r10, off32(%rbx) */
1044 				EMIT3_off32(0x4c, 0x8b, 0x93,
1045 					    offsetof(struct sk_buff, data));
1046 			EMIT1_off32(0xE8, jmp_offset); /* call */
1047 			break;
1048 
1049 		case BPF_LD | BPF_IND | BPF_H:
1050 			func = sk_load_half;
1051 			goto common_load;
1052 		case BPF_LD | BPF_ABS | BPF_H:
1053 			func = CHOOSE_LOAD_FUNC(imm32, sk_load_half);
1054 			goto common_load;
1055 		case BPF_LD | BPF_IND | BPF_B:
1056 			func = sk_load_byte;
1057 			goto common_load;
1058 		case BPF_LD | BPF_ABS | BPF_B:
1059 			func = CHOOSE_LOAD_FUNC(imm32, sk_load_byte);
1060 			goto common_load;
1061 
1062 		case BPF_JMP | BPF_EXIT:
1063 			if (seen_exit) {
1064 				jmp_offset = ctx->cleanup_addr - addrs[i];
1065 				goto emit_jmp;
1066 			}
1067 			seen_exit = true;
1068 			/* update cleanup_addr */
1069 			ctx->cleanup_addr = proglen;
1070 			/* mov rbx, qword ptr [rbp+0] */
1071 			EMIT4(0x48, 0x8B, 0x5D, 0);
1072 			/* mov r13, qword ptr [rbp+8] */
1073 			EMIT4(0x4C, 0x8B, 0x6D, 8);
1074 			/* mov r14, qword ptr [rbp+16] */
1075 			EMIT4(0x4C, 0x8B, 0x75, 16);
1076 			/* mov r15, qword ptr [rbp+24] */
1077 			EMIT4(0x4C, 0x8B, 0x7D, 24);
1078 
1079 			/* add rbp, AUX_STACK_SPACE */
1080 			EMIT4(0x48, 0x83, 0xC5, AUX_STACK_SPACE);
1081 			EMIT1(0xC9); /* leave */
1082 			EMIT1(0xC3); /* ret */
1083 			break;
1084 
1085 		default:
1086 			/* By design x64 JIT should support all BPF instructions
1087 			 * This error will be seen if new instruction was added
1088 			 * to interpreter, but not to JIT
1089 			 * or if there is junk in bpf_prog
1090 			 */
1091 			pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1092 			return -EINVAL;
1093 		}
1094 
1095 		ilen = prog - temp;
1096 		if (ilen > BPF_MAX_INSN_SIZE) {
1097 			pr_err("bpf_jit: fatal insn size error\n");
1098 			return -EFAULT;
1099 		}
1100 
1101 		if (image) {
1102 			if (unlikely(proglen + ilen > oldproglen)) {
1103 				pr_err("bpf_jit: fatal error\n");
1104 				return -EFAULT;
1105 			}
1106 			memcpy(image + proglen, temp, ilen);
1107 		}
1108 		proglen += ilen;
1109 		addrs[i] = proglen;
1110 		prog = temp;
1111 	}
1112 	return proglen;
1113 }
1114 
1115 struct x64_jit_data {
1116 	struct bpf_binary_header *header;
1117 	int *addrs;
1118 	u8 *image;
1119 	int proglen;
1120 	struct jit_context ctx;
1121 };
1122 
1123 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1124 {
1125 	struct bpf_binary_header *header = NULL;
1126 	struct bpf_prog *tmp, *orig_prog = prog;
1127 	struct x64_jit_data *jit_data;
1128 	int proglen, oldproglen = 0;
1129 	struct jit_context ctx = {};
1130 	bool tmp_blinded = false;
1131 	bool extra_pass = false;
1132 	u8 *image = NULL;
1133 	int *addrs;
1134 	int pass;
1135 	int i;
1136 
1137 	if (!prog->jit_requested)
1138 		return orig_prog;
1139 
1140 	tmp = bpf_jit_blind_constants(prog);
1141 	/* If blinding was requested and we failed during blinding,
1142 	 * we must fall back to the interpreter.
1143 	 */
1144 	if (IS_ERR(tmp))
1145 		return orig_prog;
1146 	if (tmp != prog) {
1147 		tmp_blinded = true;
1148 		prog = tmp;
1149 	}
1150 
1151 	jit_data = prog->aux->jit_data;
1152 	if (!jit_data) {
1153 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1154 		if (!jit_data) {
1155 			prog = orig_prog;
1156 			goto out;
1157 		}
1158 		prog->aux->jit_data = jit_data;
1159 	}
1160 	addrs = jit_data->addrs;
1161 	if (addrs) {
1162 		ctx = jit_data->ctx;
1163 		oldproglen = jit_data->proglen;
1164 		image = jit_data->image;
1165 		header = jit_data->header;
1166 		extra_pass = true;
1167 		goto skip_init_addrs;
1168 	}
1169 	addrs = kmalloc(prog->len * sizeof(*addrs), GFP_KERNEL);
1170 	if (!addrs) {
1171 		prog = orig_prog;
1172 		goto out_addrs;
1173 	}
1174 
1175 	/* Before first pass, make a rough estimation of addrs[]
1176 	 * each bpf instruction is translated to less than 64 bytes
1177 	 */
1178 	for (proglen = 0, i = 0; i < prog->len; i++) {
1179 		proglen += 64;
1180 		addrs[i] = proglen;
1181 	}
1182 	ctx.cleanup_addr = proglen;
1183 skip_init_addrs:
1184 
1185 	/* JITed image shrinks with every pass and the loop iterates
1186 	 * until the image stops shrinking. Very large bpf programs
1187 	 * may converge on the last pass. In such case do one more
1188 	 * pass to emit the final image
1189 	 */
1190 	for (pass = 0; pass < 10 || image; pass++) {
1191 		proglen = do_jit(prog, addrs, image, oldproglen, &ctx);
1192 		if (proglen <= 0) {
1193 			image = NULL;
1194 			if (header)
1195 				bpf_jit_binary_free(header);
1196 			prog = orig_prog;
1197 			goto out_addrs;
1198 		}
1199 		if (image) {
1200 			if (proglen != oldproglen) {
1201 				pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
1202 				       proglen, oldproglen);
1203 				prog = orig_prog;
1204 				goto out_addrs;
1205 			}
1206 			break;
1207 		}
1208 		if (proglen == oldproglen) {
1209 			header = bpf_jit_binary_alloc(proglen, &image,
1210 						      1, jit_fill_hole);
1211 			if (!header) {
1212 				prog = orig_prog;
1213 				goto out_addrs;
1214 			}
1215 		}
1216 		oldproglen = proglen;
1217 	}
1218 
1219 	if (bpf_jit_enable > 1)
1220 		bpf_jit_dump(prog->len, proglen, pass + 1, image);
1221 
1222 	if (image) {
1223 		bpf_flush_icache(header, image + proglen);
1224 		if (!prog->is_func || extra_pass) {
1225 			bpf_jit_binary_lock_ro(header);
1226 		} else {
1227 			jit_data->addrs = addrs;
1228 			jit_data->ctx = ctx;
1229 			jit_data->proglen = proglen;
1230 			jit_data->image = image;
1231 			jit_data->header = header;
1232 		}
1233 		prog->bpf_func = (void *)image;
1234 		prog->jited = 1;
1235 		prog->jited_len = proglen;
1236 	} else {
1237 		prog = orig_prog;
1238 	}
1239 
1240 	if (!prog->is_func || extra_pass) {
1241 out_addrs:
1242 		kfree(addrs);
1243 		kfree(jit_data);
1244 		prog->aux->jit_data = NULL;
1245 	}
1246 out:
1247 	if (tmp_blinded)
1248 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
1249 					   tmp : orig_prog);
1250 	return prog;
1251 }
1252