xref: /openbmc/linux/arch/x86/net/bpf_jit_comp.c (revision b9ccfda2)
1 /* bpf_jit_comp.c : BPF JIT compiler
2  *
3  * Copyright (C) 2011 Eric Dumazet (eric.dumazet@gmail.com)
4  *
5  * This program is free software; you can redistribute it and/or
6  * modify it under the terms of the GNU General Public License
7  * as published by the Free Software Foundation; version 2
8  * of the License.
9  */
10 #include <linux/moduleloader.h>
11 #include <asm/cacheflush.h>
12 #include <linux/netdevice.h>
13 #include <linux/filter.h>
14 
15 /*
16  * Conventions :
17  *  EAX : BPF A accumulator
18  *  EBX : BPF X accumulator
19  *  RDI : pointer to skb   (first argument given to JIT function)
20  *  RBP : frame pointer (even if CONFIG_FRAME_POINTER=n)
21  *  ECX,EDX,ESI : scratch registers
22  *  r9d : skb->len - skb->data_len (headlen)
23  *  r8  : skb->data
24  * -8(RBP) : saved RBX value
25  * -16(RBP)..-80(RBP) : BPF_MEMWORDS values
26  */
27 int bpf_jit_enable __read_mostly;
28 
29 /*
30  * assembly code in arch/x86/net/bpf_jit.S
31  */
32 extern u8 sk_load_word[], sk_load_half[], sk_load_byte[], sk_load_byte_msh[];
33 extern u8 sk_load_word_positive_offset[], sk_load_half_positive_offset[];
34 extern u8 sk_load_byte_positive_offset[], sk_load_byte_msh_positive_offset[];
35 extern u8 sk_load_word_negative_offset[], sk_load_half_negative_offset[];
36 extern u8 sk_load_byte_negative_offset[], sk_load_byte_msh_negative_offset[];
37 
38 static inline u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
39 {
40 	if (len == 1)
41 		*ptr = bytes;
42 	else if (len == 2)
43 		*(u16 *)ptr = bytes;
44 	else {
45 		*(u32 *)ptr = bytes;
46 		barrier();
47 	}
48 	return ptr + len;
49 }
50 
51 #define EMIT(bytes, len)	do { prog = emit_code(prog, bytes, len); } while (0)
52 
53 #define EMIT1(b1)		EMIT(b1, 1)
54 #define EMIT2(b1, b2)		EMIT((b1) + ((b2) << 8), 2)
55 #define EMIT3(b1, b2, b3)	EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
56 #define EMIT4(b1, b2, b3, b4)   EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
57 #define EMIT1_off32(b1, off)	do { EMIT1(b1); EMIT(off, 4);} while (0)
58 
59 #define CLEAR_A() EMIT2(0x31, 0xc0) /* xor %eax,%eax */
60 #define CLEAR_X() EMIT2(0x31, 0xdb) /* xor %ebx,%ebx */
61 
62 static inline bool is_imm8(int value)
63 {
64 	return value <= 127 && value >= -128;
65 }
66 
67 static inline bool is_near(int offset)
68 {
69 	return offset <= 127 && offset >= -128;
70 }
71 
72 #define EMIT_JMP(offset)						\
73 do {									\
74 	if (offset) {							\
75 		if (is_near(offset))					\
76 			EMIT2(0xeb, offset); /* jmp .+off8 */		\
77 		else							\
78 			EMIT1_off32(0xe9, offset); /* jmp .+off32 */	\
79 	}								\
80 } while (0)
81 
82 /* list of x86 cond jumps opcodes (. + s8)
83  * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
84  */
85 #define X86_JB  0x72
86 #define X86_JAE 0x73
87 #define X86_JE  0x74
88 #define X86_JNE 0x75
89 #define X86_JBE 0x76
90 #define X86_JA  0x77
91 
92 #define EMIT_COND_JMP(op, offset)				\
93 do {								\
94 	if (is_near(offset))					\
95 		EMIT2(op, offset); /* jxx .+off8 */		\
96 	else {							\
97 		EMIT2(0x0f, op + 0x10);				\
98 		EMIT(offset, 4); /* jxx .+off32 */		\
99 	}							\
100 } while (0)
101 
102 #define COND_SEL(CODE, TOP, FOP)	\
103 	case CODE:			\
104 		t_op = TOP;		\
105 		f_op = FOP;		\
106 		goto cond_branch
107 
108 
109 #define SEEN_DATAREF 1 /* might call external helpers */
110 #define SEEN_XREG    2 /* ebx is used */
111 #define SEEN_MEM     4 /* use mem[] for temporary storage */
112 
113 static inline void bpf_flush_icache(void *start, void *end)
114 {
115 	mm_segment_t old_fs = get_fs();
116 
117 	set_fs(KERNEL_DS);
118 	smp_wmb();
119 	flush_icache_range((unsigned long)start, (unsigned long)end);
120 	set_fs(old_fs);
121 }
122 
123 #define CHOOSE_LOAD_FUNC(K, func) \
124 	((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
125 
126 void bpf_jit_compile(struct sk_filter *fp)
127 {
128 	u8 temp[64];
129 	u8 *prog;
130 	unsigned int proglen, oldproglen = 0;
131 	int ilen, i;
132 	int t_offset, f_offset;
133 	u8 t_op, f_op, seen = 0, pass;
134 	u8 *image = NULL;
135 	u8 *func;
136 	int pc_ret0 = -1; /* bpf index of first RET #0 instruction (if any) */
137 	unsigned int cleanup_addr; /* epilogue code offset */
138 	unsigned int *addrs;
139 	const struct sock_filter *filter = fp->insns;
140 	int flen = fp->len;
141 
142 	if (!bpf_jit_enable)
143 		return;
144 
145 	addrs = kmalloc(flen * sizeof(*addrs), GFP_KERNEL);
146 	if (addrs == NULL)
147 		return;
148 
149 	/* Before first pass, make a rough estimation of addrs[]
150 	 * each bpf instruction is translated to less than 64 bytes
151 	 */
152 	for (proglen = 0, i = 0; i < flen; i++) {
153 		proglen += 64;
154 		addrs[i] = proglen;
155 	}
156 	cleanup_addr = proglen; /* epilogue address */
157 
158 	for (pass = 0; pass < 10; pass++) {
159 		u8 seen_or_pass0 = (pass == 0) ? (SEEN_XREG | SEEN_DATAREF | SEEN_MEM) : seen;
160 		/* no prologue/epilogue for trivial filters (RET something) */
161 		proglen = 0;
162 		prog = temp;
163 
164 		if (seen_or_pass0) {
165 			EMIT4(0x55, 0x48, 0x89, 0xe5); /* push %rbp; mov %rsp,%rbp */
166 			EMIT4(0x48, 0x83, 0xec, 96);	/* subq  $96,%rsp	*/
167 			/* note : must save %rbx in case bpf_error is hit */
168 			if (seen_or_pass0 & (SEEN_XREG | SEEN_DATAREF))
169 				EMIT4(0x48, 0x89, 0x5d, 0xf8); /* mov %rbx, -8(%rbp) */
170 			if (seen_or_pass0 & SEEN_XREG)
171 				CLEAR_X(); /* make sure we dont leek kernel memory */
172 
173 			/*
174 			 * If this filter needs to access skb data,
175 			 * loads r9 and r8 with :
176 			 *  r9 = skb->len - skb->data_len
177 			 *  r8 = skb->data
178 			 */
179 			if (seen_or_pass0 & SEEN_DATAREF) {
180 				if (offsetof(struct sk_buff, len) <= 127)
181 					/* mov    off8(%rdi),%r9d */
182 					EMIT4(0x44, 0x8b, 0x4f, offsetof(struct sk_buff, len));
183 				else {
184 					/* mov    off32(%rdi),%r9d */
185 					EMIT3(0x44, 0x8b, 0x8f);
186 					EMIT(offsetof(struct sk_buff, len), 4);
187 				}
188 				if (is_imm8(offsetof(struct sk_buff, data_len)))
189 					/* sub    off8(%rdi),%r9d */
190 					EMIT4(0x44, 0x2b, 0x4f, offsetof(struct sk_buff, data_len));
191 				else {
192 					EMIT3(0x44, 0x2b, 0x8f);
193 					EMIT(offsetof(struct sk_buff, data_len), 4);
194 				}
195 
196 				if (is_imm8(offsetof(struct sk_buff, data)))
197 					/* mov off8(%rdi),%r8 */
198 					EMIT4(0x4c, 0x8b, 0x47, offsetof(struct sk_buff, data));
199 				else {
200 					/* mov off32(%rdi),%r8 */
201 					EMIT3(0x4c, 0x8b, 0x87);
202 					EMIT(offsetof(struct sk_buff, data), 4);
203 				}
204 			}
205 		}
206 
207 		switch (filter[0].code) {
208 		case BPF_S_RET_K:
209 		case BPF_S_LD_W_LEN:
210 		case BPF_S_ANC_PROTOCOL:
211 		case BPF_S_ANC_IFINDEX:
212 		case BPF_S_ANC_MARK:
213 		case BPF_S_ANC_RXHASH:
214 		case BPF_S_ANC_CPU:
215 		case BPF_S_ANC_QUEUE:
216 		case BPF_S_LD_W_ABS:
217 		case BPF_S_LD_H_ABS:
218 		case BPF_S_LD_B_ABS:
219 			/* first instruction sets A register (or is RET 'constant') */
220 			break;
221 		default:
222 			/* make sure we dont leak kernel information to user */
223 			CLEAR_A(); /* A = 0 */
224 		}
225 
226 		for (i = 0; i < flen; i++) {
227 			unsigned int K = filter[i].k;
228 
229 			switch (filter[i].code) {
230 			case BPF_S_ALU_ADD_X: /* A += X; */
231 				seen |= SEEN_XREG;
232 				EMIT2(0x01, 0xd8);		/* add %ebx,%eax */
233 				break;
234 			case BPF_S_ALU_ADD_K: /* A += K; */
235 				if (!K)
236 					break;
237 				if (is_imm8(K))
238 					EMIT3(0x83, 0xc0, K);	/* add imm8,%eax */
239 				else
240 					EMIT1_off32(0x05, K);	/* add imm32,%eax */
241 				break;
242 			case BPF_S_ALU_SUB_X: /* A -= X; */
243 				seen |= SEEN_XREG;
244 				EMIT2(0x29, 0xd8);		/* sub    %ebx,%eax */
245 				break;
246 			case BPF_S_ALU_SUB_K: /* A -= K */
247 				if (!K)
248 					break;
249 				if (is_imm8(K))
250 					EMIT3(0x83, 0xe8, K); /* sub imm8,%eax */
251 				else
252 					EMIT1_off32(0x2d, K); /* sub imm32,%eax */
253 				break;
254 			case BPF_S_ALU_MUL_X: /* A *= X; */
255 				seen |= SEEN_XREG;
256 				EMIT3(0x0f, 0xaf, 0xc3);	/* imul %ebx,%eax */
257 				break;
258 			case BPF_S_ALU_MUL_K: /* A *= K */
259 				if (is_imm8(K))
260 					EMIT3(0x6b, 0xc0, K); /* imul imm8,%eax,%eax */
261 				else {
262 					EMIT2(0x69, 0xc0);		/* imul imm32,%eax */
263 					EMIT(K, 4);
264 				}
265 				break;
266 			case BPF_S_ALU_DIV_X: /* A /= X; */
267 				seen |= SEEN_XREG;
268 				EMIT2(0x85, 0xdb);	/* test %ebx,%ebx */
269 				if (pc_ret0 > 0) {
270 					/* addrs[pc_ret0 - 1] is start address of target
271 					 * (addrs[i] - 4) is the address following this jmp
272 					 * ("xor %edx,%edx; div %ebx" being 4 bytes long)
273 					 */
274 					EMIT_COND_JMP(X86_JE, addrs[pc_ret0 - 1] -
275 								(addrs[i] - 4));
276 				} else {
277 					EMIT_COND_JMP(X86_JNE, 2 + 5);
278 					CLEAR_A();
279 					EMIT1_off32(0xe9, cleanup_addr - (addrs[i] - 4)); /* jmp .+off32 */
280 				}
281 				EMIT4(0x31, 0xd2, 0xf7, 0xf3); /* xor %edx,%edx; div %ebx */
282 				break;
283 			case BPF_S_ALU_DIV_K: /* A = reciprocal_divide(A, K); */
284 				EMIT3(0x48, 0x69, 0xc0); /* imul imm32,%rax,%rax */
285 				EMIT(K, 4);
286 				EMIT4(0x48, 0xc1, 0xe8, 0x20); /* shr $0x20,%rax */
287 				break;
288 			case BPF_S_ALU_AND_X:
289 				seen |= SEEN_XREG;
290 				EMIT2(0x21, 0xd8);		/* and %ebx,%eax */
291 				break;
292 			case BPF_S_ALU_AND_K:
293 				if (K >= 0xFFFFFF00) {
294 					EMIT2(0x24, K & 0xFF); /* and imm8,%al */
295 				} else if (K >= 0xFFFF0000) {
296 					EMIT2(0x66, 0x25);	/* and imm16,%ax */
297 					EMIT(K, 2);
298 				} else {
299 					EMIT1_off32(0x25, K);	/* and imm32,%eax */
300 				}
301 				break;
302 			case BPF_S_ALU_OR_X:
303 				seen |= SEEN_XREG;
304 				EMIT2(0x09, 0xd8);		/* or %ebx,%eax */
305 				break;
306 			case BPF_S_ALU_OR_K:
307 				if (is_imm8(K))
308 					EMIT3(0x83, 0xc8, K); /* or imm8,%eax */
309 				else
310 					EMIT1_off32(0x0d, K);	/* or imm32,%eax */
311 				break;
312 			case BPF_S_ANC_ALU_XOR_X: /* A ^= X; */
313 				seen |= SEEN_XREG;
314 				EMIT2(0x31, 0xd8);		/* xor %ebx,%eax */
315 				break;
316 			case BPF_S_ALU_LSH_X: /* A <<= X; */
317 				seen |= SEEN_XREG;
318 				EMIT4(0x89, 0xd9, 0xd3, 0xe0);	/* mov %ebx,%ecx; shl %cl,%eax */
319 				break;
320 			case BPF_S_ALU_LSH_K:
321 				if (K == 0)
322 					break;
323 				else if (K == 1)
324 					EMIT2(0xd1, 0xe0); /* shl %eax */
325 				else
326 					EMIT3(0xc1, 0xe0, K);
327 				break;
328 			case BPF_S_ALU_RSH_X: /* A >>= X; */
329 				seen |= SEEN_XREG;
330 				EMIT4(0x89, 0xd9, 0xd3, 0xe8);	/* mov %ebx,%ecx; shr %cl,%eax */
331 				break;
332 			case BPF_S_ALU_RSH_K: /* A >>= K; */
333 				if (K == 0)
334 					break;
335 				else if (K == 1)
336 					EMIT2(0xd1, 0xe8); /* shr %eax */
337 				else
338 					EMIT3(0xc1, 0xe8, K);
339 				break;
340 			case BPF_S_ALU_NEG:
341 				EMIT2(0xf7, 0xd8);		/* neg %eax */
342 				break;
343 			case BPF_S_RET_K:
344 				if (!K) {
345 					if (pc_ret0 == -1)
346 						pc_ret0 = i;
347 					CLEAR_A();
348 				} else {
349 					EMIT1_off32(0xb8, K);	/* mov $imm32,%eax */
350 				}
351 				/* fallinto */
352 			case BPF_S_RET_A:
353 				if (seen_or_pass0) {
354 					if (i != flen - 1) {
355 						EMIT_JMP(cleanup_addr - addrs[i]);
356 						break;
357 					}
358 					if (seen_or_pass0 & SEEN_XREG)
359 						EMIT4(0x48, 0x8b, 0x5d, 0xf8);  /* mov  -8(%rbp),%rbx */
360 					EMIT1(0xc9);		/* leaveq */
361 				}
362 				EMIT1(0xc3);		/* ret */
363 				break;
364 			case BPF_S_MISC_TAX: /* X = A */
365 				seen |= SEEN_XREG;
366 				EMIT2(0x89, 0xc3);	/* mov    %eax,%ebx */
367 				break;
368 			case BPF_S_MISC_TXA: /* A = X */
369 				seen |= SEEN_XREG;
370 				EMIT2(0x89, 0xd8);	/* mov    %ebx,%eax */
371 				break;
372 			case BPF_S_LD_IMM: /* A = K */
373 				if (!K)
374 					CLEAR_A();
375 				else
376 					EMIT1_off32(0xb8, K); /* mov $imm32,%eax */
377 				break;
378 			case BPF_S_LDX_IMM: /* X = K */
379 				seen |= SEEN_XREG;
380 				if (!K)
381 					CLEAR_X();
382 				else
383 					EMIT1_off32(0xbb, K); /* mov $imm32,%ebx */
384 				break;
385 			case BPF_S_LD_MEM: /* A = mem[K] : mov off8(%rbp),%eax */
386 				seen |= SEEN_MEM;
387 				EMIT3(0x8b, 0x45, 0xf0 - K*4);
388 				break;
389 			case BPF_S_LDX_MEM: /* X = mem[K] : mov off8(%rbp),%ebx */
390 				seen |= SEEN_XREG | SEEN_MEM;
391 				EMIT3(0x8b, 0x5d, 0xf0 - K*4);
392 				break;
393 			case BPF_S_ST: /* mem[K] = A : mov %eax,off8(%rbp) */
394 				seen |= SEEN_MEM;
395 				EMIT3(0x89, 0x45, 0xf0 - K*4);
396 				break;
397 			case BPF_S_STX: /* mem[K] = X : mov %ebx,off8(%rbp) */
398 				seen |= SEEN_XREG | SEEN_MEM;
399 				EMIT3(0x89, 0x5d, 0xf0 - K*4);
400 				break;
401 			case BPF_S_LD_W_LEN: /*	A = skb->len; */
402 				BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
403 				if (is_imm8(offsetof(struct sk_buff, len)))
404 					/* mov    off8(%rdi),%eax */
405 					EMIT3(0x8b, 0x47, offsetof(struct sk_buff, len));
406 				else {
407 					EMIT2(0x8b, 0x87);
408 					EMIT(offsetof(struct sk_buff, len), 4);
409 				}
410 				break;
411 			case BPF_S_LDX_W_LEN: /* X = skb->len; */
412 				seen |= SEEN_XREG;
413 				if (is_imm8(offsetof(struct sk_buff, len)))
414 					/* mov off8(%rdi),%ebx */
415 					EMIT3(0x8b, 0x5f, offsetof(struct sk_buff, len));
416 				else {
417 					EMIT2(0x8b, 0x9f);
418 					EMIT(offsetof(struct sk_buff, len), 4);
419 				}
420 				break;
421 			case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
422 				BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
423 				if (is_imm8(offsetof(struct sk_buff, protocol))) {
424 					/* movzwl off8(%rdi),%eax */
425 					EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, protocol));
426 				} else {
427 					EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */
428 					EMIT(offsetof(struct sk_buff, protocol), 4);
429 				}
430 				EMIT2(0x86, 0xc4); /* ntohs() : xchg   %al,%ah */
431 				break;
432 			case BPF_S_ANC_IFINDEX:
433 				if (is_imm8(offsetof(struct sk_buff, dev))) {
434 					/* movq off8(%rdi),%rax */
435 					EMIT4(0x48, 0x8b, 0x47, offsetof(struct sk_buff, dev));
436 				} else {
437 					EMIT3(0x48, 0x8b, 0x87); /* movq off32(%rdi),%rax */
438 					EMIT(offsetof(struct sk_buff, dev), 4);
439 				}
440 				EMIT3(0x48, 0x85, 0xc0);	/* test %rax,%rax */
441 				EMIT_COND_JMP(X86_JE, cleanup_addr - (addrs[i] - 6));
442 				BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
443 				EMIT2(0x8b, 0x80);	/* mov off32(%rax),%eax */
444 				EMIT(offsetof(struct net_device, ifindex), 4);
445 				break;
446 			case BPF_S_ANC_MARK:
447 				BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
448 				if (is_imm8(offsetof(struct sk_buff, mark))) {
449 					/* mov off8(%rdi),%eax */
450 					EMIT3(0x8b, 0x47, offsetof(struct sk_buff, mark));
451 				} else {
452 					EMIT2(0x8b, 0x87);
453 					EMIT(offsetof(struct sk_buff, mark), 4);
454 				}
455 				break;
456 			case BPF_S_ANC_RXHASH:
457 				BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4);
458 				if (is_imm8(offsetof(struct sk_buff, rxhash))) {
459 					/* mov off8(%rdi),%eax */
460 					EMIT3(0x8b, 0x47, offsetof(struct sk_buff, rxhash));
461 				} else {
462 					EMIT2(0x8b, 0x87);
463 					EMIT(offsetof(struct sk_buff, rxhash), 4);
464 				}
465 				break;
466 			case BPF_S_ANC_QUEUE:
467 				BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
468 				if (is_imm8(offsetof(struct sk_buff, queue_mapping))) {
469 					/* movzwl off8(%rdi),%eax */
470 					EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, queue_mapping));
471 				} else {
472 					EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */
473 					EMIT(offsetof(struct sk_buff, queue_mapping), 4);
474 				}
475 				break;
476 			case BPF_S_ANC_CPU:
477 #ifdef CONFIG_SMP
478 				EMIT4(0x65, 0x8b, 0x04, 0x25); /* mov %gs:off32,%eax */
479 				EMIT((u32)(unsigned long)&cpu_number, 4); /* A = smp_processor_id(); */
480 #else
481 				CLEAR_A();
482 #endif
483 				break;
484 			case BPF_S_LD_W_ABS:
485 				func = CHOOSE_LOAD_FUNC(K, sk_load_word);
486 common_load:			seen |= SEEN_DATAREF;
487 				t_offset = func - (image + addrs[i]);
488 				EMIT1_off32(0xbe, K); /* mov imm32,%esi */
489 				EMIT1_off32(0xe8, t_offset); /* call */
490 				break;
491 			case BPF_S_LD_H_ABS:
492 				func = CHOOSE_LOAD_FUNC(K, sk_load_half);
493 				goto common_load;
494 			case BPF_S_LD_B_ABS:
495 				func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
496 				goto common_load;
497 			case BPF_S_LDX_B_MSH:
498 				func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
499 				seen |= SEEN_DATAREF | SEEN_XREG;
500 				t_offset = func - (image + addrs[i]);
501 				EMIT1_off32(0xbe, K);	/* mov imm32,%esi */
502 				EMIT1_off32(0xe8, t_offset); /* call sk_load_byte_msh */
503 				break;
504 			case BPF_S_LD_W_IND:
505 				func = sk_load_word;
506 common_load_ind:		seen |= SEEN_DATAREF | SEEN_XREG;
507 				t_offset = func - (image + addrs[i]);
508 				if (K) {
509 					if (is_imm8(K)) {
510 						EMIT3(0x8d, 0x73, K); /* lea imm8(%rbx), %esi */
511 					} else {
512 						EMIT2(0x8d, 0xb3); /* lea imm32(%rbx),%esi */
513 						EMIT(K, 4);
514 					}
515 				} else {
516 					EMIT2(0x89,0xde); /* mov %ebx,%esi */
517 				}
518 				EMIT1_off32(0xe8, t_offset);	/* call sk_load_xxx_ind */
519 				break;
520 			case BPF_S_LD_H_IND:
521 				func = sk_load_half;
522 				goto common_load_ind;
523 			case BPF_S_LD_B_IND:
524 				func = sk_load_byte;
525 				goto common_load_ind;
526 			case BPF_S_JMP_JA:
527 				t_offset = addrs[i + K] - addrs[i];
528 				EMIT_JMP(t_offset);
529 				break;
530 			COND_SEL(BPF_S_JMP_JGT_K, X86_JA, X86_JBE);
531 			COND_SEL(BPF_S_JMP_JGE_K, X86_JAE, X86_JB);
532 			COND_SEL(BPF_S_JMP_JEQ_K, X86_JE, X86_JNE);
533 			COND_SEL(BPF_S_JMP_JSET_K,X86_JNE, X86_JE);
534 			COND_SEL(BPF_S_JMP_JGT_X, X86_JA, X86_JBE);
535 			COND_SEL(BPF_S_JMP_JGE_X, X86_JAE, X86_JB);
536 			COND_SEL(BPF_S_JMP_JEQ_X, X86_JE, X86_JNE);
537 			COND_SEL(BPF_S_JMP_JSET_X,X86_JNE, X86_JE);
538 
539 cond_branch:			f_offset = addrs[i + filter[i].jf] - addrs[i];
540 				t_offset = addrs[i + filter[i].jt] - addrs[i];
541 
542 				/* same targets, can avoid doing the test :) */
543 				if (filter[i].jt == filter[i].jf) {
544 					EMIT_JMP(t_offset);
545 					break;
546 				}
547 
548 				switch (filter[i].code) {
549 				case BPF_S_JMP_JGT_X:
550 				case BPF_S_JMP_JGE_X:
551 				case BPF_S_JMP_JEQ_X:
552 					seen |= SEEN_XREG;
553 					EMIT2(0x39, 0xd8); /* cmp %ebx,%eax */
554 					break;
555 				case BPF_S_JMP_JSET_X:
556 					seen |= SEEN_XREG;
557 					EMIT2(0x85, 0xd8); /* test %ebx,%eax */
558 					break;
559 				case BPF_S_JMP_JEQ_K:
560 					if (K == 0) {
561 						EMIT2(0x85, 0xc0); /* test   %eax,%eax */
562 						break;
563 					}
564 				case BPF_S_JMP_JGT_K:
565 				case BPF_S_JMP_JGE_K:
566 					if (K <= 127)
567 						EMIT3(0x83, 0xf8, K); /* cmp imm8,%eax */
568 					else
569 						EMIT1_off32(0x3d, K); /* cmp imm32,%eax */
570 					break;
571 				case BPF_S_JMP_JSET_K:
572 					if (K <= 0xFF)
573 						EMIT2(0xa8, K); /* test imm8,%al */
574 					else if (!(K & 0xFFFF00FF))
575 						EMIT3(0xf6, 0xc4, K >> 8); /* test imm8,%ah */
576 					else if (K <= 0xFFFF) {
577 						EMIT2(0x66, 0xa9); /* test imm16,%ax */
578 						EMIT(K, 2);
579 					} else {
580 						EMIT1_off32(0xa9, K); /* test imm32,%eax */
581 					}
582 					break;
583 				}
584 				if (filter[i].jt != 0) {
585 					if (filter[i].jf && f_offset)
586 						t_offset += is_near(f_offset) ? 2 : 5;
587 					EMIT_COND_JMP(t_op, t_offset);
588 					if (filter[i].jf)
589 						EMIT_JMP(f_offset);
590 					break;
591 				}
592 				EMIT_COND_JMP(f_op, f_offset);
593 				break;
594 			default:
595 				/* hmm, too complex filter, give up with jit compiler */
596 				goto out;
597 			}
598 			ilen = prog - temp;
599 			if (image) {
600 				if (unlikely(proglen + ilen > oldproglen)) {
601 					pr_err("bpb_jit_compile fatal error\n");
602 					kfree(addrs);
603 					module_free(NULL, image);
604 					return;
605 				}
606 				memcpy(image + proglen, temp, ilen);
607 			}
608 			proglen += ilen;
609 			addrs[i] = proglen;
610 			prog = temp;
611 		}
612 		/* last bpf instruction is always a RET :
613 		 * use it to give the cleanup instruction(s) addr
614 		 */
615 		cleanup_addr = proglen - 1; /* ret */
616 		if (seen_or_pass0)
617 			cleanup_addr -= 1; /* leaveq */
618 		if (seen_or_pass0 & SEEN_XREG)
619 			cleanup_addr -= 4; /* mov  -8(%rbp),%rbx */
620 
621 		if (image) {
622 			if (proglen != oldproglen)
623 				pr_err("bpb_jit_compile proglen=%u != oldproglen=%u\n", proglen, oldproglen);
624 			break;
625 		}
626 		if (proglen == oldproglen) {
627 			image = module_alloc(max_t(unsigned int,
628 						   proglen,
629 						   sizeof(struct work_struct)));
630 			if (!image)
631 				goto out;
632 		}
633 		oldproglen = proglen;
634 	}
635 	if (bpf_jit_enable > 1)
636 		pr_err("flen=%d proglen=%u pass=%d image=%p\n",
637 		       flen, proglen, pass, image);
638 
639 	if (image) {
640 		if (bpf_jit_enable > 1)
641 			print_hex_dump(KERN_ERR, "JIT code: ", DUMP_PREFIX_ADDRESS,
642 				       16, 1, image, proglen, false);
643 
644 		bpf_flush_icache(image, image + proglen);
645 
646 		fp->bpf_func = (void *)image;
647 	}
648 out:
649 	kfree(addrs);
650 	return;
651 }
652 
653 static void jit_free_defer(struct work_struct *arg)
654 {
655 	module_free(NULL, arg);
656 }
657 
658 /* run from softirq, we must use a work_struct to call
659  * module_free() from process context
660  */
661 void bpf_jit_free(struct sk_filter *fp)
662 {
663 	if (fp->bpf_func != sk_run_filter) {
664 		struct work_struct *work = (struct work_struct *)fp->bpf_func;
665 
666 		INIT_WORK(work, jit_free_defer);
667 		schedule_work(work);
668 	}
669 }
670