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