1 /* bpf_jit_comp.c : BPF JIT compiler
3 * Copyright (C) 2011 Eric Dumazet (eric.dumazet@gmail.com)
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
10 #include <linux/moduleloader.h>
11 #include <asm/cacheflush.h>
12 #include <linux/netdevice.h>
13 #include <linux/filter.h>
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)
24 * -8(RBP) : saved RBX value
25 * -16(RBP)..-80(RBP) : BPF_MEMWORDS values
27 int bpf_jit_enable __read_mostly;
30 * assembly code in arch/x86/net/bpf_jit.S
32 extern u8 sk_load_word[], sk_load_half[], sk_load_byte[], sk_load_byte_msh[];
33 extern u8 sk_load_word_ind[], sk_load_half_ind[], sk_load_byte_ind[];
35 static inline u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
48 #define EMIT(bytes, len) do { prog = emit_code(prog, bytes, len); } while (0)
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)
56 #define CLEAR_A() EMIT2(0x31, 0xc0) /* xor %eax,%eax */
57 #define CLEAR_X() EMIT2(0x31, 0xdb) /* xor %ebx,%ebx */
59 static inline bool is_imm8(int value)
61 return value <= 127 && value >= -128;
64 static inline bool is_near(int offset)
66 return offset <= 127 && offset >= -128;
69 #define EMIT_JMP(offset) \
72 if (is_near(offset)) \
73 EMIT2(0xeb, offset); /* jmp .+off8 */ \
75 EMIT1_off32(0xe9, offset); /* jmp .+off32 */ \
79 /* list of x86 cond jumps opcodes (. + s8)
80 * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
89 #define EMIT_COND_JMP(op, offset) \
91 if (is_near(offset)) \
92 EMIT2(op, offset); /* jxx .+off8 */ \
94 EMIT2(0x0f, op + 0x10); \
95 EMIT(offset, 4); /* jxx .+off32 */ \
99 #define COND_SEL(CODE, TOP, FOP) \
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 */
110 static inline void bpf_flush_icache(void *start, void *end)
112 mm_segment_t old_fs = get_fs();
116 flush_icache_range((unsigned long)start, (unsigned long)end);
121 void bpf_jit_compile(struct sk_filter *fp)
125 unsigned int proglen, oldproglen = 0;
127 int t_offset, f_offset;
128 u8 t_op, f_op, seen = 0, pass;
131 int pc_ret0 = -1; /* bpf index of first RET #0 instruction (if any) */
132 unsigned int cleanup_addr; /* epilogue code offset */
134 const struct sock_filter *filter = fp->insns;
140 addrs = kmalloc(flen * sizeof(*addrs), GFP_KERNEL);
144 /* Before first pass, make a rough estimation of addrs[]
145 * each bpf instruction is translated to less than 64 bytes
147 for (proglen = 0, i = 0; i < flen; i++) {
151 cleanup_addr = proglen; /* epilogue address */
153 for (pass = 0; pass < 10; pass++) {
154 u8 seen_or_pass0 = (pass == 0) ? (SEEN_XREG | SEEN_DATAREF | SEEN_MEM) : seen;
155 /* no prologue/epilogue for trivial filters (RET something) */
160 EMIT4(0x55, 0x48, 0x89, 0xe5); /* push %rbp; mov %rsp,%rbp */
161 EMIT4(0x48, 0x83, 0xec, 96); /* subq $96,%rsp */
162 /* note : must save %rbx in case bpf_error is hit */
163 if (seen_or_pass0 & (SEEN_XREG | SEEN_DATAREF))
164 EMIT4(0x48, 0x89, 0x5d, 0xf8); /* mov %rbx, -8(%rbp) */
165 if (seen_or_pass0 & SEEN_XREG)
166 CLEAR_X(); /* make sure we dont leek kernel memory */
169 * If this filter needs to access skb data,
170 * loads r9 and r8 with :
171 * r9 = skb->len - skb->data_len
174 if (seen_or_pass0 & SEEN_DATAREF) {
175 if (offsetof(struct sk_buff, len) <= 127)
176 /* mov off8(%rdi),%r9d */
177 EMIT4(0x44, 0x8b, 0x4f, offsetof(struct sk_buff, len));
179 /* mov off32(%rdi),%r9d */
180 EMIT3(0x44, 0x8b, 0x8f);
181 EMIT(offsetof(struct sk_buff, len), 4);
183 if (is_imm8(offsetof(struct sk_buff, data_len)))
184 /* sub off8(%rdi),%r9d */
185 EMIT4(0x44, 0x2b, 0x4f, offsetof(struct sk_buff, data_len));
187 EMIT3(0x44, 0x2b, 0x8f);
188 EMIT(offsetof(struct sk_buff, data_len), 4);
191 if (is_imm8(offsetof(struct sk_buff, data)))
192 /* mov off8(%rdi),%r8 */
193 EMIT4(0x4c, 0x8b, 0x47, offsetof(struct sk_buff, data));
195 /* mov off32(%rdi),%r8 */
196 EMIT3(0x4c, 0x8b, 0x87);
197 EMIT(offsetof(struct sk_buff, data), 4);
202 switch (filter[0].code) {
205 case BPF_S_ANC_PROTOCOL:
206 case BPF_S_ANC_IFINDEX:
208 case BPF_S_ANC_RXHASH:
210 case BPF_S_ANC_QUEUE:
214 /* first instruction sets A register (or is RET 'constant') */
217 /* make sure we dont leak kernel information to user */
218 CLEAR_A(); /* A = 0 */
221 for (i = 0; i < flen; i++) {
222 unsigned int K = filter[i].k;
224 switch (filter[i].code) {
225 case BPF_S_ALU_ADD_X: /* A += X; */
227 EMIT2(0x01, 0xd8); /* add %ebx,%eax */
229 case BPF_S_ALU_ADD_K: /* A += K; */
233 EMIT3(0x83, 0xc0, K); /* add imm8,%eax */
235 EMIT1_off32(0x05, K); /* add imm32,%eax */
237 case BPF_S_ALU_SUB_X: /* A -= X; */
239 EMIT2(0x29, 0xd8); /* sub %ebx,%eax */
241 case BPF_S_ALU_SUB_K: /* A -= K */
245 EMIT3(0x83, 0xe8, K); /* sub imm8,%eax */
247 EMIT1_off32(0x2d, K); /* sub imm32,%eax */
249 case BPF_S_ALU_MUL_X: /* A *= X; */
251 EMIT3(0x0f, 0xaf, 0xc3); /* imul %ebx,%eax */
253 case BPF_S_ALU_MUL_K: /* A *= K */
255 EMIT3(0x6b, 0xc0, K); /* imul imm8,%eax,%eax */
257 EMIT2(0x69, 0xc0); /* imul imm32,%eax */
261 case BPF_S_ALU_DIV_X: /* A /= X; */
263 EMIT2(0x85, 0xdb); /* test %ebx,%ebx */
265 /* addrs[pc_ret0 - 1] is start address of target
266 * (addrs[i] - 4) is the address following this jmp
267 * ("xor %edx,%edx; div %ebx" being 4 bytes long)
269 EMIT_COND_JMP(X86_JE, addrs[pc_ret0 - 1] -
272 EMIT_COND_JMP(X86_JNE, 2 + 5);
274 EMIT1_off32(0xe9, cleanup_addr - (addrs[i] - 4)); /* jmp .+off32 */
276 EMIT4(0x31, 0xd2, 0xf7, 0xf3); /* xor %edx,%edx; div %ebx */
278 case BPF_S_ALU_DIV_K: /* A = reciprocal_divide(A, K); */
279 EMIT3(0x48, 0x69, 0xc0); /* imul imm32,%rax,%rax */
281 EMIT4(0x48, 0xc1, 0xe8, 0x20); /* shr $0x20,%rax */
283 case BPF_S_ALU_AND_X:
285 EMIT2(0x21, 0xd8); /* and %ebx,%eax */
287 case BPF_S_ALU_AND_K:
288 if (K >= 0xFFFFFF00) {
289 EMIT2(0x24, K & 0xFF); /* and imm8,%al */
290 } else if (K >= 0xFFFF0000) {
291 EMIT2(0x66, 0x25); /* and imm16,%ax */
294 EMIT1_off32(0x25, K); /* and imm32,%eax */
299 EMIT2(0x09, 0xd8); /* or %ebx,%eax */
303 EMIT3(0x83, 0xc8, K); /* or imm8,%eax */
305 EMIT1_off32(0x0d, K); /* or imm32,%eax */
307 case BPF_S_ALU_LSH_X: /* A <<= X; */
309 EMIT4(0x89, 0xd9, 0xd3, 0xe0); /* mov %ebx,%ecx; shl %cl,%eax */
311 case BPF_S_ALU_LSH_K:
315 EMIT2(0xd1, 0xe0); /* shl %eax */
317 EMIT3(0xc1, 0xe0, K);
319 case BPF_S_ALU_RSH_X: /* A >>= X; */
321 EMIT4(0x89, 0xd9, 0xd3, 0xe8); /* mov %ebx,%ecx; shr %cl,%eax */
323 case BPF_S_ALU_RSH_K: /* A >>= K; */
327 EMIT2(0xd1, 0xe8); /* shr %eax */
329 EMIT3(0xc1, 0xe8, K);
332 EMIT2(0xf7, 0xd8); /* neg %eax */
340 EMIT1_off32(0xb8, K); /* mov $imm32,%eax */
346 EMIT_JMP(cleanup_addr - addrs[i]);
349 if (seen_or_pass0 & SEEN_XREG)
350 EMIT4(0x48, 0x8b, 0x5d, 0xf8); /* mov -8(%rbp),%rbx */
351 EMIT1(0xc9); /* leaveq */
353 EMIT1(0xc3); /* ret */
355 case BPF_S_MISC_TAX: /* X = A */
357 EMIT2(0x89, 0xc3); /* mov %eax,%ebx */
359 case BPF_S_MISC_TXA: /* A = X */
361 EMIT2(0x89, 0xd8); /* mov %ebx,%eax */
363 case BPF_S_LD_IMM: /* A = K */
367 EMIT1_off32(0xb8, K); /* mov $imm32,%eax */
369 case BPF_S_LDX_IMM: /* X = K */
374 EMIT1_off32(0xbb, K); /* mov $imm32,%ebx */
376 case BPF_S_LD_MEM: /* A = mem[K] : mov off8(%rbp),%eax */
378 EMIT3(0x8b, 0x45, 0xf0 - K*4);
380 case BPF_S_LDX_MEM: /* X = mem[K] : mov off8(%rbp),%ebx */
381 seen |= SEEN_XREG | SEEN_MEM;
382 EMIT3(0x8b, 0x5d, 0xf0 - K*4);
384 case BPF_S_ST: /* mem[K] = A : mov %eax,off8(%rbp) */
386 EMIT3(0x89, 0x45, 0xf0 - K*4);
388 case BPF_S_STX: /* mem[K] = X : mov %ebx,off8(%rbp) */
389 seen |= SEEN_XREG | SEEN_MEM;
390 EMIT3(0x89, 0x5d, 0xf0 - K*4);
392 case BPF_S_LD_W_LEN: /* A = skb->len; */
393 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
394 if (is_imm8(offsetof(struct sk_buff, len)))
395 /* mov off8(%rdi),%eax */
396 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, len));
399 EMIT(offsetof(struct sk_buff, len), 4);
402 case BPF_S_LDX_W_LEN: /* X = skb->len; */
404 if (is_imm8(offsetof(struct sk_buff, len)))
405 /* mov off8(%rdi),%ebx */
406 EMIT3(0x8b, 0x5f, offsetof(struct sk_buff, len));
409 EMIT(offsetof(struct sk_buff, len), 4);
412 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
413 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
414 if (is_imm8(offsetof(struct sk_buff, protocol))) {
415 /* movzwl off8(%rdi),%eax */
416 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, protocol));
418 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */
419 EMIT(offsetof(struct sk_buff, protocol), 4);
421 EMIT2(0x86, 0xc4); /* ntohs() : xchg %al,%ah */
423 case BPF_S_ANC_IFINDEX:
424 if (is_imm8(offsetof(struct sk_buff, dev))) {
425 /* movq off8(%rdi),%rax */
426 EMIT4(0x48, 0x8b, 0x47, offsetof(struct sk_buff, dev));
428 EMIT3(0x48, 0x8b, 0x87); /* movq off32(%rdi),%rax */
429 EMIT(offsetof(struct sk_buff, dev), 4);
431 EMIT3(0x48, 0x85, 0xc0); /* test %rax,%rax */
432 EMIT_COND_JMP(X86_JE, cleanup_addr - (addrs[i] - 6));
433 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
434 EMIT2(0x8b, 0x80); /* mov off32(%rax),%eax */
435 EMIT(offsetof(struct net_device, ifindex), 4);
438 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
439 if (is_imm8(offsetof(struct sk_buff, mark))) {
440 /* mov off8(%rdi),%eax */
441 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, mark));
444 EMIT(offsetof(struct sk_buff, mark), 4);
447 case BPF_S_ANC_RXHASH:
448 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4);
449 if (is_imm8(offsetof(struct sk_buff, rxhash))) {
450 /* mov off8(%rdi),%eax */
451 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, rxhash));
454 EMIT(offsetof(struct sk_buff, rxhash), 4);
457 case BPF_S_ANC_QUEUE:
458 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
459 if (is_imm8(offsetof(struct sk_buff, queue_mapping))) {
460 /* movzwl off8(%rdi),%eax */
461 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, queue_mapping));
463 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */
464 EMIT(offsetof(struct sk_buff, queue_mapping), 4);
469 EMIT4(0x65, 0x8b, 0x04, 0x25); /* mov %gs:off32,%eax */
470 EMIT((u32)(unsigned long)&cpu_number, 4); /* A = smp_processor_id(); */
477 common_load: seen |= SEEN_DATAREF;
480 t_offset = func - (image + addrs[i]);
481 EMIT1_off32(0xbe, K); /* mov imm32,%esi */
482 EMIT1_off32(0xe8, t_offset); /* call */
490 case BPF_S_LDX_B_MSH:
493 /* addrs[pc_ret0 - 1] is the start address */
494 EMIT_JMP(addrs[pc_ret0 - 1] - addrs[i]);
498 EMIT_JMP(cleanup_addr - addrs[i]);
501 seen |= SEEN_DATAREF | SEEN_XREG;
502 t_offset = sk_load_byte_msh - (image + addrs[i]);
503 EMIT1_off32(0xbe, K); /* mov imm32,%esi */
504 EMIT1_off32(0xe8, t_offset); /* call sk_load_byte_msh */
507 func = sk_load_word_ind;
508 common_load_ind: seen |= SEEN_DATAREF | SEEN_XREG;
509 t_offset = func - (image + addrs[i]);
510 EMIT1_off32(0xbe, K); /* mov imm32,%esi */
511 EMIT1_off32(0xe8, t_offset); /* call sk_load_xxx_ind */
514 func = sk_load_half_ind;
515 goto common_load_ind;
517 func = sk_load_byte_ind;
518 goto common_load_ind;
520 t_offset = addrs[i + K] - addrs[i];
523 COND_SEL(BPF_S_JMP_JGT_K, X86_JA, X86_JBE);
524 COND_SEL(BPF_S_JMP_JGE_K, X86_JAE, X86_JB);
525 COND_SEL(BPF_S_JMP_JEQ_K, X86_JE, X86_JNE);
526 COND_SEL(BPF_S_JMP_JSET_K,X86_JNE, X86_JE);
527 COND_SEL(BPF_S_JMP_JGT_X, X86_JA, X86_JBE);
528 COND_SEL(BPF_S_JMP_JGE_X, X86_JAE, X86_JB);
529 COND_SEL(BPF_S_JMP_JEQ_X, X86_JE, X86_JNE);
530 COND_SEL(BPF_S_JMP_JSET_X,X86_JNE, X86_JE);
532 cond_branch: f_offset = addrs[i + filter[i].jf] - addrs[i];
533 t_offset = addrs[i + filter[i].jt] - addrs[i];
535 /* same targets, can avoid doing the test :) */
536 if (filter[i].jt == filter[i].jf) {
541 switch (filter[i].code) {
542 case BPF_S_JMP_JGT_X:
543 case BPF_S_JMP_JGE_X:
544 case BPF_S_JMP_JEQ_X:
546 EMIT2(0x39, 0xd8); /* cmp %ebx,%eax */
548 case BPF_S_JMP_JSET_X:
550 EMIT2(0x85, 0xd8); /* test %ebx,%eax */
552 case BPF_S_JMP_JEQ_K:
554 EMIT2(0x85, 0xc0); /* test %eax,%eax */
557 case BPF_S_JMP_JGT_K:
558 case BPF_S_JMP_JGE_K:
560 EMIT3(0x83, 0xf8, K); /* cmp imm8,%eax */
562 EMIT1_off32(0x3d, K); /* cmp imm32,%eax */
564 case BPF_S_JMP_JSET_K:
566 EMIT2(0xa8, K); /* test imm8,%al */
567 else if (!(K & 0xFFFF00FF))
568 EMIT3(0xf6, 0xc4, K >> 8); /* test imm8,%ah */
569 else if (K <= 0xFFFF) {
570 EMIT2(0x66, 0xa9); /* test imm16,%ax */
573 EMIT1_off32(0xa9, K); /* test imm32,%eax */
577 if (filter[i].jt != 0) {
578 if (filter[i].jf && f_offset)
579 t_offset += is_near(f_offset) ? 2 : 5;
580 EMIT_COND_JMP(t_op, t_offset);
585 EMIT_COND_JMP(f_op, f_offset);
588 /* hmm, too complex filter, give up with jit compiler */
593 if (unlikely(proglen + ilen > oldproglen)) {
594 pr_err("bpb_jit_compile fatal error\n");
596 module_free(NULL, image);
599 memcpy(image + proglen, temp, ilen);
605 /* last bpf instruction is always a RET :
606 * use it to give the cleanup instruction(s) addr
608 cleanup_addr = proglen - 1; /* ret */
610 cleanup_addr -= 1; /* leaveq */
611 if (seen_or_pass0 & SEEN_XREG)
612 cleanup_addr -= 4; /* mov -8(%rbp),%rbx */
615 if (proglen != oldproglen)
616 pr_err("bpb_jit_compile proglen=%u != oldproglen=%u\n", proglen, oldproglen);
619 if (proglen == oldproglen) {
620 image = module_alloc(max_t(unsigned int,
622 sizeof(struct work_struct)));
626 oldproglen = proglen;
628 if (bpf_jit_enable > 1)
629 pr_err("flen=%d proglen=%u pass=%d image=%p\n",
630 flen, proglen, pass, image);
633 if (bpf_jit_enable > 1)
634 print_hex_dump(KERN_ERR, "JIT code: ", DUMP_PREFIX_ADDRESS,
635 16, 1, image, proglen, false);
637 bpf_flush_icache(image, image + proglen);
639 fp->bpf_func = (void *)image;
646 static void jit_free_defer(struct work_struct *arg)
648 module_free(NULL, arg);
651 /* run from softirq, we must use a work_struct to call
652 * module_free() from process context
654 void bpf_jit_free(struct sk_filter *fp)
656 if (fp->bpf_func != sk_run_filter) {
657 struct work_struct *work = (struct work_struct *)fp->bpf_func;
659 INIT_WORK(work, jit_free_defer);