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
/* no prologue/epilogue for trivial filters (RET something) */
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 */
168
* If this filter needs to access skb data,
169
* loads r9 and r8 with :
170
* r9 = skb->len - skb->data_len
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));
178
/* mov off32(%rdi),%r9d */
179
EMIT3(0x44, 0x8b, 0x8f);
180
EMIT(offsetof(struct sk_buff, len), 4);
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));
186
EMIT3(0x44, 0x2b, 0x8f);
187
EMIT(offsetof(struct sk_buff, data_len), 4);
190
if (is_imm8(offsetof(struct sk_buff, data)))
191
/* mov off8(%rdi),%r8 */
192
EMIT4(0x4c, 0x8b, 0x47, offsetof(struct sk_buff, data));
194
/* mov off32(%rdi),%r8 */
195
EMIT3(0x4c, 0x8b, 0x87);
196
EMIT(offsetof(struct sk_buff, data), 4);
201
switch (filter[0].code) {
204
case BPF_S_ANC_PROTOCOL:
205
case BPF_S_ANC_IFINDEX:
207
case BPF_S_ANC_RXHASH:
209
case BPF_S_ANC_QUEUE:
213
/* first instruction sets A register (or is RET 'constant') */
216
/* make sure we dont leak kernel information to user */
217
CLEAR_A(); /* A = 0 */
220
for (i = 0; i < flen; i++) {
221
unsigned int K = filter[i].k;
223
switch (filter[i].code) {
224
case BPF_S_ALU_ADD_X: /* A += X; */
226
EMIT2(0x01, 0xd8); /* add %ebx,%eax */
228
case BPF_S_ALU_ADD_K: /* A += K; */
232
EMIT3(0x83, 0xc0, K); /* add imm8,%eax */
234
EMIT1_off32(0x05, K); /* add imm32,%eax */
236
case BPF_S_ALU_SUB_X: /* A -= X; */
238
EMIT2(0x29, 0xd8); /* sub %ebx,%eax */
240
case BPF_S_ALU_SUB_K: /* A -= K */
244
EMIT3(0x83, 0xe8, K); /* sub imm8,%eax */
246
EMIT1_off32(0x2d, K); /* sub imm32,%eax */
248
case BPF_S_ALU_MUL_X: /* A *= X; */
250
EMIT3(0x0f, 0xaf, 0xc3); /* imul %ebx,%eax */
252
case BPF_S_ALU_MUL_K: /* A *= K */
254
EMIT3(0x6b, 0xc0, K); /* imul imm8,%eax,%eax */
256
EMIT2(0x69, 0xc0); /* imul imm32,%eax */
260
case BPF_S_ALU_DIV_X: /* A /= X; */
262
EMIT2(0x85, 0xdb); /* test %ebx,%ebx */
264
EMIT_COND_JMP(X86_JE, addrs[pc_ret0] - (addrs[i] - 4));
266
EMIT_COND_JMP(X86_JNE, 2 + 5);
268
EMIT1_off32(0xe9, cleanup_addr - (addrs[i] - 4)); /* jmp .+off32 */
270
EMIT4(0x31, 0xd2, 0xf7, 0xf3); /* xor %edx,%edx; div %ebx */
272
case BPF_S_ALU_DIV_K: /* A = reciprocal_divide(A, K); */
273
EMIT3(0x48, 0x69, 0xc0); /* imul imm32,%rax,%rax */
275
EMIT4(0x48, 0xc1, 0xe8, 0x20); /* shr $0x20,%rax */
277
case BPF_S_ALU_AND_X:
279
EMIT2(0x21, 0xd8); /* and %ebx,%eax */
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 */
288
EMIT1_off32(0x25, K); /* and imm32,%eax */
293
EMIT2(0x09, 0xd8); /* or %ebx,%eax */
297
EMIT3(0x83, 0xc8, K); /* or imm8,%eax */
299
EMIT1_off32(0x0d, K); /* or imm32,%eax */
301
case BPF_S_ALU_LSH_X: /* A <<= X; */
303
EMIT4(0x89, 0xd9, 0xd3, 0xe0); /* mov %ebx,%ecx; shl %cl,%eax */
305
case BPF_S_ALU_LSH_K:
309
EMIT2(0xd1, 0xe0); /* shl %eax */
311
EMIT3(0xc1, 0xe0, K);
313
case BPF_S_ALU_RSH_X: /* A >>= X; */
315
EMIT4(0x89, 0xd9, 0xd3, 0xe8); /* mov %ebx,%ecx; shr %cl,%eax */
317
case BPF_S_ALU_RSH_K: /* A >>= K; */
321
EMIT2(0xd1, 0xe8); /* shr %eax */
323
EMIT3(0xc1, 0xe8, K);
326
EMIT2(0xf7, 0xd8); /* neg %eax */
334
EMIT1_off32(0xb8, K); /* mov $imm32,%eax */
340
EMIT_JMP(cleanup_addr - addrs[i]);
343
if (seen & SEEN_XREG)
344
EMIT4(0x48, 0x8b, 0x5d, 0xf8); /* mov -8(%rbp),%rbx */
345
EMIT1(0xc9); /* leaveq */
347
EMIT1(0xc3); /* ret */
349
case BPF_S_MISC_TAX: /* X = A */
351
EMIT2(0x89, 0xc3); /* mov %eax,%ebx */
353
case BPF_S_MISC_TXA: /* A = X */
355
EMIT2(0x89, 0xd8); /* mov %ebx,%eax */
357
case BPF_S_LD_IMM: /* A = K */
361
EMIT1_off32(0xb8, K); /* mov $imm32,%eax */
363
case BPF_S_LDX_IMM: /* X = K */
368
EMIT1_off32(0xbb, K); /* mov $imm32,%ebx */
370
case BPF_S_LD_MEM: /* A = mem[K] : mov off8(%rbp),%eax */
372
EMIT3(0x8b, 0x45, 0xf0 - K*4);
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);
378
case BPF_S_ST: /* mem[K] = A : mov %eax,off8(%rbp) */
380
EMIT3(0x89, 0x45, 0xf0 - K*4);
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);
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));
393
EMIT(offsetof(struct sk_buff, len), 4);
396
case BPF_S_LDX_W_LEN: /* X = skb->len; */
398
if (is_imm8(offsetof(struct sk_buff, len)))
399
/* mov off8(%rdi),%ebx */
400
EMIT3(0x8b, 0x5f, offsetof(struct sk_buff, len));
403
EMIT(offsetof(struct sk_buff, len), 4);
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));
412
EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */
413
EMIT(offsetof(struct sk_buff, protocol), 4);
415
EMIT2(0x86, 0xc4); /* ntohs() : xchg %al,%ah */
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));
422
EMIT3(0x48, 0x8b, 0x87); /* movq off32(%rdi),%rax */
423
EMIT(offsetof(struct sk_buff, dev), 4);
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);
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));
438
EMIT(offsetof(struct sk_buff, mark), 4);
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));
448
EMIT(offsetof(struct sk_buff, rxhash), 4);
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));
457
EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */
458
EMIT(offsetof(struct sk_buff, queue_mapping), 4);
463
EMIT4(0x65, 0x8b, 0x04, 0x25); /* mov %gs:off32,%eax */
464
EMIT((u32)(unsigned long)&cpu_number, 4); /* A = smp_processor_id(); */
471
common_load: seen |= SEEN_DATAREF;
474
t_offset = func - (image + addrs[i]);
475
EMIT1_off32(0xbe, K); /* mov imm32,%esi */
476
EMIT1_off32(0xe8, t_offset); /* call */
484
case BPF_S_LDX_B_MSH:
487
EMIT_JMP(addrs[pc_ret0] - addrs[i]);
491
EMIT_JMP(cleanup_addr - addrs[i]);
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 */
500
func = sk_load_word_ind;
501
common_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 */
507
func = sk_load_half_ind;
508
goto common_load_ind;
510
func = sk_load_byte_ind;
511
goto common_load_ind;
513
t_offset = addrs[i + K] - addrs[i];
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);
525
cond_branch: f_offset = addrs[i + filter[i].jf] - addrs[i];
526
t_offset = addrs[i + filter[i].jt] - addrs[i];
528
/* same targets, can avoid doing the test :) */
529
if (filter[i].jt == filter[i].jf) {
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:
539
EMIT2(0x39, 0xd8); /* cmp %ebx,%eax */
541
case BPF_S_JMP_JSET_X:
543
EMIT2(0x85, 0xd8); /* test %ebx,%eax */
545
case BPF_S_JMP_JEQ_K:
547
EMIT2(0x85, 0xc0); /* test %eax,%eax */
550
case BPF_S_JMP_JGT_K:
551
case BPF_S_JMP_JGE_K:
553
EMIT3(0x83, 0xf8, K); /* cmp imm8,%eax */
555
EMIT1_off32(0x3d, K); /* cmp imm32,%eax */
557
case BPF_S_JMP_JSET_K:
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 */
566
EMIT1_off32(0xa9, K); /* test imm32,%eax */
570
if (filter[i].jt != 0) {
572
t_offset += is_near(f_offset) ? 2 : 6;
573
EMIT_COND_JMP(t_op, t_offset);
578
EMIT_COND_JMP(f_op, f_offset);
581
/* hmm, too complex filter, give up with jit compiler */
586
if (unlikely(proglen + ilen > oldproglen)) {
587
pr_err("bpb_jit_compile fatal error\n");
589
module_free(NULL, image);
592
memcpy(image + proglen, temp, ilen);
598
/* last bpf instruction is always a RET :
599
* use it to give the cleanup instruction(s) addr
601
cleanup_addr = proglen - 1; /* ret */
603
cleanup_addr -= 1; /* leaveq */
604
if (seen & SEEN_XREG)
605
cleanup_addr -= 4; /* mov -8(%rbp),%rbx */
608
WARN_ON(proglen != oldproglen);
611
if (proglen == oldproglen) {
612
image = module_alloc(max_t(unsigned int,
614
sizeof(struct work_struct)));
618
oldproglen = proglen;
620
if (bpf_jit_enable > 1)
621
pr_err("flen=%d proglen=%u pass=%d image=%p\n",
622
flen, proglen, pass, image);
625
if (bpf_jit_enable > 1)
626
print_hex_dump(KERN_ERR, "JIT code: ", DUMP_PREFIX_ADDRESS,
627
16, 1, image, proglen, false);
629
bpf_flush_icache(image, image + proglen);
631
fp->bpf_func = (void *)image;
638
static void jit_free_defer(struct work_struct *arg)
640
module_free(NULL, arg);
643
/* run from softirq, we must use a work_struct to call
644
* module_free() from process context
646
void bpf_jit_free(struct sk_filter *fp)
648
if (fp->bpf_func != sk_run_filter) {
649
struct work_struct *work = (struct work_struct *)fp->bpf_func;
651
INIT_WORK(work, jit_free_defer);