Loading...
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 <linux/bpf.h>
15
16#include <asm/set_memory.h>
17#include <asm/nospec-branch.h>
18
19/*
20 * assembly code in arch/x86/net/bpf_jit.S
21 */
22extern u8 sk_load_word[], sk_load_half[], sk_load_byte[];
23extern u8 sk_load_word_positive_offset[], sk_load_half_positive_offset[];
24extern u8 sk_load_byte_positive_offset[];
25extern u8 sk_load_word_negative_offset[], sk_load_half_negative_offset[];
26extern u8 sk_load_byte_negative_offset[];
27
28static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
29{
30 if (len == 1)
31 *ptr = bytes;
32 else if (len == 2)
33 *(u16 *)ptr = bytes;
34 else {
35 *(u32 *)ptr = bytes;
36 barrier();
37 }
38 return ptr + len;
39}
40
41#define EMIT(bytes, len) \
42 do { prog = emit_code(prog, bytes, len); cnt += len; } while (0)
43
44#define EMIT1(b1) EMIT(b1, 1)
45#define EMIT2(b1, b2) EMIT((b1) + ((b2) << 8), 2)
46#define EMIT3(b1, b2, b3) EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
47#define EMIT4(b1, b2, b3, b4) EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
48#define EMIT1_off32(b1, off) \
49 do {EMIT1(b1); EMIT(off, 4); } while (0)
50#define EMIT2_off32(b1, b2, off) \
51 do {EMIT2(b1, b2); EMIT(off, 4); } while (0)
52#define EMIT3_off32(b1, b2, b3, off) \
53 do {EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
54#define EMIT4_off32(b1, b2, b3, b4, off) \
55 do {EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
56
57static bool is_imm8(int value)
58{
59 return value <= 127 && value >= -128;
60}
61
62static bool is_simm32(s64 value)
63{
64 return value == (s64)(s32)value;
65}
66
67static bool is_uimm32(u64 value)
68{
69 return value == (u64)(u32)value;
70}
71
72/* mov dst, src */
73#define EMIT_mov(DST, SRC) \
74 do {if (DST != SRC) \
75 EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
76 } while (0)
77
78static int bpf_size_to_x86_bytes(int bpf_size)
79{
80 if (bpf_size == BPF_W)
81 return 4;
82 else if (bpf_size == BPF_H)
83 return 2;
84 else if (bpf_size == BPF_B)
85 return 1;
86 else if (bpf_size == BPF_DW)
87 return 4; /* imm32 */
88 else
89 return 0;
90}
91
92/* list of x86 cond jumps opcodes (. + s8)
93 * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
94 */
95#define X86_JB 0x72
96#define X86_JAE 0x73
97#define X86_JE 0x74
98#define X86_JNE 0x75
99#define X86_JBE 0x76
100#define X86_JA 0x77
101#define X86_JL 0x7C
102#define X86_JGE 0x7D
103#define X86_JLE 0x7E
104#define X86_JG 0x7F
105
106#define CHOOSE_LOAD_FUNC(K, func) \
107 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
108
109/* pick a register outside of BPF range for JIT internal work */
110#define AUX_REG (MAX_BPF_JIT_REG + 1)
111
112/* The following table maps BPF registers to x64 registers.
113 *
114 * x64 register r12 is unused, since if used as base address
115 * register in load/store instructions, it always needs an
116 * extra byte of encoding and is callee saved.
117 *
118 * r9 caches skb->len - skb->data_len
119 * r10 caches skb->data, and used for blinding (if enabled)
120 */
121static const int reg2hex[] = {
122 [BPF_REG_0] = 0, /* rax */
123 [BPF_REG_1] = 7, /* rdi */
124 [BPF_REG_2] = 6, /* rsi */
125 [BPF_REG_3] = 2, /* rdx */
126 [BPF_REG_4] = 1, /* rcx */
127 [BPF_REG_5] = 0, /* r8 */
128 [BPF_REG_6] = 3, /* rbx callee saved */
129 [BPF_REG_7] = 5, /* r13 callee saved */
130 [BPF_REG_8] = 6, /* r14 callee saved */
131 [BPF_REG_9] = 7, /* r15 callee saved */
132 [BPF_REG_FP] = 5, /* rbp readonly */
133 [BPF_REG_AX] = 2, /* r10 temp register */
134 [AUX_REG] = 3, /* r11 temp register */
135};
136
137/* is_ereg() == true if BPF register 'reg' maps to x64 r8..r15
138 * which need extra byte of encoding.
139 * rax,rcx,...,rbp have simpler encoding
140 */
141static bool is_ereg(u32 reg)
142{
143 return (1 << reg) & (BIT(BPF_REG_5) |
144 BIT(AUX_REG) |
145 BIT(BPF_REG_7) |
146 BIT(BPF_REG_8) |
147 BIT(BPF_REG_9) |
148 BIT(BPF_REG_AX));
149}
150
151static bool is_axreg(u32 reg)
152{
153 return reg == BPF_REG_0;
154}
155
156/* add modifiers if 'reg' maps to x64 registers r8..r15 */
157static u8 add_1mod(u8 byte, u32 reg)
158{
159 if (is_ereg(reg))
160 byte |= 1;
161 return byte;
162}
163
164static u8 add_2mod(u8 byte, u32 r1, u32 r2)
165{
166 if (is_ereg(r1))
167 byte |= 1;
168 if (is_ereg(r2))
169 byte |= 4;
170 return byte;
171}
172
173/* encode 'dst_reg' register into x64 opcode 'byte' */
174static u8 add_1reg(u8 byte, u32 dst_reg)
175{
176 return byte + reg2hex[dst_reg];
177}
178
179/* encode 'dst_reg' and 'src_reg' registers into x64 opcode 'byte' */
180static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
181{
182 return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
183}
184
185static void jit_fill_hole(void *area, unsigned int size)
186{
187 /* fill whole space with int3 instructions */
188 memset(area, 0xcc, size);
189}
190
191struct jit_context {
192 int cleanup_addr; /* epilogue code offset */
193 bool seen_ld_abs;
194 bool seen_ax_reg;
195};
196
197/* maximum number of bytes emitted while JITing one eBPF insn */
198#define BPF_MAX_INSN_SIZE 128
199#define BPF_INSN_SAFETY 64
200
201#define AUX_STACK_SPACE \
202 (32 /* space for rbx, r13, r14, r15 */ + \
203 8 /* space for skb_copy_bits() buffer */)
204
205#define PROLOGUE_SIZE 37
206
207/* emit x64 prologue code for BPF program and check it's size.
208 * bpf_tail_call helper will skip it while jumping into another program
209 */
210static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf)
211{
212 u8 *prog = *pprog;
213 int cnt = 0;
214
215 EMIT1(0x55); /* push rbp */
216 EMIT3(0x48, 0x89, 0xE5); /* mov rbp,rsp */
217
218 /* sub rsp, rounded_stack_depth + AUX_STACK_SPACE */
219 EMIT3_off32(0x48, 0x81, 0xEC,
220 round_up(stack_depth, 8) + AUX_STACK_SPACE);
221
222 /* sub rbp, AUX_STACK_SPACE */
223 EMIT4(0x48, 0x83, 0xED, AUX_STACK_SPACE);
224
225 /* all classic BPF filters use R6(rbx) save it */
226
227 /* mov qword ptr [rbp+0],rbx */
228 EMIT4(0x48, 0x89, 0x5D, 0);
229
230 /* bpf_convert_filter() maps classic BPF register X to R7 and uses R8
231 * as temporary, so all tcpdump filters need to spill/fill R7(r13) and
232 * R8(r14). R9(r15) spill could be made conditional, but there is only
233 * one 'bpf_error' return path out of helper functions inside bpf_jit.S
234 * The overhead of extra spill is negligible for any filter other
235 * than synthetic ones. Therefore not worth adding complexity.
236 */
237
238 /* mov qword ptr [rbp+8],r13 */
239 EMIT4(0x4C, 0x89, 0x6D, 8);
240 /* mov qword ptr [rbp+16],r14 */
241 EMIT4(0x4C, 0x89, 0x75, 16);
242 /* mov qword ptr [rbp+24],r15 */
243 EMIT4(0x4C, 0x89, 0x7D, 24);
244
245 if (!ebpf_from_cbpf) {
246 /* Clear the tail call counter (tail_call_cnt): for eBPF tail
247 * calls we need to reset the counter to 0. It's done in two
248 * instructions, resetting rax register to 0, and moving it
249 * to the counter location.
250 */
251
252 /* xor eax, eax */
253 EMIT2(0x31, 0xc0);
254 /* mov qword ptr [rbp+32], rax */
255 EMIT4(0x48, 0x89, 0x45, 32);
256
257 BUILD_BUG_ON(cnt != PROLOGUE_SIZE);
258 }
259
260 *pprog = prog;
261}
262
263/* generate the following code:
264 * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
265 * if (index >= array->map.max_entries)
266 * goto out;
267 * if (++tail_call_cnt > MAX_TAIL_CALL_CNT)
268 * goto out;
269 * prog = array->ptrs[index];
270 * if (prog == NULL)
271 * goto out;
272 * goto *(prog->bpf_func + prologue_size);
273 * out:
274 */
275static void emit_bpf_tail_call(u8 **pprog)
276{
277 u8 *prog = *pprog;
278 int label1, label2, label3;
279 int cnt = 0;
280
281 /* rdi - pointer to ctx
282 * rsi - pointer to bpf_array
283 * rdx - index in bpf_array
284 */
285
286 /* if (index >= array->map.max_entries)
287 * goto out;
288 */
289 EMIT2(0x89, 0xD2); /* mov edx, edx */
290 EMIT3(0x39, 0x56, /* cmp dword ptr [rsi + 16], edx */
291 offsetof(struct bpf_array, map.max_entries));
292#define OFFSET1 (41 + RETPOLINE_RAX_BPF_JIT_SIZE) /* number of bytes to jump */
293 EMIT2(X86_JBE, OFFSET1); /* jbe out */
294 label1 = cnt;
295
296 /* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
297 * goto out;
298 */
299 EMIT2_off32(0x8B, 0x85, 36); /* mov eax, dword ptr [rbp + 36] */
300 EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */
301#define OFFSET2 (30 + RETPOLINE_RAX_BPF_JIT_SIZE)
302 EMIT2(X86_JA, OFFSET2); /* ja out */
303 label2 = cnt;
304 EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */
305 EMIT2_off32(0x89, 0x85, 36); /* mov dword ptr [rbp + 36], eax */
306
307 /* prog = array->ptrs[index]; */
308 EMIT4_off32(0x48, 0x8B, 0x84, 0xD6, /* mov rax, [rsi + rdx * 8 + offsetof(...)] */
309 offsetof(struct bpf_array, ptrs));
310
311 /* if (prog == NULL)
312 * goto out;
313 */
314 EMIT3(0x48, 0x85, 0xC0); /* test rax,rax */
315#define OFFSET3 (8 + RETPOLINE_RAX_BPF_JIT_SIZE)
316 EMIT2(X86_JE, OFFSET3); /* je out */
317 label3 = cnt;
318
319 /* goto *(prog->bpf_func + prologue_size); */
320 EMIT4(0x48, 0x8B, 0x40, /* mov rax, qword ptr [rax + 32] */
321 offsetof(struct bpf_prog, bpf_func));
322 EMIT4(0x48, 0x83, 0xC0, PROLOGUE_SIZE); /* add rax, prologue_size */
323
324 /* now we're ready to jump into next BPF program
325 * rdi == ctx (1st arg)
326 * rax == prog->bpf_func + prologue_size
327 */
328 RETPOLINE_RAX_BPF_JIT();
329
330 /* out: */
331 BUILD_BUG_ON(cnt - label1 != OFFSET1);
332 BUILD_BUG_ON(cnt - label2 != OFFSET2);
333 BUILD_BUG_ON(cnt - label3 != OFFSET3);
334 *pprog = prog;
335}
336
337
338static void emit_load_skb_data_hlen(u8 **pprog)
339{
340 u8 *prog = *pprog;
341 int cnt = 0;
342
343 /* r9d = skb->len - skb->data_len (headlen)
344 * r10 = skb->data
345 */
346 /* mov %r9d, off32(%rdi) */
347 EMIT3_off32(0x44, 0x8b, 0x8f, offsetof(struct sk_buff, len));
348
349 /* sub %r9d, off32(%rdi) */
350 EMIT3_off32(0x44, 0x2b, 0x8f, offsetof(struct sk_buff, data_len));
351
352 /* mov %r10, off32(%rdi) */
353 EMIT3_off32(0x4c, 0x8b, 0x97, offsetof(struct sk_buff, data));
354 *pprog = prog;
355}
356
357static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
358 u32 dst_reg, const u32 imm32)
359{
360 u8 *prog = *pprog;
361 u8 b1, b2, b3;
362 int cnt = 0;
363
364 /* optimization: if imm32 is positive, use 'mov %eax, imm32'
365 * (which zero-extends imm32) to save 2 bytes.
366 */
367 if (sign_propagate && (s32)imm32 < 0) {
368 /* 'mov %rax, imm32' sign extends imm32 */
369 b1 = add_1mod(0x48, dst_reg);
370 b2 = 0xC7;
371 b3 = 0xC0;
372 EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
373 goto done;
374 }
375
376 /* optimization: if imm32 is zero, use 'xor %eax, %eax'
377 * to save 3 bytes.
378 */
379 if (imm32 == 0) {
380 if (is_ereg(dst_reg))
381 EMIT1(add_2mod(0x40, dst_reg, dst_reg));
382 b2 = 0x31; /* xor */
383 b3 = 0xC0;
384 EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
385 goto done;
386 }
387
388 /* mov %eax, imm32 */
389 if (is_ereg(dst_reg))
390 EMIT1(add_1mod(0x40, dst_reg));
391 EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
392done:
393 *pprog = prog;
394}
395
396static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
397 const u32 imm32_hi, const u32 imm32_lo)
398{
399 u8 *prog = *pprog;
400 int cnt = 0;
401
402 if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
403 /* For emitting plain u32, where sign bit must not be
404 * propagated LLVM tends to load imm64 over mov32
405 * directly, so save couple of bytes by just doing
406 * 'mov %eax, imm32' instead.
407 */
408 emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
409 } else {
410 /* movabsq %rax, imm64 */
411 EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
412 EMIT(imm32_lo, 4);
413 EMIT(imm32_hi, 4);
414 }
415
416 *pprog = prog;
417}
418
419static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
420{
421 u8 *prog = *pprog;
422 int cnt = 0;
423
424 if (is64) {
425 /* mov dst, src */
426 EMIT_mov(dst_reg, src_reg);
427 } else {
428 /* mov32 dst, src */
429 if (is_ereg(dst_reg) || is_ereg(src_reg))
430 EMIT1(add_2mod(0x40, dst_reg, src_reg));
431 EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
432 }
433
434 *pprog = prog;
435}
436
437static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
438 int oldproglen, struct jit_context *ctx)
439{
440 struct bpf_insn *insn = bpf_prog->insnsi;
441 int insn_cnt = bpf_prog->len;
442 bool seen_ld_abs = ctx->seen_ld_abs | (oldproglen == 0);
443 bool seen_ax_reg = ctx->seen_ax_reg | (oldproglen == 0);
444 bool seen_exit = false;
445 u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
446 int i, cnt = 0;
447 int proglen = 0;
448 u8 *prog = temp;
449
450 emit_prologue(&prog, bpf_prog->aux->stack_depth,
451 bpf_prog_was_classic(bpf_prog));
452
453 if (seen_ld_abs)
454 emit_load_skb_data_hlen(&prog);
455
456 for (i = 0; i < insn_cnt; i++, insn++) {
457 const s32 imm32 = insn->imm;
458 u32 dst_reg = insn->dst_reg;
459 u32 src_reg = insn->src_reg;
460 u8 b2 = 0, b3 = 0;
461 s64 jmp_offset;
462 u8 jmp_cond;
463 bool reload_skb_data;
464 int ilen;
465 u8 *func;
466
467 if (dst_reg == BPF_REG_AX || src_reg == BPF_REG_AX)
468 ctx->seen_ax_reg = seen_ax_reg = true;
469
470 switch (insn->code) {
471 /* ALU */
472 case BPF_ALU | BPF_ADD | BPF_X:
473 case BPF_ALU | BPF_SUB | BPF_X:
474 case BPF_ALU | BPF_AND | BPF_X:
475 case BPF_ALU | BPF_OR | BPF_X:
476 case BPF_ALU | BPF_XOR | BPF_X:
477 case BPF_ALU64 | BPF_ADD | BPF_X:
478 case BPF_ALU64 | BPF_SUB | BPF_X:
479 case BPF_ALU64 | BPF_AND | BPF_X:
480 case BPF_ALU64 | BPF_OR | BPF_X:
481 case BPF_ALU64 | BPF_XOR | BPF_X:
482 switch (BPF_OP(insn->code)) {
483 case BPF_ADD: b2 = 0x01; break;
484 case BPF_SUB: b2 = 0x29; break;
485 case BPF_AND: b2 = 0x21; break;
486 case BPF_OR: b2 = 0x09; break;
487 case BPF_XOR: b2 = 0x31; break;
488 }
489 if (BPF_CLASS(insn->code) == BPF_ALU64)
490 EMIT1(add_2mod(0x48, dst_reg, src_reg));
491 else if (is_ereg(dst_reg) || is_ereg(src_reg))
492 EMIT1(add_2mod(0x40, dst_reg, src_reg));
493 EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
494 break;
495
496 case BPF_ALU64 | BPF_MOV | BPF_X:
497 case BPF_ALU | BPF_MOV | BPF_X:
498 emit_mov_reg(&prog,
499 BPF_CLASS(insn->code) == BPF_ALU64,
500 dst_reg, src_reg);
501 break;
502
503 /* neg dst */
504 case BPF_ALU | BPF_NEG:
505 case BPF_ALU64 | BPF_NEG:
506 if (BPF_CLASS(insn->code) == BPF_ALU64)
507 EMIT1(add_1mod(0x48, dst_reg));
508 else if (is_ereg(dst_reg))
509 EMIT1(add_1mod(0x40, dst_reg));
510 EMIT2(0xF7, add_1reg(0xD8, dst_reg));
511 break;
512
513 case BPF_ALU | BPF_ADD | BPF_K:
514 case BPF_ALU | BPF_SUB | BPF_K:
515 case BPF_ALU | BPF_AND | BPF_K:
516 case BPF_ALU | BPF_OR | BPF_K:
517 case BPF_ALU | BPF_XOR | BPF_K:
518 case BPF_ALU64 | BPF_ADD | BPF_K:
519 case BPF_ALU64 | BPF_SUB | BPF_K:
520 case BPF_ALU64 | BPF_AND | BPF_K:
521 case BPF_ALU64 | BPF_OR | BPF_K:
522 case BPF_ALU64 | BPF_XOR | BPF_K:
523 if (BPF_CLASS(insn->code) == BPF_ALU64)
524 EMIT1(add_1mod(0x48, dst_reg));
525 else if (is_ereg(dst_reg))
526 EMIT1(add_1mod(0x40, dst_reg));
527
528 /* b3 holds 'normal' opcode, b2 short form only valid
529 * in case dst is eax/rax.
530 */
531 switch (BPF_OP(insn->code)) {
532 case BPF_ADD:
533 b3 = 0xC0;
534 b2 = 0x05;
535 break;
536 case BPF_SUB:
537 b3 = 0xE8;
538 b2 = 0x2D;
539 break;
540 case BPF_AND:
541 b3 = 0xE0;
542 b2 = 0x25;
543 break;
544 case BPF_OR:
545 b3 = 0xC8;
546 b2 = 0x0D;
547 break;
548 case BPF_XOR:
549 b3 = 0xF0;
550 b2 = 0x35;
551 break;
552 }
553
554 if (is_imm8(imm32))
555 EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
556 else if (is_axreg(dst_reg))
557 EMIT1_off32(b2, imm32);
558 else
559 EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
560 break;
561
562 case BPF_ALU64 | BPF_MOV | BPF_K:
563 case BPF_ALU | BPF_MOV | BPF_K:
564 emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
565 dst_reg, imm32);
566 break;
567
568 case BPF_LD | BPF_IMM | BPF_DW:
569 emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
570 insn++;
571 i++;
572 break;
573
574 /* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
575 case BPF_ALU | BPF_MOD | BPF_X:
576 case BPF_ALU | BPF_DIV | BPF_X:
577 case BPF_ALU | BPF_MOD | BPF_K:
578 case BPF_ALU | BPF_DIV | BPF_K:
579 case BPF_ALU64 | BPF_MOD | BPF_X:
580 case BPF_ALU64 | BPF_DIV | BPF_X:
581 case BPF_ALU64 | BPF_MOD | BPF_K:
582 case BPF_ALU64 | BPF_DIV | BPF_K:
583 EMIT1(0x50); /* push rax */
584 EMIT1(0x52); /* push rdx */
585
586 if (BPF_SRC(insn->code) == BPF_X)
587 /* mov r11, src_reg */
588 EMIT_mov(AUX_REG, src_reg);
589 else
590 /* mov r11, imm32 */
591 EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
592
593 /* mov rax, dst_reg */
594 EMIT_mov(BPF_REG_0, dst_reg);
595
596 /* xor edx, edx
597 * equivalent to 'xor rdx, rdx', but one byte less
598 */
599 EMIT2(0x31, 0xd2);
600
601 if (BPF_CLASS(insn->code) == BPF_ALU64)
602 /* div r11 */
603 EMIT3(0x49, 0xF7, 0xF3);
604 else
605 /* div r11d */
606 EMIT3(0x41, 0xF7, 0xF3);
607
608 if (BPF_OP(insn->code) == BPF_MOD)
609 /* mov r11, rdx */
610 EMIT3(0x49, 0x89, 0xD3);
611 else
612 /* mov r11, rax */
613 EMIT3(0x49, 0x89, 0xC3);
614
615 EMIT1(0x5A); /* pop rdx */
616 EMIT1(0x58); /* pop rax */
617
618 /* mov dst_reg, r11 */
619 EMIT_mov(dst_reg, AUX_REG);
620 break;
621
622 case BPF_ALU | BPF_MUL | BPF_K:
623 case BPF_ALU | BPF_MUL | BPF_X:
624 case BPF_ALU64 | BPF_MUL | BPF_K:
625 case BPF_ALU64 | BPF_MUL | BPF_X:
626 {
627 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
628
629 if (dst_reg != BPF_REG_0)
630 EMIT1(0x50); /* push rax */
631 if (dst_reg != BPF_REG_3)
632 EMIT1(0x52); /* push rdx */
633
634 /* mov r11, dst_reg */
635 EMIT_mov(AUX_REG, dst_reg);
636
637 if (BPF_SRC(insn->code) == BPF_X)
638 emit_mov_reg(&prog, is64, BPF_REG_0, src_reg);
639 else
640 emit_mov_imm32(&prog, is64, BPF_REG_0, imm32);
641
642 if (is64)
643 EMIT1(add_1mod(0x48, AUX_REG));
644 else if (is_ereg(AUX_REG))
645 EMIT1(add_1mod(0x40, AUX_REG));
646 /* mul(q) r11 */
647 EMIT2(0xF7, add_1reg(0xE0, AUX_REG));
648
649 if (dst_reg != BPF_REG_3)
650 EMIT1(0x5A); /* pop rdx */
651 if (dst_reg != BPF_REG_0) {
652 /* mov dst_reg, rax */
653 EMIT_mov(dst_reg, BPF_REG_0);
654 EMIT1(0x58); /* pop rax */
655 }
656 break;
657 }
658 /* shifts */
659 case BPF_ALU | BPF_LSH | BPF_K:
660 case BPF_ALU | BPF_RSH | BPF_K:
661 case BPF_ALU | BPF_ARSH | BPF_K:
662 case BPF_ALU64 | BPF_LSH | BPF_K:
663 case BPF_ALU64 | BPF_RSH | BPF_K:
664 case BPF_ALU64 | BPF_ARSH | BPF_K:
665 if (BPF_CLASS(insn->code) == BPF_ALU64)
666 EMIT1(add_1mod(0x48, dst_reg));
667 else if (is_ereg(dst_reg))
668 EMIT1(add_1mod(0x40, dst_reg));
669
670 switch (BPF_OP(insn->code)) {
671 case BPF_LSH: b3 = 0xE0; break;
672 case BPF_RSH: b3 = 0xE8; break;
673 case BPF_ARSH: b3 = 0xF8; break;
674 }
675
676 if (imm32 == 1)
677 EMIT2(0xD1, add_1reg(b3, dst_reg));
678 else
679 EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
680 break;
681
682 case BPF_ALU | BPF_LSH | BPF_X:
683 case BPF_ALU | BPF_RSH | BPF_X:
684 case BPF_ALU | BPF_ARSH | BPF_X:
685 case BPF_ALU64 | BPF_LSH | BPF_X:
686 case BPF_ALU64 | BPF_RSH | BPF_X:
687 case BPF_ALU64 | BPF_ARSH | BPF_X:
688
689 /* check for bad case when dst_reg == rcx */
690 if (dst_reg == BPF_REG_4) {
691 /* mov r11, dst_reg */
692 EMIT_mov(AUX_REG, dst_reg);
693 dst_reg = AUX_REG;
694 }
695
696 if (src_reg != BPF_REG_4) { /* common case */
697 EMIT1(0x51); /* push rcx */
698
699 /* mov rcx, src_reg */
700 EMIT_mov(BPF_REG_4, src_reg);
701 }
702
703 /* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
704 if (BPF_CLASS(insn->code) == BPF_ALU64)
705 EMIT1(add_1mod(0x48, dst_reg));
706 else if (is_ereg(dst_reg))
707 EMIT1(add_1mod(0x40, dst_reg));
708
709 switch (BPF_OP(insn->code)) {
710 case BPF_LSH: b3 = 0xE0; break;
711 case BPF_RSH: b3 = 0xE8; break;
712 case BPF_ARSH: b3 = 0xF8; break;
713 }
714 EMIT2(0xD3, add_1reg(b3, dst_reg));
715
716 if (src_reg != BPF_REG_4)
717 EMIT1(0x59); /* pop rcx */
718
719 if (insn->dst_reg == BPF_REG_4)
720 /* mov dst_reg, r11 */
721 EMIT_mov(insn->dst_reg, AUX_REG);
722 break;
723
724 case BPF_ALU | BPF_END | BPF_FROM_BE:
725 switch (imm32) {
726 case 16:
727 /* emit 'ror %ax, 8' to swap lower 2 bytes */
728 EMIT1(0x66);
729 if (is_ereg(dst_reg))
730 EMIT1(0x41);
731 EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
732
733 /* emit 'movzwl eax, ax' */
734 if (is_ereg(dst_reg))
735 EMIT3(0x45, 0x0F, 0xB7);
736 else
737 EMIT2(0x0F, 0xB7);
738 EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
739 break;
740 case 32:
741 /* emit 'bswap eax' to swap lower 4 bytes */
742 if (is_ereg(dst_reg))
743 EMIT2(0x41, 0x0F);
744 else
745 EMIT1(0x0F);
746 EMIT1(add_1reg(0xC8, dst_reg));
747 break;
748 case 64:
749 /* emit 'bswap rax' to swap 8 bytes */
750 EMIT3(add_1mod(0x48, dst_reg), 0x0F,
751 add_1reg(0xC8, dst_reg));
752 break;
753 }
754 break;
755
756 case BPF_ALU | BPF_END | BPF_FROM_LE:
757 switch (imm32) {
758 case 16:
759 /* emit 'movzwl eax, ax' to zero extend 16-bit
760 * into 64 bit
761 */
762 if (is_ereg(dst_reg))
763 EMIT3(0x45, 0x0F, 0xB7);
764 else
765 EMIT2(0x0F, 0xB7);
766 EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
767 break;
768 case 32:
769 /* emit 'mov eax, eax' to clear upper 32-bits */
770 if (is_ereg(dst_reg))
771 EMIT1(0x45);
772 EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
773 break;
774 case 64:
775 /* nop */
776 break;
777 }
778 break;
779
780 /* ST: *(u8*)(dst_reg + off) = imm */
781 case BPF_ST | BPF_MEM | BPF_B:
782 if (is_ereg(dst_reg))
783 EMIT2(0x41, 0xC6);
784 else
785 EMIT1(0xC6);
786 goto st;
787 case BPF_ST | BPF_MEM | BPF_H:
788 if (is_ereg(dst_reg))
789 EMIT3(0x66, 0x41, 0xC7);
790 else
791 EMIT2(0x66, 0xC7);
792 goto st;
793 case BPF_ST | BPF_MEM | BPF_W:
794 if (is_ereg(dst_reg))
795 EMIT2(0x41, 0xC7);
796 else
797 EMIT1(0xC7);
798 goto st;
799 case BPF_ST | BPF_MEM | BPF_DW:
800 EMIT2(add_1mod(0x48, dst_reg), 0xC7);
801
802st: if (is_imm8(insn->off))
803 EMIT2(add_1reg(0x40, dst_reg), insn->off);
804 else
805 EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
806
807 EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
808 break;
809
810 /* STX: *(u8*)(dst_reg + off) = src_reg */
811 case BPF_STX | BPF_MEM | BPF_B:
812 /* emit 'mov byte ptr [rax + off], al' */
813 if (is_ereg(dst_reg) || is_ereg(src_reg) ||
814 /* have to add extra byte for x86 SIL, DIL regs */
815 src_reg == BPF_REG_1 || src_reg == BPF_REG_2)
816 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
817 else
818 EMIT1(0x88);
819 goto stx;
820 case BPF_STX | BPF_MEM | BPF_H:
821 if (is_ereg(dst_reg) || is_ereg(src_reg))
822 EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
823 else
824 EMIT2(0x66, 0x89);
825 goto stx;
826 case BPF_STX | BPF_MEM | BPF_W:
827 if (is_ereg(dst_reg) || is_ereg(src_reg))
828 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
829 else
830 EMIT1(0x89);
831 goto stx;
832 case BPF_STX | BPF_MEM | BPF_DW:
833 EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
834stx: if (is_imm8(insn->off))
835 EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off);
836 else
837 EMIT1_off32(add_2reg(0x80, dst_reg, src_reg),
838 insn->off);
839 break;
840
841 /* LDX: dst_reg = *(u8*)(src_reg + off) */
842 case BPF_LDX | BPF_MEM | BPF_B:
843 /* emit 'movzx rax, byte ptr [rax + off]' */
844 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
845 goto ldx;
846 case BPF_LDX | BPF_MEM | BPF_H:
847 /* emit 'movzx rax, word ptr [rax + off]' */
848 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
849 goto ldx;
850 case BPF_LDX | BPF_MEM | BPF_W:
851 /* emit 'mov eax, dword ptr [rax+0x14]' */
852 if (is_ereg(dst_reg) || is_ereg(src_reg))
853 EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
854 else
855 EMIT1(0x8B);
856 goto ldx;
857 case BPF_LDX | BPF_MEM | BPF_DW:
858 /* emit 'mov rax, qword ptr [rax+0x14]' */
859 EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
860ldx: /* if insn->off == 0 we can save one extra byte, but
861 * special case of x86 r13 which always needs an offset
862 * is not worth the hassle
863 */
864 if (is_imm8(insn->off))
865 EMIT2(add_2reg(0x40, src_reg, dst_reg), insn->off);
866 else
867 EMIT1_off32(add_2reg(0x80, src_reg, dst_reg),
868 insn->off);
869 break;
870
871 /* STX XADD: lock *(u32*)(dst_reg + off) += src_reg */
872 case BPF_STX | BPF_XADD | BPF_W:
873 /* emit 'lock add dword ptr [rax + off], eax' */
874 if (is_ereg(dst_reg) || is_ereg(src_reg))
875 EMIT3(0xF0, add_2mod(0x40, dst_reg, src_reg), 0x01);
876 else
877 EMIT2(0xF0, 0x01);
878 goto xadd;
879 case BPF_STX | BPF_XADD | BPF_DW:
880 EMIT3(0xF0, add_2mod(0x48, dst_reg, src_reg), 0x01);
881xadd: if (is_imm8(insn->off))
882 EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off);
883 else
884 EMIT1_off32(add_2reg(0x80, dst_reg, src_reg),
885 insn->off);
886 break;
887
888 /* call */
889 case BPF_JMP | BPF_CALL:
890 func = (u8 *) __bpf_call_base + imm32;
891 jmp_offset = func - (image + addrs[i]);
892 if (seen_ld_abs) {
893 reload_skb_data = bpf_helper_changes_pkt_data(func);
894 if (reload_skb_data) {
895 EMIT1(0x57); /* push %rdi */
896 jmp_offset += 22; /* pop, mov, sub, mov */
897 } else {
898 EMIT2(0x41, 0x52); /* push %r10 */
899 EMIT2(0x41, 0x51); /* push %r9 */
900 /* need to adjust jmp offset, since
901 * pop %r9, pop %r10 take 4 bytes after call insn
902 */
903 jmp_offset += 4;
904 }
905 }
906 if (!imm32 || !is_simm32(jmp_offset)) {
907 pr_err("unsupported bpf func %d addr %p image %p\n",
908 imm32, func, image);
909 return -EINVAL;
910 }
911 EMIT1_off32(0xE8, jmp_offset);
912 if (seen_ld_abs) {
913 if (reload_skb_data) {
914 EMIT1(0x5F); /* pop %rdi */
915 emit_load_skb_data_hlen(&prog);
916 } else {
917 EMIT2(0x41, 0x59); /* pop %r9 */
918 EMIT2(0x41, 0x5A); /* pop %r10 */
919 }
920 }
921 break;
922
923 case BPF_JMP | BPF_TAIL_CALL:
924 emit_bpf_tail_call(&prog);
925 break;
926
927 /* cond jump */
928 case BPF_JMP | BPF_JEQ | BPF_X:
929 case BPF_JMP | BPF_JNE | BPF_X:
930 case BPF_JMP | BPF_JGT | BPF_X:
931 case BPF_JMP | BPF_JLT | BPF_X:
932 case BPF_JMP | BPF_JGE | BPF_X:
933 case BPF_JMP | BPF_JLE | BPF_X:
934 case BPF_JMP | BPF_JSGT | BPF_X:
935 case BPF_JMP | BPF_JSLT | BPF_X:
936 case BPF_JMP | BPF_JSGE | BPF_X:
937 case BPF_JMP | BPF_JSLE | BPF_X:
938 /* cmp dst_reg, src_reg */
939 EMIT3(add_2mod(0x48, dst_reg, src_reg), 0x39,
940 add_2reg(0xC0, dst_reg, src_reg));
941 goto emit_cond_jmp;
942
943 case BPF_JMP | BPF_JSET | BPF_X:
944 /* test dst_reg, src_reg */
945 EMIT3(add_2mod(0x48, dst_reg, src_reg), 0x85,
946 add_2reg(0xC0, dst_reg, src_reg));
947 goto emit_cond_jmp;
948
949 case BPF_JMP | BPF_JSET | BPF_K:
950 /* test dst_reg, imm32 */
951 EMIT1(add_1mod(0x48, dst_reg));
952 EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
953 goto emit_cond_jmp;
954
955 case BPF_JMP | BPF_JEQ | BPF_K:
956 case BPF_JMP | BPF_JNE | BPF_K:
957 case BPF_JMP | BPF_JGT | BPF_K:
958 case BPF_JMP | BPF_JLT | BPF_K:
959 case BPF_JMP | BPF_JGE | BPF_K:
960 case BPF_JMP | BPF_JLE | BPF_K:
961 case BPF_JMP | BPF_JSGT | BPF_K:
962 case BPF_JMP | BPF_JSLT | BPF_K:
963 case BPF_JMP | BPF_JSGE | BPF_K:
964 case BPF_JMP | BPF_JSLE | BPF_K:
965 /* cmp dst_reg, imm8/32 */
966 EMIT1(add_1mod(0x48, dst_reg));
967
968 if (is_imm8(imm32))
969 EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
970 else
971 EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
972
973emit_cond_jmp: /* convert BPF opcode to x86 */
974 switch (BPF_OP(insn->code)) {
975 case BPF_JEQ:
976 jmp_cond = X86_JE;
977 break;
978 case BPF_JSET:
979 case BPF_JNE:
980 jmp_cond = X86_JNE;
981 break;
982 case BPF_JGT:
983 /* GT is unsigned '>', JA in x86 */
984 jmp_cond = X86_JA;
985 break;
986 case BPF_JLT:
987 /* LT is unsigned '<', JB in x86 */
988 jmp_cond = X86_JB;
989 break;
990 case BPF_JGE:
991 /* GE is unsigned '>=', JAE in x86 */
992 jmp_cond = X86_JAE;
993 break;
994 case BPF_JLE:
995 /* LE is unsigned '<=', JBE in x86 */
996 jmp_cond = X86_JBE;
997 break;
998 case BPF_JSGT:
999 /* signed '>', GT in x86 */
1000 jmp_cond = X86_JG;
1001 break;
1002 case BPF_JSLT:
1003 /* signed '<', LT in x86 */
1004 jmp_cond = X86_JL;
1005 break;
1006 case BPF_JSGE:
1007 /* signed '>=', GE in x86 */
1008 jmp_cond = X86_JGE;
1009 break;
1010 case BPF_JSLE:
1011 /* signed '<=', LE in x86 */
1012 jmp_cond = X86_JLE;
1013 break;
1014 default: /* to silence gcc warning */
1015 return -EFAULT;
1016 }
1017 jmp_offset = addrs[i + insn->off] - addrs[i];
1018 if (is_imm8(jmp_offset)) {
1019 EMIT2(jmp_cond, jmp_offset);
1020 } else if (is_simm32(jmp_offset)) {
1021 EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1022 } else {
1023 pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1024 return -EFAULT;
1025 }
1026
1027 break;
1028
1029 case BPF_JMP | BPF_JA:
1030 if (insn->off == -1)
1031 /* -1 jmp instructions will always jump
1032 * backwards two bytes. Explicitly handling
1033 * this case avoids wasting too many passes
1034 * when there are long sequences of replaced
1035 * dead code.
1036 */
1037 jmp_offset = -2;
1038 else
1039 jmp_offset = addrs[i + insn->off] - addrs[i];
1040
1041 if (!jmp_offset)
1042 /* optimize out nop jumps */
1043 break;
1044emit_jmp:
1045 if (is_imm8(jmp_offset)) {
1046 EMIT2(0xEB, jmp_offset);
1047 } else if (is_simm32(jmp_offset)) {
1048 EMIT1_off32(0xE9, jmp_offset);
1049 } else {
1050 pr_err("jmp gen bug %llx\n", jmp_offset);
1051 return -EFAULT;
1052 }
1053 break;
1054
1055 case BPF_LD | BPF_IND | BPF_W:
1056 func = sk_load_word;
1057 goto common_load;
1058 case BPF_LD | BPF_ABS | BPF_W:
1059 func = CHOOSE_LOAD_FUNC(imm32, sk_load_word);
1060common_load:
1061 ctx->seen_ld_abs = seen_ld_abs = true;
1062 jmp_offset = func - (image + addrs[i]);
1063 if (!func || !is_simm32(jmp_offset)) {
1064 pr_err("unsupported bpf func %d addr %p image %p\n",
1065 imm32, func, image);
1066 return -EINVAL;
1067 }
1068 if (BPF_MODE(insn->code) == BPF_ABS) {
1069 /* mov %esi, imm32 */
1070 EMIT1_off32(0xBE, imm32);
1071 } else {
1072 /* mov %rsi, src_reg */
1073 EMIT_mov(BPF_REG_2, src_reg);
1074 if (imm32) {
1075 if (is_imm8(imm32))
1076 /* add %esi, imm8 */
1077 EMIT3(0x83, 0xC6, imm32);
1078 else
1079 /* add %esi, imm32 */
1080 EMIT2_off32(0x81, 0xC6, imm32);
1081 }
1082 }
1083 /* skb pointer is in R6 (%rbx), it will be copied into
1084 * %rdi if skb_copy_bits() call is necessary.
1085 * sk_load_* helpers also use %r10 and %r9d.
1086 * See bpf_jit.S
1087 */
1088 if (seen_ax_reg)
1089 /* r10 = skb->data, mov %r10, off32(%rbx) */
1090 EMIT3_off32(0x4c, 0x8b, 0x93,
1091 offsetof(struct sk_buff, data));
1092 EMIT1_off32(0xE8, jmp_offset); /* call */
1093 break;
1094
1095 case BPF_LD | BPF_IND | BPF_H:
1096 func = sk_load_half;
1097 goto common_load;
1098 case BPF_LD | BPF_ABS | BPF_H:
1099 func = CHOOSE_LOAD_FUNC(imm32, sk_load_half);
1100 goto common_load;
1101 case BPF_LD | BPF_IND | BPF_B:
1102 func = sk_load_byte;
1103 goto common_load;
1104 case BPF_LD | BPF_ABS | BPF_B:
1105 func = CHOOSE_LOAD_FUNC(imm32, sk_load_byte);
1106 goto common_load;
1107
1108 case BPF_JMP | BPF_EXIT:
1109 if (seen_exit) {
1110 jmp_offset = ctx->cleanup_addr - addrs[i];
1111 goto emit_jmp;
1112 }
1113 seen_exit = true;
1114 /* update cleanup_addr */
1115 ctx->cleanup_addr = proglen;
1116 /* mov rbx, qword ptr [rbp+0] */
1117 EMIT4(0x48, 0x8B, 0x5D, 0);
1118 /* mov r13, qword ptr [rbp+8] */
1119 EMIT4(0x4C, 0x8B, 0x6D, 8);
1120 /* mov r14, qword ptr [rbp+16] */
1121 EMIT4(0x4C, 0x8B, 0x75, 16);
1122 /* mov r15, qword ptr [rbp+24] */
1123 EMIT4(0x4C, 0x8B, 0x7D, 24);
1124
1125 /* add rbp, AUX_STACK_SPACE */
1126 EMIT4(0x48, 0x83, 0xC5, AUX_STACK_SPACE);
1127 EMIT1(0xC9); /* leave */
1128 EMIT1(0xC3); /* ret */
1129 break;
1130
1131 default:
1132 /* By design x64 JIT should support all BPF instructions
1133 * This error will be seen if new instruction was added
1134 * to interpreter, but not to JIT
1135 * or if there is junk in bpf_prog
1136 */
1137 pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1138 return -EINVAL;
1139 }
1140
1141 ilen = prog - temp;
1142 if (ilen > BPF_MAX_INSN_SIZE) {
1143 pr_err("bpf_jit: fatal insn size error\n");
1144 return -EFAULT;
1145 }
1146
1147 if (image) {
1148 if (unlikely(proglen + ilen > oldproglen)) {
1149 pr_err("bpf_jit: fatal error\n");
1150 return -EFAULT;
1151 }
1152 memcpy(image + proglen, temp, ilen);
1153 }
1154 proglen += ilen;
1155 addrs[i] = proglen;
1156 prog = temp;
1157 }
1158 return proglen;
1159}
1160
1161struct x64_jit_data {
1162 struct bpf_binary_header *header;
1163 int *addrs;
1164 u8 *image;
1165 int proglen;
1166 struct jit_context ctx;
1167};
1168
1169struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1170{
1171 struct bpf_binary_header *header = NULL;
1172 struct bpf_prog *tmp, *orig_prog = prog;
1173 struct x64_jit_data *jit_data;
1174 int proglen, oldproglen = 0;
1175 struct jit_context ctx = {};
1176 bool tmp_blinded = false;
1177 bool extra_pass = false;
1178 u8 *image = NULL;
1179 int *addrs;
1180 int pass;
1181 int i;
1182
1183 if (!prog->jit_requested)
1184 return orig_prog;
1185
1186 tmp = bpf_jit_blind_constants(prog);
1187 /* If blinding was requested and we failed during blinding,
1188 * we must fall back to the interpreter.
1189 */
1190 if (IS_ERR(tmp))
1191 return orig_prog;
1192 if (tmp != prog) {
1193 tmp_blinded = true;
1194 prog = tmp;
1195 }
1196
1197 jit_data = prog->aux->jit_data;
1198 if (!jit_data) {
1199 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1200 if (!jit_data) {
1201 prog = orig_prog;
1202 goto out;
1203 }
1204 prog->aux->jit_data = jit_data;
1205 }
1206 addrs = jit_data->addrs;
1207 if (addrs) {
1208 ctx = jit_data->ctx;
1209 oldproglen = jit_data->proglen;
1210 image = jit_data->image;
1211 header = jit_data->header;
1212 extra_pass = true;
1213 goto skip_init_addrs;
1214 }
1215 addrs = kmalloc(prog->len * sizeof(*addrs), GFP_KERNEL);
1216 if (!addrs) {
1217 prog = orig_prog;
1218 goto out_addrs;
1219 }
1220
1221 /* Before first pass, make a rough estimation of addrs[]
1222 * each bpf instruction is translated to less than 64 bytes
1223 */
1224 for (proglen = 0, i = 0; i < prog->len; i++) {
1225 proglen += 64;
1226 addrs[i] = proglen;
1227 }
1228 ctx.cleanup_addr = proglen;
1229skip_init_addrs:
1230
1231 /* JITed image shrinks with every pass and the loop iterates
1232 * until the image stops shrinking. Very large bpf programs
1233 * may converge on the last pass. In such case do one more
1234 * pass to emit the final image
1235 */
1236 for (pass = 0; pass < 20 || image; pass++) {
1237 proglen = do_jit(prog, addrs, image, oldproglen, &ctx);
1238 if (proglen <= 0) {
1239out_image:
1240 image = NULL;
1241 if (header)
1242 bpf_jit_binary_free(header);
1243 prog = orig_prog;
1244 goto out_addrs;
1245 }
1246 if (image) {
1247 if (proglen != oldproglen) {
1248 pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
1249 proglen, oldproglen);
1250 goto out_image;
1251 }
1252 break;
1253 }
1254 if (proglen == oldproglen) {
1255 header = bpf_jit_binary_alloc(proglen, &image,
1256 1, jit_fill_hole);
1257 if (!header) {
1258 prog = orig_prog;
1259 goto out_addrs;
1260 }
1261 }
1262 oldproglen = proglen;
1263 cond_resched();
1264 }
1265
1266 if (bpf_jit_enable > 1)
1267 bpf_jit_dump(prog->len, proglen, pass + 1, image);
1268
1269 if (image) {
1270 if (!prog->is_func || extra_pass) {
1271 bpf_jit_binary_lock_ro(header);
1272 } else {
1273 jit_data->addrs = addrs;
1274 jit_data->ctx = ctx;
1275 jit_data->proglen = proglen;
1276 jit_data->image = image;
1277 jit_data->header = header;
1278 }
1279 prog->bpf_func = (void *)image;
1280 prog->jited = 1;
1281 prog->jited_len = proglen;
1282 } else {
1283 prog = orig_prog;
1284 }
1285
1286 if (!image || !prog->is_func || extra_pass) {
1287out_addrs:
1288 kfree(addrs);
1289 kfree(jit_data);
1290 prog->aux->jit_data = NULL;
1291 }
1292out:
1293 if (tmp_blinded)
1294 bpf_jit_prog_release_other(prog, prog == orig_prog ?
1295 tmp : orig_prog);
1296 return prog;
1297}
1/* bpf_jit_comp.c : BPF JIT compiler
2 *
3 * Copyright (C) 2011-2013 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#include <linux/if_vlan.h>
15#include <linux/random.h>
16
17/*
18 * Conventions :
19 * EAX : BPF A accumulator
20 * EBX : BPF X accumulator
21 * RDI : pointer to skb (first argument given to JIT function)
22 * RBP : frame pointer (even if CONFIG_FRAME_POINTER=n)
23 * ECX,EDX,ESI : scratch registers
24 * r9d : skb->len - skb->data_len (headlen)
25 * r8 : skb->data
26 * -8(RBP) : saved RBX value
27 * -16(RBP)..-80(RBP) : BPF_MEMWORDS values
28 */
29int bpf_jit_enable __read_mostly;
30
31/*
32 * assembly code in arch/x86/net/bpf_jit.S
33 */
34extern u8 sk_load_word[], sk_load_half[], sk_load_byte[], sk_load_byte_msh[];
35extern u8 sk_load_word_positive_offset[], sk_load_half_positive_offset[];
36extern u8 sk_load_byte_positive_offset[], sk_load_byte_msh_positive_offset[];
37extern u8 sk_load_word_negative_offset[], sk_load_half_negative_offset[];
38extern u8 sk_load_byte_negative_offset[], sk_load_byte_msh_negative_offset[];
39
40static inline u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
41{
42 if (len == 1)
43 *ptr = bytes;
44 else if (len == 2)
45 *(u16 *)ptr = bytes;
46 else {
47 *(u32 *)ptr = bytes;
48 barrier();
49 }
50 return ptr + len;
51}
52
53#define EMIT(bytes, len) do { prog = emit_code(prog, bytes, len); } while (0)
54
55#define EMIT1(b1) EMIT(b1, 1)
56#define EMIT2(b1, b2) EMIT((b1) + ((b2) << 8), 2)
57#define EMIT3(b1, b2, b3) EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
58#define EMIT4(b1, b2, b3, b4) EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
59#define EMIT1_off32(b1, off) do { EMIT1(b1); EMIT(off, 4);} while (0)
60
61#define CLEAR_A() EMIT2(0x31, 0xc0) /* xor %eax,%eax */
62#define CLEAR_X() EMIT2(0x31, 0xdb) /* xor %ebx,%ebx */
63
64static inline bool is_imm8(int value)
65{
66 return value <= 127 && value >= -128;
67}
68
69static inline bool is_near(int offset)
70{
71 return offset <= 127 && offset >= -128;
72}
73
74#define EMIT_JMP(offset) \
75do { \
76 if (offset) { \
77 if (is_near(offset)) \
78 EMIT2(0xeb, offset); /* jmp .+off8 */ \
79 else \
80 EMIT1_off32(0xe9, offset); /* jmp .+off32 */ \
81 } \
82} while (0)
83
84/* list of x86 cond jumps opcodes (. + s8)
85 * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
86 */
87#define X86_JB 0x72
88#define X86_JAE 0x73
89#define X86_JE 0x74
90#define X86_JNE 0x75
91#define X86_JBE 0x76
92#define X86_JA 0x77
93
94#define EMIT_COND_JMP(op, offset) \
95do { \
96 if (is_near(offset)) \
97 EMIT2(op, offset); /* jxx .+off8 */ \
98 else { \
99 EMIT2(0x0f, op + 0x10); \
100 EMIT(offset, 4); /* jxx .+off32 */ \
101 } \
102} while (0)
103
104#define COND_SEL(CODE, TOP, FOP) \
105 case CODE: \
106 t_op = TOP; \
107 f_op = FOP; \
108 goto cond_branch
109
110
111#define SEEN_DATAREF 1 /* might call external helpers */
112#define SEEN_XREG 2 /* ebx is used */
113#define SEEN_MEM 4 /* use mem[] for temporary storage */
114
115static inline void bpf_flush_icache(void *start, void *end)
116{
117 mm_segment_t old_fs = get_fs();
118
119 set_fs(KERNEL_DS);
120 smp_wmb();
121 flush_icache_range((unsigned long)start, (unsigned long)end);
122 set_fs(old_fs);
123}
124
125#define CHOOSE_LOAD_FUNC(K, func) \
126 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative_offset : func) : func##_positive_offset)
127
128/* Helper to find the offset of pkt_type in sk_buff
129 * We want to make sure its still a 3bit field starting at a byte boundary.
130 */
131#define PKT_TYPE_MAX 7
132static int pkt_type_offset(void)
133{
134 struct sk_buff skb_probe = {
135 .pkt_type = ~0,
136 };
137 char *ct = (char *)&skb_probe;
138 unsigned int off;
139
140 for (off = 0; off < sizeof(struct sk_buff); off++) {
141 if (ct[off] == PKT_TYPE_MAX)
142 return off;
143 }
144 pr_err_once("Please fix pkt_type_offset(), as pkt_type couldn't be found\n");
145 return -1;
146}
147
148struct bpf_binary_header {
149 unsigned int pages;
150 /* Note : for security reasons, bpf code will follow a randomly
151 * sized amount of int3 instructions
152 */
153 u8 image[];
154};
155
156static struct bpf_binary_header *bpf_alloc_binary(unsigned int proglen,
157 u8 **image_ptr)
158{
159 unsigned int sz, hole;
160 struct bpf_binary_header *header;
161
162 /* Most of BPF filters are really small,
163 * but if some of them fill a page, allow at least
164 * 128 extra bytes to insert a random section of int3
165 */
166 sz = round_up(proglen + sizeof(*header) + 128, PAGE_SIZE);
167 header = module_alloc(sz);
168 if (!header)
169 return NULL;
170
171 memset(header, 0xcc, sz); /* fill whole space with int3 instructions */
172
173 header->pages = sz / PAGE_SIZE;
174 hole = min(sz - (proglen + sizeof(*header)), PAGE_SIZE - sizeof(*header));
175
176 /* insert a random number of int3 instructions before BPF code */
177 *image_ptr = &header->image[prandom_u32() % hole];
178 return header;
179}
180
181void bpf_jit_compile(struct sk_filter *fp)
182{
183 u8 temp[64];
184 u8 *prog;
185 unsigned int proglen, oldproglen = 0;
186 int ilen, i;
187 int t_offset, f_offset;
188 u8 t_op, f_op, seen = 0, pass;
189 u8 *image = NULL;
190 struct bpf_binary_header *header = NULL;
191 u8 *func;
192 int pc_ret0 = -1; /* bpf index of first RET #0 instruction (if any) */
193 unsigned int cleanup_addr; /* epilogue code offset */
194 unsigned int *addrs;
195 const struct sock_filter *filter = fp->insns;
196 int flen = fp->len;
197
198 if (!bpf_jit_enable)
199 return;
200
201 addrs = kmalloc(flen * sizeof(*addrs), GFP_KERNEL);
202 if (addrs == NULL)
203 return;
204
205 /* Before first pass, make a rough estimation of addrs[]
206 * each bpf instruction is translated to less than 64 bytes
207 */
208 for (proglen = 0, i = 0; i < flen; i++) {
209 proglen += 64;
210 addrs[i] = proglen;
211 }
212 cleanup_addr = proglen; /* epilogue address */
213
214 for (pass = 0; pass < 10; pass++) {
215 u8 seen_or_pass0 = (pass == 0) ? (SEEN_XREG | SEEN_DATAREF | SEEN_MEM) : seen;
216 /* no prologue/epilogue for trivial filters (RET something) */
217 proglen = 0;
218 prog = temp;
219
220 if (seen_or_pass0) {
221 EMIT4(0x55, 0x48, 0x89, 0xe5); /* push %rbp; mov %rsp,%rbp */
222 EMIT4(0x48, 0x83, 0xec, 96); /* subq $96,%rsp */
223 /* note : must save %rbx in case bpf_error is hit */
224 if (seen_or_pass0 & (SEEN_XREG | SEEN_DATAREF))
225 EMIT4(0x48, 0x89, 0x5d, 0xf8); /* mov %rbx, -8(%rbp) */
226 if (seen_or_pass0 & SEEN_XREG)
227 CLEAR_X(); /* make sure we dont leek kernel memory */
228
229 /*
230 * If this filter needs to access skb data,
231 * loads r9 and r8 with :
232 * r9 = skb->len - skb->data_len
233 * r8 = skb->data
234 */
235 if (seen_or_pass0 & SEEN_DATAREF) {
236 if (offsetof(struct sk_buff, len) <= 127)
237 /* mov off8(%rdi),%r9d */
238 EMIT4(0x44, 0x8b, 0x4f, offsetof(struct sk_buff, len));
239 else {
240 /* mov off32(%rdi),%r9d */
241 EMIT3(0x44, 0x8b, 0x8f);
242 EMIT(offsetof(struct sk_buff, len), 4);
243 }
244 if (is_imm8(offsetof(struct sk_buff, data_len)))
245 /* sub off8(%rdi),%r9d */
246 EMIT4(0x44, 0x2b, 0x4f, offsetof(struct sk_buff, data_len));
247 else {
248 EMIT3(0x44, 0x2b, 0x8f);
249 EMIT(offsetof(struct sk_buff, data_len), 4);
250 }
251
252 if (is_imm8(offsetof(struct sk_buff, data)))
253 /* mov off8(%rdi),%r8 */
254 EMIT4(0x4c, 0x8b, 0x47, offsetof(struct sk_buff, data));
255 else {
256 /* mov off32(%rdi),%r8 */
257 EMIT3(0x4c, 0x8b, 0x87);
258 EMIT(offsetof(struct sk_buff, data), 4);
259 }
260 }
261 }
262
263 switch (filter[0].code) {
264 case BPF_S_RET_K:
265 case BPF_S_LD_W_LEN:
266 case BPF_S_ANC_PROTOCOL:
267 case BPF_S_ANC_IFINDEX:
268 case BPF_S_ANC_MARK:
269 case BPF_S_ANC_RXHASH:
270 case BPF_S_ANC_CPU:
271 case BPF_S_ANC_VLAN_TAG:
272 case BPF_S_ANC_VLAN_TAG_PRESENT:
273 case BPF_S_ANC_QUEUE:
274 case BPF_S_ANC_PKTTYPE:
275 case BPF_S_LD_W_ABS:
276 case BPF_S_LD_H_ABS:
277 case BPF_S_LD_B_ABS:
278 /* first instruction sets A register (or is RET 'constant') */
279 break;
280 default:
281 /* make sure we dont leak kernel information to user */
282 CLEAR_A(); /* A = 0 */
283 }
284
285 for (i = 0; i < flen; i++) {
286 unsigned int K = filter[i].k;
287
288 switch (filter[i].code) {
289 case BPF_S_ALU_ADD_X: /* A += X; */
290 seen |= SEEN_XREG;
291 EMIT2(0x01, 0xd8); /* add %ebx,%eax */
292 break;
293 case BPF_S_ALU_ADD_K: /* A += K; */
294 if (!K)
295 break;
296 if (is_imm8(K))
297 EMIT3(0x83, 0xc0, K); /* add imm8,%eax */
298 else
299 EMIT1_off32(0x05, K); /* add imm32,%eax */
300 break;
301 case BPF_S_ALU_SUB_X: /* A -= X; */
302 seen |= SEEN_XREG;
303 EMIT2(0x29, 0xd8); /* sub %ebx,%eax */
304 break;
305 case BPF_S_ALU_SUB_K: /* A -= K */
306 if (!K)
307 break;
308 if (is_imm8(K))
309 EMIT3(0x83, 0xe8, K); /* sub imm8,%eax */
310 else
311 EMIT1_off32(0x2d, K); /* sub imm32,%eax */
312 break;
313 case BPF_S_ALU_MUL_X: /* A *= X; */
314 seen |= SEEN_XREG;
315 EMIT3(0x0f, 0xaf, 0xc3); /* imul %ebx,%eax */
316 break;
317 case BPF_S_ALU_MUL_K: /* A *= K */
318 if (is_imm8(K))
319 EMIT3(0x6b, 0xc0, K); /* imul imm8,%eax,%eax */
320 else {
321 EMIT2(0x69, 0xc0); /* imul imm32,%eax */
322 EMIT(K, 4);
323 }
324 break;
325 case BPF_S_ALU_DIV_X: /* A /= X; */
326 seen |= SEEN_XREG;
327 EMIT2(0x85, 0xdb); /* test %ebx,%ebx */
328 if (pc_ret0 > 0) {
329 /* addrs[pc_ret0 - 1] is start address of target
330 * (addrs[i] - 4) is the address following this jmp
331 * ("xor %edx,%edx; div %ebx" being 4 bytes long)
332 */
333 EMIT_COND_JMP(X86_JE, addrs[pc_ret0 - 1] -
334 (addrs[i] - 4));
335 } else {
336 EMIT_COND_JMP(X86_JNE, 2 + 5);
337 CLEAR_A();
338 EMIT1_off32(0xe9, cleanup_addr - (addrs[i] - 4)); /* jmp .+off32 */
339 }
340 EMIT4(0x31, 0xd2, 0xf7, 0xf3); /* xor %edx,%edx; div %ebx */
341 break;
342 case BPF_S_ALU_MOD_X: /* A %= X; */
343 seen |= SEEN_XREG;
344 EMIT2(0x85, 0xdb); /* test %ebx,%ebx */
345 if (pc_ret0 > 0) {
346 /* addrs[pc_ret0 - 1] is start address of target
347 * (addrs[i] - 6) is the address following this jmp
348 * ("xor %edx,%edx; div %ebx;mov %edx,%eax" being 6 bytes long)
349 */
350 EMIT_COND_JMP(X86_JE, addrs[pc_ret0 - 1] -
351 (addrs[i] - 6));
352 } else {
353 EMIT_COND_JMP(X86_JNE, 2 + 5);
354 CLEAR_A();
355 EMIT1_off32(0xe9, cleanup_addr - (addrs[i] - 6)); /* jmp .+off32 */
356 }
357 EMIT2(0x31, 0xd2); /* xor %edx,%edx */
358 EMIT2(0xf7, 0xf3); /* div %ebx */
359 EMIT2(0x89, 0xd0); /* mov %edx,%eax */
360 break;
361 case BPF_S_ALU_MOD_K: /* A %= K; */
362 if (K == 1) {
363 CLEAR_A();
364 break;
365 }
366 EMIT2(0x31, 0xd2); /* xor %edx,%edx */
367 EMIT1(0xb9);EMIT(K, 4); /* mov imm32,%ecx */
368 EMIT2(0xf7, 0xf1); /* div %ecx */
369 EMIT2(0x89, 0xd0); /* mov %edx,%eax */
370 break;
371 case BPF_S_ALU_DIV_K: /* A /= K */
372 if (K == 1)
373 break;
374 EMIT2(0x31, 0xd2); /* xor %edx,%edx */
375 EMIT1(0xb9);EMIT(K, 4); /* mov imm32,%ecx */
376 EMIT2(0xf7, 0xf1); /* div %ecx */
377 break;
378 case BPF_S_ALU_AND_X:
379 seen |= SEEN_XREG;
380 EMIT2(0x21, 0xd8); /* and %ebx,%eax */
381 break;
382 case BPF_S_ALU_AND_K:
383 if (K >= 0xFFFFFF00) {
384 EMIT2(0x24, K & 0xFF); /* and imm8,%al */
385 } else if (K >= 0xFFFF0000) {
386 EMIT2(0x66, 0x25); /* and imm16,%ax */
387 EMIT(K, 2);
388 } else {
389 EMIT1_off32(0x25, K); /* and imm32,%eax */
390 }
391 break;
392 case BPF_S_ALU_OR_X:
393 seen |= SEEN_XREG;
394 EMIT2(0x09, 0xd8); /* or %ebx,%eax */
395 break;
396 case BPF_S_ALU_OR_K:
397 if (is_imm8(K))
398 EMIT3(0x83, 0xc8, K); /* or imm8,%eax */
399 else
400 EMIT1_off32(0x0d, K); /* or imm32,%eax */
401 break;
402 case BPF_S_ANC_ALU_XOR_X: /* A ^= X; */
403 case BPF_S_ALU_XOR_X:
404 seen |= SEEN_XREG;
405 EMIT2(0x31, 0xd8); /* xor %ebx,%eax */
406 break;
407 case BPF_S_ALU_XOR_K: /* A ^= K; */
408 if (K == 0)
409 break;
410 if (is_imm8(K))
411 EMIT3(0x83, 0xf0, K); /* xor imm8,%eax */
412 else
413 EMIT1_off32(0x35, K); /* xor imm32,%eax */
414 break;
415 case BPF_S_ALU_LSH_X: /* A <<= X; */
416 seen |= SEEN_XREG;
417 EMIT4(0x89, 0xd9, 0xd3, 0xe0); /* mov %ebx,%ecx; shl %cl,%eax */
418 break;
419 case BPF_S_ALU_LSH_K:
420 if (K == 0)
421 break;
422 else if (K == 1)
423 EMIT2(0xd1, 0xe0); /* shl %eax */
424 else
425 EMIT3(0xc1, 0xe0, K);
426 break;
427 case BPF_S_ALU_RSH_X: /* A >>= X; */
428 seen |= SEEN_XREG;
429 EMIT4(0x89, 0xd9, 0xd3, 0xe8); /* mov %ebx,%ecx; shr %cl,%eax */
430 break;
431 case BPF_S_ALU_RSH_K: /* A >>= K; */
432 if (K == 0)
433 break;
434 else if (K == 1)
435 EMIT2(0xd1, 0xe8); /* shr %eax */
436 else
437 EMIT3(0xc1, 0xe8, K);
438 break;
439 case BPF_S_ALU_NEG:
440 EMIT2(0xf7, 0xd8); /* neg %eax */
441 break;
442 case BPF_S_RET_K:
443 if (!K) {
444 if (pc_ret0 == -1)
445 pc_ret0 = i;
446 CLEAR_A();
447 } else {
448 EMIT1_off32(0xb8, K); /* mov $imm32,%eax */
449 }
450 /* fallinto */
451 case BPF_S_RET_A:
452 if (seen_or_pass0) {
453 if (i != flen - 1) {
454 EMIT_JMP(cleanup_addr - addrs[i]);
455 break;
456 }
457 if (seen_or_pass0 & SEEN_XREG)
458 EMIT4(0x48, 0x8b, 0x5d, 0xf8); /* mov -8(%rbp),%rbx */
459 EMIT1(0xc9); /* leaveq */
460 }
461 EMIT1(0xc3); /* ret */
462 break;
463 case BPF_S_MISC_TAX: /* X = A */
464 seen |= SEEN_XREG;
465 EMIT2(0x89, 0xc3); /* mov %eax,%ebx */
466 break;
467 case BPF_S_MISC_TXA: /* A = X */
468 seen |= SEEN_XREG;
469 EMIT2(0x89, 0xd8); /* mov %ebx,%eax */
470 break;
471 case BPF_S_LD_IMM: /* A = K */
472 if (!K)
473 CLEAR_A();
474 else
475 EMIT1_off32(0xb8, K); /* mov $imm32,%eax */
476 break;
477 case BPF_S_LDX_IMM: /* X = K */
478 seen |= SEEN_XREG;
479 if (!K)
480 CLEAR_X();
481 else
482 EMIT1_off32(0xbb, K); /* mov $imm32,%ebx */
483 break;
484 case BPF_S_LD_MEM: /* A = mem[K] : mov off8(%rbp),%eax */
485 seen |= SEEN_MEM;
486 EMIT3(0x8b, 0x45, 0xf0 - K*4);
487 break;
488 case BPF_S_LDX_MEM: /* X = mem[K] : mov off8(%rbp),%ebx */
489 seen |= SEEN_XREG | SEEN_MEM;
490 EMIT3(0x8b, 0x5d, 0xf0 - K*4);
491 break;
492 case BPF_S_ST: /* mem[K] = A : mov %eax,off8(%rbp) */
493 seen |= SEEN_MEM;
494 EMIT3(0x89, 0x45, 0xf0 - K*4);
495 break;
496 case BPF_S_STX: /* mem[K] = X : mov %ebx,off8(%rbp) */
497 seen |= SEEN_XREG | SEEN_MEM;
498 EMIT3(0x89, 0x5d, 0xf0 - K*4);
499 break;
500 case BPF_S_LD_W_LEN: /* A = skb->len; */
501 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
502 if (is_imm8(offsetof(struct sk_buff, len)))
503 /* mov off8(%rdi),%eax */
504 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, len));
505 else {
506 EMIT2(0x8b, 0x87);
507 EMIT(offsetof(struct sk_buff, len), 4);
508 }
509 break;
510 case BPF_S_LDX_W_LEN: /* X = skb->len; */
511 seen |= SEEN_XREG;
512 if (is_imm8(offsetof(struct sk_buff, len)))
513 /* mov off8(%rdi),%ebx */
514 EMIT3(0x8b, 0x5f, offsetof(struct sk_buff, len));
515 else {
516 EMIT2(0x8b, 0x9f);
517 EMIT(offsetof(struct sk_buff, len), 4);
518 }
519 break;
520 case BPF_S_ANC_PROTOCOL: /* A = ntohs(skb->protocol); */
521 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, protocol) != 2);
522 if (is_imm8(offsetof(struct sk_buff, protocol))) {
523 /* movzwl off8(%rdi),%eax */
524 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, protocol));
525 } else {
526 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */
527 EMIT(offsetof(struct sk_buff, protocol), 4);
528 }
529 EMIT2(0x86, 0xc4); /* ntohs() : xchg %al,%ah */
530 break;
531 case BPF_S_ANC_IFINDEX:
532 if (is_imm8(offsetof(struct sk_buff, dev))) {
533 /* movq off8(%rdi),%rax */
534 EMIT4(0x48, 0x8b, 0x47, offsetof(struct sk_buff, dev));
535 } else {
536 EMIT3(0x48, 0x8b, 0x87); /* movq off32(%rdi),%rax */
537 EMIT(offsetof(struct sk_buff, dev), 4);
538 }
539 EMIT3(0x48, 0x85, 0xc0); /* test %rax,%rax */
540 EMIT_COND_JMP(X86_JE, cleanup_addr - (addrs[i] - 6));
541 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, ifindex) != 4);
542 EMIT2(0x8b, 0x80); /* mov off32(%rax),%eax */
543 EMIT(offsetof(struct net_device, ifindex), 4);
544 break;
545 case BPF_S_ANC_MARK:
546 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
547 if (is_imm8(offsetof(struct sk_buff, mark))) {
548 /* mov off8(%rdi),%eax */
549 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, mark));
550 } else {
551 EMIT2(0x8b, 0x87);
552 EMIT(offsetof(struct sk_buff, mark), 4);
553 }
554 break;
555 case BPF_S_ANC_RXHASH:
556 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
557 if (is_imm8(offsetof(struct sk_buff, hash))) {
558 /* mov off8(%rdi),%eax */
559 EMIT3(0x8b, 0x47, offsetof(struct sk_buff, hash));
560 } else {
561 EMIT2(0x8b, 0x87);
562 EMIT(offsetof(struct sk_buff, hash), 4);
563 }
564 break;
565 case BPF_S_ANC_QUEUE:
566 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, queue_mapping) != 2);
567 if (is_imm8(offsetof(struct sk_buff, queue_mapping))) {
568 /* movzwl off8(%rdi),%eax */
569 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, queue_mapping));
570 } else {
571 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */
572 EMIT(offsetof(struct sk_buff, queue_mapping), 4);
573 }
574 break;
575 case BPF_S_ANC_CPU:
576#ifdef CONFIG_SMP
577 EMIT4(0x65, 0x8b, 0x04, 0x25); /* mov %gs:off32,%eax */
578 EMIT((u32)(unsigned long)&cpu_number, 4); /* A = smp_processor_id(); */
579#else
580 CLEAR_A();
581#endif
582 break;
583 case BPF_S_ANC_VLAN_TAG:
584 case BPF_S_ANC_VLAN_TAG_PRESENT:
585 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
586 if (is_imm8(offsetof(struct sk_buff, vlan_tci))) {
587 /* movzwl off8(%rdi),%eax */
588 EMIT4(0x0f, 0xb7, 0x47, offsetof(struct sk_buff, vlan_tci));
589 } else {
590 EMIT3(0x0f, 0xb7, 0x87); /* movzwl off32(%rdi),%eax */
591 EMIT(offsetof(struct sk_buff, vlan_tci), 4);
592 }
593 BUILD_BUG_ON(VLAN_TAG_PRESENT != 0x1000);
594 if (filter[i].code == BPF_S_ANC_VLAN_TAG) {
595 EMIT3(0x80, 0xe4, 0xef); /* and $0xef,%ah */
596 } else {
597 EMIT3(0xc1, 0xe8, 0x0c); /* shr $0xc,%eax */
598 EMIT3(0x83, 0xe0, 0x01); /* and $0x1,%eax */
599 }
600 break;
601 case BPF_S_ANC_PKTTYPE:
602 {
603 int off = pkt_type_offset();
604
605 if (off < 0)
606 goto out;
607 if (is_imm8(off)) {
608 /* movzbl off8(%rdi),%eax */
609 EMIT4(0x0f, 0xb6, 0x47, off);
610 } else {
611 /* movbl off32(%rdi),%eax */
612 EMIT3(0x0f, 0xb6, 0x87);
613 EMIT(off, 4);
614 }
615 EMIT3(0x83, 0xe0, PKT_TYPE_MAX); /* and $0x7,%eax */
616 break;
617 }
618 case BPF_S_LD_W_ABS:
619 func = CHOOSE_LOAD_FUNC(K, sk_load_word);
620common_load: seen |= SEEN_DATAREF;
621 t_offset = func - (image + addrs[i]);
622 EMIT1_off32(0xbe, K); /* mov imm32,%esi */
623 EMIT1_off32(0xe8, t_offset); /* call */
624 break;
625 case BPF_S_LD_H_ABS:
626 func = CHOOSE_LOAD_FUNC(K, sk_load_half);
627 goto common_load;
628 case BPF_S_LD_B_ABS:
629 func = CHOOSE_LOAD_FUNC(K, sk_load_byte);
630 goto common_load;
631 case BPF_S_LDX_B_MSH:
632 func = CHOOSE_LOAD_FUNC(K, sk_load_byte_msh);
633 seen |= SEEN_DATAREF | SEEN_XREG;
634 t_offset = func - (image + addrs[i]);
635 EMIT1_off32(0xbe, K); /* mov imm32,%esi */
636 EMIT1_off32(0xe8, t_offset); /* call sk_load_byte_msh */
637 break;
638 case BPF_S_LD_W_IND:
639 func = sk_load_word;
640common_load_ind: seen |= SEEN_DATAREF | SEEN_XREG;
641 t_offset = func - (image + addrs[i]);
642 if (K) {
643 if (is_imm8(K)) {
644 EMIT3(0x8d, 0x73, K); /* lea imm8(%rbx), %esi */
645 } else {
646 EMIT2(0x8d, 0xb3); /* lea imm32(%rbx),%esi */
647 EMIT(K, 4);
648 }
649 } else {
650 EMIT2(0x89,0xde); /* mov %ebx,%esi */
651 }
652 EMIT1_off32(0xe8, t_offset); /* call sk_load_xxx_ind */
653 break;
654 case BPF_S_LD_H_IND:
655 func = sk_load_half;
656 goto common_load_ind;
657 case BPF_S_LD_B_IND:
658 func = sk_load_byte;
659 goto common_load_ind;
660 case BPF_S_JMP_JA:
661 t_offset = addrs[i + K] - addrs[i];
662 EMIT_JMP(t_offset);
663 break;
664 COND_SEL(BPF_S_JMP_JGT_K, X86_JA, X86_JBE);
665 COND_SEL(BPF_S_JMP_JGE_K, X86_JAE, X86_JB);
666 COND_SEL(BPF_S_JMP_JEQ_K, X86_JE, X86_JNE);
667 COND_SEL(BPF_S_JMP_JSET_K,X86_JNE, X86_JE);
668 COND_SEL(BPF_S_JMP_JGT_X, X86_JA, X86_JBE);
669 COND_SEL(BPF_S_JMP_JGE_X, X86_JAE, X86_JB);
670 COND_SEL(BPF_S_JMP_JEQ_X, X86_JE, X86_JNE);
671 COND_SEL(BPF_S_JMP_JSET_X,X86_JNE, X86_JE);
672
673cond_branch: f_offset = addrs[i + filter[i].jf] - addrs[i];
674 t_offset = addrs[i + filter[i].jt] - addrs[i];
675
676 /* same targets, can avoid doing the test :) */
677 if (filter[i].jt == filter[i].jf) {
678 EMIT_JMP(t_offset);
679 break;
680 }
681
682 switch (filter[i].code) {
683 case BPF_S_JMP_JGT_X:
684 case BPF_S_JMP_JGE_X:
685 case BPF_S_JMP_JEQ_X:
686 seen |= SEEN_XREG;
687 EMIT2(0x39, 0xd8); /* cmp %ebx,%eax */
688 break;
689 case BPF_S_JMP_JSET_X:
690 seen |= SEEN_XREG;
691 EMIT2(0x85, 0xd8); /* test %ebx,%eax */
692 break;
693 case BPF_S_JMP_JEQ_K:
694 if (K == 0) {
695 EMIT2(0x85, 0xc0); /* test %eax,%eax */
696 break;
697 }
698 case BPF_S_JMP_JGT_K:
699 case BPF_S_JMP_JGE_K:
700 if (K <= 127)
701 EMIT3(0x83, 0xf8, K); /* cmp imm8,%eax */
702 else
703 EMIT1_off32(0x3d, K); /* cmp imm32,%eax */
704 break;
705 case BPF_S_JMP_JSET_K:
706 if (K <= 0xFF)
707 EMIT2(0xa8, K); /* test imm8,%al */
708 else if (!(K & 0xFFFF00FF))
709 EMIT3(0xf6, 0xc4, K >> 8); /* test imm8,%ah */
710 else if (K <= 0xFFFF) {
711 EMIT2(0x66, 0xa9); /* test imm16,%ax */
712 EMIT(K, 2);
713 } else {
714 EMIT1_off32(0xa9, K); /* test imm32,%eax */
715 }
716 break;
717 }
718 if (filter[i].jt != 0) {
719 if (filter[i].jf && f_offset)
720 t_offset += is_near(f_offset) ? 2 : 5;
721 EMIT_COND_JMP(t_op, t_offset);
722 if (filter[i].jf)
723 EMIT_JMP(f_offset);
724 break;
725 }
726 EMIT_COND_JMP(f_op, f_offset);
727 break;
728 default:
729 /* hmm, too complex filter, give up with jit compiler */
730 goto out;
731 }
732 ilen = prog - temp;
733 if (image) {
734 if (unlikely(proglen + ilen > oldproglen)) {
735 pr_err("bpb_jit_compile fatal error\n");
736 kfree(addrs);
737 module_free(NULL, header);
738 return;
739 }
740 memcpy(image + proglen, temp, ilen);
741 }
742 proglen += ilen;
743 addrs[i] = proglen;
744 prog = temp;
745 }
746 /* last bpf instruction is always a RET :
747 * use it to give the cleanup instruction(s) addr
748 */
749 cleanup_addr = proglen - 1; /* ret */
750 if (seen_or_pass0)
751 cleanup_addr -= 1; /* leaveq */
752 if (seen_or_pass0 & SEEN_XREG)
753 cleanup_addr -= 4; /* mov -8(%rbp),%rbx */
754
755 if (image) {
756 if (proglen != oldproglen)
757 pr_err("bpb_jit_compile proglen=%u != oldproglen=%u\n", proglen, oldproglen);
758 break;
759 }
760 if (proglen == oldproglen) {
761 header = bpf_alloc_binary(proglen, &image);
762 if (!header)
763 goto out;
764 }
765 oldproglen = proglen;
766 }
767
768 if (bpf_jit_enable > 1)
769 bpf_jit_dump(flen, proglen, pass, image);
770
771 if (image) {
772 bpf_flush_icache(header, image + proglen);
773 set_memory_ro((unsigned long)header, header->pages);
774 fp->bpf_func = (void *)image;
775 fp->jited = 1;
776 }
777out:
778 kfree(addrs);
779 return;
780}
781
782static void bpf_jit_free_deferred(struct work_struct *work)
783{
784 struct sk_filter *fp = container_of(work, struct sk_filter, work);
785 unsigned long addr = (unsigned long)fp->bpf_func & PAGE_MASK;
786 struct bpf_binary_header *header = (void *)addr;
787
788 set_memory_rw(addr, header->pages);
789 module_free(NULL, header);
790 kfree(fp);
791}
792
793void bpf_jit_free(struct sk_filter *fp)
794{
795 if (fp->jited) {
796 INIT_WORK(&fp->work, bpf_jit_free_deferred);
797 schedule_work(&fp->work);
798 } else {
799 kfree(fp);
800 }
801}