]> git.ipfire.org Git - thirdparty/linux.git/blob - arch/x86/net/bpf_jit_comp.c
projects / thirdparty / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
riscv: select ARCH_HAS_STRICT_KERNEL_RWX only if MMU
[thirdparty/linux.git] / arch / x86 / net / bpf_jit_comp.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * bpf_jit_comp.c: BPF JIT compiler
4 *
5 * Copyright (C) 2011-2013 Eric Dumazet (eric.dumazet@gmail.com)
6 * Internal BPF Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
7 */
8 #include <linux/netdevice.h>
9 #include <linux/filter.h>
10 #include <linux/if_vlan.h>
11 #include <linux/bpf.h>
12 #include <linux/memory.h>
13 #include <linux/sort.h>
14 #include <asm/extable.h>
15 #include <asm/set_memory.h>
16 #include <asm/nospec-branch.h>
17 #include <asm/text-patching.h>
18 #include <asm/asm-prototypes.h>
19
20 static u8 *emit_code(u8 *ptr, u32 bytes, unsigned int len)
21 {
22 if (len == 1)
23 *ptr = bytes;
24 else if (len == 2)
25 *(u16 *)ptr = bytes;
26 else {
27 *(u32 *)ptr = bytes;
28 barrier();
29 }
30 return ptr + len;
31 }
32
33 #define EMIT(bytes, len) \
34 do { prog = emit_code(prog, bytes, len); cnt += len; } while (0)
35
36 #define EMIT1(b1) EMIT(b1, 1)
37 #define EMIT2(b1, b2) EMIT((b1) + ((b2) << 8), 2)
38 #define EMIT3(b1, b2, b3) EMIT((b1) + ((b2) << 8) + ((b3) << 16), 3)
39 #define EMIT4(b1, b2, b3, b4) EMIT((b1) + ((b2) << 8) + ((b3) << 16) + ((b4) << 24), 4)
40
41 #define EMIT1_off32(b1, off) \
42 do { EMIT1(b1); EMIT(off, 4); } while (0)
43 #define EMIT2_off32(b1, b2, off) \
44 do { EMIT2(b1, b2); EMIT(off, 4); } while (0)
45 #define EMIT3_off32(b1, b2, b3, off) \
46 do { EMIT3(b1, b2, b3); EMIT(off, 4); } while (0)
47 #define EMIT4_off32(b1, b2, b3, b4, off) \
48 do { EMIT4(b1, b2, b3, b4); EMIT(off, 4); } while (0)
49
50 static bool is_imm8(int value)
51 {
52 return value <= 127 && value >= -128;
53 }
54
55 static bool is_simm32(s64 value)
56 {
57 return value == (s64)(s32)value;
58 }
59
60 static bool is_uimm32(u64 value)
61 {
62 return value == (u64)(u32)value;
63 }
64
65 /* mov dst, src */
66 #define EMIT_mov(DST, SRC) \
67 do { \
68 if (DST != SRC) \
69 EMIT3(add_2mod(0x48, DST, SRC), 0x89, add_2reg(0xC0, DST, SRC)); \
70 } while (0)
71
72 static int bpf_size_to_x86_bytes(int bpf_size)
73 {
74 if (bpf_size == BPF_W)
75 return 4;
76 else if (bpf_size == BPF_H)
77 return 2;
78 else if (bpf_size == BPF_B)
79 return 1;
80 else if (bpf_size == BPF_DW)
81 return 4; /* imm32 */
82 else
83 return 0;
84 }
85
86 /*
87 * List of x86 cond jumps opcodes (. + s8)
88 * Add 0x10 (and an extra 0x0f) to generate far jumps (. + s32)
89 */
90 #define X86_JB 0x72
91 #define X86_JAE 0x73
92 #define X86_JE 0x74
93 #define X86_JNE 0x75
94 #define X86_JBE 0x76
95 #define X86_JA 0x77
96 #define X86_JL 0x7C
97 #define X86_JGE 0x7D
98 #define X86_JLE 0x7E
99 #define X86_JG 0x7F
100
101 /* Pick a register outside of BPF range for JIT internal work */
102 #define AUX_REG (MAX_BPF_JIT_REG + 1)
103 #define X86_REG_R9 (MAX_BPF_JIT_REG + 2)
104
105 /*
106 * The following table maps BPF registers to x86-64 registers.
107 *
108 * x86-64 register R12 is unused, since if used as base address
109 * register in load/store instructions, it always needs an
110 * extra byte of encoding and is callee saved.
111 *
112 * x86-64 register R9 is not used by BPF programs, but can be used by BPF
113 * trampoline. x86-64 register R10 is used for blinding (if enabled).
114 */
115 static const int reg2hex[] = {
116 [BPF_REG_0] = 0, /* RAX */
117 [BPF_REG_1] = 7, /* RDI */
118 [BPF_REG_2] = 6, /* RSI */
119 [BPF_REG_3] = 2, /* RDX */
120 [BPF_REG_4] = 1, /* RCX */
121 [BPF_REG_5] = 0, /* R8 */
122 [BPF_REG_6] = 3, /* RBX callee saved */
123 [BPF_REG_7] = 5, /* R13 callee saved */
124 [BPF_REG_8] = 6, /* R14 callee saved */
125 [BPF_REG_9] = 7, /* R15 callee saved */
126 [BPF_REG_FP] = 5, /* RBP readonly */
127 [BPF_REG_AX] = 2, /* R10 temp register */
128 [AUX_REG] = 3, /* R11 temp register */
129 [X86_REG_R9] = 1, /* R9 register, 6th function argument */
130 };
131
132 static const int reg2pt_regs[] = {
133 [BPF_REG_0] = offsetof(struct pt_regs, ax),
134 [BPF_REG_1] = offsetof(struct pt_regs, di),
135 [BPF_REG_2] = offsetof(struct pt_regs, si),
136 [BPF_REG_3] = offsetof(struct pt_regs, dx),
137 [BPF_REG_4] = offsetof(struct pt_regs, cx),
138 [BPF_REG_5] = offsetof(struct pt_regs, r8),
139 [BPF_REG_6] = offsetof(struct pt_regs, bx),
140 [BPF_REG_7] = offsetof(struct pt_regs, r13),
141 [BPF_REG_8] = offsetof(struct pt_regs, r14),
142 [BPF_REG_9] = offsetof(struct pt_regs, r15),
143 };
144
145 /*
146 * is_ereg() == true if BPF register 'reg' maps to x86-64 r8..r15
147 * which need extra byte of encoding.
148 * rax,rcx,...,rbp have simpler encoding
149 */
150 static bool is_ereg(u32 reg)
151 {
152 return (1 << reg) & (BIT(BPF_REG_5) |
153 BIT(AUX_REG) |
154 BIT(BPF_REG_7) |
155 BIT(BPF_REG_8) |
156 BIT(BPF_REG_9) |
157 BIT(X86_REG_R9) |
158 BIT(BPF_REG_AX));
159 }
160
161 static bool is_axreg(u32 reg)
162 {
163 return reg == BPF_REG_0;
164 }
165
166 /* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
167 static u8 add_1mod(u8 byte, u32 reg)
168 {
169 if (is_ereg(reg))
170 byte |= 1;
171 return byte;
172 }
173
174 static u8 add_2mod(u8 byte, u32 r1, u32 r2)
175 {
176 if (is_ereg(r1))
177 byte |= 1;
178 if (is_ereg(r2))
179 byte |= 4;
180 return byte;
181 }
182
183 /* Encode 'dst_reg' register into x86-64 opcode 'byte' */
184 static u8 add_1reg(u8 byte, u32 dst_reg)
185 {
186 return byte + reg2hex[dst_reg];
187 }
188
189 /* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
190 static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
191 {
192 return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
193 }
194
195 static void jit_fill_hole(void *area, unsigned int size)
196 {
197 /* Fill whole space with INT3 instructions */
198 memset(area, 0xcc, size);
199 }
200
201 struct jit_context {
202 int cleanup_addr; /* Epilogue code offset */
203 };
204
205 /* Maximum number of bytes emitted while JITing one eBPF insn */
206 #define BPF_MAX_INSN_SIZE 128
207 #define BPF_INSN_SAFETY 64
208
209 /* Number of bytes emit_patch() needs to generate instructions */
210 #define X86_PATCH_SIZE 5
211
212 #define PROLOGUE_SIZE 25
213
214 /*
215 * Emit x86-64 prologue code for BPF program and check its size.
216 * bpf_tail_call helper will skip it while jumping into another program
217 */
218 static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf)
219 {
220 u8 *prog = *pprog;
221 int cnt = X86_PATCH_SIZE;
222
223 /* BPF trampoline can be made to work without these nops,
224 * but let's waste 5 bytes for now and optimize later
225 */
226 memcpy(prog, ideal_nops[NOP_ATOMIC5], cnt);
227 prog += cnt;
228 EMIT1(0x55); /* push rbp */
229 EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
230 /* sub rsp, rounded_stack_depth */
231 EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
232 EMIT1(0x53); /* push rbx */
233 EMIT2(0x41, 0x55); /* push r13 */
234 EMIT2(0x41, 0x56); /* push r14 */
235 EMIT2(0x41, 0x57); /* push r15 */
236 if (!ebpf_from_cbpf) {
237 /* zero init tail_call_cnt */
238 EMIT2(0x6a, 0x00);
239 BUILD_BUG_ON(cnt != PROLOGUE_SIZE);
240 }
241 *pprog = prog;
242 }
243
244 static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
245 {
246 u8 *prog = *pprog;
247 int cnt = 0;
248 s64 offset;
249
250 offset = func - (ip + X86_PATCH_SIZE);
251 if (!is_simm32(offset)) {
252 pr_err("Target call %p is out of range\n", func);
253 return -ERANGE;
254 }
255 EMIT1_off32(opcode, offset);
256 *pprog = prog;
257 return 0;
258 }
259
260 static int emit_call(u8 **pprog, void *func, void *ip)
261 {
262 return emit_patch(pprog, func, ip, 0xE8);
263 }
264
265 static int emit_jump(u8 **pprog, void *func, void *ip)
266 {
267 return emit_patch(pprog, func, ip, 0xE9);
268 }
269
270 static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
271 void *old_addr, void *new_addr,
272 const bool text_live)
273 {
274 const u8 *nop_insn = ideal_nops[NOP_ATOMIC5];
275 u8 old_insn[X86_PATCH_SIZE];
276 u8 new_insn[X86_PATCH_SIZE];
277 u8 *prog;
278 int ret;
279
280 memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
281 if (old_addr) {
282 prog = old_insn;
283 ret = t == BPF_MOD_CALL ?
284 emit_call(&prog, old_addr, ip) :
285 emit_jump(&prog, old_addr, ip);
286 if (ret)
287 return ret;
288 }
289
290 memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
291 if (new_addr) {
292 prog = new_insn;
293 ret = t == BPF_MOD_CALL ?
294 emit_call(&prog, new_addr, ip) :
295 emit_jump(&prog, new_addr, ip);
296 if (ret)
297 return ret;
298 }
299
300 ret = -EBUSY;
301 mutex_lock(&text_mutex);
302 if (memcmp(ip, old_insn, X86_PATCH_SIZE))
303 goto out;
304 if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
305 if (text_live)
306 text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
307 else
308 memcpy(ip, new_insn, X86_PATCH_SIZE);
309 }
310 ret = 0;
311 out:
312 mutex_unlock(&text_mutex);
313 return ret;
314 }
315
316 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
317 void *old_addr, void *new_addr)
318 {
319 if (!is_kernel_text((long)ip) &&
320 !is_bpf_text_address((long)ip))
321 /* BPF poking in modules is not supported */
322 return -EINVAL;
323
324 return __bpf_arch_text_poke(ip, t, old_addr, new_addr, true);
325 }
326
327 /*
328 * Generate the following code:
329 *
330 * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
331 * if (index >= array->map.max_entries)
332 * goto out;
333 * if (++tail_call_cnt > MAX_TAIL_CALL_CNT)
334 * goto out;
335 * prog = array->ptrs[index];
336 * if (prog == NULL)
337 * goto out;
338 * goto *(prog->bpf_func + prologue_size);
339 * out:
340 */
341 static void emit_bpf_tail_call_indirect(u8 **pprog)
342 {
343 u8 *prog = *pprog;
344 int label1, label2, label3;
345 int cnt = 0;
346
347 /*
348 * rdi - pointer to ctx
349 * rsi - pointer to bpf_array
350 * rdx - index in bpf_array
351 */
352
353 /*
354 * if (index >= array->map.max_entries)
355 * goto out;
356 */
357 EMIT2(0x89, 0xD2); /* mov edx, edx */
358 EMIT3(0x39, 0x56, /* cmp dword ptr [rsi + 16], edx */
359 offsetof(struct bpf_array, map.max_entries));
360 #define OFFSET1 (41 + RETPOLINE_RAX_BPF_JIT_SIZE) /* Number of bytes to jump */
361 EMIT2(X86_JBE, OFFSET1); /* jbe out */
362 label1 = cnt;
363
364 /*
365 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
366 * goto out;
367 */
368 EMIT2_off32(0x8B, 0x85, -36 - MAX_BPF_STACK); /* mov eax, dword ptr [rbp - 548] */
369 EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */
370 #define OFFSET2 (30 + RETPOLINE_RAX_BPF_JIT_SIZE)
371 EMIT2(X86_JA, OFFSET2); /* ja out */
372 label2 = cnt;
373 EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */
374 EMIT2_off32(0x89, 0x85, -36 - MAX_BPF_STACK); /* mov dword ptr [rbp -548], eax */
375
376 /* prog = array->ptrs[index]; */
377 EMIT4_off32(0x48, 0x8B, 0x84, 0xD6, /* mov rax, [rsi + rdx * 8 + offsetof(...)] */
378 offsetof(struct bpf_array, ptrs));
379
380 /*
381 * if (prog == NULL)
382 * goto out;
383 */
384 EMIT3(0x48, 0x85, 0xC0); /* test rax,rax */
385 #define OFFSET3 (8 + RETPOLINE_RAX_BPF_JIT_SIZE)
386 EMIT2(X86_JE, OFFSET3); /* je out */
387 label3 = cnt;
388
389 /* goto *(prog->bpf_func + prologue_size); */
390 EMIT4(0x48, 0x8B, 0x40, /* mov rax, qword ptr [rax + 32] */
391 offsetof(struct bpf_prog, bpf_func));
392 EMIT4(0x48, 0x83, 0xC0, PROLOGUE_SIZE); /* add rax, prologue_size */
393
394 /*
395 * Wow we're ready to jump into next BPF program
396 * rdi == ctx (1st arg)
397 * rax == prog->bpf_func + prologue_size
398 */
399 RETPOLINE_RAX_BPF_JIT();
400
401 /* out: */
402 BUILD_BUG_ON(cnt - label1 != OFFSET1);
403 BUILD_BUG_ON(cnt - label2 != OFFSET2);
404 BUILD_BUG_ON(cnt - label3 != OFFSET3);
405 *pprog = prog;
406 }
407
408 static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
409 u8 **pprog, int addr, u8 *image)
410 {
411 u8 *prog = *pprog;
412 int cnt = 0;
413
414 /*
415 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
416 * goto out;
417 */
418 EMIT2_off32(0x8B, 0x85, -36 - MAX_BPF_STACK); /* mov eax, dword ptr [rbp - 548] */
419 EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */
420 EMIT2(X86_JA, 14); /* ja out */
421 EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */
422 EMIT2_off32(0x89, 0x85, -36 - MAX_BPF_STACK); /* mov dword ptr [rbp -548], eax */
423
424 poke->ip = image + (addr - X86_PATCH_SIZE);
425 poke->adj_off = PROLOGUE_SIZE;
426
427 memcpy(prog, ideal_nops[NOP_ATOMIC5], X86_PATCH_SIZE);
428 prog += X86_PATCH_SIZE;
429 /* out: */
430
431 *pprog = prog;
432 }
433
434 static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
435 {
436 struct bpf_jit_poke_descriptor *poke;
437 struct bpf_array *array;
438 struct bpf_prog *target;
439 int i, ret;
440
441 for (i = 0; i < prog->aux->size_poke_tab; i++) {
442 poke = &prog->aux->poke_tab[i];
443 WARN_ON_ONCE(READ_ONCE(poke->ip_stable));
444
445 if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
446 continue;
447
448 array = container_of(poke->tail_call.map, struct bpf_array, map);
449 mutex_lock(&array->aux->poke_mutex);
450 target = array->ptrs[poke->tail_call.key];
451 if (target) {
452 /* Plain memcpy is used when image is not live yet
453 * and still not locked as read-only. Once poke
454 * location is active (poke->ip_stable), any parallel
455 * bpf_arch_text_poke() might occur still on the
456 * read-write image until we finally locked it as
457 * read-only. Both modifications on the given image
458 * are under text_mutex to avoid interference.
459 */
460 ret = __bpf_arch_text_poke(poke->ip, BPF_MOD_JUMP, NULL,
461 (u8 *)target->bpf_func +
462 poke->adj_off, false);
463 BUG_ON(ret < 0);
464 }
465 WRITE_ONCE(poke->ip_stable, true);
466 mutex_unlock(&array->aux->poke_mutex);
467 }
468 }
469
470 static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
471 u32 dst_reg, const u32 imm32)
472 {
473 u8 *prog = *pprog;
474 u8 b1, b2, b3;
475 int cnt = 0;
476
477 /*
478 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
479 * (which zero-extends imm32) to save 2 bytes.
480 */
481 if (sign_propagate && (s32)imm32 < 0) {
482 /* 'mov %rax, imm32' sign extends imm32 */
483 b1 = add_1mod(0x48, dst_reg);
484 b2 = 0xC7;
485 b3 = 0xC0;
486 EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
487 goto done;
488 }
489
490 /*
491 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
492 * to save 3 bytes.
493 */
494 if (imm32 == 0) {
495 if (is_ereg(dst_reg))
496 EMIT1(add_2mod(0x40, dst_reg, dst_reg));
497 b2 = 0x31; /* xor */
498 b3 = 0xC0;
499 EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
500 goto done;
501 }
502
503 /* mov %eax, imm32 */
504 if (is_ereg(dst_reg))
505 EMIT1(add_1mod(0x40, dst_reg));
506 EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
507 done:
508 *pprog = prog;
509 }
510
511 static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
512 const u32 imm32_hi, const u32 imm32_lo)
513 {
514 u8 *prog = *pprog;
515 int cnt = 0;
516
517 if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
518 /*
519 * For emitting plain u32, where sign bit must not be
520 * propagated LLVM tends to load imm64 over mov32
521 * directly, so save couple of bytes by just doing
522 * 'mov %eax, imm32' instead.
523 */
524 emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
525 } else {
526 /* movabsq %rax, imm64 */
527 EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
528 EMIT(imm32_lo, 4);
529 EMIT(imm32_hi, 4);
530 }
531
532 *pprog = prog;
533 }
534
535 static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
536 {
537 u8 *prog = *pprog;
538 int cnt = 0;
539
540 if (is64) {
541 /* mov dst, src */
542 EMIT_mov(dst_reg, src_reg);
543 } else {
544 /* mov32 dst, src */
545 if (is_ereg(dst_reg) || is_ereg(src_reg))
546 EMIT1(add_2mod(0x40, dst_reg, src_reg));
547 EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
548 }
549
550 *pprog = prog;
551 }
552
553 /* LDX: dst_reg = *(u8*)(src_reg + off) */
554 static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
555 {
556 u8 *prog = *pprog;
557 int cnt = 0;
558
559 switch (size) {
560 case BPF_B:
561 /* Emit 'movzx rax, byte ptr [rax + off]' */
562 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
563 break;
564 case BPF_H:
565 /* Emit 'movzx rax, word ptr [rax + off]' */
566 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
567 break;
568 case BPF_W:
569 /* Emit 'mov eax, dword ptr [rax+0x14]' */
570 if (is_ereg(dst_reg) || is_ereg(src_reg))
571 EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
572 else
573 EMIT1(0x8B);
574 break;
575 case BPF_DW:
576 /* Emit 'mov rax, qword ptr [rax+0x14]' */
577 EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
578 break;
579 }
580 /*
581 * If insn->off == 0 we can save one extra byte, but
582 * special case of x86 R13 which always needs an offset
583 * is not worth the hassle
584 */
585 if (is_imm8(off))
586 EMIT2(add_2reg(0x40, src_reg, dst_reg), off);
587 else
588 EMIT1_off32(add_2reg(0x80, src_reg, dst_reg), off);
589 *pprog = prog;
590 }
591
592 /* STX: *(u8*)(dst_reg + off) = src_reg */
593 static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
594 {
595 u8 *prog = *pprog;
596 int cnt = 0;
597
598 switch (size) {
599 case BPF_B:
600 /* Emit 'mov byte ptr [rax + off], al' */
601 if (is_ereg(dst_reg) || is_ereg(src_reg) ||
602 /* We have to add extra byte for x86 SIL, DIL regs */
603 src_reg == BPF_REG_1 || src_reg == BPF_REG_2)
604 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
605 else
606 EMIT1(0x88);
607 break;
608 case BPF_H:
609 if (is_ereg(dst_reg) || is_ereg(src_reg))
610 EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
611 else
612 EMIT2(0x66, 0x89);
613 break;
614 case BPF_W:
615 if (is_ereg(dst_reg) || is_ereg(src_reg))
616 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
617 else
618 EMIT1(0x89);
619 break;
620 case BPF_DW:
621 EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
622 break;
623 }
624 if (is_imm8(off))
625 EMIT2(add_2reg(0x40, dst_reg, src_reg), off);
626 else
627 EMIT1_off32(add_2reg(0x80, dst_reg, src_reg), off);
628 *pprog = prog;
629 }
630
631 static bool ex_handler_bpf(const struct exception_table_entry *x,
632 struct pt_regs *regs, int trapnr,
633 unsigned long error_code, unsigned long fault_addr)
634 {
635 u32 reg = x->fixup >> 8;
636
637 /* jump over faulting load and clear dest register */
638 *(unsigned long *)((void *)regs + reg) = 0;
639 regs->ip += x->fixup & 0xff;
640 return true;
641 }
642
643 static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
644 int oldproglen, struct jit_context *ctx)
645 {
646 struct bpf_insn *insn = bpf_prog->insnsi;
647 int insn_cnt = bpf_prog->len;
648 bool seen_exit = false;
649 u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
650 int i, cnt = 0, excnt = 0;
651 int proglen = 0;
652 u8 *prog = temp;
653
654 emit_prologue(&prog, bpf_prog->aux->stack_depth,
655 bpf_prog_was_classic(bpf_prog));
656 addrs[0] = prog - temp;
657
658 for (i = 1; i <= insn_cnt; i++, insn++) {
659 const s32 imm32 = insn->imm;
660 u32 dst_reg = insn->dst_reg;
661 u32 src_reg = insn->src_reg;
662 u8 b2 = 0, b3 = 0;
663 s64 jmp_offset;
664 u8 jmp_cond;
665 int ilen;
666 u8 *func;
667
668 switch (insn->code) {
669 /* ALU */
670 case BPF_ALU | BPF_ADD | BPF_X:
671 case BPF_ALU | BPF_SUB | BPF_X:
672 case BPF_ALU | BPF_AND | BPF_X:
673 case BPF_ALU | BPF_OR | BPF_X:
674 case BPF_ALU | BPF_XOR | BPF_X:
675 case BPF_ALU64 | BPF_ADD | BPF_X:
676 case BPF_ALU64 | BPF_SUB | BPF_X:
677 case BPF_ALU64 | BPF_AND | BPF_X:
678 case BPF_ALU64 | BPF_OR | BPF_X:
679 case BPF_ALU64 | BPF_XOR | BPF_X:
680 switch (BPF_OP(insn->code)) {
681 case BPF_ADD: b2 = 0x01; break;
682 case BPF_SUB: b2 = 0x29; break;
683 case BPF_AND: b2 = 0x21; break;
684 case BPF_OR: b2 = 0x09; break;
685 case BPF_XOR: b2 = 0x31; break;
686 }
687 if (BPF_CLASS(insn->code) == BPF_ALU64)
688 EMIT1(add_2mod(0x48, dst_reg, src_reg));
689 else if (is_ereg(dst_reg) || is_ereg(src_reg))
690 EMIT1(add_2mod(0x40, dst_reg, src_reg));
691 EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
692 break;
693
694 case BPF_ALU64 | BPF_MOV | BPF_X:
695 case BPF_ALU | BPF_MOV | BPF_X:
696 emit_mov_reg(&prog,
697 BPF_CLASS(insn->code) == BPF_ALU64,
698 dst_reg, src_reg);
699 break;
700
701 /* neg dst */
702 case BPF_ALU | BPF_NEG:
703 case BPF_ALU64 | BPF_NEG:
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 EMIT2(0xF7, add_1reg(0xD8, dst_reg));
709 break;
710
711 case BPF_ALU | BPF_ADD | BPF_K:
712 case BPF_ALU | BPF_SUB | BPF_K:
713 case BPF_ALU | BPF_AND | BPF_K:
714 case BPF_ALU | BPF_OR | BPF_K:
715 case BPF_ALU | BPF_XOR | BPF_K:
716 case BPF_ALU64 | BPF_ADD | BPF_K:
717 case BPF_ALU64 | BPF_SUB | BPF_K:
718 case BPF_ALU64 | BPF_AND | BPF_K:
719 case BPF_ALU64 | BPF_OR | BPF_K:
720 case BPF_ALU64 | BPF_XOR | BPF_K:
721 if (BPF_CLASS(insn->code) == BPF_ALU64)
722 EMIT1(add_1mod(0x48, dst_reg));
723 else if (is_ereg(dst_reg))
724 EMIT1(add_1mod(0x40, dst_reg));
725
726 /*
727 * b3 holds 'normal' opcode, b2 short form only valid
728 * in case dst is eax/rax.
729 */
730 switch (BPF_OP(insn->code)) {
731 case BPF_ADD:
732 b3 = 0xC0;
733 b2 = 0x05;
734 break;
735 case BPF_SUB:
736 b3 = 0xE8;
737 b2 = 0x2D;
738 break;
739 case BPF_AND:
740 b3 = 0xE0;
741 b2 = 0x25;
742 break;
743 case BPF_OR:
744 b3 = 0xC8;
745 b2 = 0x0D;
746 break;
747 case BPF_XOR:
748 b3 = 0xF0;
749 b2 = 0x35;
750 break;
751 }
752
753 if (is_imm8(imm32))
754 EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
755 else if (is_axreg(dst_reg))
756 EMIT1_off32(b2, imm32);
757 else
758 EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
759 break;
760
761 case BPF_ALU64 | BPF_MOV | BPF_K:
762 case BPF_ALU | BPF_MOV | BPF_K:
763 emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
764 dst_reg, imm32);
765 break;
766
767 case BPF_LD | BPF_IMM | BPF_DW:
768 emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
769 insn++;
770 i++;
771 break;
772
773 /* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
774 case BPF_ALU | BPF_MOD | BPF_X:
775 case BPF_ALU | BPF_DIV | BPF_X:
776 case BPF_ALU | BPF_MOD | BPF_K:
777 case BPF_ALU | BPF_DIV | BPF_K:
778 case BPF_ALU64 | BPF_MOD | BPF_X:
779 case BPF_ALU64 | BPF_DIV | BPF_X:
780 case BPF_ALU64 | BPF_MOD | BPF_K:
781 case BPF_ALU64 | BPF_DIV | BPF_K:
782 EMIT1(0x50); /* push rax */
783 EMIT1(0x52); /* push rdx */
784
785 if (BPF_SRC(insn->code) == BPF_X)
786 /* mov r11, src_reg */
787 EMIT_mov(AUX_REG, src_reg);
788 else
789 /* mov r11, imm32 */
790 EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
791
792 /* mov rax, dst_reg */
793 EMIT_mov(BPF_REG_0, dst_reg);
794
795 /*
796 * xor edx, edx
797 * equivalent to 'xor rdx, rdx', but one byte less
798 */
799 EMIT2(0x31, 0xd2);
800
801 if (BPF_CLASS(insn->code) == BPF_ALU64)
802 /* div r11 */
803 EMIT3(0x49, 0xF7, 0xF3);
804 else
805 /* div r11d */
806 EMIT3(0x41, 0xF7, 0xF3);
807
808 if (BPF_OP(insn->code) == BPF_MOD)
809 /* mov r11, rdx */
810 EMIT3(0x49, 0x89, 0xD3);
811 else
812 /* mov r11, rax */
813 EMIT3(0x49, 0x89, 0xC3);
814
815 EMIT1(0x5A); /* pop rdx */
816 EMIT1(0x58); /* pop rax */
817
818 /* mov dst_reg, r11 */
819 EMIT_mov(dst_reg, AUX_REG);
820 break;
821
822 case BPF_ALU | BPF_MUL | BPF_K:
823 case BPF_ALU | BPF_MUL | BPF_X:
824 case BPF_ALU64 | BPF_MUL | BPF_K:
825 case BPF_ALU64 | BPF_MUL | BPF_X:
826 {
827 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
828
829 if (dst_reg != BPF_REG_0)
830 EMIT1(0x50); /* push rax */
831 if (dst_reg != BPF_REG_3)
832 EMIT1(0x52); /* push rdx */
833
834 /* mov r11, dst_reg */
835 EMIT_mov(AUX_REG, dst_reg);
836
837 if (BPF_SRC(insn->code) == BPF_X)
838 emit_mov_reg(&prog, is64, BPF_REG_0, src_reg);
839 else
840 emit_mov_imm32(&prog, is64, BPF_REG_0, imm32);
841
842 if (is64)
843 EMIT1(add_1mod(0x48, AUX_REG));
844 else if (is_ereg(AUX_REG))
845 EMIT1(add_1mod(0x40, AUX_REG));
846 /* mul(q) r11 */
847 EMIT2(0xF7, add_1reg(0xE0, AUX_REG));
848
849 if (dst_reg != BPF_REG_3)
850 EMIT1(0x5A); /* pop rdx */
851 if (dst_reg != BPF_REG_0) {
852 /* mov dst_reg, rax */
853 EMIT_mov(dst_reg, BPF_REG_0);
854 EMIT1(0x58); /* pop rax */
855 }
856 break;
857 }
858 /* Shifts */
859 case BPF_ALU | BPF_LSH | BPF_K:
860 case BPF_ALU | BPF_RSH | BPF_K:
861 case BPF_ALU | BPF_ARSH | BPF_K:
862 case BPF_ALU64 | BPF_LSH | BPF_K:
863 case BPF_ALU64 | BPF_RSH | BPF_K:
864 case BPF_ALU64 | BPF_ARSH | BPF_K:
865 if (BPF_CLASS(insn->code) == BPF_ALU64)
866 EMIT1(add_1mod(0x48, dst_reg));
867 else if (is_ereg(dst_reg))
868 EMIT1(add_1mod(0x40, dst_reg));
869
870 switch (BPF_OP(insn->code)) {
871 case BPF_LSH: b3 = 0xE0; break;
872 case BPF_RSH: b3 = 0xE8; break;
873 case BPF_ARSH: b3 = 0xF8; break;
874 }
875
876 if (imm32 == 1)
877 EMIT2(0xD1, add_1reg(b3, dst_reg));
878 else
879 EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
880 break;
881
882 case BPF_ALU | BPF_LSH | BPF_X:
883 case BPF_ALU | BPF_RSH | BPF_X:
884 case BPF_ALU | BPF_ARSH | BPF_X:
885 case BPF_ALU64 | BPF_LSH | BPF_X:
886 case BPF_ALU64 | BPF_RSH | BPF_X:
887 case BPF_ALU64 | BPF_ARSH | BPF_X:
888
889 /* Check for bad case when dst_reg == rcx */
890 if (dst_reg == BPF_REG_4) {
891 /* mov r11, dst_reg */
892 EMIT_mov(AUX_REG, dst_reg);
893 dst_reg = AUX_REG;
894 }
895
896 if (src_reg != BPF_REG_4) { /* common case */
897 EMIT1(0x51); /* push rcx */
898
899 /* mov rcx, src_reg */
900 EMIT_mov(BPF_REG_4, src_reg);
901 }
902
903 /* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
904 if (BPF_CLASS(insn->code) == BPF_ALU64)
905 EMIT1(add_1mod(0x48, dst_reg));
906 else if (is_ereg(dst_reg))
907 EMIT1(add_1mod(0x40, dst_reg));
908
909 switch (BPF_OP(insn->code)) {
910 case BPF_LSH: b3 = 0xE0; break;
911 case BPF_RSH: b3 = 0xE8; break;
912 case BPF_ARSH: b3 = 0xF8; break;
913 }
914 EMIT2(0xD3, add_1reg(b3, dst_reg));
915
916 if (src_reg != BPF_REG_4)
917 EMIT1(0x59); /* pop rcx */
918
919 if (insn->dst_reg == BPF_REG_4)
920 /* mov dst_reg, r11 */
921 EMIT_mov(insn->dst_reg, AUX_REG);
922 break;
923
924 case BPF_ALU | BPF_END | BPF_FROM_BE:
925 switch (imm32) {
926 case 16:
927 /* Emit 'ror %ax, 8' to swap lower 2 bytes */
928 EMIT1(0x66);
929 if (is_ereg(dst_reg))
930 EMIT1(0x41);
931 EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
932
933 /* Emit 'movzwl eax, ax' */
934 if (is_ereg(dst_reg))
935 EMIT3(0x45, 0x0F, 0xB7);
936 else
937 EMIT2(0x0F, 0xB7);
938 EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
939 break;
940 case 32:
941 /* Emit 'bswap eax' to swap lower 4 bytes */
942 if (is_ereg(dst_reg))
943 EMIT2(0x41, 0x0F);
944 else
945 EMIT1(0x0F);
946 EMIT1(add_1reg(0xC8, dst_reg));
947 break;
948 case 64:
949 /* Emit 'bswap rax' to swap 8 bytes */
950 EMIT3(add_1mod(0x48, dst_reg), 0x0F,
951 add_1reg(0xC8, dst_reg));
952 break;
953 }
954 break;
955
956 case BPF_ALU | BPF_END | BPF_FROM_LE:
957 switch (imm32) {
958 case 16:
959 /*
960 * Emit 'movzwl eax, ax' to zero extend 16-bit
961 * into 64 bit
962 */
963 if (is_ereg(dst_reg))
964 EMIT3(0x45, 0x0F, 0xB7);
965 else
966 EMIT2(0x0F, 0xB7);
967 EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
968 break;
969 case 32:
970 /* Emit 'mov eax, eax' to clear upper 32-bits */
971 if (is_ereg(dst_reg))
972 EMIT1(0x45);
973 EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
974 break;
975 case 64:
976 /* nop */
977 break;
978 }
979 break;
980
981 /* ST: *(u8*)(dst_reg + off) = imm */
982 case BPF_ST | BPF_MEM | BPF_B:
983 if (is_ereg(dst_reg))
984 EMIT2(0x41, 0xC6);
985 else
986 EMIT1(0xC6);
987 goto st;
988 case BPF_ST | BPF_MEM | BPF_H:
989 if (is_ereg(dst_reg))
990 EMIT3(0x66, 0x41, 0xC7);
991 else
992 EMIT2(0x66, 0xC7);
993 goto st;
994 case BPF_ST | BPF_MEM | BPF_W:
995 if (is_ereg(dst_reg))
996 EMIT2(0x41, 0xC7);
997 else
998 EMIT1(0xC7);
999 goto st;
1000 case BPF_ST | BPF_MEM | BPF_DW:
1001 EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1002
1003 st: if (is_imm8(insn->off))
1004 EMIT2(add_1reg(0x40, dst_reg), insn->off);
1005 else
1006 EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1007
1008 EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1009 break;
1010
1011 /* STX: *(u8*)(dst_reg + off) = src_reg */
1012 case BPF_STX | BPF_MEM | BPF_B:
1013 case BPF_STX | BPF_MEM | BPF_H:
1014 case BPF_STX | BPF_MEM | BPF_W:
1015 case BPF_STX | BPF_MEM | BPF_DW:
1016 emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1017 break;
1018
1019 /* LDX: dst_reg = *(u8*)(src_reg + off) */
1020 case BPF_LDX | BPF_MEM | BPF_B:
1021 case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1022 case BPF_LDX | BPF_MEM | BPF_H:
1023 case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1024 case BPF_LDX | BPF_MEM | BPF_W:
1025 case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1026 case BPF_LDX | BPF_MEM | BPF_DW:
1027 case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1028 emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1029 if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1030 struct exception_table_entry *ex;
1031 u8 *_insn = image + proglen;
1032 s64 delta;
1033
1034 if (!bpf_prog->aux->extable)
1035 break;
1036
1037 if (excnt >= bpf_prog->aux->num_exentries) {
1038 pr_err("ex gen bug\n");
1039 return -EFAULT;
1040 }
1041 ex = &bpf_prog->aux->extable[excnt++];
1042
1043 delta = _insn - (u8 *)&ex->insn;
1044 if (!is_simm32(delta)) {
1045 pr_err("extable->insn doesn't fit into 32-bit\n");
1046 return -EFAULT;
1047 }
1048 ex->insn = delta;
1049
1050 delta = (u8 *)ex_handler_bpf - (u8 *)&ex->handler;
1051 if (!is_simm32(delta)) {
1052 pr_err("extable->handler doesn't fit into 32-bit\n");
1053 return -EFAULT;
1054 }
1055 ex->handler = delta;
1056
1057 if (dst_reg > BPF_REG_9) {
1058 pr_err("verifier error\n");
1059 return -EFAULT;
1060 }
1061 /*
1062 * Compute size of x86 insn and its target dest x86 register.
1063 * ex_handler_bpf() will use lower 8 bits to adjust
1064 * pt_regs->ip to jump over this x86 instruction
1065 * and upper bits to figure out which pt_regs to zero out.
1066 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1067 * of 4 bytes will be ignored and rbx will be zero inited.
1068 */
1069 ex->fixup = (prog - temp) | (reg2pt_regs[dst_reg] << 8);
1070 }
1071 break;
1072
1073 /* STX XADD: lock *(u32*)(dst_reg + off) += src_reg */
1074 case BPF_STX | BPF_XADD | BPF_W:
1075 /* Emit 'lock add dword ptr [rax + off], eax' */
1076 if (is_ereg(dst_reg) || is_ereg(src_reg))
1077 EMIT3(0xF0, add_2mod(0x40, dst_reg, src_reg), 0x01);
1078 else
1079 EMIT2(0xF0, 0x01);
1080 goto xadd;
1081 case BPF_STX | BPF_XADD | BPF_DW:
1082 EMIT3(0xF0, add_2mod(0x48, dst_reg, src_reg), 0x01);
1083 xadd: if (is_imm8(insn->off))
1084 EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off);
1085 else
1086 EMIT1_off32(add_2reg(0x80, dst_reg, src_reg),
1087 insn->off);
1088 break;
1089
1090 /* call */
1091 case BPF_JMP | BPF_CALL:
1092 func = (u8 *) __bpf_call_base + imm32;
1093 if (!imm32 || emit_call(&prog, func, image + addrs[i - 1]))
1094 return -EINVAL;
1095 break;
1096
1097 case BPF_JMP | BPF_TAIL_CALL:
1098 if (imm32)
1099 emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1100 &prog, addrs[i], image);
1101 else
1102 emit_bpf_tail_call_indirect(&prog);
1103 break;
1104
1105 /* cond jump */
1106 case BPF_JMP | BPF_JEQ | BPF_X:
1107 case BPF_JMP | BPF_JNE | BPF_X:
1108 case BPF_JMP | BPF_JGT | BPF_X:
1109 case BPF_JMP | BPF_JLT | BPF_X:
1110 case BPF_JMP | BPF_JGE | BPF_X:
1111 case BPF_JMP | BPF_JLE | BPF_X:
1112 case BPF_JMP | BPF_JSGT | BPF_X:
1113 case BPF_JMP | BPF_JSLT | BPF_X:
1114 case BPF_JMP | BPF_JSGE | BPF_X:
1115 case BPF_JMP | BPF_JSLE | BPF_X:
1116 case BPF_JMP32 | BPF_JEQ | BPF_X:
1117 case BPF_JMP32 | BPF_JNE | BPF_X:
1118 case BPF_JMP32 | BPF_JGT | BPF_X:
1119 case BPF_JMP32 | BPF_JLT | BPF_X:
1120 case BPF_JMP32 | BPF_JGE | BPF_X:
1121 case BPF_JMP32 | BPF_JLE | BPF_X:
1122 case BPF_JMP32 | BPF_JSGT | BPF_X:
1123 case BPF_JMP32 | BPF_JSLT | BPF_X:
1124 case BPF_JMP32 | BPF_JSGE | BPF_X:
1125 case BPF_JMP32 | BPF_JSLE | BPF_X:
1126 /* cmp dst_reg, src_reg */
1127 if (BPF_CLASS(insn->code) == BPF_JMP)
1128 EMIT1(add_2mod(0x48, dst_reg, src_reg));
1129 else if (is_ereg(dst_reg) || is_ereg(src_reg))
1130 EMIT1(add_2mod(0x40, dst_reg, src_reg));
1131 EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1132 goto emit_cond_jmp;
1133
1134 case BPF_JMP | BPF_JSET | BPF_X:
1135 case BPF_JMP32 | BPF_JSET | BPF_X:
1136 /* test dst_reg, src_reg */
1137 if (BPF_CLASS(insn->code) == BPF_JMP)
1138 EMIT1(add_2mod(0x48, dst_reg, src_reg));
1139 else if (is_ereg(dst_reg) || is_ereg(src_reg))
1140 EMIT1(add_2mod(0x40, dst_reg, src_reg));
1141 EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1142 goto emit_cond_jmp;
1143
1144 case BPF_JMP | BPF_JSET | BPF_K:
1145 case BPF_JMP32 | BPF_JSET | BPF_K:
1146 /* test dst_reg, imm32 */
1147 if (BPF_CLASS(insn->code) == BPF_JMP)
1148 EMIT1(add_1mod(0x48, dst_reg));
1149 else if (is_ereg(dst_reg))
1150 EMIT1(add_1mod(0x40, dst_reg));
1151 EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1152 goto emit_cond_jmp;
1153
1154 case BPF_JMP | BPF_JEQ | BPF_K:
1155 case BPF_JMP | BPF_JNE | BPF_K:
1156 case BPF_JMP | BPF_JGT | BPF_K:
1157 case BPF_JMP | BPF_JLT | BPF_K:
1158 case BPF_JMP | BPF_JGE | BPF_K:
1159 case BPF_JMP | BPF_JLE | BPF_K:
1160 case BPF_JMP | BPF_JSGT | BPF_K:
1161 case BPF_JMP | BPF_JSLT | BPF_K:
1162 case BPF_JMP | BPF_JSGE | BPF_K:
1163 case BPF_JMP | BPF_JSLE | BPF_K:
1164 case BPF_JMP32 | BPF_JEQ | BPF_K:
1165 case BPF_JMP32 | BPF_JNE | BPF_K:
1166 case BPF_JMP32 | BPF_JGT | BPF_K:
1167 case BPF_JMP32 | BPF_JLT | BPF_K:
1168 case BPF_JMP32 | BPF_JGE | BPF_K:
1169 case BPF_JMP32 | BPF_JLE | BPF_K:
1170 case BPF_JMP32 | BPF_JSGT | BPF_K:
1171 case BPF_JMP32 | BPF_JSLT | BPF_K:
1172 case BPF_JMP32 | BPF_JSGE | BPF_K:
1173 case BPF_JMP32 | BPF_JSLE | BPF_K:
1174 /* test dst_reg, dst_reg to save one extra byte */
1175 if (imm32 == 0) {
1176 if (BPF_CLASS(insn->code) == BPF_JMP)
1177 EMIT1(add_2mod(0x48, dst_reg, dst_reg));
1178 else if (is_ereg(dst_reg))
1179 EMIT1(add_2mod(0x40, dst_reg, dst_reg));
1180 EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1181 goto emit_cond_jmp;
1182 }
1183
1184 /* cmp dst_reg, imm8/32 */
1185 if (BPF_CLASS(insn->code) == BPF_JMP)
1186 EMIT1(add_1mod(0x48, dst_reg));
1187 else if (is_ereg(dst_reg))
1188 EMIT1(add_1mod(0x40, dst_reg));
1189
1190 if (is_imm8(imm32))
1191 EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1192 else
1193 EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1194
1195 emit_cond_jmp: /* Convert BPF opcode to x86 */
1196 switch (BPF_OP(insn->code)) {
1197 case BPF_JEQ:
1198 jmp_cond = X86_JE;
1199 break;
1200 case BPF_JSET:
1201 case BPF_JNE:
1202 jmp_cond = X86_JNE;
1203 break;
1204 case BPF_JGT:
1205 /* GT is unsigned '>', JA in x86 */
1206 jmp_cond = X86_JA;
1207 break;
1208 case BPF_JLT:
1209 /* LT is unsigned '<', JB in x86 */
1210 jmp_cond = X86_JB;
1211 break;
1212 case BPF_JGE:
1213 /* GE is unsigned '>=', JAE in x86 */
1214 jmp_cond = X86_JAE;
1215 break;
1216 case BPF_JLE:
1217 /* LE is unsigned '<=', JBE in x86 */
1218 jmp_cond = X86_JBE;
1219 break;
1220 case BPF_JSGT:
1221 /* Signed '>', GT in x86 */
1222 jmp_cond = X86_JG;
1223 break;
1224 case BPF_JSLT:
1225 /* Signed '<', LT in x86 */
1226 jmp_cond = X86_JL;
1227 break;
1228 case BPF_JSGE:
1229 /* Signed '>=', GE in x86 */
1230 jmp_cond = X86_JGE;
1231 break;
1232 case BPF_JSLE:
1233 /* Signed '<=', LE in x86 */
1234 jmp_cond = X86_JLE;
1235 break;
1236 default: /* to silence GCC warning */
1237 return -EFAULT;
1238 }
1239 jmp_offset = addrs[i + insn->off] - addrs[i];
1240 if (is_imm8(jmp_offset)) {
1241 EMIT2(jmp_cond, jmp_offset);
1242 } else if (is_simm32(jmp_offset)) {
1243 EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1244 } else {
1245 pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1246 return -EFAULT;
1247 }
1248
1249 break;
1250
1251 case BPF_JMP | BPF_JA:
1252 if (insn->off == -1)
1253 /* -1 jmp instructions will always jump
1254 * backwards two bytes. Explicitly handling
1255 * this case avoids wasting too many passes
1256 * when there are long sequences of replaced
1257 * dead code.
1258 */
1259 jmp_offset = -2;
1260 else
1261 jmp_offset = addrs[i + insn->off] - addrs[i];
1262
1263 if (!jmp_offset)
1264 /* Optimize out nop jumps */
1265 break;
1266 emit_jmp:
1267 if (is_imm8(jmp_offset)) {
1268 EMIT2(0xEB, jmp_offset);
1269 } else if (is_simm32(jmp_offset)) {
1270 EMIT1_off32(0xE9, jmp_offset);
1271 } else {
1272 pr_err("jmp gen bug %llx\n", jmp_offset);
1273 return -EFAULT;
1274 }
1275 break;
1276
1277 case BPF_JMP | BPF_EXIT:
1278 if (seen_exit) {
1279 jmp_offset = ctx->cleanup_addr - addrs[i];
1280 goto emit_jmp;
1281 }
1282 seen_exit = true;
1283 /* Update cleanup_addr */
1284 ctx->cleanup_addr = proglen;
1285 if (!bpf_prog_was_classic(bpf_prog))
1286 EMIT1(0x5B); /* get rid of tail_call_cnt */
1287 EMIT2(0x41, 0x5F); /* pop r15 */
1288 EMIT2(0x41, 0x5E); /* pop r14 */
1289 EMIT2(0x41, 0x5D); /* pop r13 */
1290 EMIT1(0x5B); /* pop rbx */
1291 EMIT1(0xC9); /* leave */
1292 EMIT1(0xC3); /* ret */
1293 break;
1294
1295 default:
1296 /*
1297 * By design x86-64 JIT should support all BPF instructions.
1298 * This error will be seen if new instruction was added
1299 * to the interpreter, but not to the JIT, or if there is
1300 * junk in bpf_prog.
1301 */
1302 pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1303 return -EINVAL;
1304 }
1305
1306 ilen = prog - temp;
1307 if (ilen > BPF_MAX_INSN_SIZE) {
1308 pr_err("bpf_jit: fatal insn size error\n");
1309 return -EFAULT;
1310 }
1311
1312 if (image) {
1313 if (unlikely(proglen + ilen > oldproglen)) {
1314 pr_err("bpf_jit: fatal error\n");
1315 return -EFAULT;
1316 }
1317 memcpy(image + proglen, temp, ilen);
1318 }
1319 proglen += ilen;
1320 addrs[i] = proglen;
1321 prog = temp;
1322 }
1323
1324 if (image && excnt != bpf_prog->aux->num_exentries) {
1325 pr_err("extable is not populated\n");
1326 return -EFAULT;
1327 }
1328 return proglen;
1329 }
1330
1331 static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1332 int stack_size)
1333 {
1334 int i;
1335 /* Store function arguments to stack.
1336 * For a function that accepts two pointers the sequence will be:
1337 * mov QWORD PTR [rbp-0x10],rdi
1338 * mov QWORD PTR [rbp-0x8],rsi
1339 */
1340 for (i = 0; i < min(nr_args, 6); i++)
1341 emit_stx(prog, bytes_to_bpf_size(m->arg_size[i]),
1342 BPF_REG_FP,
1343 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1344 -(stack_size - i * 8));
1345 }
1346
1347 static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1348 int stack_size)
1349 {
1350 int i;
1351
1352 /* Restore function arguments from stack.
1353 * For a function that accepts two pointers the sequence will be:
1354 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1355 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1356 */
1357 for (i = 0; i < min(nr_args, 6); i++)
1358 emit_ldx(prog, bytes_to_bpf_size(m->arg_size[i]),
1359 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1360 BPF_REG_FP,
1361 -(stack_size - i * 8));
1362 }
1363
1364 static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1365 struct bpf_prog *p, int stack_size, bool mod_ret)
1366 {
1367 u8 *prog = *pprog;
1368 int cnt = 0;
1369
1370 if (emit_call(&prog, __bpf_prog_enter, prog))
1371 return -EINVAL;
1372 /* remember prog start time returned by __bpf_prog_enter */
1373 emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1374
1375 /* arg1: lea rdi, [rbp - stack_size] */
1376 EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1377 /* arg2: progs[i]->insnsi for interpreter */
1378 if (!p->jited)
1379 emit_mov_imm64(&prog, BPF_REG_2,
1380 (long) p->insnsi >> 32,
1381 (u32) (long) p->insnsi);
1382 /* call JITed bpf program or interpreter */
1383 if (emit_call(&prog, p->bpf_func, prog))
1384 return -EINVAL;
1385
1386 /* BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1387 * of the previous call which is then passed on the stack to
1388 * the next BPF program.
1389 */
1390 if (mod_ret)
1391 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1392
1393 /* arg1: mov rdi, progs[i] */
1394 emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32,
1395 (u32) (long) p);
1396 /* arg2: mov rsi, rbx <- start time in nsec */
1397 emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1398 if (emit_call(&prog, __bpf_prog_exit, prog))
1399 return -EINVAL;
1400
1401 *pprog = prog;
1402 return 0;
1403 }
1404
1405 static void emit_nops(u8 **pprog, unsigned int len)
1406 {
1407 unsigned int i, noplen;
1408 u8 *prog = *pprog;
1409 int cnt = 0;
1410
1411 while (len > 0) {
1412 noplen = len;
1413
1414 if (noplen > ASM_NOP_MAX)
1415 noplen = ASM_NOP_MAX;
1416
1417 for (i = 0; i < noplen; i++)
1418 EMIT1(ideal_nops[noplen][i]);
1419 len -= noplen;
1420 }
1421
1422 *pprog = prog;
1423 }
1424
1425 static void emit_align(u8 **pprog, u32 align)
1426 {
1427 u8 *target, *prog = *pprog;
1428
1429 target = PTR_ALIGN(prog, align);
1430 if (target != prog)
1431 emit_nops(&prog, target - prog);
1432
1433 *pprog = prog;
1434 }
1435
1436 static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
1437 {
1438 u8 *prog = *pprog;
1439 int cnt = 0;
1440 s64 offset;
1441
1442 offset = func - (ip + 2 + 4);
1443 if (!is_simm32(offset)) {
1444 pr_err("Target %p is out of range\n", func);
1445 return -EINVAL;
1446 }
1447 EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
1448 *pprog = prog;
1449 return 0;
1450 }
1451
1452 static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
1453 struct bpf_tramp_progs *tp, int stack_size)
1454 {
1455 int i;
1456 u8 *prog = *pprog;
1457
1458 for (i = 0; i < tp->nr_progs; i++) {
1459 if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size, false))
1460 return -EINVAL;
1461 }
1462 *pprog = prog;
1463 return 0;
1464 }
1465
1466 static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
1467 struct bpf_tramp_progs *tp, int stack_size,
1468 u8 **branches)
1469 {
1470 u8 *prog = *pprog;
1471 int i, cnt = 0;
1472
1473 /* The first fmod_ret program will receive a garbage return value.
1474 * Set this to 0 to avoid confusing the program.
1475 */
1476 emit_mov_imm32(&prog, false, BPF_REG_0, 0);
1477 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1478 for (i = 0; i < tp->nr_progs; i++) {
1479 if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size, true))
1480 return -EINVAL;
1481
1482 /* mod_ret prog stored return value into [rbp - 8]. Emit:
1483 * if (*(u64 *)(rbp - 8) != 0)
1484 * goto do_fexit;
1485 */
1486 /* cmp QWORD PTR [rbp - 0x8], 0x0 */
1487 EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
1488
1489 /* Save the location of the branch and Generate 6 nops
1490 * (4 bytes for an offset and 2 bytes for the jump) These nops
1491 * are replaced with a conditional jump once do_fexit (i.e. the
1492 * start of the fexit invocation) is finalized.
1493 */
1494 branches[i] = prog;
1495 emit_nops(&prog, 4 + 2);
1496 }
1497
1498 *pprog = prog;
1499 return 0;
1500 }
1501
1502 /* Example:
1503 * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
1504 * its 'struct btf_func_model' will be nr_args=2
1505 * The assembly code when eth_type_trans is executing after trampoline:
1506 *
1507 * push rbp
1508 * mov rbp, rsp
1509 * sub rsp, 16 // space for skb and dev
1510 * push rbx // temp regs to pass start time
1511 * mov qword ptr [rbp - 16], rdi // save skb pointer to stack
1512 * mov qword ptr [rbp - 8], rsi // save dev pointer to stack
1513 * call __bpf_prog_enter // rcu_read_lock and preempt_disable
1514 * mov rbx, rax // remember start time in bpf stats are enabled
1515 * lea rdi, [rbp - 16] // R1==ctx of bpf prog
1516 * call addr_of_jited_FENTRY_prog
1517 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
1518 * mov rsi, rbx // prog start time
1519 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
1520 * mov rdi, qword ptr [rbp - 16] // restore skb pointer from stack
1521 * mov rsi, qword ptr [rbp - 8] // restore dev pointer from stack
1522 * pop rbx
1523 * leave
1524 * ret
1525 *
1526 * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
1527 * replaced with 'call generated_bpf_trampoline'. When it returns
1528 * eth_type_trans will continue executing with original skb and dev pointers.
1529 *
1530 * The assembly code when eth_type_trans is called from trampoline:
1531 *
1532 * push rbp
1533 * mov rbp, rsp
1534 * sub rsp, 24 // space for skb, dev, return value
1535 * push rbx // temp regs to pass start time
1536 * mov qword ptr [rbp - 24], rdi // save skb pointer to stack
1537 * mov qword ptr [rbp - 16], rsi // save dev pointer to stack
1538 * call __bpf_prog_enter // rcu_read_lock and preempt_disable
1539 * mov rbx, rax // remember start time if bpf stats are enabled
1540 * lea rdi, [rbp - 24] // R1==ctx of bpf prog
1541 * call addr_of_jited_FENTRY_prog // bpf prog can access skb and dev
1542 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
1543 * mov rsi, rbx // prog start time
1544 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
1545 * mov rdi, qword ptr [rbp - 24] // restore skb pointer from stack
1546 * mov rsi, qword ptr [rbp - 16] // restore dev pointer from stack
1547 * call eth_type_trans+5 // execute body of eth_type_trans
1548 * mov qword ptr [rbp - 8], rax // save return value
1549 * call __bpf_prog_enter // rcu_read_lock and preempt_disable
1550 * mov rbx, rax // remember start time in bpf stats are enabled
1551 * lea rdi, [rbp - 24] // R1==ctx of bpf prog
1552 * call addr_of_jited_FEXIT_prog // bpf prog can access skb, dev, return value
1553 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
1554 * mov rsi, rbx // prog start time
1555 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
1556 * mov rax, qword ptr [rbp - 8] // restore eth_type_trans's return value
1557 * pop rbx
1558 * leave
1559 * add rsp, 8 // skip eth_type_trans's frame
1560 * ret // return to its caller
1561 */
1562 int arch_prepare_bpf_trampoline(void *image, void *image_end,
1563 const struct btf_func_model *m, u32 flags,
1564 struct bpf_tramp_progs *tprogs,
1565 void *orig_call)
1566 {
1567 int ret, i, cnt = 0, nr_args = m->nr_args;
1568 int stack_size = nr_args * 8;
1569 struct bpf_tramp_progs *fentry = &tprogs[BPF_TRAMP_FENTRY];
1570 struct bpf_tramp_progs *fexit = &tprogs[BPF_TRAMP_FEXIT];
1571 struct bpf_tramp_progs *fmod_ret = &tprogs[BPF_TRAMP_MODIFY_RETURN];
1572 u8 **branches = NULL;
1573 u8 *prog;
1574
1575 /* x86-64 supports up to 6 arguments. 7+ can be added in the future */
1576 if (nr_args > 6)
1577 return -ENOTSUPP;
1578
1579 if ((flags & BPF_TRAMP_F_RESTORE_REGS) &&
1580 (flags & BPF_TRAMP_F_SKIP_FRAME))
1581 return -EINVAL;
1582
1583 if (flags & BPF_TRAMP_F_CALL_ORIG)
1584 stack_size += 8; /* room for return value of orig_call */
1585
1586 if (flags & BPF_TRAMP_F_SKIP_FRAME)
1587 /* skip patched call instruction and point orig_call to actual
1588 * body of the kernel function.
1589 */
1590 orig_call += X86_PATCH_SIZE;
1591
1592 prog = image;
1593
1594 EMIT1(0x55); /* push rbp */
1595 EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
1596 EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
1597 EMIT1(0x53); /* push rbx */
1598
1599 save_regs(m, &prog, nr_args, stack_size);
1600
1601 if (fentry->nr_progs)
1602 if (invoke_bpf(m, &prog, fentry, stack_size))
1603 return -EINVAL;
1604
1605 if (fmod_ret->nr_progs) {
1606 branches = kcalloc(fmod_ret->nr_progs, sizeof(u8 *),
1607 GFP_KERNEL);
1608 if (!branches)
1609 return -ENOMEM;
1610
1611 if (invoke_bpf_mod_ret(m, &prog, fmod_ret, stack_size,
1612 branches)) {
1613 ret = -EINVAL;
1614 goto cleanup;
1615 }
1616 }
1617
1618 if (flags & BPF_TRAMP_F_CALL_ORIG) {
1619 if (fentry->nr_progs || fmod_ret->nr_progs)
1620 restore_regs(m, &prog, nr_args, stack_size);
1621
1622 /* call original function */
1623 if (emit_call(&prog, orig_call, prog)) {
1624 ret = -EINVAL;
1625 goto cleanup;
1626 }
1627 /* remember return value in a stack for bpf prog to access */
1628 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1629 }
1630
1631 if (fmod_ret->nr_progs) {
1632 /* From Intel 64 and IA-32 Architectures Optimization
1633 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
1634 * Coding Rule 11: All branch targets should be 16-byte
1635 * aligned.
1636 */
1637 emit_align(&prog, 16);
1638 /* Update the branches saved in invoke_bpf_mod_ret with the
1639 * aligned address of do_fexit.
1640 */
1641 for (i = 0; i < fmod_ret->nr_progs; i++)
1642 emit_cond_near_jump(&branches[i], prog, branches[i],
1643 X86_JNE);
1644 }
1645
1646 if (fexit->nr_progs)
1647 if (invoke_bpf(m, &prog, fexit, stack_size)) {
1648 ret = -EINVAL;
1649 goto cleanup;
1650 }
1651
1652 if (flags & BPF_TRAMP_F_RESTORE_REGS)
1653 restore_regs(m, &prog, nr_args, stack_size);
1654
1655 /* This needs to be done regardless. If there were fmod_ret programs,
1656 * the return value is only updated on the stack and still needs to be
1657 * restored to R0.
1658 */
1659 if (flags & BPF_TRAMP_F_CALL_ORIG)
1660 /* restore original return value back into RAX */
1661 emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
1662
1663 EMIT1(0x5B); /* pop rbx */
1664 EMIT1(0xC9); /* leave */
1665 if (flags & BPF_TRAMP_F_SKIP_FRAME)
1666 /* skip our return address and return to parent */
1667 EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
1668 EMIT1(0xC3); /* ret */
1669 /* Make sure the trampoline generation logic doesn't overflow */
1670 if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
1671 ret = -EFAULT;
1672 goto cleanup;
1673 }
1674 ret = prog - (u8 *)image;
1675
1676 cleanup:
1677 kfree(branches);
1678 return ret;
1679 }
1680
1681 static int emit_fallback_jump(u8 **pprog)
1682 {
1683 u8 *prog = *pprog;
1684 int err = 0;
1685
1686 #ifdef CONFIG_RETPOLINE
1687 /* Note that this assumes the the compiler uses external
1688 * thunks for indirect calls. Both clang and GCC use the same
1689 * naming convention for external thunks.
1690 */
1691 err = emit_jump(&prog, __x86_indirect_thunk_rdx, prog);
1692 #else
1693 int cnt = 0;
1694
1695 EMIT2(0xFF, 0xE2); /* jmp rdx */
1696 #endif
1697 *pprog = prog;
1698 return err;
1699 }
1700
1701 static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs)
1702 {
1703 u8 *jg_reloc, *prog = *pprog;
1704 int pivot, err, jg_bytes = 1, cnt = 0;
1705 s64 jg_offset;
1706
1707 if (a == b) {
1708 /* Leaf node of recursion, i.e. not a range of indices
1709 * anymore.
1710 */
1711 EMIT1(add_1mod(0x48, BPF_REG_3)); /* cmp rdx,func */
1712 if (!is_simm32(progs[a]))
1713 return -1;
1714 EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
1715 progs[a]);
1716 err = emit_cond_near_jump(&prog, /* je func */
1717 (void *)progs[a], prog,
1718 X86_JE);
1719 if (err)
1720 return err;
1721
1722 err = emit_fallback_jump(&prog); /* jmp thunk/indirect */
1723 if (err)
1724 return err;
1725
1726 *pprog = prog;
1727 return 0;
1728 }
1729
1730 /* Not a leaf node, so we pivot, and recursively descend into
1731 * the lower and upper ranges.
1732 */
1733 pivot = (b - a) / 2;
1734 EMIT1(add_1mod(0x48, BPF_REG_3)); /* cmp rdx,func */
1735 if (!is_simm32(progs[a + pivot]))
1736 return -1;
1737 EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
1738
1739 if (pivot > 2) { /* jg upper_part */
1740 /* Require near jump. */
1741 jg_bytes = 4;
1742 EMIT2_off32(0x0F, X86_JG + 0x10, 0);
1743 } else {
1744 EMIT2(X86_JG, 0);
1745 }
1746 jg_reloc = prog;
1747
1748 err = emit_bpf_dispatcher(&prog, a, a + pivot, /* emit lower_part */
1749 progs);
1750 if (err)
1751 return err;
1752
1753 /* From Intel 64 and IA-32 Architectures Optimization
1754 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
1755 * Coding Rule 11: All branch targets should be 16-byte
1756 * aligned.
1757 */
1758 emit_align(&prog, 16);
1759 jg_offset = prog - jg_reloc;
1760 emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
1761
1762 err = emit_bpf_dispatcher(&prog, a + pivot + 1, /* emit upper_part */
1763 b, progs);
1764 if (err)
1765 return err;
1766
1767 *pprog = prog;
1768 return 0;
1769 }
1770
1771 static int cmp_ips(const void *a, const void *b)
1772 {
1773 const s64 *ipa = a;
1774 const s64 *ipb = b;
1775
1776 if (*ipa > *ipb)
1777 return 1;
1778 if (*ipa < *ipb)
1779 return -1;
1780 return 0;
1781 }
1782
1783 int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs)
1784 {
1785 u8 *prog = image;
1786
1787 sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
1788 return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs);
1789 }
1790
1791 struct x64_jit_data {
1792 struct bpf_binary_header *header;
1793 int *addrs;
1794 u8 *image;
1795 int proglen;
1796 struct jit_context ctx;
1797 };
1798
1799 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1800 {
1801 struct bpf_binary_header *header = NULL;
1802 struct bpf_prog *tmp, *orig_prog = prog;
1803 struct x64_jit_data *jit_data;
1804 int proglen, oldproglen = 0;
1805 struct jit_context ctx = {};
1806 bool tmp_blinded = false;
1807 bool extra_pass = false;
1808 u8 *image = NULL;
1809 int *addrs;
1810 int pass;
1811 int i;
1812
1813 if (!prog->jit_requested)
1814 return orig_prog;
1815
1816 tmp = bpf_jit_blind_constants(prog);
1817 /*
1818 * If blinding was requested and we failed during blinding,
1819 * we must fall back to the interpreter.
1820 */
1821 if (IS_ERR(tmp))
1822 return orig_prog;
1823 if (tmp != prog) {
1824 tmp_blinded = true;
1825 prog = tmp;
1826 }
1827
1828 jit_data = prog->aux->jit_data;
1829 if (!jit_data) {
1830 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1831 if (!jit_data) {
1832 prog = orig_prog;
1833 goto out;
1834 }
1835 prog->aux->jit_data = jit_data;
1836 }
1837 addrs = jit_data->addrs;
1838 if (addrs) {
1839 ctx = jit_data->ctx;
1840 oldproglen = jit_data->proglen;
1841 image = jit_data->image;
1842 header = jit_data->header;
1843 extra_pass = true;
1844 goto skip_init_addrs;
1845 }
1846 addrs = kmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
1847 if (!addrs) {
1848 prog = orig_prog;
1849 goto out_addrs;
1850 }
1851
1852 /*
1853 * Before first pass, make a rough estimation of addrs[]
1854 * each BPF instruction is translated to less than 64 bytes
1855 */
1856 for (proglen = 0, i = 0; i <= prog->len; i++) {
1857 proglen += 64;
1858 addrs[i] = proglen;
1859 }
1860 ctx.cleanup_addr = proglen;
1861 skip_init_addrs:
1862
1863 /*
1864 * JITed image shrinks with every pass and the loop iterates
1865 * until the image stops shrinking. Very large BPF programs
1866 * may converge on the last pass. In such case do one more
1867 * pass to emit the final image.
1868 */
1869 for (pass = 0; pass < 20 || image; pass++) {
1870 proglen = do_jit(prog, addrs, image, oldproglen, &ctx);
1871 if (proglen <= 0) {
1872 out_image:
1873 image = NULL;
1874 if (header)
1875 bpf_jit_binary_free(header);
1876 prog = orig_prog;
1877 goto out_addrs;
1878 }
1879 if (image) {
1880 if (proglen != oldproglen) {
1881 pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
1882 proglen, oldproglen);
1883 goto out_image;
1884 }
1885 break;
1886 }
1887 if (proglen == oldproglen) {
1888 /*
1889 * The number of entries in extable is the number of BPF_LDX
1890 * insns that access kernel memory via "pointer to BTF type".
1891 * The verifier changed their opcode from LDX|MEM|size
1892 * to LDX|PROBE_MEM|size to make JITing easier.
1893 */
1894 u32 align = __alignof__(struct exception_table_entry);
1895 u32 extable_size = prog->aux->num_exentries *
1896 sizeof(struct exception_table_entry);
1897
1898 /* allocate module memory for x86 insns and extable */
1899 header = bpf_jit_binary_alloc(roundup(proglen, align) + extable_size,
1900 &image, align, jit_fill_hole);
1901 if (!header) {
1902 prog = orig_prog;
1903 goto out_addrs;
1904 }
1905 prog->aux->extable = (void *) image + roundup(proglen, align);
1906 }
1907 oldproglen = proglen;
1908 cond_resched();
1909 }
1910
1911 if (bpf_jit_enable > 1)
1912 bpf_jit_dump(prog->len, proglen, pass + 1, image);
1913
1914 if (image) {
1915 if (!prog->is_func || extra_pass) {
1916 bpf_tail_call_direct_fixup(prog);
1917 bpf_jit_binary_lock_ro(header);
1918 } else {
1919 jit_data->addrs = addrs;
1920 jit_data->ctx = ctx;
1921 jit_data->proglen = proglen;
1922 jit_data->image = image;
1923 jit_data->header = header;
1924 }
1925 prog->bpf_func = (void *)image;
1926 prog->jited = 1;
1927 prog->jited_len = proglen;
1928 } else {
1929 prog = orig_prog;
1930 }
1931
1932 if (!image || !prog->is_func || extra_pass) {
1933 if (image)
1934 bpf_prog_fill_jited_linfo(prog, addrs + 1);
1935 out_addrs:
1936 kfree(addrs);
1937 kfree(jit_data);
1938 prog->aux->jit_data = NULL;
1939 }
1940 out:
1941 if (tmp_blinded)
1942 bpf_jit_prog_release_other(prog, prog == orig_prog ?
1943 tmp : orig_prog);
1944 return prog;
1945 }
A mirror of Linus' kernel repository