Loading...
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}
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
20static 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
50static bool is_imm8(int value)
51{
52 return value <= 127 && value >= -128;
53}
54
55static bool is_simm32(s64 value)
56{
57 return value == (s64)(s32)value;
58}
59
60static 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
72static 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 */
115static 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
132static 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 */
150static 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/*
162 * is_ereg_8l() == true if BPF register 'reg' is mapped to access x86-64
163 * lower 8-bit registers dil,sil,bpl,spl,r8b..r15b, which need extra byte
164 * of encoding. al,cl,dl,bl have simpler encoding.
165 */
166static bool is_ereg_8l(u32 reg)
167{
168 return is_ereg(reg) ||
169 (1 << reg) & (BIT(BPF_REG_1) |
170 BIT(BPF_REG_2) |
171 BIT(BPF_REG_FP));
172}
173
174static bool is_axreg(u32 reg)
175{
176 return reg == BPF_REG_0;
177}
178
179/* Add modifiers if 'reg' maps to x86-64 registers R8..R15 */
180static u8 add_1mod(u8 byte, u32 reg)
181{
182 if (is_ereg(reg))
183 byte |= 1;
184 return byte;
185}
186
187static u8 add_2mod(u8 byte, u32 r1, u32 r2)
188{
189 if (is_ereg(r1))
190 byte |= 1;
191 if (is_ereg(r2))
192 byte |= 4;
193 return byte;
194}
195
196/* Encode 'dst_reg' register into x86-64 opcode 'byte' */
197static u8 add_1reg(u8 byte, u32 dst_reg)
198{
199 return byte + reg2hex[dst_reg];
200}
201
202/* Encode 'dst_reg' and 'src_reg' registers into x86-64 opcode 'byte' */
203static u8 add_2reg(u8 byte, u32 dst_reg, u32 src_reg)
204{
205 return byte + reg2hex[dst_reg] + (reg2hex[src_reg] << 3);
206}
207
208static void jit_fill_hole(void *area, unsigned int size)
209{
210 /* Fill whole space with INT3 instructions */
211 memset(area, 0xcc, size);
212}
213
214struct jit_context {
215 int cleanup_addr; /* Epilogue code offset */
216};
217
218/* Maximum number of bytes emitted while JITing one eBPF insn */
219#define BPF_MAX_INSN_SIZE 128
220#define BPF_INSN_SAFETY 64
221
222/* Number of bytes emit_patch() needs to generate instructions */
223#define X86_PATCH_SIZE 5
224
225#define PROLOGUE_SIZE 25
226
227/*
228 * Emit x86-64 prologue code for BPF program and check its size.
229 * bpf_tail_call helper will skip it while jumping into another program
230 */
231static void emit_prologue(u8 **pprog, u32 stack_depth, bool ebpf_from_cbpf)
232{
233 u8 *prog = *pprog;
234 int cnt = X86_PATCH_SIZE;
235
236 /* BPF trampoline can be made to work without these nops,
237 * but let's waste 5 bytes for now and optimize later
238 */
239 memcpy(prog, ideal_nops[NOP_ATOMIC5], cnt);
240 prog += cnt;
241 EMIT1(0x55); /* push rbp */
242 EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
243 /* sub rsp, rounded_stack_depth */
244 EMIT3_off32(0x48, 0x81, 0xEC, round_up(stack_depth, 8));
245 EMIT1(0x53); /* push rbx */
246 EMIT2(0x41, 0x55); /* push r13 */
247 EMIT2(0x41, 0x56); /* push r14 */
248 EMIT2(0x41, 0x57); /* push r15 */
249 if (!ebpf_from_cbpf) {
250 /* zero init tail_call_cnt */
251 EMIT2(0x6a, 0x00);
252 BUILD_BUG_ON(cnt != PROLOGUE_SIZE);
253 }
254 *pprog = prog;
255}
256
257static int emit_patch(u8 **pprog, void *func, void *ip, u8 opcode)
258{
259 u8 *prog = *pprog;
260 int cnt = 0;
261 s64 offset;
262
263 offset = func - (ip + X86_PATCH_SIZE);
264 if (!is_simm32(offset)) {
265 pr_err("Target call %p is out of range\n", func);
266 return -ERANGE;
267 }
268 EMIT1_off32(opcode, offset);
269 *pprog = prog;
270 return 0;
271}
272
273static int emit_call(u8 **pprog, void *func, void *ip)
274{
275 return emit_patch(pprog, func, ip, 0xE8);
276}
277
278static int emit_jump(u8 **pprog, void *func, void *ip)
279{
280 return emit_patch(pprog, func, ip, 0xE9);
281}
282
283static int __bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
284 void *old_addr, void *new_addr,
285 const bool text_live)
286{
287 const u8 *nop_insn = ideal_nops[NOP_ATOMIC5];
288 u8 old_insn[X86_PATCH_SIZE];
289 u8 new_insn[X86_PATCH_SIZE];
290 u8 *prog;
291 int ret;
292
293 memcpy(old_insn, nop_insn, X86_PATCH_SIZE);
294 if (old_addr) {
295 prog = old_insn;
296 ret = t == BPF_MOD_CALL ?
297 emit_call(&prog, old_addr, ip) :
298 emit_jump(&prog, old_addr, ip);
299 if (ret)
300 return ret;
301 }
302
303 memcpy(new_insn, nop_insn, X86_PATCH_SIZE);
304 if (new_addr) {
305 prog = new_insn;
306 ret = t == BPF_MOD_CALL ?
307 emit_call(&prog, new_addr, ip) :
308 emit_jump(&prog, new_addr, ip);
309 if (ret)
310 return ret;
311 }
312
313 ret = -EBUSY;
314 mutex_lock(&text_mutex);
315 if (memcmp(ip, old_insn, X86_PATCH_SIZE))
316 goto out;
317 if (memcmp(ip, new_insn, X86_PATCH_SIZE)) {
318 if (text_live)
319 text_poke_bp(ip, new_insn, X86_PATCH_SIZE, NULL);
320 else
321 memcpy(ip, new_insn, X86_PATCH_SIZE);
322 }
323 ret = 0;
324out:
325 mutex_unlock(&text_mutex);
326 return ret;
327}
328
329int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type t,
330 void *old_addr, void *new_addr)
331{
332 if (!is_kernel_text((long)ip) &&
333 !is_bpf_text_address((long)ip))
334 /* BPF poking in modules is not supported */
335 return -EINVAL;
336
337 return __bpf_arch_text_poke(ip, t, old_addr, new_addr, true);
338}
339
340/*
341 * Generate the following code:
342 *
343 * ... bpf_tail_call(void *ctx, struct bpf_array *array, u64 index) ...
344 * if (index >= array->map.max_entries)
345 * goto out;
346 * if (++tail_call_cnt > MAX_TAIL_CALL_CNT)
347 * goto out;
348 * prog = array->ptrs[index];
349 * if (prog == NULL)
350 * goto out;
351 * goto *(prog->bpf_func + prologue_size);
352 * out:
353 */
354static void emit_bpf_tail_call_indirect(u8 **pprog)
355{
356 u8 *prog = *pprog;
357 int label1, label2, label3;
358 int cnt = 0;
359
360 /*
361 * rdi - pointer to ctx
362 * rsi - pointer to bpf_array
363 * rdx - index in bpf_array
364 */
365
366 /*
367 * if (index >= array->map.max_entries)
368 * goto out;
369 */
370 EMIT2(0x89, 0xD2); /* mov edx, edx */
371 EMIT3(0x39, 0x56, /* cmp dword ptr [rsi + 16], edx */
372 offsetof(struct bpf_array, map.max_entries));
373#define OFFSET1 (41 + RETPOLINE_RAX_BPF_JIT_SIZE) /* Number of bytes to jump */
374 EMIT2(X86_JBE, OFFSET1); /* jbe out */
375 label1 = cnt;
376
377 /*
378 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
379 * goto out;
380 */
381 EMIT2_off32(0x8B, 0x85, -36 - MAX_BPF_STACK); /* mov eax, dword ptr [rbp - 548] */
382 EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */
383#define OFFSET2 (30 + RETPOLINE_RAX_BPF_JIT_SIZE)
384 EMIT2(X86_JA, OFFSET2); /* ja out */
385 label2 = cnt;
386 EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */
387 EMIT2_off32(0x89, 0x85, -36 - MAX_BPF_STACK); /* mov dword ptr [rbp -548], eax */
388
389 /* prog = array->ptrs[index]; */
390 EMIT4_off32(0x48, 0x8B, 0x84, 0xD6, /* mov rax, [rsi + rdx * 8 + offsetof(...)] */
391 offsetof(struct bpf_array, ptrs));
392
393 /*
394 * if (prog == NULL)
395 * goto out;
396 */
397 EMIT3(0x48, 0x85, 0xC0); /* test rax,rax */
398#define OFFSET3 (8 + RETPOLINE_RAX_BPF_JIT_SIZE)
399 EMIT2(X86_JE, OFFSET3); /* je out */
400 label3 = cnt;
401
402 /* goto *(prog->bpf_func + prologue_size); */
403 EMIT4(0x48, 0x8B, 0x40, /* mov rax, qword ptr [rax + 32] */
404 offsetof(struct bpf_prog, bpf_func));
405 EMIT4(0x48, 0x83, 0xC0, PROLOGUE_SIZE); /* add rax, prologue_size */
406
407 /*
408 * Wow we're ready to jump into next BPF program
409 * rdi == ctx (1st arg)
410 * rax == prog->bpf_func + prologue_size
411 */
412 RETPOLINE_RAX_BPF_JIT();
413
414 /* out: */
415 BUILD_BUG_ON(cnt - label1 != OFFSET1);
416 BUILD_BUG_ON(cnt - label2 != OFFSET2);
417 BUILD_BUG_ON(cnt - label3 != OFFSET3);
418 *pprog = prog;
419}
420
421static void emit_bpf_tail_call_direct(struct bpf_jit_poke_descriptor *poke,
422 u8 **pprog, int addr, u8 *image)
423{
424 u8 *prog = *pprog;
425 int cnt = 0;
426
427 /*
428 * if (tail_call_cnt > MAX_TAIL_CALL_CNT)
429 * goto out;
430 */
431 EMIT2_off32(0x8B, 0x85, -36 - MAX_BPF_STACK); /* mov eax, dword ptr [rbp - 548] */
432 EMIT3(0x83, 0xF8, MAX_TAIL_CALL_CNT); /* cmp eax, MAX_TAIL_CALL_CNT */
433 EMIT2(X86_JA, 14); /* ja out */
434 EMIT3(0x83, 0xC0, 0x01); /* add eax, 1 */
435 EMIT2_off32(0x89, 0x85, -36 - MAX_BPF_STACK); /* mov dword ptr [rbp -548], eax */
436
437 poke->ip = image + (addr - X86_PATCH_SIZE);
438 poke->adj_off = PROLOGUE_SIZE;
439
440 memcpy(prog, ideal_nops[NOP_ATOMIC5], X86_PATCH_SIZE);
441 prog += X86_PATCH_SIZE;
442 /* out: */
443
444 *pprog = prog;
445}
446
447static void bpf_tail_call_direct_fixup(struct bpf_prog *prog)
448{
449 struct bpf_jit_poke_descriptor *poke;
450 struct bpf_array *array;
451 struct bpf_prog *target;
452 int i, ret;
453
454 for (i = 0; i < prog->aux->size_poke_tab; i++) {
455 poke = &prog->aux->poke_tab[i];
456 WARN_ON_ONCE(READ_ONCE(poke->ip_stable));
457
458 if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
459 continue;
460
461 array = container_of(poke->tail_call.map, struct bpf_array, map);
462 mutex_lock(&array->aux->poke_mutex);
463 target = array->ptrs[poke->tail_call.key];
464 if (target) {
465 /* Plain memcpy is used when image is not live yet
466 * and still not locked as read-only. Once poke
467 * location is active (poke->ip_stable), any parallel
468 * bpf_arch_text_poke() might occur still on the
469 * read-write image until we finally locked it as
470 * read-only. Both modifications on the given image
471 * are under text_mutex to avoid interference.
472 */
473 ret = __bpf_arch_text_poke(poke->ip, BPF_MOD_JUMP, NULL,
474 (u8 *)target->bpf_func +
475 poke->adj_off, false);
476 BUG_ON(ret < 0);
477 }
478 WRITE_ONCE(poke->ip_stable, true);
479 mutex_unlock(&array->aux->poke_mutex);
480 }
481}
482
483static void emit_mov_imm32(u8 **pprog, bool sign_propagate,
484 u32 dst_reg, const u32 imm32)
485{
486 u8 *prog = *pprog;
487 u8 b1, b2, b3;
488 int cnt = 0;
489
490 /*
491 * Optimization: if imm32 is positive, use 'mov %eax, imm32'
492 * (which zero-extends imm32) to save 2 bytes.
493 */
494 if (sign_propagate && (s32)imm32 < 0) {
495 /* 'mov %rax, imm32' sign extends imm32 */
496 b1 = add_1mod(0x48, dst_reg);
497 b2 = 0xC7;
498 b3 = 0xC0;
499 EMIT3_off32(b1, b2, add_1reg(b3, dst_reg), imm32);
500 goto done;
501 }
502
503 /*
504 * Optimization: if imm32 is zero, use 'xor %eax, %eax'
505 * to save 3 bytes.
506 */
507 if (imm32 == 0) {
508 if (is_ereg(dst_reg))
509 EMIT1(add_2mod(0x40, dst_reg, dst_reg));
510 b2 = 0x31; /* xor */
511 b3 = 0xC0;
512 EMIT2(b2, add_2reg(b3, dst_reg, dst_reg));
513 goto done;
514 }
515
516 /* mov %eax, imm32 */
517 if (is_ereg(dst_reg))
518 EMIT1(add_1mod(0x40, dst_reg));
519 EMIT1_off32(add_1reg(0xB8, dst_reg), imm32);
520done:
521 *pprog = prog;
522}
523
524static void emit_mov_imm64(u8 **pprog, u32 dst_reg,
525 const u32 imm32_hi, const u32 imm32_lo)
526{
527 u8 *prog = *pprog;
528 int cnt = 0;
529
530 if (is_uimm32(((u64)imm32_hi << 32) | (u32)imm32_lo)) {
531 /*
532 * For emitting plain u32, where sign bit must not be
533 * propagated LLVM tends to load imm64 over mov32
534 * directly, so save couple of bytes by just doing
535 * 'mov %eax, imm32' instead.
536 */
537 emit_mov_imm32(&prog, false, dst_reg, imm32_lo);
538 } else {
539 /* movabsq %rax, imm64 */
540 EMIT2(add_1mod(0x48, dst_reg), add_1reg(0xB8, dst_reg));
541 EMIT(imm32_lo, 4);
542 EMIT(imm32_hi, 4);
543 }
544
545 *pprog = prog;
546}
547
548static void emit_mov_reg(u8 **pprog, bool is64, u32 dst_reg, u32 src_reg)
549{
550 u8 *prog = *pprog;
551 int cnt = 0;
552
553 if (is64) {
554 /* mov dst, src */
555 EMIT_mov(dst_reg, src_reg);
556 } else {
557 /* mov32 dst, src */
558 if (is_ereg(dst_reg) || is_ereg(src_reg))
559 EMIT1(add_2mod(0x40, dst_reg, src_reg));
560 EMIT2(0x89, add_2reg(0xC0, dst_reg, src_reg));
561 }
562
563 *pprog = prog;
564}
565
566/* LDX: dst_reg = *(u8*)(src_reg + off) */
567static void emit_ldx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
568{
569 u8 *prog = *pprog;
570 int cnt = 0;
571
572 switch (size) {
573 case BPF_B:
574 /* Emit 'movzx rax, byte ptr [rax + off]' */
575 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB6);
576 break;
577 case BPF_H:
578 /* Emit 'movzx rax, word ptr [rax + off]' */
579 EMIT3(add_2mod(0x48, src_reg, dst_reg), 0x0F, 0xB7);
580 break;
581 case BPF_W:
582 /* Emit 'mov eax, dword ptr [rax+0x14]' */
583 if (is_ereg(dst_reg) || is_ereg(src_reg))
584 EMIT2(add_2mod(0x40, src_reg, dst_reg), 0x8B);
585 else
586 EMIT1(0x8B);
587 break;
588 case BPF_DW:
589 /* Emit 'mov rax, qword ptr [rax+0x14]' */
590 EMIT2(add_2mod(0x48, src_reg, dst_reg), 0x8B);
591 break;
592 }
593 /*
594 * If insn->off == 0 we can save one extra byte, but
595 * special case of x86 R13 which always needs an offset
596 * is not worth the hassle
597 */
598 if (is_imm8(off))
599 EMIT2(add_2reg(0x40, src_reg, dst_reg), off);
600 else
601 EMIT1_off32(add_2reg(0x80, src_reg, dst_reg), off);
602 *pprog = prog;
603}
604
605/* STX: *(u8*)(dst_reg + off) = src_reg */
606static void emit_stx(u8 **pprog, u32 size, u32 dst_reg, u32 src_reg, int off)
607{
608 u8 *prog = *pprog;
609 int cnt = 0;
610
611 switch (size) {
612 case BPF_B:
613 /* Emit 'mov byte ptr [rax + off], al' */
614 if (is_ereg(dst_reg) || is_ereg_8l(src_reg))
615 /* Add extra byte for eregs or SIL,DIL,BPL in src_reg */
616 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x88);
617 else
618 EMIT1(0x88);
619 break;
620 case BPF_H:
621 if (is_ereg(dst_reg) || is_ereg(src_reg))
622 EMIT3(0x66, add_2mod(0x40, dst_reg, src_reg), 0x89);
623 else
624 EMIT2(0x66, 0x89);
625 break;
626 case BPF_W:
627 if (is_ereg(dst_reg) || is_ereg(src_reg))
628 EMIT2(add_2mod(0x40, dst_reg, src_reg), 0x89);
629 else
630 EMIT1(0x89);
631 break;
632 case BPF_DW:
633 EMIT2(add_2mod(0x48, dst_reg, src_reg), 0x89);
634 break;
635 }
636 if (is_imm8(off))
637 EMIT2(add_2reg(0x40, dst_reg, src_reg), off);
638 else
639 EMIT1_off32(add_2reg(0x80, dst_reg, src_reg), off);
640 *pprog = prog;
641}
642
643static bool ex_handler_bpf(const struct exception_table_entry *x,
644 struct pt_regs *regs, int trapnr,
645 unsigned long error_code, unsigned long fault_addr)
646{
647 u32 reg = x->fixup >> 8;
648
649 /* jump over faulting load and clear dest register */
650 *(unsigned long *)((void *)regs + reg) = 0;
651 regs->ip += x->fixup & 0xff;
652 return true;
653}
654
655static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image,
656 int oldproglen, struct jit_context *ctx)
657{
658 struct bpf_insn *insn = bpf_prog->insnsi;
659 int insn_cnt = bpf_prog->len;
660 bool seen_exit = false;
661 u8 temp[BPF_MAX_INSN_SIZE + BPF_INSN_SAFETY];
662 int i, cnt = 0, excnt = 0;
663 int proglen = 0;
664 u8 *prog = temp;
665
666 emit_prologue(&prog, bpf_prog->aux->stack_depth,
667 bpf_prog_was_classic(bpf_prog));
668 addrs[0] = prog - temp;
669
670 for (i = 1; i <= insn_cnt; i++, insn++) {
671 const s32 imm32 = insn->imm;
672 u32 dst_reg = insn->dst_reg;
673 u32 src_reg = insn->src_reg;
674 u8 b2 = 0, b3 = 0;
675 s64 jmp_offset;
676 u8 jmp_cond;
677 int ilen;
678 u8 *func;
679
680 switch (insn->code) {
681 /* ALU */
682 case BPF_ALU | BPF_ADD | BPF_X:
683 case BPF_ALU | BPF_SUB | BPF_X:
684 case BPF_ALU | BPF_AND | BPF_X:
685 case BPF_ALU | BPF_OR | BPF_X:
686 case BPF_ALU | BPF_XOR | BPF_X:
687 case BPF_ALU64 | BPF_ADD | BPF_X:
688 case BPF_ALU64 | BPF_SUB | BPF_X:
689 case BPF_ALU64 | BPF_AND | BPF_X:
690 case BPF_ALU64 | BPF_OR | BPF_X:
691 case BPF_ALU64 | BPF_XOR | BPF_X:
692 switch (BPF_OP(insn->code)) {
693 case BPF_ADD: b2 = 0x01; break;
694 case BPF_SUB: b2 = 0x29; break;
695 case BPF_AND: b2 = 0x21; break;
696 case BPF_OR: b2 = 0x09; break;
697 case BPF_XOR: b2 = 0x31; break;
698 }
699 if (BPF_CLASS(insn->code) == BPF_ALU64)
700 EMIT1(add_2mod(0x48, dst_reg, src_reg));
701 else if (is_ereg(dst_reg) || is_ereg(src_reg))
702 EMIT1(add_2mod(0x40, dst_reg, src_reg));
703 EMIT2(b2, add_2reg(0xC0, dst_reg, src_reg));
704 break;
705
706 case BPF_ALU64 | BPF_MOV | BPF_X:
707 case BPF_ALU | BPF_MOV | BPF_X:
708 emit_mov_reg(&prog,
709 BPF_CLASS(insn->code) == BPF_ALU64,
710 dst_reg, src_reg);
711 break;
712
713 /* neg dst */
714 case BPF_ALU | BPF_NEG:
715 case BPF_ALU64 | BPF_NEG:
716 if (BPF_CLASS(insn->code) == BPF_ALU64)
717 EMIT1(add_1mod(0x48, dst_reg));
718 else if (is_ereg(dst_reg))
719 EMIT1(add_1mod(0x40, dst_reg));
720 EMIT2(0xF7, add_1reg(0xD8, dst_reg));
721 break;
722
723 case BPF_ALU | BPF_ADD | BPF_K:
724 case BPF_ALU | BPF_SUB | BPF_K:
725 case BPF_ALU | BPF_AND | BPF_K:
726 case BPF_ALU | BPF_OR | BPF_K:
727 case BPF_ALU | BPF_XOR | BPF_K:
728 case BPF_ALU64 | BPF_ADD | BPF_K:
729 case BPF_ALU64 | BPF_SUB | BPF_K:
730 case BPF_ALU64 | BPF_AND | BPF_K:
731 case BPF_ALU64 | BPF_OR | BPF_K:
732 case BPF_ALU64 | BPF_XOR | BPF_K:
733 if (BPF_CLASS(insn->code) == BPF_ALU64)
734 EMIT1(add_1mod(0x48, dst_reg));
735 else if (is_ereg(dst_reg))
736 EMIT1(add_1mod(0x40, dst_reg));
737
738 /*
739 * b3 holds 'normal' opcode, b2 short form only valid
740 * in case dst is eax/rax.
741 */
742 switch (BPF_OP(insn->code)) {
743 case BPF_ADD:
744 b3 = 0xC0;
745 b2 = 0x05;
746 break;
747 case BPF_SUB:
748 b3 = 0xE8;
749 b2 = 0x2D;
750 break;
751 case BPF_AND:
752 b3 = 0xE0;
753 b2 = 0x25;
754 break;
755 case BPF_OR:
756 b3 = 0xC8;
757 b2 = 0x0D;
758 break;
759 case BPF_XOR:
760 b3 = 0xF0;
761 b2 = 0x35;
762 break;
763 }
764
765 if (is_imm8(imm32))
766 EMIT3(0x83, add_1reg(b3, dst_reg), imm32);
767 else if (is_axreg(dst_reg))
768 EMIT1_off32(b2, imm32);
769 else
770 EMIT2_off32(0x81, add_1reg(b3, dst_reg), imm32);
771 break;
772
773 case BPF_ALU64 | BPF_MOV | BPF_K:
774 case BPF_ALU | BPF_MOV | BPF_K:
775 emit_mov_imm32(&prog, BPF_CLASS(insn->code) == BPF_ALU64,
776 dst_reg, imm32);
777 break;
778
779 case BPF_LD | BPF_IMM | BPF_DW:
780 emit_mov_imm64(&prog, dst_reg, insn[1].imm, insn[0].imm);
781 insn++;
782 i++;
783 break;
784
785 /* dst %= src, dst /= src, dst %= imm32, dst /= imm32 */
786 case BPF_ALU | BPF_MOD | BPF_X:
787 case BPF_ALU | BPF_DIV | BPF_X:
788 case BPF_ALU | BPF_MOD | BPF_K:
789 case BPF_ALU | BPF_DIV | BPF_K:
790 case BPF_ALU64 | BPF_MOD | BPF_X:
791 case BPF_ALU64 | BPF_DIV | BPF_X:
792 case BPF_ALU64 | BPF_MOD | BPF_K:
793 case BPF_ALU64 | BPF_DIV | BPF_K:
794 EMIT1(0x50); /* push rax */
795 EMIT1(0x52); /* push rdx */
796
797 if (BPF_SRC(insn->code) == BPF_X)
798 /* mov r11, src_reg */
799 EMIT_mov(AUX_REG, src_reg);
800 else
801 /* mov r11, imm32 */
802 EMIT3_off32(0x49, 0xC7, 0xC3, imm32);
803
804 /* mov rax, dst_reg */
805 EMIT_mov(BPF_REG_0, dst_reg);
806
807 /*
808 * xor edx, edx
809 * equivalent to 'xor rdx, rdx', but one byte less
810 */
811 EMIT2(0x31, 0xd2);
812
813 if (BPF_CLASS(insn->code) == BPF_ALU64)
814 /* div r11 */
815 EMIT3(0x49, 0xF7, 0xF3);
816 else
817 /* div r11d */
818 EMIT3(0x41, 0xF7, 0xF3);
819
820 if (BPF_OP(insn->code) == BPF_MOD)
821 /* mov r11, rdx */
822 EMIT3(0x49, 0x89, 0xD3);
823 else
824 /* mov r11, rax */
825 EMIT3(0x49, 0x89, 0xC3);
826
827 EMIT1(0x5A); /* pop rdx */
828 EMIT1(0x58); /* pop rax */
829
830 /* mov dst_reg, r11 */
831 EMIT_mov(dst_reg, AUX_REG);
832 break;
833
834 case BPF_ALU | BPF_MUL | BPF_K:
835 case BPF_ALU | BPF_MUL | BPF_X:
836 case BPF_ALU64 | BPF_MUL | BPF_K:
837 case BPF_ALU64 | BPF_MUL | BPF_X:
838 {
839 bool is64 = BPF_CLASS(insn->code) == BPF_ALU64;
840
841 if (dst_reg != BPF_REG_0)
842 EMIT1(0x50); /* push rax */
843 if (dst_reg != BPF_REG_3)
844 EMIT1(0x52); /* push rdx */
845
846 /* mov r11, dst_reg */
847 EMIT_mov(AUX_REG, dst_reg);
848
849 if (BPF_SRC(insn->code) == BPF_X)
850 emit_mov_reg(&prog, is64, BPF_REG_0, src_reg);
851 else
852 emit_mov_imm32(&prog, is64, BPF_REG_0, imm32);
853
854 if (is64)
855 EMIT1(add_1mod(0x48, AUX_REG));
856 else if (is_ereg(AUX_REG))
857 EMIT1(add_1mod(0x40, AUX_REG));
858 /* mul(q) r11 */
859 EMIT2(0xF7, add_1reg(0xE0, AUX_REG));
860
861 if (dst_reg != BPF_REG_3)
862 EMIT1(0x5A); /* pop rdx */
863 if (dst_reg != BPF_REG_0) {
864 /* mov dst_reg, rax */
865 EMIT_mov(dst_reg, BPF_REG_0);
866 EMIT1(0x58); /* pop rax */
867 }
868 break;
869 }
870 /* Shifts */
871 case BPF_ALU | BPF_LSH | BPF_K:
872 case BPF_ALU | BPF_RSH | BPF_K:
873 case BPF_ALU | BPF_ARSH | BPF_K:
874 case BPF_ALU64 | BPF_LSH | BPF_K:
875 case BPF_ALU64 | BPF_RSH | BPF_K:
876 case BPF_ALU64 | BPF_ARSH | BPF_K:
877 if (BPF_CLASS(insn->code) == BPF_ALU64)
878 EMIT1(add_1mod(0x48, dst_reg));
879 else if (is_ereg(dst_reg))
880 EMIT1(add_1mod(0x40, dst_reg));
881
882 switch (BPF_OP(insn->code)) {
883 case BPF_LSH: b3 = 0xE0; break;
884 case BPF_RSH: b3 = 0xE8; break;
885 case BPF_ARSH: b3 = 0xF8; break;
886 }
887
888 if (imm32 == 1)
889 EMIT2(0xD1, add_1reg(b3, dst_reg));
890 else
891 EMIT3(0xC1, add_1reg(b3, dst_reg), imm32);
892 break;
893
894 case BPF_ALU | BPF_LSH | BPF_X:
895 case BPF_ALU | BPF_RSH | BPF_X:
896 case BPF_ALU | BPF_ARSH | BPF_X:
897 case BPF_ALU64 | BPF_LSH | BPF_X:
898 case BPF_ALU64 | BPF_RSH | BPF_X:
899 case BPF_ALU64 | BPF_ARSH | BPF_X:
900
901 /* Check for bad case when dst_reg == rcx */
902 if (dst_reg == BPF_REG_4) {
903 /* mov r11, dst_reg */
904 EMIT_mov(AUX_REG, dst_reg);
905 dst_reg = AUX_REG;
906 }
907
908 if (src_reg != BPF_REG_4) { /* common case */
909 EMIT1(0x51); /* push rcx */
910
911 /* mov rcx, src_reg */
912 EMIT_mov(BPF_REG_4, src_reg);
913 }
914
915 /* shl %rax, %cl | shr %rax, %cl | sar %rax, %cl */
916 if (BPF_CLASS(insn->code) == BPF_ALU64)
917 EMIT1(add_1mod(0x48, dst_reg));
918 else if (is_ereg(dst_reg))
919 EMIT1(add_1mod(0x40, dst_reg));
920
921 switch (BPF_OP(insn->code)) {
922 case BPF_LSH: b3 = 0xE0; break;
923 case BPF_RSH: b3 = 0xE8; break;
924 case BPF_ARSH: b3 = 0xF8; break;
925 }
926 EMIT2(0xD3, add_1reg(b3, dst_reg));
927
928 if (src_reg != BPF_REG_4)
929 EMIT1(0x59); /* pop rcx */
930
931 if (insn->dst_reg == BPF_REG_4)
932 /* mov dst_reg, r11 */
933 EMIT_mov(insn->dst_reg, AUX_REG);
934 break;
935
936 case BPF_ALU | BPF_END | BPF_FROM_BE:
937 switch (imm32) {
938 case 16:
939 /* Emit 'ror %ax, 8' to swap lower 2 bytes */
940 EMIT1(0x66);
941 if (is_ereg(dst_reg))
942 EMIT1(0x41);
943 EMIT3(0xC1, add_1reg(0xC8, dst_reg), 8);
944
945 /* Emit 'movzwl eax, ax' */
946 if (is_ereg(dst_reg))
947 EMIT3(0x45, 0x0F, 0xB7);
948 else
949 EMIT2(0x0F, 0xB7);
950 EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
951 break;
952 case 32:
953 /* Emit 'bswap eax' to swap lower 4 bytes */
954 if (is_ereg(dst_reg))
955 EMIT2(0x41, 0x0F);
956 else
957 EMIT1(0x0F);
958 EMIT1(add_1reg(0xC8, dst_reg));
959 break;
960 case 64:
961 /* Emit 'bswap rax' to swap 8 bytes */
962 EMIT3(add_1mod(0x48, dst_reg), 0x0F,
963 add_1reg(0xC8, dst_reg));
964 break;
965 }
966 break;
967
968 case BPF_ALU | BPF_END | BPF_FROM_LE:
969 switch (imm32) {
970 case 16:
971 /*
972 * Emit 'movzwl eax, ax' to zero extend 16-bit
973 * into 64 bit
974 */
975 if (is_ereg(dst_reg))
976 EMIT3(0x45, 0x0F, 0xB7);
977 else
978 EMIT2(0x0F, 0xB7);
979 EMIT1(add_2reg(0xC0, dst_reg, dst_reg));
980 break;
981 case 32:
982 /* Emit 'mov eax, eax' to clear upper 32-bits */
983 if (is_ereg(dst_reg))
984 EMIT1(0x45);
985 EMIT2(0x89, add_2reg(0xC0, dst_reg, dst_reg));
986 break;
987 case 64:
988 /* nop */
989 break;
990 }
991 break;
992
993 /* ST: *(u8*)(dst_reg + off) = imm */
994 case BPF_ST | BPF_MEM | BPF_B:
995 if (is_ereg(dst_reg))
996 EMIT2(0x41, 0xC6);
997 else
998 EMIT1(0xC6);
999 goto st;
1000 case BPF_ST | BPF_MEM | BPF_H:
1001 if (is_ereg(dst_reg))
1002 EMIT3(0x66, 0x41, 0xC7);
1003 else
1004 EMIT2(0x66, 0xC7);
1005 goto st;
1006 case BPF_ST | BPF_MEM | BPF_W:
1007 if (is_ereg(dst_reg))
1008 EMIT2(0x41, 0xC7);
1009 else
1010 EMIT1(0xC7);
1011 goto st;
1012 case BPF_ST | BPF_MEM | BPF_DW:
1013 EMIT2(add_1mod(0x48, dst_reg), 0xC7);
1014
1015st: if (is_imm8(insn->off))
1016 EMIT2(add_1reg(0x40, dst_reg), insn->off);
1017 else
1018 EMIT1_off32(add_1reg(0x80, dst_reg), insn->off);
1019
1020 EMIT(imm32, bpf_size_to_x86_bytes(BPF_SIZE(insn->code)));
1021 break;
1022
1023 /* STX: *(u8*)(dst_reg + off) = src_reg */
1024 case BPF_STX | BPF_MEM | BPF_B:
1025 case BPF_STX | BPF_MEM | BPF_H:
1026 case BPF_STX | BPF_MEM | BPF_W:
1027 case BPF_STX | BPF_MEM | BPF_DW:
1028 emit_stx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1029 break;
1030
1031 /* LDX: dst_reg = *(u8*)(src_reg + off) */
1032 case BPF_LDX | BPF_MEM | BPF_B:
1033 case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1034 case BPF_LDX | BPF_MEM | BPF_H:
1035 case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1036 case BPF_LDX | BPF_MEM | BPF_W:
1037 case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1038 case BPF_LDX | BPF_MEM | BPF_DW:
1039 case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1040 emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
1041 if (BPF_MODE(insn->code) == BPF_PROBE_MEM) {
1042 struct exception_table_entry *ex;
1043 u8 *_insn = image + proglen;
1044 s64 delta;
1045
1046 if (!bpf_prog->aux->extable)
1047 break;
1048
1049 if (excnt >= bpf_prog->aux->num_exentries) {
1050 pr_err("ex gen bug\n");
1051 return -EFAULT;
1052 }
1053 ex = &bpf_prog->aux->extable[excnt++];
1054
1055 delta = _insn - (u8 *)&ex->insn;
1056 if (!is_simm32(delta)) {
1057 pr_err("extable->insn doesn't fit into 32-bit\n");
1058 return -EFAULT;
1059 }
1060 ex->insn = delta;
1061
1062 delta = (u8 *)ex_handler_bpf - (u8 *)&ex->handler;
1063 if (!is_simm32(delta)) {
1064 pr_err("extable->handler doesn't fit into 32-bit\n");
1065 return -EFAULT;
1066 }
1067 ex->handler = delta;
1068
1069 if (dst_reg > BPF_REG_9) {
1070 pr_err("verifier error\n");
1071 return -EFAULT;
1072 }
1073 /*
1074 * Compute size of x86 insn and its target dest x86 register.
1075 * ex_handler_bpf() will use lower 8 bits to adjust
1076 * pt_regs->ip to jump over this x86 instruction
1077 * and upper bits to figure out which pt_regs to zero out.
1078 * End result: x86 insn "mov rbx, qword ptr [rax+0x14]"
1079 * of 4 bytes will be ignored and rbx will be zero inited.
1080 */
1081 ex->fixup = (prog - temp) | (reg2pt_regs[dst_reg] << 8);
1082 }
1083 break;
1084
1085 /* STX XADD: lock *(u32*)(dst_reg + off) += src_reg */
1086 case BPF_STX | BPF_XADD | BPF_W:
1087 /* Emit 'lock add dword ptr [rax + off], eax' */
1088 if (is_ereg(dst_reg) || is_ereg(src_reg))
1089 EMIT3(0xF0, add_2mod(0x40, dst_reg, src_reg), 0x01);
1090 else
1091 EMIT2(0xF0, 0x01);
1092 goto xadd;
1093 case BPF_STX | BPF_XADD | BPF_DW:
1094 EMIT3(0xF0, add_2mod(0x48, dst_reg, src_reg), 0x01);
1095xadd: if (is_imm8(insn->off))
1096 EMIT2(add_2reg(0x40, dst_reg, src_reg), insn->off);
1097 else
1098 EMIT1_off32(add_2reg(0x80, dst_reg, src_reg),
1099 insn->off);
1100 break;
1101
1102 /* call */
1103 case BPF_JMP | BPF_CALL:
1104 func = (u8 *) __bpf_call_base + imm32;
1105 if (!imm32 || emit_call(&prog, func, image + addrs[i - 1]))
1106 return -EINVAL;
1107 break;
1108
1109 case BPF_JMP | BPF_TAIL_CALL:
1110 if (imm32)
1111 emit_bpf_tail_call_direct(&bpf_prog->aux->poke_tab[imm32 - 1],
1112 &prog, addrs[i], image);
1113 else
1114 emit_bpf_tail_call_indirect(&prog);
1115 break;
1116
1117 /* cond jump */
1118 case BPF_JMP | BPF_JEQ | BPF_X:
1119 case BPF_JMP | BPF_JNE | BPF_X:
1120 case BPF_JMP | BPF_JGT | BPF_X:
1121 case BPF_JMP | BPF_JLT | BPF_X:
1122 case BPF_JMP | BPF_JGE | BPF_X:
1123 case BPF_JMP | BPF_JLE | BPF_X:
1124 case BPF_JMP | BPF_JSGT | BPF_X:
1125 case BPF_JMP | BPF_JSLT | BPF_X:
1126 case BPF_JMP | BPF_JSGE | BPF_X:
1127 case BPF_JMP | BPF_JSLE | BPF_X:
1128 case BPF_JMP32 | BPF_JEQ | BPF_X:
1129 case BPF_JMP32 | BPF_JNE | BPF_X:
1130 case BPF_JMP32 | BPF_JGT | BPF_X:
1131 case BPF_JMP32 | BPF_JLT | BPF_X:
1132 case BPF_JMP32 | BPF_JGE | BPF_X:
1133 case BPF_JMP32 | BPF_JLE | BPF_X:
1134 case BPF_JMP32 | BPF_JSGT | BPF_X:
1135 case BPF_JMP32 | BPF_JSLT | BPF_X:
1136 case BPF_JMP32 | BPF_JSGE | BPF_X:
1137 case BPF_JMP32 | BPF_JSLE | BPF_X:
1138 /* cmp dst_reg, src_reg */
1139 if (BPF_CLASS(insn->code) == BPF_JMP)
1140 EMIT1(add_2mod(0x48, dst_reg, src_reg));
1141 else if (is_ereg(dst_reg) || is_ereg(src_reg))
1142 EMIT1(add_2mod(0x40, dst_reg, src_reg));
1143 EMIT2(0x39, add_2reg(0xC0, dst_reg, src_reg));
1144 goto emit_cond_jmp;
1145
1146 case BPF_JMP | BPF_JSET | BPF_X:
1147 case BPF_JMP32 | BPF_JSET | BPF_X:
1148 /* test dst_reg, src_reg */
1149 if (BPF_CLASS(insn->code) == BPF_JMP)
1150 EMIT1(add_2mod(0x48, dst_reg, src_reg));
1151 else if (is_ereg(dst_reg) || is_ereg(src_reg))
1152 EMIT1(add_2mod(0x40, dst_reg, src_reg));
1153 EMIT2(0x85, add_2reg(0xC0, dst_reg, src_reg));
1154 goto emit_cond_jmp;
1155
1156 case BPF_JMP | BPF_JSET | BPF_K:
1157 case BPF_JMP32 | BPF_JSET | BPF_K:
1158 /* test dst_reg, imm32 */
1159 if (BPF_CLASS(insn->code) == BPF_JMP)
1160 EMIT1(add_1mod(0x48, dst_reg));
1161 else if (is_ereg(dst_reg))
1162 EMIT1(add_1mod(0x40, dst_reg));
1163 EMIT2_off32(0xF7, add_1reg(0xC0, dst_reg), imm32);
1164 goto emit_cond_jmp;
1165
1166 case BPF_JMP | BPF_JEQ | BPF_K:
1167 case BPF_JMP | BPF_JNE | BPF_K:
1168 case BPF_JMP | BPF_JGT | BPF_K:
1169 case BPF_JMP | BPF_JLT | BPF_K:
1170 case BPF_JMP | BPF_JGE | BPF_K:
1171 case BPF_JMP | BPF_JLE | BPF_K:
1172 case BPF_JMP | BPF_JSGT | BPF_K:
1173 case BPF_JMP | BPF_JSLT | BPF_K:
1174 case BPF_JMP | BPF_JSGE | BPF_K:
1175 case BPF_JMP | BPF_JSLE | BPF_K:
1176 case BPF_JMP32 | BPF_JEQ | BPF_K:
1177 case BPF_JMP32 | BPF_JNE | BPF_K:
1178 case BPF_JMP32 | BPF_JGT | BPF_K:
1179 case BPF_JMP32 | BPF_JLT | BPF_K:
1180 case BPF_JMP32 | BPF_JGE | BPF_K:
1181 case BPF_JMP32 | BPF_JLE | BPF_K:
1182 case BPF_JMP32 | BPF_JSGT | BPF_K:
1183 case BPF_JMP32 | BPF_JSLT | BPF_K:
1184 case BPF_JMP32 | BPF_JSGE | BPF_K:
1185 case BPF_JMP32 | BPF_JSLE | BPF_K:
1186 /* test dst_reg, dst_reg to save one extra byte */
1187 if (imm32 == 0) {
1188 if (BPF_CLASS(insn->code) == BPF_JMP)
1189 EMIT1(add_2mod(0x48, dst_reg, dst_reg));
1190 else if (is_ereg(dst_reg))
1191 EMIT1(add_2mod(0x40, dst_reg, dst_reg));
1192 EMIT2(0x85, add_2reg(0xC0, dst_reg, dst_reg));
1193 goto emit_cond_jmp;
1194 }
1195
1196 /* cmp dst_reg, imm8/32 */
1197 if (BPF_CLASS(insn->code) == BPF_JMP)
1198 EMIT1(add_1mod(0x48, dst_reg));
1199 else if (is_ereg(dst_reg))
1200 EMIT1(add_1mod(0x40, dst_reg));
1201
1202 if (is_imm8(imm32))
1203 EMIT3(0x83, add_1reg(0xF8, dst_reg), imm32);
1204 else
1205 EMIT2_off32(0x81, add_1reg(0xF8, dst_reg), imm32);
1206
1207emit_cond_jmp: /* Convert BPF opcode to x86 */
1208 switch (BPF_OP(insn->code)) {
1209 case BPF_JEQ:
1210 jmp_cond = X86_JE;
1211 break;
1212 case BPF_JSET:
1213 case BPF_JNE:
1214 jmp_cond = X86_JNE;
1215 break;
1216 case BPF_JGT:
1217 /* GT is unsigned '>', JA in x86 */
1218 jmp_cond = X86_JA;
1219 break;
1220 case BPF_JLT:
1221 /* LT is unsigned '<', JB in x86 */
1222 jmp_cond = X86_JB;
1223 break;
1224 case BPF_JGE:
1225 /* GE is unsigned '>=', JAE in x86 */
1226 jmp_cond = X86_JAE;
1227 break;
1228 case BPF_JLE:
1229 /* LE is unsigned '<=', JBE in x86 */
1230 jmp_cond = X86_JBE;
1231 break;
1232 case BPF_JSGT:
1233 /* Signed '>', GT in x86 */
1234 jmp_cond = X86_JG;
1235 break;
1236 case BPF_JSLT:
1237 /* Signed '<', LT in x86 */
1238 jmp_cond = X86_JL;
1239 break;
1240 case BPF_JSGE:
1241 /* Signed '>=', GE in x86 */
1242 jmp_cond = X86_JGE;
1243 break;
1244 case BPF_JSLE:
1245 /* Signed '<=', LE in x86 */
1246 jmp_cond = X86_JLE;
1247 break;
1248 default: /* to silence GCC warning */
1249 return -EFAULT;
1250 }
1251 jmp_offset = addrs[i + insn->off] - addrs[i];
1252 if (is_imm8(jmp_offset)) {
1253 EMIT2(jmp_cond, jmp_offset);
1254 } else if (is_simm32(jmp_offset)) {
1255 EMIT2_off32(0x0F, jmp_cond + 0x10, jmp_offset);
1256 } else {
1257 pr_err("cond_jmp gen bug %llx\n", jmp_offset);
1258 return -EFAULT;
1259 }
1260
1261 break;
1262
1263 case BPF_JMP | BPF_JA:
1264 if (insn->off == -1)
1265 /* -1 jmp instructions will always jump
1266 * backwards two bytes. Explicitly handling
1267 * this case avoids wasting too many passes
1268 * when there are long sequences of replaced
1269 * dead code.
1270 */
1271 jmp_offset = -2;
1272 else
1273 jmp_offset = addrs[i + insn->off] - addrs[i];
1274
1275 if (!jmp_offset)
1276 /* Optimize out nop jumps */
1277 break;
1278emit_jmp:
1279 if (is_imm8(jmp_offset)) {
1280 EMIT2(0xEB, jmp_offset);
1281 } else if (is_simm32(jmp_offset)) {
1282 EMIT1_off32(0xE9, jmp_offset);
1283 } else {
1284 pr_err("jmp gen bug %llx\n", jmp_offset);
1285 return -EFAULT;
1286 }
1287 break;
1288
1289 case BPF_JMP | BPF_EXIT:
1290 if (seen_exit) {
1291 jmp_offset = ctx->cleanup_addr - addrs[i];
1292 goto emit_jmp;
1293 }
1294 seen_exit = true;
1295 /* Update cleanup_addr */
1296 ctx->cleanup_addr = proglen;
1297 if (!bpf_prog_was_classic(bpf_prog))
1298 EMIT1(0x5B); /* get rid of tail_call_cnt */
1299 EMIT2(0x41, 0x5F); /* pop r15 */
1300 EMIT2(0x41, 0x5E); /* pop r14 */
1301 EMIT2(0x41, 0x5D); /* pop r13 */
1302 EMIT1(0x5B); /* pop rbx */
1303 EMIT1(0xC9); /* leave */
1304 EMIT1(0xC3); /* ret */
1305 break;
1306
1307 default:
1308 /*
1309 * By design x86-64 JIT should support all BPF instructions.
1310 * This error will be seen if new instruction was added
1311 * to the interpreter, but not to the JIT, or if there is
1312 * junk in bpf_prog.
1313 */
1314 pr_err("bpf_jit: unknown opcode %02x\n", insn->code);
1315 return -EINVAL;
1316 }
1317
1318 ilen = prog - temp;
1319 if (ilen > BPF_MAX_INSN_SIZE) {
1320 pr_err("bpf_jit: fatal insn size error\n");
1321 return -EFAULT;
1322 }
1323
1324 if (image) {
1325 if (unlikely(proglen + ilen > oldproglen)) {
1326 pr_err("bpf_jit: fatal error\n");
1327 return -EFAULT;
1328 }
1329 memcpy(image + proglen, temp, ilen);
1330 }
1331 proglen += ilen;
1332 addrs[i] = proglen;
1333 prog = temp;
1334 }
1335
1336 if (image && excnt != bpf_prog->aux->num_exentries) {
1337 pr_err("extable is not populated\n");
1338 return -EFAULT;
1339 }
1340 return proglen;
1341}
1342
1343static void save_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1344 int stack_size)
1345{
1346 int i;
1347 /* Store function arguments to stack.
1348 * For a function that accepts two pointers the sequence will be:
1349 * mov QWORD PTR [rbp-0x10],rdi
1350 * mov QWORD PTR [rbp-0x8],rsi
1351 */
1352 for (i = 0; i < min(nr_args, 6); i++)
1353 emit_stx(prog, bytes_to_bpf_size(m->arg_size[i]),
1354 BPF_REG_FP,
1355 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1356 -(stack_size - i * 8));
1357}
1358
1359static void restore_regs(const struct btf_func_model *m, u8 **prog, int nr_args,
1360 int stack_size)
1361{
1362 int i;
1363
1364 /* Restore function arguments from stack.
1365 * For a function that accepts two pointers the sequence will be:
1366 * EMIT4(0x48, 0x8B, 0x7D, 0xF0); mov rdi,QWORD PTR [rbp-0x10]
1367 * EMIT4(0x48, 0x8B, 0x75, 0xF8); mov rsi,QWORD PTR [rbp-0x8]
1368 */
1369 for (i = 0; i < min(nr_args, 6); i++)
1370 emit_ldx(prog, bytes_to_bpf_size(m->arg_size[i]),
1371 i == 5 ? X86_REG_R9 : BPF_REG_1 + i,
1372 BPF_REG_FP,
1373 -(stack_size - i * 8));
1374}
1375
1376static int invoke_bpf_prog(const struct btf_func_model *m, u8 **pprog,
1377 struct bpf_prog *p, int stack_size, bool mod_ret)
1378{
1379 u8 *prog = *pprog;
1380 int cnt = 0;
1381
1382 if (emit_call(&prog, __bpf_prog_enter, prog))
1383 return -EINVAL;
1384 /* remember prog start time returned by __bpf_prog_enter */
1385 emit_mov_reg(&prog, true, BPF_REG_6, BPF_REG_0);
1386
1387 /* arg1: lea rdi, [rbp - stack_size] */
1388 EMIT4(0x48, 0x8D, 0x7D, -stack_size);
1389 /* arg2: progs[i]->insnsi for interpreter */
1390 if (!p->jited)
1391 emit_mov_imm64(&prog, BPF_REG_2,
1392 (long) p->insnsi >> 32,
1393 (u32) (long) p->insnsi);
1394 /* call JITed bpf program or interpreter */
1395 if (emit_call(&prog, p->bpf_func, prog))
1396 return -EINVAL;
1397
1398 /* BPF_TRAMP_MODIFY_RETURN trampolines can modify the return
1399 * of the previous call which is then passed on the stack to
1400 * the next BPF program.
1401 */
1402 if (mod_ret)
1403 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1404
1405 /* arg1: mov rdi, progs[i] */
1406 emit_mov_imm64(&prog, BPF_REG_1, (long) p >> 32,
1407 (u32) (long) p);
1408 /* arg2: mov rsi, rbx <- start time in nsec */
1409 emit_mov_reg(&prog, true, BPF_REG_2, BPF_REG_6);
1410 if (emit_call(&prog, __bpf_prog_exit, prog))
1411 return -EINVAL;
1412
1413 *pprog = prog;
1414 return 0;
1415}
1416
1417static void emit_nops(u8 **pprog, unsigned int len)
1418{
1419 unsigned int i, noplen;
1420 u8 *prog = *pprog;
1421 int cnt = 0;
1422
1423 while (len > 0) {
1424 noplen = len;
1425
1426 if (noplen > ASM_NOP_MAX)
1427 noplen = ASM_NOP_MAX;
1428
1429 for (i = 0; i < noplen; i++)
1430 EMIT1(ideal_nops[noplen][i]);
1431 len -= noplen;
1432 }
1433
1434 *pprog = prog;
1435}
1436
1437static void emit_align(u8 **pprog, u32 align)
1438{
1439 u8 *target, *prog = *pprog;
1440
1441 target = PTR_ALIGN(prog, align);
1442 if (target != prog)
1443 emit_nops(&prog, target - prog);
1444
1445 *pprog = prog;
1446}
1447
1448static int emit_cond_near_jump(u8 **pprog, void *func, void *ip, u8 jmp_cond)
1449{
1450 u8 *prog = *pprog;
1451 int cnt = 0;
1452 s64 offset;
1453
1454 offset = func - (ip + 2 + 4);
1455 if (!is_simm32(offset)) {
1456 pr_err("Target %p is out of range\n", func);
1457 return -EINVAL;
1458 }
1459 EMIT2_off32(0x0F, jmp_cond + 0x10, offset);
1460 *pprog = prog;
1461 return 0;
1462}
1463
1464static int invoke_bpf(const struct btf_func_model *m, u8 **pprog,
1465 struct bpf_tramp_progs *tp, int stack_size)
1466{
1467 int i;
1468 u8 *prog = *pprog;
1469
1470 for (i = 0; i < tp->nr_progs; i++) {
1471 if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size, false))
1472 return -EINVAL;
1473 }
1474 *pprog = prog;
1475 return 0;
1476}
1477
1478static int invoke_bpf_mod_ret(const struct btf_func_model *m, u8 **pprog,
1479 struct bpf_tramp_progs *tp, int stack_size,
1480 u8 **branches)
1481{
1482 u8 *prog = *pprog;
1483 int i, cnt = 0;
1484
1485 /* The first fmod_ret program will receive a garbage return value.
1486 * Set this to 0 to avoid confusing the program.
1487 */
1488 emit_mov_imm32(&prog, false, BPF_REG_0, 0);
1489 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1490 for (i = 0; i < tp->nr_progs; i++) {
1491 if (invoke_bpf_prog(m, &prog, tp->progs[i], stack_size, true))
1492 return -EINVAL;
1493
1494 /* mod_ret prog stored return value into [rbp - 8]. Emit:
1495 * if (*(u64 *)(rbp - 8) != 0)
1496 * goto do_fexit;
1497 */
1498 /* cmp QWORD PTR [rbp - 0x8], 0x0 */
1499 EMIT4(0x48, 0x83, 0x7d, 0xf8); EMIT1(0x00);
1500
1501 /* Save the location of the branch and Generate 6 nops
1502 * (4 bytes for an offset and 2 bytes for the jump) These nops
1503 * are replaced with a conditional jump once do_fexit (i.e. the
1504 * start of the fexit invocation) is finalized.
1505 */
1506 branches[i] = prog;
1507 emit_nops(&prog, 4 + 2);
1508 }
1509
1510 *pprog = prog;
1511 return 0;
1512}
1513
1514/* Example:
1515 * __be16 eth_type_trans(struct sk_buff *skb, struct net_device *dev);
1516 * its 'struct btf_func_model' will be nr_args=2
1517 * The assembly code when eth_type_trans is executing after trampoline:
1518 *
1519 * push rbp
1520 * mov rbp, rsp
1521 * sub rsp, 16 // space for skb and dev
1522 * push rbx // temp regs to pass start time
1523 * mov qword ptr [rbp - 16], rdi // save skb pointer to stack
1524 * mov qword ptr [rbp - 8], rsi // save dev pointer to stack
1525 * call __bpf_prog_enter // rcu_read_lock and preempt_disable
1526 * mov rbx, rax // remember start time in bpf stats are enabled
1527 * lea rdi, [rbp - 16] // R1==ctx of bpf prog
1528 * call addr_of_jited_FENTRY_prog
1529 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
1530 * mov rsi, rbx // prog start time
1531 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
1532 * mov rdi, qword ptr [rbp - 16] // restore skb pointer from stack
1533 * mov rsi, qword ptr [rbp - 8] // restore dev pointer from stack
1534 * pop rbx
1535 * leave
1536 * ret
1537 *
1538 * eth_type_trans has 5 byte nop at the beginning. These 5 bytes will be
1539 * replaced with 'call generated_bpf_trampoline'. When it returns
1540 * eth_type_trans will continue executing with original skb and dev pointers.
1541 *
1542 * The assembly code when eth_type_trans is called from trampoline:
1543 *
1544 * push rbp
1545 * mov rbp, rsp
1546 * sub rsp, 24 // space for skb, dev, return value
1547 * push rbx // temp regs to pass start time
1548 * mov qword ptr [rbp - 24], rdi // save skb pointer to stack
1549 * mov qword ptr [rbp - 16], rsi // save dev pointer to stack
1550 * call __bpf_prog_enter // rcu_read_lock and preempt_disable
1551 * mov rbx, rax // remember start time if bpf stats are enabled
1552 * lea rdi, [rbp - 24] // R1==ctx of bpf prog
1553 * call addr_of_jited_FENTRY_prog // bpf prog can access skb and dev
1554 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
1555 * mov rsi, rbx // prog start time
1556 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
1557 * mov rdi, qword ptr [rbp - 24] // restore skb pointer from stack
1558 * mov rsi, qword ptr [rbp - 16] // restore dev pointer from stack
1559 * call eth_type_trans+5 // execute body of eth_type_trans
1560 * mov qword ptr [rbp - 8], rax // save return value
1561 * call __bpf_prog_enter // rcu_read_lock and preempt_disable
1562 * mov rbx, rax // remember start time in bpf stats are enabled
1563 * lea rdi, [rbp - 24] // R1==ctx of bpf prog
1564 * call addr_of_jited_FEXIT_prog // bpf prog can access skb, dev, return value
1565 * movabsq rdi, 64bit_addr_of_struct_bpf_prog // unused if bpf stats are off
1566 * mov rsi, rbx // prog start time
1567 * call __bpf_prog_exit // rcu_read_unlock, preempt_enable and stats math
1568 * mov rax, qword ptr [rbp - 8] // restore eth_type_trans's return value
1569 * pop rbx
1570 * leave
1571 * add rsp, 8 // skip eth_type_trans's frame
1572 * ret // return to its caller
1573 */
1574int arch_prepare_bpf_trampoline(void *image, void *image_end,
1575 const struct btf_func_model *m, u32 flags,
1576 struct bpf_tramp_progs *tprogs,
1577 void *orig_call)
1578{
1579 int ret, i, cnt = 0, nr_args = m->nr_args;
1580 int stack_size = nr_args * 8;
1581 struct bpf_tramp_progs *fentry = &tprogs[BPF_TRAMP_FENTRY];
1582 struct bpf_tramp_progs *fexit = &tprogs[BPF_TRAMP_FEXIT];
1583 struct bpf_tramp_progs *fmod_ret = &tprogs[BPF_TRAMP_MODIFY_RETURN];
1584 u8 **branches = NULL;
1585 u8 *prog;
1586
1587 /* x86-64 supports up to 6 arguments. 7+ can be added in the future */
1588 if (nr_args > 6)
1589 return -ENOTSUPP;
1590
1591 if ((flags & BPF_TRAMP_F_RESTORE_REGS) &&
1592 (flags & BPF_TRAMP_F_SKIP_FRAME))
1593 return -EINVAL;
1594
1595 if (flags & BPF_TRAMP_F_CALL_ORIG)
1596 stack_size += 8; /* room for return value of orig_call */
1597
1598 if (flags & BPF_TRAMP_F_SKIP_FRAME)
1599 /* skip patched call instruction and point orig_call to actual
1600 * body of the kernel function.
1601 */
1602 orig_call += X86_PATCH_SIZE;
1603
1604 prog = image;
1605
1606 EMIT1(0x55); /* push rbp */
1607 EMIT3(0x48, 0x89, 0xE5); /* mov rbp, rsp */
1608 EMIT4(0x48, 0x83, 0xEC, stack_size); /* sub rsp, stack_size */
1609 EMIT1(0x53); /* push rbx */
1610
1611 save_regs(m, &prog, nr_args, stack_size);
1612
1613 if (fentry->nr_progs)
1614 if (invoke_bpf(m, &prog, fentry, stack_size))
1615 return -EINVAL;
1616
1617 if (fmod_ret->nr_progs) {
1618 branches = kcalloc(fmod_ret->nr_progs, sizeof(u8 *),
1619 GFP_KERNEL);
1620 if (!branches)
1621 return -ENOMEM;
1622
1623 if (invoke_bpf_mod_ret(m, &prog, fmod_ret, stack_size,
1624 branches)) {
1625 ret = -EINVAL;
1626 goto cleanup;
1627 }
1628 }
1629
1630 if (flags & BPF_TRAMP_F_CALL_ORIG) {
1631 if (fentry->nr_progs || fmod_ret->nr_progs)
1632 restore_regs(m, &prog, nr_args, stack_size);
1633
1634 /* call original function */
1635 if (emit_call(&prog, orig_call, prog)) {
1636 ret = -EINVAL;
1637 goto cleanup;
1638 }
1639 /* remember return value in a stack for bpf prog to access */
1640 emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
1641 }
1642
1643 if (fmod_ret->nr_progs) {
1644 /* From Intel 64 and IA-32 Architectures Optimization
1645 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
1646 * Coding Rule 11: All branch targets should be 16-byte
1647 * aligned.
1648 */
1649 emit_align(&prog, 16);
1650 /* Update the branches saved in invoke_bpf_mod_ret with the
1651 * aligned address of do_fexit.
1652 */
1653 for (i = 0; i < fmod_ret->nr_progs; i++)
1654 emit_cond_near_jump(&branches[i], prog, branches[i],
1655 X86_JNE);
1656 }
1657
1658 if (fexit->nr_progs)
1659 if (invoke_bpf(m, &prog, fexit, stack_size)) {
1660 ret = -EINVAL;
1661 goto cleanup;
1662 }
1663
1664 if (flags & BPF_TRAMP_F_RESTORE_REGS)
1665 restore_regs(m, &prog, nr_args, stack_size);
1666
1667 /* This needs to be done regardless. If there were fmod_ret programs,
1668 * the return value is only updated on the stack and still needs to be
1669 * restored to R0.
1670 */
1671 if (flags & BPF_TRAMP_F_CALL_ORIG)
1672 /* restore original return value back into RAX */
1673 emit_ldx(&prog, BPF_DW, BPF_REG_0, BPF_REG_FP, -8);
1674
1675 EMIT1(0x5B); /* pop rbx */
1676 EMIT1(0xC9); /* leave */
1677 if (flags & BPF_TRAMP_F_SKIP_FRAME)
1678 /* skip our return address and return to parent */
1679 EMIT4(0x48, 0x83, 0xC4, 8); /* add rsp, 8 */
1680 EMIT1(0xC3); /* ret */
1681 /* Make sure the trampoline generation logic doesn't overflow */
1682 if (WARN_ON_ONCE(prog > (u8 *)image_end - BPF_INSN_SAFETY)) {
1683 ret = -EFAULT;
1684 goto cleanup;
1685 }
1686 ret = prog - (u8 *)image;
1687
1688cleanup:
1689 kfree(branches);
1690 return ret;
1691}
1692
1693static int emit_fallback_jump(u8 **pprog)
1694{
1695 u8 *prog = *pprog;
1696 int err = 0;
1697
1698#ifdef CONFIG_RETPOLINE
1699 /* Note that this assumes the the compiler uses external
1700 * thunks for indirect calls. Both clang and GCC use the same
1701 * naming convention for external thunks.
1702 */
1703 err = emit_jump(&prog, __x86_indirect_thunk_rdx, prog);
1704#else
1705 int cnt = 0;
1706
1707 EMIT2(0xFF, 0xE2); /* jmp rdx */
1708#endif
1709 *pprog = prog;
1710 return err;
1711}
1712
1713static int emit_bpf_dispatcher(u8 **pprog, int a, int b, s64 *progs)
1714{
1715 u8 *jg_reloc, *prog = *pprog;
1716 int pivot, err, jg_bytes = 1, cnt = 0;
1717 s64 jg_offset;
1718
1719 if (a == b) {
1720 /* Leaf node of recursion, i.e. not a range of indices
1721 * anymore.
1722 */
1723 EMIT1(add_1mod(0x48, BPF_REG_3)); /* cmp rdx,func */
1724 if (!is_simm32(progs[a]))
1725 return -1;
1726 EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3),
1727 progs[a]);
1728 err = emit_cond_near_jump(&prog, /* je func */
1729 (void *)progs[a], prog,
1730 X86_JE);
1731 if (err)
1732 return err;
1733
1734 err = emit_fallback_jump(&prog); /* jmp thunk/indirect */
1735 if (err)
1736 return err;
1737
1738 *pprog = prog;
1739 return 0;
1740 }
1741
1742 /* Not a leaf node, so we pivot, and recursively descend into
1743 * the lower and upper ranges.
1744 */
1745 pivot = (b - a) / 2;
1746 EMIT1(add_1mod(0x48, BPF_REG_3)); /* cmp rdx,func */
1747 if (!is_simm32(progs[a + pivot]))
1748 return -1;
1749 EMIT2_off32(0x81, add_1reg(0xF8, BPF_REG_3), progs[a + pivot]);
1750
1751 if (pivot > 2) { /* jg upper_part */
1752 /* Require near jump. */
1753 jg_bytes = 4;
1754 EMIT2_off32(0x0F, X86_JG + 0x10, 0);
1755 } else {
1756 EMIT2(X86_JG, 0);
1757 }
1758 jg_reloc = prog;
1759
1760 err = emit_bpf_dispatcher(&prog, a, a + pivot, /* emit lower_part */
1761 progs);
1762 if (err)
1763 return err;
1764
1765 /* From Intel 64 and IA-32 Architectures Optimization
1766 * Reference Manual, 3.4.1.4 Code Alignment, Assembly/Compiler
1767 * Coding Rule 11: All branch targets should be 16-byte
1768 * aligned.
1769 */
1770 emit_align(&prog, 16);
1771 jg_offset = prog - jg_reloc;
1772 emit_code(jg_reloc - jg_bytes, jg_offset, jg_bytes);
1773
1774 err = emit_bpf_dispatcher(&prog, a + pivot + 1, /* emit upper_part */
1775 b, progs);
1776 if (err)
1777 return err;
1778
1779 *pprog = prog;
1780 return 0;
1781}
1782
1783static int cmp_ips(const void *a, const void *b)
1784{
1785 const s64 *ipa = a;
1786 const s64 *ipb = b;
1787
1788 if (*ipa > *ipb)
1789 return 1;
1790 if (*ipa < *ipb)
1791 return -1;
1792 return 0;
1793}
1794
1795int arch_prepare_bpf_dispatcher(void *image, s64 *funcs, int num_funcs)
1796{
1797 u8 *prog = image;
1798
1799 sort(funcs, num_funcs, sizeof(funcs[0]), cmp_ips, NULL);
1800 return emit_bpf_dispatcher(&prog, 0, num_funcs - 1, funcs);
1801}
1802
1803struct x64_jit_data {
1804 struct bpf_binary_header *header;
1805 int *addrs;
1806 u8 *image;
1807 int proglen;
1808 struct jit_context ctx;
1809};
1810
1811struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1812{
1813 struct bpf_binary_header *header = NULL;
1814 struct bpf_prog *tmp, *orig_prog = prog;
1815 struct x64_jit_data *jit_data;
1816 int proglen, oldproglen = 0;
1817 struct jit_context ctx = {};
1818 bool tmp_blinded = false;
1819 bool extra_pass = false;
1820 u8 *image = NULL;
1821 int *addrs;
1822 int pass;
1823 int i;
1824
1825 if (!prog->jit_requested)
1826 return orig_prog;
1827
1828 tmp = bpf_jit_blind_constants(prog);
1829 /*
1830 * If blinding was requested and we failed during blinding,
1831 * we must fall back to the interpreter.
1832 */
1833 if (IS_ERR(tmp))
1834 return orig_prog;
1835 if (tmp != prog) {
1836 tmp_blinded = true;
1837 prog = tmp;
1838 }
1839
1840 jit_data = prog->aux->jit_data;
1841 if (!jit_data) {
1842 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1843 if (!jit_data) {
1844 prog = orig_prog;
1845 goto out;
1846 }
1847 prog->aux->jit_data = jit_data;
1848 }
1849 addrs = jit_data->addrs;
1850 if (addrs) {
1851 ctx = jit_data->ctx;
1852 oldproglen = jit_data->proglen;
1853 image = jit_data->image;
1854 header = jit_data->header;
1855 extra_pass = true;
1856 goto skip_init_addrs;
1857 }
1858 addrs = kmalloc_array(prog->len + 1, sizeof(*addrs), GFP_KERNEL);
1859 if (!addrs) {
1860 prog = orig_prog;
1861 goto out_addrs;
1862 }
1863
1864 /*
1865 * Before first pass, make a rough estimation of addrs[]
1866 * each BPF instruction is translated to less than 64 bytes
1867 */
1868 for (proglen = 0, i = 0; i <= prog->len; i++) {
1869 proglen += 64;
1870 addrs[i] = proglen;
1871 }
1872 ctx.cleanup_addr = proglen;
1873skip_init_addrs:
1874
1875 /*
1876 * JITed image shrinks with every pass and the loop iterates
1877 * until the image stops shrinking. Very large BPF programs
1878 * may converge on the last pass. In such case do one more
1879 * pass to emit the final image.
1880 */
1881 for (pass = 0; pass < 20 || image; pass++) {
1882 proglen = do_jit(prog, addrs, image, oldproglen, &ctx);
1883 if (proglen <= 0) {
1884out_image:
1885 image = NULL;
1886 if (header)
1887 bpf_jit_binary_free(header);
1888 prog = orig_prog;
1889 goto out_addrs;
1890 }
1891 if (image) {
1892 if (proglen != oldproglen) {
1893 pr_err("bpf_jit: proglen=%d != oldproglen=%d\n",
1894 proglen, oldproglen);
1895 goto out_image;
1896 }
1897 break;
1898 }
1899 if (proglen == oldproglen) {
1900 /*
1901 * The number of entries in extable is the number of BPF_LDX
1902 * insns that access kernel memory via "pointer to BTF type".
1903 * The verifier changed their opcode from LDX|MEM|size
1904 * to LDX|PROBE_MEM|size to make JITing easier.
1905 */
1906 u32 align = __alignof__(struct exception_table_entry);
1907 u32 extable_size = prog->aux->num_exentries *
1908 sizeof(struct exception_table_entry);
1909
1910 /* allocate module memory for x86 insns and extable */
1911 header = bpf_jit_binary_alloc(roundup(proglen, align) + extable_size,
1912 &image, align, jit_fill_hole);
1913 if (!header) {
1914 prog = orig_prog;
1915 goto out_addrs;
1916 }
1917 prog->aux->extable = (void *) image + roundup(proglen, align);
1918 }
1919 oldproglen = proglen;
1920 cond_resched();
1921 }
1922
1923 if (bpf_jit_enable > 1)
1924 bpf_jit_dump(prog->len, proglen, pass + 1, image);
1925
1926 if (image) {
1927 if (!prog->is_func || extra_pass) {
1928 bpf_tail_call_direct_fixup(prog);
1929 bpf_jit_binary_lock_ro(header);
1930 } else {
1931 jit_data->addrs = addrs;
1932 jit_data->ctx = ctx;
1933 jit_data->proglen = proglen;
1934 jit_data->image = image;
1935 jit_data->header = header;
1936 }
1937 prog->bpf_func = (void *)image;
1938 prog->jited = 1;
1939 prog->jited_len = proglen;
1940 } else {
1941 prog = orig_prog;
1942 }
1943
1944 if (!image || !prog->is_func || extra_pass) {
1945 if (image)
1946 bpf_prog_fill_jited_linfo(prog, addrs + 1);
1947out_addrs:
1948 kfree(addrs);
1949 kfree(jit_data);
1950 prog->aux->jit_data = NULL;
1951 }
1952out:
1953 if (tmp_blinded)
1954 bpf_jit_prog_release_other(prog, prog == orig_prog ?
1955 tmp : orig_prog);
1956 return prog;
1957}