?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel...
[cris-mirror.git] / arch / arm64 / net / bpf_jit_comp.c
blobb2fc97a2c56c80fe633777bbb18322c980949953
1 /*
2 * BPF JIT compiler for ARM64
3 *
4 * Copyright (C) 2014-2016 Zi Shen Lim <zlim.lnx@gmail.com>
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program. If not, see <http://www.gnu.org/licenses/>.
17 */
18
19 #define pr_fmt(fmt) "bpf_jit: " fmt
20
21 #include <linux/bpf.h>
22 #include <linux/filter.h>
23 #include <linux/printk.h>
24 #include <linux/skbuff.h>
25 #include <linux/slab.h>
26
27 #include <asm/byteorder.h>
28 #include <asm/cacheflush.h>
29 #include <asm/debug-monitors.h>
30
31 #include "bpf_jit.h"
32
33 int bpf_jit_enable __read_mostly;
34
35 #define TMP_REG_1 (MAX_BPF_JIT_REG + 0)
36 #define TMP_REG_2 (MAX_BPF_JIT_REG + 1)
37 #define TCALL_CNT (MAX_BPF_JIT_REG + 2)
38
39 /* Map BPF registers to A64 registers */
40 static const int bpf2a64[] = {
41 /* return value from in-kernel function, and exit value from eBPF */
42 [BPF_REG_0] = A64_R(7),
43 /* arguments from eBPF program to in-kernel function */
44 [BPF_REG_1] = A64_R(0),
45 [BPF_REG_2] = A64_R(1),
46 [BPF_REG_3] = A64_R(2),
47 [BPF_REG_4] = A64_R(3),
48 [BPF_REG_5] = A64_R(4),
49 /* callee saved registers that in-kernel function will preserve */
50 [BPF_REG_6] = A64_R(19),
51 [BPF_REG_7] = A64_R(20),
52 [BPF_REG_8] = A64_R(21),
53 [BPF_REG_9] = A64_R(22),
54 /* read-only frame pointer to access stack */
55 [BPF_REG_FP] = A64_R(25),
56 /* temporary registers for internal BPF JIT */
57 [TMP_REG_1] = A64_R(10),
58 [TMP_REG_2] = A64_R(11),
59 /* tail_call_cnt */
60 [TCALL_CNT] = A64_R(26),
61 /* temporary register for blinding constants */
62 [BPF_REG_AX] = A64_R(9),
63 };
64
65 struct jit_ctx {
66 const struct bpf_prog *prog;
67 int idx;
68 int epilogue_offset;
69 int *offset;
70 u32 *image;
71 };
72
73 static inline void emit(const u32 insn, struct jit_ctx *ctx)
74 {
75 if (ctx->image != NULL)
76 ctx->image[ctx->idx] = cpu_to_le32(insn);
77
78 ctx->idx++;
79 }
80
81 static inline void emit_a64_mov_i64(const int reg, const u64 val,
82 struct jit_ctx *ctx)
83 {
84 u64 tmp = val;
85 int shift = 0;
86
87 emit(A64_MOVZ(1, reg, tmp & 0xffff, shift), ctx);
88 tmp >>= 16;
89 shift += 16;
90 while (tmp) {
91 if (tmp & 0xffff)
92 emit(A64_MOVK(1, reg, tmp & 0xffff, shift), ctx);
93 tmp >>= 16;
94 shift += 16;
95 }
96 }
97
98 static inline void emit_a64_mov_i(const int is64, const int reg,
99 const s32 val, struct jit_ctx *ctx)
100 {
101 u16 hi = val >> 16;
102 u16 lo = val & 0xffff;
103
104 if (hi & 0x8000) {
105 if (hi == 0xffff) {
106 emit(A64_MOVN(is64, reg, (u16)~lo, 0), ctx);
107 } else {
108 emit(A64_MOVN(is64, reg, (u16)~hi, 16), ctx);
109 emit(A64_MOVK(is64, reg, lo, 0), ctx);
110 }
111 } else {
112 emit(A64_MOVZ(is64, reg, lo, 0), ctx);
113 if (hi)
114 emit(A64_MOVK(is64, reg, hi, 16), ctx);
115 }
116 }
117
118 static inline int bpf2a64_offset(int bpf_to, int bpf_from,
119 const struct jit_ctx *ctx)
120 {
121 int to = ctx->offset[bpf_to];
122 /* -1 to account for the Branch instruction */
123 int from = ctx->offset[bpf_from] - 1;
124
125 return to - from;
126 }
127
128 static void jit_fill_hole(void *area, unsigned int size)
129 {
130 u32 *ptr;
131 /* We are guaranteed to have aligned memory. */
132 for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
133 *ptr++ = cpu_to_le32(AARCH64_BREAK_FAULT);
134 }
135
136 static inline int epilogue_offset(const struct jit_ctx *ctx)
137 {
138 int to = ctx->epilogue_offset;
139 int from = ctx->idx;
140
141 return to - from;
142 }
143
144 /* Stack must be multiples of 16B */
145 #define STACK_ALIGN(sz) (((sz) + 15) & ~15)
146
147 #define _STACK_SIZE \
148 (MAX_BPF_STACK \
149 + 4 /* extra for skb_copy_bits buffer */)
150
151 #define STACK_SIZE STACK_ALIGN(_STACK_SIZE)
152
153 #define PROLOGUE_OFFSET 8
154
155 static int build_prologue(struct jit_ctx *ctx)
156 {
157 const u8 r6 = bpf2a64[BPF_REG_6];
158 const u8 r7 = bpf2a64[BPF_REG_7];
159 const u8 r8 = bpf2a64[BPF_REG_8];
160 const u8 r9 = bpf2a64[BPF_REG_9];
161 const u8 fp = bpf2a64[BPF_REG_FP];
162 const u8 tcc = bpf2a64[TCALL_CNT];
163 const int idx0 = ctx->idx;
164 int cur_offset;
165
166 /*
167 * BPF prog stack layout
168 *
169 * high
170 * original A64_SP => 0:+-----+ BPF prologue
171 * |FP/LR|
172 * current A64_FP => -16:+-----+
173 * | ... | callee saved registers
174 * BPF fp register => -64:+-----+ <= (BPF_FP)
175 * | |
176 * | ... | BPF prog stack
177 * | |
178 * +-----+ <= (BPF_FP - MAX_BPF_STACK)
179 * |RSVD | JIT scratchpad
180 * current A64_SP => +-----+ <= (BPF_FP - STACK_SIZE)
181 * | |
182 * | ... | Function call stack
183 * | |
184 * +-----+
185 * low
186 *
187 */
188
189 /* Save FP and LR registers to stay align with ARM64 AAPCS */
190 emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx);
191 emit(A64_MOV(1, A64_FP, A64_SP), ctx);
192
193 /* Save callee-saved registers */
194 emit(A64_PUSH(r6, r7, A64_SP), ctx);
195 emit(A64_PUSH(r8, r9, A64_SP), ctx);
196 emit(A64_PUSH(fp, tcc, A64_SP), ctx);
197
198 /* Set up BPF prog stack base register */
199 emit(A64_MOV(1, fp, A64_SP), ctx);
200
201 /* Initialize tail_call_cnt */
202 emit(A64_MOVZ(1, tcc, 0, 0), ctx);
203
204 /* Set up function call stack */
205 emit(A64_SUB_I(1, A64_SP, A64_SP, STACK_SIZE), ctx);
206
207 cur_offset = ctx->idx - idx0;
208 if (cur_offset != PROLOGUE_OFFSET) {
209 pr_err_once("PROLOGUE_OFFSET = %d, expected %d!\n",
210 cur_offset, PROLOGUE_OFFSET);
211 return -1;
212 }
213 return 0;
214 }
215
216 static int out_offset = -1; /* initialized on the first pass of build_body() */
217 static int emit_bpf_tail_call(struct jit_ctx *ctx)
218 {
219 /* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
220 const u8 r2 = bpf2a64[BPF_REG_2];
221 const u8 r3 = bpf2a64[BPF_REG_3];
222
223 const u8 tmp = bpf2a64[TMP_REG_1];
224 const u8 prg = bpf2a64[TMP_REG_2];
225 const u8 tcc = bpf2a64[TCALL_CNT];
226 const int idx0 = ctx->idx;
227 #define cur_offset (ctx->idx - idx0)
228 #define jmp_offset (out_offset - (cur_offset))
229 size_t off;
230
231 /* if (index >= array->map.max_entries)
232 * goto out;
233 */
234 off = offsetof(struct bpf_array, map.max_entries);
235 emit_a64_mov_i64(tmp, off, ctx);
236 emit(A64_LDR32(tmp, r2, tmp), ctx);
237 emit(A64_CMP(0, r3, tmp), ctx);
238 emit(A64_B_(A64_COND_GE, jmp_offset), ctx);
239
240 /* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
241 * goto out;
242 * tail_call_cnt++;
243 */
244 emit_a64_mov_i64(tmp, MAX_TAIL_CALL_CNT, ctx);
245 emit(A64_CMP(1, tcc, tmp), ctx);
246 emit(A64_B_(A64_COND_GT, jmp_offset), ctx);
247 emit(A64_ADD_I(1, tcc, tcc, 1), ctx);
248
249 /* prog = array->ptrs[index];
250 * if (prog == NULL)
251 * goto out;
252 */
253 off = offsetof(struct bpf_array, ptrs);
254 emit_a64_mov_i64(tmp, off, ctx);
255 emit(A64_LDR64(tmp, r2, tmp), ctx);
256 emit(A64_LDR64(prg, tmp, r3), ctx);
257 emit(A64_CBZ(1, prg, jmp_offset), ctx);
258
259 /* goto *(prog->bpf_func + prologue_size); */
260 off = offsetof(struct bpf_prog, bpf_func);
261 emit_a64_mov_i64(tmp, off, ctx);
262 emit(A64_LDR64(tmp, prg, tmp), ctx);
263 emit(A64_ADD_I(1, tmp, tmp, sizeof(u32) * PROLOGUE_OFFSET), ctx);
264 emit(A64_BR(tmp), ctx);
265
266 /* out: */
267 if (out_offset == -1)
268 out_offset = cur_offset;
269 if (cur_offset != out_offset) {
270 pr_err_once("tail_call out_offset = %d, expected %d!\n",
271 cur_offset, out_offset);
272 return -1;
273 }
274 return 0;
275 #undef cur_offset
276 #undef jmp_offset
277 }
278
279 static void build_epilogue(struct jit_ctx *ctx)
280 {
281 const u8 r0 = bpf2a64[BPF_REG_0];
282 const u8 r6 = bpf2a64[BPF_REG_6];
283 const u8 r7 = bpf2a64[BPF_REG_7];
284 const u8 r8 = bpf2a64[BPF_REG_8];
285 const u8 r9 = bpf2a64[BPF_REG_9];
286 const u8 fp = bpf2a64[BPF_REG_FP];
287
288 /* We're done with BPF stack */
289 emit(A64_ADD_I(1, A64_SP, A64_SP, STACK_SIZE), ctx);
290
291 /* Restore fs (x25) and x26 */
292 emit(A64_POP(fp, A64_R(26), A64_SP), ctx);
293
294 /* Restore callee-saved register */
295 emit(A64_POP(r8, r9, A64_SP), ctx);
296 emit(A64_POP(r6, r7, A64_SP), ctx);
297
298 /* Restore FP/LR registers */
299 emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx);
300
301 /* Set return value */
302 emit(A64_MOV(1, A64_R(0), r0), ctx);
303
304 emit(A64_RET(A64_LR), ctx);
305 }
306
307 /* JITs an eBPF instruction.
308 * Returns:
309 * 0 - successfully JITed an 8-byte eBPF instruction.
310 * >0 - successfully JITed a 16-byte eBPF instruction.
311 * <0 - failed to JIT.
312 */
313 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx)
314 {
315 const u8 code = insn->code;
316 const u8 dst = bpf2a64[insn->dst_reg];
317 const u8 src = bpf2a64[insn->src_reg];
318 const u8 tmp = bpf2a64[TMP_REG_1];
319 const u8 tmp2 = bpf2a64[TMP_REG_2];
320 const s16 off = insn->off;
321 const s32 imm = insn->imm;
322 const int i = insn - ctx->prog->insnsi;
323 const bool is64 = BPF_CLASS(code) == BPF_ALU64;
324 u8 jmp_cond;
325 s32 jmp_offset;
326
327 #define check_imm(bits, imm) do { \
328 if ((((imm) > 0) && ((imm) >> (bits))) || \
329 (((imm) < 0) && (~(imm) >> (bits)))) { \
330 pr_info("[%2d] imm=%d(0x%x) out of range\n", \
331 i, imm, imm); \
332 return -EINVAL; \
333 } \
334 } while (0)
335 #define check_imm19(imm) check_imm(19, imm)
336 #define check_imm26(imm) check_imm(26, imm)
337
338 switch (code) {
339 /* dst = src */
340 case BPF_ALU | BPF_MOV | BPF_X:
341 case BPF_ALU64 | BPF_MOV | BPF_X:
342 emit(A64_MOV(is64, dst, src), ctx);
343 break;
344 /* dst = dst OP src */
345 case BPF_ALU | BPF_ADD | BPF_X:
346 case BPF_ALU64 | BPF_ADD | BPF_X:
347 emit(A64_ADD(is64, dst, dst, src), ctx);
348 break;
349 case BPF_ALU | BPF_SUB | BPF_X:
350 case BPF_ALU64 | BPF_SUB | BPF_X:
351 emit(A64_SUB(is64, dst, dst, src), ctx);
352 break;
353 case BPF_ALU | BPF_AND | BPF_X:
354 case BPF_ALU64 | BPF_AND | BPF_X:
355 emit(A64_AND(is64, dst, dst, src), ctx);
356 break;
357 case BPF_ALU | BPF_OR | BPF_X:
358 case BPF_ALU64 | BPF_OR | BPF_X:
359 emit(A64_ORR(is64, dst, dst, src), ctx);
360 break;
361 case BPF_ALU | BPF_XOR | BPF_X:
362 case BPF_ALU64 | BPF_XOR | BPF_X:
363 emit(A64_EOR(is64, dst, dst, src), ctx);
364 break;
365 case BPF_ALU | BPF_MUL | BPF_X:
366 case BPF_ALU64 | BPF_MUL | BPF_X:
367 emit(A64_MUL(is64, dst, dst, src), ctx);
368 break;
369 case BPF_ALU | BPF_DIV | BPF_X:
370 case BPF_ALU64 | BPF_DIV | BPF_X:
371 case BPF_ALU | BPF_MOD | BPF_X:
372 case BPF_ALU64 | BPF_MOD | BPF_X:
373 {
374 const u8 r0 = bpf2a64[BPF_REG_0];
375
376 /* if (src == 0) return 0 */
377 jmp_offset = 3; /* skip ahead to else path */
378 check_imm19(jmp_offset);
379 emit(A64_CBNZ(is64, src, jmp_offset), ctx);
380 emit(A64_MOVZ(1, r0, 0, 0), ctx);
381 jmp_offset = epilogue_offset(ctx);
382 check_imm26(jmp_offset);
383 emit(A64_B(jmp_offset), ctx);
384 /* else */
385 switch (BPF_OP(code)) {
386 case BPF_DIV:
387 emit(A64_UDIV(is64, dst, dst, src), ctx);
388 break;
389 case BPF_MOD:
390 emit(A64_UDIV(is64, tmp, dst, src), ctx);
391 emit(A64_MUL(is64, tmp, tmp, src), ctx);
392 emit(A64_SUB(is64, dst, dst, tmp), ctx);
393 break;
394 }
395 break;
396 }
397 case BPF_ALU | BPF_LSH | BPF_X:
398 case BPF_ALU64 | BPF_LSH | BPF_X:
399 emit(A64_LSLV(is64, dst, dst, src), ctx);
400 break;
401 case BPF_ALU | BPF_RSH | BPF_X:
402 case BPF_ALU64 | BPF_RSH | BPF_X:
403 emit(A64_LSRV(is64, dst, dst, src), ctx);
404 break;
405 case BPF_ALU | BPF_ARSH | BPF_X:
406 case BPF_ALU64 | BPF_ARSH | BPF_X:
407 emit(A64_ASRV(is64, dst, dst, src), ctx);
408 break;
409 /* dst = -dst */
410 case BPF_ALU | BPF_NEG:
411 case BPF_ALU64 | BPF_NEG:
412 emit(A64_NEG(is64, dst, dst), ctx);
413 break;
414 /* dst = BSWAP##imm(dst) */
415 case BPF_ALU | BPF_END | BPF_FROM_LE:
416 case BPF_ALU | BPF_END | BPF_FROM_BE:
417 #ifdef CONFIG_CPU_BIG_ENDIAN
418 if (BPF_SRC(code) == BPF_FROM_BE)
419 goto emit_bswap_uxt;
420 #else /* !CONFIG_CPU_BIG_ENDIAN */
421 if (BPF_SRC(code) == BPF_FROM_LE)
422 goto emit_bswap_uxt;
423 #endif
424 switch (imm) {
425 case 16:
426 emit(A64_REV16(is64, dst, dst), ctx);
427 /* zero-extend 16 bits into 64 bits */
428 emit(A64_UXTH(is64, dst, dst), ctx);
429 break;
430 case 32:
431 emit(A64_REV32(is64, dst, dst), ctx);
432 /* upper 32 bits already cleared */
433 break;
434 case 64:
435 emit(A64_REV64(dst, dst), ctx);
436 break;
437 }
438 break;
439 emit_bswap_uxt:
440 switch (imm) {
441 case 16:
442 /* zero-extend 16 bits into 64 bits */
443 emit(A64_UXTH(is64, dst, dst), ctx);
444 break;
445 case 32:
446 /* zero-extend 32 bits into 64 bits */
447 emit(A64_UXTW(is64, dst, dst), ctx);
448 break;
449 case 64:
450 /* nop */
451 break;
452 }
453 break;
454 /* dst = imm */
455 case BPF_ALU | BPF_MOV | BPF_K:
456 case BPF_ALU64 | BPF_MOV | BPF_K:
457 emit_a64_mov_i(is64, dst, imm, ctx);
458 break;
459 /* dst = dst OP imm */
460 case BPF_ALU | BPF_ADD | BPF_K:
461 case BPF_ALU64 | BPF_ADD | BPF_K:
462 emit_a64_mov_i(is64, tmp, imm, ctx);
463 emit(A64_ADD(is64, dst, dst, tmp), ctx);
464 break;
465 case BPF_ALU | BPF_SUB | BPF_K:
466 case BPF_ALU64 | BPF_SUB | BPF_K:
467 emit_a64_mov_i(is64, tmp, imm, ctx);
468 emit(A64_SUB(is64, dst, dst, tmp), ctx);
469 break;
470 case BPF_ALU | BPF_AND | BPF_K:
471 case BPF_ALU64 | BPF_AND | BPF_K:
472 emit_a64_mov_i(is64, tmp, imm, ctx);
473 emit(A64_AND(is64, dst, dst, tmp), ctx);
474 break;
475 case BPF_ALU | BPF_OR | BPF_K:
476 case BPF_ALU64 | BPF_OR | BPF_K:
477 emit_a64_mov_i(is64, tmp, imm, ctx);
478 emit(A64_ORR(is64, dst, dst, tmp), ctx);
479 break;
480 case BPF_ALU | BPF_XOR | BPF_K:
481 case BPF_ALU64 | BPF_XOR | BPF_K:
482 emit_a64_mov_i(is64, tmp, imm, ctx);
483 emit(A64_EOR(is64, dst, dst, tmp), ctx);
484 break;
485 case BPF_ALU | BPF_MUL | BPF_K:
486 case BPF_ALU64 | BPF_MUL | BPF_K:
487 emit_a64_mov_i(is64, tmp, imm, ctx);
488 emit(A64_MUL(is64, dst, dst, tmp), ctx);
489 break;
490 case BPF_ALU | BPF_DIV | BPF_K:
491 case BPF_ALU64 | BPF_DIV | BPF_K:
492 emit_a64_mov_i(is64, tmp, imm, ctx);
493 emit(A64_UDIV(is64, dst, dst, tmp), ctx);
494 break;
495 case BPF_ALU | BPF_MOD | BPF_K:
496 case BPF_ALU64 | BPF_MOD | BPF_K:
497 emit_a64_mov_i(is64, tmp2, imm, ctx);
498 emit(A64_UDIV(is64, tmp, dst, tmp2), ctx);
499 emit(A64_MUL(is64, tmp, tmp, tmp2), ctx);
500 emit(A64_SUB(is64, dst, dst, tmp), ctx);
501 break;
502 case BPF_ALU | BPF_LSH | BPF_K:
503 case BPF_ALU64 | BPF_LSH | BPF_K:
504 emit(A64_LSL(is64, dst, dst, imm), ctx);
505 break;
506 case BPF_ALU | BPF_RSH | BPF_K:
507 case BPF_ALU64 | BPF_RSH | BPF_K:
508 emit(A64_LSR(is64, dst, dst, imm), ctx);
509 break;
510 case BPF_ALU | BPF_ARSH | BPF_K:
511 case BPF_ALU64 | BPF_ARSH | BPF_K:
512 emit(A64_ASR(is64, dst, dst, imm), ctx);
513 break;
514
515 /* JUMP off */
516 case BPF_JMP | BPF_JA:
517 jmp_offset = bpf2a64_offset(i + off, i, ctx);
518 check_imm26(jmp_offset);
519 emit(A64_B(jmp_offset), ctx);
520 break;
521 /* IF (dst COND src) JUMP off */
522 case BPF_JMP | BPF_JEQ | BPF_X:
523 case BPF_JMP | BPF_JGT | BPF_X:
524 case BPF_JMP | BPF_JGE | BPF_X:
525 case BPF_JMP | BPF_JNE | BPF_X:
526 case BPF_JMP | BPF_JSGT | BPF_X:
527 case BPF_JMP | BPF_JSGE | BPF_X:
528 emit(A64_CMP(1, dst, src), ctx);
529 emit_cond_jmp:
530 jmp_offset = bpf2a64_offset(i + off, i, ctx);
531 check_imm19(jmp_offset);
532 switch (BPF_OP(code)) {
533 case BPF_JEQ:
534 jmp_cond = A64_COND_EQ;
535 break;
536 case BPF_JGT:
537 jmp_cond = A64_COND_HI;
538 break;
539 case BPF_JGE:
540 jmp_cond = A64_COND_CS;
541 break;
542 case BPF_JSET:
543 case BPF_JNE:
544 jmp_cond = A64_COND_NE;
545 break;
546 case BPF_JSGT:
547 jmp_cond = A64_COND_GT;
548 break;
549 case BPF_JSGE:
550 jmp_cond = A64_COND_GE;
551 break;
552 default:
553 return -EFAULT;
554 }
555 emit(A64_B_(jmp_cond, jmp_offset), ctx);
556 break;
557 case BPF_JMP | BPF_JSET | BPF_X:
558 emit(A64_TST(1, dst, src), ctx);
559 goto emit_cond_jmp;
560 /* IF (dst COND imm) JUMP off */
561 case BPF_JMP | BPF_JEQ | BPF_K:
562 case BPF_JMP | BPF_JGT | BPF_K:
563 case BPF_JMP | BPF_JGE | BPF_K:
564 case BPF_JMP | BPF_JNE | BPF_K:
565 case BPF_JMP | BPF_JSGT | BPF_K:
566 case BPF_JMP | BPF_JSGE | BPF_K:
567 emit_a64_mov_i(1, tmp, imm, ctx);
568 emit(A64_CMP(1, dst, tmp), ctx);
569 goto emit_cond_jmp;
570 case BPF_JMP | BPF_JSET | BPF_K:
571 emit_a64_mov_i(1, tmp, imm, ctx);
572 emit(A64_TST(1, dst, tmp), ctx);
573 goto emit_cond_jmp;
574 /* function call */
575 case BPF_JMP | BPF_CALL:
576 {
577 const u8 r0 = bpf2a64[BPF_REG_0];
578 const u64 func = (u64)__bpf_call_base + imm;
579
580 emit_a64_mov_i64(tmp, func, ctx);
581 emit(A64_BLR(tmp), ctx);
582 emit(A64_MOV(1, r0, A64_R(0)), ctx);
583 break;
584 }
585 /* tail call */
586 case BPF_JMP | BPF_CALL | BPF_X:
587 if (emit_bpf_tail_call(ctx))
588 return -EFAULT;
589 break;
590 /* function return */
591 case BPF_JMP | BPF_EXIT:
592 /* Optimization: when last instruction is EXIT,
593 simply fallthrough to epilogue. */
594 if (i == ctx->prog->len - 1)
595 break;
596 jmp_offset = epilogue_offset(ctx);
597 check_imm26(jmp_offset);
598 emit(A64_B(jmp_offset), ctx);
599 break;
600
601 /* dst = imm64 */
602 case BPF_LD | BPF_IMM | BPF_DW:
603 {
604 const struct bpf_insn insn1 = insn[1];
605 u64 imm64;
606
607 if (insn1.code != 0 || insn1.src_reg != 0 ||
608 insn1.dst_reg != 0 || insn1.off != 0) {
609 /* Note: verifier in BPF core must catch invalid
610 * instructions.
611 */
612 pr_err_once("Invalid BPF_LD_IMM64 instruction\n");
613 return -EINVAL;
614 }
615
616 imm64 = (u64)insn1.imm << 32 | (u32)imm;
617 emit_a64_mov_i64(dst, imm64, ctx);
618
619 return 1;
620 }
621
622 /* LDX: dst = *(size *)(src + off) */
623 case BPF_LDX | BPF_MEM | BPF_W:
624 case BPF_LDX | BPF_MEM | BPF_H:
625 case BPF_LDX | BPF_MEM | BPF_B:
626 case BPF_LDX | BPF_MEM | BPF_DW:
627 emit_a64_mov_i(1, tmp, off, ctx);
628 switch (BPF_SIZE(code)) {
629 case BPF_W:
630 emit(A64_LDR32(dst, src, tmp), ctx);
631 break;
632 case BPF_H:
633 emit(A64_LDRH(dst, src, tmp), ctx);
634 break;
635 case BPF_B:
636 emit(A64_LDRB(dst, src, tmp), ctx);
637 break;
638 case BPF_DW:
639 emit(A64_LDR64(dst, src, tmp), ctx);
640 break;
641 }
642 break;
643
644 /* ST: *(size *)(dst + off) = imm */
645 case BPF_ST | BPF_MEM | BPF_W:
646 case BPF_ST | BPF_MEM | BPF_H:
647 case BPF_ST | BPF_MEM | BPF_B:
648 case BPF_ST | BPF_MEM | BPF_DW:
649 /* Load imm to a register then store it */
650 emit_a64_mov_i(1, tmp2, off, ctx);
651 emit_a64_mov_i(1, tmp, imm, ctx);
652 switch (BPF_SIZE(code)) {
653 case BPF_W:
654 emit(A64_STR32(tmp, dst, tmp2), ctx);
655 break;
656 case BPF_H:
657 emit(A64_STRH(tmp, dst, tmp2), ctx);
658 break;
659 case BPF_B:
660 emit(A64_STRB(tmp, dst, tmp2), ctx);
661 break;
662 case BPF_DW:
663 emit(A64_STR64(tmp, dst, tmp2), ctx);
664 break;
665 }
666 break;
667
668 /* STX: *(size *)(dst + off) = src */
669 case BPF_STX | BPF_MEM | BPF_W:
670 case BPF_STX | BPF_MEM | BPF_H:
671 case BPF_STX | BPF_MEM | BPF_B:
672 case BPF_STX | BPF_MEM | BPF_DW:
673 emit_a64_mov_i(1, tmp, off, ctx);
674 switch (BPF_SIZE(code)) {
675 case BPF_W:
676 emit(A64_STR32(src, dst, tmp), ctx);
677 break;
678 case BPF_H:
679 emit(A64_STRH(src, dst, tmp), ctx);
680 break;
681 case BPF_B:
682 emit(A64_STRB(src, dst, tmp), ctx);
683 break;
684 case BPF_DW:
685 emit(A64_STR64(src, dst, tmp), ctx);
686 break;
687 }
688 break;
689 /* STX XADD: lock *(u32 *)(dst + off) += src */
690 case BPF_STX | BPF_XADD | BPF_W:
691 /* STX XADD: lock *(u64 *)(dst + off) += src */
692 case BPF_STX | BPF_XADD | BPF_DW:
693 goto notyet;
694
695 /* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + imm)) */
696 case BPF_LD | BPF_ABS | BPF_W:
697 case BPF_LD | BPF_ABS | BPF_H:
698 case BPF_LD | BPF_ABS | BPF_B:
699 /* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + src + imm)) */
700 case BPF_LD | BPF_IND | BPF_W:
701 case BPF_LD | BPF_IND | BPF_H:
702 case BPF_LD | BPF_IND | BPF_B:
703 {
704 const u8 r0 = bpf2a64[BPF_REG_0]; /* r0 = return value */
705 const u8 r6 = bpf2a64[BPF_REG_6]; /* r6 = pointer to sk_buff */
706 const u8 fp = bpf2a64[BPF_REG_FP];
707 const u8 r1 = bpf2a64[BPF_REG_1]; /* r1: struct sk_buff *skb */
708 const u8 r2 = bpf2a64[BPF_REG_2]; /* r2: int k */
709 const u8 r3 = bpf2a64[BPF_REG_3]; /* r3: unsigned int size */
710 const u8 r4 = bpf2a64[BPF_REG_4]; /* r4: void *buffer */
711 const u8 r5 = bpf2a64[BPF_REG_5]; /* r5: void *(*func)(...) */
712 int size;
713
714 emit(A64_MOV(1, r1, r6), ctx);
715 emit_a64_mov_i(0, r2, imm, ctx);
716 if (BPF_MODE(code) == BPF_IND)
717 emit(A64_ADD(0, r2, r2, src), ctx);
718 switch (BPF_SIZE(code)) {
719 case BPF_W:
720 size = 4;
721 break;
722 case BPF_H:
723 size = 2;
724 break;
725 case BPF_B:
726 size = 1;
727 break;
728 default:
729 return -EINVAL;
730 }
731 emit_a64_mov_i64(r3, size, ctx);
732 emit(A64_SUB_I(1, r4, fp, STACK_SIZE), ctx);
733 emit_a64_mov_i64(r5, (unsigned long)bpf_load_pointer, ctx);
734 emit(A64_BLR(r5), ctx);
735 emit(A64_MOV(1, r0, A64_R(0)), ctx);
736
737 jmp_offset = epilogue_offset(ctx);
738 check_imm19(jmp_offset);
739 emit(A64_CBZ(1, r0, jmp_offset), ctx);
740 emit(A64_MOV(1, r5, r0), ctx);
741 switch (BPF_SIZE(code)) {
742 case BPF_W:
743 emit(A64_LDR32(r0, r5, A64_ZR), ctx);
744 #ifndef CONFIG_CPU_BIG_ENDIAN
745 emit(A64_REV32(0, r0, r0), ctx);
746 #endif
747 break;
748 case BPF_H:
749 emit(A64_LDRH(r0, r5, A64_ZR), ctx);
750 #ifndef CONFIG_CPU_BIG_ENDIAN
751 emit(A64_REV16(0, r0, r0), ctx);
752 #endif
753 break;
754 case BPF_B:
755 emit(A64_LDRB(r0, r5, A64_ZR), ctx);
756 break;
757 }
758 break;
759 }
760 notyet:
761 pr_info_once("*** NOT YET: opcode %02x ***\n", code);
762 return -EFAULT;
763
764 default:
765 pr_err_once("unknown opcode %02x\n", code);
766 return -EINVAL;
767 }
768
769 return 0;
770 }
771
772 static int build_body(struct jit_ctx *ctx)
773 {
774 const struct bpf_prog *prog = ctx->prog;
775 int i;
776
777 for (i = 0; i < prog->len; i++) {
778 const struct bpf_insn *insn = &prog->insnsi[i];
779 int ret;
780
781 ret = build_insn(insn, ctx);
782
783 if (ctx->image == NULL)
784 ctx->offset[i] = ctx->idx;
785
786 if (ret > 0) {
787 i++;
788 continue;
789 }
790 if (ret)
791 return ret;
792 }
793
794 return 0;
795 }
796
797 static int validate_code(struct jit_ctx *ctx)
798 {
799 int i;
800
801 for (i = 0; i < ctx->idx; i++) {
802 u32 a64_insn = le32_to_cpu(ctx->image[i]);
803
804 if (a64_insn == AARCH64_BREAK_FAULT)
805 return -1;
806 }
807
808 return 0;
809 }
810
811 static inline void bpf_flush_icache(void *start, void *end)
812 {
813 flush_icache_range((unsigned long)start, (unsigned long)end);
814 }
815
816 void bpf_jit_compile(struct bpf_prog *prog)
817 {
818 /* Nothing to do here. We support Internal BPF. */
819 }
820
821 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
822 {
823 struct bpf_prog *tmp, *orig_prog = prog;
824 struct bpf_binary_header *header;
825 bool tmp_blinded = false;
826 struct jit_ctx ctx;
827 int image_size;
828 u8 *image_ptr;
829
830 if (!bpf_jit_enable)
831 return orig_prog;
832
833 tmp = bpf_jit_blind_constants(prog);
834 /* If blinding was requested and we failed during blinding,
835 * we must fall back to the interpreter.
836 */
837 if (IS_ERR(tmp))
838 return orig_prog;
839 if (tmp != prog) {
840 tmp_blinded = true;
841 prog = tmp;
842 }
843
844 memset(&ctx, 0, sizeof(ctx));
845 ctx.prog = prog;
846
847 ctx.offset = kcalloc(prog->len, sizeof(int), GFP_KERNEL);
848 if (ctx.offset == NULL) {
849 prog = orig_prog;
850 goto out;
851 }
852
853 /* 1. Initial fake pass to compute ctx->idx. */
854
855 /* Fake pass to fill in ctx->offset. */
856 if (build_body(&ctx)) {
857 prog = orig_prog;
858 goto out_off;
859 }
860
861 if (build_prologue(&ctx)) {
862 prog = orig_prog;
863 goto out_off;
864 }
865
866 ctx.epilogue_offset = ctx.idx;
867 build_epilogue(&ctx);
868
869 /* Now we know the actual image size. */
870 image_size = sizeof(u32) * ctx.idx;
871 header = bpf_jit_binary_alloc(image_size, &image_ptr,
872 sizeof(u32), jit_fill_hole);
873 if (header == NULL) {
874 prog = orig_prog;
875 goto out_off;
876 }
877
878 /* 2. Now, the actual pass. */
879
880 ctx.image = (u32 *)image_ptr;
881 ctx.idx = 0;
882
883 build_prologue(&ctx);
884
885 if (build_body(&ctx)) {
886 bpf_jit_binary_free(header);
887 prog = orig_prog;
888 goto out_off;
889 }
890
891 build_epilogue(&ctx);
892
893 /* 3. Extra pass to validate JITed code. */
894 if (validate_code(&ctx)) {
895 bpf_jit_binary_free(header);
896 prog = orig_prog;
897 goto out_off;
898 }
899
900 /* And we're done. */
901 if (bpf_jit_enable > 1)
902 bpf_jit_dump(prog->len, image_size, 2, ctx.image);
903
904 bpf_flush_icache(header, ctx.image + ctx.idx);
905
906 set_memory_ro((unsigned long)header, header->pages);
907 prog->bpf_func = (void *)ctx.image;
908 prog->jited = 1;
909
910 out_off:
911 kfree(ctx.offset);
912 out:
913 if (tmp_blinded)
914 bpf_jit_prog_release_other(prog, prog == orig_prog ?
915 tmp : orig_prog);
916 return prog;
917 }
918
919 void bpf_jit_free(struct bpf_prog *prog)
920 {
921 unsigned long addr = (unsigned long)prog->bpf_func & PAGE_MASK;
922 struct bpf_binary_header *header = (void *)addr;
923
924 if (!prog->jited)
925 goto free_filter;
926
927 set_memory_rw(addr, header->pages);
928 bpf_jit_binary_free(header);
929
930 free_filter:
931 bpf_prog_unlock_free(prog);
932 }
CRIS linux kernel tree