1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * BPF JIT compiler for ARM64
4 *
5 * Copyright (C) 2014-2016 Zi Shen Lim <zlim.lnx@gmail.com>
6 */
7
8 #define pr_fmt(fmt) "bpf_jit: " fmt
9
10 #include <linux/bpf.h>
11 #include <linux/filter.h>
12 #include <linux/printk.h>
13 #include <linux/slab.h>
14
15 #include <asm/byteorder.h>
16 #include <asm/cacheflush.h>
17 #include <asm/debug-monitors.h>
18 #include <asm/set_memory.h>
19
20 #include "bpf_jit.h"
21
22 #define TMP_REG_1 (MAX_BPF_JIT_REG + 0)
23 #define TMP_REG_2 (MAX_BPF_JIT_REG + 1)
24 #define TCALL_CNT (MAX_BPF_JIT_REG + 2)
25 #define TMP_REG_3 (MAX_BPF_JIT_REG + 3)
26
27 /* Map BPF registers to A64 registers */
28 static const int bpf2a64[] = {
29 /* return value from in-kernel function, and exit value from eBPF */
30 [BPF_REG_0] = A64_R(7),
31 /* arguments from eBPF program to in-kernel function */
32 [BPF_REG_1] = A64_R(0),
33 [BPF_REG_2] = A64_R(1),
34 [BPF_REG_3] = A64_R(2),
35 [BPF_REG_4] = A64_R(3),
36 [BPF_REG_5] = A64_R(4),
37 /* callee saved registers that in-kernel function will preserve */
38 [BPF_REG_6] = A64_R(19),
39 [BPF_REG_7] = A64_R(20),
40 [BPF_REG_8] = A64_R(21),
41 [BPF_REG_9] = A64_R(22),
42 /* read-only frame pointer to access stack */
43 [BPF_REG_FP] = A64_R(25),
44 /* temporary registers for internal BPF JIT */
45 [TMP_REG_1] = A64_R(10),
46 [TMP_REG_2] = A64_R(11),
47 [TMP_REG_3] = A64_R(12),
48 /* tail_call_cnt */
49 [TCALL_CNT] = A64_R(26),
50 /* temporary register for blinding constants */
51 [BPF_REG_AX] = A64_R(9),
52 };
53
54 struct jit_ctx {
55 const struct bpf_prog *prog;
56 int idx;
57 int epilogue_offset;
58 int *offset;
59 __le32 *image;
60 u32 stack_size;
61 };
62
emit(const u32 insn,struct jit_ctx * ctx)63 static inline void emit(const u32 insn, struct jit_ctx *ctx)
64 {
65 if (ctx->image != NULL)
66 ctx->image[ctx->idx] = cpu_to_le32(insn);
67
68 ctx->idx++;
69 }
70
emit_a64_mov_i(const int is64,const int reg,const s32 val,struct jit_ctx * ctx)71 static inline void emit_a64_mov_i(const int is64, const int reg,
72 const s32 val, struct jit_ctx *ctx)
73 {
74 u16 hi = val >> 16;
75 u16 lo = val & 0xffff;
76
77 if (hi & 0x8000) {
78 if (hi == 0xffff) {
79 emit(A64_MOVN(is64, reg, (u16)~lo, 0), ctx);
80 } else {
81 emit(A64_MOVN(is64, reg, (u16)~hi, 16), ctx);
82 if (lo != 0xffff)
83 emit(A64_MOVK(is64, reg, lo, 0), ctx);
84 }
85 } else {
86 emit(A64_MOVZ(is64, reg, lo, 0), ctx);
87 if (hi)
88 emit(A64_MOVK(is64, reg, hi, 16), ctx);
89 }
90 }
91
i64_i16_blocks(const u64 val,bool inverse)92 static int i64_i16_blocks(const u64 val, bool inverse)
93 {
94 return (((val >> 0) & 0xffff) != (inverse ? 0xffff : 0x0000)) +
95 (((val >> 16) & 0xffff) != (inverse ? 0xffff : 0x0000)) +
96 (((val >> 32) & 0xffff) != (inverse ? 0xffff : 0x0000)) +
97 (((val >> 48) & 0xffff) != (inverse ? 0xffff : 0x0000));
98 }
99
emit_a64_mov_i64(const int reg,const u64 val,struct jit_ctx * ctx)100 static inline void emit_a64_mov_i64(const int reg, const u64 val,
101 struct jit_ctx *ctx)
102 {
103 u64 nrm_tmp = val, rev_tmp = ~val;
104 bool inverse;
105 int shift;
106
107 if (!(nrm_tmp >> 32))
108 return emit_a64_mov_i(0, reg, (u32)val, ctx);
109
110 inverse = i64_i16_blocks(nrm_tmp, true) < i64_i16_blocks(nrm_tmp, false);
111 shift = max(round_down((inverse ? (fls64(rev_tmp) - 1) :
112 (fls64(nrm_tmp) - 1)), 16), 0);
113 if (inverse)
114 emit(A64_MOVN(1, reg, (rev_tmp >> shift) & 0xffff, shift), ctx);
115 else
116 emit(A64_MOVZ(1, reg, (nrm_tmp >> shift) & 0xffff, shift), ctx);
117 shift -= 16;
118 while (shift >= 0) {
119 if (((nrm_tmp >> shift) & 0xffff) != (inverse ? 0xffff : 0x0000))
120 emit(A64_MOVK(1, reg, (nrm_tmp >> shift) & 0xffff, shift), ctx);
121 shift -= 16;
122 }
123 }
124
125 /*
126 * Kernel addresses in the vmalloc space use at most 48 bits, and the
127 * remaining bits are guaranteed to be 0x1. So we can compose the address
128 * with a fixed length movn/movk/movk sequence.
129 */
emit_addr_mov_i64(const int reg,const u64 val,struct jit_ctx * ctx)130 static inline void emit_addr_mov_i64(const int reg, const u64 val,
131 struct jit_ctx *ctx)
132 {
133 u64 tmp = val;
134 int shift = 0;
135
136 emit(A64_MOVN(1, reg, ~tmp & 0xffff, shift), ctx);
137 while (shift < 32) {
138 tmp >>= 16;
139 shift += 16;
140 emit(A64_MOVK(1, reg, tmp & 0xffff, shift), ctx);
141 }
142 }
143
bpf2a64_offset(int bpf_insn,int off,const struct jit_ctx * ctx)144 static inline int bpf2a64_offset(int bpf_insn, int off,
145 const struct jit_ctx *ctx)
146 {
147 /* BPF JMP offset is relative to the next instruction */
148 bpf_insn++;
149 /*
150 * Whereas arm64 branch instructions encode the offset
151 * from the branch itself, so we must subtract 1 from the
152 * instruction offset.
153 */
154 return ctx->offset[bpf_insn + off] - (ctx->offset[bpf_insn] - 1);
155 }
156
jit_fill_hole(void * area,unsigned int size)157 static void jit_fill_hole(void *area, unsigned int size)
158 {
159 __le32 *ptr;
160 /* We are guaranteed to have aligned memory. */
161 for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
162 *ptr++ = cpu_to_le32(AARCH64_BREAK_FAULT);
163 }
164
epilogue_offset(const struct jit_ctx * ctx)165 static inline int epilogue_offset(const struct jit_ctx *ctx)
166 {
167 int to = ctx->epilogue_offset;
168 int from = ctx->idx;
169
170 return to - from;
171 }
172
173 /* Stack must be multiples of 16B */
174 #define STACK_ALIGN(sz) (((sz) + 15) & ~15)
175
176 /* Tail call offset to jump into */
177 #define PROLOGUE_OFFSET 7
178
build_prologue(struct jit_ctx * ctx,bool ebpf_from_cbpf)179 static int build_prologue(struct jit_ctx *ctx, bool ebpf_from_cbpf)
180 {
181 const struct bpf_prog *prog = ctx->prog;
182 const u8 r6 = bpf2a64[BPF_REG_6];
183 const u8 r7 = bpf2a64[BPF_REG_7];
184 const u8 r8 = bpf2a64[BPF_REG_8];
185 const u8 r9 = bpf2a64[BPF_REG_9];
186 const u8 fp = bpf2a64[BPF_REG_FP];
187 const u8 tcc = bpf2a64[TCALL_CNT];
188 const int idx0 = ctx->idx;
189 int cur_offset;
190
191 /*
192 * BPF prog stack layout
193 *
194 * high
195 * original A64_SP => 0:+-----+ BPF prologue
196 * |FP/LR|
197 * current A64_FP => -16:+-----+
198 * | ... | callee saved registers
199 * BPF fp register => -64:+-----+ <= (BPF_FP)
200 * | |
201 * | ... | BPF prog stack
202 * | |
203 * +-----+ <= (BPF_FP - prog->aux->stack_depth)
204 * |RSVD | padding
205 * current A64_SP => +-----+ <= (BPF_FP - ctx->stack_size)
206 * | |
207 * | ... | Function call stack
208 * | |
209 * +-----+
210 * low
211 *
212 */
213
214 /* Save FP and LR registers to stay align with ARM64 AAPCS */
215 emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx);
216 emit(A64_MOV(1, A64_FP, A64_SP), ctx);
217
218 /* Save callee-saved registers */
219 emit(A64_PUSH(r6, r7, A64_SP), ctx);
220 emit(A64_PUSH(r8, r9, A64_SP), ctx);
221 emit(A64_PUSH(fp, tcc, A64_SP), ctx);
222
223 /* Set up BPF prog stack base register */
224 emit(A64_MOV(1, fp, A64_SP), ctx);
225
226 if (!ebpf_from_cbpf) {
227 /* Initialize tail_call_cnt */
228 emit(A64_MOVZ(1, tcc, 0, 0), ctx);
229
230 cur_offset = ctx->idx - idx0;
231 if (cur_offset != PROLOGUE_OFFSET) {
232 pr_err_once("PROLOGUE_OFFSET = %d, expected %d!\n",
233 cur_offset, PROLOGUE_OFFSET);
234 return -1;
235 }
236 }
237
238 ctx->stack_size = STACK_ALIGN(prog->aux->stack_depth);
239
240 /* Set up function call stack */
241 emit(A64_SUB_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
242 return 0;
243 }
244
245 static int out_offset = -1; /* initialized on the first pass of build_body() */
emit_bpf_tail_call(struct jit_ctx * ctx)246 static int emit_bpf_tail_call(struct jit_ctx *ctx)
247 {
248 /* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
249 const u8 r2 = bpf2a64[BPF_REG_2];
250 const u8 r3 = bpf2a64[BPF_REG_3];
251
252 const u8 tmp = bpf2a64[TMP_REG_1];
253 const u8 prg = bpf2a64[TMP_REG_2];
254 const u8 tcc = bpf2a64[TCALL_CNT];
255 const int idx0 = ctx->idx;
256 #define cur_offset (ctx->idx - idx0)
257 #define jmp_offset (out_offset - (cur_offset))
258 size_t off;
259
260 /* if (index >= array->map.max_entries)
261 * goto out;
262 */
263 off = offsetof(struct bpf_array, map.max_entries);
264 emit_a64_mov_i64(tmp, off, ctx);
265 emit(A64_LDR32(tmp, r2, tmp), ctx);
266 emit(A64_MOV(0, r3, r3), ctx);
267 emit(A64_CMP(0, r3, tmp), ctx);
268 emit(A64_B_(A64_COND_CS, jmp_offset), ctx);
269
270 /* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
271 * goto out;
272 * tail_call_cnt++;
273 */
274 emit_a64_mov_i64(tmp, MAX_TAIL_CALL_CNT, ctx);
275 emit(A64_CMP(1, tcc, tmp), ctx);
276 emit(A64_B_(A64_COND_HI, jmp_offset), ctx);
277 emit(A64_ADD_I(1, tcc, tcc, 1), ctx);
278
279 /* prog = array->ptrs[index];
280 * if (prog == NULL)
281 * goto out;
282 */
283 off = offsetof(struct bpf_array, ptrs);
284 emit_a64_mov_i64(tmp, off, ctx);
285 emit(A64_ADD(1, tmp, r2, tmp), ctx);
286 emit(A64_LSL(1, prg, r3, 3), ctx);
287 emit(A64_LDR64(prg, tmp, prg), ctx);
288 emit(A64_CBZ(1, prg, jmp_offset), ctx);
289
290 /* goto *(prog->bpf_func + prologue_offset); */
291 off = offsetof(struct bpf_prog, bpf_func);
292 emit_a64_mov_i64(tmp, off, ctx);
293 emit(A64_LDR64(tmp, prg, tmp), ctx);
294 emit(A64_ADD_I(1, tmp, tmp, sizeof(u32) * PROLOGUE_OFFSET), ctx);
295 emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
296 emit(A64_BR(tmp), ctx);
297
298 /* out: */
299 if (out_offset == -1)
300 out_offset = cur_offset;
301 if (cur_offset != out_offset) {
302 pr_err_once("tail_call out_offset = %d, expected %d!\n",
303 cur_offset, out_offset);
304 return -1;
305 }
306 return 0;
307 #undef cur_offset
308 #undef jmp_offset
309 }
310
build_epilogue(struct jit_ctx * ctx)311 static void build_epilogue(struct jit_ctx *ctx)
312 {
313 const u8 r0 = bpf2a64[BPF_REG_0];
314 const u8 r6 = bpf2a64[BPF_REG_6];
315 const u8 r7 = bpf2a64[BPF_REG_7];
316 const u8 r8 = bpf2a64[BPF_REG_8];
317 const u8 r9 = bpf2a64[BPF_REG_9];
318 const u8 fp = bpf2a64[BPF_REG_FP];
319
320 /* We're done with BPF stack */
321 emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
322
323 /* Restore fs (x25) and x26 */
324 emit(A64_POP(fp, A64_R(26), A64_SP), ctx);
325
326 /* Restore callee-saved register */
327 emit(A64_POP(r8, r9, A64_SP), ctx);
328 emit(A64_POP(r6, r7, A64_SP), ctx);
329
330 /* Restore FP/LR registers */
331 emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx);
332
333 /* Set return value */
334 emit(A64_MOV(1, A64_R(0), r0), ctx);
335
336 emit(A64_RET(A64_LR), ctx);
337 }
338
339 /* JITs an eBPF instruction.
340 * Returns:
341 * 0 - successfully JITed an 8-byte eBPF instruction.
342 * >0 - successfully JITed a 16-byte eBPF instruction.
343 * <0 - failed to JIT.
344 */
build_insn(const struct bpf_insn * insn,struct jit_ctx * ctx,bool extra_pass)345 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx,
346 bool extra_pass)
347 {
348 const u8 code = insn->code;
349 const u8 dst = bpf2a64[insn->dst_reg];
350 const u8 src = bpf2a64[insn->src_reg];
351 const u8 tmp = bpf2a64[TMP_REG_1];
352 const u8 tmp2 = bpf2a64[TMP_REG_2];
353 const u8 tmp3 = bpf2a64[TMP_REG_3];
354 const s16 off = insn->off;
355 const s32 imm = insn->imm;
356 const int i = insn - ctx->prog->insnsi;
357 const bool is64 = BPF_CLASS(code) == BPF_ALU64 ||
358 BPF_CLASS(code) == BPF_JMP;
359 const bool isdw = BPF_SIZE(code) == BPF_DW;
360 u8 jmp_cond, reg;
361 s32 jmp_offset;
362
363 #define check_imm(bits, imm) do { \
364 if ((((imm) > 0) && ((imm) >> (bits))) || \
365 (((imm) < 0) && (~(imm) >> (bits)))) { \
366 pr_info("[%2d] imm=%d(0x%x) out of range\n", \
367 i, imm, imm); \
368 return -EINVAL; \
369 } \
370 } while (0)
371 #define check_imm19(imm) check_imm(19, imm)
372 #define check_imm26(imm) check_imm(26, imm)
373
374 switch (code) {
375 /* dst = src */
376 case BPF_ALU | BPF_MOV | BPF_X:
377 case BPF_ALU64 | BPF_MOV | BPF_X:
378 emit(A64_MOV(is64, dst, src), ctx);
379 break;
380 /* dst = dst OP src */
381 case BPF_ALU | BPF_ADD | BPF_X:
382 case BPF_ALU64 | BPF_ADD | BPF_X:
383 emit(A64_ADD(is64, dst, dst, src), ctx);
384 break;
385 case BPF_ALU | BPF_SUB | BPF_X:
386 case BPF_ALU64 | BPF_SUB | BPF_X:
387 emit(A64_SUB(is64, dst, dst, src), ctx);
388 break;
389 case BPF_ALU | BPF_AND | BPF_X:
390 case BPF_ALU64 | BPF_AND | BPF_X:
391 emit(A64_AND(is64, dst, dst, src), ctx);
392 break;
393 case BPF_ALU | BPF_OR | BPF_X:
394 case BPF_ALU64 | BPF_OR | BPF_X:
395 emit(A64_ORR(is64, dst, dst, src), ctx);
396 break;
397 case BPF_ALU | BPF_XOR | BPF_X:
398 case BPF_ALU64 | BPF_XOR | BPF_X:
399 emit(A64_EOR(is64, dst, dst, src), ctx);
400 break;
401 case BPF_ALU | BPF_MUL | BPF_X:
402 case BPF_ALU64 | BPF_MUL | BPF_X:
403 emit(A64_MUL(is64, dst, dst, src), ctx);
404 break;
405 case BPF_ALU | BPF_DIV | BPF_X:
406 case BPF_ALU64 | BPF_DIV | BPF_X:
407 case BPF_ALU | BPF_MOD | BPF_X:
408 case BPF_ALU64 | BPF_MOD | BPF_X:
409 switch (BPF_OP(code)) {
410 case BPF_DIV:
411 emit(A64_UDIV(is64, dst, dst, src), ctx);
412 break;
413 case BPF_MOD:
414 emit(A64_UDIV(is64, tmp, dst, src), ctx);
415 emit(A64_MSUB(is64, dst, dst, tmp, src), ctx);
416 break;
417 }
418 break;
419 case BPF_ALU | BPF_LSH | BPF_X:
420 case BPF_ALU64 | BPF_LSH | BPF_X:
421 emit(A64_LSLV(is64, dst, dst, src), ctx);
422 break;
423 case BPF_ALU | BPF_RSH | BPF_X:
424 case BPF_ALU64 | BPF_RSH | BPF_X:
425 emit(A64_LSRV(is64, dst, dst, src), ctx);
426 break;
427 case BPF_ALU | BPF_ARSH | BPF_X:
428 case BPF_ALU64 | BPF_ARSH | BPF_X:
429 emit(A64_ASRV(is64, dst, dst, src), ctx);
430 break;
431 /* dst = -dst */
432 case BPF_ALU | BPF_NEG:
433 case BPF_ALU64 | BPF_NEG:
434 emit(A64_NEG(is64, dst, dst), ctx);
435 break;
436 /* dst = BSWAP##imm(dst) */
437 case BPF_ALU | BPF_END | BPF_FROM_LE:
438 case BPF_ALU | BPF_END | BPF_FROM_BE:
439 #ifdef CONFIG_CPU_BIG_ENDIAN
440 if (BPF_SRC(code) == BPF_FROM_BE)
441 goto emit_bswap_uxt;
442 #else /* !CONFIG_CPU_BIG_ENDIAN */
443 if (BPF_SRC(code) == BPF_FROM_LE)
444 goto emit_bswap_uxt;
445 #endif
446 switch (imm) {
447 case 16:
448 emit(A64_REV16(is64, dst, dst), ctx);
449 /* zero-extend 16 bits into 64 bits */
450 emit(A64_UXTH(is64, dst, dst), ctx);
451 break;
452 case 32:
453 emit(A64_REV32(is64, dst, dst), ctx);
454 /* upper 32 bits already cleared */
455 break;
456 case 64:
457 emit(A64_REV64(dst, dst), ctx);
458 break;
459 }
460 break;
461 emit_bswap_uxt:
462 switch (imm) {
463 case 16:
464 /* zero-extend 16 bits into 64 bits */
465 emit(A64_UXTH(is64, dst, dst), ctx);
466 break;
467 case 32:
468 /* zero-extend 32 bits into 64 bits */
469 emit(A64_UXTW(is64, dst, dst), ctx);
470 break;
471 case 64:
472 /* nop */
473 break;
474 }
475 break;
476 /* dst = imm */
477 case BPF_ALU | BPF_MOV | BPF_K:
478 case BPF_ALU64 | BPF_MOV | BPF_K:
479 emit_a64_mov_i(is64, dst, imm, ctx);
480 break;
481 /* dst = dst OP imm */
482 case BPF_ALU | BPF_ADD | BPF_K:
483 case BPF_ALU64 | BPF_ADD | BPF_K:
484 emit_a64_mov_i(is64, tmp, imm, ctx);
485 emit(A64_ADD(is64, dst, dst, tmp), ctx);
486 break;
487 case BPF_ALU | BPF_SUB | BPF_K:
488 case BPF_ALU64 | BPF_SUB | BPF_K:
489 emit_a64_mov_i(is64, tmp, imm, ctx);
490 emit(A64_SUB(is64, dst, dst, tmp), ctx);
491 break;
492 case BPF_ALU | BPF_AND | BPF_K:
493 case BPF_ALU64 | BPF_AND | BPF_K:
494 emit_a64_mov_i(is64, tmp, imm, ctx);
495 emit(A64_AND(is64, dst, dst, tmp), ctx);
496 break;
497 case BPF_ALU | BPF_OR | BPF_K:
498 case BPF_ALU64 | BPF_OR | BPF_K:
499 emit_a64_mov_i(is64, tmp, imm, ctx);
500 emit(A64_ORR(is64, dst, dst, tmp), ctx);
501 break;
502 case BPF_ALU | BPF_XOR | BPF_K:
503 case BPF_ALU64 | BPF_XOR | BPF_K:
504 emit_a64_mov_i(is64, tmp, imm, ctx);
505 emit(A64_EOR(is64, dst, dst, tmp), ctx);
506 break;
507 case BPF_ALU | BPF_MUL | BPF_K:
508 case BPF_ALU64 | BPF_MUL | BPF_K:
509 emit_a64_mov_i(is64, tmp, imm, ctx);
510 emit(A64_MUL(is64, dst, dst, tmp), ctx);
511 break;
512 case BPF_ALU | BPF_DIV | BPF_K:
513 case BPF_ALU64 | BPF_DIV | BPF_K:
514 emit_a64_mov_i(is64, tmp, imm, ctx);
515 emit(A64_UDIV(is64, dst, dst, tmp), ctx);
516 break;
517 case BPF_ALU | BPF_MOD | BPF_K:
518 case BPF_ALU64 | BPF_MOD | BPF_K:
519 emit_a64_mov_i(is64, tmp2, imm, ctx);
520 emit(A64_UDIV(is64, tmp, dst, tmp2), ctx);
521 emit(A64_MSUB(is64, dst, dst, tmp, tmp2), ctx);
522 break;
523 case BPF_ALU | BPF_LSH | BPF_K:
524 case BPF_ALU64 | BPF_LSH | BPF_K:
525 emit(A64_LSL(is64, dst, dst, imm), ctx);
526 break;
527 case BPF_ALU | BPF_RSH | BPF_K:
528 case BPF_ALU64 | BPF_RSH | BPF_K:
529 emit(A64_LSR(is64, dst, dst, imm), ctx);
530 break;
531 case BPF_ALU | BPF_ARSH | BPF_K:
532 case BPF_ALU64 | BPF_ARSH | BPF_K:
533 emit(A64_ASR(is64, dst, dst, imm), ctx);
534 break;
535
536 /* JUMP off */
537 case BPF_JMP | BPF_JA:
538 jmp_offset = bpf2a64_offset(i, off, ctx);
539 check_imm26(jmp_offset);
540 emit(A64_B(jmp_offset), ctx);
541 break;
542 /* IF (dst COND src) JUMP off */
543 case BPF_JMP | BPF_JEQ | BPF_X:
544 case BPF_JMP | BPF_JGT | BPF_X:
545 case BPF_JMP | BPF_JLT | BPF_X:
546 case BPF_JMP | BPF_JGE | BPF_X:
547 case BPF_JMP | BPF_JLE | BPF_X:
548 case BPF_JMP | BPF_JNE | BPF_X:
549 case BPF_JMP | BPF_JSGT | BPF_X:
550 case BPF_JMP | BPF_JSLT | BPF_X:
551 case BPF_JMP | BPF_JSGE | BPF_X:
552 case BPF_JMP | BPF_JSLE | BPF_X:
553 case BPF_JMP32 | BPF_JEQ | BPF_X:
554 case BPF_JMP32 | BPF_JGT | BPF_X:
555 case BPF_JMP32 | BPF_JLT | BPF_X:
556 case BPF_JMP32 | BPF_JGE | BPF_X:
557 case BPF_JMP32 | BPF_JLE | BPF_X:
558 case BPF_JMP32 | BPF_JNE | BPF_X:
559 case BPF_JMP32 | BPF_JSGT | BPF_X:
560 case BPF_JMP32 | BPF_JSLT | BPF_X:
561 case BPF_JMP32 | BPF_JSGE | BPF_X:
562 case BPF_JMP32 | BPF_JSLE | BPF_X:
563 emit(A64_CMP(is64, dst, src), ctx);
564 emit_cond_jmp:
565 jmp_offset = bpf2a64_offset(i, off, ctx);
566 check_imm19(jmp_offset);
567 switch (BPF_OP(code)) {
568 case BPF_JEQ:
569 jmp_cond = A64_COND_EQ;
570 break;
571 case BPF_JGT:
572 jmp_cond = A64_COND_HI;
573 break;
574 case BPF_JLT:
575 jmp_cond = A64_COND_CC;
576 break;
577 case BPF_JGE:
578 jmp_cond = A64_COND_CS;
579 break;
580 case BPF_JLE:
581 jmp_cond = A64_COND_LS;
582 break;
583 case BPF_JSET:
584 case BPF_JNE:
585 jmp_cond = A64_COND_NE;
586 break;
587 case BPF_JSGT:
588 jmp_cond = A64_COND_GT;
589 break;
590 case BPF_JSLT:
591 jmp_cond = A64_COND_LT;
592 break;
593 case BPF_JSGE:
594 jmp_cond = A64_COND_GE;
595 break;
596 case BPF_JSLE:
597 jmp_cond = A64_COND_LE;
598 break;
599 default:
600 return -EFAULT;
601 }
602 emit(A64_B_(jmp_cond, jmp_offset), ctx);
603 break;
604 case BPF_JMP | BPF_JSET | BPF_X:
605 case BPF_JMP32 | BPF_JSET | BPF_X:
606 emit(A64_TST(is64, dst, src), ctx);
607 goto emit_cond_jmp;
608 /* IF (dst COND imm) JUMP off */
609 case BPF_JMP | BPF_JEQ | BPF_K:
610 case BPF_JMP | BPF_JGT | BPF_K:
611 case BPF_JMP | BPF_JLT | BPF_K:
612 case BPF_JMP | BPF_JGE | BPF_K:
613 case BPF_JMP | BPF_JLE | BPF_K:
614 case BPF_JMP | BPF_JNE | BPF_K:
615 case BPF_JMP | BPF_JSGT | BPF_K:
616 case BPF_JMP | BPF_JSLT | BPF_K:
617 case BPF_JMP | BPF_JSGE | BPF_K:
618 case BPF_JMP | BPF_JSLE | BPF_K:
619 case BPF_JMP32 | BPF_JEQ | BPF_K:
620 case BPF_JMP32 | BPF_JGT | BPF_K:
621 case BPF_JMP32 | BPF_JLT | BPF_K:
622 case BPF_JMP32 | BPF_JGE | BPF_K:
623 case BPF_JMP32 | BPF_JLE | BPF_K:
624 case BPF_JMP32 | BPF_JNE | BPF_K:
625 case BPF_JMP32 | BPF_JSGT | BPF_K:
626 case BPF_JMP32 | BPF_JSLT | BPF_K:
627 case BPF_JMP32 | BPF_JSGE | BPF_K:
628 case BPF_JMP32 | BPF_JSLE | BPF_K:
629 emit_a64_mov_i(is64, tmp, imm, ctx);
630 emit(A64_CMP(is64, dst, tmp), ctx);
631 goto emit_cond_jmp;
632 case BPF_JMP | BPF_JSET | BPF_K:
633 case BPF_JMP32 | BPF_JSET | BPF_K:
634 emit_a64_mov_i(is64, tmp, imm, ctx);
635 emit(A64_TST(is64, dst, tmp), ctx);
636 goto emit_cond_jmp;
637 /* function call */
638 case BPF_JMP | BPF_CALL:
639 {
640 const u8 r0 = bpf2a64[BPF_REG_0];
641 bool func_addr_fixed;
642 u64 func_addr;
643 int ret;
644
645 ret = bpf_jit_get_func_addr(ctx->prog, insn, extra_pass,
646 &func_addr, &func_addr_fixed);
647 if (ret < 0)
648 return ret;
649 emit_addr_mov_i64(tmp, func_addr, ctx);
650 emit(A64_BLR(tmp), ctx);
651 emit(A64_MOV(1, r0, A64_R(0)), ctx);
652 break;
653 }
654 /* tail call */
655 case BPF_JMP | BPF_TAIL_CALL:
656 if (emit_bpf_tail_call(ctx))
657 return -EFAULT;
658 break;
659 /* function return */
660 case BPF_JMP | BPF_EXIT:
661 /* Optimization: when last instruction is EXIT,
662 simply fallthrough to epilogue. */
663 if (i == ctx->prog->len - 1)
664 break;
665 jmp_offset = epilogue_offset(ctx);
666 check_imm26(jmp_offset);
667 emit(A64_B(jmp_offset), ctx);
668 break;
669
670 /* dst = imm64 */
671 case BPF_LD | BPF_IMM | BPF_DW:
672 {
673 const struct bpf_insn insn1 = insn[1];
674 u64 imm64;
675
676 imm64 = (u64)insn1.imm << 32 | (u32)imm;
677 emit_a64_mov_i64(dst, imm64, ctx);
678
679 return 1;
680 }
681
682 /* LDX: dst = *(size *)(src + off) */
683 case BPF_LDX | BPF_MEM | BPF_W:
684 case BPF_LDX | BPF_MEM | BPF_H:
685 case BPF_LDX | BPF_MEM | BPF_B:
686 case BPF_LDX | BPF_MEM | BPF_DW:
687 emit_a64_mov_i(1, tmp, off, ctx);
688 switch (BPF_SIZE(code)) {
689 case BPF_W:
690 emit(A64_LDR32(dst, src, tmp), ctx);
691 break;
692 case BPF_H:
693 emit(A64_LDRH(dst, src, tmp), ctx);
694 break;
695 case BPF_B:
696 emit(A64_LDRB(dst, src, tmp), ctx);
697 break;
698 case BPF_DW:
699 emit(A64_LDR64(dst, src, tmp), ctx);
700 break;
701 }
702 break;
703
704 /* speculation barrier */
705 case BPF_ST | BPF_NOSPEC:
706 /*
707 * Nothing required here.
708 *
709 * In case of arm64, we rely on the firmware mitigation of
710 * Speculative Store Bypass as controlled via the ssbd kernel
711 * parameter. Whenever the mitigation is enabled, it works
712 * for all of the kernel code with no need to provide any
713 * additional instructions.
714 */
715 break;
716
717 /* ST: *(size *)(dst + off) = imm */
718 case BPF_ST | BPF_MEM | BPF_W:
719 case BPF_ST | BPF_MEM | BPF_H:
720 case BPF_ST | BPF_MEM | BPF_B:
721 case BPF_ST | BPF_MEM | BPF_DW:
722 /* Load imm to a register then store it */
723 emit_a64_mov_i(1, tmp2, off, ctx);
724 emit_a64_mov_i(1, tmp, imm, ctx);
725 switch (BPF_SIZE(code)) {
726 case BPF_W:
727 emit(A64_STR32(tmp, dst, tmp2), ctx);
728 break;
729 case BPF_H:
730 emit(A64_STRH(tmp, dst, tmp2), ctx);
731 break;
732 case BPF_B:
733 emit(A64_STRB(tmp, dst, tmp2), ctx);
734 break;
735 case BPF_DW:
736 emit(A64_STR64(tmp, dst, tmp2), ctx);
737 break;
738 }
739 break;
740
741 /* STX: *(size *)(dst + off) = src */
742 case BPF_STX | BPF_MEM | BPF_W:
743 case BPF_STX | BPF_MEM | BPF_H:
744 case BPF_STX | BPF_MEM | BPF_B:
745 case BPF_STX | BPF_MEM | BPF_DW:
746 emit_a64_mov_i(1, tmp, off, ctx);
747 switch (BPF_SIZE(code)) {
748 case BPF_W:
749 emit(A64_STR32(src, dst, tmp), ctx);
750 break;
751 case BPF_H:
752 emit(A64_STRH(src, dst, tmp), ctx);
753 break;
754 case BPF_B:
755 emit(A64_STRB(src, dst, tmp), ctx);
756 break;
757 case BPF_DW:
758 emit(A64_STR64(src, dst, tmp), ctx);
759 break;
760 }
761 break;
762
763 /* STX XADD: lock *(u32 *)(dst + off) += src */
764 case BPF_STX | BPF_XADD | BPF_W:
765 /* STX XADD: lock *(u64 *)(dst + off) += src */
766 case BPF_STX | BPF_XADD | BPF_DW:
767 if (!off) {
768 reg = dst;
769 } else {
770 emit_a64_mov_i(1, tmp, off, ctx);
771 emit(A64_ADD(1, tmp, tmp, dst), ctx);
772 reg = tmp;
773 }
774 if (cpus_have_cap(ARM64_HAS_LSE_ATOMICS)) {
775 emit(A64_STADD(isdw, reg, src), ctx);
776 } else {
777 emit(A64_LDXR(isdw, tmp2, reg), ctx);
778 emit(A64_ADD(isdw, tmp2, tmp2, src), ctx);
779 emit(A64_STXR(isdw, tmp2, reg, tmp3), ctx);
780 jmp_offset = -3;
781 check_imm19(jmp_offset);
782 emit(A64_CBNZ(0, tmp3, jmp_offset), ctx);
783 }
784 break;
785
786 default:
787 pr_err_once("unknown opcode %02x\n", code);
788 return -EINVAL;
789 }
790
791 return 0;
792 }
793
build_body(struct jit_ctx * ctx,bool extra_pass)794 static int build_body(struct jit_ctx *ctx, bool extra_pass)
795 {
796 const struct bpf_prog *prog = ctx->prog;
797 int i;
798
799 /*
800 * - offset[0] offset of the end of prologue,
801 * start of the 1st instruction.
802 * - offset[1] - offset of the end of 1st instruction,
803 * start of the 2nd instruction
804 * [....]
805 * - offset[3] - offset of the end of 3rd instruction,
806 * start of 4th instruction
807 */
808 for (i = 0; i < prog->len; i++) {
809 const struct bpf_insn *insn = &prog->insnsi[i];
810 int ret;
811
812 if (ctx->image == NULL)
813 ctx->offset[i] = ctx->idx;
814 ret = build_insn(insn, ctx, extra_pass);
815 if (ret > 0) {
816 i++;
817 if (ctx->image == NULL)
818 ctx->offset[i] = ctx->idx;
819 continue;
820 }
821 if (ret)
822 return ret;
823 }
824 /*
825 * offset is allocated with prog->len + 1 so fill in
826 * the last element with the offset after the last
827 * instruction (end of program)
828 */
829 if (ctx->image == NULL)
830 ctx->offset[i] = ctx->idx;
831
832 return 0;
833 }
834
validate_code(struct jit_ctx * ctx)835 static int validate_code(struct jit_ctx *ctx)
836 {
837 int i;
838
839 for (i = 0; i < ctx->idx; i++) {
840 u32 a64_insn = le32_to_cpu(ctx->image[i]);
841
842 if (a64_insn == AARCH64_BREAK_FAULT)
843 return -1;
844 }
845
846 return 0;
847 }
848
bpf_flush_icache(void * start,void * end)849 static inline void bpf_flush_icache(void *start, void *end)
850 {
851 flush_icache_range((unsigned long)start, (unsigned long)end);
852 }
853
854 struct arm64_jit_data {
855 struct bpf_binary_header *header;
856 u8 *image;
857 struct jit_ctx ctx;
858 };
859
bpf_int_jit_compile(struct bpf_prog * prog)860 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
861 {
862 struct bpf_prog *tmp, *orig_prog = prog;
863 struct bpf_binary_header *header;
864 struct arm64_jit_data *jit_data;
865 bool was_classic = bpf_prog_was_classic(prog);
866 bool tmp_blinded = false;
867 bool extra_pass = false;
868 struct jit_ctx ctx;
869 int image_size;
870 u8 *image_ptr;
871
872 if (!prog->jit_requested)
873 return orig_prog;
874
875 tmp = bpf_jit_blind_constants(prog);
876 /* If blinding was requested and we failed during blinding,
877 * we must fall back to the interpreter.
878 */
879 if (IS_ERR(tmp))
880 return orig_prog;
881 if (tmp != prog) {
882 tmp_blinded = true;
883 prog = tmp;
884 }
885
886 jit_data = prog->aux->jit_data;
887 if (!jit_data) {
888 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
889 if (!jit_data) {
890 prog = orig_prog;
891 goto out;
892 }
893 prog->aux->jit_data = jit_data;
894 }
895 if (jit_data->ctx.offset) {
896 ctx = jit_data->ctx;
897 image_ptr = jit_data->image;
898 header = jit_data->header;
899 extra_pass = true;
900 image_size = sizeof(u32) * ctx.idx;
901 goto skip_init_ctx;
902 }
903 memset(&ctx, 0, sizeof(ctx));
904 ctx.prog = prog;
905
906 ctx.offset = kcalloc(prog->len + 1, sizeof(int), GFP_KERNEL);
907 if (ctx.offset == NULL) {
908 prog = orig_prog;
909 goto out_off;
910 }
911
912 /*
913 * 1. Initial fake pass to compute ctx->idx and ctx->offset.
914 *
915 * BPF line info needs ctx->offset[i] to be the offset of
916 * instruction[i] in jited image, so build prologue first.
917 */
918 if (build_prologue(&ctx, was_classic)) {
919 prog = orig_prog;
920 goto out_off;
921 }
922
923 if (build_body(&ctx, extra_pass)) {
924 prog = orig_prog;
925 goto out_off;
926 }
927
928 ctx.epilogue_offset = ctx.idx;
929 build_epilogue(&ctx);
930
931 /* Now we know the actual image size. */
932 image_size = sizeof(u32) * ctx.idx;
933 header = bpf_jit_binary_alloc(image_size, &image_ptr,
934 sizeof(u32), jit_fill_hole);
935 if (header == NULL) {
936 prog = orig_prog;
937 goto out_off;
938 }
939
940 /* 2. Now, the actual pass. */
941
942 ctx.image = (__le32 *)image_ptr;
943 skip_init_ctx:
944 ctx.idx = 0;
945
946 build_prologue(&ctx, was_classic);
947
948 if (build_body(&ctx, extra_pass)) {
949 bpf_jit_binary_free(header);
950 prog = orig_prog;
951 goto out_off;
952 }
953
954 build_epilogue(&ctx);
955
956 /* 3. Extra pass to validate JITed code. */
957 if (validate_code(&ctx)) {
958 bpf_jit_binary_free(header);
959 prog = orig_prog;
960 goto out_off;
961 }
962
963 /* And we're done. */
964 if (bpf_jit_enable > 1)
965 bpf_jit_dump(prog->len, image_size, 2, ctx.image);
966
967 bpf_flush_icache(header, ctx.image + ctx.idx);
968
969 if (!prog->is_func || extra_pass) {
970 if (extra_pass && ctx.idx != jit_data->ctx.idx) {
971 pr_err_once("multi-func JIT bug %d != %d\n",
972 ctx.idx, jit_data->ctx.idx);
973 bpf_jit_binary_free(header);
974 prog->bpf_func = NULL;
975 prog->jited = 0;
976 prog->jited_len = 0;
977 goto out_off;
978 }
979 bpf_jit_binary_lock_ro(header);
980 } else {
981 jit_data->ctx = ctx;
982 jit_data->image = image_ptr;
983 jit_data->header = header;
984 }
985 prog->bpf_func = (void *)ctx.image;
986 prog->jited = 1;
987 prog->jited_len = image_size;
988
989 if (!prog->is_func || extra_pass) {
990 int i;
991
992 /* offset[prog->len] is the size of program */
993 for (i = 0; i <= prog->len; i++)
994 ctx.offset[i] *= AARCH64_INSN_SIZE;
995 bpf_prog_fill_jited_linfo(prog, ctx.offset + 1);
996 out_off:
997 kfree(ctx.offset);
998 kfree(jit_data);
999 prog->aux->jit_data = NULL;
1000 }
1001 out:
1002 if (tmp_blinded)
1003 bpf_jit_prog_release_other(prog, prog == orig_prog ?
1004 tmp : orig_prog);
1005 return prog;
1006 }
1007
bpf_jit_alloc_exec_limit(void)1008 u64 bpf_jit_alloc_exec_limit(void)
1009 {
1010 return BPF_JIT_REGION_SIZE;
1011 }
1012
bpf_jit_alloc_exec(unsigned long size)1013 void *bpf_jit_alloc_exec(unsigned long size)
1014 {
1015 return __vmalloc_node_range(size, PAGE_SIZE, BPF_JIT_REGION_START,
1016 BPF_JIT_REGION_END, GFP_KERNEL,
1017 PAGE_KERNEL, 0, NUMA_NO_NODE,
1018 __builtin_return_address(0));
1019 }
1020
bpf_jit_free_exec(void * addr)1021 void bpf_jit_free_exec(void *addr)
1022 {
1023 return vfree(addr);
1024 }
1025
1026 #ifdef CONFIG_CFI_CLANG
arch_bpf_jit_check_func(const struct bpf_prog * prog)1027 bool arch_bpf_jit_check_func(const struct bpf_prog *prog)
1028 {
1029 const uintptr_t func = (const uintptr_t)prog->bpf_func;
1030
1031 /* bpf_func must be correctly aligned and within the BPF JIT region */
1032 return (func >= BPF_JIT_REGION_START && func < BPF_JIT_REGION_END &&
1033 IS_ALIGNED(func, sizeof(u32)));
1034 }
1035 #endif
1036