xref: /openbmc/linux/arch/arm64/net/bpf_jit_comp.c (revision e00a844a)
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 #include <asm/set_memory.h>
31 
32 #include "bpf_jit.h"
33 
34 int bpf_jit_enable __read_mostly;
35 
36 #define TMP_REG_1 (MAX_BPF_JIT_REG + 0)
37 #define TMP_REG_2 (MAX_BPF_JIT_REG + 1)
38 #define TCALL_CNT (MAX_BPF_JIT_REG + 2)
39 #define TMP_REG_3 (MAX_BPF_JIT_REG + 3)
40 
41 /* Map BPF registers to A64 registers */
42 static const int bpf2a64[] = {
43 	/* return value from in-kernel function, and exit value from eBPF */
44 	[BPF_REG_0] = A64_R(7),
45 	/* arguments from eBPF program to in-kernel function */
46 	[BPF_REG_1] = A64_R(0),
47 	[BPF_REG_2] = A64_R(1),
48 	[BPF_REG_3] = A64_R(2),
49 	[BPF_REG_4] = A64_R(3),
50 	[BPF_REG_5] = A64_R(4),
51 	/* callee saved registers that in-kernel function will preserve */
52 	[BPF_REG_6] = A64_R(19),
53 	[BPF_REG_7] = A64_R(20),
54 	[BPF_REG_8] = A64_R(21),
55 	[BPF_REG_9] = A64_R(22),
56 	/* read-only frame pointer to access stack */
57 	[BPF_REG_FP] = A64_R(25),
58 	/* temporary registers for internal BPF JIT */
59 	[TMP_REG_1] = A64_R(10),
60 	[TMP_REG_2] = A64_R(11),
61 	[TMP_REG_3] = A64_R(12),
62 	/* tail_call_cnt */
63 	[TCALL_CNT] = A64_R(26),
64 	/* temporary register for blinding constants */
65 	[BPF_REG_AX] = A64_R(9),
66 };
67 
68 struct jit_ctx {
69 	const struct bpf_prog *prog;
70 	int idx;
71 	int epilogue_offset;
72 	int *offset;
73 	__le32 *image;
74 	u32 stack_size;
75 };
76 
77 static inline void emit(const u32 insn, struct jit_ctx *ctx)
78 {
79 	if (ctx->image != NULL)
80 		ctx->image[ctx->idx] = cpu_to_le32(insn);
81 
82 	ctx->idx++;
83 }
84 
85 static inline void emit_a64_mov_i64(const int reg, const u64 val,
86 				    struct jit_ctx *ctx)
87 {
88 	u64 tmp = val;
89 	int shift = 0;
90 
91 	emit(A64_MOVZ(1, reg, tmp & 0xffff, shift), ctx);
92 	tmp >>= 16;
93 	shift += 16;
94 	while (tmp) {
95 		if (tmp & 0xffff)
96 			emit(A64_MOVK(1, reg, tmp & 0xffff, shift), ctx);
97 		tmp >>= 16;
98 		shift += 16;
99 	}
100 }
101 
102 static inline void emit_a64_mov_i(const int is64, const int reg,
103 				  const s32 val, struct jit_ctx *ctx)
104 {
105 	u16 hi = val >> 16;
106 	u16 lo = val & 0xffff;
107 
108 	if (hi & 0x8000) {
109 		if (hi == 0xffff) {
110 			emit(A64_MOVN(is64, reg, (u16)~lo, 0), ctx);
111 		} else {
112 			emit(A64_MOVN(is64, reg, (u16)~hi, 16), ctx);
113 			emit(A64_MOVK(is64, reg, lo, 0), ctx);
114 		}
115 	} else {
116 		emit(A64_MOVZ(is64, reg, lo, 0), ctx);
117 		if (hi)
118 			emit(A64_MOVK(is64, reg, hi, 16), ctx);
119 	}
120 }
121 
122 static inline int bpf2a64_offset(int bpf_to, int bpf_from,
123 				 const struct jit_ctx *ctx)
124 {
125 	int to = ctx->offset[bpf_to];
126 	/* -1 to account for the Branch instruction */
127 	int from = ctx->offset[bpf_from] - 1;
128 
129 	return to - from;
130 }
131 
132 static void jit_fill_hole(void *area, unsigned int size)
133 {
134 	__le32 *ptr;
135 	/* We are guaranteed to have aligned memory. */
136 	for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
137 		*ptr++ = cpu_to_le32(AARCH64_BREAK_FAULT);
138 }
139 
140 static inline int epilogue_offset(const struct jit_ctx *ctx)
141 {
142 	int to = ctx->epilogue_offset;
143 	int from = ctx->idx;
144 
145 	return to - from;
146 }
147 
148 /* Stack must be multiples of 16B */
149 #define STACK_ALIGN(sz) (((sz) + 15) & ~15)
150 
151 #define PROLOGUE_OFFSET 8
152 
153 static int build_prologue(struct jit_ctx *ctx)
154 {
155 	const struct bpf_prog *prog = ctx->prog;
156 	const u8 r6 = bpf2a64[BPF_REG_6];
157 	const u8 r7 = bpf2a64[BPF_REG_7];
158 	const u8 r8 = bpf2a64[BPF_REG_8];
159 	const u8 r9 = bpf2a64[BPF_REG_9];
160 	const u8 fp = bpf2a64[BPF_REG_FP];
161 	const u8 tcc = bpf2a64[TCALL_CNT];
162 	const int idx0 = ctx->idx;
163 	int cur_offset;
164 
165 	/*
166 	 * BPF prog stack layout
167 	 *
168 	 *                         high
169 	 * original A64_SP =>   0:+-----+ BPF prologue
170 	 *                        |FP/LR|
171 	 * current A64_FP =>  -16:+-----+
172 	 *                        | ... | callee saved registers
173 	 * BPF fp register => -64:+-----+ <= (BPF_FP)
174 	 *                        |     |
175 	 *                        | ... | BPF prog stack
176 	 *                        |     |
177 	 *                        +-----+ <= (BPF_FP - prog->aux->stack_depth)
178 	 *                        |RSVD | JIT scratchpad
179 	 * current A64_SP =>      +-----+ <= (BPF_FP - ctx->stack_size)
180 	 *                        |     |
181 	 *                        | ... | Function call stack
182 	 *                        |     |
183 	 *                        +-----+
184 	 *                          low
185 	 *
186 	 */
187 
188 	/* Save FP and LR registers to stay align with ARM64 AAPCS */
189 	emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx);
190 	emit(A64_MOV(1, A64_FP, A64_SP), ctx);
191 
192 	/* Save callee-saved registers */
193 	emit(A64_PUSH(r6, r7, A64_SP), ctx);
194 	emit(A64_PUSH(r8, r9, A64_SP), ctx);
195 	emit(A64_PUSH(fp, tcc, A64_SP), ctx);
196 
197 	/* Set up BPF prog stack base register */
198 	emit(A64_MOV(1, fp, A64_SP), ctx);
199 
200 	/* Initialize tail_call_cnt */
201 	emit(A64_MOVZ(1, tcc, 0, 0), ctx);
202 
203 	/* 4 byte extra for skb_copy_bits buffer */
204 	ctx->stack_size = prog->aux->stack_depth + 4;
205 	ctx->stack_size = STACK_ALIGN(ctx->stack_size);
206 
207 	/* Set up function call stack */
208 	emit(A64_SUB_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
209 
210 	cur_offset = ctx->idx - idx0;
211 	if (cur_offset != PROLOGUE_OFFSET) {
212 		pr_err_once("PROLOGUE_OFFSET = %d, expected %d!\n",
213 			    cur_offset, PROLOGUE_OFFSET);
214 		return -1;
215 	}
216 	return 0;
217 }
218 
219 static int out_offset = -1; /* initialized on the first pass of build_body() */
220 static int emit_bpf_tail_call(struct jit_ctx *ctx)
221 {
222 	/* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
223 	const u8 r2 = bpf2a64[BPF_REG_2];
224 	const u8 r3 = bpf2a64[BPF_REG_3];
225 
226 	const u8 tmp = bpf2a64[TMP_REG_1];
227 	const u8 prg = bpf2a64[TMP_REG_2];
228 	const u8 tcc = bpf2a64[TCALL_CNT];
229 	const int idx0 = ctx->idx;
230 #define cur_offset (ctx->idx - idx0)
231 #define jmp_offset (out_offset - (cur_offset))
232 	size_t off;
233 
234 	/* if (index >= array->map.max_entries)
235 	 *     goto out;
236 	 */
237 	off = offsetof(struct bpf_array, map.max_entries);
238 	emit_a64_mov_i64(tmp, off, ctx);
239 	emit(A64_LDR32(tmp, r2, tmp), ctx);
240 	emit(A64_CMP(0, r3, tmp), ctx);
241 	emit(A64_B_(A64_COND_GE, jmp_offset), ctx);
242 
243 	/* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
244 	 *     goto out;
245 	 * tail_call_cnt++;
246 	 */
247 	emit_a64_mov_i64(tmp, MAX_TAIL_CALL_CNT, ctx);
248 	emit(A64_CMP(1, tcc, tmp), ctx);
249 	emit(A64_B_(A64_COND_GT, jmp_offset), ctx);
250 	emit(A64_ADD_I(1, tcc, tcc, 1), ctx);
251 
252 	/* prog = array->ptrs[index];
253 	 * if (prog == NULL)
254 	 *     goto out;
255 	 */
256 	off = offsetof(struct bpf_array, ptrs);
257 	emit_a64_mov_i64(tmp, off, ctx);
258 	emit(A64_ADD(1, tmp, r2, tmp), ctx);
259 	emit(A64_LSL(1, prg, r3, 3), ctx);
260 	emit(A64_LDR64(prg, tmp, prg), ctx);
261 	emit(A64_CBZ(1, prg, jmp_offset), ctx);
262 
263 	/* goto *(prog->bpf_func + prologue_size); */
264 	off = offsetof(struct bpf_prog, bpf_func);
265 	emit_a64_mov_i64(tmp, off, ctx);
266 	emit(A64_LDR64(tmp, prg, tmp), ctx);
267 	emit(A64_ADD_I(1, tmp, tmp, sizeof(u32) * PROLOGUE_OFFSET), ctx);
268 	emit(A64_BR(tmp), ctx);
269 
270 	/* out: */
271 	if (out_offset == -1)
272 		out_offset = cur_offset;
273 	if (cur_offset != out_offset) {
274 		pr_err_once("tail_call out_offset = %d, expected %d!\n",
275 			    cur_offset, out_offset);
276 		return -1;
277 	}
278 	return 0;
279 #undef cur_offset
280 #undef jmp_offset
281 }
282 
283 static void build_epilogue(struct jit_ctx *ctx)
284 {
285 	const u8 r0 = bpf2a64[BPF_REG_0];
286 	const u8 r6 = bpf2a64[BPF_REG_6];
287 	const u8 r7 = bpf2a64[BPF_REG_7];
288 	const u8 r8 = bpf2a64[BPF_REG_8];
289 	const u8 r9 = bpf2a64[BPF_REG_9];
290 	const u8 fp = bpf2a64[BPF_REG_FP];
291 
292 	/* We're done with BPF stack */
293 	emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
294 
295 	/* Restore fs (x25) and x26 */
296 	emit(A64_POP(fp, A64_R(26), A64_SP), ctx);
297 
298 	/* Restore callee-saved register */
299 	emit(A64_POP(r8, r9, A64_SP), ctx);
300 	emit(A64_POP(r6, r7, A64_SP), ctx);
301 
302 	/* Restore FP/LR registers */
303 	emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx);
304 
305 	/* Set return value */
306 	emit(A64_MOV(1, A64_R(0), r0), ctx);
307 
308 	emit(A64_RET(A64_LR), ctx);
309 }
310 
311 /* JITs an eBPF instruction.
312  * Returns:
313  * 0  - successfully JITed an 8-byte eBPF instruction.
314  * >0 - successfully JITed a 16-byte eBPF instruction.
315  * <0 - failed to JIT.
316  */
317 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx)
318 {
319 	const u8 code = insn->code;
320 	const u8 dst = bpf2a64[insn->dst_reg];
321 	const u8 src = bpf2a64[insn->src_reg];
322 	const u8 tmp = bpf2a64[TMP_REG_1];
323 	const u8 tmp2 = bpf2a64[TMP_REG_2];
324 	const u8 tmp3 = bpf2a64[TMP_REG_3];
325 	const s16 off = insn->off;
326 	const s32 imm = insn->imm;
327 	const int i = insn - ctx->prog->insnsi;
328 	const bool is64 = BPF_CLASS(code) == BPF_ALU64;
329 	const bool isdw = BPF_SIZE(code) == BPF_DW;
330 	u8 jmp_cond;
331 	s32 jmp_offset;
332 
333 #define check_imm(bits, imm) do {				\
334 	if ((((imm) > 0) && ((imm) >> (bits))) ||		\
335 	    (((imm) < 0) && (~(imm) >> (bits)))) {		\
336 		pr_info("[%2d] imm=%d(0x%x) out of range\n",	\
337 			i, imm, imm);				\
338 		return -EINVAL;					\
339 	}							\
340 } while (0)
341 #define check_imm19(imm) check_imm(19, imm)
342 #define check_imm26(imm) check_imm(26, imm)
343 
344 	switch (code) {
345 	/* dst = src */
346 	case BPF_ALU | BPF_MOV | BPF_X:
347 	case BPF_ALU64 | BPF_MOV | BPF_X:
348 		emit(A64_MOV(is64, dst, src), ctx);
349 		break;
350 	/* dst = dst OP src */
351 	case BPF_ALU | BPF_ADD | BPF_X:
352 	case BPF_ALU64 | BPF_ADD | BPF_X:
353 		emit(A64_ADD(is64, dst, dst, src), ctx);
354 		break;
355 	case BPF_ALU | BPF_SUB | BPF_X:
356 	case BPF_ALU64 | BPF_SUB | BPF_X:
357 		emit(A64_SUB(is64, dst, dst, src), ctx);
358 		break;
359 	case BPF_ALU | BPF_AND | BPF_X:
360 	case BPF_ALU64 | BPF_AND | BPF_X:
361 		emit(A64_AND(is64, dst, dst, src), ctx);
362 		break;
363 	case BPF_ALU | BPF_OR | BPF_X:
364 	case BPF_ALU64 | BPF_OR | BPF_X:
365 		emit(A64_ORR(is64, dst, dst, src), ctx);
366 		break;
367 	case BPF_ALU | BPF_XOR | BPF_X:
368 	case BPF_ALU64 | BPF_XOR | BPF_X:
369 		emit(A64_EOR(is64, dst, dst, src), ctx);
370 		break;
371 	case BPF_ALU | BPF_MUL | BPF_X:
372 	case BPF_ALU64 | BPF_MUL | BPF_X:
373 		emit(A64_MUL(is64, dst, dst, src), ctx);
374 		break;
375 	case BPF_ALU | BPF_DIV | BPF_X:
376 	case BPF_ALU64 | BPF_DIV | BPF_X:
377 	case BPF_ALU | BPF_MOD | BPF_X:
378 	case BPF_ALU64 | BPF_MOD | BPF_X:
379 	{
380 		const u8 r0 = bpf2a64[BPF_REG_0];
381 
382 		/* if (src == 0) return 0 */
383 		jmp_offset = 3; /* skip ahead to else path */
384 		check_imm19(jmp_offset);
385 		emit(A64_CBNZ(is64, src, jmp_offset), ctx);
386 		emit(A64_MOVZ(1, r0, 0, 0), ctx);
387 		jmp_offset = epilogue_offset(ctx);
388 		check_imm26(jmp_offset);
389 		emit(A64_B(jmp_offset), ctx);
390 		/* else */
391 		switch (BPF_OP(code)) {
392 		case BPF_DIV:
393 			emit(A64_UDIV(is64, dst, dst, src), ctx);
394 			break;
395 		case BPF_MOD:
396 			emit(A64_UDIV(is64, tmp, dst, src), ctx);
397 			emit(A64_MUL(is64, tmp, tmp, src), ctx);
398 			emit(A64_SUB(is64, dst, dst, tmp), ctx);
399 			break;
400 		}
401 		break;
402 	}
403 	case BPF_ALU | BPF_LSH | BPF_X:
404 	case BPF_ALU64 | BPF_LSH | BPF_X:
405 		emit(A64_LSLV(is64, dst, dst, src), ctx);
406 		break;
407 	case BPF_ALU | BPF_RSH | BPF_X:
408 	case BPF_ALU64 | BPF_RSH | BPF_X:
409 		emit(A64_LSRV(is64, dst, dst, src), ctx);
410 		break;
411 	case BPF_ALU | BPF_ARSH | BPF_X:
412 	case BPF_ALU64 | BPF_ARSH | BPF_X:
413 		emit(A64_ASRV(is64, dst, dst, src), ctx);
414 		break;
415 	/* dst = -dst */
416 	case BPF_ALU | BPF_NEG:
417 	case BPF_ALU64 | BPF_NEG:
418 		emit(A64_NEG(is64, dst, dst), ctx);
419 		break;
420 	/* dst = BSWAP##imm(dst) */
421 	case BPF_ALU | BPF_END | BPF_FROM_LE:
422 	case BPF_ALU | BPF_END | BPF_FROM_BE:
423 #ifdef CONFIG_CPU_BIG_ENDIAN
424 		if (BPF_SRC(code) == BPF_FROM_BE)
425 			goto emit_bswap_uxt;
426 #else /* !CONFIG_CPU_BIG_ENDIAN */
427 		if (BPF_SRC(code) == BPF_FROM_LE)
428 			goto emit_bswap_uxt;
429 #endif
430 		switch (imm) {
431 		case 16:
432 			emit(A64_REV16(is64, dst, dst), ctx);
433 			/* zero-extend 16 bits into 64 bits */
434 			emit(A64_UXTH(is64, dst, dst), ctx);
435 			break;
436 		case 32:
437 			emit(A64_REV32(is64, dst, dst), ctx);
438 			/* upper 32 bits already cleared */
439 			break;
440 		case 64:
441 			emit(A64_REV64(dst, dst), ctx);
442 			break;
443 		}
444 		break;
445 emit_bswap_uxt:
446 		switch (imm) {
447 		case 16:
448 			/* zero-extend 16 bits into 64 bits */
449 			emit(A64_UXTH(is64, dst, dst), ctx);
450 			break;
451 		case 32:
452 			/* zero-extend 32 bits into 64 bits */
453 			emit(A64_UXTW(is64, dst, dst), ctx);
454 			break;
455 		case 64:
456 			/* nop */
457 			break;
458 		}
459 		break;
460 	/* dst = imm */
461 	case BPF_ALU | BPF_MOV | BPF_K:
462 	case BPF_ALU64 | BPF_MOV | BPF_K:
463 		emit_a64_mov_i(is64, dst, imm, ctx);
464 		break;
465 	/* dst = dst OP imm */
466 	case BPF_ALU | BPF_ADD | BPF_K:
467 	case BPF_ALU64 | BPF_ADD | BPF_K:
468 		emit_a64_mov_i(is64, tmp, imm, ctx);
469 		emit(A64_ADD(is64, dst, dst, tmp), ctx);
470 		break;
471 	case BPF_ALU | BPF_SUB | BPF_K:
472 	case BPF_ALU64 | BPF_SUB | BPF_K:
473 		emit_a64_mov_i(is64, tmp, imm, ctx);
474 		emit(A64_SUB(is64, dst, dst, tmp), ctx);
475 		break;
476 	case BPF_ALU | BPF_AND | BPF_K:
477 	case BPF_ALU64 | BPF_AND | BPF_K:
478 		emit_a64_mov_i(is64, tmp, imm, ctx);
479 		emit(A64_AND(is64, dst, dst, tmp), ctx);
480 		break;
481 	case BPF_ALU | BPF_OR | BPF_K:
482 	case BPF_ALU64 | BPF_OR | BPF_K:
483 		emit_a64_mov_i(is64, tmp, imm, ctx);
484 		emit(A64_ORR(is64, dst, dst, tmp), ctx);
485 		break;
486 	case BPF_ALU | BPF_XOR | BPF_K:
487 	case BPF_ALU64 | BPF_XOR | BPF_K:
488 		emit_a64_mov_i(is64, tmp, imm, ctx);
489 		emit(A64_EOR(is64, dst, dst, tmp), ctx);
490 		break;
491 	case BPF_ALU | BPF_MUL | BPF_K:
492 	case BPF_ALU64 | BPF_MUL | BPF_K:
493 		emit_a64_mov_i(is64, tmp, imm, ctx);
494 		emit(A64_MUL(is64, dst, dst, tmp), ctx);
495 		break;
496 	case BPF_ALU | BPF_DIV | BPF_K:
497 	case BPF_ALU64 | BPF_DIV | BPF_K:
498 		emit_a64_mov_i(is64, tmp, imm, ctx);
499 		emit(A64_UDIV(is64, dst, dst, tmp), ctx);
500 		break;
501 	case BPF_ALU | BPF_MOD | BPF_K:
502 	case BPF_ALU64 | BPF_MOD | BPF_K:
503 		emit_a64_mov_i(is64, tmp2, imm, ctx);
504 		emit(A64_UDIV(is64, tmp, dst, tmp2), ctx);
505 		emit(A64_MUL(is64, tmp, tmp, tmp2), ctx);
506 		emit(A64_SUB(is64, dst, dst, tmp), ctx);
507 		break;
508 	case BPF_ALU | BPF_LSH | BPF_K:
509 	case BPF_ALU64 | BPF_LSH | BPF_K:
510 		emit(A64_LSL(is64, dst, dst, imm), ctx);
511 		break;
512 	case BPF_ALU | BPF_RSH | BPF_K:
513 	case BPF_ALU64 | BPF_RSH | BPF_K:
514 		emit(A64_LSR(is64, dst, dst, imm), ctx);
515 		break;
516 	case BPF_ALU | BPF_ARSH | BPF_K:
517 	case BPF_ALU64 | BPF_ARSH | BPF_K:
518 		emit(A64_ASR(is64, dst, dst, imm), ctx);
519 		break;
520 
521 	/* JUMP off */
522 	case BPF_JMP | BPF_JA:
523 		jmp_offset = bpf2a64_offset(i + off, i, ctx);
524 		check_imm26(jmp_offset);
525 		emit(A64_B(jmp_offset), ctx);
526 		break;
527 	/* IF (dst COND src) JUMP off */
528 	case BPF_JMP | BPF_JEQ | BPF_X:
529 	case BPF_JMP | BPF_JGT | BPF_X:
530 	case BPF_JMP | BPF_JLT | BPF_X:
531 	case BPF_JMP | BPF_JGE | BPF_X:
532 	case BPF_JMP | BPF_JLE | BPF_X:
533 	case BPF_JMP | BPF_JNE | BPF_X:
534 	case BPF_JMP | BPF_JSGT | BPF_X:
535 	case BPF_JMP | BPF_JSLT | BPF_X:
536 	case BPF_JMP | BPF_JSGE | BPF_X:
537 	case BPF_JMP | BPF_JSLE | BPF_X:
538 		emit(A64_CMP(1, dst, src), ctx);
539 emit_cond_jmp:
540 		jmp_offset = bpf2a64_offset(i + off, i, ctx);
541 		check_imm19(jmp_offset);
542 		switch (BPF_OP(code)) {
543 		case BPF_JEQ:
544 			jmp_cond = A64_COND_EQ;
545 			break;
546 		case BPF_JGT:
547 			jmp_cond = A64_COND_HI;
548 			break;
549 		case BPF_JLT:
550 			jmp_cond = A64_COND_CC;
551 			break;
552 		case BPF_JGE:
553 			jmp_cond = A64_COND_CS;
554 			break;
555 		case BPF_JLE:
556 			jmp_cond = A64_COND_LS;
557 			break;
558 		case BPF_JSET:
559 		case BPF_JNE:
560 			jmp_cond = A64_COND_NE;
561 			break;
562 		case BPF_JSGT:
563 			jmp_cond = A64_COND_GT;
564 			break;
565 		case BPF_JSLT:
566 			jmp_cond = A64_COND_LT;
567 			break;
568 		case BPF_JSGE:
569 			jmp_cond = A64_COND_GE;
570 			break;
571 		case BPF_JSLE:
572 			jmp_cond = A64_COND_LE;
573 			break;
574 		default:
575 			return -EFAULT;
576 		}
577 		emit(A64_B_(jmp_cond, jmp_offset), ctx);
578 		break;
579 	case BPF_JMP | BPF_JSET | BPF_X:
580 		emit(A64_TST(1, dst, src), ctx);
581 		goto emit_cond_jmp;
582 	/* IF (dst COND imm) JUMP off */
583 	case BPF_JMP | BPF_JEQ | BPF_K:
584 	case BPF_JMP | BPF_JGT | BPF_K:
585 	case BPF_JMP | BPF_JLT | BPF_K:
586 	case BPF_JMP | BPF_JGE | BPF_K:
587 	case BPF_JMP | BPF_JLE | BPF_K:
588 	case BPF_JMP | BPF_JNE | BPF_K:
589 	case BPF_JMP | BPF_JSGT | BPF_K:
590 	case BPF_JMP | BPF_JSLT | BPF_K:
591 	case BPF_JMP | BPF_JSGE | BPF_K:
592 	case BPF_JMP | BPF_JSLE | BPF_K:
593 		emit_a64_mov_i(1, tmp, imm, ctx);
594 		emit(A64_CMP(1, dst, tmp), ctx);
595 		goto emit_cond_jmp;
596 	case BPF_JMP | BPF_JSET | BPF_K:
597 		emit_a64_mov_i(1, tmp, imm, ctx);
598 		emit(A64_TST(1, dst, tmp), ctx);
599 		goto emit_cond_jmp;
600 	/* function call */
601 	case BPF_JMP | BPF_CALL:
602 	{
603 		const u8 r0 = bpf2a64[BPF_REG_0];
604 		const u64 func = (u64)__bpf_call_base + imm;
605 
606 		emit_a64_mov_i64(tmp, func, ctx);
607 		emit(A64_BLR(tmp), ctx);
608 		emit(A64_MOV(1, r0, A64_R(0)), ctx);
609 		break;
610 	}
611 	/* tail call */
612 	case BPF_JMP | BPF_TAIL_CALL:
613 		if (emit_bpf_tail_call(ctx))
614 			return -EFAULT;
615 		break;
616 	/* function return */
617 	case BPF_JMP | BPF_EXIT:
618 		/* Optimization: when last instruction is EXIT,
619 		   simply fallthrough to epilogue. */
620 		if (i == ctx->prog->len - 1)
621 			break;
622 		jmp_offset = epilogue_offset(ctx);
623 		check_imm26(jmp_offset);
624 		emit(A64_B(jmp_offset), ctx);
625 		break;
626 
627 	/* dst = imm64 */
628 	case BPF_LD | BPF_IMM | BPF_DW:
629 	{
630 		const struct bpf_insn insn1 = insn[1];
631 		u64 imm64;
632 
633 		imm64 = (u64)insn1.imm << 32 | (u32)imm;
634 		emit_a64_mov_i64(dst, imm64, ctx);
635 
636 		return 1;
637 	}
638 
639 	/* LDX: dst = *(size *)(src + off) */
640 	case BPF_LDX | BPF_MEM | BPF_W:
641 	case BPF_LDX | BPF_MEM | BPF_H:
642 	case BPF_LDX | BPF_MEM | BPF_B:
643 	case BPF_LDX | BPF_MEM | BPF_DW:
644 		emit_a64_mov_i(1, tmp, off, ctx);
645 		switch (BPF_SIZE(code)) {
646 		case BPF_W:
647 			emit(A64_LDR32(dst, src, tmp), ctx);
648 			break;
649 		case BPF_H:
650 			emit(A64_LDRH(dst, src, tmp), ctx);
651 			break;
652 		case BPF_B:
653 			emit(A64_LDRB(dst, src, tmp), ctx);
654 			break;
655 		case BPF_DW:
656 			emit(A64_LDR64(dst, src, tmp), ctx);
657 			break;
658 		}
659 		break;
660 
661 	/* ST: *(size *)(dst + off) = imm */
662 	case BPF_ST | BPF_MEM | BPF_W:
663 	case BPF_ST | BPF_MEM | BPF_H:
664 	case BPF_ST | BPF_MEM | BPF_B:
665 	case BPF_ST | BPF_MEM | BPF_DW:
666 		/* Load imm to a register then store it */
667 		emit_a64_mov_i(1, tmp2, off, ctx);
668 		emit_a64_mov_i(1, tmp, imm, ctx);
669 		switch (BPF_SIZE(code)) {
670 		case BPF_W:
671 			emit(A64_STR32(tmp, dst, tmp2), ctx);
672 			break;
673 		case BPF_H:
674 			emit(A64_STRH(tmp, dst, tmp2), ctx);
675 			break;
676 		case BPF_B:
677 			emit(A64_STRB(tmp, dst, tmp2), ctx);
678 			break;
679 		case BPF_DW:
680 			emit(A64_STR64(tmp, dst, tmp2), ctx);
681 			break;
682 		}
683 		break;
684 
685 	/* STX: *(size *)(dst + off) = src */
686 	case BPF_STX | BPF_MEM | BPF_W:
687 	case BPF_STX | BPF_MEM | BPF_H:
688 	case BPF_STX | BPF_MEM | BPF_B:
689 	case BPF_STX | BPF_MEM | BPF_DW:
690 		emit_a64_mov_i(1, tmp, off, ctx);
691 		switch (BPF_SIZE(code)) {
692 		case BPF_W:
693 			emit(A64_STR32(src, dst, tmp), ctx);
694 			break;
695 		case BPF_H:
696 			emit(A64_STRH(src, dst, tmp), ctx);
697 			break;
698 		case BPF_B:
699 			emit(A64_STRB(src, dst, tmp), ctx);
700 			break;
701 		case BPF_DW:
702 			emit(A64_STR64(src, dst, tmp), ctx);
703 			break;
704 		}
705 		break;
706 	/* STX XADD: lock *(u32 *)(dst + off) += src */
707 	case BPF_STX | BPF_XADD | BPF_W:
708 	/* STX XADD: lock *(u64 *)(dst + off) += src */
709 	case BPF_STX | BPF_XADD | BPF_DW:
710 		emit_a64_mov_i(1, tmp, off, ctx);
711 		emit(A64_ADD(1, tmp, tmp, dst), ctx);
712 		emit(A64_PRFM(tmp, PST, L1, STRM), ctx);
713 		emit(A64_LDXR(isdw, tmp2, tmp), ctx);
714 		emit(A64_ADD(isdw, tmp2, tmp2, src), ctx);
715 		emit(A64_STXR(isdw, tmp2, tmp, tmp3), ctx);
716 		jmp_offset = -3;
717 		check_imm19(jmp_offset);
718 		emit(A64_CBNZ(0, tmp3, jmp_offset), ctx);
719 		break;
720 
721 	/* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + imm)) */
722 	case BPF_LD | BPF_ABS | BPF_W:
723 	case BPF_LD | BPF_ABS | BPF_H:
724 	case BPF_LD | BPF_ABS | BPF_B:
725 	/* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + src + imm)) */
726 	case BPF_LD | BPF_IND | BPF_W:
727 	case BPF_LD | BPF_IND | BPF_H:
728 	case BPF_LD | BPF_IND | BPF_B:
729 	{
730 		const u8 r0 = bpf2a64[BPF_REG_0]; /* r0 = return value */
731 		const u8 r6 = bpf2a64[BPF_REG_6]; /* r6 = pointer to sk_buff */
732 		const u8 fp = bpf2a64[BPF_REG_FP];
733 		const u8 r1 = bpf2a64[BPF_REG_1]; /* r1: struct sk_buff *skb */
734 		const u8 r2 = bpf2a64[BPF_REG_2]; /* r2: int k */
735 		const u8 r3 = bpf2a64[BPF_REG_3]; /* r3: unsigned int size */
736 		const u8 r4 = bpf2a64[BPF_REG_4]; /* r4: void *buffer */
737 		const u8 r5 = bpf2a64[BPF_REG_5]; /* r5: void *(*func)(...) */
738 		int size;
739 
740 		emit(A64_MOV(1, r1, r6), ctx);
741 		emit_a64_mov_i(0, r2, imm, ctx);
742 		if (BPF_MODE(code) == BPF_IND)
743 			emit(A64_ADD(0, r2, r2, src), ctx);
744 		switch (BPF_SIZE(code)) {
745 		case BPF_W:
746 			size = 4;
747 			break;
748 		case BPF_H:
749 			size = 2;
750 			break;
751 		case BPF_B:
752 			size = 1;
753 			break;
754 		default:
755 			return -EINVAL;
756 		}
757 		emit_a64_mov_i64(r3, size, ctx);
758 		emit(A64_SUB_I(1, r4, fp, ctx->stack_size), ctx);
759 		emit_a64_mov_i64(r5, (unsigned long)bpf_load_pointer, ctx);
760 		emit(A64_BLR(r5), ctx);
761 		emit(A64_MOV(1, r0, A64_R(0)), ctx);
762 
763 		jmp_offset = epilogue_offset(ctx);
764 		check_imm19(jmp_offset);
765 		emit(A64_CBZ(1, r0, jmp_offset), ctx);
766 		emit(A64_MOV(1, r5, r0), ctx);
767 		switch (BPF_SIZE(code)) {
768 		case BPF_W:
769 			emit(A64_LDR32(r0, r5, A64_ZR), ctx);
770 #ifndef CONFIG_CPU_BIG_ENDIAN
771 			emit(A64_REV32(0, r0, r0), ctx);
772 #endif
773 			break;
774 		case BPF_H:
775 			emit(A64_LDRH(r0, r5, A64_ZR), ctx);
776 #ifndef CONFIG_CPU_BIG_ENDIAN
777 			emit(A64_REV16(0, r0, r0), ctx);
778 #endif
779 			break;
780 		case BPF_B:
781 			emit(A64_LDRB(r0, r5, A64_ZR), ctx);
782 			break;
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 
794 static int build_body(struct jit_ctx *ctx)
795 {
796 	const struct bpf_prog *prog = ctx->prog;
797 	int i;
798 
799 	for (i = 0; i < prog->len; i++) {
800 		const struct bpf_insn *insn = &prog->insnsi[i];
801 		int ret;
802 
803 		ret = build_insn(insn, ctx);
804 		if (ret > 0) {
805 			i++;
806 			if (ctx->image == NULL)
807 				ctx->offset[i] = ctx->idx;
808 			continue;
809 		}
810 		if (ctx->image == NULL)
811 			ctx->offset[i] = ctx->idx;
812 		if (ret)
813 			return ret;
814 	}
815 
816 	return 0;
817 }
818 
819 static int validate_code(struct jit_ctx *ctx)
820 {
821 	int i;
822 
823 	for (i = 0; i < ctx->idx; i++) {
824 		u32 a64_insn = le32_to_cpu(ctx->image[i]);
825 
826 		if (a64_insn == AARCH64_BREAK_FAULT)
827 			return -1;
828 	}
829 
830 	return 0;
831 }
832 
833 static inline void bpf_flush_icache(void *start, void *end)
834 {
835 	flush_icache_range((unsigned long)start, (unsigned long)end);
836 }
837 
838 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
839 {
840 	struct bpf_prog *tmp, *orig_prog = prog;
841 	struct bpf_binary_header *header;
842 	bool tmp_blinded = false;
843 	struct jit_ctx ctx;
844 	int image_size;
845 	u8 *image_ptr;
846 
847 	if (!bpf_jit_enable)
848 		return orig_prog;
849 
850 	tmp = bpf_jit_blind_constants(prog);
851 	/* If blinding was requested and we failed during blinding,
852 	 * we must fall back to the interpreter.
853 	 */
854 	if (IS_ERR(tmp))
855 		return orig_prog;
856 	if (tmp != prog) {
857 		tmp_blinded = true;
858 		prog = tmp;
859 	}
860 
861 	memset(&ctx, 0, sizeof(ctx));
862 	ctx.prog = prog;
863 
864 	ctx.offset = kcalloc(prog->len, sizeof(int), GFP_KERNEL);
865 	if (ctx.offset == NULL) {
866 		prog = orig_prog;
867 		goto out;
868 	}
869 
870 	/* 1. Initial fake pass to compute ctx->idx. */
871 
872 	/* Fake pass to fill in ctx->offset. */
873 	if (build_body(&ctx)) {
874 		prog = orig_prog;
875 		goto out_off;
876 	}
877 
878 	if (build_prologue(&ctx)) {
879 		prog = orig_prog;
880 		goto out_off;
881 	}
882 
883 	ctx.epilogue_offset = ctx.idx;
884 	build_epilogue(&ctx);
885 
886 	/* Now we know the actual image size. */
887 	image_size = sizeof(u32) * ctx.idx;
888 	header = bpf_jit_binary_alloc(image_size, &image_ptr,
889 				      sizeof(u32), jit_fill_hole);
890 	if (header == NULL) {
891 		prog = orig_prog;
892 		goto out_off;
893 	}
894 
895 	/* 2. Now, the actual pass. */
896 
897 	ctx.image = (__le32 *)image_ptr;
898 	ctx.idx = 0;
899 
900 	build_prologue(&ctx);
901 
902 	if (build_body(&ctx)) {
903 		bpf_jit_binary_free(header);
904 		prog = orig_prog;
905 		goto out_off;
906 	}
907 
908 	build_epilogue(&ctx);
909 
910 	/* 3. Extra pass to validate JITed code. */
911 	if (validate_code(&ctx)) {
912 		bpf_jit_binary_free(header);
913 		prog = orig_prog;
914 		goto out_off;
915 	}
916 
917 	/* And we're done. */
918 	if (bpf_jit_enable > 1)
919 		bpf_jit_dump(prog->len, image_size, 2, ctx.image);
920 
921 	bpf_flush_icache(header, ctx.image + ctx.idx);
922 
923 	bpf_jit_binary_lock_ro(header);
924 	prog->bpf_func = (void *)ctx.image;
925 	prog->jited = 1;
926 	prog->jited_len = image_size;
927 
928 out_off:
929 	kfree(ctx.offset);
930 out:
931 	if (tmp_blinded)
932 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
933 					   tmp : orig_prog);
934 	return prog;
935 }
936