xref: /openbmc/linux/arch/arm64/net/bpf_jit_comp.c (revision 6774def6)
1 /*
2  * BPF JIT compiler for ARM64
3  *
4  * Copyright (C) 2014 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/filter.h>
22 #include <linux/printk.h>
23 #include <linux/skbuff.h>
24 #include <linux/slab.h>
25 
26 #include <asm/byteorder.h>
27 #include <asm/cacheflush.h>
28 #include <asm/debug-monitors.h>
29 
30 #include "bpf_jit.h"
31 
32 int bpf_jit_enable __read_mostly;
33 
34 #define TMP_REG_1 (MAX_BPF_REG + 0)
35 #define TMP_REG_2 (MAX_BPF_REG + 1)
36 
37 /* Map BPF registers to A64 registers */
38 static const int bpf2a64[] = {
39 	/* return value from in-kernel function, and exit value from eBPF */
40 	[BPF_REG_0] = A64_R(7),
41 	/* arguments from eBPF program to in-kernel function */
42 	[BPF_REG_1] = A64_R(0),
43 	[BPF_REG_2] = A64_R(1),
44 	[BPF_REG_3] = A64_R(2),
45 	[BPF_REG_4] = A64_R(3),
46 	[BPF_REG_5] = A64_R(4),
47 	/* callee saved registers that in-kernel function will preserve */
48 	[BPF_REG_6] = A64_R(19),
49 	[BPF_REG_7] = A64_R(20),
50 	[BPF_REG_8] = A64_R(21),
51 	[BPF_REG_9] = A64_R(22),
52 	/* read-only frame pointer to access stack */
53 	[BPF_REG_FP] = A64_FP,
54 	/* temporary register for internal BPF JIT */
55 	[TMP_REG_1] = A64_R(23),
56 	[TMP_REG_2] = A64_R(24),
57 };
58 
59 struct jit_ctx {
60 	const struct bpf_prog *prog;
61 	int idx;
62 	int tmp_used;
63 	int body_offset;
64 	int *offset;
65 	u32 *image;
66 };
67 
68 static inline void emit(const u32 insn, struct jit_ctx *ctx)
69 {
70 	if (ctx->image != NULL)
71 		ctx->image[ctx->idx] = cpu_to_le32(insn);
72 
73 	ctx->idx++;
74 }
75 
76 static inline void emit_a64_mov_i64(const int reg, const u64 val,
77 				    struct jit_ctx *ctx)
78 {
79 	u64 tmp = val;
80 	int shift = 0;
81 
82 	emit(A64_MOVZ(1, reg, tmp & 0xffff, shift), ctx);
83 	tmp >>= 16;
84 	shift += 16;
85 	while (tmp) {
86 		if (tmp & 0xffff)
87 			emit(A64_MOVK(1, reg, tmp & 0xffff, shift), ctx);
88 		tmp >>= 16;
89 		shift += 16;
90 	}
91 }
92 
93 static inline void emit_a64_mov_i(const int is64, const int reg,
94 				  const s32 val, struct jit_ctx *ctx)
95 {
96 	u16 hi = val >> 16;
97 	u16 lo = val & 0xffff;
98 
99 	if (hi & 0x8000) {
100 		if (hi == 0xffff) {
101 			emit(A64_MOVN(is64, reg, (u16)~lo, 0), ctx);
102 		} else {
103 			emit(A64_MOVN(is64, reg, (u16)~hi, 16), ctx);
104 			emit(A64_MOVK(is64, reg, lo, 0), ctx);
105 		}
106 	} else {
107 		emit(A64_MOVZ(is64, reg, lo, 0), ctx);
108 		if (hi)
109 			emit(A64_MOVK(is64, reg, hi, 16), ctx);
110 	}
111 }
112 
113 static inline int bpf2a64_offset(int bpf_to, int bpf_from,
114 				 const struct jit_ctx *ctx)
115 {
116 	int to = ctx->offset[bpf_to + 1];
117 	/* -1 to account for the Branch instruction */
118 	int from = ctx->offset[bpf_from + 1] - 1;
119 
120 	return to - from;
121 }
122 
123 static void jit_fill_hole(void *area, unsigned int size)
124 {
125 	u32 *ptr;
126 	/* We are guaranteed to have aligned memory. */
127 	for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
128 		*ptr++ = cpu_to_le32(AARCH64_BREAK_FAULT);
129 }
130 
131 static inline int epilogue_offset(const struct jit_ctx *ctx)
132 {
133 	int to = ctx->offset[ctx->prog->len - 1];
134 	int from = ctx->idx - ctx->body_offset;
135 
136 	return to - from;
137 }
138 
139 /* Stack must be multiples of 16B */
140 #define STACK_ALIGN(sz) (((sz) + 15) & ~15)
141 
142 static void build_prologue(struct jit_ctx *ctx)
143 {
144 	const u8 r6 = bpf2a64[BPF_REG_6];
145 	const u8 r7 = bpf2a64[BPF_REG_7];
146 	const u8 r8 = bpf2a64[BPF_REG_8];
147 	const u8 r9 = bpf2a64[BPF_REG_9];
148 	const u8 fp = bpf2a64[BPF_REG_FP];
149 	const u8 ra = bpf2a64[BPF_REG_A];
150 	const u8 rx = bpf2a64[BPF_REG_X];
151 	const u8 tmp1 = bpf2a64[TMP_REG_1];
152 	const u8 tmp2 = bpf2a64[TMP_REG_2];
153 	int stack_size = MAX_BPF_STACK;
154 
155 	stack_size += 4; /* extra for skb_copy_bits buffer */
156 	stack_size = STACK_ALIGN(stack_size);
157 
158 	/* Save callee-saved register */
159 	emit(A64_PUSH(r6, r7, A64_SP), ctx);
160 	emit(A64_PUSH(r8, r9, A64_SP), ctx);
161 	if (ctx->tmp_used)
162 		emit(A64_PUSH(tmp1, tmp2, A64_SP), ctx);
163 
164 	/* Set up BPF stack */
165 	emit(A64_SUB_I(1, A64_SP, A64_SP, stack_size), ctx);
166 
167 	/* Set up frame pointer */
168 	emit(A64_MOV(1, fp, A64_SP), ctx);
169 
170 	/* Clear registers A and X */
171 	emit_a64_mov_i64(ra, 0, ctx);
172 	emit_a64_mov_i64(rx, 0, ctx);
173 }
174 
175 static void build_epilogue(struct jit_ctx *ctx)
176 {
177 	const u8 r0 = bpf2a64[BPF_REG_0];
178 	const u8 r6 = bpf2a64[BPF_REG_6];
179 	const u8 r7 = bpf2a64[BPF_REG_7];
180 	const u8 r8 = bpf2a64[BPF_REG_8];
181 	const u8 r9 = bpf2a64[BPF_REG_9];
182 	const u8 fp = bpf2a64[BPF_REG_FP];
183 	const u8 tmp1 = bpf2a64[TMP_REG_1];
184 	const u8 tmp2 = bpf2a64[TMP_REG_2];
185 	int stack_size = MAX_BPF_STACK;
186 
187 	stack_size += 4; /* extra for skb_copy_bits buffer */
188 	stack_size = STACK_ALIGN(stack_size);
189 
190 	/* We're done with BPF stack */
191 	emit(A64_ADD_I(1, A64_SP, A64_SP, stack_size), ctx);
192 
193 	/* Restore callee-saved register */
194 	if (ctx->tmp_used)
195 		emit(A64_POP(tmp1, tmp2, A64_SP), ctx);
196 	emit(A64_POP(r8, r9, A64_SP), ctx);
197 	emit(A64_POP(r6, r7, A64_SP), ctx);
198 
199 	/* Restore frame pointer */
200 	emit(A64_MOV(1, fp, A64_SP), ctx);
201 
202 	/* Set return value */
203 	emit(A64_MOV(1, A64_R(0), r0), ctx);
204 
205 	emit(A64_RET(A64_LR), ctx);
206 }
207 
208 /* JITs an eBPF instruction.
209  * Returns:
210  * 0  - successfully JITed an 8-byte eBPF instruction.
211  * >0 - successfully JITed a 16-byte eBPF instruction.
212  * <0 - failed to JIT.
213  */
214 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx)
215 {
216 	const u8 code = insn->code;
217 	const u8 dst = bpf2a64[insn->dst_reg];
218 	const u8 src = bpf2a64[insn->src_reg];
219 	const u8 tmp = bpf2a64[TMP_REG_1];
220 	const u8 tmp2 = bpf2a64[TMP_REG_2];
221 	const s16 off = insn->off;
222 	const s32 imm = insn->imm;
223 	const int i = insn - ctx->prog->insnsi;
224 	const bool is64 = BPF_CLASS(code) == BPF_ALU64;
225 	u8 jmp_cond;
226 	s32 jmp_offset;
227 
228 	switch (code) {
229 	/* dst = src */
230 	case BPF_ALU | BPF_MOV | BPF_X:
231 	case BPF_ALU64 | BPF_MOV | BPF_X:
232 		emit(A64_MOV(is64, dst, src), ctx);
233 		break;
234 	/* dst = dst OP src */
235 	case BPF_ALU | BPF_ADD | BPF_X:
236 	case BPF_ALU64 | BPF_ADD | BPF_X:
237 		emit(A64_ADD(is64, dst, dst, src), ctx);
238 		break;
239 	case BPF_ALU | BPF_SUB | BPF_X:
240 	case BPF_ALU64 | BPF_SUB | BPF_X:
241 		emit(A64_SUB(is64, dst, dst, src), ctx);
242 		break;
243 	case BPF_ALU | BPF_AND | BPF_X:
244 	case BPF_ALU64 | BPF_AND | BPF_X:
245 		emit(A64_AND(is64, dst, dst, src), ctx);
246 		break;
247 	case BPF_ALU | BPF_OR | BPF_X:
248 	case BPF_ALU64 | BPF_OR | BPF_X:
249 		emit(A64_ORR(is64, dst, dst, src), ctx);
250 		break;
251 	case BPF_ALU | BPF_XOR | BPF_X:
252 	case BPF_ALU64 | BPF_XOR | BPF_X:
253 		emit(A64_EOR(is64, dst, dst, src), ctx);
254 		break;
255 	case BPF_ALU | BPF_MUL | BPF_X:
256 	case BPF_ALU64 | BPF_MUL | BPF_X:
257 		emit(A64_MUL(is64, dst, dst, src), ctx);
258 		break;
259 	case BPF_ALU | BPF_DIV | BPF_X:
260 	case BPF_ALU64 | BPF_DIV | BPF_X:
261 		emit(A64_UDIV(is64, dst, dst, src), ctx);
262 		break;
263 	case BPF_ALU | BPF_MOD | BPF_X:
264 	case BPF_ALU64 | BPF_MOD | BPF_X:
265 		ctx->tmp_used = 1;
266 		emit(A64_UDIV(is64, tmp, dst, src), ctx);
267 		emit(A64_MUL(is64, tmp, tmp, src), ctx);
268 		emit(A64_SUB(is64, dst, dst, tmp), ctx);
269 		break;
270 	case BPF_ALU | BPF_LSH | BPF_X:
271 	case BPF_ALU64 | BPF_LSH | BPF_X:
272 		emit(A64_LSLV(is64, dst, dst, src), ctx);
273 		break;
274 	case BPF_ALU | BPF_RSH | BPF_X:
275 	case BPF_ALU64 | BPF_RSH | BPF_X:
276 		emit(A64_LSRV(is64, dst, dst, src), ctx);
277 		break;
278 	case BPF_ALU | BPF_ARSH | BPF_X:
279 	case BPF_ALU64 | BPF_ARSH | BPF_X:
280 		emit(A64_ASRV(is64, dst, dst, src), ctx);
281 		break;
282 	/* dst = -dst */
283 	case BPF_ALU | BPF_NEG:
284 	case BPF_ALU64 | BPF_NEG:
285 		emit(A64_NEG(is64, dst, dst), ctx);
286 		break;
287 	/* dst = BSWAP##imm(dst) */
288 	case BPF_ALU | BPF_END | BPF_FROM_LE:
289 	case BPF_ALU | BPF_END | BPF_FROM_BE:
290 #ifdef CONFIG_CPU_BIG_ENDIAN
291 		if (BPF_SRC(code) == BPF_FROM_BE)
292 			break;
293 #else /* !CONFIG_CPU_BIG_ENDIAN */
294 		if (BPF_SRC(code) == BPF_FROM_LE)
295 			break;
296 #endif
297 		switch (imm) {
298 		case 16:
299 			emit(A64_REV16(is64, dst, dst), ctx);
300 			break;
301 		case 32:
302 			emit(A64_REV32(is64, dst, dst), ctx);
303 			break;
304 		case 64:
305 			emit(A64_REV64(dst, dst), ctx);
306 			break;
307 		}
308 		break;
309 	/* dst = imm */
310 	case BPF_ALU | BPF_MOV | BPF_K:
311 	case BPF_ALU64 | BPF_MOV | BPF_K:
312 		emit_a64_mov_i(is64, dst, imm, ctx);
313 		break;
314 	/* dst = dst OP imm */
315 	case BPF_ALU | BPF_ADD | BPF_K:
316 	case BPF_ALU64 | BPF_ADD | BPF_K:
317 		ctx->tmp_used = 1;
318 		emit_a64_mov_i(is64, tmp, imm, ctx);
319 		emit(A64_ADD(is64, dst, dst, tmp), ctx);
320 		break;
321 	case BPF_ALU | BPF_SUB | BPF_K:
322 	case BPF_ALU64 | BPF_SUB | BPF_K:
323 		ctx->tmp_used = 1;
324 		emit_a64_mov_i(is64, tmp, imm, ctx);
325 		emit(A64_SUB(is64, dst, dst, tmp), ctx);
326 		break;
327 	case BPF_ALU | BPF_AND | BPF_K:
328 	case BPF_ALU64 | BPF_AND | BPF_K:
329 		ctx->tmp_used = 1;
330 		emit_a64_mov_i(is64, tmp, imm, ctx);
331 		emit(A64_AND(is64, dst, dst, tmp), ctx);
332 		break;
333 	case BPF_ALU | BPF_OR | BPF_K:
334 	case BPF_ALU64 | BPF_OR | BPF_K:
335 		ctx->tmp_used = 1;
336 		emit_a64_mov_i(is64, tmp, imm, ctx);
337 		emit(A64_ORR(is64, dst, dst, tmp), ctx);
338 		break;
339 	case BPF_ALU | BPF_XOR | BPF_K:
340 	case BPF_ALU64 | BPF_XOR | BPF_K:
341 		ctx->tmp_used = 1;
342 		emit_a64_mov_i(is64, tmp, imm, ctx);
343 		emit(A64_EOR(is64, dst, dst, tmp), ctx);
344 		break;
345 	case BPF_ALU | BPF_MUL | BPF_K:
346 	case BPF_ALU64 | BPF_MUL | BPF_K:
347 		ctx->tmp_used = 1;
348 		emit_a64_mov_i(is64, tmp, imm, ctx);
349 		emit(A64_MUL(is64, dst, dst, tmp), ctx);
350 		break;
351 	case BPF_ALU | BPF_DIV | BPF_K:
352 	case BPF_ALU64 | BPF_DIV | BPF_K:
353 		ctx->tmp_used = 1;
354 		emit_a64_mov_i(is64, tmp, imm, ctx);
355 		emit(A64_UDIV(is64, dst, dst, tmp), ctx);
356 		break;
357 	case BPF_ALU | BPF_MOD | BPF_K:
358 	case BPF_ALU64 | BPF_MOD | BPF_K:
359 		ctx->tmp_used = 1;
360 		emit_a64_mov_i(is64, tmp2, imm, ctx);
361 		emit(A64_UDIV(is64, tmp, dst, tmp2), ctx);
362 		emit(A64_MUL(is64, tmp, tmp, tmp2), ctx);
363 		emit(A64_SUB(is64, dst, dst, tmp), ctx);
364 		break;
365 	case BPF_ALU | BPF_LSH | BPF_K:
366 	case BPF_ALU64 | BPF_LSH | BPF_K:
367 		emit(A64_LSL(is64, dst, dst, imm), ctx);
368 		break;
369 	case BPF_ALU | BPF_RSH | BPF_K:
370 	case BPF_ALU64 | BPF_RSH | BPF_K:
371 		emit(A64_LSR(is64, dst, dst, imm), ctx);
372 		break;
373 	case BPF_ALU | BPF_ARSH | BPF_K:
374 	case BPF_ALU64 | BPF_ARSH | BPF_K:
375 		emit(A64_ASR(is64, dst, dst, imm), ctx);
376 		break;
377 
378 #define check_imm(bits, imm) do {				\
379 	if ((((imm) > 0) && ((imm) >> (bits))) ||		\
380 	    (((imm) < 0) && (~(imm) >> (bits)))) {		\
381 		pr_info("[%2d] imm=%d(0x%x) out of range\n",	\
382 			i, imm, imm);				\
383 		return -EINVAL;					\
384 	}							\
385 } while (0)
386 #define check_imm19(imm) check_imm(19, imm)
387 #define check_imm26(imm) check_imm(26, imm)
388 
389 	/* JUMP off */
390 	case BPF_JMP | BPF_JA:
391 		jmp_offset = bpf2a64_offset(i + off, i, ctx);
392 		check_imm26(jmp_offset);
393 		emit(A64_B(jmp_offset), ctx);
394 		break;
395 	/* IF (dst COND src) JUMP off */
396 	case BPF_JMP | BPF_JEQ | BPF_X:
397 	case BPF_JMP | BPF_JGT | BPF_X:
398 	case BPF_JMP | BPF_JGE | BPF_X:
399 	case BPF_JMP | BPF_JNE | BPF_X:
400 	case BPF_JMP | BPF_JSGT | BPF_X:
401 	case BPF_JMP | BPF_JSGE | BPF_X:
402 		emit(A64_CMP(1, dst, src), ctx);
403 emit_cond_jmp:
404 		jmp_offset = bpf2a64_offset(i + off, i, ctx);
405 		check_imm19(jmp_offset);
406 		switch (BPF_OP(code)) {
407 		case BPF_JEQ:
408 			jmp_cond = A64_COND_EQ;
409 			break;
410 		case BPF_JGT:
411 			jmp_cond = A64_COND_HI;
412 			break;
413 		case BPF_JGE:
414 			jmp_cond = A64_COND_CS;
415 			break;
416 		case BPF_JNE:
417 			jmp_cond = A64_COND_NE;
418 			break;
419 		case BPF_JSGT:
420 			jmp_cond = A64_COND_GT;
421 			break;
422 		case BPF_JSGE:
423 			jmp_cond = A64_COND_GE;
424 			break;
425 		default:
426 			return -EFAULT;
427 		}
428 		emit(A64_B_(jmp_cond, jmp_offset), ctx);
429 		break;
430 	case BPF_JMP | BPF_JSET | BPF_X:
431 		emit(A64_TST(1, dst, src), ctx);
432 		goto emit_cond_jmp;
433 	/* IF (dst COND imm) JUMP off */
434 	case BPF_JMP | BPF_JEQ | BPF_K:
435 	case BPF_JMP | BPF_JGT | BPF_K:
436 	case BPF_JMP | BPF_JGE | BPF_K:
437 	case BPF_JMP | BPF_JNE | BPF_K:
438 	case BPF_JMP | BPF_JSGT | BPF_K:
439 	case BPF_JMP | BPF_JSGE | BPF_K:
440 		ctx->tmp_used = 1;
441 		emit_a64_mov_i(1, tmp, imm, ctx);
442 		emit(A64_CMP(1, dst, tmp), ctx);
443 		goto emit_cond_jmp;
444 	case BPF_JMP | BPF_JSET | BPF_K:
445 		ctx->tmp_used = 1;
446 		emit_a64_mov_i(1, tmp, imm, ctx);
447 		emit(A64_TST(1, dst, tmp), ctx);
448 		goto emit_cond_jmp;
449 	/* function call */
450 	case BPF_JMP | BPF_CALL:
451 	{
452 		const u8 r0 = bpf2a64[BPF_REG_0];
453 		const u64 func = (u64)__bpf_call_base + imm;
454 
455 		ctx->tmp_used = 1;
456 		emit_a64_mov_i64(tmp, func, ctx);
457 		emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx);
458 		emit(A64_MOV(1, A64_FP, A64_SP), ctx);
459 		emit(A64_BLR(tmp), ctx);
460 		emit(A64_MOV(1, r0, A64_R(0)), ctx);
461 		emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx);
462 		break;
463 	}
464 	/* function return */
465 	case BPF_JMP | BPF_EXIT:
466 		if (i == ctx->prog->len - 1)
467 			break;
468 		jmp_offset = epilogue_offset(ctx);
469 		check_imm26(jmp_offset);
470 		emit(A64_B(jmp_offset), ctx);
471 		break;
472 
473 	/* dst = imm64 */
474 	case BPF_LD | BPF_IMM | BPF_DW:
475 	{
476 		const struct bpf_insn insn1 = insn[1];
477 		u64 imm64;
478 
479 		if (insn1.code != 0 || insn1.src_reg != 0 ||
480 		    insn1.dst_reg != 0 || insn1.off != 0) {
481 			/* Note: verifier in BPF core must catch invalid
482 			 * instructions.
483 			 */
484 			pr_err_once("Invalid BPF_LD_IMM64 instruction\n");
485 			return -EINVAL;
486 		}
487 
488 		imm64 = (u64)insn1.imm << 32 | imm;
489 		emit_a64_mov_i64(dst, imm64, ctx);
490 
491 		return 1;
492 	}
493 
494 	/* LDX: dst = *(size *)(src + off) */
495 	case BPF_LDX | BPF_MEM | BPF_W:
496 	case BPF_LDX | BPF_MEM | BPF_H:
497 	case BPF_LDX | BPF_MEM | BPF_B:
498 	case BPF_LDX | BPF_MEM | BPF_DW:
499 		ctx->tmp_used = 1;
500 		emit_a64_mov_i(1, tmp, off, ctx);
501 		switch (BPF_SIZE(code)) {
502 		case BPF_W:
503 			emit(A64_LDR32(dst, src, tmp), ctx);
504 			break;
505 		case BPF_H:
506 			emit(A64_LDRH(dst, src, tmp), ctx);
507 			break;
508 		case BPF_B:
509 			emit(A64_LDRB(dst, src, tmp), ctx);
510 			break;
511 		case BPF_DW:
512 			emit(A64_LDR64(dst, src, tmp), ctx);
513 			break;
514 		}
515 		break;
516 
517 	/* ST: *(size *)(dst + off) = imm */
518 	case BPF_ST | BPF_MEM | BPF_W:
519 	case BPF_ST | BPF_MEM | BPF_H:
520 	case BPF_ST | BPF_MEM | BPF_B:
521 	case BPF_ST | BPF_MEM | BPF_DW:
522 		goto notyet;
523 
524 	/* STX: *(size *)(dst + off) = src */
525 	case BPF_STX | BPF_MEM | BPF_W:
526 	case BPF_STX | BPF_MEM | BPF_H:
527 	case BPF_STX | BPF_MEM | BPF_B:
528 	case BPF_STX | BPF_MEM | BPF_DW:
529 		ctx->tmp_used = 1;
530 		emit_a64_mov_i(1, tmp, off, ctx);
531 		switch (BPF_SIZE(code)) {
532 		case BPF_W:
533 			emit(A64_STR32(src, dst, tmp), ctx);
534 			break;
535 		case BPF_H:
536 			emit(A64_STRH(src, dst, tmp), ctx);
537 			break;
538 		case BPF_B:
539 			emit(A64_STRB(src, dst, tmp), ctx);
540 			break;
541 		case BPF_DW:
542 			emit(A64_STR64(src, dst, tmp), ctx);
543 			break;
544 		}
545 		break;
546 	/* STX XADD: lock *(u32 *)(dst + off) += src */
547 	case BPF_STX | BPF_XADD | BPF_W:
548 	/* STX XADD: lock *(u64 *)(dst + off) += src */
549 	case BPF_STX | BPF_XADD | BPF_DW:
550 		goto notyet;
551 
552 	/* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + imm)) */
553 	case BPF_LD | BPF_ABS | BPF_W:
554 	case BPF_LD | BPF_ABS | BPF_H:
555 	case BPF_LD | BPF_ABS | BPF_B:
556 	/* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + src + imm)) */
557 	case BPF_LD | BPF_IND | BPF_W:
558 	case BPF_LD | BPF_IND | BPF_H:
559 	case BPF_LD | BPF_IND | BPF_B:
560 	{
561 		const u8 r0 = bpf2a64[BPF_REG_0]; /* r0 = return value */
562 		const u8 r6 = bpf2a64[BPF_REG_6]; /* r6 = pointer to sk_buff */
563 		const u8 fp = bpf2a64[BPF_REG_FP];
564 		const u8 r1 = bpf2a64[BPF_REG_1]; /* r1: struct sk_buff *skb */
565 		const u8 r2 = bpf2a64[BPF_REG_2]; /* r2: int k */
566 		const u8 r3 = bpf2a64[BPF_REG_3]; /* r3: unsigned int size */
567 		const u8 r4 = bpf2a64[BPF_REG_4]; /* r4: void *buffer */
568 		const u8 r5 = bpf2a64[BPF_REG_5]; /* r5: void *(*func)(...) */
569 		int size;
570 
571 		emit(A64_MOV(1, r1, r6), ctx);
572 		emit_a64_mov_i(0, r2, imm, ctx);
573 		if (BPF_MODE(code) == BPF_IND)
574 			emit(A64_ADD(0, r2, r2, src), ctx);
575 		switch (BPF_SIZE(code)) {
576 		case BPF_W:
577 			size = 4;
578 			break;
579 		case BPF_H:
580 			size = 2;
581 			break;
582 		case BPF_B:
583 			size = 1;
584 			break;
585 		default:
586 			return -EINVAL;
587 		}
588 		emit_a64_mov_i64(r3, size, ctx);
589 		emit(A64_ADD_I(1, r4, fp, MAX_BPF_STACK), ctx);
590 		emit_a64_mov_i64(r5, (unsigned long)bpf_load_pointer, ctx);
591 		emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx);
592 		emit(A64_MOV(1, A64_FP, A64_SP), ctx);
593 		emit(A64_BLR(r5), ctx);
594 		emit(A64_MOV(1, r0, A64_R(0)), ctx);
595 		emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx);
596 
597 		jmp_offset = epilogue_offset(ctx);
598 		check_imm19(jmp_offset);
599 		emit(A64_CBZ(1, r0, jmp_offset), ctx);
600 		emit(A64_MOV(1, r5, r0), ctx);
601 		switch (BPF_SIZE(code)) {
602 		case BPF_W:
603 			emit(A64_LDR32(r0, r5, A64_ZR), ctx);
604 #ifndef CONFIG_CPU_BIG_ENDIAN
605 			emit(A64_REV32(0, r0, r0), ctx);
606 #endif
607 			break;
608 		case BPF_H:
609 			emit(A64_LDRH(r0, r5, A64_ZR), ctx);
610 #ifndef CONFIG_CPU_BIG_ENDIAN
611 			emit(A64_REV16(0, r0, r0), ctx);
612 #endif
613 			break;
614 		case BPF_B:
615 			emit(A64_LDRB(r0, r5, A64_ZR), ctx);
616 			break;
617 		}
618 		break;
619 	}
620 notyet:
621 		pr_info_once("*** NOT YET: opcode %02x ***\n", code);
622 		return -EFAULT;
623 
624 	default:
625 		pr_err_once("unknown opcode %02x\n", code);
626 		return -EINVAL;
627 	}
628 
629 	return 0;
630 }
631 
632 static int build_body(struct jit_ctx *ctx)
633 {
634 	const struct bpf_prog *prog = ctx->prog;
635 	int i;
636 
637 	for (i = 0; i < prog->len; i++) {
638 		const struct bpf_insn *insn = &prog->insnsi[i];
639 		int ret;
640 
641 		if (ctx->image == NULL)
642 			ctx->offset[i] = ctx->idx;
643 
644 		ret = build_insn(insn, ctx);
645 		if (ret > 0) {
646 			i++;
647 			continue;
648 		}
649 		if (ret)
650 			return ret;
651 	}
652 
653 	return 0;
654 }
655 
656 static inline void bpf_flush_icache(void *start, void *end)
657 {
658 	flush_icache_range((unsigned long)start, (unsigned long)end);
659 }
660 
661 void bpf_jit_compile(struct bpf_prog *prog)
662 {
663 	/* Nothing to do here. We support Internal BPF. */
664 }
665 
666 void bpf_int_jit_compile(struct bpf_prog *prog)
667 {
668 	struct bpf_binary_header *header;
669 	struct jit_ctx ctx;
670 	int image_size;
671 	u8 *image_ptr;
672 
673 	if (!bpf_jit_enable)
674 		return;
675 
676 	if (!prog || !prog->len)
677 		return;
678 
679 	memset(&ctx, 0, sizeof(ctx));
680 	ctx.prog = prog;
681 
682 	ctx.offset = kcalloc(prog->len, sizeof(int), GFP_KERNEL);
683 	if (ctx.offset == NULL)
684 		return;
685 
686 	/* 1. Initial fake pass to compute ctx->idx. */
687 
688 	/* Fake pass to fill in ctx->offset. */
689 	if (build_body(&ctx))
690 		goto out;
691 
692 	build_prologue(&ctx);
693 	build_epilogue(&ctx);
694 
695 	/* Now we know the actual image size. */
696 	image_size = sizeof(u32) * ctx.idx;
697 	header = bpf_jit_binary_alloc(image_size, &image_ptr,
698 				      sizeof(u32), jit_fill_hole);
699 	if (header == NULL)
700 		goto out;
701 
702 	/* 2. Now, the actual pass. */
703 
704 	ctx.image = (u32 *)image_ptr;
705 	ctx.idx = 0;
706 
707 	build_prologue(&ctx);
708 
709 	ctx.body_offset = ctx.idx;
710 	if (build_body(&ctx)) {
711 		bpf_jit_binary_free(header);
712 		goto out;
713 	}
714 
715 	build_epilogue(&ctx);
716 
717 	/* And we're done. */
718 	if (bpf_jit_enable > 1)
719 		bpf_jit_dump(prog->len, image_size, 2, ctx.image);
720 
721 	bpf_flush_icache(ctx.image, ctx.image + ctx.idx);
722 
723 	set_memory_ro((unsigned long)header, header->pages);
724 	prog->bpf_func = (void *)ctx.image;
725 	prog->jited = true;
726 out:
727 	kfree(ctx.offset);
728 }
729 
730 void bpf_jit_free(struct bpf_prog *prog)
731 {
732 	unsigned long addr = (unsigned long)prog->bpf_func & PAGE_MASK;
733 	struct bpf_binary_header *header = (void *)addr;
734 
735 	if (!prog->jited)
736 		goto free_filter;
737 
738 	set_memory_rw(addr, header->pages);
739 	bpf_jit_binary_free(header);
740 
741 free_filter:
742 	bpf_prog_unlock_free(prog);
743 }
744