xref: /linux/arch/arm64/net/bpf_jit_comp.c (revision a7f7f6248d9740d710fd6bd190293fe5e16410ac)
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 
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 
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 
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 
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  */
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 
144 static inline int bpf2a64_offset(int bpf_to, int bpf_from,
145 				 const struct jit_ctx *ctx)
146 {
147 	int to = ctx->offset[bpf_to];
148 	/* -1 to account for the Branch instruction */
149 	int from = ctx->offset[bpf_from] - 1;
150 
151 	return to - from;
152 }
153 
154 static void jit_fill_hole(void *area, unsigned int size)
155 {
156 	__le32 *ptr;
157 	/* We are guaranteed to have aligned memory. */
158 	for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
159 		*ptr++ = cpu_to_le32(AARCH64_BREAK_FAULT);
160 }
161 
162 static inline int epilogue_offset(const struct jit_ctx *ctx)
163 {
164 	int to = ctx->epilogue_offset;
165 	int from = ctx->idx;
166 
167 	return to - from;
168 }
169 
170 static bool is_addsub_imm(u32 imm)
171 {
172 	/* Either imm12 or shifted imm12. */
173 	return !(imm & ~0xfff) || !(imm & ~0xfff000);
174 }
175 
176 /* Stack must be multiples of 16B */
177 #define STACK_ALIGN(sz) (((sz) + 15) & ~15)
178 
179 /* Tail call offset to jump into */
180 #if IS_ENABLED(CONFIG_ARM64_BTI_KERNEL)
181 #define PROLOGUE_OFFSET 8
182 #else
183 #define PROLOGUE_OFFSET 7
184 #endif
185 
186 static int build_prologue(struct jit_ctx *ctx, bool ebpf_from_cbpf)
187 {
188 	const struct bpf_prog *prog = ctx->prog;
189 	const u8 r6 = bpf2a64[BPF_REG_6];
190 	const u8 r7 = bpf2a64[BPF_REG_7];
191 	const u8 r8 = bpf2a64[BPF_REG_8];
192 	const u8 r9 = bpf2a64[BPF_REG_9];
193 	const u8 fp = bpf2a64[BPF_REG_FP];
194 	const u8 tcc = bpf2a64[TCALL_CNT];
195 	const int idx0 = ctx->idx;
196 	int cur_offset;
197 
198 	/*
199 	 * BPF prog stack layout
200 	 *
201 	 *                         high
202 	 * original A64_SP =>   0:+-----+ BPF prologue
203 	 *                        |FP/LR|
204 	 * current A64_FP =>  -16:+-----+
205 	 *                        | ... | callee saved registers
206 	 * BPF fp register => -64:+-----+ <= (BPF_FP)
207 	 *                        |     |
208 	 *                        | ... | BPF prog stack
209 	 *                        |     |
210 	 *                        +-----+ <= (BPF_FP - prog->aux->stack_depth)
211 	 *                        |RSVD | padding
212 	 * current A64_SP =>      +-----+ <= (BPF_FP - ctx->stack_size)
213 	 *                        |     |
214 	 *                        | ... | Function call stack
215 	 *                        |     |
216 	 *                        +-----+
217 	 *                          low
218 	 *
219 	 */
220 
221 	/* BTI landing pad */
222 	if (IS_ENABLED(CONFIG_ARM64_BTI_KERNEL))
223 		emit(A64_BTI_C, ctx);
224 
225 	/* Save FP and LR registers to stay align with ARM64 AAPCS */
226 	emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx);
227 	emit(A64_MOV(1, A64_FP, A64_SP), ctx);
228 
229 	/* Save callee-saved registers */
230 	emit(A64_PUSH(r6, r7, A64_SP), ctx);
231 	emit(A64_PUSH(r8, r9, A64_SP), ctx);
232 	emit(A64_PUSH(fp, tcc, A64_SP), ctx);
233 
234 	/* Set up BPF prog stack base register */
235 	emit(A64_MOV(1, fp, A64_SP), ctx);
236 
237 	if (!ebpf_from_cbpf) {
238 		/* Initialize tail_call_cnt */
239 		emit(A64_MOVZ(1, tcc, 0, 0), ctx);
240 
241 		cur_offset = ctx->idx - idx0;
242 		if (cur_offset != PROLOGUE_OFFSET) {
243 			pr_err_once("PROLOGUE_OFFSET = %d, expected %d!\n",
244 				    cur_offset, PROLOGUE_OFFSET);
245 			return -1;
246 		}
247 
248 		/* BTI landing pad for the tail call, done with a BR */
249 		if (IS_ENABLED(CONFIG_ARM64_BTI_KERNEL))
250 			emit(A64_BTI_J, ctx);
251 	}
252 
253 	ctx->stack_size = STACK_ALIGN(prog->aux->stack_depth);
254 
255 	/* Set up function call stack */
256 	emit(A64_SUB_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
257 	return 0;
258 }
259 
260 static int out_offset = -1; /* initialized on the first pass of build_body() */
261 static int emit_bpf_tail_call(struct jit_ctx *ctx)
262 {
263 	/* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
264 	const u8 r2 = bpf2a64[BPF_REG_2];
265 	const u8 r3 = bpf2a64[BPF_REG_3];
266 
267 	const u8 tmp = bpf2a64[TMP_REG_1];
268 	const u8 prg = bpf2a64[TMP_REG_2];
269 	const u8 tcc = bpf2a64[TCALL_CNT];
270 	const int idx0 = ctx->idx;
271 #define cur_offset (ctx->idx - idx0)
272 #define jmp_offset (out_offset - (cur_offset))
273 	size_t off;
274 
275 	/* if (index >= array->map.max_entries)
276 	 *     goto out;
277 	 */
278 	off = offsetof(struct bpf_array, map.max_entries);
279 	emit_a64_mov_i64(tmp, off, ctx);
280 	emit(A64_LDR32(tmp, r2, tmp), ctx);
281 	emit(A64_MOV(0, r3, r3), ctx);
282 	emit(A64_CMP(0, r3, tmp), ctx);
283 	emit(A64_B_(A64_COND_CS, jmp_offset), ctx);
284 
285 	/* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
286 	 *     goto out;
287 	 * tail_call_cnt++;
288 	 */
289 	emit_a64_mov_i64(tmp, MAX_TAIL_CALL_CNT, ctx);
290 	emit(A64_CMP(1, tcc, tmp), ctx);
291 	emit(A64_B_(A64_COND_HI, jmp_offset), ctx);
292 	emit(A64_ADD_I(1, tcc, tcc, 1), ctx);
293 
294 	/* prog = array->ptrs[index];
295 	 * if (prog == NULL)
296 	 *     goto out;
297 	 */
298 	off = offsetof(struct bpf_array, ptrs);
299 	emit_a64_mov_i64(tmp, off, ctx);
300 	emit(A64_ADD(1, tmp, r2, tmp), ctx);
301 	emit(A64_LSL(1, prg, r3, 3), ctx);
302 	emit(A64_LDR64(prg, tmp, prg), ctx);
303 	emit(A64_CBZ(1, prg, jmp_offset), ctx);
304 
305 	/* goto *(prog->bpf_func + prologue_offset); */
306 	off = offsetof(struct bpf_prog, bpf_func);
307 	emit_a64_mov_i64(tmp, off, ctx);
308 	emit(A64_LDR64(tmp, prg, tmp), ctx);
309 	emit(A64_ADD_I(1, tmp, tmp, sizeof(u32) * PROLOGUE_OFFSET), ctx);
310 	emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
311 	emit(A64_BR(tmp), ctx);
312 
313 	/* out: */
314 	if (out_offset == -1)
315 		out_offset = cur_offset;
316 	if (cur_offset != out_offset) {
317 		pr_err_once("tail_call out_offset = %d, expected %d!\n",
318 			    cur_offset, out_offset);
319 		return -1;
320 	}
321 	return 0;
322 #undef cur_offset
323 #undef jmp_offset
324 }
325 
326 static void build_epilogue(struct jit_ctx *ctx)
327 {
328 	const u8 r0 = bpf2a64[BPF_REG_0];
329 	const u8 r6 = bpf2a64[BPF_REG_6];
330 	const u8 r7 = bpf2a64[BPF_REG_7];
331 	const u8 r8 = bpf2a64[BPF_REG_8];
332 	const u8 r9 = bpf2a64[BPF_REG_9];
333 	const u8 fp = bpf2a64[BPF_REG_FP];
334 
335 	/* We're done with BPF stack */
336 	emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
337 
338 	/* Restore fs (x25) and x26 */
339 	emit(A64_POP(fp, A64_R(26), A64_SP), ctx);
340 
341 	/* Restore callee-saved register */
342 	emit(A64_POP(r8, r9, A64_SP), ctx);
343 	emit(A64_POP(r6, r7, A64_SP), ctx);
344 
345 	/* Restore FP/LR registers */
346 	emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx);
347 
348 	/* Set return value */
349 	emit(A64_MOV(1, A64_R(0), r0), ctx);
350 
351 	emit(A64_RET(A64_LR), ctx);
352 }
353 
354 /* JITs an eBPF instruction.
355  * Returns:
356  * 0  - successfully JITed an 8-byte eBPF instruction.
357  * >0 - successfully JITed a 16-byte eBPF instruction.
358  * <0 - failed to JIT.
359  */
360 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx,
361 		      bool extra_pass)
362 {
363 	const u8 code = insn->code;
364 	const u8 dst = bpf2a64[insn->dst_reg];
365 	const u8 src = bpf2a64[insn->src_reg];
366 	const u8 tmp = bpf2a64[TMP_REG_1];
367 	const u8 tmp2 = bpf2a64[TMP_REG_2];
368 	const u8 tmp3 = bpf2a64[TMP_REG_3];
369 	const s16 off = insn->off;
370 	const s32 imm = insn->imm;
371 	const int i = insn - ctx->prog->insnsi;
372 	const bool is64 = BPF_CLASS(code) == BPF_ALU64 ||
373 			  BPF_CLASS(code) == BPF_JMP;
374 	const bool isdw = BPF_SIZE(code) == BPF_DW;
375 	u8 jmp_cond, reg;
376 	s32 jmp_offset;
377 	u32 a64_insn;
378 
379 #define check_imm(bits, imm) do {				\
380 	if ((((imm) > 0) && ((imm) >> (bits))) ||		\
381 	    (((imm) < 0) && (~(imm) >> (bits)))) {		\
382 		pr_info("[%2d] imm=%d(0x%x) out of range\n",	\
383 			i, imm, imm);				\
384 		return -EINVAL;					\
385 	}							\
386 } while (0)
387 #define check_imm19(imm) check_imm(19, imm)
388 #define check_imm26(imm) check_imm(26, imm)
389 
390 	switch (code) {
391 	/* dst = src */
392 	case BPF_ALU | BPF_MOV | BPF_X:
393 	case BPF_ALU64 | BPF_MOV | BPF_X:
394 		emit(A64_MOV(is64, dst, src), ctx);
395 		break;
396 	/* dst = dst OP src */
397 	case BPF_ALU | BPF_ADD | BPF_X:
398 	case BPF_ALU64 | BPF_ADD | BPF_X:
399 		emit(A64_ADD(is64, dst, dst, src), ctx);
400 		break;
401 	case BPF_ALU | BPF_SUB | BPF_X:
402 	case BPF_ALU64 | BPF_SUB | BPF_X:
403 		emit(A64_SUB(is64, dst, dst, src), ctx);
404 		break;
405 	case BPF_ALU | BPF_AND | BPF_X:
406 	case BPF_ALU64 | BPF_AND | BPF_X:
407 		emit(A64_AND(is64, dst, dst, src), ctx);
408 		break;
409 	case BPF_ALU | BPF_OR | BPF_X:
410 	case BPF_ALU64 | BPF_OR | BPF_X:
411 		emit(A64_ORR(is64, dst, dst, src), ctx);
412 		break;
413 	case BPF_ALU | BPF_XOR | BPF_X:
414 	case BPF_ALU64 | BPF_XOR | BPF_X:
415 		emit(A64_EOR(is64, dst, dst, src), ctx);
416 		break;
417 	case BPF_ALU | BPF_MUL | BPF_X:
418 	case BPF_ALU64 | BPF_MUL | BPF_X:
419 		emit(A64_MUL(is64, dst, dst, src), ctx);
420 		break;
421 	case BPF_ALU | BPF_DIV | BPF_X:
422 	case BPF_ALU64 | BPF_DIV | BPF_X:
423 	case BPF_ALU | BPF_MOD | BPF_X:
424 	case BPF_ALU64 | BPF_MOD | BPF_X:
425 		switch (BPF_OP(code)) {
426 		case BPF_DIV:
427 			emit(A64_UDIV(is64, dst, dst, src), ctx);
428 			break;
429 		case BPF_MOD:
430 			emit(A64_UDIV(is64, tmp, dst, src), ctx);
431 			emit(A64_MSUB(is64, dst, dst, tmp, src), ctx);
432 			break;
433 		}
434 		break;
435 	case BPF_ALU | BPF_LSH | BPF_X:
436 	case BPF_ALU64 | BPF_LSH | BPF_X:
437 		emit(A64_LSLV(is64, dst, dst, src), ctx);
438 		break;
439 	case BPF_ALU | BPF_RSH | BPF_X:
440 	case BPF_ALU64 | BPF_RSH | BPF_X:
441 		emit(A64_LSRV(is64, dst, dst, src), ctx);
442 		break;
443 	case BPF_ALU | BPF_ARSH | BPF_X:
444 	case BPF_ALU64 | BPF_ARSH | BPF_X:
445 		emit(A64_ASRV(is64, dst, dst, src), ctx);
446 		break;
447 	/* dst = -dst */
448 	case BPF_ALU | BPF_NEG:
449 	case BPF_ALU64 | BPF_NEG:
450 		emit(A64_NEG(is64, dst, dst), ctx);
451 		break;
452 	/* dst = BSWAP##imm(dst) */
453 	case BPF_ALU | BPF_END | BPF_FROM_LE:
454 	case BPF_ALU | BPF_END | BPF_FROM_BE:
455 #ifdef CONFIG_CPU_BIG_ENDIAN
456 		if (BPF_SRC(code) == BPF_FROM_BE)
457 			goto emit_bswap_uxt;
458 #else /* !CONFIG_CPU_BIG_ENDIAN */
459 		if (BPF_SRC(code) == BPF_FROM_LE)
460 			goto emit_bswap_uxt;
461 #endif
462 		switch (imm) {
463 		case 16:
464 			emit(A64_REV16(is64, dst, dst), ctx);
465 			/* zero-extend 16 bits into 64 bits */
466 			emit(A64_UXTH(is64, dst, dst), ctx);
467 			break;
468 		case 32:
469 			emit(A64_REV32(is64, dst, dst), ctx);
470 			/* upper 32 bits already cleared */
471 			break;
472 		case 64:
473 			emit(A64_REV64(dst, dst), ctx);
474 			break;
475 		}
476 		break;
477 emit_bswap_uxt:
478 		switch (imm) {
479 		case 16:
480 			/* zero-extend 16 bits into 64 bits */
481 			emit(A64_UXTH(is64, dst, dst), ctx);
482 			break;
483 		case 32:
484 			/* zero-extend 32 bits into 64 bits */
485 			emit(A64_UXTW(is64, dst, dst), ctx);
486 			break;
487 		case 64:
488 			/* nop */
489 			break;
490 		}
491 		break;
492 	/* dst = imm */
493 	case BPF_ALU | BPF_MOV | BPF_K:
494 	case BPF_ALU64 | BPF_MOV | BPF_K:
495 		emit_a64_mov_i(is64, dst, imm, ctx);
496 		break;
497 	/* dst = dst OP imm */
498 	case BPF_ALU | BPF_ADD | BPF_K:
499 	case BPF_ALU64 | BPF_ADD | BPF_K:
500 		if (is_addsub_imm(imm)) {
501 			emit(A64_ADD_I(is64, dst, dst, imm), ctx);
502 		} else if (is_addsub_imm(-imm)) {
503 			emit(A64_SUB_I(is64, dst, dst, -imm), ctx);
504 		} else {
505 			emit_a64_mov_i(is64, tmp, imm, ctx);
506 			emit(A64_ADD(is64, dst, dst, tmp), ctx);
507 		}
508 		break;
509 	case BPF_ALU | BPF_SUB | BPF_K:
510 	case BPF_ALU64 | BPF_SUB | BPF_K:
511 		if (is_addsub_imm(imm)) {
512 			emit(A64_SUB_I(is64, dst, dst, imm), ctx);
513 		} else if (is_addsub_imm(-imm)) {
514 			emit(A64_ADD_I(is64, dst, dst, -imm), ctx);
515 		} else {
516 			emit_a64_mov_i(is64, tmp, imm, ctx);
517 			emit(A64_SUB(is64, dst, dst, tmp), ctx);
518 		}
519 		break;
520 	case BPF_ALU | BPF_AND | BPF_K:
521 	case BPF_ALU64 | BPF_AND | BPF_K:
522 		a64_insn = A64_AND_I(is64, dst, dst, imm);
523 		if (a64_insn != AARCH64_BREAK_FAULT) {
524 			emit(a64_insn, ctx);
525 		} else {
526 			emit_a64_mov_i(is64, tmp, imm, ctx);
527 			emit(A64_AND(is64, dst, dst, tmp), ctx);
528 		}
529 		break;
530 	case BPF_ALU | BPF_OR | BPF_K:
531 	case BPF_ALU64 | BPF_OR | BPF_K:
532 		a64_insn = A64_ORR_I(is64, dst, dst, imm);
533 		if (a64_insn != AARCH64_BREAK_FAULT) {
534 			emit(a64_insn, ctx);
535 		} else {
536 			emit_a64_mov_i(is64, tmp, imm, ctx);
537 			emit(A64_ORR(is64, dst, dst, tmp), ctx);
538 		}
539 		break;
540 	case BPF_ALU | BPF_XOR | BPF_K:
541 	case BPF_ALU64 | BPF_XOR | BPF_K:
542 		a64_insn = A64_EOR_I(is64, dst, dst, imm);
543 		if (a64_insn != AARCH64_BREAK_FAULT) {
544 			emit(a64_insn, ctx);
545 		} else {
546 			emit_a64_mov_i(is64, tmp, imm, ctx);
547 			emit(A64_EOR(is64, dst, dst, tmp), ctx);
548 		}
549 		break;
550 	case BPF_ALU | BPF_MUL | BPF_K:
551 	case BPF_ALU64 | BPF_MUL | BPF_K:
552 		emit_a64_mov_i(is64, tmp, imm, ctx);
553 		emit(A64_MUL(is64, dst, dst, tmp), ctx);
554 		break;
555 	case BPF_ALU | BPF_DIV | BPF_K:
556 	case BPF_ALU64 | BPF_DIV | BPF_K:
557 		emit_a64_mov_i(is64, tmp, imm, ctx);
558 		emit(A64_UDIV(is64, dst, dst, tmp), ctx);
559 		break;
560 	case BPF_ALU | BPF_MOD | BPF_K:
561 	case BPF_ALU64 | BPF_MOD | BPF_K:
562 		emit_a64_mov_i(is64, tmp2, imm, ctx);
563 		emit(A64_UDIV(is64, tmp, dst, tmp2), ctx);
564 		emit(A64_MSUB(is64, dst, dst, tmp, tmp2), ctx);
565 		break;
566 	case BPF_ALU | BPF_LSH | BPF_K:
567 	case BPF_ALU64 | BPF_LSH | BPF_K:
568 		emit(A64_LSL(is64, dst, dst, imm), ctx);
569 		break;
570 	case BPF_ALU | BPF_RSH | BPF_K:
571 	case BPF_ALU64 | BPF_RSH | BPF_K:
572 		emit(A64_LSR(is64, dst, dst, imm), ctx);
573 		break;
574 	case BPF_ALU | BPF_ARSH | BPF_K:
575 	case BPF_ALU64 | BPF_ARSH | BPF_K:
576 		emit(A64_ASR(is64, dst, dst, imm), ctx);
577 		break;
578 
579 	/* JUMP off */
580 	case BPF_JMP | BPF_JA:
581 		jmp_offset = bpf2a64_offset(i + off, i, ctx);
582 		check_imm26(jmp_offset);
583 		emit(A64_B(jmp_offset), ctx);
584 		break;
585 	/* IF (dst COND src) JUMP off */
586 	case BPF_JMP | BPF_JEQ | BPF_X:
587 	case BPF_JMP | BPF_JGT | BPF_X:
588 	case BPF_JMP | BPF_JLT | BPF_X:
589 	case BPF_JMP | BPF_JGE | BPF_X:
590 	case BPF_JMP | BPF_JLE | BPF_X:
591 	case BPF_JMP | BPF_JNE | BPF_X:
592 	case BPF_JMP | BPF_JSGT | BPF_X:
593 	case BPF_JMP | BPF_JSLT | BPF_X:
594 	case BPF_JMP | BPF_JSGE | BPF_X:
595 	case BPF_JMP | BPF_JSLE | BPF_X:
596 	case BPF_JMP32 | BPF_JEQ | BPF_X:
597 	case BPF_JMP32 | BPF_JGT | BPF_X:
598 	case BPF_JMP32 | BPF_JLT | BPF_X:
599 	case BPF_JMP32 | BPF_JGE | BPF_X:
600 	case BPF_JMP32 | BPF_JLE | BPF_X:
601 	case BPF_JMP32 | BPF_JNE | BPF_X:
602 	case BPF_JMP32 | BPF_JSGT | BPF_X:
603 	case BPF_JMP32 | BPF_JSLT | BPF_X:
604 	case BPF_JMP32 | BPF_JSGE | BPF_X:
605 	case BPF_JMP32 | BPF_JSLE | BPF_X:
606 		emit(A64_CMP(is64, dst, src), ctx);
607 emit_cond_jmp:
608 		jmp_offset = bpf2a64_offset(i + off, i, ctx);
609 		check_imm19(jmp_offset);
610 		switch (BPF_OP(code)) {
611 		case BPF_JEQ:
612 			jmp_cond = A64_COND_EQ;
613 			break;
614 		case BPF_JGT:
615 			jmp_cond = A64_COND_HI;
616 			break;
617 		case BPF_JLT:
618 			jmp_cond = A64_COND_CC;
619 			break;
620 		case BPF_JGE:
621 			jmp_cond = A64_COND_CS;
622 			break;
623 		case BPF_JLE:
624 			jmp_cond = A64_COND_LS;
625 			break;
626 		case BPF_JSET:
627 		case BPF_JNE:
628 			jmp_cond = A64_COND_NE;
629 			break;
630 		case BPF_JSGT:
631 			jmp_cond = A64_COND_GT;
632 			break;
633 		case BPF_JSLT:
634 			jmp_cond = A64_COND_LT;
635 			break;
636 		case BPF_JSGE:
637 			jmp_cond = A64_COND_GE;
638 			break;
639 		case BPF_JSLE:
640 			jmp_cond = A64_COND_LE;
641 			break;
642 		default:
643 			return -EFAULT;
644 		}
645 		emit(A64_B_(jmp_cond, jmp_offset), ctx);
646 		break;
647 	case BPF_JMP | BPF_JSET | BPF_X:
648 	case BPF_JMP32 | BPF_JSET | BPF_X:
649 		emit(A64_TST(is64, dst, src), ctx);
650 		goto emit_cond_jmp;
651 	/* IF (dst COND imm) JUMP off */
652 	case BPF_JMP | BPF_JEQ | BPF_K:
653 	case BPF_JMP | BPF_JGT | BPF_K:
654 	case BPF_JMP | BPF_JLT | BPF_K:
655 	case BPF_JMP | BPF_JGE | BPF_K:
656 	case BPF_JMP | BPF_JLE | BPF_K:
657 	case BPF_JMP | BPF_JNE | BPF_K:
658 	case BPF_JMP | BPF_JSGT | BPF_K:
659 	case BPF_JMP | BPF_JSLT | BPF_K:
660 	case BPF_JMP | BPF_JSGE | BPF_K:
661 	case BPF_JMP | BPF_JSLE | BPF_K:
662 	case BPF_JMP32 | BPF_JEQ | BPF_K:
663 	case BPF_JMP32 | BPF_JGT | BPF_K:
664 	case BPF_JMP32 | BPF_JLT | BPF_K:
665 	case BPF_JMP32 | BPF_JGE | BPF_K:
666 	case BPF_JMP32 | BPF_JLE | BPF_K:
667 	case BPF_JMP32 | BPF_JNE | BPF_K:
668 	case BPF_JMP32 | BPF_JSGT | BPF_K:
669 	case BPF_JMP32 | BPF_JSLT | BPF_K:
670 	case BPF_JMP32 | BPF_JSGE | BPF_K:
671 	case BPF_JMP32 | BPF_JSLE | BPF_K:
672 		if (is_addsub_imm(imm)) {
673 			emit(A64_CMP_I(is64, dst, imm), ctx);
674 		} else if (is_addsub_imm(-imm)) {
675 			emit(A64_CMN_I(is64, dst, -imm), ctx);
676 		} else {
677 			emit_a64_mov_i(is64, tmp, imm, ctx);
678 			emit(A64_CMP(is64, dst, tmp), ctx);
679 		}
680 		goto emit_cond_jmp;
681 	case BPF_JMP | BPF_JSET | BPF_K:
682 	case BPF_JMP32 | BPF_JSET | BPF_K:
683 		a64_insn = A64_TST_I(is64, dst, imm);
684 		if (a64_insn != AARCH64_BREAK_FAULT) {
685 			emit(a64_insn, ctx);
686 		} else {
687 			emit_a64_mov_i(is64, tmp, imm, ctx);
688 			emit(A64_TST(is64, dst, tmp), ctx);
689 		}
690 		goto emit_cond_jmp;
691 	/* function call */
692 	case BPF_JMP | BPF_CALL:
693 	{
694 		const u8 r0 = bpf2a64[BPF_REG_0];
695 		bool func_addr_fixed;
696 		u64 func_addr;
697 		int ret;
698 
699 		ret = bpf_jit_get_func_addr(ctx->prog, insn, extra_pass,
700 					    &func_addr, &func_addr_fixed);
701 		if (ret < 0)
702 			return ret;
703 		emit_addr_mov_i64(tmp, func_addr, ctx);
704 		emit(A64_BLR(tmp), ctx);
705 		emit(A64_MOV(1, r0, A64_R(0)), ctx);
706 		break;
707 	}
708 	/* tail call */
709 	case BPF_JMP | BPF_TAIL_CALL:
710 		if (emit_bpf_tail_call(ctx))
711 			return -EFAULT;
712 		break;
713 	/* function return */
714 	case BPF_JMP | BPF_EXIT:
715 		/* Optimization: when last instruction is EXIT,
716 		   simply fallthrough to epilogue. */
717 		if (i == ctx->prog->len - 1)
718 			break;
719 		jmp_offset = epilogue_offset(ctx);
720 		check_imm26(jmp_offset);
721 		emit(A64_B(jmp_offset), ctx);
722 		break;
723 
724 	/* dst = imm64 */
725 	case BPF_LD | BPF_IMM | BPF_DW:
726 	{
727 		const struct bpf_insn insn1 = insn[1];
728 		u64 imm64;
729 
730 		imm64 = (u64)insn1.imm << 32 | (u32)imm;
731 		emit_a64_mov_i64(dst, imm64, ctx);
732 
733 		return 1;
734 	}
735 
736 	/* LDX: dst = *(size *)(src + off) */
737 	case BPF_LDX | BPF_MEM | BPF_W:
738 	case BPF_LDX | BPF_MEM | BPF_H:
739 	case BPF_LDX | BPF_MEM | BPF_B:
740 	case BPF_LDX | BPF_MEM | BPF_DW:
741 		emit_a64_mov_i(1, tmp, off, ctx);
742 		switch (BPF_SIZE(code)) {
743 		case BPF_W:
744 			emit(A64_LDR32(dst, src, tmp), ctx);
745 			break;
746 		case BPF_H:
747 			emit(A64_LDRH(dst, src, tmp), ctx);
748 			break;
749 		case BPF_B:
750 			emit(A64_LDRB(dst, src, tmp), ctx);
751 			break;
752 		case BPF_DW:
753 			emit(A64_LDR64(dst, src, tmp), ctx);
754 			break;
755 		}
756 		break;
757 
758 	/* ST: *(size *)(dst + off) = imm */
759 	case BPF_ST | BPF_MEM | BPF_W:
760 	case BPF_ST | BPF_MEM | BPF_H:
761 	case BPF_ST | BPF_MEM | BPF_B:
762 	case BPF_ST | BPF_MEM | BPF_DW:
763 		/* Load imm to a register then store it */
764 		emit_a64_mov_i(1, tmp2, off, ctx);
765 		emit_a64_mov_i(1, tmp, imm, ctx);
766 		switch (BPF_SIZE(code)) {
767 		case BPF_W:
768 			emit(A64_STR32(tmp, dst, tmp2), ctx);
769 			break;
770 		case BPF_H:
771 			emit(A64_STRH(tmp, dst, tmp2), ctx);
772 			break;
773 		case BPF_B:
774 			emit(A64_STRB(tmp, dst, tmp2), ctx);
775 			break;
776 		case BPF_DW:
777 			emit(A64_STR64(tmp, dst, tmp2), ctx);
778 			break;
779 		}
780 		break;
781 
782 	/* STX: *(size *)(dst + off) = src */
783 	case BPF_STX | BPF_MEM | BPF_W:
784 	case BPF_STX | BPF_MEM | BPF_H:
785 	case BPF_STX | BPF_MEM | BPF_B:
786 	case BPF_STX | BPF_MEM | BPF_DW:
787 		emit_a64_mov_i(1, tmp, off, ctx);
788 		switch (BPF_SIZE(code)) {
789 		case BPF_W:
790 			emit(A64_STR32(src, dst, tmp), ctx);
791 			break;
792 		case BPF_H:
793 			emit(A64_STRH(src, dst, tmp), ctx);
794 			break;
795 		case BPF_B:
796 			emit(A64_STRB(src, dst, tmp), ctx);
797 			break;
798 		case BPF_DW:
799 			emit(A64_STR64(src, dst, tmp), ctx);
800 			break;
801 		}
802 		break;
803 
804 	/* STX XADD: lock *(u32 *)(dst + off) += src */
805 	case BPF_STX | BPF_XADD | BPF_W:
806 	/* STX XADD: lock *(u64 *)(dst + off) += src */
807 	case BPF_STX | BPF_XADD | BPF_DW:
808 		if (!off) {
809 			reg = dst;
810 		} else {
811 			emit_a64_mov_i(1, tmp, off, ctx);
812 			emit(A64_ADD(1, tmp, tmp, dst), ctx);
813 			reg = tmp;
814 		}
815 		if (cpus_have_cap(ARM64_HAS_LSE_ATOMICS)) {
816 			emit(A64_STADD(isdw, reg, src), ctx);
817 		} else {
818 			emit(A64_LDXR(isdw, tmp2, reg), ctx);
819 			emit(A64_ADD(isdw, tmp2, tmp2, src), ctx);
820 			emit(A64_STXR(isdw, tmp2, reg, tmp3), ctx);
821 			jmp_offset = -3;
822 			check_imm19(jmp_offset);
823 			emit(A64_CBNZ(0, tmp3, jmp_offset), ctx);
824 		}
825 		break;
826 
827 	default:
828 		pr_err_once("unknown opcode %02x\n", code);
829 		return -EINVAL;
830 	}
831 
832 	return 0;
833 }
834 
835 static int build_body(struct jit_ctx *ctx, bool extra_pass)
836 {
837 	const struct bpf_prog *prog = ctx->prog;
838 	int i;
839 
840 	for (i = 0; i < prog->len; i++) {
841 		const struct bpf_insn *insn = &prog->insnsi[i];
842 		int ret;
843 
844 		ret = build_insn(insn, ctx, extra_pass);
845 		if (ret > 0) {
846 			i++;
847 			if (ctx->image == NULL)
848 				ctx->offset[i] = ctx->idx;
849 			continue;
850 		}
851 		if (ctx->image == NULL)
852 			ctx->offset[i] = ctx->idx;
853 		if (ret)
854 			return ret;
855 	}
856 
857 	return 0;
858 }
859 
860 static int validate_code(struct jit_ctx *ctx)
861 {
862 	int i;
863 
864 	for (i = 0; i < ctx->idx; i++) {
865 		u32 a64_insn = le32_to_cpu(ctx->image[i]);
866 
867 		if (a64_insn == AARCH64_BREAK_FAULT)
868 			return -1;
869 	}
870 
871 	return 0;
872 }
873 
874 static inline void bpf_flush_icache(void *start, void *end)
875 {
876 	flush_icache_range((unsigned long)start, (unsigned long)end);
877 }
878 
879 struct arm64_jit_data {
880 	struct bpf_binary_header *header;
881 	u8 *image;
882 	struct jit_ctx ctx;
883 };
884 
885 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
886 {
887 	struct bpf_prog *tmp, *orig_prog = prog;
888 	struct bpf_binary_header *header;
889 	struct arm64_jit_data *jit_data;
890 	bool was_classic = bpf_prog_was_classic(prog);
891 	bool tmp_blinded = false;
892 	bool extra_pass = false;
893 	struct jit_ctx ctx;
894 	int image_size;
895 	u8 *image_ptr;
896 
897 	if (!prog->jit_requested)
898 		return orig_prog;
899 
900 	tmp = bpf_jit_blind_constants(prog);
901 	/* If blinding was requested and we failed during blinding,
902 	 * we must fall back to the interpreter.
903 	 */
904 	if (IS_ERR(tmp))
905 		return orig_prog;
906 	if (tmp != prog) {
907 		tmp_blinded = true;
908 		prog = tmp;
909 	}
910 
911 	jit_data = prog->aux->jit_data;
912 	if (!jit_data) {
913 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
914 		if (!jit_data) {
915 			prog = orig_prog;
916 			goto out;
917 		}
918 		prog->aux->jit_data = jit_data;
919 	}
920 	if (jit_data->ctx.offset) {
921 		ctx = jit_data->ctx;
922 		image_ptr = jit_data->image;
923 		header = jit_data->header;
924 		extra_pass = true;
925 		image_size = sizeof(u32) * ctx.idx;
926 		goto skip_init_ctx;
927 	}
928 	memset(&ctx, 0, sizeof(ctx));
929 	ctx.prog = prog;
930 
931 	ctx.offset = kcalloc(prog->len, sizeof(int), GFP_KERNEL);
932 	if (ctx.offset == NULL) {
933 		prog = orig_prog;
934 		goto out_off;
935 	}
936 
937 	/* 1. Initial fake pass to compute ctx->idx. */
938 
939 	/* Fake pass to fill in ctx->offset. */
940 	if (build_body(&ctx, extra_pass)) {
941 		prog = orig_prog;
942 		goto out_off;
943 	}
944 
945 	if (build_prologue(&ctx, was_classic)) {
946 		prog = orig_prog;
947 		goto out_off;
948 	}
949 
950 	ctx.epilogue_offset = ctx.idx;
951 	build_epilogue(&ctx);
952 
953 	/* Now we know the actual image size. */
954 	image_size = sizeof(u32) * ctx.idx;
955 	header = bpf_jit_binary_alloc(image_size, &image_ptr,
956 				      sizeof(u32), jit_fill_hole);
957 	if (header == NULL) {
958 		prog = orig_prog;
959 		goto out_off;
960 	}
961 
962 	/* 2. Now, the actual pass. */
963 
964 	ctx.image = (__le32 *)image_ptr;
965 skip_init_ctx:
966 	ctx.idx = 0;
967 
968 	build_prologue(&ctx, was_classic);
969 
970 	if (build_body(&ctx, extra_pass)) {
971 		bpf_jit_binary_free(header);
972 		prog = orig_prog;
973 		goto out_off;
974 	}
975 
976 	build_epilogue(&ctx);
977 
978 	/* 3. Extra pass to validate JITed code. */
979 	if (validate_code(&ctx)) {
980 		bpf_jit_binary_free(header);
981 		prog = orig_prog;
982 		goto out_off;
983 	}
984 
985 	/* And we're done. */
986 	if (bpf_jit_enable > 1)
987 		bpf_jit_dump(prog->len, image_size, 2, ctx.image);
988 
989 	bpf_flush_icache(header, ctx.image + ctx.idx);
990 
991 	if (!prog->is_func || extra_pass) {
992 		if (extra_pass && ctx.idx != jit_data->ctx.idx) {
993 			pr_err_once("multi-func JIT bug %d != %d\n",
994 				    ctx.idx, jit_data->ctx.idx);
995 			bpf_jit_binary_free(header);
996 			prog->bpf_func = NULL;
997 			prog->jited = 0;
998 			goto out_off;
999 		}
1000 		bpf_jit_binary_lock_ro(header);
1001 	} else {
1002 		jit_data->ctx = ctx;
1003 		jit_data->image = image_ptr;
1004 		jit_data->header = header;
1005 	}
1006 	prog->bpf_func = (void *)ctx.image;
1007 	prog->jited = 1;
1008 	prog->jited_len = image_size;
1009 
1010 	if (!prog->is_func || extra_pass) {
1011 		bpf_prog_fill_jited_linfo(prog, ctx.offset);
1012 out_off:
1013 		kfree(ctx.offset);
1014 		kfree(jit_data);
1015 		prog->aux->jit_data = NULL;
1016 	}
1017 out:
1018 	if (tmp_blinded)
1019 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
1020 					   tmp : orig_prog);
1021 	return prog;
1022 }
1023 
1024 void *bpf_jit_alloc_exec(unsigned long size)
1025 {
1026 	return __vmalloc_node_range(size, PAGE_SIZE, BPF_JIT_REGION_START,
1027 				    BPF_JIT_REGION_END, GFP_KERNEL,
1028 				    PAGE_KERNEL, 0, NUMA_NO_NODE,
1029 				    __builtin_return_address(0));
1030 }
1031 
1032 void bpf_jit_free_exec(void *addr)
1033 {
1034 	return vfree(addr);
1035 }
1036