xref: /openbmc/linux/arch/arm64/net/bpf_jit_comp.c (revision 6db6b729)
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/bitfield.h>
11 #include <linux/bpf.h>
12 #include <linux/filter.h>
13 #include <linux/memory.h>
14 #include <linux/printk.h>
15 #include <linux/slab.h>
16 
17 #include <asm/asm-extable.h>
18 #include <asm/byteorder.h>
19 #include <asm/cacheflush.h>
20 #include <asm/debug-monitors.h>
21 #include <asm/insn.h>
22 #include <asm/patching.h>
23 #include <asm/set_memory.h>
24 
25 #include "bpf_jit.h"
26 
27 #define TMP_REG_1 (MAX_BPF_JIT_REG + 0)
28 #define TMP_REG_2 (MAX_BPF_JIT_REG + 1)
29 #define TCALL_CNT (MAX_BPF_JIT_REG + 2)
30 #define TMP_REG_3 (MAX_BPF_JIT_REG + 3)
31 #define FP_BOTTOM (MAX_BPF_JIT_REG + 4)
32 
33 #define check_imm(bits, imm) do {				\
34 	if ((((imm) > 0) && ((imm) >> (bits))) ||		\
35 	    (((imm) < 0) && (~(imm) >> (bits)))) {		\
36 		pr_info("[%2d] imm=%d(0x%x) out of range\n",	\
37 			i, imm, imm);				\
38 		return -EINVAL;					\
39 	}							\
40 } while (0)
41 #define check_imm19(imm) check_imm(19, imm)
42 #define check_imm26(imm) check_imm(26, imm)
43 
44 /* Map BPF registers to A64 registers */
45 static const int bpf2a64[] = {
46 	/* return value from in-kernel function, and exit value from eBPF */
47 	[BPF_REG_0] = A64_R(7),
48 	/* arguments from eBPF program to in-kernel function */
49 	[BPF_REG_1] = A64_R(0),
50 	[BPF_REG_2] = A64_R(1),
51 	[BPF_REG_3] = A64_R(2),
52 	[BPF_REG_4] = A64_R(3),
53 	[BPF_REG_5] = A64_R(4),
54 	/* callee saved registers that in-kernel function will preserve */
55 	[BPF_REG_6] = A64_R(19),
56 	[BPF_REG_7] = A64_R(20),
57 	[BPF_REG_8] = A64_R(21),
58 	[BPF_REG_9] = A64_R(22),
59 	/* read-only frame pointer to access stack */
60 	[BPF_REG_FP] = A64_R(25),
61 	/* temporary registers for BPF JIT */
62 	[TMP_REG_1] = A64_R(10),
63 	[TMP_REG_2] = A64_R(11),
64 	[TMP_REG_3] = A64_R(12),
65 	/* tail_call_cnt */
66 	[TCALL_CNT] = A64_R(26),
67 	/* temporary register for blinding constants */
68 	[BPF_REG_AX] = A64_R(9),
69 	[FP_BOTTOM] = A64_R(27),
70 };
71 
72 struct jit_ctx {
73 	const struct bpf_prog *prog;
74 	int idx;
75 	int epilogue_offset;
76 	int *offset;
77 	int exentry_idx;
78 	__le32 *image;
79 	u32 stack_size;
80 	int fpb_offset;
81 };
82 
83 struct bpf_plt {
84 	u32 insn_ldr; /* load target */
85 	u32 insn_br;  /* branch to target */
86 	u64 target;   /* target value */
87 };
88 
89 #define PLT_TARGET_SIZE   sizeof_field(struct bpf_plt, target)
90 #define PLT_TARGET_OFFSET offsetof(struct bpf_plt, target)
91 
92 static inline void emit(const u32 insn, struct jit_ctx *ctx)
93 {
94 	if (ctx->image != NULL)
95 		ctx->image[ctx->idx] = cpu_to_le32(insn);
96 
97 	ctx->idx++;
98 }
99 
100 static inline void emit_a64_mov_i(const int is64, const int reg,
101 				  const s32 val, struct jit_ctx *ctx)
102 {
103 	u16 hi = val >> 16;
104 	u16 lo = val & 0xffff;
105 
106 	if (hi & 0x8000) {
107 		if (hi == 0xffff) {
108 			emit(A64_MOVN(is64, reg, (u16)~lo, 0), ctx);
109 		} else {
110 			emit(A64_MOVN(is64, reg, (u16)~hi, 16), ctx);
111 			if (lo != 0xffff)
112 				emit(A64_MOVK(is64, reg, lo, 0), ctx);
113 		}
114 	} else {
115 		emit(A64_MOVZ(is64, reg, lo, 0), ctx);
116 		if (hi)
117 			emit(A64_MOVK(is64, reg, hi, 16), ctx);
118 	}
119 }
120 
121 static int i64_i16_blocks(const u64 val, bool inverse)
122 {
123 	return (((val >>  0) & 0xffff) != (inverse ? 0xffff : 0x0000)) +
124 	       (((val >> 16) & 0xffff) != (inverse ? 0xffff : 0x0000)) +
125 	       (((val >> 32) & 0xffff) != (inverse ? 0xffff : 0x0000)) +
126 	       (((val >> 48) & 0xffff) != (inverse ? 0xffff : 0x0000));
127 }
128 
129 static inline void emit_a64_mov_i64(const int reg, const u64 val,
130 				    struct jit_ctx *ctx)
131 {
132 	u64 nrm_tmp = val, rev_tmp = ~val;
133 	bool inverse;
134 	int shift;
135 
136 	if (!(nrm_tmp >> 32))
137 		return emit_a64_mov_i(0, reg, (u32)val, ctx);
138 
139 	inverse = i64_i16_blocks(nrm_tmp, true) < i64_i16_blocks(nrm_tmp, false);
140 	shift = max(round_down((inverse ? (fls64(rev_tmp) - 1) :
141 					  (fls64(nrm_tmp) - 1)), 16), 0);
142 	if (inverse)
143 		emit(A64_MOVN(1, reg, (rev_tmp >> shift) & 0xffff, shift), ctx);
144 	else
145 		emit(A64_MOVZ(1, reg, (nrm_tmp >> shift) & 0xffff, shift), ctx);
146 	shift -= 16;
147 	while (shift >= 0) {
148 		if (((nrm_tmp >> shift) & 0xffff) != (inverse ? 0xffff : 0x0000))
149 			emit(A64_MOVK(1, reg, (nrm_tmp >> shift) & 0xffff, shift), ctx);
150 		shift -= 16;
151 	}
152 }
153 
154 static inline void emit_bti(u32 insn, struct jit_ctx *ctx)
155 {
156 	if (IS_ENABLED(CONFIG_ARM64_BTI_KERNEL))
157 		emit(insn, ctx);
158 }
159 
160 /*
161  * Kernel addresses in the vmalloc space use at most 48 bits, and the
162  * remaining bits are guaranteed to be 0x1. So we can compose the address
163  * with a fixed length movn/movk/movk sequence.
164  */
165 static inline void emit_addr_mov_i64(const int reg, const u64 val,
166 				     struct jit_ctx *ctx)
167 {
168 	u64 tmp = val;
169 	int shift = 0;
170 
171 	emit(A64_MOVN(1, reg, ~tmp & 0xffff, shift), ctx);
172 	while (shift < 32) {
173 		tmp >>= 16;
174 		shift += 16;
175 		emit(A64_MOVK(1, reg, tmp & 0xffff, shift), ctx);
176 	}
177 }
178 
179 static inline void emit_call(u64 target, struct jit_ctx *ctx)
180 {
181 	u8 tmp = bpf2a64[TMP_REG_1];
182 
183 	emit_addr_mov_i64(tmp, target, ctx);
184 	emit(A64_BLR(tmp), ctx);
185 }
186 
187 static inline int bpf2a64_offset(int bpf_insn, int off,
188 				 const struct jit_ctx *ctx)
189 {
190 	/* BPF JMP offset is relative to the next instruction */
191 	bpf_insn++;
192 	/*
193 	 * Whereas arm64 branch instructions encode the offset
194 	 * from the branch itself, so we must subtract 1 from the
195 	 * instruction offset.
196 	 */
197 	return ctx->offset[bpf_insn + off] - (ctx->offset[bpf_insn] - 1);
198 }
199 
200 static void jit_fill_hole(void *area, unsigned int size)
201 {
202 	__le32 *ptr;
203 	/* We are guaranteed to have aligned memory. */
204 	for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
205 		*ptr++ = cpu_to_le32(AARCH64_BREAK_FAULT);
206 }
207 
208 static inline int epilogue_offset(const struct jit_ctx *ctx)
209 {
210 	int to = ctx->epilogue_offset;
211 	int from = ctx->idx;
212 
213 	return to - from;
214 }
215 
216 static bool is_addsub_imm(u32 imm)
217 {
218 	/* Either imm12 or shifted imm12. */
219 	return !(imm & ~0xfff) || !(imm & ~0xfff000);
220 }
221 
222 /*
223  * There are 3 types of AArch64 LDR/STR (immediate) instruction:
224  * Post-index, Pre-index, Unsigned offset.
225  *
226  * For BPF ldr/str, the "unsigned offset" type is sufficient.
227  *
228  * "Unsigned offset" type LDR(immediate) format:
229  *
230  *    3                   2                   1                   0
231  *  1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0
232  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
233  * |x x|1 1 1 0 0 1 0 1|         imm12         |    Rn   |    Rt   |
234  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
235  * scale
236  *
237  * "Unsigned offset" type STR(immediate) format:
238  *    3                   2                   1                   0
239  *  1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0
240  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
241  * |x x|1 1 1 0 0 1 0 0|         imm12         |    Rn   |    Rt   |
242  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
243  * scale
244  *
245  * The offset is calculated from imm12 and scale in the following way:
246  *
247  * offset = (u64)imm12 << scale
248  */
249 static bool is_lsi_offset(int offset, int scale)
250 {
251 	if (offset < 0)
252 		return false;
253 
254 	if (offset > (0xFFF << scale))
255 		return false;
256 
257 	if (offset & ((1 << scale) - 1))
258 		return false;
259 
260 	return true;
261 }
262 
263 /* generated prologue:
264  *      bti c // if CONFIG_ARM64_BTI_KERNEL
265  *      mov x9, lr
266  *      nop  // POKE_OFFSET
267  *      paciasp // if CONFIG_ARM64_PTR_AUTH_KERNEL
268  *      stp x29, lr, [sp, #-16]!
269  *      mov x29, sp
270  *      stp x19, x20, [sp, #-16]!
271  *      stp x21, x22, [sp, #-16]!
272  *      stp x25, x26, [sp, #-16]!
273  *      stp x27, x28, [sp, #-16]!
274  *      mov x25, sp
275  *      mov tcc, #0
276  *      // PROLOGUE_OFFSET
277  */
278 
279 #define BTI_INSNS (IS_ENABLED(CONFIG_ARM64_BTI_KERNEL) ? 1 : 0)
280 #define PAC_INSNS (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL) ? 1 : 0)
281 
282 /* Offset of nop instruction in bpf prog entry to be poked */
283 #define POKE_OFFSET (BTI_INSNS + 1)
284 
285 /* Tail call offset to jump into */
286 #define PROLOGUE_OFFSET (BTI_INSNS + 2 + PAC_INSNS + 8)
287 
288 static int build_prologue(struct jit_ctx *ctx, bool ebpf_from_cbpf)
289 {
290 	const struct bpf_prog *prog = ctx->prog;
291 	const bool is_main_prog = prog->aux->func_idx == 0;
292 	const u8 r6 = bpf2a64[BPF_REG_6];
293 	const u8 r7 = bpf2a64[BPF_REG_7];
294 	const u8 r8 = bpf2a64[BPF_REG_8];
295 	const u8 r9 = bpf2a64[BPF_REG_9];
296 	const u8 fp = bpf2a64[BPF_REG_FP];
297 	const u8 tcc = bpf2a64[TCALL_CNT];
298 	const u8 fpb = bpf2a64[FP_BOTTOM];
299 	const int idx0 = ctx->idx;
300 	int cur_offset;
301 
302 	/*
303 	 * BPF prog stack layout
304 	 *
305 	 *                         high
306 	 * original A64_SP =>   0:+-----+ BPF prologue
307 	 *                        |FP/LR|
308 	 * current A64_FP =>  -16:+-----+
309 	 *                        | ... | callee saved registers
310 	 * BPF fp register => -64:+-----+ <= (BPF_FP)
311 	 *                        |     |
312 	 *                        | ... | BPF prog stack
313 	 *                        |     |
314 	 *                        +-----+ <= (BPF_FP - prog->aux->stack_depth)
315 	 *                        |RSVD | padding
316 	 * current A64_SP =>      +-----+ <= (BPF_FP - ctx->stack_size)
317 	 *                        |     |
318 	 *                        | ... | Function call stack
319 	 *                        |     |
320 	 *                        +-----+
321 	 *                          low
322 	 *
323 	 */
324 
325 	/* bpf function may be invoked by 3 instruction types:
326 	 * 1. bl, attached via freplace to bpf prog via short jump
327 	 * 2. br, attached via freplace to bpf prog via long jump
328 	 * 3. blr, working as a function pointer, used by emit_call.
329 	 * So BTI_JC should used here to support both br and blr.
330 	 */
331 	emit_bti(A64_BTI_JC, ctx);
332 
333 	emit(A64_MOV(1, A64_R(9), A64_LR), ctx);
334 	emit(A64_NOP, ctx);
335 
336 	/* Sign lr */
337 	if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL))
338 		emit(A64_PACIASP, ctx);
339 
340 	/* Save FP and LR registers to stay align with ARM64 AAPCS */
341 	emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx);
342 	emit(A64_MOV(1, A64_FP, A64_SP), ctx);
343 
344 	/* Save callee-saved registers */
345 	emit(A64_PUSH(r6, r7, A64_SP), ctx);
346 	emit(A64_PUSH(r8, r9, A64_SP), ctx);
347 	emit(A64_PUSH(fp, tcc, A64_SP), ctx);
348 	emit(A64_PUSH(fpb, A64_R(28), A64_SP), ctx);
349 
350 	/* Set up BPF prog stack base register */
351 	emit(A64_MOV(1, fp, A64_SP), ctx);
352 
353 	if (!ebpf_from_cbpf && is_main_prog) {
354 		/* Initialize tail_call_cnt */
355 		emit(A64_MOVZ(1, tcc, 0, 0), ctx);
356 
357 		cur_offset = ctx->idx - idx0;
358 		if (cur_offset != PROLOGUE_OFFSET) {
359 			pr_err_once("PROLOGUE_OFFSET = %d, expected %d!\n",
360 				    cur_offset, PROLOGUE_OFFSET);
361 			return -1;
362 		}
363 
364 		/* BTI landing pad for the tail call, done with a BR */
365 		emit_bti(A64_BTI_J, ctx);
366 	}
367 
368 	emit(A64_SUB_I(1, fpb, fp, ctx->fpb_offset), ctx);
369 
370 	/* Stack must be multiples of 16B */
371 	ctx->stack_size = round_up(prog->aux->stack_depth, 16);
372 
373 	/* Set up function call stack */
374 	emit(A64_SUB_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
375 	return 0;
376 }
377 
378 static int out_offset = -1; /* initialized on the first pass of build_body() */
379 static int emit_bpf_tail_call(struct jit_ctx *ctx)
380 {
381 	/* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
382 	const u8 r2 = bpf2a64[BPF_REG_2];
383 	const u8 r3 = bpf2a64[BPF_REG_3];
384 
385 	const u8 tmp = bpf2a64[TMP_REG_1];
386 	const u8 prg = bpf2a64[TMP_REG_2];
387 	const u8 tcc = bpf2a64[TCALL_CNT];
388 	const int idx0 = ctx->idx;
389 #define cur_offset (ctx->idx - idx0)
390 #define jmp_offset (out_offset - (cur_offset))
391 	size_t off;
392 
393 	/* if (index >= array->map.max_entries)
394 	 *     goto out;
395 	 */
396 	off = offsetof(struct bpf_array, map.max_entries);
397 	emit_a64_mov_i64(tmp, off, ctx);
398 	emit(A64_LDR32(tmp, r2, tmp), ctx);
399 	emit(A64_MOV(0, r3, r3), ctx);
400 	emit(A64_CMP(0, r3, tmp), ctx);
401 	emit(A64_B_(A64_COND_CS, jmp_offset), ctx);
402 
403 	/*
404 	 * if (tail_call_cnt >= MAX_TAIL_CALL_CNT)
405 	 *     goto out;
406 	 * tail_call_cnt++;
407 	 */
408 	emit_a64_mov_i64(tmp, MAX_TAIL_CALL_CNT, ctx);
409 	emit(A64_CMP(1, tcc, tmp), ctx);
410 	emit(A64_B_(A64_COND_CS, jmp_offset), ctx);
411 	emit(A64_ADD_I(1, tcc, tcc, 1), ctx);
412 
413 	/* prog = array->ptrs[index];
414 	 * if (prog == NULL)
415 	 *     goto out;
416 	 */
417 	off = offsetof(struct bpf_array, ptrs);
418 	emit_a64_mov_i64(tmp, off, ctx);
419 	emit(A64_ADD(1, tmp, r2, tmp), ctx);
420 	emit(A64_LSL(1, prg, r3, 3), ctx);
421 	emit(A64_LDR64(prg, tmp, prg), ctx);
422 	emit(A64_CBZ(1, prg, jmp_offset), ctx);
423 
424 	/* goto *(prog->bpf_func + prologue_offset); */
425 	off = offsetof(struct bpf_prog, bpf_func);
426 	emit_a64_mov_i64(tmp, off, ctx);
427 	emit(A64_LDR64(tmp, prg, tmp), ctx);
428 	emit(A64_ADD_I(1, tmp, tmp, sizeof(u32) * PROLOGUE_OFFSET), ctx);
429 	emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
430 	emit(A64_BR(tmp), ctx);
431 
432 	/* out: */
433 	if (out_offset == -1)
434 		out_offset = cur_offset;
435 	if (cur_offset != out_offset) {
436 		pr_err_once("tail_call out_offset = %d, expected %d!\n",
437 			    cur_offset, out_offset);
438 		return -1;
439 	}
440 	return 0;
441 #undef cur_offset
442 #undef jmp_offset
443 }
444 
445 #ifdef CONFIG_ARM64_LSE_ATOMICS
446 static int emit_lse_atomic(const struct bpf_insn *insn, struct jit_ctx *ctx)
447 {
448 	const u8 code = insn->code;
449 	const u8 dst = bpf2a64[insn->dst_reg];
450 	const u8 src = bpf2a64[insn->src_reg];
451 	const u8 tmp = bpf2a64[TMP_REG_1];
452 	const u8 tmp2 = bpf2a64[TMP_REG_2];
453 	const bool isdw = BPF_SIZE(code) == BPF_DW;
454 	const s16 off = insn->off;
455 	u8 reg;
456 
457 	if (!off) {
458 		reg = dst;
459 	} else {
460 		emit_a64_mov_i(1, tmp, off, ctx);
461 		emit(A64_ADD(1, tmp, tmp, dst), ctx);
462 		reg = tmp;
463 	}
464 
465 	switch (insn->imm) {
466 	/* lock *(u32/u64 *)(dst_reg + off) <op>= src_reg */
467 	case BPF_ADD:
468 		emit(A64_STADD(isdw, reg, src), ctx);
469 		break;
470 	case BPF_AND:
471 		emit(A64_MVN(isdw, tmp2, src), ctx);
472 		emit(A64_STCLR(isdw, reg, tmp2), ctx);
473 		break;
474 	case BPF_OR:
475 		emit(A64_STSET(isdw, reg, src), ctx);
476 		break;
477 	case BPF_XOR:
478 		emit(A64_STEOR(isdw, reg, src), ctx);
479 		break;
480 	/* src_reg = atomic_fetch_<op>(dst_reg + off, src_reg) */
481 	case BPF_ADD | BPF_FETCH:
482 		emit(A64_LDADDAL(isdw, src, reg, src), ctx);
483 		break;
484 	case BPF_AND | BPF_FETCH:
485 		emit(A64_MVN(isdw, tmp2, src), ctx);
486 		emit(A64_LDCLRAL(isdw, src, reg, tmp2), ctx);
487 		break;
488 	case BPF_OR | BPF_FETCH:
489 		emit(A64_LDSETAL(isdw, src, reg, src), ctx);
490 		break;
491 	case BPF_XOR | BPF_FETCH:
492 		emit(A64_LDEORAL(isdw, src, reg, src), ctx);
493 		break;
494 	/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
495 	case BPF_XCHG:
496 		emit(A64_SWPAL(isdw, src, reg, src), ctx);
497 		break;
498 	/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
499 	case BPF_CMPXCHG:
500 		emit(A64_CASAL(isdw, src, reg, bpf2a64[BPF_REG_0]), ctx);
501 		break;
502 	default:
503 		pr_err_once("unknown atomic op code %02x\n", insn->imm);
504 		return -EINVAL;
505 	}
506 
507 	return 0;
508 }
509 #else
510 static inline int emit_lse_atomic(const struct bpf_insn *insn, struct jit_ctx *ctx)
511 {
512 	return -EINVAL;
513 }
514 #endif
515 
516 static int emit_ll_sc_atomic(const struct bpf_insn *insn, struct jit_ctx *ctx)
517 {
518 	const u8 code = insn->code;
519 	const u8 dst = bpf2a64[insn->dst_reg];
520 	const u8 src = bpf2a64[insn->src_reg];
521 	const u8 tmp = bpf2a64[TMP_REG_1];
522 	const u8 tmp2 = bpf2a64[TMP_REG_2];
523 	const u8 tmp3 = bpf2a64[TMP_REG_3];
524 	const int i = insn - ctx->prog->insnsi;
525 	const s32 imm = insn->imm;
526 	const s16 off = insn->off;
527 	const bool isdw = BPF_SIZE(code) == BPF_DW;
528 	u8 reg;
529 	s32 jmp_offset;
530 
531 	if (!off) {
532 		reg = dst;
533 	} else {
534 		emit_a64_mov_i(1, tmp, off, ctx);
535 		emit(A64_ADD(1, tmp, tmp, dst), ctx);
536 		reg = tmp;
537 	}
538 
539 	if (imm == BPF_ADD || imm == BPF_AND ||
540 	    imm == BPF_OR || imm == BPF_XOR) {
541 		/* lock *(u32/u64 *)(dst_reg + off) <op>= src_reg */
542 		emit(A64_LDXR(isdw, tmp2, reg), ctx);
543 		if (imm == BPF_ADD)
544 			emit(A64_ADD(isdw, tmp2, tmp2, src), ctx);
545 		else if (imm == BPF_AND)
546 			emit(A64_AND(isdw, tmp2, tmp2, src), ctx);
547 		else if (imm == BPF_OR)
548 			emit(A64_ORR(isdw, tmp2, tmp2, src), ctx);
549 		else
550 			emit(A64_EOR(isdw, tmp2, tmp2, src), ctx);
551 		emit(A64_STXR(isdw, tmp2, reg, tmp3), ctx);
552 		jmp_offset = -3;
553 		check_imm19(jmp_offset);
554 		emit(A64_CBNZ(0, tmp3, jmp_offset), ctx);
555 	} else if (imm == (BPF_ADD | BPF_FETCH) ||
556 		   imm == (BPF_AND | BPF_FETCH) ||
557 		   imm == (BPF_OR | BPF_FETCH) ||
558 		   imm == (BPF_XOR | BPF_FETCH)) {
559 		/* src_reg = atomic_fetch_<op>(dst_reg + off, src_reg) */
560 		const u8 ax = bpf2a64[BPF_REG_AX];
561 
562 		emit(A64_MOV(isdw, ax, src), ctx);
563 		emit(A64_LDXR(isdw, src, reg), ctx);
564 		if (imm == (BPF_ADD | BPF_FETCH))
565 			emit(A64_ADD(isdw, tmp2, src, ax), ctx);
566 		else if (imm == (BPF_AND | BPF_FETCH))
567 			emit(A64_AND(isdw, tmp2, src, ax), ctx);
568 		else if (imm == (BPF_OR | BPF_FETCH))
569 			emit(A64_ORR(isdw, tmp2, src, ax), ctx);
570 		else
571 			emit(A64_EOR(isdw, tmp2, src, ax), ctx);
572 		emit(A64_STLXR(isdw, tmp2, reg, tmp3), ctx);
573 		jmp_offset = -3;
574 		check_imm19(jmp_offset);
575 		emit(A64_CBNZ(0, tmp3, jmp_offset), ctx);
576 		emit(A64_DMB_ISH, ctx);
577 	} else if (imm == BPF_XCHG) {
578 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
579 		emit(A64_MOV(isdw, tmp2, src), ctx);
580 		emit(A64_LDXR(isdw, src, reg), ctx);
581 		emit(A64_STLXR(isdw, tmp2, reg, tmp3), ctx);
582 		jmp_offset = -2;
583 		check_imm19(jmp_offset);
584 		emit(A64_CBNZ(0, tmp3, jmp_offset), ctx);
585 		emit(A64_DMB_ISH, ctx);
586 	} else if (imm == BPF_CMPXCHG) {
587 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
588 		const u8 r0 = bpf2a64[BPF_REG_0];
589 
590 		emit(A64_MOV(isdw, tmp2, r0), ctx);
591 		emit(A64_LDXR(isdw, r0, reg), ctx);
592 		emit(A64_EOR(isdw, tmp3, r0, tmp2), ctx);
593 		jmp_offset = 4;
594 		check_imm19(jmp_offset);
595 		emit(A64_CBNZ(isdw, tmp3, jmp_offset), ctx);
596 		emit(A64_STLXR(isdw, src, reg, tmp3), ctx);
597 		jmp_offset = -4;
598 		check_imm19(jmp_offset);
599 		emit(A64_CBNZ(0, tmp3, jmp_offset), ctx);
600 		emit(A64_DMB_ISH, ctx);
601 	} else {
602 		pr_err_once("unknown atomic op code %02x\n", imm);
603 		return -EINVAL;
604 	}
605 
606 	return 0;
607 }
608 
609 void dummy_tramp(void);
610 
611 asm (
612 "	.pushsection .text, \"ax\", @progbits\n"
613 "	.global dummy_tramp\n"
614 "	.type dummy_tramp, %function\n"
615 "dummy_tramp:"
616 #if IS_ENABLED(CONFIG_ARM64_BTI_KERNEL)
617 "	bti j\n" /* dummy_tramp is called via "br x10" */
618 #endif
619 "	mov x10, x30\n"
620 "	mov x30, x9\n"
621 "	ret x10\n"
622 "	.size dummy_tramp, .-dummy_tramp\n"
623 "	.popsection\n"
624 );
625 
626 /* build a plt initialized like this:
627  *
628  * plt:
629  *      ldr tmp, target
630  *      br tmp
631  * target:
632  *      .quad dummy_tramp
633  *
634  * when a long jump trampoline is attached, target is filled with the
635  * trampoline address, and when the trampoline is removed, target is
636  * restored to dummy_tramp address.
637  */
638 static void build_plt(struct jit_ctx *ctx)
639 {
640 	const u8 tmp = bpf2a64[TMP_REG_1];
641 	struct bpf_plt *plt = NULL;
642 
643 	/* make sure target is 64-bit aligned */
644 	if ((ctx->idx + PLT_TARGET_OFFSET / AARCH64_INSN_SIZE) % 2)
645 		emit(A64_NOP, ctx);
646 
647 	plt = (struct bpf_plt *)(ctx->image + ctx->idx);
648 	/* plt is called via bl, no BTI needed here */
649 	emit(A64_LDR64LIT(tmp, 2 * AARCH64_INSN_SIZE), ctx);
650 	emit(A64_BR(tmp), ctx);
651 
652 	if (ctx->image)
653 		plt->target = (u64)&dummy_tramp;
654 }
655 
656 static void build_epilogue(struct jit_ctx *ctx)
657 {
658 	const u8 r0 = bpf2a64[BPF_REG_0];
659 	const u8 r6 = bpf2a64[BPF_REG_6];
660 	const u8 r7 = bpf2a64[BPF_REG_7];
661 	const u8 r8 = bpf2a64[BPF_REG_8];
662 	const u8 r9 = bpf2a64[BPF_REG_9];
663 	const u8 fp = bpf2a64[BPF_REG_FP];
664 	const u8 fpb = bpf2a64[FP_BOTTOM];
665 
666 	/* We're done with BPF stack */
667 	emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
668 
669 	/* Restore x27 and x28 */
670 	emit(A64_POP(fpb, A64_R(28), A64_SP), ctx);
671 	/* Restore fs (x25) and x26 */
672 	emit(A64_POP(fp, A64_R(26), A64_SP), ctx);
673 
674 	/* Restore callee-saved register */
675 	emit(A64_POP(r8, r9, A64_SP), ctx);
676 	emit(A64_POP(r6, r7, A64_SP), ctx);
677 
678 	/* Restore FP/LR registers */
679 	emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx);
680 
681 	/* Set return value */
682 	emit(A64_MOV(1, A64_R(0), r0), ctx);
683 
684 	/* Authenticate lr */
685 	if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL))
686 		emit(A64_AUTIASP, ctx);
687 
688 	emit(A64_RET(A64_LR), ctx);
689 }
690 
691 #define BPF_FIXUP_OFFSET_MASK	GENMASK(26, 0)
692 #define BPF_FIXUP_REG_MASK	GENMASK(31, 27)
693 
694 bool ex_handler_bpf(const struct exception_table_entry *ex,
695 		    struct pt_regs *regs)
696 {
697 	off_t offset = FIELD_GET(BPF_FIXUP_OFFSET_MASK, ex->fixup);
698 	int dst_reg = FIELD_GET(BPF_FIXUP_REG_MASK, ex->fixup);
699 
700 	regs->regs[dst_reg] = 0;
701 	regs->pc = (unsigned long)&ex->fixup - offset;
702 	return true;
703 }
704 
705 /* For accesses to BTF pointers, add an entry to the exception table */
706 static int add_exception_handler(const struct bpf_insn *insn,
707 				 struct jit_ctx *ctx,
708 				 int dst_reg)
709 {
710 	off_t offset;
711 	unsigned long pc;
712 	struct exception_table_entry *ex;
713 
714 	if (!ctx->image)
715 		/* First pass */
716 		return 0;
717 
718 	if (BPF_MODE(insn->code) != BPF_PROBE_MEM &&
719 		BPF_MODE(insn->code) != BPF_PROBE_MEMSX)
720 		return 0;
721 
722 	if (!ctx->prog->aux->extable ||
723 	    WARN_ON_ONCE(ctx->exentry_idx >= ctx->prog->aux->num_exentries))
724 		return -EINVAL;
725 
726 	ex = &ctx->prog->aux->extable[ctx->exentry_idx];
727 	pc = (unsigned long)&ctx->image[ctx->idx - 1];
728 
729 	offset = pc - (long)&ex->insn;
730 	if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
731 		return -ERANGE;
732 	ex->insn = offset;
733 
734 	/*
735 	 * Since the extable follows the program, the fixup offset is always
736 	 * negative and limited to BPF_JIT_REGION_SIZE. Store a positive value
737 	 * to keep things simple, and put the destination register in the upper
738 	 * bits. We don't need to worry about buildtime or runtime sort
739 	 * modifying the upper bits because the table is already sorted, and
740 	 * isn't part of the main exception table.
741 	 */
742 	offset = (long)&ex->fixup - (pc + AARCH64_INSN_SIZE);
743 	if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, offset))
744 		return -ERANGE;
745 
746 	ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, offset) |
747 		    FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg);
748 
749 	ex->type = EX_TYPE_BPF;
750 
751 	ctx->exentry_idx++;
752 	return 0;
753 }
754 
755 /* JITs an eBPF instruction.
756  * Returns:
757  * 0  - successfully JITed an 8-byte eBPF instruction.
758  * >0 - successfully JITed a 16-byte eBPF instruction.
759  * <0 - failed to JIT.
760  */
761 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx,
762 		      bool extra_pass)
763 {
764 	const u8 code = insn->code;
765 	const u8 dst = bpf2a64[insn->dst_reg];
766 	const u8 src = bpf2a64[insn->src_reg];
767 	const u8 tmp = bpf2a64[TMP_REG_1];
768 	const u8 tmp2 = bpf2a64[TMP_REG_2];
769 	const u8 fp = bpf2a64[BPF_REG_FP];
770 	const u8 fpb = bpf2a64[FP_BOTTOM];
771 	const s16 off = insn->off;
772 	const s32 imm = insn->imm;
773 	const int i = insn - ctx->prog->insnsi;
774 	const bool is64 = BPF_CLASS(code) == BPF_ALU64 ||
775 			  BPF_CLASS(code) == BPF_JMP;
776 	u8 jmp_cond;
777 	s32 jmp_offset;
778 	u32 a64_insn;
779 	u8 src_adj;
780 	u8 dst_adj;
781 	int off_adj;
782 	int ret;
783 	bool sign_extend;
784 
785 	switch (code) {
786 	/* dst = src */
787 	case BPF_ALU | BPF_MOV | BPF_X:
788 	case BPF_ALU64 | BPF_MOV | BPF_X:
789 		switch (insn->off) {
790 		case 0:
791 			emit(A64_MOV(is64, dst, src), ctx);
792 			break;
793 		case 8:
794 			emit(A64_SXTB(is64, dst, src), ctx);
795 			break;
796 		case 16:
797 			emit(A64_SXTH(is64, dst, src), ctx);
798 			break;
799 		case 32:
800 			emit(A64_SXTW(is64, dst, src), ctx);
801 			break;
802 		}
803 		break;
804 	/* dst = dst OP src */
805 	case BPF_ALU | BPF_ADD | BPF_X:
806 	case BPF_ALU64 | BPF_ADD | BPF_X:
807 		emit(A64_ADD(is64, dst, dst, src), ctx);
808 		break;
809 	case BPF_ALU | BPF_SUB | BPF_X:
810 	case BPF_ALU64 | BPF_SUB | BPF_X:
811 		emit(A64_SUB(is64, dst, dst, src), ctx);
812 		break;
813 	case BPF_ALU | BPF_AND | BPF_X:
814 	case BPF_ALU64 | BPF_AND | BPF_X:
815 		emit(A64_AND(is64, dst, dst, src), ctx);
816 		break;
817 	case BPF_ALU | BPF_OR | BPF_X:
818 	case BPF_ALU64 | BPF_OR | BPF_X:
819 		emit(A64_ORR(is64, dst, dst, src), ctx);
820 		break;
821 	case BPF_ALU | BPF_XOR | BPF_X:
822 	case BPF_ALU64 | BPF_XOR | BPF_X:
823 		emit(A64_EOR(is64, dst, dst, src), ctx);
824 		break;
825 	case BPF_ALU | BPF_MUL | BPF_X:
826 	case BPF_ALU64 | BPF_MUL | BPF_X:
827 		emit(A64_MUL(is64, dst, dst, src), ctx);
828 		break;
829 	case BPF_ALU | BPF_DIV | BPF_X:
830 	case BPF_ALU64 | BPF_DIV | BPF_X:
831 		if (!off)
832 			emit(A64_UDIV(is64, dst, dst, src), ctx);
833 		else
834 			emit(A64_SDIV(is64, dst, dst, src), ctx);
835 		break;
836 	case BPF_ALU | BPF_MOD | BPF_X:
837 	case BPF_ALU64 | BPF_MOD | BPF_X:
838 		if (!off)
839 			emit(A64_UDIV(is64, tmp, dst, src), ctx);
840 		else
841 			emit(A64_SDIV(is64, tmp, dst, src), ctx);
842 		emit(A64_MSUB(is64, dst, dst, tmp, src), ctx);
843 		break;
844 	case BPF_ALU | BPF_LSH | BPF_X:
845 	case BPF_ALU64 | BPF_LSH | BPF_X:
846 		emit(A64_LSLV(is64, dst, dst, src), ctx);
847 		break;
848 	case BPF_ALU | BPF_RSH | BPF_X:
849 	case BPF_ALU64 | BPF_RSH | BPF_X:
850 		emit(A64_LSRV(is64, dst, dst, src), ctx);
851 		break;
852 	case BPF_ALU | BPF_ARSH | BPF_X:
853 	case BPF_ALU64 | BPF_ARSH | BPF_X:
854 		emit(A64_ASRV(is64, dst, dst, src), ctx);
855 		break;
856 	/* dst = -dst */
857 	case BPF_ALU | BPF_NEG:
858 	case BPF_ALU64 | BPF_NEG:
859 		emit(A64_NEG(is64, dst, dst), ctx);
860 		break;
861 	/* dst = BSWAP##imm(dst) */
862 	case BPF_ALU | BPF_END | BPF_FROM_LE:
863 	case BPF_ALU | BPF_END | BPF_FROM_BE:
864 	case BPF_ALU64 | BPF_END | BPF_FROM_LE:
865 #ifdef CONFIG_CPU_BIG_ENDIAN
866 		if (BPF_CLASS(code) == BPF_ALU && BPF_SRC(code) == BPF_FROM_BE)
867 			goto emit_bswap_uxt;
868 #else /* !CONFIG_CPU_BIG_ENDIAN */
869 		if (BPF_CLASS(code) == BPF_ALU && BPF_SRC(code) == BPF_FROM_LE)
870 			goto emit_bswap_uxt;
871 #endif
872 		switch (imm) {
873 		case 16:
874 			emit(A64_REV16(is64, dst, dst), ctx);
875 			/* zero-extend 16 bits into 64 bits */
876 			emit(A64_UXTH(is64, dst, dst), ctx);
877 			break;
878 		case 32:
879 			emit(A64_REV32(is64, dst, dst), ctx);
880 			/* upper 32 bits already cleared */
881 			break;
882 		case 64:
883 			emit(A64_REV64(dst, dst), ctx);
884 			break;
885 		}
886 		break;
887 emit_bswap_uxt:
888 		switch (imm) {
889 		case 16:
890 			/* zero-extend 16 bits into 64 bits */
891 			emit(A64_UXTH(is64, dst, dst), ctx);
892 			break;
893 		case 32:
894 			/* zero-extend 32 bits into 64 bits */
895 			emit(A64_UXTW(is64, dst, dst), ctx);
896 			break;
897 		case 64:
898 			/* nop */
899 			break;
900 		}
901 		break;
902 	/* dst = imm */
903 	case BPF_ALU | BPF_MOV | BPF_K:
904 	case BPF_ALU64 | BPF_MOV | BPF_K:
905 		emit_a64_mov_i(is64, dst, imm, ctx);
906 		break;
907 	/* dst = dst OP imm */
908 	case BPF_ALU | BPF_ADD | BPF_K:
909 	case BPF_ALU64 | BPF_ADD | BPF_K:
910 		if (is_addsub_imm(imm)) {
911 			emit(A64_ADD_I(is64, dst, dst, imm), ctx);
912 		} else if (is_addsub_imm(-imm)) {
913 			emit(A64_SUB_I(is64, dst, dst, -imm), ctx);
914 		} else {
915 			emit_a64_mov_i(is64, tmp, imm, ctx);
916 			emit(A64_ADD(is64, dst, dst, tmp), ctx);
917 		}
918 		break;
919 	case BPF_ALU | BPF_SUB | BPF_K:
920 	case BPF_ALU64 | BPF_SUB | BPF_K:
921 		if (is_addsub_imm(imm)) {
922 			emit(A64_SUB_I(is64, dst, dst, imm), ctx);
923 		} else if (is_addsub_imm(-imm)) {
924 			emit(A64_ADD_I(is64, dst, dst, -imm), ctx);
925 		} else {
926 			emit_a64_mov_i(is64, tmp, imm, ctx);
927 			emit(A64_SUB(is64, dst, dst, tmp), ctx);
928 		}
929 		break;
930 	case BPF_ALU | BPF_AND | BPF_K:
931 	case BPF_ALU64 | BPF_AND | BPF_K:
932 		a64_insn = A64_AND_I(is64, dst, dst, imm);
933 		if (a64_insn != AARCH64_BREAK_FAULT) {
934 			emit(a64_insn, ctx);
935 		} else {
936 			emit_a64_mov_i(is64, tmp, imm, ctx);
937 			emit(A64_AND(is64, dst, dst, tmp), ctx);
938 		}
939 		break;
940 	case BPF_ALU | BPF_OR | BPF_K:
941 	case BPF_ALU64 | BPF_OR | BPF_K:
942 		a64_insn = A64_ORR_I(is64, dst, dst, imm);
943 		if (a64_insn != AARCH64_BREAK_FAULT) {
944 			emit(a64_insn, ctx);
945 		} else {
946 			emit_a64_mov_i(is64, tmp, imm, ctx);
947 			emit(A64_ORR(is64, dst, dst, tmp), ctx);
948 		}
949 		break;
950 	case BPF_ALU | BPF_XOR | BPF_K:
951 	case BPF_ALU64 | BPF_XOR | BPF_K:
952 		a64_insn = A64_EOR_I(is64, dst, dst, imm);
953 		if (a64_insn != AARCH64_BREAK_FAULT) {
954 			emit(a64_insn, ctx);
955 		} else {
956 			emit_a64_mov_i(is64, tmp, imm, ctx);
957 			emit(A64_EOR(is64, dst, dst, tmp), ctx);
958 		}
959 		break;
960 	case BPF_ALU | BPF_MUL | BPF_K:
961 	case BPF_ALU64 | BPF_MUL | BPF_K:
962 		emit_a64_mov_i(is64, tmp, imm, ctx);
963 		emit(A64_MUL(is64, dst, dst, tmp), ctx);
964 		break;
965 	case BPF_ALU | BPF_DIV | BPF_K:
966 	case BPF_ALU64 | BPF_DIV | BPF_K:
967 		emit_a64_mov_i(is64, tmp, imm, ctx);
968 		if (!off)
969 			emit(A64_UDIV(is64, dst, dst, tmp), ctx);
970 		else
971 			emit(A64_SDIV(is64, dst, dst, tmp), ctx);
972 		break;
973 	case BPF_ALU | BPF_MOD | BPF_K:
974 	case BPF_ALU64 | BPF_MOD | BPF_K:
975 		emit_a64_mov_i(is64, tmp2, imm, ctx);
976 		if (!off)
977 			emit(A64_UDIV(is64, tmp, dst, tmp2), ctx);
978 		else
979 			emit(A64_SDIV(is64, tmp, dst, tmp2), ctx);
980 		emit(A64_MSUB(is64, dst, dst, tmp, tmp2), ctx);
981 		break;
982 	case BPF_ALU | BPF_LSH | BPF_K:
983 	case BPF_ALU64 | BPF_LSH | BPF_K:
984 		emit(A64_LSL(is64, dst, dst, imm), ctx);
985 		break;
986 	case BPF_ALU | BPF_RSH | BPF_K:
987 	case BPF_ALU64 | BPF_RSH | BPF_K:
988 		emit(A64_LSR(is64, dst, dst, imm), ctx);
989 		break;
990 	case BPF_ALU | BPF_ARSH | BPF_K:
991 	case BPF_ALU64 | BPF_ARSH | BPF_K:
992 		emit(A64_ASR(is64, dst, dst, imm), ctx);
993 		break;
994 
995 	/* JUMP off */
996 	case BPF_JMP | BPF_JA:
997 	case BPF_JMP32 | BPF_JA:
998 		if (BPF_CLASS(code) == BPF_JMP)
999 			jmp_offset = bpf2a64_offset(i, off, ctx);
1000 		else
1001 			jmp_offset = bpf2a64_offset(i, imm, ctx);
1002 		check_imm26(jmp_offset);
1003 		emit(A64_B(jmp_offset), ctx);
1004 		break;
1005 	/* IF (dst COND src) JUMP off */
1006 	case BPF_JMP | BPF_JEQ | BPF_X:
1007 	case BPF_JMP | BPF_JGT | BPF_X:
1008 	case BPF_JMP | BPF_JLT | BPF_X:
1009 	case BPF_JMP | BPF_JGE | BPF_X:
1010 	case BPF_JMP | BPF_JLE | BPF_X:
1011 	case BPF_JMP | BPF_JNE | BPF_X:
1012 	case BPF_JMP | BPF_JSGT | BPF_X:
1013 	case BPF_JMP | BPF_JSLT | BPF_X:
1014 	case BPF_JMP | BPF_JSGE | BPF_X:
1015 	case BPF_JMP | BPF_JSLE | BPF_X:
1016 	case BPF_JMP32 | BPF_JEQ | BPF_X:
1017 	case BPF_JMP32 | BPF_JGT | BPF_X:
1018 	case BPF_JMP32 | BPF_JLT | BPF_X:
1019 	case BPF_JMP32 | BPF_JGE | BPF_X:
1020 	case BPF_JMP32 | BPF_JLE | BPF_X:
1021 	case BPF_JMP32 | BPF_JNE | BPF_X:
1022 	case BPF_JMP32 | BPF_JSGT | BPF_X:
1023 	case BPF_JMP32 | BPF_JSLT | BPF_X:
1024 	case BPF_JMP32 | BPF_JSGE | BPF_X:
1025 	case BPF_JMP32 | BPF_JSLE | BPF_X:
1026 		emit(A64_CMP(is64, dst, src), ctx);
1027 emit_cond_jmp:
1028 		jmp_offset = bpf2a64_offset(i, off, ctx);
1029 		check_imm19(jmp_offset);
1030 		switch (BPF_OP(code)) {
1031 		case BPF_JEQ:
1032 			jmp_cond = A64_COND_EQ;
1033 			break;
1034 		case BPF_JGT:
1035 			jmp_cond = A64_COND_HI;
1036 			break;
1037 		case BPF_JLT:
1038 			jmp_cond = A64_COND_CC;
1039 			break;
1040 		case BPF_JGE:
1041 			jmp_cond = A64_COND_CS;
1042 			break;
1043 		case BPF_JLE:
1044 			jmp_cond = A64_COND_LS;
1045 			break;
1046 		case BPF_JSET:
1047 		case BPF_JNE:
1048 			jmp_cond = A64_COND_NE;
1049 			break;
1050 		case BPF_JSGT:
1051 			jmp_cond = A64_COND_GT;
1052 			break;
1053 		case BPF_JSLT:
1054 			jmp_cond = A64_COND_LT;
1055 			break;
1056 		case BPF_JSGE:
1057 			jmp_cond = A64_COND_GE;
1058 			break;
1059 		case BPF_JSLE:
1060 			jmp_cond = A64_COND_LE;
1061 			break;
1062 		default:
1063 			return -EFAULT;
1064 		}
1065 		emit(A64_B_(jmp_cond, jmp_offset), ctx);
1066 		break;
1067 	case BPF_JMP | BPF_JSET | BPF_X:
1068 	case BPF_JMP32 | BPF_JSET | BPF_X:
1069 		emit(A64_TST(is64, dst, src), ctx);
1070 		goto emit_cond_jmp;
1071 	/* IF (dst COND imm) JUMP off */
1072 	case BPF_JMP | BPF_JEQ | BPF_K:
1073 	case BPF_JMP | BPF_JGT | BPF_K:
1074 	case BPF_JMP | BPF_JLT | BPF_K:
1075 	case BPF_JMP | BPF_JGE | BPF_K:
1076 	case BPF_JMP | BPF_JLE | BPF_K:
1077 	case BPF_JMP | BPF_JNE | BPF_K:
1078 	case BPF_JMP | BPF_JSGT | BPF_K:
1079 	case BPF_JMP | BPF_JSLT | BPF_K:
1080 	case BPF_JMP | BPF_JSGE | BPF_K:
1081 	case BPF_JMP | BPF_JSLE | BPF_K:
1082 	case BPF_JMP32 | BPF_JEQ | BPF_K:
1083 	case BPF_JMP32 | BPF_JGT | BPF_K:
1084 	case BPF_JMP32 | BPF_JLT | BPF_K:
1085 	case BPF_JMP32 | BPF_JGE | BPF_K:
1086 	case BPF_JMP32 | BPF_JLE | BPF_K:
1087 	case BPF_JMP32 | BPF_JNE | BPF_K:
1088 	case BPF_JMP32 | BPF_JSGT | BPF_K:
1089 	case BPF_JMP32 | BPF_JSLT | BPF_K:
1090 	case BPF_JMP32 | BPF_JSGE | BPF_K:
1091 	case BPF_JMP32 | BPF_JSLE | BPF_K:
1092 		if (is_addsub_imm(imm)) {
1093 			emit(A64_CMP_I(is64, dst, imm), ctx);
1094 		} else if (is_addsub_imm(-imm)) {
1095 			emit(A64_CMN_I(is64, dst, -imm), ctx);
1096 		} else {
1097 			emit_a64_mov_i(is64, tmp, imm, ctx);
1098 			emit(A64_CMP(is64, dst, tmp), ctx);
1099 		}
1100 		goto emit_cond_jmp;
1101 	case BPF_JMP | BPF_JSET | BPF_K:
1102 	case BPF_JMP32 | BPF_JSET | BPF_K:
1103 		a64_insn = A64_TST_I(is64, dst, imm);
1104 		if (a64_insn != AARCH64_BREAK_FAULT) {
1105 			emit(a64_insn, ctx);
1106 		} else {
1107 			emit_a64_mov_i(is64, tmp, imm, ctx);
1108 			emit(A64_TST(is64, dst, tmp), ctx);
1109 		}
1110 		goto emit_cond_jmp;
1111 	/* function call */
1112 	case BPF_JMP | BPF_CALL:
1113 	{
1114 		const u8 r0 = bpf2a64[BPF_REG_0];
1115 		bool func_addr_fixed;
1116 		u64 func_addr;
1117 
1118 		ret = bpf_jit_get_func_addr(ctx->prog, insn, extra_pass,
1119 					    &func_addr, &func_addr_fixed);
1120 		if (ret < 0)
1121 			return ret;
1122 		emit_call(func_addr, ctx);
1123 		emit(A64_MOV(1, r0, A64_R(0)), ctx);
1124 		break;
1125 	}
1126 	/* tail call */
1127 	case BPF_JMP | BPF_TAIL_CALL:
1128 		if (emit_bpf_tail_call(ctx))
1129 			return -EFAULT;
1130 		break;
1131 	/* function return */
1132 	case BPF_JMP | BPF_EXIT:
1133 		/* Optimization: when last instruction is EXIT,
1134 		   simply fallthrough to epilogue. */
1135 		if (i == ctx->prog->len - 1)
1136 			break;
1137 		jmp_offset = epilogue_offset(ctx);
1138 		check_imm26(jmp_offset);
1139 		emit(A64_B(jmp_offset), ctx);
1140 		break;
1141 
1142 	/* dst = imm64 */
1143 	case BPF_LD | BPF_IMM | BPF_DW:
1144 	{
1145 		const struct bpf_insn insn1 = insn[1];
1146 		u64 imm64;
1147 
1148 		imm64 = (u64)insn1.imm << 32 | (u32)imm;
1149 		if (bpf_pseudo_func(insn))
1150 			emit_addr_mov_i64(dst, imm64, ctx);
1151 		else
1152 			emit_a64_mov_i64(dst, imm64, ctx);
1153 
1154 		return 1;
1155 	}
1156 
1157 	/* LDX: dst = (u64)*(unsigned size *)(src + off) */
1158 	case BPF_LDX | BPF_MEM | BPF_W:
1159 	case BPF_LDX | BPF_MEM | BPF_H:
1160 	case BPF_LDX | BPF_MEM | BPF_B:
1161 	case BPF_LDX | BPF_MEM | BPF_DW:
1162 	case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1163 	case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1164 	case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1165 	case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1166 	/* LDXS: dst_reg = (s64)*(signed size *)(src_reg + off) */
1167 	case BPF_LDX | BPF_MEMSX | BPF_B:
1168 	case BPF_LDX | BPF_MEMSX | BPF_H:
1169 	case BPF_LDX | BPF_MEMSX | BPF_W:
1170 	case BPF_LDX | BPF_PROBE_MEMSX | BPF_B:
1171 	case BPF_LDX | BPF_PROBE_MEMSX | BPF_H:
1172 	case BPF_LDX | BPF_PROBE_MEMSX | BPF_W:
1173 		if (ctx->fpb_offset > 0 && src == fp) {
1174 			src_adj = fpb;
1175 			off_adj = off + ctx->fpb_offset;
1176 		} else {
1177 			src_adj = src;
1178 			off_adj = off;
1179 		}
1180 		sign_extend = (BPF_MODE(insn->code) == BPF_MEMSX ||
1181 				BPF_MODE(insn->code) == BPF_PROBE_MEMSX);
1182 		switch (BPF_SIZE(code)) {
1183 		case BPF_W:
1184 			if (is_lsi_offset(off_adj, 2)) {
1185 				if (sign_extend)
1186 					emit(A64_LDRSWI(dst, src_adj, off_adj), ctx);
1187 				else
1188 					emit(A64_LDR32I(dst, src_adj, off_adj), ctx);
1189 			} else {
1190 				emit_a64_mov_i(1, tmp, off, ctx);
1191 				if (sign_extend)
1192 					emit(A64_LDRSW(dst, src_adj, off_adj), ctx);
1193 				else
1194 					emit(A64_LDR32(dst, src, tmp), ctx);
1195 			}
1196 			break;
1197 		case BPF_H:
1198 			if (is_lsi_offset(off_adj, 1)) {
1199 				if (sign_extend)
1200 					emit(A64_LDRSHI(dst, src_adj, off_adj), ctx);
1201 				else
1202 					emit(A64_LDRHI(dst, src_adj, off_adj), ctx);
1203 			} else {
1204 				emit_a64_mov_i(1, tmp, off, ctx);
1205 				if (sign_extend)
1206 					emit(A64_LDRSH(dst, src, tmp), ctx);
1207 				else
1208 					emit(A64_LDRH(dst, src, tmp), ctx);
1209 			}
1210 			break;
1211 		case BPF_B:
1212 			if (is_lsi_offset(off_adj, 0)) {
1213 				if (sign_extend)
1214 					emit(A64_LDRSBI(dst, src_adj, off_adj), ctx);
1215 				else
1216 					emit(A64_LDRBI(dst, src_adj, off_adj), ctx);
1217 			} else {
1218 				emit_a64_mov_i(1, tmp, off, ctx);
1219 				if (sign_extend)
1220 					emit(A64_LDRSB(dst, src, tmp), ctx);
1221 				else
1222 					emit(A64_LDRB(dst, src, tmp), ctx);
1223 			}
1224 			break;
1225 		case BPF_DW:
1226 			if (is_lsi_offset(off_adj, 3)) {
1227 				emit(A64_LDR64I(dst, src_adj, off_adj), ctx);
1228 			} else {
1229 				emit_a64_mov_i(1, tmp, off, ctx);
1230 				emit(A64_LDR64(dst, src, tmp), ctx);
1231 			}
1232 			break;
1233 		}
1234 
1235 		ret = add_exception_handler(insn, ctx, dst);
1236 		if (ret)
1237 			return ret;
1238 		break;
1239 
1240 	/* speculation barrier */
1241 	case BPF_ST | BPF_NOSPEC:
1242 		/*
1243 		 * Nothing required here.
1244 		 *
1245 		 * In case of arm64, we rely on the firmware mitigation of
1246 		 * Speculative Store Bypass as controlled via the ssbd kernel
1247 		 * parameter. Whenever the mitigation is enabled, it works
1248 		 * for all of the kernel code with no need to provide any
1249 		 * additional instructions.
1250 		 */
1251 		break;
1252 
1253 	/* ST: *(size *)(dst + off) = imm */
1254 	case BPF_ST | BPF_MEM | BPF_W:
1255 	case BPF_ST | BPF_MEM | BPF_H:
1256 	case BPF_ST | BPF_MEM | BPF_B:
1257 	case BPF_ST | BPF_MEM | BPF_DW:
1258 		if (ctx->fpb_offset > 0 && dst == fp) {
1259 			dst_adj = fpb;
1260 			off_adj = off + ctx->fpb_offset;
1261 		} else {
1262 			dst_adj = dst;
1263 			off_adj = off;
1264 		}
1265 		/* Load imm to a register then store it */
1266 		emit_a64_mov_i(1, tmp, imm, ctx);
1267 		switch (BPF_SIZE(code)) {
1268 		case BPF_W:
1269 			if (is_lsi_offset(off_adj, 2)) {
1270 				emit(A64_STR32I(tmp, dst_adj, off_adj), ctx);
1271 			} else {
1272 				emit_a64_mov_i(1, tmp2, off, ctx);
1273 				emit(A64_STR32(tmp, dst, tmp2), ctx);
1274 			}
1275 			break;
1276 		case BPF_H:
1277 			if (is_lsi_offset(off_adj, 1)) {
1278 				emit(A64_STRHI(tmp, dst_adj, off_adj), ctx);
1279 			} else {
1280 				emit_a64_mov_i(1, tmp2, off, ctx);
1281 				emit(A64_STRH(tmp, dst, tmp2), ctx);
1282 			}
1283 			break;
1284 		case BPF_B:
1285 			if (is_lsi_offset(off_adj, 0)) {
1286 				emit(A64_STRBI(tmp, dst_adj, off_adj), ctx);
1287 			} else {
1288 				emit_a64_mov_i(1, tmp2, off, ctx);
1289 				emit(A64_STRB(tmp, dst, tmp2), ctx);
1290 			}
1291 			break;
1292 		case BPF_DW:
1293 			if (is_lsi_offset(off_adj, 3)) {
1294 				emit(A64_STR64I(tmp, dst_adj, off_adj), ctx);
1295 			} else {
1296 				emit_a64_mov_i(1, tmp2, off, ctx);
1297 				emit(A64_STR64(tmp, dst, tmp2), ctx);
1298 			}
1299 			break;
1300 		}
1301 		break;
1302 
1303 	/* STX: *(size *)(dst + off) = src */
1304 	case BPF_STX | BPF_MEM | BPF_W:
1305 	case BPF_STX | BPF_MEM | BPF_H:
1306 	case BPF_STX | BPF_MEM | BPF_B:
1307 	case BPF_STX | BPF_MEM | BPF_DW:
1308 		if (ctx->fpb_offset > 0 && dst == fp) {
1309 			dst_adj = fpb;
1310 			off_adj = off + ctx->fpb_offset;
1311 		} else {
1312 			dst_adj = dst;
1313 			off_adj = off;
1314 		}
1315 		switch (BPF_SIZE(code)) {
1316 		case BPF_W:
1317 			if (is_lsi_offset(off_adj, 2)) {
1318 				emit(A64_STR32I(src, dst_adj, off_adj), ctx);
1319 			} else {
1320 				emit_a64_mov_i(1, tmp, off, ctx);
1321 				emit(A64_STR32(src, dst, tmp), ctx);
1322 			}
1323 			break;
1324 		case BPF_H:
1325 			if (is_lsi_offset(off_adj, 1)) {
1326 				emit(A64_STRHI(src, dst_adj, off_adj), ctx);
1327 			} else {
1328 				emit_a64_mov_i(1, tmp, off, ctx);
1329 				emit(A64_STRH(src, dst, tmp), ctx);
1330 			}
1331 			break;
1332 		case BPF_B:
1333 			if (is_lsi_offset(off_adj, 0)) {
1334 				emit(A64_STRBI(src, dst_adj, off_adj), ctx);
1335 			} else {
1336 				emit_a64_mov_i(1, tmp, off, ctx);
1337 				emit(A64_STRB(src, dst, tmp), ctx);
1338 			}
1339 			break;
1340 		case BPF_DW:
1341 			if (is_lsi_offset(off_adj, 3)) {
1342 				emit(A64_STR64I(src, dst_adj, off_adj), ctx);
1343 			} else {
1344 				emit_a64_mov_i(1, tmp, off, ctx);
1345 				emit(A64_STR64(src, dst, tmp), ctx);
1346 			}
1347 			break;
1348 		}
1349 		break;
1350 
1351 	case BPF_STX | BPF_ATOMIC | BPF_W:
1352 	case BPF_STX | BPF_ATOMIC | BPF_DW:
1353 		if (cpus_have_cap(ARM64_HAS_LSE_ATOMICS))
1354 			ret = emit_lse_atomic(insn, ctx);
1355 		else
1356 			ret = emit_ll_sc_atomic(insn, ctx);
1357 		if (ret)
1358 			return ret;
1359 		break;
1360 
1361 	default:
1362 		pr_err_once("unknown opcode %02x\n", code);
1363 		return -EINVAL;
1364 	}
1365 
1366 	return 0;
1367 }
1368 
1369 /*
1370  * Return 0 if FP may change at runtime, otherwise find the minimum negative
1371  * offset to FP, converts it to positive number, and align down to 8 bytes.
1372  */
1373 static int find_fpb_offset(struct bpf_prog *prog)
1374 {
1375 	int i;
1376 	int offset = 0;
1377 
1378 	for (i = 0; i < prog->len; i++) {
1379 		const struct bpf_insn *insn = &prog->insnsi[i];
1380 		const u8 class = BPF_CLASS(insn->code);
1381 		const u8 mode = BPF_MODE(insn->code);
1382 		const u8 src = insn->src_reg;
1383 		const u8 dst = insn->dst_reg;
1384 		const s32 imm = insn->imm;
1385 		const s16 off = insn->off;
1386 
1387 		switch (class) {
1388 		case BPF_STX:
1389 		case BPF_ST:
1390 			/* fp holds atomic operation result */
1391 			if (class == BPF_STX && mode == BPF_ATOMIC &&
1392 			    ((imm == BPF_XCHG ||
1393 			      imm == (BPF_FETCH | BPF_ADD) ||
1394 			      imm == (BPF_FETCH | BPF_AND) ||
1395 			      imm == (BPF_FETCH | BPF_XOR) ||
1396 			      imm == (BPF_FETCH | BPF_OR)) &&
1397 			     src == BPF_REG_FP))
1398 				return 0;
1399 
1400 			if (mode == BPF_MEM && dst == BPF_REG_FP &&
1401 			    off < offset)
1402 				offset = insn->off;
1403 			break;
1404 
1405 		case BPF_JMP32:
1406 		case BPF_JMP:
1407 			break;
1408 
1409 		case BPF_LDX:
1410 		case BPF_LD:
1411 			/* fp holds load result */
1412 			if (dst == BPF_REG_FP)
1413 				return 0;
1414 
1415 			if (class == BPF_LDX && mode == BPF_MEM &&
1416 			    src == BPF_REG_FP && off < offset)
1417 				offset = off;
1418 			break;
1419 
1420 		case BPF_ALU:
1421 		case BPF_ALU64:
1422 		default:
1423 			/* fp holds ALU result */
1424 			if (dst == BPF_REG_FP)
1425 				return 0;
1426 		}
1427 	}
1428 
1429 	if (offset < 0) {
1430 		/*
1431 		 * safely be converted to a positive 'int', since insn->off
1432 		 * is 's16'
1433 		 */
1434 		offset = -offset;
1435 		/* align down to 8 bytes */
1436 		offset = ALIGN_DOWN(offset, 8);
1437 	}
1438 
1439 	return offset;
1440 }
1441 
1442 static int build_body(struct jit_ctx *ctx, bool extra_pass)
1443 {
1444 	const struct bpf_prog *prog = ctx->prog;
1445 	int i;
1446 
1447 	/*
1448 	 * - offset[0] offset of the end of prologue,
1449 	 *   start of the 1st instruction.
1450 	 * - offset[1] - offset of the end of 1st instruction,
1451 	 *   start of the 2nd instruction
1452 	 * [....]
1453 	 * - offset[3] - offset of the end of 3rd instruction,
1454 	 *   start of 4th instruction
1455 	 */
1456 	for (i = 0; i < prog->len; i++) {
1457 		const struct bpf_insn *insn = &prog->insnsi[i];
1458 		int ret;
1459 
1460 		if (ctx->image == NULL)
1461 			ctx->offset[i] = ctx->idx;
1462 		ret = build_insn(insn, ctx, extra_pass);
1463 		if (ret > 0) {
1464 			i++;
1465 			if (ctx->image == NULL)
1466 				ctx->offset[i] = ctx->idx;
1467 			continue;
1468 		}
1469 		if (ret)
1470 			return ret;
1471 	}
1472 	/*
1473 	 * offset is allocated with prog->len + 1 so fill in
1474 	 * the last element with the offset after the last
1475 	 * instruction (end of program)
1476 	 */
1477 	if (ctx->image == NULL)
1478 		ctx->offset[i] = ctx->idx;
1479 
1480 	return 0;
1481 }
1482 
1483 static int validate_code(struct jit_ctx *ctx)
1484 {
1485 	int i;
1486 
1487 	for (i = 0; i < ctx->idx; i++) {
1488 		u32 a64_insn = le32_to_cpu(ctx->image[i]);
1489 
1490 		if (a64_insn == AARCH64_BREAK_FAULT)
1491 			return -1;
1492 	}
1493 	return 0;
1494 }
1495 
1496 static int validate_ctx(struct jit_ctx *ctx)
1497 {
1498 	if (validate_code(ctx))
1499 		return -1;
1500 
1501 	if (WARN_ON_ONCE(ctx->exentry_idx != ctx->prog->aux->num_exentries))
1502 		return -1;
1503 
1504 	return 0;
1505 }
1506 
1507 static inline void bpf_flush_icache(void *start, void *end)
1508 {
1509 	flush_icache_range((unsigned long)start, (unsigned long)end);
1510 }
1511 
1512 struct arm64_jit_data {
1513 	struct bpf_binary_header *header;
1514 	u8 *image;
1515 	struct jit_ctx ctx;
1516 };
1517 
1518 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1519 {
1520 	int image_size, prog_size, extable_size, extable_align, extable_offset;
1521 	struct bpf_prog *tmp, *orig_prog = prog;
1522 	struct bpf_binary_header *header;
1523 	struct arm64_jit_data *jit_data;
1524 	bool was_classic = bpf_prog_was_classic(prog);
1525 	bool tmp_blinded = false;
1526 	bool extra_pass = false;
1527 	struct jit_ctx ctx;
1528 	u8 *image_ptr;
1529 
1530 	if (!prog->jit_requested)
1531 		return orig_prog;
1532 
1533 	tmp = bpf_jit_blind_constants(prog);
1534 	/* If blinding was requested and we failed during blinding,
1535 	 * we must fall back to the interpreter.
1536 	 */
1537 	if (IS_ERR(tmp))
1538 		return orig_prog;
1539 	if (tmp != prog) {
1540 		tmp_blinded = true;
1541 		prog = tmp;
1542 	}
1543 
1544 	jit_data = prog->aux->jit_data;
1545 	if (!jit_data) {
1546 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1547 		if (!jit_data) {
1548 			prog = orig_prog;
1549 			goto out;
1550 		}
1551 		prog->aux->jit_data = jit_data;
1552 	}
1553 	if (jit_data->ctx.offset) {
1554 		ctx = jit_data->ctx;
1555 		image_ptr = jit_data->image;
1556 		header = jit_data->header;
1557 		extra_pass = true;
1558 		prog_size = sizeof(u32) * ctx.idx;
1559 		goto skip_init_ctx;
1560 	}
1561 	memset(&ctx, 0, sizeof(ctx));
1562 	ctx.prog = prog;
1563 
1564 	ctx.offset = kvcalloc(prog->len + 1, sizeof(int), GFP_KERNEL);
1565 	if (ctx.offset == NULL) {
1566 		prog = orig_prog;
1567 		goto out_off;
1568 	}
1569 
1570 	ctx.fpb_offset = find_fpb_offset(prog);
1571 
1572 	/*
1573 	 * 1. Initial fake pass to compute ctx->idx and ctx->offset.
1574 	 *
1575 	 * BPF line info needs ctx->offset[i] to be the offset of
1576 	 * instruction[i] in jited image, so build prologue first.
1577 	 */
1578 	if (build_prologue(&ctx, was_classic)) {
1579 		prog = orig_prog;
1580 		goto out_off;
1581 	}
1582 
1583 	if (build_body(&ctx, extra_pass)) {
1584 		prog = orig_prog;
1585 		goto out_off;
1586 	}
1587 
1588 	ctx.epilogue_offset = ctx.idx;
1589 	build_epilogue(&ctx);
1590 	build_plt(&ctx);
1591 
1592 	extable_align = __alignof__(struct exception_table_entry);
1593 	extable_size = prog->aux->num_exentries *
1594 		sizeof(struct exception_table_entry);
1595 
1596 	/* Now we know the actual image size. */
1597 	prog_size = sizeof(u32) * ctx.idx;
1598 	/* also allocate space for plt target */
1599 	extable_offset = round_up(prog_size + PLT_TARGET_SIZE, extable_align);
1600 	image_size = extable_offset + extable_size;
1601 	header = bpf_jit_binary_alloc(image_size, &image_ptr,
1602 				      sizeof(u32), jit_fill_hole);
1603 	if (header == NULL) {
1604 		prog = orig_prog;
1605 		goto out_off;
1606 	}
1607 
1608 	/* 2. Now, the actual pass. */
1609 
1610 	ctx.image = (__le32 *)image_ptr;
1611 	if (extable_size)
1612 		prog->aux->extable = (void *)image_ptr + extable_offset;
1613 skip_init_ctx:
1614 	ctx.idx = 0;
1615 	ctx.exentry_idx = 0;
1616 
1617 	build_prologue(&ctx, was_classic);
1618 
1619 	if (build_body(&ctx, extra_pass)) {
1620 		bpf_jit_binary_free(header);
1621 		prog = orig_prog;
1622 		goto out_off;
1623 	}
1624 
1625 	build_epilogue(&ctx);
1626 	build_plt(&ctx);
1627 
1628 	/* 3. Extra pass to validate JITed code. */
1629 	if (validate_ctx(&ctx)) {
1630 		bpf_jit_binary_free(header);
1631 		prog = orig_prog;
1632 		goto out_off;
1633 	}
1634 
1635 	/* And we're done. */
1636 	if (bpf_jit_enable > 1)
1637 		bpf_jit_dump(prog->len, prog_size, 2, ctx.image);
1638 
1639 	bpf_flush_icache(header, ctx.image + ctx.idx);
1640 
1641 	if (!prog->is_func || extra_pass) {
1642 		if (extra_pass && ctx.idx != jit_data->ctx.idx) {
1643 			pr_err_once("multi-func JIT bug %d != %d\n",
1644 				    ctx.idx, jit_data->ctx.idx);
1645 			bpf_jit_binary_free(header);
1646 			prog->bpf_func = NULL;
1647 			prog->jited = 0;
1648 			prog->jited_len = 0;
1649 			goto out_off;
1650 		}
1651 		bpf_jit_binary_lock_ro(header);
1652 	} else {
1653 		jit_data->ctx = ctx;
1654 		jit_data->image = image_ptr;
1655 		jit_data->header = header;
1656 	}
1657 	prog->bpf_func = (void *)ctx.image;
1658 	prog->jited = 1;
1659 	prog->jited_len = prog_size;
1660 
1661 	if (!prog->is_func || extra_pass) {
1662 		int i;
1663 
1664 		/* offset[prog->len] is the size of program */
1665 		for (i = 0; i <= prog->len; i++)
1666 			ctx.offset[i] *= AARCH64_INSN_SIZE;
1667 		bpf_prog_fill_jited_linfo(prog, ctx.offset + 1);
1668 out_off:
1669 		kvfree(ctx.offset);
1670 		kfree(jit_data);
1671 		prog->aux->jit_data = NULL;
1672 	}
1673 out:
1674 	if (tmp_blinded)
1675 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
1676 					   tmp : orig_prog);
1677 	return prog;
1678 }
1679 
1680 bool bpf_jit_supports_kfunc_call(void)
1681 {
1682 	return true;
1683 }
1684 
1685 u64 bpf_jit_alloc_exec_limit(void)
1686 {
1687 	return VMALLOC_END - VMALLOC_START;
1688 }
1689 
1690 void *bpf_jit_alloc_exec(unsigned long size)
1691 {
1692 	/* Memory is intended to be executable, reset the pointer tag. */
1693 	return kasan_reset_tag(vmalloc(size));
1694 }
1695 
1696 void bpf_jit_free_exec(void *addr)
1697 {
1698 	return vfree(addr);
1699 }
1700 
1701 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
1702 bool bpf_jit_supports_subprog_tailcalls(void)
1703 {
1704 	return true;
1705 }
1706 
1707 static void invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_link *l,
1708 			    int args_off, int retval_off, int run_ctx_off,
1709 			    bool save_ret)
1710 {
1711 	__le32 *branch;
1712 	u64 enter_prog;
1713 	u64 exit_prog;
1714 	struct bpf_prog *p = l->link.prog;
1715 	int cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
1716 
1717 	enter_prog = (u64)bpf_trampoline_enter(p);
1718 	exit_prog = (u64)bpf_trampoline_exit(p);
1719 
1720 	if (l->cookie == 0) {
1721 		/* if cookie is zero, one instruction is enough to store it */
1722 		emit(A64_STR64I(A64_ZR, A64_SP, run_ctx_off + cookie_off), ctx);
1723 	} else {
1724 		emit_a64_mov_i64(A64_R(10), l->cookie, ctx);
1725 		emit(A64_STR64I(A64_R(10), A64_SP, run_ctx_off + cookie_off),
1726 		     ctx);
1727 	}
1728 
1729 	/* save p to callee saved register x19 to avoid loading p with mov_i64
1730 	 * each time.
1731 	 */
1732 	emit_addr_mov_i64(A64_R(19), (const u64)p, ctx);
1733 
1734 	/* arg1: prog */
1735 	emit(A64_MOV(1, A64_R(0), A64_R(19)), ctx);
1736 	/* arg2: &run_ctx */
1737 	emit(A64_ADD_I(1, A64_R(1), A64_SP, run_ctx_off), ctx);
1738 
1739 	emit_call(enter_prog, ctx);
1740 
1741 	/* if (__bpf_prog_enter(prog) == 0)
1742 	 *         goto skip_exec_of_prog;
1743 	 */
1744 	branch = ctx->image + ctx->idx;
1745 	emit(A64_NOP, ctx);
1746 
1747 	/* save return value to callee saved register x20 */
1748 	emit(A64_MOV(1, A64_R(20), A64_R(0)), ctx);
1749 
1750 	emit(A64_ADD_I(1, A64_R(0), A64_SP, args_off), ctx);
1751 	if (!p->jited)
1752 		emit_addr_mov_i64(A64_R(1), (const u64)p->insnsi, ctx);
1753 
1754 	emit_call((const u64)p->bpf_func, ctx);
1755 
1756 	if (save_ret)
1757 		emit(A64_STR64I(A64_R(0), A64_SP, retval_off), ctx);
1758 
1759 	if (ctx->image) {
1760 		int offset = &ctx->image[ctx->idx] - branch;
1761 		*branch = cpu_to_le32(A64_CBZ(1, A64_R(0), offset));
1762 	}
1763 
1764 	/* arg1: prog */
1765 	emit(A64_MOV(1, A64_R(0), A64_R(19)), ctx);
1766 	/* arg2: start time */
1767 	emit(A64_MOV(1, A64_R(1), A64_R(20)), ctx);
1768 	/* arg3: &run_ctx */
1769 	emit(A64_ADD_I(1, A64_R(2), A64_SP, run_ctx_off), ctx);
1770 
1771 	emit_call(exit_prog, ctx);
1772 }
1773 
1774 static void invoke_bpf_mod_ret(struct jit_ctx *ctx, struct bpf_tramp_links *tl,
1775 			       int args_off, int retval_off, int run_ctx_off,
1776 			       __le32 **branches)
1777 {
1778 	int i;
1779 
1780 	/* The first fmod_ret program will receive a garbage return value.
1781 	 * Set this to 0 to avoid confusing the program.
1782 	 */
1783 	emit(A64_STR64I(A64_ZR, A64_SP, retval_off), ctx);
1784 	for (i = 0; i < tl->nr_links; i++) {
1785 		invoke_bpf_prog(ctx, tl->links[i], args_off, retval_off,
1786 				run_ctx_off, true);
1787 		/* if (*(u64 *)(sp + retval_off) !=  0)
1788 		 *	goto do_fexit;
1789 		 */
1790 		emit(A64_LDR64I(A64_R(10), A64_SP, retval_off), ctx);
1791 		/* Save the location of branch, and generate a nop.
1792 		 * This nop will be replaced with a cbnz later.
1793 		 */
1794 		branches[i] = ctx->image + ctx->idx;
1795 		emit(A64_NOP, ctx);
1796 	}
1797 }
1798 
1799 static void save_args(struct jit_ctx *ctx, int args_off, int nregs)
1800 {
1801 	int i;
1802 
1803 	for (i = 0; i < nregs; i++) {
1804 		emit(A64_STR64I(i, A64_SP, args_off), ctx);
1805 		args_off += 8;
1806 	}
1807 }
1808 
1809 static void restore_args(struct jit_ctx *ctx, int args_off, int nregs)
1810 {
1811 	int i;
1812 
1813 	for (i = 0; i < nregs; i++) {
1814 		emit(A64_LDR64I(i, A64_SP, args_off), ctx);
1815 		args_off += 8;
1816 	}
1817 }
1818 
1819 /* Based on the x86's implementation of arch_prepare_bpf_trampoline().
1820  *
1821  * bpf prog and function entry before bpf trampoline hooked:
1822  *   mov x9, lr
1823  *   nop
1824  *
1825  * bpf prog and function entry after bpf trampoline hooked:
1826  *   mov x9, lr
1827  *   bl  <bpf_trampoline or plt>
1828  *
1829  */
1830 static int prepare_trampoline(struct jit_ctx *ctx, struct bpf_tramp_image *im,
1831 			      struct bpf_tramp_links *tlinks, void *orig_call,
1832 			      int nregs, u32 flags)
1833 {
1834 	int i;
1835 	int stack_size;
1836 	int retaddr_off;
1837 	int regs_off;
1838 	int retval_off;
1839 	int args_off;
1840 	int nregs_off;
1841 	int ip_off;
1842 	int run_ctx_off;
1843 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
1844 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
1845 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
1846 	bool save_ret;
1847 	__le32 **branches = NULL;
1848 
1849 	/* trampoline stack layout:
1850 	 *                  [ parent ip         ]
1851 	 *                  [ FP                ]
1852 	 * SP + retaddr_off [ self ip           ]
1853 	 *                  [ FP                ]
1854 	 *
1855 	 *                  [ padding           ] align SP to multiples of 16
1856 	 *
1857 	 *                  [ x20               ] callee saved reg x20
1858 	 * SP + regs_off    [ x19               ] callee saved reg x19
1859 	 *
1860 	 * SP + retval_off  [ return value      ] BPF_TRAMP_F_CALL_ORIG or
1861 	 *                                        BPF_TRAMP_F_RET_FENTRY_RET
1862 	 *
1863 	 *                  [ arg reg N         ]
1864 	 *                  [ ...               ]
1865 	 * SP + args_off    [ arg reg 1         ]
1866 	 *
1867 	 * SP + nregs_off   [ arg regs count    ]
1868 	 *
1869 	 * SP + ip_off      [ traced function   ] BPF_TRAMP_F_IP_ARG flag
1870 	 *
1871 	 * SP + run_ctx_off [ bpf_tramp_run_ctx ]
1872 	 */
1873 
1874 	stack_size = 0;
1875 	run_ctx_off = stack_size;
1876 	/* room for bpf_tramp_run_ctx */
1877 	stack_size += round_up(sizeof(struct bpf_tramp_run_ctx), 8);
1878 
1879 	ip_off = stack_size;
1880 	/* room for IP address argument */
1881 	if (flags & BPF_TRAMP_F_IP_ARG)
1882 		stack_size += 8;
1883 
1884 	nregs_off = stack_size;
1885 	/* room for args count */
1886 	stack_size += 8;
1887 
1888 	args_off = stack_size;
1889 	/* room for args */
1890 	stack_size += nregs * 8;
1891 
1892 	/* room for return value */
1893 	retval_off = stack_size;
1894 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
1895 	if (save_ret)
1896 		stack_size += 8;
1897 
1898 	/* room for callee saved registers, currently x19 and x20 are used */
1899 	regs_off = stack_size;
1900 	stack_size += 16;
1901 
1902 	/* round up to multiples of 16 to avoid SPAlignmentFault */
1903 	stack_size = round_up(stack_size, 16);
1904 
1905 	/* return address locates above FP */
1906 	retaddr_off = stack_size + 8;
1907 
1908 	/* bpf trampoline may be invoked by 3 instruction types:
1909 	 * 1. bl, attached to bpf prog or kernel function via short jump
1910 	 * 2. br, attached to bpf prog or kernel function via long jump
1911 	 * 3. blr, working as a function pointer, used by struct_ops.
1912 	 * So BTI_JC should used here to support both br and blr.
1913 	 */
1914 	emit_bti(A64_BTI_JC, ctx);
1915 
1916 	/* frame for parent function */
1917 	emit(A64_PUSH(A64_FP, A64_R(9), A64_SP), ctx);
1918 	emit(A64_MOV(1, A64_FP, A64_SP), ctx);
1919 
1920 	/* frame for patched function */
1921 	emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx);
1922 	emit(A64_MOV(1, A64_FP, A64_SP), ctx);
1923 
1924 	/* allocate stack space */
1925 	emit(A64_SUB_I(1, A64_SP, A64_SP, stack_size), ctx);
1926 
1927 	if (flags & BPF_TRAMP_F_IP_ARG) {
1928 		/* save ip address of the traced function */
1929 		emit_addr_mov_i64(A64_R(10), (const u64)orig_call, ctx);
1930 		emit(A64_STR64I(A64_R(10), A64_SP, ip_off), ctx);
1931 	}
1932 
1933 	/* save arg regs count*/
1934 	emit(A64_MOVZ(1, A64_R(10), nregs, 0), ctx);
1935 	emit(A64_STR64I(A64_R(10), A64_SP, nregs_off), ctx);
1936 
1937 	/* save arg regs */
1938 	save_args(ctx, args_off, nregs);
1939 
1940 	/* save callee saved registers */
1941 	emit(A64_STR64I(A64_R(19), A64_SP, regs_off), ctx);
1942 	emit(A64_STR64I(A64_R(20), A64_SP, regs_off + 8), ctx);
1943 
1944 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
1945 		emit_addr_mov_i64(A64_R(0), (const u64)im, ctx);
1946 		emit_call((const u64)__bpf_tramp_enter, ctx);
1947 	}
1948 
1949 	for (i = 0; i < fentry->nr_links; i++)
1950 		invoke_bpf_prog(ctx, fentry->links[i], args_off,
1951 				retval_off, run_ctx_off,
1952 				flags & BPF_TRAMP_F_RET_FENTRY_RET);
1953 
1954 	if (fmod_ret->nr_links) {
1955 		branches = kcalloc(fmod_ret->nr_links, sizeof(__le32 *),
1956 				   GFP_KERNEL);
1957 		if (!branches)
1958 			return -ENOMEM;
1959 
1960 		invoke_bpf_mod_ret(ctx, fmod_ret, args_off, retval_off,
1961 				   run_ctx_off, branches);
1962 	}
1963 
1964 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
1965 		restore_args(ctx, args_off, nregs);
1966 		/* call original func */
1967 		emit(A64_LDR64I(A64_R(10), A64_SP, retaddr_off), ctx);
1968 		emit(A64_ADR(A64_LR, AARCH64_INSN_SIZE * 2), ctx);
1969 		emit(A64_RET(A64_R(10)), ctx);
1970 		/* store return value */
1971 		emit(A64_STR64I(A64_R(0), A64_SP, retval_off), ctx);
1972 		/* reserve a nop for bpf_tramp_image_put */
1973 		im->ip_after_call = ctx->image + ctx->idx;
1974 		emit(A64_NOP, ctx);
1975 	}
1976 
1977 	/* update the branches saved in invoke_bpf_mod_ret with cbnz */
1978 	for (i = 0; i < fmod_ret->nr_links && ctx->image != NULL; i++) {
1979 		int offset = &ctx->image[ctx->idx] - branches[i];
1980 		*branches[i] = cpu_to_le32(A64_CBNZ(1, A64_R(10), offset));
1981 	}
1982 
1983 	for (i = 0; i < fexit->nr_links; i++)
1984 		invoke_bpf_prog(ctx, fexit->links[i], args_off, retval_off,
1985 				run_ctx_off, false);
1986 
1987 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
1988 		im->ip_epilogue = ctx->image + ctx->idx;
1989 		emit_addr_mov_i64(A64_R(0), (const u64)im, ctx);
1990 		emit_call((const u64)__bpf_tramp_exit, ctx);
1991 	}
1992 
1993 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
1994 		restore_args(ctx, args_off, nregs);
1995 
1996 	/* restore callee saved register x19 and x20 */
1997 	emit(A64_LDR64I(A64_R(19), A64_SP, regs_off), ctx);
1998 	emit(A64_LDR64I(A64_R(20), A64_SP, regs_off + 8), ctx);
1999 
2000 	if (save_ret)
2001 		emit(A64_LDR64I(A64_R(0), A64_SP, retval_off), ctx);
2002 
2003 	/* reset SP  */
2004 	emit(A64_MOV(1, A64_SP, A64_FP), ctx);
2005 
2006 	/* pop frames  */
2007 	emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx);
2008 	emit(A64_POP(A64_FP, A64_R(9), A64_SP), ctx);
2009 
2010 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2011 		/* skip patched function, return to parent */
2012 		emit(A64_MOV(1, A64_LR, A64_R(9)), ctx);
2013 		emit(A64_RET(A64_R(9)), ctx);
2014 	} else {
2015 		/* return to patched function */
2016 		emit(A64_MOV(1, A64_R(10), A64_LR), ctx);
2017 		emit(A64_MOV(1, A64_LR, A64_R(9)), ctx);
2018 		emit(A64_RET(A64_R(10)), ctx);
2019 	}
2020 
2021 	if (ctx->image)
2022 		bpf_flush_icache(ctx->image, ctx->image + ctx->idx);
2023 
2024 	kfree(branches);
2025 
2026 	return ctx->idx;
2027 }
2028 
2029 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image,
2030 				void *image_end, const struct btf_func_model *m,
2031 				u32 flags, struct bpf_tramp_links *tlinks,
2032 				void *orig_call)
2033 {
2034 	int i, ret;
2035 	int nregs = m->nr_args;
2036 	int max_insns = ((long)image_end - (long)image) / AARCH64_INSN_SIZE;
2037 	struct jit_ctx ctx = {
2038 		.image = NULL,
2039 		.idx = 0,
2040 	};
2041 
2042 	/* extra registers needed for struct argument */
2043 	for (i = 0; i < MAX_BPF_FUNC_ARGS; i++) {
2044 		/* The arg_size is at most 16 bytes, enforced by the verifier. */
2045 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
2046 			nregs += (m->arg_size[i] + 7) / 8 - 1;
2047 	}
2048 
2049 	/* the first 8 registers are used for arguments */
2050 	if (nregs > 8)
2051 		return -ENOTSUPP;
2052 
2053 	ret = prepare_trampoline(&ctx, im, tlinks, orig_call, nregs, flags);
2054 	if (ret < 0)
2055 		return ret;
2056 
2057 	if (ret > max_insns)
2058 		return -EFBIG;
2059 
2060 	ctx.image = image;
2061 	ctx.idx = 0;
2062 
2063 	jit_fill_hole(image, (unsigned int)(image_end - image));
2064 	ret = prepare_trampoline(&ctx, im, tlinks, orig_call, nregs, flags);
2065 
2066 	if (ret > 0 && validate_code(&ctx) < 0)
2067 		ret = -EINVAL;
2068 
2069 	if (ret > 0)
2070 		ret *= AARCH64_INSN_SIZE;
2071 
2072 	return ret;
2073 }
2074 
2075 static bool is_long_jump(void *ip, void *target)
2076 {
2077 	long offset;
2078 
2079 	/* NULL target means this is a NOP */
2080 	if (!target)
2081 		return false;
2082 
2083 	offset = (long)target - (long)ip;
2084 	return offset < -SZ_128M || offset >= SZ_128M;
2085 }
2086 
2087 static int gen_branch_or_nop(enum aarch64_insn_branch_type type, void *ip,
2088 			     void *addr, void *plt, u32 *insn)
2089 {
2090 	void *target;
2091 
2092 	if (!addr) {
2093 		*insn = aarch64_insn_gen_nop();
2094 		return 0;
2095 	}
2096 
2097 	if (is_long_jump(ip, addr))
2098 		target = plt;
2099 	else
2100 		target = addr;
2101 
2102 	*insn = aarch64_insn_gen_branch_imm((unsigned long)ip,
2103 					    (unsigned long)target,
2104 					    type);
2105 
2106 	return *insn != AARCH64_BREAK_FAULT ? 0 : -EFAULT;
2107 }
2108 
2109 /* Replace the branch instruction from @ip to @old_addr in a bpf prog or a bpf
2110  * trampoline with the branch instruction from @ip to @new_addr. If @old_addr
2111  * or @new_addr is NULL, the old or new instruction is NOP.
2112  *
2113  * When @ip is the bpf prog entry, a bpf trampoline is being attached or
2114  * detached. Since bpf trampoline and bpf prog are allocated separately with
2115  * vmalloc, the address distance may exceed 128MB, the maximum branch range.
2116  * So long jump should be handled.
2117  *
2118  * When a bpf prog is constructed, a plt pointing to empty trampoline
2119  * dummy_tramp is placed at the end:
2120  *
2121  *      bpf_prog:
2122  *              mov x9, lr
2123  *              nop // patchsite
2124  *              ...
2125  *              ret
2126  *
2127  *      plt:
2128  *              ldr x10, target
2129  *              br x10
2130  *      target:
2131  *              .quad dummy_tramp // plt target
2132  *
2133  * This is also the state when no trampoline is attached.
2134  *
2135  * When a short-jump bpf trampoline is attached, the patchsite is patched
2136  * to a bl instruction to the trampoline directly:
2137  *
2138  *      bpf_prog:
2139  *              mov x9, lr
2140  *              bl <short-jump bpf trampoline address> // patchsite
2141  *              ...
2142  *              ret
2143  *
2144  *      plt:
2145  *              ldr x10, target
2146  *              br x10
2147  *      target:
2148  *              .quad dummy_tramp // plt target
2149  *
2150  * When a long-jump bpf trampoline is attached, the plt target is filled with
2151  * the trampoline address and the patchsite is patched to a bl instruction to
2152  * the plt:
2153  *
2154  *      bpf_prog:
2155  *              mov x9, lr
2156  *              bl plt // patchsite
2157  *              ...
2158  *              ret
2159  *
2160  *      plt:
2161  *              ldr x10, target
2162  *              br x10
2163  *      target:
2164  *              .quad <long-jump bpf trampoline address> // plt target
2165  *
2166  * The dummy_tramp is used to prevent another CPU from jumping to unknown
2167  * locations during the patching process, making the patching process easier.
2168  */
2169 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type poke_type,
2170 		       void *old_addr, void *new_addr)
2171 {
2172 	int ret;
2173 	u32 old_insn;
2174 	u32 new_insn;
2175 	u32 replaced;
2176 	struct bpf_plt *plt = NULL;
2177 	unsigned long size = 0UL;
2178 	unsigned long offset = ~0UL;
2179 	enum aarch64_insn_branch_type branch_type;
2180 	char namebuf[KSYM_NAME_LEN];
2181 	void *image = NULL;
2182 	u64 plt_target = 0ULL;
2183 	bool poking_bpf_entry;
2184 
2185 	if (!__bpf_address_lookup((unsigned long)ip, &size, &offset, namebuf))
2186 		/* Only poking bpf text is supported. Since kernel function
2187 		 * entry is set up by ftrace, we reply on ftrace to poke kernel
2188 		 * functions.
2189 		 */
2190 		return -ENOTSUPP;
2191 
2192 	image = ip - offset;
2193 	/* zero offset means we're poking bpf prog entry */
2194 	poking_bpf_entry = (offset == 0UL);
2195 
2196 	/* bpf prog entry, find plt and the real patchsite */
2197 	if (poking_bpf_entry) {
2198 		/* plt locates at the end of bpf prog */
2199 		plt = image + size - PLT_TARGET_OFFSET;
2200 
2201 		/* skip to the nop instruction in bpf prog entry:
2202 		 * bti c // if BTI enabled
2203 		 * mov x9, x30
2204 		 * nop
2205 		 */
2206 		ip = image + POKE_OFFSET * AARCH64_INSN_SIZE;
2207 	}
2208 
2209 	/* long jump is only possible at bpf prog entry */
2210 	if (WARN_ON((is_long_jump(ip, new_addr) || is_long_jump(ip, old_addr)) &&
2211 		    !poking_bpf_entry))
2212 		return -EINVAL;
2213 
2214 	if (poke_type == BPF_MOD_CALL)
2215 		branch_type = AARCH64_INSN_BRANCH_LINK;
2216 	else
2217 		branch_type = AARCH64_INSN_BRANCH_NOLINK;
2218 
2219 	if (gen_branch_or_nop(branch_type, ip, old_addr, plt, &old_insn) < 0)
2220 		return -EFAULT;
2221 
2222 	if (gen_branch_or_nop(branch_type, ip, new_addr, plt, &new_insn) < 0)
2223 		return -EFAULT;
2224 
2225 	if (is_long_jump(ip, new_addr))
2226 		plt_target = (u64)new_addr;
2227 	else if (is_long_jump(ip, old_addr))
2228 		/* if the old target is a long jump and the new target is not,
2229 		 * restore the plt target to dummy_tramp, so there is always a
2230 		 * legal and harmless address stored in plt target, and we'll
2231 		 * never jump from plt to an unknown place.
2232 		 */
2233 		plt_target = (u64)&dummy_tramp;
2234 
2235 	if (plt_target) {
2236 		/* non-zero plt_target indicates we're patching a bpf prog,
2237 		 * which is read only.
2238 		 */
2239 		if (set_memory_rw(PAGE_MASK & ((uintptr_t)&plt->target), 1))
2240 			return -EFAULT;
2241 		WRITE_ONCE(plt->target, plt_target);
2242 		set_memory_ro(PAGE_MASK & ((uintptr_t)&plt->target), 1);
2243 		/* since plt target points to either the new trampoline
2244 		 * or dummy_tramp, even if another CPU reads the old plt
2245 		 * target value before fetching the bl instruction to plt,
2246 		 * it will be brought back by dummy_tramp, so no barrier is
2247 		 * required here.
2248 		 */
2249 	}
2250 
2251 	/* if the old target and the new target are both long jumps, no
2252 	 * patching is required
2253 	 */
2254 	if (old_insn == new_insn)
2255 		return 0;
2256 
2257 	mutex_lock(&text_mutex);
2258 	if (aarch64_insn_read(ip, &replaced)) {
2259 		ret = -EFAULT;
2260 		goto out;
2261 	}
2262 
2263 	if (replaced != old_insn) {
2264 		ret = -EFAULT;
2265 		goto out;
2266 	}
2267 
2268 	/* We call aarch64_insn_patch_text_nosync() to replace instruction
2269 	 * atomically, so no other CPUs will fetch a half-new and half-old
2270 	 * instruction. But there is chance that another CPU executes the
2271 	 * old instruction after the patching operation finishes (e.g.,
2272 	 * pipeline not flushed, or icache not synchronized yet).
2273 	 *
2274 	 * 1. when a new trampoline is attached, it is not a problem for
2275 	 *    different CPUs to jump to different trampolines temporarily.
2276 	 *
2277 	 * 2. when an old trampoline is freed, we should wait for all other
2278 	 *    CPUs to exit the trampoline and make sure the trampoline is no
2279 	 *    longer reachable, since bpf_tramp_image_put() function already
2280 	 *    uses percpu_ref and task-based rcu to do the sync, no need to call
2281 	 *    the sync version here, see bpf_tramp_image_put() for details.
2282 	 */
2283 	ret = aarch64_insn_patch_text_nosync(ip, new_insn);
2284 out:
2285 	mutex_unlock(&text_mutex);
2286 
2287 	return ret;
2288 }
2289