xref: /linux/arch/arm64/net/bpf_jit_comp.c (revision 7255fcc80d4b525cc10cfaaf7f485830d4ed2000)
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 	__le32 *ro_image;
80 	u32 stack_size;
81 	int fpb_offset;
82 };
83 
84 struct bpf_plt {
85 	u32 insn_ldr; /* load target */
86 	u32 insn_br;  /* branch to target */
87 	u64 target;   /* target value */
88 };
89 
90 #define PLT_TARGET_SIZE   sizeof_field(struct bpf_plt, target)
91 #define PLT_TARGET_OFFSET offsetof(struct bpf_plt, target)
92 
93 static inline void emit(const u32 insn, struct jit_ctx *ctx)
94 {
95 	if (ctx->image != NULL)
96 		ctx->image[ctx->idx] = cpu_to_le32(insn);
97 
98 	ctx->idx++;
99 }
100 
101 static inline void emit_a64_mov_i(const int is64, const int reg,
102 				  const s32 val, struct jit_ctx *ctx)
103 {
104 	u16 hi = val >> 16;
105 	u16 lo = val & 0xffff;
106 
107 	if (hi & 0x8000) {
108 		if (hi == 0xffff) {
109 			emit(A64_MOVN(is64, reg, (u16)~lo, 0), ctx);
110 		} else {
111 			emit(A64_MOVN(is64, reg, (u16)~hi, 16), ctx);
112 			if (lo != 0xffff)
113 				emit(A64_MOVK(is64, reg, lo, 0), ctx);
114 		}
115 	} else {
116 		emit(A64_MOVZ(is64, reg, lo, 0), ctx);
117 		if (hi)
118 			emit(A64_MOVK(is64, reg, hi, 16), ctx);
119 	}
120 }
121 
122 static int i64_i16_blocks(const u64 val, bool inverse)
123 {
124 	return (((val >>  0) & 0xffff) != (inverse ? 0xffff : 0x0000)) +
125 	       (((val >> 16) & 0xffff) != (inverse ? 0xffff : 0x0000)) +
126 	       (((val >> 32) & 0xffff) != (inverse ? 0xffff : 0x0000)) +
127 	       (((val >> 48) & 0xffff) != (inverse ? 0xffff : 0x0000));
128 }
129 
130 static inline void emit_a64_mov_i64(const int reg, const u64 val,
131 				    struct jit_ctx *ctx)
132 {
133 	u64 nrm_tmp = val, rev_tmp = ~val;
134 	bool inverse;
135 	int shift;
136 
137 	if (!(nrm_tmp >> 32))
138 		return emit_a64_mov_i(0, reg, (u32)val, ctx);
139 
140 	inverse = i64_i16_blocks(nrm_tmp, true) < i64_i16_blocks(nrm_tmp, false);
141 	shift = max(round_down((inverse ? (fls64(rev_tmp) - 1) :
142 					  (fls64(nrm_tmp) - 1)), 16), 0);
143 	if (inverse)
144 		emit(A64_MOVN(1, reg, (rev_tmp >> shift) & 0xffff, shift), ctx);
145 	else
146 		emit(A64_MOVZ(1, reg, (nrm_tmp >> shift) & 0xffff, shift), ctx);
147 	shift -= 16;
148 	while (shift >= 0) {
149 		if (((nrm_tmp >> shift) & 0xffff) != (inverse ? 0xffff : 0x0000))
150 			emit(A64_MOVK(1, reg, (nrm_tmp >> shift) & 0xffff, shift), ctx);
151 		shift -= 16;
152 	}
153 }
154 
155 static inline void emit_bti(u32 insn, struct jit_ctx *ctx)
156 {
157 	if (IS_ENABLED(CONFIG_ARM64_BTI_KERNEL))
158 		emit(insn, ctx);
159 }
160 
161 /*
162  * Kernel addresses in the vmalloc space use at most 48 bits, and the
163  * remaining bits are guaranteed to be 0x1. So we can compose the address
164  * with a fixed length movn/movk/movk sequence.
165  */
166 static inline void emit_addr_mov_i64(const int reg, const u64 val,
167 				     struct jit_ctx *ctx)
168 {
169 	u64 tmp = val;
170 	int shift = 0;
171 
172 	emit(A64_MOVN(1, reg, ~tmp & 0xffff, shift), ctx);
173 	while (shift < 32) {
174 		tmp >>= 16;
175 		shift += 16;
176 		emit(A64_MOVK(1, reg, tmp & 0xffff, shift), ctx);
177 	}
178 }
179 
180 static inline void emit_call(u64 target, struct jit_ctx *ctx)
181 {
182 	u8 tmp = bpf2a64[TMP_REG_1];
183 
184 	emit_addr_mov_i64(tmp, target, ctx);
185 	emit(A64_BLR(tmp), ctx);
186 }
187 
188 static inline int bpf2a64_offset(int bpf_insn, int off,
189 				 const struct jit_ctx *ctx)
190 {
191 	/* BPF JMP offset is relative to the next instruction */
192 	bpf_insn++;
193 	/*
194 	 * Whereas arm64 branch instructions encode the offset
195 	 * from the branch itself, so we must subtract 1 from the
196 	 * instruction offset.
197 	 */
198 	return ctx->offset[bpf_insn + off] - (ctx->offset[bpf_insn] - 1);
199 }
200 
201 static void jit_fill_hole(void *area, unsigned int size)
202 {
203 	__le32 *ptr;
204 	/* We are guaranteed to have aligned memory. */
205 	for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
206 		*ptr++ = cpu_to_le32(AARCH64_BREAK_FAULT);
207 }
208 
209 int bpf_arch_text_invalidate(void *dst, size_t len)
210 {
211 	if (!aarch64_insn_set(dst, AARCH64_BREAK_FAULT, len))
212 		return -EINVAL;
213 
214 	return 0;
215 }
216 
217 static inline int epilogue_offset(const struct jit_ctx *ctx)
218 {
219 	int to = ctx->epilogue_offset;
220 	int from = ctx->idx;
221 
222 	return to - from;
223 }
224 
225 static bool is_addsub_imm(u32 imm)
226 {
227 	/* Either imm12 or shifted imm12. */
228 	return !(imm & ~0xfff) || !(imm & ~0xfff000);
229 }
230 
231 /*
232  * There are 3 types of AArch64 LDR/STR (immediate) instruction:
233  * Post-index, Pre-index, Unsigned offset.
234  *
235  * For BPF ldr/str, the "unsigned offset" type is sufficient.
236  *
237  * "Unsigned offset" type LDR(immediate) format:
238  *
239  *    3                   2                   1                   0
240  *  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
241  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
242  * |x x|1 1 1 0 0 1 0 1|         imm12         |    Rn   |    Rt   |
243  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
244  * scale
245  *
246  * "Unsigned offset" type STR(immediate) format:
247  *    3                   2                   1                   0
248  *  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
249  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
250  * |x x|1 1 1 0 0 1 0 0|         imm12         |    Rn   |    Rt   |
251  * +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
252  * scale
253  *
254  * The offset is calculated from imm12 and scale in the following way:
255  *
256  * offset = (u64)imm12 << scale
257  */
258 static bool is_lsi_offset(int offset, int scale)
259 {
260 	if (offset < 0)
261 		return false;
262 
263 	if (offset > (0xFFF << scale))
264 		return false;
265 
266 	if (offset & ((1 << scale) - 1))
267 		return false;
268 
269 	return true;
270 }
271 
272 /* generated prologue:
273  *      bti c // if CONFIG_ARM64_BTI_KERNEL
274  *      mov x9, lr
275  *      nop  // POKE_OFFSET
276  *      paciasp // if CONFIG_ARM64_PTR_AUTH_KERNEL
277  *      stp x29, lr, [sp, #-16]!
278  *      mov x29, sp
279  *      stp x19, x20, [sp, #-16]!
280  *      stp x21, x22, [sp, #-16]!
281  *      stp x25, x26, [sp, #-16]!
282  *      stp x27, x28, [sp, #-16]!
283  *      mov x25, sp
284  *      mov tcc, #0
285  *      // PROLOGUE_OFFSET
286  */
287 
288 #define BTI_INSNS (IS_ENABLED(CONFIG_ARM64_BTI_KERNEL) ? 1 : 0)
289 #define PAC_INSNS (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL) ? 1 : 0)
290 
291 /* Offset of nop instruction in bpf prog entry to be poked */
292 #define POKE_OFFSET (BTI_INSNS + 1)
293 
294 /* Tail call offset to jump into */
295 #define PROLOGUE_OFFSET (BTI_INSNS + 2 + PAC_INSNS + 8)
296 
297 static int build_prologue(struct jit_ctx *ctx, bool ebpf_from_cbpf,
298 			  bool is_exception_cb)
299 {
300 	const struct bpf_prog *prog = ctx->prog;
301 	const bool is_main_prog = !bpf_is_subprog(prog);
302 	const u8 r6 = bpf2a64[BPF_REG_6];
303 	const u8 r7 = bpf2a64[BPF_REG_7];
304 	const u8 r8 = bpf2a64[BPF_REG_8];
305 	const u8 r9 = bpf2a64[BPF_REG_9];
306 	const u8 fp = bpf2a64[BPF_REG_FP];
307 	const u8 tcc = bpf2a64[TCALL_CNT];
308 	const u8 fpb = bpf2a64[FP_BOTTOM];
309 	const int idx0 = ctx->idx;
310 	int cur_offset;
311 
312 	/*
313 	 * BPF prog stack layout
314 	 *
315 	 *                         high
316 	 * original A64_SP =>   0:+-----+ BPF prologue
317 	 *                        |FP/LR|
318 	 * current A64_FP =>  -16:+-----+
319 	 *                        | ... | callee saved registers
320 	 * BPF fp register => -64:+-----+ <= (BPF_FP)
321 	 *                        |     |
322 	 *                        | ... | BPF prog stack
323 	 *                        |     |
324 	 *                        +-----+ <= (BPF_FP - prog->aux->stack_depth)
325 	 *                        |RSVD | padding
326 	 * current A64_SP =>      +-----+ <= (BPF_FP - ctx->stack_size)
327 	 *                        |     |
328 	 *                        | ... | Function call stack
329 	 *                        |     |
330 	 *                        +-----+
331 	 *                          low
332 	 *
333 	 */
334 
335 	/* bpf function may be invoked by 3 instruction types:
336 	 * 1. bl, attached via freplace to bpf prog via short jump
337 	 * 2. br, attached via freplace to bpf prog via long jump
338 	 * 3. blr, working as a function pointer, used by emit_call.
339 	 * So BTI_JC should used here to support both br and blr.
340 	 */
341 	emit_bti(A64_BTI_JC, ctx);
342 
343 	emit(A64_MOV(1, A64_R(9), A64_LR), ctx);
344 	emit(A64_NOP, ctx);
345 
346 	if (!is_exception_cb) {
347 		/* Sign lr */
348 		if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL))
349 			emit(A64_PACIASP, ctx);
350 		/* Save FP and LR registers to stay align with ARM64 AAPCS */
351 		emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx);
352 		emit(A64_MOV(1, A64_FP, A64_SP), ctx);
353 
354 		/* Save callee-saved registers */
355 		emit(A64_PUSH(r6, r7, A64_SP), ctx);
356 		emit(A64_PUSH(r8, r9, A64_SP), ctx);
357 		emit(A64_PUSH(fp, tcc, A64_SP), ctx);
358 		emit(A64_PUSH(fpb, A64_R(28), A64_SP), ctx);
359 	} else {
360 		/*
361 		 * Exception callback receives FP of Main Program as third
362 		 * parameter
363 		 */
364 		emit(A64_MOV(1, A64_FP, A64_R(2)), ctx);
365 		/*
366 		 * Main Program already pushed the frame record and the
367 		 * callee-saved registers. The exception callback will not push
368 		 * anything and re-use the main program's stack.
369 		 *
370 		 * 10 registers are on the stack
371 		 */
372 		emit(A64_SUB_I(1, A64_SP, A64_FP, 80), ctx);
373 	}
374 
375 	/* Set up BPF prog stack base register */
376 	emit(A64_MOV(1, fp, A64_SP), ctx);
377 
378 	if (!ebpf_from_cbpf && is_main_prog) {
379 		/* Initialize tail_call_cnt */
380 		emit(A64_MOVZ(1, tcc, 0, 0), ctx);
381 
382 		cur_offset = ctx->idx - idx0;
383 		if (cur_offset != PROLOGUE_OFFSET) {
384 			pr_err_once("PROLOGUE_OFFSET = %d, expected %d!\n",
385 				    cur_offset, PROLOGUE_OFFSET);
386 			return -1;
387 		}
388 
389 		/* BTI landing pad for the tail call, done with a BR */
390 		emit_bti(A64_BTI_J, ctx);
391 	}
392 
393 	/*
394 	 * Program acting as exception boundary should save all ARM64
395 	 * Callee-saved registers as the exception callback needs to recover
396 	 * all ARM64 Callee-saved registers in its epilogue.
397 	 */
398 	if (prog->aux->exception_boundary) {
399 		/*
400 		 * As we are pushing two more registers, BPF_FP should be moved
401 		 * 16 bytes
402 		 */
403 		emit(A64_SUB_I(1, fp, fp, 16), ctx);
404 		emit(A64_PUSH(A64_R(23), A64_R(24), A64_SP), ctx);
405 	}
406 
407 	emit(A64_SUB_I(1, fpb, fp, ctx->fpb_offset), ctx);
408 
409 	/* Stack must be multiples of 16B */
410 	ctx->stack_size = round_up(prog->aux->stack_depth, 16);
411 
412 	/* Set up function call stack */
413 	emit(A64_SUB_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
414 	return 0;
415 }
416 
417 static int out_offset = -1; /* initialized on the first pass of build_body() */
418 static int emit_bpf_tail_call(struct jit_ctx *ctx)
419 {
420 	/* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
421 	const u8 r2 = bpf2a64[BPF_REG_2];
422 	const u8 r3 = bpf2a64[BPF_REG_3];
423 
424 	const u8 tmp = bpf2a64[TMP_REG_1];
425 	const u8 prg = bpf2a64[TMP_REG_2];
426 	const u8 tcc = bpf2a64[TCALL_CNT];
427 	const int idx0 = ctx->idx;
428 #define cur_offset (ctx->idx - idx0)
429 #define jmp_offset (out_offset - (cur_offset))
430 	size_t off;
431 
432 	/* if (index >= array->map.max_entries)
433 	 *     goto out;
434 	 */
435 	off = offsetof(struct bpf_array, map.max_entries);
436 	emit_a64_mov_i64(tmp, off, ctx);
437 	emit(A64_LDR32(tmp, r2, tmp), ctx);
438 	emit(A64_MOV(0, r3, r3), ctx);
439 	emit(A64_CMP(0, r3, tmp), ctx);
440 	emit(A64_B_(A64_COND_CS, jmp_offset), ctx);
441 
442 	/*
443 	 * if (tail_call_cnt >= MAX_TAIL_CALL_CNT)
444 	 *     goto out;
445 	 * tail_call_cnt++;
446 	 */
447 	emit_a64_mov_i64(tmp, MAX_TAIL_CALL_CNT, ctx);
448 	emit(A64_CMP(1, tcc, tmp), ctx);
449 	emit(A64_B_(A64_COND_CS, jmp_offset), ctx);
450 	emit(A64_ADD_I(1, tcc, tcc, 1), ctx);
451 
452 	/* prog = array->ptrs[index];
453 	 * if (prog == NULL)
454 	 *     goto out;
455 	 */
456 	off = offsetof(struct bpf_array, ptrs);
457 	emit_a64_mov_i64(tmp, off, ctx);
458 	emit(A64_ADD(1, tmp, r2, tmp), ctx);
459 	emit(A64_LSL(1, prg, r3, 3), ctx);
460 	emit(A64_LDR64(prg, tmp, prg), ctx);
461 	emit(A64_CBZ(1, prg, jmp_offset), ctx);
462 
463 	/* goto *(prog->bpf_func + prologue_offset); */
464 	off = offsetof(struct bpf_prog, bpf_func);
465 	emit_a64_mov_i64(tmp, off, ctx);
466 	emit(A64_LDR64(tmp, prg, tmp), ctx);
467 	emit(A64_ADD_I(1, tmp, tmp, sizeof(u32) * PROLOGUE_OFFSET), ctx);
468 	emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
469 	emit(A64_BR(tmp), ctx);
470 
471 	/* out: */
472 	if (out_offset == -1)
473 		out_offset = cur_offset;
474 	if (cur_offset != out_offset) {
475 		pr_err_once("tail_call out_offset = %d, expected %d!\n",
476 			    cur_offset, out_offset);
477 		return -1;
478 	}
479 	return 0;
480 #undef cur_offset
481 #undef jmp_offset
482 }
483 
484 #ifdef CONFIG_ARM64_LSE_ATOMICS
485 static int emit_lse_atomic(const struct bpf_insn *insn, struct jit_ctx *ctx)
486 {
487 	const u8 code = insn->code;
488 	const u8 dst = bpf2a64[insn->dst_reg];
489 	const u8 src = bpf2a64[insn->src_reg];
490 	const u8 tmp = bpf2a64[TMP_REG_1];
491 	const u8 tmp2 = bpf2a64[TMP_REG_2];
492 	const bool isdw = BPF_SIZE(code) == BPF_DW;
493 	const s16 off = insn->off;
494 	u8 reg;
495 
496 	if (!off) {
497 		reg = dst;
498 	} else {
499 		emit_a64_mov_i(1, tmp, off, ctx);
500 		emit(A64_ADD(1, tmp, tmp, dst), ctx);
501 		reg = tmp;
502 	}
503 
504 	switch (insn->imm) {
505 	/* lock *(u32/u64 *)(dst_reg + off) <op>= src_reg */
506 	case BPF_ADD:
507 		emit(A64_STADD(isdw, reg, src), ctx);
508 		break;
509 	case BPF_AND:
510 		emit(A64_MVN(isdw, tmp2, src), ctx);
511 		emit(A64_STCLR(isdw, reg, tmp2), ctx);
512 		break;
513 	case BPF_OR:
514 		emit(A64_STSET(isdw, reg, src), ctx);
515 		break;
516 	case BPF_XOR:
517 		emit(A64_STEOR(isdw, reg, src), ctx);
518 		break;
519 	/* src_reg = atomic_fetch_<op>(dst_reg + off, src_reg) */
520 	case BPF_ADD | BPF_FETCH:
521 		emit(A64_LDADDAL(isdw, src, reg, src), ctx);
522 		break;
523 	case BPF_AND | BPF_FETCH:
524 		emit(A64_MVN(isdw, tmp2, src), ctx);
525 		emit(A64_LDCLRAL(isdw, src, reg, tmp2), ctx);
526 		break;
527 	case BPF_OR | BPF_FETCH:
528 		emit(A64_LDSETAL(isdw, src, reg, src), ctx);
529 		break;
530 	case BPF_XOR | BPF_FETCH:
531 		emit(A64_LDEORAL(isdw, src, reg, src), ctx);
532 		break;
533 	/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
534 	case BPF_XCHG:
535 		emit(A64_SWPAL(isdw, src, reg, src), ctx);
536 		break;
537 	/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
538 	case BPF_CMPXCHG:
539 		emit(A64_CASAL(isdw, src, reg, bpf2a64[BPF_REG_0]), ctx);
540 		break;
541 	default:
542 		pr_err_once("unknown atomic op code %02x\n", insn->imm);
543 		return -EINVAL;
544 	}
545 
546 	return 0;
547 }
548 #else
549 static inline int emit_lse_atomic(const struct bpf_insn *insn, struct jit_ctx *ctx)
550 {
551 	return -EINVAL;
552 }
553 #endif
554 
555 static int emit_ll_sc_atomic(const struct bpf_insn *insn, struct jit_ctx *ctx)
556 {
557 	const u8 code = insn->code;
558 	const u8 dst = bpf2a64[insn->dst_reg];
559 	const u8 src = bpf2a64[insn->src_reg];
560 	const u8 tmp = bpf2a64[TMP_REG_1];
561 	const u8 tmp2 = bpf2a64[TMP_REG_2];
562 	const u8 tmp3 = bpf2a64[TMP_REG_3];
563 	const int i = insn - ctx->prog->insnsi;
564 	const s32 imm = insn->imm;
565 	const s16 off = insn->off;
566 	const bool isdw = BPF_SIZE(code) == BPF_DW;
567 	u8 reg;
568 	s32 jmp_offset;
569 
570 	if (!off) {
571 		reg = dst;
572 	} else {
573 		emit_a64_mov_i(1, tmp, off, ctx);
574 		emit(A64_ADD(1, tmp, tmp, dst), ctx);
575 		reg = tmp;
576 	}
577 
578 	if (imm == BPF_ADD || imm == BPF_AND ||
579 	    imm == BPF_OR || imm == BPF_XOR) {
580 		/* lock *(u32/u64 *)(dst_reg + off) <op>= src_reg */
581 		emit(A64_LDXR(isdw, tmp2, reg), ctx);
582 		if (imm == BPF_ADD)
583 			emit(A64_ADD(isdw, tmp2, tmp2, src), ctx);
584 		else if (imm == BPF_AND)
585 			emit(A64_AND(isdw, tmp2, tmp2, src), ctx);
586 		else if (imm == BPF_OR)
587 			emit(A64_ORR(isdw, tmp2, tmp2, src), ctx);
588 		else
589 			emit(A64_EOR(isdw, tmp2, tmp2, src), ctx);
590 		emit(A64_STXR(isdw, tmp2, reg, tmp3), ctx);
591 		jmp_offset = -3;
592 		check_imm19(jmp_offset);
593 		emit(A64_CBNZ(0, tmp3, jmp_offset), ctx);
594 	} else if (imm == (BPF_ADD | BPF_FETCH) ||
595 		   imm == (BPF_AND | BPF_FETCH) ||
596 		   imm == (BPF_OR | BPF_FETCH) ||
597 		   imm == (BPF_XOR | BPF_FETCH)) {
598 		/* src_reg = atomic_fetch_<op>(dst_reg + off, src_reg) */
599 		const u8 ax = bpf2a64[BPF_REG_AX];
600 
601 		emit(A64_MOV(isdw, ax, src), ctx);
602 		emit(A64_LDXR(isdw, src, reg), ctx);
603 		if (imm == (BPF_ADD | BPF_FETCH))
604 			emit(A64_ADD(isdw, tmp2, src, ax), ctx);
605 		else if (imm == (BPF_AND | BPF_FETCH))
606 			emit(A64_AND(isdw, tmp2, src, ax), ctx);
607 		else if (imm == (BPF_OR | BPF_FETCH))
608 			emit(A64_ORR(isdw, tmp2, src, ax), ctx);
609 		else
610 			emit(A64_EOR(isdw, tmp2, src, ax), ctx);
611 		emit(A64_STLXR(isdw, tmp2, reg, tmp3), ctx);
612 		jmp_offset = -3;
613 		check_imm19(jmp_offset);
614 		emit(A64_CBNZ(0, tmp3, jmp_offset), ctx);
615 		emit(A64_DMB_ISH, ctx);
616 	} else if (imm == BPF_XCHG) {
617 		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
618 		emit(A64_MOV(isdw, tmp2, src), ctx);
619 		emit(A64_LDXR(isdw, src, reg), ctx);
620 		emit(A64_STLXR(isdw, tmp2, reg, tmp3), ctx);
621 		jmp_offset = -2;
622 		check_imm19(jmp_offset);
623 		emit(A64_CBNZ(0, tmp3, jmp_offset), ctx);
624 		emit(A64_DMB_ISH, ctx);
625 	} else if (imm == BPF_CMPXCHG) {
626 		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
627 		const u8 r0 = bpf2a64[BPF_REG_0];
628 
629 		emit(A64_MOV(isdw, tmp2, r0), ctx);
630 		emit(A64_LDXR(isdw, r0, reg), ctx);
631 		emit(A64_EOR(isdw, tmp3, r0, tmp2), ctx);
632 		jmp_offset = 4;
633 		check_imm19(jmp_offset);
634 		emit(A64_CBNZ(isdw, tmp3, jmp_offset), ctx);
635 		emit(A64_STLXR(isdw, src, reg, tmp3), ctx);
636 		jmp_offset = -4;
637 		check_imm19(jmp_offset);
638 		emit(A64_CBNZ(0, tmp3, jmp_offset), ctx);
639 		emit(A64_DMB_ISH, ctx);
640 	} else {
641 		pr_err_once("unknown atomic op code %02x\n", imm);
642 		return -EINVAL;
643 	}
644 
645 	return 0;
646 }
647 
648 void dummy_tramp(void);
649 
650 asm (
651 "	.pushsection .text, \"ax\", @progbits\n"
652 "	.global dummy_tramp\n"
653 "	.type dummy_tramp, %function\n"
654 "dummy_tramp:"
655 #if IS_ENABLED(CONFIG_ARM64_BTI_KERNEL)
656 "	bti j\n" /* dummy_tramp is called via "br x10" */
657 #endif
658 "	mov x10, x30\n"
659 "	mov x30, x9\n"
660 "	ret x10\n"
661 "	.size dummy_tramp, .-dummy_tramp\n"
662 "	.popsection\n"
663 );
664 
665 /* build a plt initialized like this:
666  *
667  * plt:
668  *      ldr tmp, target
669  *      br tmp
670  * target:
671  *      .quad dummy_tramp
672  *
673  * when a long jump trampoline is attached, target is filled with the
674  * trampoline address, and when the trampoline is removed, target is
675  * restored to dummy_tramp address.
676  */
677 static void build_plt(struct jit_ctx *ctx)
678 {
679 	const u8 tmp = bpf2a64[TMP_REG_1];
680 	struct bpf_plt *plt = NULL;
681 
682 	/* make sure target is 64-bit aligned */
683 	if ((ctx->idx + PLT_TARGET_OFFSET / AARCH64_INSN_SIZE) % 2)
684 		emit(A64_NOP, ctx);
685 
686 	plt = (struct bpf_plt *)(ctx->image + ctx->idx);
687 	/* plt is called via bl, no BTI needed here */
688 	emit(A64_LDR64LIT(tmp, 2 * AARCH64_INSN_SIZE), ctx);
689 	emit(A64_BR(tmp), ctx);
690 
691 	if (ctx->image)
692 		plt->target = (u64)&dummy_tramp;
693 }
694 
695 static void build_epilogue(struct jit_ctx *ctx, bool is_exception_cb)
696 {
697 	const u8 r0 = bpf2a64[BPF_REG_0];
698 	const u8 r6 = bpf2a64[BPF_REG_6];
699 	const u8 r7 = bpf2a64[BPF_REG_7];
700 	const u8 r8 = bpf2a64[BPF_REG_8];
701 	const u8 r9 = bpf2a64[BPF_REG_9];
702 	const u8 fp = bpf2a64[BPF_REG_FP];
703 	const u8 fpb = bpf2a64[FP_BOTTOM];
704 
705 	/* We're done with BPF stack */
706 	emit(A64_ADD_I(1, A64_SP, A64_SP, ctx->stack_size), ctx);
707 
708 	/*
709 	 * Program acting as exception boundary pushes R23 and R24 in addition
710 	 * to BPF callee-saved registers. Exception callback uses the boundary
711 	 * program's stack frame, so recover these extra registers in the above
712 	 * two cases.
713 	 */
714 	if (ctx->prog->aux->exception_boundary || is_exception_cb)
715 		emit(A64_POP(A64_R(23), A64_R(24), A64_SP), ctx);
716 
717 	/* Restore x27 and x28 */
718 	emit(A64_POP(fpb, A64_R(28), A64_SP), ctx);
719 	/* Restore fs (x25) and x26 */
720 	emit(A64_POP(fp, A64_R(26), A64_SP), ctx);
721 
722 	/* Restore callee-saved register */
723 	emit(A64_POP(r8, r9, A64_SP), ctx);
724 	emit(A64_POP(r6, r7, A64_SP), ctx);
725 
726 	/* Restore FP/LR registers */
727 	emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx);
728 
729 	/* Set return value */
730 	emit(A64_MOV(1, A64_R(0), r0), ctx);
731 
732 	/* Authenticate lr */
733 	if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL))
734 		emit(A64_AUTIASP, ctx);
735 
736 	emit(A64_RET(A64_LR), ctx);
737 }
738 
739 #define BPF_FIXUP_OFFSET_MASK	GENMASK(26, 0)
740 #define BPF_FIXUP_REG_MASK	GENMASK(31, 27)
741 
742 bool ex_handler_bpf(const struct exception_table_entry *ex,
743 		    struct pt_regs *regs)
744 {
745 	off_t offset = FIELD_GET(BPF_FIXUP_OFFSET_MASK, ex->fixup);
746 	int dst_reg = FIELD_GET(BPF_FIXUP_REG_MASK, ex->fixup);
747 
748 	regs->regs[dst_reg] = 0;
749 	regs->pc = (unsigned long)&ex->fixup - offset;
750 	return true;
751 }
752 
753 /* For accesses to BTF pointers, add an entry to the exception table */
754 static int add_exception_handler(const struct bpf_insn *insn,
755 				 struct jit_ctx *ctx,
756 				 int dst_reg)
757 {
758 	off_t ins_offset;
759 	off_t fixup_offset;
760 	unsigned long pc;
761 	struct exception_table_entry *ex;
762 
763 	if (!ctx->image)
764 		/* First pass */
765 		return 0;
766 
767 	if (BPF_MODE(insn->code) != BPF_PROBE_MEM &&
768 		BPF_MODE(insn->code) != BPF_PROBE_MEMSX)
769 		return 0;
770 
771 	if (!ctx->prog->aux->extable ||
772 	    WARN_ON_ONCE(ctx->exentry_idx >= ctx->prog->aux->num_exentries))
773 		return -EINVAL;
774 
775 	ex = &ctx->prog->aux->extable[ctx->exentry_idx];
776 	pc = (unsigned long)&ctx->ro_image[ctx->idx - 1];
777 
778 	/*
779 	 * This is the relative offset of the instruction that may fault from
780 	 * the exception table itself. This will be written to the exception
781 	 * table and if this instruction faults, the destination register will
782 	 * be set to '0' and the execution will jump to the next instruction.
783 	 */
784 	ins_offset = pc - (long)&ex->insn;
785 	if (WARN_ON_ONCE(ins_offset >= 0 || ins_offset < INT_MIN))
786 		return -ERANGE;
787 
788 	/*
789 	 * Since the extable follows the program, the fixup offset is always
790 	 * negative and limited to BPF_JIT_REGION_SIZE. Store a positive value
791 	 * to keep things simple, and put the destination register in the upper
792 	 * bits. We don't need to worry about buildtime or runtime sort
793 	 * modifying the upper bits because the table is already sorted, and
794 	 * isn't part of the main exception table.
795 	 *
796 	 * The fixup_offset is set to the next instruction from the instruction
797 	 * that may fault. The execution will jump to this after handling the
798 	 * fault.
799 	 */
800 	fixup_offset = (long)&ex->fixup - (pc + AARCH64_INSN_SIZE);
801 	if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, fixup_offset))
802 		return -ERANGE;
803 
804 	/*
805 	 * The offsets above have been calculated using the RO buffer but we
806 	 * need to use the R/W buffer for writes.
807 	 * switch ex to rw buffer for writing.
808 	 */
809 	ex = (void *)ctx->image + ((void *)ex - (void *)ctx->ro_image);
810 
811 	ex->insn = ins_offset;
812 
813 	ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, fixup_offset) |
814 		    FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg);
815 
816 	ex->type = EX_TYPE_BPF;
817 
818 	ctx->exentry_idx++;
819 	return 0;
820 }
821 
822 /* JITs an eBPF instruction.
823  * Returns:
824  * 0  - successfully JITed an 8-byte eBPF instruction.
825  * >0 - successfully JITed a 16-byte eBPF instruction.
826  * <0 - failed to JIT.
827  */
828 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx,
829 		      bool extra_pass)
830 {
831 	const u8 code = insn->code;
832 	const u8 dst = bpf2a64[insn->dst_reg];
833 	const u8 src = bpf2a64[insn->src_reg];
834 	const u8 tmp = bpf2a64[TMP_REG_1];
835 	const u8 tmp2 = bpf2a64[TMP_REG_2];
836 	const u8 fp = bpf2a64[BPF_REG_FP];
837 	const u8 fpb = bpf2a64[FP_BOTTOM];
838 	const s16 off = insn->off;
839 	const s32 imm = insn->imm;
840 	const int i = insn - ctx->prog->insnsi;
841 	const bool is64 = BPF_CLASS(code) == BPF_ALU64 ||
842 			  BPF_CLASS(code) == BPF_JMP;
843 	u8 jmp_cond;
844 	s32 jmp_offset;
845 	u32 a64_insn;
846 	u8 src_adj;
847 	u8 dst_adj;
848 	int off_adj;
849 	int ret;
850 	bool sign_extend;
851 
852 	switch (code) {
853 	/* dst = src */
854 	case BPF_ALU | BPF_MOV | BPF_X:
855 	case BPF_ALU64 | BPF_MOV | BPF_X:
856 		switch (insn->off) {
857 		case 0:
858 			emit(A64_MOV(is64, dst, src), ctx);
859 			break;
860 		case 8:
861 			emit(A64_SXTB(is64, dst, src), ctx);
862 			break;
863 		case 16:
864 			emit(A64_SXTH(is64, dst, src), ctx);
865 			break;
866 		case 32:
867 			emit(A64_SXTW(is64, dst, src), ctx);
868 			break;
869 		}
870 		break;
871 	/* dst = dst OP src */
872 	case BPF_ALU | BPF_ADD | BPF_X:
873 	case BPF_ALU64 | BPF_ADD | BPF_X:
874 		emit(A64_ADD(is64, dst, dst, src), ctx);
875 		break;
876 	case BPF_ALU | BPF_SUB | BPF_X:
877 	case BPF_ALU64 | BPF_SUB | BPF_X:
878 		emit(A64_SUB(is64, dst, dst, src), ctx);
879 		break;
880 	case BPF_ALU | BPF_AND | BPF_X:
881 	case BPF_ALU64 | BPF_AND | BPF_X:
882 		emit(A64_AND(is64, dst, dst, src), ctx);
883 		break;
884 	case BPF_ALU | BPF_OR | BPF_X:
885 	case BPF_ALU64 | BPF_OR | BPF_X:
886 		emit(A64_ORR(is64, dst, dst, src), ctx);
887 		break;
888 	case BPF_ALU | BPF_XOR | BPF_X:
889 	case BPF_ALU64 | BPF_XOR | BPF_X:
890 		emit(A64_EOR(is64, dst, dst, src), ctx);
891 		break;
892 	case BPF_ALU | BPF_MUL | BPF_X:
893 	case BPF_ALU64 | BPF_MUL | BPF_X:
894 		emit(A64_MUL(is64, dst, dst, src), ctx);
895 		break;
896 	case BPF_ALU | BPF_DIV | BPF_X:
897 	case BPF_ALU64 | BPF_DIV | BPF_X:
898 		if (!off)
899 			emit(A64_UDIV(is64, dst, dst, src), ctx);
900 		else
901 			emit(A64_SDIV(is64, dst, dst, src), ctx);
902 		break;
903 	case BPF_ALU | BPF_MOD | BPF_X:
904 	case BPF_ALU64 | BPF_MOD | BPF_X:
905 		if (!off)
906 			emit(A64_UDIV(is64, tmp, dst, src), ctx);
907 		else
908 			emit(A64_SDIV(is64, tmp, dst, src), ctx);
909 		emit(A64_MSUB(is64, dst, dst, tmp, src), ctx);
910 		break;
911 	case BPF_ALU | BPF_LSH | BPF_X:
912 	case BPF_ALU64 | BPF_LSH | BPF_X:
913 		emit(A64_LSLV(is64, dst, dst, src), ctx);
914 		break;
915 	case BPF_ALU | BPF_RSH | BPF_X:
916 	case BPF_ALU64 | BPF_RSH | BPF_X:
917 		emit(A64_LSRV(is64, dst, dst, src), ctx);
918 		break;
919 	case BPF_ALU | BPF_ARSH | BPF_X:
920 	case BPF_ALU64 | BPF_ARSH | BPF_X:
921 		emit(A64_ASRV(is64, dst, dst, src), ctx);
922 		break;
923 	/* dst = -dst */
924 	case BPF_ALU | BPF_NEG:
925 	case BPF_ALU64 | BPF_NEG:
926 		emit(A64_NEG(is64, dst, dst), ctx);
927 		break;
928 	/* dst = BSWAP##imm(dst) */
929 	case BPF_ALU | BPF_END | BPF_FROM_LE:
930 	case BPF_ALU | BPF_END | BPF_FROM_BE:
931 	case BPF_ALU64 | BPF_END | BPF_FROM_LE:
932 #ifdef CONFIG_CPU_BIG_ENDIAN
933 		if (BPF_CLASS(code) == BPF_ALU && BPF_SRC(code) == BPF_FROM_BE)
934 			goto emit_bswap_uxt;
935 #else /* !CONFIG_CPU_BIG_ENDIAN */
936 		if (BPF_CLASS(code) == BPF_ALU && BPF_SRC(code) == BPF_FROM_LE)
937 			goto emit_bswap_uxt;
938 #endif
939 		switch (imm) {
940 		case 16:
941 			emit(A64_REV16(is64, dst, dst), ctx);
942 			/* zero-extend 16 bits into 64 bits */
943 			emit(A64_UXTH(is64, dst, dst), ctx);
944 			break;
945 		case 32:
946 			emit(A64_REV32(0, dst, dst), ctx);
947 			/* upper 32 bits already cleared */
948 			break;
949 		case 64:
950 			emit(A64_REV64(dst, dst), ctx);
951 			break;
952 		}
953 		break;
954 emit_bswap_uxt:
955 		switch (imm) {
956 		case 16:
957 			/* zero-extend 16 bits into 64 bits */
958 			emit(A64_UXTH(is64, dst, dst), ctx);
959 			break;
960 		case 32:
961 			/* zero-extend 32 bits into 64 bits */
962 			emit(A64_UXTW(is64, dst, dst), ctx);
963 			break;
964 		case 64:
965 			/* nop */
966 			break;
967 		}
968 		break;
969 	/* dst = imm */
970 	case BPF_ALU | BPF_MOV | BPF_K:
971 	case BPF_ALU64 | BPF_MOV | BPF_K:
972 		emit_a64_mov_i(is64, dst, imm, ctx);
973 		break;
974 	/* dst = dst OP imm */
975 	case BPF_ALU | BPF_ADD | BPF_K:
976 	case BPF_ALU64 | BPF_ADD | BPF_K:
977 		if (is_addsub_imm(imm)) {
978 			emit(A64_ADD_I(is64, dst, dst, imm), ctx);
979 		} else if (is_addsub_imm(-imm)) {
980 			emit(A64_SUB_I(is64, dst, dst, -imm), ctx);
981 		} else {
982 			emit_a64_mov_i(is64, tmp, imm, ctx);
983 			emit(A64_ADD(is64, dst, dst, tmp), ctx);
984 		}
985 		break;
986 	case BPF_ALU | BPF_SUB | BPF_K:
987 	case BPF_ALU64 | BPF_SUB | BPF_K:
988 		if (is_addsub_imm(imm)) {
989 			emit(A64_SUB_I(is64, dst, dst, imm), ctx);
990 		} else if (is_addsub_imm(-imm)) {
991 			emit(A64_ADD_I(is64, dst, dst, -imm), ctx);
992 		} else {
993 			emit_a64_mov_i(is64, tmp, imm, ctx);
994 			emit(A64_SUB(is64, dst, dst, tmp), ctx);
995 		}
996 		break;
997 	case BPF_ALU | BPF_AND | BPF_K:
998 	case BPF_ALU64 | BPF_AND | BPF_K:
999 		a64_insn = A64_AND_I(is64, dst, dst, imm);
1000 		if (a64_insn != AARCH64_BREAK_FAULT) {
1001 			emit(a64_insn, ctx);
1002 		} else {
1003 			emit_a64_mov_i(is64, tmp, imm, ctx);
1004 			emit(A64_AND(is64, dst, dst, tmp), ctx);
1005 		}
1006 		break;
1007 	case BPF_ALU | BPF_OR | BPF_K:
1008 	case BPF_ALU64 | BPF_OR | BPF_K:
1009 		a64_insn = A64_ORR_I(is64, dst, dst, imm);
1010 		if (a64_insn != AARCH64_BREAK_FAULT) {
1011 			emit(a64_insn, ctx);
1012 		} else {
1013 			emit_a64_mov_i(is64, tmp, imm, ctx);
1014 			emit(A64_ORR(is64, dst, dst, tmp), ctx);
1015 		}
1016 		break;
1017 	case BPF_ALU | BPF_XOR | BPF_K:
1018 	case BPF_ALU64 | BPF_XOR | BPF_K:
1019 		a64_insn = A64_EOR_I(is64, dst, dst, imm);
1020 		if (a64_insn != AARCH64_BREAK_FAULT) {
1021 			emit(a64_insn, ctx);
1022 		} else {
1023 			emit_a64_mov_i(is64, tmp, imm, ctx);
1024 			emit(A64_EOR(is64, dst, dst, tmp), ctx);
1025 		}
1026 		break;
1027 	case BPF_ALU | BPF_MUL | BPF_K:
1028 	case BPF_ALU64 | BPF_MUL | BPF_K:
1029 		emit_a64_mov_i(is64, tmp, imm, ctx);
1030 		emit(A64_MUL(is64, dst, dst, tmp), ctx);
1031 		break;
1032 	case BPF_ALU | BPF_DIV | BPF_K:
1033 	case BPF_ALU64 | BPF_DIV | BPF_K:
1034 		emit_a64_mov_i(is64, tmp, imm, ctx);
1035 		if (!off)
1036 			emit(A64_UDIV(is64, dst, dst, tmp), ctx);
1037 		else
1038 			emit(A64_SDIV(is64, dst, dst, tmp), ctx);
1039 		break;
1040 	case BPF_ALU | BPF_MOD | BPF_K:
1041 	case BPF_ALU64 | BPF_MOD | BPF_K:
1042 		emit_a64_mov_i(is64, tmp2, imm, ctx);
1043 		if (!off)
1044 			emit(A64_UDIV(is64, tmp, dst, tmp2), ctx);
1045 		else
1046 			emit(A64_SDIV(is64, tmp, dst, tmp2), ctx);
1047 		emit(A64_MSUB(is64, dst, dst, tmp, tmp2), ctx);
1048 		break;
1049 	case BPF_ALU | BPF_LSH | BPF_K:
1050 	case BPF_ALU64 | BPF_LSH | BPF_K:
1051 		emit(A64_LSL(is64, dst, dst, imm), ctx);
1052 		break;
1053 	case BPF_ALU | BPF_RSH | BPF_K:
1054 	case BPF_ALU64 | BPF_RSH | BPF_K:
1055 		emit(A64_LSR(is64, dst, dst, imm), ctx);
1056 		break;
1057 	case BPF_ALU | BPF_ARSH | BPF_K:
1058 	case BPF_ALU64 | BPF_ARSH | BPF_K:
1059 		emit(A64_ASR(is64, dst, dst, imm), ctx);
1060 		break;
1061 
1062 	/* JUMP off */
1063 	case BPF_JMP | BPF_JA:
1064 	case BPF_JMP32 | BPF_JA:
1065 		if (BPF_CLASS(code) == BPF_JMP)
1066 			jmp_offset = bpf2a64_offset(i, off, ctx);
1067 		else
1068 			jmp_offset = bpf2a64_offset(i, imm, ctx);
1069 		check_imm26(jmp_offset);
1070 		emit(A64_B(jmp_offset), ctx);
1071 		break;
1072 	/* IF (dst COND src) JUMP off */
1073 	case BPF_JMP | BPF_JEQ | BPF_X:
1074 	case BPF_JMP | BPF_JGT | BPF_X:
1075 	case BPF_JMP | BPF_JLT | BPF_X:
1076 	case BPF_JMP | BPF_JGE | BPF_X:
1077 	case BPF_JMP | BPF_JLE | BPF_X:
1078 	case BPF_JMP | BPF_JNE | BPF_X:
1079 	case BPF_JMP | BPF_JSGT | BPF_X:
1080 	case BPF_JMP | BPF_JSLT | BPF_X:
1081 	case BPF_JMP | BPF_JSGE | BPF_X:
1082 	case BPF_JMP | BPF_JSLE | BPF_X:
1083 	case BPF_JMP32 | BPF_JEQ | BPF_X:
1084 	case BPF_JMP32 | BPF_JGT | BPF_X:
1085 	case BPF_JMP32 | BPF_JLT | BPF_X:
1086 	case BPF_JMP32 | BPF_JGE | BPF_X:
1087 	case BPF_JMP32 | BPF_JLE | BPF_X:
1088 	case BPF_JMP32 | BPF_JNE | BPF_X:
1089 	case BPF_JMP32 | BPF_JSGT | BPF_X:
1090 	case BPF_JMP32 | BPF_JSLT | BPF_X:
1091 	case BPF_JMP32 | BPF_JSGE | BPF_X:
1092 	case BPF_JMP32 | BPF_JSLE | BPF_X:
1093 		emit(A64_CMP(is64, dst, src), ctx);
1094 emit_cond_jmp:
1095 		jmp_offset = bpf2a64_offset(i, off, ctx);
1096 		check_imm19(jmp_offset);
1097 		switch (BPF_OP(code)) {
1098 		case BPF_JEQ:
1099 			jmp_cond = A64_COND_EQ;
1100 			break;
1101 		case BPF_JGT:
1102 			jmp_cond = A64_COND_HI;
1103 			break;
1104 		case BPF_JLT:
1105 			jmp_cond = A64_COND_CC;
1106 			break;
1107 		case BPF_JGE:
1108 			jmp_cond = A64_COND_CS;
1109 			break;
1110 		case BPF_JLE:
1111 			jmp_cond = A64_COND_LS;
1112 			break;
1113 		case BPF_JSET:
1114 		case BPF_JNE:
1115 			jmp_cond = A64_COND_NE;
1116 			break;
1117 		case BPF_JSGT:
1118 			jmp_cond = A64_COND_GT;
1119 			break;
1120 		case BPF_JSLT:
1121 			jmp_cond = A64_COND_LT;
1122 			break;
1123 		case BPF_JSGE:
1124 			jmp_cond = A64_COND_GE;
1125 			break;
1126 		case BPF_JSLE:
1127 			jmp_cond = A64_COND_LE;
1128 			break;
1129 		default:
1130 			return -EFAULT;
1131 		}
1132 		emit(A64_B_(jmp_cond, jmp_offset), ctx);
1133 		break;
1134 	case BPF_JMP | BPF_JSET | BPF_X:
1135 	case BPF_JMP32 | BPF_JSET | BPF_X:
1136 		emit(A64_TST(is64, dst, src), ctx);
1137 		goto emit_cond_jmp;
1138 	/* IF (dst COND imm) JUMP off */
1139 	case BPF_JMP | BPF_JEQ | BPF_K:
1140 	case BPF_JMP | BPF_JGT | BPF_K:
1141 	case BPF_JMP | BPF_JLT | BPF_K:
1142 	case BPF_JMP | BPF_JGE | BPF_K:
1143 	case BPF_JMP | BPF_JLE | BPF_K:
1144 	case BPF_JMP | BPF_JNE | BPF_K:
1145 	case BPF_JMP | BPF_JSGT | BPF_K:
1146 	case BPF_JMP | BPF_JSLT | BPF_K:
1147 	case BPF_JMP | BPF_JSGE | BPF_K:
1148 	case BPF_JMP | BPF_JSLE | BPF_K:
1149 	case BPF_JMP32 | BPF_JEQ | BPF_K:
1150 	case BPF_JMP32 | BPF_JGT | BPF_K:
1151 	case BPF_JMP32 | BPF_JLT | BPF_K:
1152 	case BPF_JMP32 | BPF_JGE | BPF_K:
1153 	case BPF_JMP32 | BPF_JLE | BPF_K:
1154 	case BPF_JMP32 | BPF_JNE | BPF_K:
1155 	case BPF_JMP32 | BPF_JSGT | BPF_K:
1156 	case BPF_JMP32 | BPF_JSLT | BPF_K:
1157 	case BPF_JMP32 | BPF_JSGE | BPF_K:
1158 	case BPF_JMP32 | BPF_JSLE | BPF_K:
1159 		if (is_addsub_imm(imm)) {
1160 			emit(A64_CMP_I(is64, dst, imm), ctx);
1161 		} else if (is_addsub_imm(-imm)) {
1162 			emit(A64_CMN_I(is64, dst, -imm), ctx);
1163 		} else {
1164 			emit_a64_mov_i(is64, tmp, imm, ctx);
1165 			emit(A64_CMP(is64, dst, tmp), ctx);
1166 		}
1167 		goto emit_cond_jmp;
1168 	case BPF_JMP | BPF_JSET | BPF_K:
1169 	case BPF_JMP32 | BPF_JSET | BPF_K:
1170 		a64_insn = A64_TST_I(is64, dst, imm);
1171 		if (a64_insn != AARCH64_BREAK_FAULT) {
1172 			emit(a64_insn, ctx);
1173 		} else {
1174 			emit_a64_mov_i(is64, tmp, imm, ctx);
1175 			emit(A64_TST(is64, dst, tmp), ctx);
1176 		}
1177 		goto emit_cond_jmp;
1178 	/* function call */
1179 	case BPF_JMP | BPF_CALL:
1180 	{
1181 		const u8 r0 = bpf2a64[BPF_REG_0];
1182 		bool func_addr_fixed;
1183 		u64 func_addr;
1184 
1185 		ret = bpf_jit_get_func_addr(ctx->prog, insn, extra_pass,
1186 					    &func_addr, &func_addr_fixed);
1187 		if (ret < 0)
1188 			return ret;
1189 		emit_call(func_addr, ctx);
1190 		emit(A64_MOV(1, r0, A64_R(0)), ctx);
1191 		break;
1192 	}
1193 	/* tail call */
1194 	case BPF_JMP | BPF_TAIL_CALL:
1195 		if (emit_bpf_tail_call(ctx))
1196 			return -EFAULT;
1197 		break;
1198 	/* function return */
1199 	case BPF_JMP | BPF_EXIT:
1200 		/* Optimization: when last instruction is EXIT,
1201 		   simply fallthrough to epilogue. */
1202 		if (i == ctx->prog->len - 1)
1203 			break;
1204 		jmp_offset = epilogue_offset(ctx);
1205 		check_imm26(jmp_offset);
1206 		emit(A64_B(jmp_offset), ctx);
1207 		break;
1208 
1209 	/* dst = imm64 */
1210 	case BPF_LD | BPF_IMM | BPF_DW:
1211 	{
1212 		const struct bpf_insn insn1 = insn[1];
1213 		u64 imm64;
1214 
1215 		imm64 = (u64)insn1.imm << 32 | (u32)imm;
1216 		if (bpf_pseudo_func(insn))
1217 			emit_addr_mov_i64(dst, imm64, ctx);
1218 		else
1219 			emit_a64_mov_i64(dst, imm64, ctx);
1220 
1221 		return 1;
1222 	}
1223 
1224 	/* LDX: dst = (u64)*(unsigned size *)(src + off) */
1225 	case BPF_LDX | BPF_MEM | BPF_W:
1226 	case BPF_LDX | BPF_MEM | BPF_H:
1227 	case BPF_LDX | BPF_MEM | BPF_B:
1228 	case BPF_LDX | BPF_MEM | BPF_DW:
1229 	case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
1230 	case BPF_LDX | BPF_PROBE_MEM | BPF_W:
1231 	case BPF_LDX | BPF_PROBE_MEM | BPF_H:
1232 	case BPF_LDX | BPF_PROBE_MEM | BPF_B:
1233 	/* LDXS: dst_reg = (s64)*(signed size *)(src_reg + off) */
1234 	case BPF_LDX | BPF_MEMSX | BPF_B:
1235 	case BPF_LDX | BPF_MEMSX | BPF_H:
1236 	case BPF_LDX | BPF_MEMSX | BPF_W:
1237 	case BPF_LDX | BPF_PROBE_MEMSX | BPF_B:
1238 	case BPF_LDX | BPF_PROBE_MEMSX | BPF_H:
1239 	case BPF_LDX | BPF_PROBE_MEMSX | BPF_W:
1240 		if (ctx->fpb_offset > 0 && src == fp) {
1241 			src_adj = fpb;
1242 			off_adj = off + ctx->fpb_offset;
1243 		} else {
1244 			src_adj = src;
1245 			off_adj = off;
1246 		}
1247 		sign_extend = (BPF_MODE(insn->code) == BPF_MEMSX ||
1248 				BPF_MODE(insn->code) == BPF_PROBE_MEMSX);
1249 		switch (BPF_SIZE(code)) {
1250 		case BPF_W:
1251 			if (is_lsi_offset(off_adj, 2)) {
1252 				if (sign_extend)
1253 					emit(A64_LDRSWI(dst, src_adj, off_adj), ctx);
1254 				else
1255 					emit(A64_LDR32I(dst, src_adj, off_adj), ctx);
1256 			} else {
1257 				emit_a64_mov_i(1, tmp, off, ctx);
1258 				if (sign_extend)
1259 					emit(A64_LDRSW(dst, src, tmp), ctx);
1260 				else
1261 					emit(A64_LDR32(dst, src, tmp), ctx);
1262 			}
1263 			break;
1264 		case BPF_H:
1265 			if (is_lsi_offset(off_adj, 1)) {
1266 				if (sign_extend)
1267 					emit(A64_LDRSHI(dst, src_adj, off_adj), ctx);
1268 				else
1269 					emit(A64_LDRHI(dst, src_adj, off_adj), ctx);
1270 			} else {
1271 				emit_a64_mov_i(1, tmp, off, ctx);
1272 				if (sign_extend)
1273 					emit(A64_LDRSH(dst, src, tmp), ctx);
1274 				else
1275 					emit(A64_LDRH(dst, src, tmp), ctx);
1276 			}
1277 			break;
1278 		case BPF_B:
1279 			if (is_lsi_offset(off_adj, 0)) {
1280 				if (sign_extend)
1281 					emit(A64_LDRSBI(dst, src_adj, off_adj), ctx);
1282 				else
1283 					emit(A64_LDRBI(dst, src_adj, off_adj), ctx);
1284 			} else {
1285 				emit_a64_mov_i(1, tmp, off, ctx);
1286 				if (sign_extend)
1287 					emit(A64_LDRSB(dst, src, tmp), ctx);
1288 				else
1289 					emit(A64_LDRB(dst, src, tmp), ctx);
1290 			}
1291 			break;
1292 		case BPF_DW:
1293 			if (is_lsi_offset(off_adj, 3)) {
1294 				emit(A64_LDR64I(dst, src_adj, off_adj), ctx);
1295 			} else {
1296 				emit_a64_mov_i(1, tmp, off, ctx);
1297 				emit(A64_LDR64(dst, src, tmp), ctx);
1298 			}
1299 			break;
1300 		}
1301 
1302 		ret = add_exception_handler(insn, ctx, dst);
1303 		if (ret)
1304 			return ret;
1305 		break;
1306 
1307 	/* speculation barrier */
1308 	case BPF_ST | BPF_NOSPEC:
1309 		/*
1310 		 * Nothing required here.
1311 		 *
1312 		 * In case of arm64, we rely on the firmware mitigation of
1313 		 * Speculative Store Bypass as controlled via the ssbd kernel
1314 		 * parameter. Whenever the mitigation is enabled, it works
1315 		 * for all of the kernel code with no need to provide any
1316 		 * additional instructions.
1317 		 */
1318 		break;
1319 
1320 	/* ST: *(size *)(dst + off) = imm */
1321 	case BPF_ST | BPF_MEM | BPF_W:
1322 	case BPF_ST | BPF_MEM | BPF_H:
1323 	case BPF_ST | BPF_MEM | BPF_B:
1324 	case BPF_ST | BPF_MEM | BPF_DW:
1325 		if (ctx->fpb_offset > 0 && dst == fp) {
1326 			dst_adj = fpb;
1327 			off_adj = off + ctx->fpb_offset;
1328 		} else {
1329 			dst_adj = dst;
1330 			off_adj = off;
1331 		}
1332 		/* Load imm to a register then store it */
1333 		emit_a64_mov_i(1, tmp, imm, ctx);
1334 		switch (BPF_SIZE(code)) {
1335 		case BPF_W:
1336 			if (is_lsi_offset(off_adj, 2)) {
1337 				emit(A64_STR32I(tmp, dst_adj, off_adj), ctx);
1338 			} else {
1339 				emit_a64_mov_i(1, tmp2, off, ctx);
1340 				emit(A64_STR32(tmp, dst, tmp2), ctx);
1341 			}
1342 			break;
1343 		case BPF_H:
1344 			if (is_lsi_offset(off_adj, 1)) {
1345 				emit(A64_STRHI(tmp, dst_adj, off_adj), ctx);
1346 			} else {
1347 				emit_a64_mov_i(1, tmp2, off, ctx);
1348 				emit(A64_STRH(tmp, dst, tmp2), ctx);
1349 			}
1350 			break;
1351 		case BPF_B:
1352 			if (is_lsi_offset(off_adj, 0)) {
1353 				emit(A64_STRBI(tmp, dst_adj, off_adj), ctx);
1354 			} else {
1355 				emit_a64_mov_i(1, tmp2, off, ctx);
1356 				emit(A64_STRB(tmp, dst, tmp2), ctx);
1357 			}
1358 			break;
1359 		case BPF_DW:
1360 			if (is_lsi_offset(off_adj, 3)) {
1361 				emit(A64_STR64I(tmp, dst_adj, off_adj), ctx);
1362 			} else {
1363 				emit_a64_mov_i(1, tmp2, off, ctx);
1364 				emit(A64_STR64(tmp, dst, tmp2), ctx);
1365 			}
1366 			break;
1367 		}
1368 		break;
1369 
1370 	/* STX: *(size *)(dst + off) = src */
1371 	case BPF_STX | BPF_MEM | BPF_W:
1372 	case BPF_STX | BPF_MEM | BPF_H:
1373 	case BPF_STX | BPF_MEM | BPF_B:
1374 	case BPF_STX | BPF_MEM | BPF_DW:
1375 		if (ctx->fpb_offset > 0 && dst == fp) {
1376 			dst_adj = fpb;
1377 			off_adj = off + ctx->fpb_offset;
1378 		} else {
1379 			dst_adj = dst;
1380 			off_adj = off;
1381 		}
1382 		switch (BPF_SIZE(code)) {
1383 		case BPF_W:
1384 			if (is_lsi_offset(off_adj, 2)) {
1385 				emit(A64_STR32I(src, dst_adj, off_adj), ctx);
1386 			} else {
1387 				emit_a64_mov_i(1, tmp, off, ctx);
1388 				emit(A64_STR32(src, dst, tmp), ctx);
1389 			}
1390 			break;
1391 		case BPF_H:
1392 			if (is_lsi_offset(off_adj, 1)) {
1393 				emit(A64_STRHI(src, dst_adj, off_adj), ctx);
1394 			} else {
1395 				emit_a64_mov_i(1, tmp, off, ctx);
1396 				emit(A64_STRH(src, dst, tmp), ctx);
1397 			}
1398 			break;
1399 		case BPF_B:
1400 			if (is_lsi_offset(off_adj, 0)) {
1401 				emit(A64_STRBI(src, dst_adj, off_adj), ctx);
1402 			} else {
1403 				emit_a64_mov_i(1, tmp, off, ctx);
1404 				emit(A64_STRB(src, dst, tmp), ctx);
1405 			}
1406 			break;
1407 		case BPF_DW:
1408 			if (is_lsi_offset(off_adj, 3)) {
1409 				emit(A64_STR64I(src, dst_adj, off_adj), ctx);
1410 			} else {
1411 				emit_a64_mov_i(1, tmp, off, ctx);
1412 				emit(A64_STR64(src, dst, tmp), ctx);
1413 			}
1414 			break;
1415 		}
1416 		break;
1417 
1418 	case BPF_STX | BPF_ATOMIC | BPF_W:
1419 	case BPF_STX | BPF_ATOMIC | BPF_DW:
1420 		if (cpus_have_cap(ARM64_HAS_LSE_ATOMICS))
1421 			ret = emit_lse_atomic(insn, ctx);
1422 		else
1423 			ret = emit_ll_sc_atomic(insn, ctx);
1424 		if (ret)
1425 			return ret;
1426 		break;
1427 
1428 	default:
1429 		pr_err_once("unknown opcode %02x\n", code);
1430 		return -EINVAL;
1431 	}
1432 
1433 	return 0;
1434 }
1435 
1436 /*
1437  * Return 0 if FP may change at runtime, otherwise find the minimum negative
1438  * offset to FP, converts it to positive number, and align down to 8 bytes.
1439  */
1440 static int find_fpb_offset(struct bpf_prog *prog)
1441 {
1442 	int i;
1443 	int offset = 0;
1444 
1445 	for (i = 0; i < prog->len; i++) {
1446 		const struct bpf_insn *insn = &prog->insnsi[i];
1447 		const u8 class = BPF_CLASS(insn->code);
1448 		const u8 mode = BPF_MODE(insn->code);
1449 		const u8 src = insn->src_reg;
1450 		const u8 dst = insn->dst_reg;
1451 		const s32 imm = insn->imm;
1452 		const s16 off = insn->off;
1453 
1454 		switch (class) {
1455 		case BPF_STX:
1456 		case BPF_ST:
1457 			/* fp holds atomic operation result */
1458 			if (class == BPF_STX && mode == BPF_ATOMIC &&
1459 			    ((imm == BPF_XCHG ||
1460 			      imm == (BPF_FETCH | BPF_ADD) ||
1461 			      imm == (BPF_FETCH | BPF_AND) ||
1462 			      imm == (BPF_FETCH | BPF_XOR) ||
1463 			      imm == (BPF_FETCH | BPF_OR)) &&
1464 			     src == BPF_REG_FP))
1465 				return 0;
1466 
1467 			if (mode == BPF_MEM && dst == BPF_REG_FP &&
1468 			    off < offset)
1469 				offset = insn->off;
1470 			break;
1471 
1472 		case BPF_JMP32:
1473 		case BPF_JMP:
1474 			break;
1475 
1476 		case BPF_LDX:
1477 		case BPF_LD:
1478 			/* fp holds load result */
1479 			if (dst == BPF_REG_FP)
1480 				return 0;
1481 
1482 			if (class == BPF_LDX && mode == BPF_MEM &&
1483 			    src == BPF_REG_FP && off < offset)
1484 				offset = off;
1485 			break;
1486 
1487 		case BPF_ALU:
1488 		case BPF_ALU64:
1489 		default:
1490 			/* fp holds ALU result */
1491 			if (dst == BPF_REG_FP)
1492 				return 0;
1493 		}
1494 	}
1495 
1496 	if (offset < 0) {
1497 		/*
1498 		 * safely be converted to a positive 'int', since insn->off
1499 		 * is 's16'
1500 		 */
1501 		offset = -offset;
1502 		/* align down to 8 bytes */
1503 		offset = ALIGN_DOWN(offset, 8);
1504 	}
1505 
1506 	return offset;
1507 }
1508 
1509 static int build_body(struct jit_ctx *ctx, bool extra_pass)
1510 {
1511 	const struct bpf_prog *prog = ctx->prog;
1512 	int i;
1513 
1514 	/*
1515 	 * - offset[0] offset of the end of prologue,
1516 	 *   start of the 1st instruction.
1517 	 * - offset[1] - offset of the end of 1st instruction,
1518 	 *   start of the 2nd instruction
1519 	 * [....]
1520 	 * - offset[3] - offset of the end of 3rd instruction,
1521 	 *   start of 4th instruction
1522 	 */
1523 	for (i = 0; i < prog->len; i++) {
1524 		const struct bpf_insn *insn = &prog->insnsi[i];
1525 		int ret;
1526 
1527 		if (ctx->image == NULL)
1528 			ctx->offset[i] = ctx->idx;
1529 		ret = build_insn(insn, ctx, extra_pass);
1530 		if (ret > 0) {
1531 			i++;
1532 			if (ctx->image == NULL)
1533 				ctx->offset[i] = ctx->idx;
1534 			continue;
1535 		}
1536 		if (ret)
1537 			return ret;
1538 	}
1539 	/*
1540 	 * offset is allocated with prog->len + 1 so fill in
1541 	 * the last element with the offset after the last
1542 	 * instruction (end of program)
1543 	 */
1544 	if (ctx->image == NULL)
1545 		ctx->offset[i] = ctx->idx;
1546 
1547 	return 0;
1548 }
1549 
1550 static int validate_code(struct jit_ctx *ctx)
1551 {
1552 	int i;
1553 
1554 	for (i = 0; i < ctx->idx; i++) {
1555 		u32 a64_insn = le32_to_cpu(ctx->image[i]);
1556 
1557 		if (a64_insn == AARCH64_BREAK_FAULT)
1558 			return -1;
1559 	}
1560 	return 0;
1561 }
1562 
1563 static int validate_ctx(struct jit_ctx *ctx)
1564 {
1565 	if (validate_code(ctx))
1566 		return -1;
1567 
1568 	if (WARN_ON_ONCE(ctx->exentry_idx != ctx->prog->aux->num_exentries))
1569 		return -1;
1570 
1571 	return 0;
1572 }
1573 
1574 static inline void bpf_flush_icache(void *start, void *end)
1575 {
1576 	flush_icache_range((unsigned long)start, (unsigned long)end);
1577 }
1578 
1579 struct arm64_jit_data {
1580 	struct bpf_binary_header *header;
1581 	u8 *ro_image;
1582 	struct bpf_binary_header *ro_header;
1583 	struct jit_ctx ctx;
1584 };
1585 
1586 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1587 {
1588 	int image_size, prog_size, extable_size, extable_align, extable_offset;
1589 	struct bpf_prog *tmp, *orig_prog = prog;
1590 	struct bpf_binary_header *header;
1591 	struct bpf_binary_header *ro_header;
1592 	struct arm64_jit_data *jit_data;
1593 	bool was_classic = bpf_prog_was_classic(prog);
1594 	bool tmp_blinded = false;
1595 	bool extra_pass = false;
1596 	struct jit_ctx ctx;
1597 	u8 *image_ptr;
1598 	u8 *ro_image_ptr;
1599 
1600 	if (!prog->jit_requested)
1601 		return orig_prog;
1602 
1603 	tmp = bpf_jit_blind_constants(prog);
1604 	/* If blinding was requested and we failed during blinding,
1605 	 * we must fall back to the interpreter.
1606 	 */
1607 	if (IS_ERR(tmp))
1608 		return orig_prog;
1609 	if (tmp != prog) {
1610 		tmp_blinded = true;
1611 		prog = tmp;
1612 	}
1613 
1614 	jit_data = prog->aux->jit_data;
1615 	if (!jit_data) {
1616 		jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1617 		if (!jit_data) {
1618 			prog = orig_prog;
1619 			goto out;
1620 		}
1621 		prog->aux->jit_data = jit_data;
1622 	}
1623 	if (jit_data->ctx.offset) {
1624 		ctx = jit_data->ctx;
1625 		ro_image_ptr = jit_data->ro_image;
1626 		ro_header = jit_data->ro_header;
1627 		header = jit_data->header;
1628 		image_ptr = (void *)header + ((void *)ro_image_ptr
1629 						 - (void *)ro_header);
1630 		extra_pass = true;
1631 		prog_size = sizeof(u32) * ctx.idx;
1632 		goto skip_init_ctx;
1633 	}
1634 	memset(&ctx, 0, sizeof(ctx));
1635 	ctx.prog = prog;
1636 
1637 	ctx.offset = kvcalloc(prog->len + 1, sizeof(int), GFP_KERNEL);
1638 	if (ctx.offset == NULL) {
1639 		prog = orig_prog;
1640 		goto out_off;
1641 	}
1642 
1643 	ctx.fpb_offset = find_fpb_offset(prog);
1644 
1645 	/*
1646 	 * 1. Initial fake pass to compute ctx->idx and ctx->offset.
1647 	 *
1648 	 * BPF line info needs ctx->offset[i] to be the offset of
1649 	 * instruction[i] in jited image, so build prologue first.
1650 	 */
1651 	if (build_prologue(&ctx, was_classic, prog->aux->exception_cb)) {
1652 		prog = orig_prog;
1653 		goto out_off;
1654 	}
1655 
1656 	if (build_body(&ctx, extra_pass)) {
1657 		prog = orig_prog;
1658 		goto out_off;
1659 	}
1660 
1661 	ctx.epilogue_offset = ctx.idx;
1662 	build_epilogue(&ctx, prog->aux->exception_cb);
1663 	build_plt(&ctx);
1664 
1665 	extable_align = __alignof__(struct exception_table_entry);
1666 	extable_size = prog->aux->num_exentries *
1667 		sizeof(struct exception_table_entry);
1668 
1669 	/* Now we know the actual image size. */
1670 	prog_size = sizeof(u32) * ctx.idx;
1671 	/* also allocate space for plt target */
1672 	extable_offset = round_up(prog_size + PLT_TARGET_SIZE, extable_align);
1673 	image_size = extable_offset + extable_size;
1674 	ro_header = bpf_jit_binary_pack_alloc(image_size, &ro_image_ptr,
1675 					      sizeof(u32), &header, &image_ptr,
1676 					      jit_fill_hole);
1677 	if (!ro_header) {
1678 		prog = orig_prog;
1679 		goto out_off;
1680 	}
1681 
1682 	/* 2. Now, the actual pass. */
1683 
1684 	/*
1685 	 * Use the image(RW) for writing the JITed instructions. But also save
1686 	 * the ro_image(RX) for calculating the offsets in the image. The RW
1687 	 * image will be later copied to the RX image from where the program
1688 	 * will run. The bpf_jit_binary_pack_finalize() will do this copy in the
1689 	 * final step.
1690 	 */
1691 	ctx.image = (__le32 *)image_ptr;
1692 	ctx.ro_image = (__le32 *)ro_image_ptr;
1693 	if (extable_size)
1694 		prog->aux->extable = (void *)ro_image_ptr + extable_offset;
1695 skip_init_ctx:
1696 	ctx.idx = 0;
1697 	ctx.exentry_idx = 0;
1698 
1699 	build_prologue(&ctx, was_classic, prog->aux->exception_cb);
1700 
1701 	if (build_body(&ctx, extra_pass)) {
1702 		prog = orig_prog;
1703 		goto out_free_hdr;
1704 	}
1705 
1706 	build_epilogue(&ctx, prog->aux->exception_cb);
1707 	build_plt(&ctx);
1708 
1709 	/* 3. Extra pass to validate JITed code. */
1710 	if (validate_ctx(&ctx)) {
1711 		prog = orig_prog;
1712 		goto out_free_hdr;
1713 	}
1714 
1715 	/* And we're done. */
1716 	if (bpf_jit_enable > 1)
1717 		bpf_jit_dump(prog->len, prog_size, 2, ctx.image);
1718 
1719 	if (!prog->is_func || extra_pass) {
1720 		if (extra_pass && ctx.idx != jit_data->ctx.idx) {
1721 			pr_err_once("multi-func JIT bug %d != %d\n",
1722 				    ctx.idx, jit_data->ctx.idx);
1723 			prog->bpf_func = NULL;
1724 			prog->jited = 0;
1725 			prog->jited_len = 0;
1726 			goto out_free_hdr;
1727 		}
1728 		if (WARN_ON(bpf_jit_binary_pack_finalize(prog, ro_header,
1729 							 header))) {
1730 			/* ro_header has been freed */
1731 			ro_header = NULL;
1732 			prog = orig_prog;
1733 			goto out_off;
1734 		}
1735 		/*
1736 		 * The instructions have now been copied to the ROX region from
1737 		 * where they will execute. Now the data cache has to be cleaned to
1738 		 * the PoU and the I-cache has to be invalidated for the VAs.
1739 		 */
1740 		bpf_flush_icache(ro_header, ctx.ro_image + ctx.idx);
1741 	} else {
1742 		jit_data->ctx = ctx;
1743 		jit_data->ro_image = ro_image_ptr;
1744 		jit_data->header = header;
1745 		jit_data->ro_header = ro_header;
1746 	}
1747 
1748 	prog->bpf_func = (void *)ctx.ro_image;
1749 	prog->jited = 1;
1750 	prog->jited_len = prog_size;
1751 
1752 	if (!prog->is_func || extra_pass) {
1753 		int i;
1754 
1755 		/* offset[prog->len] is the size of program */
1756 		for (i = 0; i <= prog->len; i++)
1757 			ctx.offset[i] *= AARCH64_INSN_SIZE;
1758 		bpf_prog_fill_jited_linfo(prog, ctx.offset + 1);
1759 out_off:
1760 		kvfree(ctx.offset);
1761 		kfree(jit_data);
1762 		prog->aux->jit_data = NULL;
1763 	}
1764 out:
1765 	if (tmp_blinded)
1766 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
1767 					   tmp : orig_prog);
1768 	return prog;
1769 
1770 out_free_hdr:
1771 	if (header) {
1772 		bpf_arch_text_copy(&ro_header->size, &header->size,
1773 				   sizeof(header->size));
1774 		bpf_jit_binary_pack_free(ro_header, header);
1775 	}
1776 	goto out_off;
1777 }
1778 
1779 bool bpf_jit_supports_kfunc_call(void)
1780 {
1781 	return true;
1782 }
1783 
1784 void *bpf_arch_text_copy(void *dst, void *src, size_t len)
1785 {
1786 	if (!aarch64_insn_copy(dst, src, len))
1787 		return ERR_PTR(-EINVAL);
1788 	return dst;
1789 }
1790 
1791 u64 bpf_jit_alloc_exec_limit(void)
1792 {
1793 	return VMALLOC_END - VMALLOC_START;
1794 }
1795 
1796 void *bpf_jit_alloc_exec(unsigned long size)
1797 {
1798 	/* Memory is intended to be executable, reset the pointer tag. */
1799 	return kasan_reset_tag(vmalloc(size));
1800 }
1801 
1802 void bpf_jit_free_exec(void *addr)
1803 {
1804 	return vfree(addr);
1805 }
1806 
1807 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
1808 bool bpf_jit_supports_subprog_tailcalls(void)
1809 {
1810 	return true;
1811 }
1812 
1813 static void invoke_bpf_prog(struct jit_ctx *ctx, struct bpf_tramp_link *l,
1814 			    int args_off, int retval_off, int run_ctx_off,
1815 			    bool save_ret)
1816 {
1817 	__le32 *branch;
1818 	u64 enter_prog;
1819 	u64 exit_prog;
1820 	struct bpf_prog *p = l->link.prog;
1821 	int cookie_off = offsetof(struct bpf_tramp_run_ctx, bpf_cookie);
1822 
1823 	enter_prog = (u64)bpf_trampoline_enter(p);
1824 	exit_prog = (u64)bpf_trampoline_exit(p);
1825 
1826 	if (l->cookie == 0) {
1827 		/* if cookie is zero, one instruction is enough to store it */
1828 		emit(A64_STR64I(A64_ZR, A64_SP, run_ctx_off + cookie_off), ctx);
1829 	} else {
1830 		emit_a64_mov_i64(A64_R(10), l->cookie, ctx);
1831 		emit(A64_STR64I(A64_R(10), A64_SP, run_ctx_off + cookie_off),
1832 		     ctx);
1833 	}
1834 
1835 	/* save p to callee saved register x19 to avoid loading p with mov_i64
1836 	 * each time.
1837 	 */
1838 	emit_addr_mov_i64(A64_R(19), (const u64)p, ctx);
1839 
1840 	/* arg1: prog */
1841 	emit(A64_MOV(1, A64_R(0), A64_R(19)), ctx);
1842 	/* arg2: &run_ctx */
1843 	emit(A64_ADD_I(1, A64_R(1), A64_SP, run_ctx_off), ctx);
1844 
1845 	emit_call(enter_prog, ctx);
1846 
1847 	/* if (__bpf_prog_enter(prog) == 0)
1848 	 *         goto skip_exec_of_prog;
1849 	 */
1850 	branch = ctx->image + ctx->idx;
1851 	emit(A64_NOP, ctx);
1852 
1853 	/* save return value to callee saved register x20 */
1854 	emit(A64_MOV(1, A64_R(20), A64_R(0)), ctx);
1855 
1856 	emit(A64_ADD_I(1, A64_R(0), A64_SP, args_off), ctx);
1857 	if (!p->jited)
1858 		emit_addr_mov_i64(A64_R(1), (const u64)p->insnsi, ctx);
1859 
1860 	emit_call((const u64)p->bpf_func, ctx);
1861 
1862 	if (save_ret)
1863 		emit(A64_STR64I(A64_R(0), A64_SP, retval_off), ctx);
1864 
1865 	if (ctx->image) {
1866 		int offset = &ctx->image[ctx->idx] - branch;
1867 		*branch = cpu_to_le32(A64_CBZ(1, A64_R(0), offset));
1868 	}
1869 
1870 	/* arg1: prog */
1871 	emit(A64_MOV(1, A64_R(0), A64_R(19)), ctx);
1872 	/* arg2: start time */
1873 	emit(A64_MOV(1, A64_R(1), A64_R(20)), ctx);
1874 	/* arg3: &run_ctx */
1875 	emit(A64_ADD_I(1, A64_R(2), A64_SP, run_ctx_off), ctx);
1876 
1877 	emit_call(exit_prog, ctx);
1878 }
1879 
1880 static void invoke_bpf_mod_ret(struct jit_ctx *ctx, struct bpf_tramp_links *tl,
1881 			       int args_off, int retval_off, int run_ctx_off,
1882 			       __le32 **branches)
1883 {
1884 	int i;
1885 
1886 	/* The first fmod_ret program will receive a garbage return value.
1887 	 * Set this to 0 to avoid confusing the program.
1888 	 */
1889 	emit(A64_STR64I(A64_ZR, A64_SP, retval_off), ctx);
1890 	for (i = 0; i < tl->nr_links; i++) {
1891 		invoke_bpf_prog(ctx, tl->links[i], args_off, retval_off,
1892 				run_ctx_off, true);
1893 		/* if (*(u64 *)(sp + retval_off) !=  0)
1894 		 *	goto do_fexit;
1895 		 */
1896 		emit(A64_LDR64I(A64_R(10), A64_SP, retval_off), ctx);
1897 		/* Save the location of branch, and generate a nop.
1898 		 * This nop will be replaced with a cbnz later.
1899 		 */
1900 		branches[i] = ctx->image + ctx->idx;
1901 		emit(A64_NOP, ctx);
1902 	}
1903 }
1904 
1905 static void save_args(struct jit_ctx *ctx, int args_off, int nregs)
1906 {
1907 	int i;
1908 
1909 	for (i = 0; i < nregs; i++) {
1910 		emit(A64_STR64I(i, A64_SP, args_off), ctx);
1911 		args_off += 8;
1912 	}
1913 }
1914 
1915 static void restore_args(struct jit_ctx *ctx, int args_off, int nregs)
1916 {
1917 	int i;
1918 
1919 	for (i = 0; i < nregs; i++) {
1920 		emit(A64_LDR64I(i, A64_SP, args_off), ctx);
1921 		args_off += 8;
1922 	}
1923 }
1924 
1925 /* Based on the x86's implementation of arch_prepare_bpf_trampoline().
1926  *
1927  * bpf prog and function entry before bpf trampoline hooked:
1928  *   mov x9, lr
1929  *   nop
1930  *
1931  * bpf prog and function entry after bpf trampoline hooked:
1932  *   mov x9, lr
1933  *   bl  <bpf_trampoline or plt>
1934  *
1935  */
1936 static int prepare_trampoline(struct jit_ctx *ctx, struct bpf_tramp_image *im,
1937 			      struct bpf_tramp_links *tlinks, void *func_addr,
1938 			      int nregs, u32 flags)
1939 {
1940 	int i;
1941 	int stack_size;
1942 	int retaddr_off;
1943 	int regs_off;
1944 	int retval_off;
1945 	int args_off;
1946 	int nregs_off;
1947 	int ip_off;
1948 	int run_ctx_off;
1949 	struct bpf_tramp_links *fentry = &tlinks[BPF_TRAMP_FENTRY];
1950 	struct bpf_tramp_links *fexit = &tlinks[BPF_TRAMP_FEXIT];
1951 	struct bpf_tramp_links *fmod_ret = &tlinks[BPF_TRAMP_MODIFY_RETURN];
1952 	bool save_ret;
1953 	__le32 **branches = NULL;
1954 
1955 	/* trampoline stack layout:
1956 	 *                  [ parent ip         ]
1957 	 *                  [ FP                ]
1958 	 * SP + retaddr_off [ self ip           ]
1959 	 *                  [ FP                ]
1960 	 *
1961 	 *                  [ padding           ] align SP to multiples of 16
1962 	 *
1963 	 *                  [ x20               ] callee saved reg x20
1964 	 * SP + regs_off    [ x19               ] callee saved reg x19
1965 	 *
1966 	 * SP + retval_off  [ return value      ] BPF_TRAMP_F_CALL_ORIG or
1967 	 *                                        BPF_TRAMP_F_RET_FENTRY_RET
1968 	 *
1969 	 *                  [ arg reg N         ]
1970 	 *                  [ ...               ]
1971 	 * SP + args_off    [ arg reg 1         ]
1972 	 *
1973 	 * SP + nregs_off   [ arg regs count    ]
1974 	 *
1975 	 * SP + ip_off      [ traced function   ] BPF_TRAMP_F_IP_ARG flag
1976 	 *
1977 	 * SP + run_ctx_off [ bpf_tramp_run_ctx ]
1978 	 */
1979 
1980 	stack_size = 0;
1981 	run_ctx_off = stack_size;
1982 	/* room for bpf_tramp_run_ctx */
1983 	stack_size += round_up(sizeof(struct bpf_tramp_run_ctx), 8);
1984 
1985 	ip_off = stack_size;
1986 	/* room for IP address argument */
1987 	if (flags & BPF_TRAMP_F_IP_ARG)
1988 		stack_size += 8;
1989 
1990 	nregs_off = stack_size;
1991 	/* room for args count */
1992 	stack_size += 8;
1993 
1994 	args_off = stack_size;
1995 	/* room for args */
1996 	stack_size += nregs * 8;
1997 
1998 	/* room for return value */
1999 	retval_off = stack_size;
2000 	save_ret = flags & (BPF_TRAMP_F_CALL_ORIG | BPF_TRAMP_F_RET_FENTRY_RET);
2001 	if (save_ret)
2002 		stack_size += 8;
2003 
2004 	/* room for callee saved registers, currently x19 and x20 are used */
2005 	regs_off = stack_size;
2006 	stack_size += 16;
2007 
2008 	/* round up to multiples of 16 to avoid SPAlignmentFault */
2009 	stack_size = round_up(stack_size, 16);
2010 
2011 	/* return address locates above FP */
2012 	retaddr_off = stack_size + 8;
2013 
2014 	/* bpf trampoline may be invoked by 3 instruction types:
2015 	 * 1. bl, attached to bpf prog or kernel function via short jump
2016 	 * 2. br, attached to bpf prog or kernel function via long jump
2017 	 * 3. blr, working as a function pointer, used by struct_ops.
2018 	 * So BTI_JC should used here to support both br and blr.
2019 	 */
2020 	emit_bti(A64_BTI_JC, ctx);
2021 
2022 	/* frame for parent function */
2023 	emit(A64_PUSH(A64_FP, A64_R(9), A64_SP), ctx);
2024 	emit(A64_MOV(1, A64_FP, A64_SP), ctx);
2025 
2026 	/* frame for patched function */
2027 	emit(A64_PUSH(A64_FP, A64_LR, A64_SP), ctx);
2028 	emit(A64_MOV(1, A64_FP, A64_SP), ctx);
2029 
2030 	/* allocate stack space */
2031 	emit(A64_SUB_I(1, A64_SP, A64_SP, stack_size), ctx);
2032 
2033 	if (flags & BPF_TRAMP_F_IP_ARG) {
2034 		/* save ip address of the traced function */
2035 		emit_addr_mov_i64(A64_R(10), (const u64)func_addr, ctx);
2036 		emit(A64_STR64I(A64_R(10), A64_SP, ip_off), ctx);
2037 	}
2038 
2039 	/* save arg regs count*/
2040 	emit(A64_MOVZ(1, A64_R(10), nregs, 0), ctx);
2041 	emit(A64_STR64I(A64_R(10), A64_SP, nregs_off), ctx);
2042 
2043 	/* save arg regs */
2044 	save_args(ctx, args_off, nregs);
2045 
2046 	/* save callee saved registers */
2047 	emit(A64_STR64I(A64_R(19), A64_SP, regs_off), ctx);
2048 	emit(A64_STR64I(A64_R(20), A64_SP, regs_off + 8), ctx);
2049 
2050 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2051 		emit_addr_mov_i64(A64_R(0), (const u64)im, ctx);
2052 		emit_call((const u64)__bpf_tramp_enter, ctx);
2053 	}
2054 
2055 	for (i = 0; i < fentry->nr_links; i++)
2056 		invoke_bpf_prog(ctx, fentry->links[i], args_off,
2057 				retval_off, run_ctx_off,
2058 				flags & BPF_TRAMP_F_RET_FENTRY_RET);
2059 
2060 	if (fmod_ret->nr_links) {
2061 		branches = kcalloc(fmod_ret->nr_links, sizeof(__le32 *),
2062 				   GFP_KERNEL);
2063 		if (!branches)
2064 			return -ENOMEM;
2065 
2066 		invoke_bpf_mod_ret(ctx, fmod_ret, args_off, retval_off,
2067 				   run_ctx_off, branches);
2068 	}
2069 
2070 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2071 		restore_args(ctx, args_off, nregs);
2072 		/* call original func */
2073 		emit(A64_LDR64I(A64_R(10), A64_SP, retaddr_off), ctx);
2074 		emit(A64_ADR(A64_LR, AARCH64_INSN_SIZE * 2), ctx);
2075 		emit(A64_RET(A64_R(10)), ctx);
2076 		/* store return value */
2077 		emit(A64_STR64I(A64_R(0), A64_SP, retval_off), ctx);
2078 		/* reserve a nop for bpf_tramp_image_put */
2079 		im->ip_after_call = ctx->ro_image + ctx->idx;
2080 		emit(A64_NOP, ctx);
2081 	}
2082 
2083 	/* update the branches saved in invoke_bpf_mod_ret with cbnz */
2084 	for (i = 0; i < fmod_ret->nr_links && ctx->image != NULL; i++) {
2085 		int offset = &ctx->image[ctx->idx] - branches[i];
2086 		*branches[i] = cpu_to_le32(A64_CBNZ(1, A64_R(10), offset));
2087 	}
2088 
2089 	for (i = 0; i < fexit->nr_links; i++)
2090 		invoke_bpf_prog(ctx, fexit->links[i], args_off, retval_off,
2091 				run_ctx_off, false);
2092 
2093 	if (flags & BPF_TRAMP_F_CALL_ORIG) {
2094 		im->ip_epilogue = ctx->ro_image + ctx->idx;
2095 		emit_addr_mov_i64(A64_R(0), (const u64)im, ctx);
2096 		emit_call((const u64)__bpf_tramp_exit, ctx);
2097 	}
2098 
2099 	if (flags & BPF_TRAMP_F_RESTORE_REGS)
2100 		restore_args(ctx, args_off, nregs);
2101 
2102 	/* restore callee saved register x19 and x20 */
2103 	emit(A64_LDR64I(A64_R(19), A64_SP, regs_off), ctx);
2104 	emit(A64_LDR64I(A64_R(20), A64_SP, regs_off + 8), ctx);
2105 
2106 	if (save_ret)
2107 		emit(A64_LDR64I(A64_R(0), A64_SP, retval_off), ctx);
2108 
2109 	/* reset SP  */
2110 	emit(A64_MOV(1, A64_SP, A64_FP), ctx);
2111 
2112 	/* pop frames  */
2113 	emit(A64_POP(A64_FP, A64_LR, A64_SP), ctx);
2114 	emit(A64_POP(A64_FP, A64_R(9), A64_SP), ctx);
2115 
2116 	if (flags & BPF_TRAMP_F_SKIP_FRAME) {
2117 		/* skip patched function, return to parent */
2118 		emit(A64_MOV(1, A64_LR, A64_R(9)), ctx);
2119 		emit(A64_RET(A64_R(9)), ctx);
2120 	} else {
2121 		/* return to patched function */
2122 		emit(A64_MOV(1, A64_R(10), A64_LR), ctx);
2123 		emit(A64_MOV(1, A64_LR, A64_R(9)), ctx);
2124 		emit(A64_RET(A64_R(10)), ctx);
2125 	}
2126 
2127 	kfree(branches);
2128 
2129 	return ctx->idx;
2130 }
2131 
2132 static int btf_func_model_nregs(const struct btf_func_model *m)
2133 {
2134 	int nregs = m->nr_args;
2135 	int i;
2136 
2137 	/* extra registers needed for struct argument */
2138 	for (i = 0; i < MAX_BPF_FUNC_ARGS; i++) {
2139 		/* The arg_size is at most 16 bytes, enforced by the verifier. */
2140 		if (m->arg_flags[i] & BTF_FMODEL_STRUCT_ARG)
2141 			nregs += (m->arg_size[i] + 7) / 8 - 1;
2142 	}
2143 
2144 	return nregs;
2145 }
2146 
2147 int arch_bpf_trampoline_size(const struct btf_func_model *m, u32 flags,
2148 			     struct bpf_tramp_links *tlinks, void *func_addr)
2149 {
2150 	struct jit_ctx ctx = {
2151 		.image = NULL,
2152 		.idx = 0,
2153 	};
2154 	struct bpf_tramp_image im;
2155 	int nregs, ret;
2156 
2157 	nregs = btf_func_model_nregs(m);
2158 	/* the first 8 registers are used for arguments */
2159 	if (nregs > 8)
2160 		return -ENOTSUPP;
2161 
2162 	ret = prepare_trampoline(&ctx, &im, tlinks, func_addr, nregs, flags);
2163 	if (ret < 0)
2164 		return ret;
2165 
2166 	return ret < 0 ? ret : ret * AARCH64_INSN_SIZE;
2167 }
2168 
2169 void *arch_alloc_bpf_trampoline(unsigned int size)
2170 {
2171 	return bpf_prog_pack_alloc(size, jit_fill_hole);
2172 }
2173 
2174 void arch_free_bpf_trampoline(void *image, unsigned int size)
2175 {
2176 	bpf_prog_pack_free(image, size);
2177 }
2178 
2179 void arch_protect_bpf_trampoline(void *image, unsigned int size)
2180 {
2181 }
2182 
2183 void arch_unprotect_bpf_trampoline(void *image, unsigned int size)
2184 {
2185 }
2186 
2187 int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *ro_image,
2188 				void *ro_image_end, const struct btf_func_model *m,
2189 				u32 flags, struct bpf_tramp_links *tlinks,
2190 				void *func_addr)
2191 {
2192 	int ret, nregs;
2193 	void *image, *tmp;
2194 	u32 size = ro_image_end - ro_image;
2195 
2196 	/* image doesn't need to be in module memory range, so we can
2197 	 * use kvmalloc.
2198 	 */
2199 	image = kvmalloc(size, GFP_KERNEL);
2200 	if (!image)
2201 		return -ENOMEM;
2202 
2203 	struct jit_ctx ctx = {
2204 		.image = image,
2205 		.ro_image = ro_image,
2206 		.idx = 0,
2207 	};
2208 
2209 	nregs = btf_func_model_nregs(m);
2210 	/* the first 8 registers are used for arguments */
2211 	if (nregs > 8)
2212 		return -ENOTSUPP;
2213 
2214 	jit_fill_hole(image, (unsigned int)(ro_image_end - ro_image));
2215 	ret = prepare_trampoline(&ctx, im, tlinks, func_addr, nregs, flags);
2216 
2217 	if (ret > 0 && validate_code(&ctx) < 0) {
2218 		ret = -EINVAL;
2219 		goto out;
2220 	}
2221 
2222 	if (ret > 0)
2223 		ret *= AARCH64_INSN_SIZE;
2224 
2225 	tmp = bpf_arch_text_copy(ro_image, image, size);
2226 	if (IS_ERR(tmp)) {
2227 		ret = PTR_ERR(tmp);
2228 		goto out;
2229 	}
2230 
2231 	bpf_flush_icache(ro_image, ro_image + size);
2232 out:
2233 	kvfree(image);
2234 	return ret;
2235 }
2236 
2237 static bool is_long_jump(void *ip, void *target)
2238 {
2239 	long offset;
2240 
2241 	/* NULL target means this is a NOP */
2242 	if (!target)
2243 		return false;
2244 
2245 	offset = (long)target - (long)ip;
2246 	return offset < -SZ_128M || offset >= SZ_128M;
2247 }
2248 
2249 static int gen_branch_or_nop(enum aarch64_insn_branch_type type, void *ip,
2250 			     void *addr, void *plt, u32 *insn)
2251 {
2252 	void *target;
2253 
2254 	if (!addr) {
2255 		*insn = aarch64_insn_gen_nop();
2256 		return 0;
2257 	}
2258 
2259 	if (is_long_jump(ip, addr))
2260 		target = plt;
2261 	else
2262 		target = addr;
2263 
2264 	*insn = aarch64_insn_gen_branch_imm((unsigned long)ip,
2265 					    (unsigned long)target,
2266 					    type);
2267 
2268 	return *insn != AARCH64_BREAK_FAULT ? 0 : -EFAULT;
2269 }
2270 
2271 /* Replace the branch instruction from @ip to @old_addr in a bpf prog or a bpf
2272  * trampoline with the branch instruction from @ip to @new_addr. If @old_addr
2273  * or @new_addr is NULL, the old or new instruction is NOP.
2274  *
2275  * When @ip is the bpf prog entry, a bpf trampoline is being attached or
2276  * detached. Since bpf trampoline and bpf prog are allocated separately with
2277  * vmalloc, the address distance may exceed 128MB, the maximum branch range.
2278  * So long jump should be handled.
2279  *
2280  * When a bpf prog is constructed, a plt pointing to empty trampoline
2281  * dummy_tramp is placed at the end:
2282  *
2283  *      bpf_prog:
2284  *              mov x9, lr
2285  *              nop // patchsite
2286  *              ...
2287  *              ret
2288  *
2289  *      plt:
2290  *              ldr x10, target
2291  *              br x10
2292  *      target:
2293  *              .quad dummy_tramp // plt target
2294  *
2295  * This is also the state when no trampoline is attached.
2296  *
2297  * When a short-jump bpf trampoline is attached, the patchsite is patched
2298  * to a bl instruction to the trampoline directly:
2299  *
2300  *      bpf_prog:
2301  *              mov x9, lr
2302  *              bl <short-jump bpf trampoline address> // patchsite
2303  *              ...
2304  *              ret
2305  *
2306  *      plt:
2307  *              ldr x10, target
2308  *              br x10
2309  *      target:
2310  *              .quad dummy_tramp // plt target
2311  *
2312  * When a long-jump bpf trampoline is attached, the plt target is filled with
2313  * the trampoline address and the patchsite is patched to a bl instruction to
2314  * the plt:
2315  *
2316  *      bpf_prog:
2317  *              mov x9, lr
2318  *              bl plt // patchsite
2319  *              ...
2320  *              ret
2321  *
2322  *      plt:
2323  *              ldr x10, target
2324  *              br x10
2325  *      target:
2326  *              .quad <long-jump bpf trampoline address> // plt target
2327  *
2328  * The dummy_tramp is used to prevent another CPU from jumping to unknown
2329  * locations during the patching process, making the patching process easier.
2330  */
2331 int bpf_arch_text_poke(void *ip, enum bpf_text_poke_type poke_type,
2332 		       void *old_addr, void *new_addr)
2333 {
2334 	int ret;
2335 	u32 old_insn;
2336 	u32 new_insn;
2337 	u32 replaced;
2338 	struct bpf_plt *plt = NULL;
2339 	unsigned long size = 0UL;
2340 	unsigned long offset = ~0UL;
2341 	enum aarch64_insn_branch_type branch_type;
2342 	char namebuf[KSYM_NAME_LEN];
2343 	void *image = NULL;
2344 	u64 plt_target = 0ULL;
2345 	bool poking_bpf_entry;
2346 
2347 	if (!__bpf_address_lookup((unsigned long)ip, &size, &offset, namebuf))
2348 		/* Only poking bpf text is supported. Since kernel function
2349 		 * entry is set up by ftrace, we reply on ftrace to poke kernel
2350 		 * functions.
2351 		 */
2352 		return -ENOTSUPP;
2353 
2354 	image = ip - offset;
2355 	/* zero offset means we're poking bpf prog entry */
2356 	poking_bpf_entry = (offset == 0UL);
2357 
2358 	/* bpf prog entry, find plt and the real patchsite */
2359 	if (poking_bpf_entry) {
2360 		/* plt locates at the end of bpf prog */
2361 		plt = image + size - PLT_TARGET_OFFSET;
2362 
2363 		/* skip to the nop instruction in bpf prog entry:
2364 		 * bti c // if BTI enabled
2365 		 * mov x9, x30
2366 		 * nop
2367 		 */
2368 		ip = image + POKE_OFFSET * AARCH64_INSN_SIZE;
2369 	}
2370 
2371 	/* long jump is only possible at bpf prog entry */
2372 	if (WARN_ON((is_long_jump(ip, new_addr) || is_long_jump(ip, old_addr)) &&
2373 		    !poking_bpf_entry))
2374 		return -EINVAL;
2375 
2376 	if (poke_type == BPF_MOD_CALL)
2377 		branch_type = AARCH64_INSN_BRANCH_LINK;
2378 	else
2379 		branch_type = AARCH64_INSN_BRANCH_NOLINK;
2380 
2381 	if (gen_branch_or_nop(branch_type, ip, old_addr, plt, &old_insn) < 0)
2382 		return -EFAULT;
2383 
2384 	if (gen_branch_or_nop(branch_type, ip, new_addr, plt, &new_insn) < 0)
2385 		return -EFAULT;
2386 
2387 	if (is_long_jump(ip, new_addr))
2388 		plt_target = (u64)new_addr;
2389 	else if (is_long_jump(ip, old_addr))
2390 		/* if the old target is a long jump and the new target is not,
2391 		 * restore the plt target to dummy_tramp, so there is always a
2392 		 * legal and harmless address stored in plt target, and we'll
2393 		 * never jump from plt to an unknown place.
2394 		 */
2395 		plt_target = (u64)&dummy_tramp;
2396 
2397 	if (plt_target) {
2398 		/* non-zero plt_target indicates we're patching a bpf prog,
2399 		 * which is read only.
2400 		 */
2401 		if (set_memory_rw(PAGE_MASK & ((uintptr_t)&plt->target), 1))
2402 			return -EFAULT;
2403 		WRITE_ONCE(plt->target, plt_target);
2404 		set_memory_ro(PAGE_MASK & ((uintptr_t)&plt->target), 1);
2405 		/* since plt target points to either the new trampoline
2406 		 * or dummy_tramp, even if another CPU reads the old plt
2407 		 * target value before fetching the bl instruction to plt,
2408 		 * it will be brought back by dummy_tramp, so no barrier is
2409 		 * required here.
2410 		 */
2411 	}
2412 
2413 	/* if the old target and the new target are both long jumps, no
2414 	 * patching is required
2415 	 */
2416 	if (old_insn == new_insn)
2417 		return 0;
2418 
2419 	mutex_lock(&text_mutex);
2420 	if (aarch64_insn_read(ip, &replaced)) {
2421 		ret = -EFAULT;
2422 		goto out;
2423 	}
2424 
2425 	if (replaced != old_insn) {
2426 		ret = -EFAULT;
2427 		goto out;
2428 	}
2429 
2430 	/* We call aarch64_insn_patch_text_nosync() to replace instruction
2431 	 * atomically, so no other CPUs will fetch a half-new and half-old
2432 	 * instruction. But there is chance that another CPU executes the
2433 	 * old instruction after the patching operation finishes (e.g.,
2434 	 * pipeline not flushed, or icache not synchronized yet).
2435 	 *
2436 	 * 1. when a new trampoline is attached, it is not a problem for
2437 	 *    different CPUs to jump to different trampolines temporarily.
2438 	 *
2439 	 * 2. when an old trampoline is freed, we should wait for all other
2440 	 *    CPUs to exit the trampoline and make sure the trampoline is no
2441 	 *    longer reachable, since bpf_tramp_image_put() function already
2442 	 *    uses percpu_ref and task-based rcu to do the sync, no need to call
2443 	 *    the sync version here, see bpf_tramp_image_put() for details.
2444 	 */
2445 	ret = aarch64_insn_patch_text_nosync(ip, new_insn);
2446 out:
2447 	mutex_unlock(&text_mutex);
2448 
2449 	return ret;
2450 }
2451 
2452 bool bpf_jit_supports_ptr_xchg(void)
2453 {
2454 	return true;
2455 }
2456 
2457 bool bpf_jit_supports_exceptions(void)
2458 {
2459 	/* We unwind through both kernel frames starting from within bpf_throw
2460 	 * call and BPF frames. Therefore we require FP unwinder to be enabled
2461 	 * to walk kernel frames and reach BPF frames in the stack trace.
2462 	 * ARM64 kernel is aways compiled with CONFIG_FRAME_POINTER=y
2463 	 */
2464 	return true;
2465 }
2466 
2467 void bpf_jit_free(struct bpf_prog *prog)
2468 {
2469 	if (prog->jited) {
2470 		struct arm64_jit_data *jit_data = prog->aux->jit_data;
2471 		struct bpf_binary_header *hdr;
2472 
2473 		/*
2474 		 * If we fail the final pass of JIT (from jit_subprogs),
2475 		 * the program may not be finalized yet. Call finalize here
2476 		 * before freeing it.
2477 		 */
2478 		if (jit_data) {
2479 			bpf_arch_text_copy(&jit_data->ro_header->size, &jit_data->header->size,
2480 					   sizeof(jit_data->header->size));
2481 			kfree(jit_data);
2482 		}
2483 		hdr = bpf_jit_binary_pack_hdr(prog);
2484 		bpf_jit_binary_pack_free(hdr, NULL);
2485 		WARN_ON_ONCE(!bpf_prog_kallsyms_verify_off(prog));
2486 	}
2487 
2488 	bpf_prog_unlock_free(prog);
2489 }
2490