xref: /linux/arch/arm/net/bpf_jit_32.c (revision 140eb5227767c6754742020a16d2691222b9c19b)
1 /*
2  * Just-In-Time compiler for eBPF filters on 32bit ARM
3  *
4  * Copyright (c) 2017 Shubham Bansal <illusionist.neo@gmail.com>
5  * Copyright (c) 2011 Mircea Gherzan <mgherzan@gmail.com>
6  *
7  * This program is free software; you can redistribute it and/or modify it
8  * under the terms of the GNU General Public License as published by the
9  * Free Software Foundation; version 2 of the License.
10  */
11 
12 #include <linux/bpf.h>
13 #include <linux/bitops.h>
14 #include <linux/compiler.h>
15 #include <linux/errno.h>
16 #include <linux/filter.h>
17 #include <linux/netdevice.h>
18 #include <linux/string.h>
19 #include <linux/slab.h>
20 #include <linux/if_vlan.h>
21 
22 #include <asm/cacheflush.h>
23 #include <asm/hwcap.h>
24 #include <asm/opcodes.h>
25 
26 #include "bpf_jit_32.h"
27 
28 int bpf_jit_enable __read_mostly;
29 
30 /*
31  * eBPF prog stack layout:
32  *
33  *                         high
34  * original ARM_SP =>     +-----+
35  *                        |     | callee saved registers
36  *                        +-----+ <= (BPF_FP + SCRATCH_SIZE)
37  *                        | ... | eBPF JIT scratch space
38  * eBPF fp register =>    +-----+
39  *   (BPF_FP)             | ... | eBPF prog stack
40  *                        +-----+
41  *                        |RSVD | JIT scratchpad
42  * current ARM_SP =>      +-----+ <= (BPF_FP - STACK_SIZE + SCRATCH_SIZE)
43  *                        |     |
44  *                        | ... | Function call stack
45  *                        |     |
46  *                        +-----+
47  *                          low
48  *
49  * The callee saved registers depends on whether frame pointers are enabled.
50  * With frame pointers (to be compliant with the ABI):
51  *
52  *                                high
53  * original ARM_SP =>     +------------------+ \
54  *                        |        pc        | |
55  * current ARM_FP =>      +------------------+ } callee saved registers
56  *                        |r4-r8,r10,fp,ip,lr| |
57  *                        +------------------+ /
58  *                                low
59  *
60  * Without frame pointers:
61  *
62  *                                high
63  * original ARM_SP =>     +------------------+
64  *                        | r4-r8,r10,fp,lr  | callee saved registers
65  * current ARM_FP =>      +------------------+
66  *                                low
67  *
68  * When popping registers off the stack at the end of a BPF function, we
69  * reference them via the current ARM_FP register.
70  */
71 #define CALLEE_MASK	(1 << ARM_R4 | 1 << ARM_R5 | 1 << ARM_R6 | \
72 			 1 << ARM_R7 | 1 << ARM_R8 | 1 << ARM_R10 | \
73 			 1 << ARM_FP)
74 #define CALLEE_PUSH_MASK (CALLEE_MASK | 1 << ARM_LR)
75 #define CALLEE_POP_MASK  (CALLEE_MASK | 1 << ARM_PC)
76 
77 #define STACK_OFFSET(k)	(k)
78 #define TMP_REG_1	(MAX_BPF_JIT_REG + 0)	/* TEMP Register 1 */
79 #define TMP_REG_2	(MAX_BPF_JIT_REG + 1)	/* TEMP Register 2 */
80 #define TCALL_CNT	(MAX_BPF_JIT_REG + 2)	/* Tail Call Count */
81 
82 #define FLAG_IMM_OVERFLOW	(1 << 0)
83 
84 /*
85  * Map eBPF registers to ARM 32bit registers or stack scratch space.
86  *
87  * 1. First argument is passed using the arm 32bit registers and rest of the
88  * arguments are passed on stack scratch space.
89  * 2. First callee-saved arugument is mapped to arm 32 bit registers and rest
90  * arguments are mapped to scratch space on stack.
91  * 3. We need two 64 bit temp registers to do complex operations on eBPF
92  * registers.
93  *
94  * As the eBPF registers are all 64 bit registers and arm has only 32 bit
95  * registers, we have to map each eBPF registers with two arm 32 bit regs or
96  * scratch memory space and we have to build eBPF 64 bit register from those.
97  *
98  */
99 static const u8 bpf2a32[][2] = {
100 	/* return value from in-kernel function, and exit value from eBPF */
101 	[BPF_REG_0] = {ARM_R1, ARM_R0},
102 	/* arguments from eBPF program to in-kernel function */
103 	[BPF_REG_1] = {ARM_R3, ARM_R2},
104 	/* Stored on stack scratch space */
105 	[BPF_REG_2] = {STACK_OFFSET(0), STACK_OFFSET(4)},
106 	[BPF_REG_3] = {STACK_OFFSET(8), STACK_OFFSET(12)},
107 	[BPF_REG_4] = {STACK_OFFSET(16), STACK_OFFSET(20)},
108 	[BPF_REG_5] = {STACK_OFFSET(24), STACK_OFFSET(28)},
109 	/* callee saved registers that in-kernel function will preserve */
110 	[BPF_REG_6] = {ARM_R5, ARM_R4},
111 	/* Stored on stack scratch space */
112 	[BPF_REG_7] = {STACK_OFFSET(32), STACK_OFFSET(36)},
113 	[BPF_REG_8] = {STACK_OFFSET(40), STACK_OFFSET(44)},
114 	[BPF_REG_9] = {STACK_OFFSET(48), STACK_OFFSET(52)},
115 	/* Read only Frame Pointer to access Stack */
116 	[BPF_REG_FP] = {STACK_OFFSET(56), STACK_OFFSET(60)},
117 	/* Temporary Register for internal BPF JIT, can be used
118 	 * for constant blindings and others.
119 	 */
120 	[TMP_REG_1] = {ARM_R7, ARM_R6},
121 	[TMP_REG_2] = {ARM_R10, ARM_R8},
122 	/* Tail call count. Stored on stack scratch space. */
123 	[TCALL_CNT] = {STACK_OFFSET(64), STACK_OFFSET(68)},
124 	/* temporary register for blinding constants.
125 	 * Stored on stack scratch space.
126 	 */
127 	[BPF_REG_AX] = {STACK_OFFSET(72), STACK_OFFSET(76)},
128 };
129 
130 #define	dst_lo	dst[1]
131 #define dst_hi	dst[0]
132 #define src_lo	src[1]
133 #define src_hi	src[0]
134 
135 /*
136  * JIT Context:
137  *
138  * prog			:	bpf_prog
139  * idx			:	index of current last JITed instruction.
140  * prologue_bytes	:	bytes used in prologue.
141  * epilogue_offset	:	offset of epilogue starting.
142  * offsets		:	array of eBPF instruction offsets in
143  *				JITed code.
144  * target		:	final JITed code.
145  * epilogue_bytes	:	no of bytes used in epilogue.
146  * imm_count		:	no of immediate counts used for global
147  *				variables.
148  * imms			:	array of global variable addresses.
149  */
150 
151 struct jit_ctx {
152 	const struct bpf_prog *prog;
153 	unsigned int idx;
154 	unsigned int prologue_bytes;
155 	unsigned int epilogue_offset;
156 	u32 flags;
157 	u32 *offsets;
158 	u32 *target;
159 	u32 stack_size;
160 #if __LINUX_ARM_ARCH__ < 7
161 	u16 epilogue_bytes;
162 	u16 imm_count;
163 	u32 *imms;
164 #endif
165 };
166 
167 /*
168  * Wrappers which handle both OABI and EABI and assures Thumb2 interworking
169  * (where the assembly routines like __aeabi_uidiv could cause problems).
170  */
171 static u32 jit_udiv32(u32 dividend, u32 divisor)
172 {
173 	return dividend / divisor;
174 }
175 
176 static u32 jit_mod32(u32 dividend, u32 divisor)
177 {
178 	return dividend % divisor;
179 }
180 
181 static inline void _emit(int cond, u32 inst, struct jit_ctx *ctx)
182 {
183 	inst |= (cond << 28);
184 	inst = __opcode_to_mem_arm(inst);
185 
186 	if (ctx->target != NULL)
187 		ctx->target[ctx->idx] = inst;
188 
189 	ctx->idx++;
190 }
191 
192 /*
193  * Emit an instruction that will be executed unconditionally.
194  */
195 static inline void emit(u32 inst, struct jit_ctx *ctx)
196 {
197 	_emit(ARM_COND_AL, inst, ctx);
198 }
199 
200 /*
201  * Checks if immediate value can be converted to imm12(12 bits) value.
202  */
203 static int16_t imm8m(u32 x)
204 {
205 	u32 rot;
206 
207 	for (rot = 0; rot < 16; rot++)
208 		if ((x & ~ror32(0xff, 2 * rot)) == 0)
209 			return rol32(x, 2 * rot) | (rot << 8);
210 	return -1;
211 }
212 
213 /*
214  * Initializes the JIT space with undefined instructions.
215  */
216 static void jit_fill_hole(void *area, unsigned int size)
217 {
218 	u32 *ptr;
219 	/* We are guaranteed to have aligned memory. */
220 	for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
221 		*ptr++ = __opcode_to_mem_arm(ARM_INST_UDF);
222 }
223 
224 #if defined(CONFIG_AEABI) && (__LINUX_ARM_ARCH__ >= 5)
225 /* EABI requires the stack to be aligned to 64-bit boundaries */
226 #define STACK_ALIGNMENT	8
227 #else
228 /* Stack must be aligned to 32-bit boundaries */
229 #define STACK_ALIGNMENT	4
230 #endif
231 
232 /* Stack space for BPF_REG_2, BPF_REG_3, BPF_REG_4,
233  * BPF_REG_5, BPF_REG_7, BPF_REG_8, BPF_REG_9,
234  * BPF_REG_FP and Tail call counts.
235  */
236 #define SCRATCH_SIZE 80
237 
238 /* total stack size used in JITed code */
239 #define _STACK_SIZE \
240 	(ctx->prog->aux->stack_depth + \
241 	 + SCRATCH_SIZE + \
242 	 + 4 /* extra for skb_copy_bits buffer */)
243 
244 #define STACK_SIZE ALIGN(_STACK_SIZE, STACK_ALIGNMENT)
245 
246 /* Get the offset of eBPF REGISTERs stored on scratch space. */
247 #define STACK_VAR(off) (STACK_SIZE-off-4)
248 
249 /* Offset of skb_copy_bits buffer */
250 #define SKB_BUFFER STACK_VAR(SCRATCH_SIZE)
251 
252 #if __LINUX_ARM_ARCH__ < 7
253 
254 static u16 imm_offset(u32 k, struct jit_ctx *ctx)
255 {
256 	unsigned int i = 0, offset;
257 	u16 imm;
258 
259 	/* on the "fake" run we just count them (duplicates included) */
260 	if (ctx->target == NULL) {
261 		ctx->imm_count++;
262 		return 0;
263 	}
264 
265 	while ((i < ctx->imm_count) && ctx->imms[i]) {
266 		if (ctx->imms[i] == k)
267 			break;
268 		i++;
269 	}
270 
271 	if (ctx->imms[i] == 0)
272 		ctx->imms[i] = k;
273 
274 	/* constants go just after the epilogue */
275 	offset =  ctx->offsets[ctx->prog->len - 1] * 4;
276 	offset += ctx->prologue_bytes;
277 	offset += ctx->epilogue_bytes;
278 	offset += i * 4;
279 
280 	ctx->target[offset / 4] = k;
281 
282 	/* PC in ARM mode == address of the instruction + 8 */
283 	imm = offset - (8 + ctx->idx * 4);
284 
285 	if (imm & ~0xfff) {
286 		/*
287 		 * literal pool is too far, signal it into flags. we
288 		 * can only detect it on the second pass unfortunately.
289 		 */
290 		ctx->flags |= FLAG_IMM_OVERFLOW;
291 		return 0;
292 	}
293 
294 	return imm;
295 }
296 
297 #endif /* __LINUX_ARM_ARCH__ */
298 
299 static inline int bpf2a32_offset(int bpf_to, int bpf_from,
300 				 const struct jit_ctx *ctx) {
301 	int to, from;
302 
303 	if (ctx->target == NULL)
304 		return 0;
305 	to = ctx->offsets[bpf_to];
306 	from = ctx->offsets[bpf_from];
307 
308 	return to - from - 1;
309 }
310 
311 /*
312  * Move an immediate that's not an imm8m to a core register.
313  */
314 static inline void emit_mov_i_no8m(const u8 rd, u32 val, struct jit_ctx *ctx)
315 {
316 #if __LINUX_ARM_ARCH__ < 7
317 	emit(ARM_LDR_I(rd, ARM_PC, imm_offset(val, ctx)), ctx);
318 #else
319 	emit(ARM_MOVW(rd, val & 0xffff), ctx);
320 	if (val > 0xffff)
321 		emit(ARM_MOVT(rd, val >> 16), ctx);
322 #endif
323 }
324 
325 static inline void emit_mov_i(const u8 rd, u32 val, struct jit_ctx *ctx)
326 {
327 	int imm12 = imm8m(val);
328 
329 	if (imm12 >= 0)
330 		emit(ARM_MOV_I(rd, imm12), ctx);
331 	else
332 		emit_mov_i_no8m(rd, val, ctx);
333 }
334 
335 static void emit_bx_r(u8 tgt_reg, struct jit_ctx *ctx)
336 {
337 	if (elf_hwcap & HWCAP_THUMB)
338 		emit(ARM_BX(tgt_reg), ctx);
339 	else
340 		emit(ARM_MOV_R(ARM_PC, tgt_reg), ctx);
341 }
342 
343 static inline void emit_blx_r(u8 tgt_reg, struct jit_ctx *ctx)
344 {
345 #if __LINUX_ARM_ARCH__ < 5
346 	emit(ARM_MOV_R(ARM_LR, ARM_PC), ctx);
347 	emit_bx_r(tgt_reg, ctx);
348 #else
349 	emit(ARM_BLX_R(tgt_reg), ctx);
350 #endif
351 }
352 
353 static inline int epilogue_offset(const struct jit_ctx *ctx)
354 {
355 	int to, from;
356 	/* No need for 1st dummy run */
357 	if (ctx->target == NULL)
358 		return 0;
359 	to = ctx->epilogue_offset;
360 	from = ctx->idx;
361 
362 	return to - from - 2;
363 }
364 
365 static inline void emit_udivmod(u8 rd, u8 rm, u8 rn, struct jit_ctx *ctx, u8 op)
366 {
367 	const u8 *tmp = bpf2a32[TMP_REG_1];
368 	s32 jmp_offset;
369 
370 	/* checks if divisor is zero or not. If it is, then
371 	 * exit directly.
372 	 */
373 	emit(ARM_CMP_I(rn, 0), ctx);
374 	_emit(ARM_COND_EQ, ARM_MOV_I(ARM_R0, 0), ctx);
375 	jmp_offset = epilogue_offset(ctx);
376 	_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
377 #if __LINUX_ARM_ARCH__ == 7
378 	if (elf_hwcap & HWCAP_IDIVA) {
379 		if (op == BPF_DIV)
380 			emit(ARM_UDIV(rd, rm, rn), ctx);
381 		else {
382 			emit(ARM_UDIV(ARM_IP, rm, rn), ctx);
383 			emit(ARM_MLS(rd, rn, ARM_IP, rm), ctx);
384 		}
385 		return;
386 	}
387 #endif
388 
389 	/*
390 	 * For BPF_ALU | BPF_DIV | BPF_K instructions
391 	 * As ARM_R1 and ARM_R0 contains 1st argument of bpf
392 	 * function, we need to save it on caller side to save
393 	 * it from getting destroyed within callee.
394 	 * After the return from the callee, we restore ARM_R0
395 	 * ARM_R1.
396 	 */
397 	if (rn != ARM_R1) {
398 		emit(ARM_MOV_R(tmp[0], ARM_R1), ctx);
399 		emit(ARM_MOV_R(ARM_R1, rn), ctx);
400 	}
401 	if (rm != ARM_R0) {
402 		emit(ARM_MOV_R(tmp[1], ARM_R0), ctx);
403 		emit(ARM_MOV_R(ARM_R0, rm), ctx);
404 	}
405 
406 	/* Call appropriate function */
407 	emit_mov_i(ARM_IP, op == BPF_DIV ?
408 		   (u32)jit_udiv32 : (u32)jit_mod32, ctx);
409 	emit_blx_r(ARM_IP, ctx);
410 
411 	/* Save return value */
412 	if (rd != ARM_R0)
413 		emit(ARM_MOV_R(rd, ARM_R0), ctx);
414 
415 	/* Restore ARM_R0 and ARM_R1 */
416 	if (rn != ARM_R1)
417 		emit(ARM_MOV_R(ARM_R1, tmp[0]), ctx);
418 	if (rm != ARM_R0)
419 		emit(ARM_MOV_R(ARM_R0, tmp[1]), ctx);
420 }
421 
422 /* Checks whether BPF register is on scratch stack space or not. */
423 static inline bool is_on_stack(u8 bpf_reg)
424 {
425 	static u8 stack_regs[] = {BPF_REG_AX, BPF_REG_3, BPF_REG_4, BPF_REG_5,
426 				BPF_REG_7, BPF_REG_8, BPF_REG_9, TCALL_CNT,
427 				BPF_REG_2, BPF_REG_FP};
428 	int i, reg_len = sizeof(stack_regs);
429 
430 	for (i = 0 ; i < reg_len ; i++) {
431 		if (bpf_reg == stack_regs[i])
432 			return true;
433 	}
434 	return false;
435 }
436 
437 static inline void emit_a32_mov_i(const u8 dst, const u32 val,
438 				  bool dstk, struct jit_ctx *ctx)
439 {
440 	const u8 *tmp = bpf2a32[TMP_REG_1];
441 
442 	if (dstk) {
443 		emit_mov_i(tmp[1], val, ctx);
444 		emit(ARM_STR_I(tmp[1], ARM_SP, STACK_VAR(dst)), ctx);
445 	} else {
446 		emit_mov_i(dst, val, ctx);
447 	}
448 }
449 
450 /* Sign extended move */
451 static inline void emit_a32_mov_i64(const bool is64, const u8 dst[],
452 				  const u32 val, bool dstk,
453 				  struct jit_ctx *ctx) {
454 	u32 hi = 0;
455 
456 	if (is64 && (val & (1<<31)))
457 		hi = (u32)~0;
458 	emit_a32_mov_i(dst_lo, val, dstk, ctx);
459 	emit_a32_mov_i(dst_hi, hi, dstk, ctx);
460 }
461 
462 static inline void emit_a32_add_r(const u8 dst, const u8 src,
463 			      const bool is64, const bool hi,
464 			      struct jit_ctx *ctx) {
465 	/* 64 bit :
466 	 *	adds dst_lo, dst_lo, src_lo
467 	 *	adc dst_hi, dst_hi, src_hi
468 	 * 32 bit :
469 	 *	add dst_lo, dst_lo, src_lo
470 	 */
471 	if (!hi && is64)
472 		emit(ARM_ADDS_R(dst, dst, src), ctx);
473 	else if (hi && is64)
474 		emit(ARM_ADC_R(dst, dst, src), ctx);
475 	else
476 		emit(ARM_ADD_R(dst, dst, src), ctx);
477 }
478 
479 static inline void emit_a32_sub_r(const u8 dst, const u8 src,
480 				  const bool is64, const bool hi,
481 				  struct jit_ctx *ctx) {
482 	/* 64 bit :
483 	 *	subs dst_lo, dst_lo, src_lo
484 	 *	sbc dst_hi, dst_hi, src_hi
485 	 * 32 bit :
486 	 *	sub dst_lo, dst_lo, src_lo
487 	 */
488 	if (!hi && is64)
489 		emit(ARM_SUBS_R(dst, dst, src), ctx);
490 	else if (hi && is64)
491 		emit(ARM_SBC_R(dst, dst, src), ctx);
492 	else
493 		emit(ARM_SUB_R(dst, dst, src), ctx);
494 }
495 
496 static inline void emit_alu_r(const u8 dst, const u8 src, const bool is64,
497 			      const bool hi, const u8 op, struct jit_ctx *ctx){
498 	switch (BPF_OP(op)) {
499 	/* dst = dst + src */
500 	case BPF_ADD:
501 		emit_a32_add_r(dst, src, is64, hi, ctx);
502 		break;
503 	/* dst = dst - src */
504 	case BPF_SUB:
505 		emit_a32_sub_r(dst, src, is64, hi, ctx);
506 		break;
507 	/* dst = dst | src */
508 	case BPF_OR:
509 		emit(ARM_ORR_R(dst, dst, src), ctx);
510 		break;
511 	/* dst = dst & src */
512 	case BPF_AND:
513 		emit(ARM_AND_R(dst, dst, src), ctx);
514 		break;
515 	/* dst = dst ^ src */
516 	case BPF_XOR:
517 		emit(ARM_EOR_R(dst, dst, src), ctx);
518 		break;
519 	/* dst = dst * src */
520 	case BPF_MUL:
521 		emit(ARM_MUL(dst, dst, src), ctx);
522 		break;
523 	/* dst = dst << src */
524 	case BPF_LSH:
525 		emit(ARM_LSL_R(dst, dst, src), ctx);
526 		break;
527 	/* dst = dst >> src */
528 	case BPF_RSH:
529 		emit(ARM_LSR_R(dst, dst, src), ctx);
530 		break;
531 	/* dst = dst >> src (signed)*/
532 	case BPF_ARSH:
533 		emit(ARM_MOV_SR(dst, dst, SRTYPE_ASR, src), ctx);
534 		break;
535 	}
536 }
537 
538 /* ALU operation (32 bit)
539  * dst = dst (op) src
540  */
541 static inline void emit_a32_alu_r(const u8 dst, const u8 src,
542 				  bool dstk, bool sstk,
543 				  struct jit_ctx *ctx, const bool is64,
544 				  const bool hi, const u8 op) {
545 	const u8 *tmp = bpf2a32[TMP_REG_1];
546 	u8 rn = sstk ? tmp[1] : src;
547 
548 	if (sstk)
549 		emit(ARM_LDR_I(rn, ARM_SP, STACK_VAR(src)), ctx);
550 
551 	/* ALU operation */
552 	if (dstk) {
553 		emit(ARM_LDR_I(tmp[0], ARM_SP, STACK_VAR(dst)), ctx);
554 		emit_alu_r(tmp[0], rn, is64, hi, op, ctx);
555 		emit(ARM_STR_I(tmp[0], ARM_SP, STACK_VAR(dst)), ctx);
556 	} else {
557 		emit_alu_r(dst, rn, is64, hi, op, ctx);
558 	}
559 }
560 
561 /* ALU operation (64 bit) */
562 static inline void emit_a32_alu_r64(const bool is64, const u8 dst[],
563 				  const u8 src[], bool dstk,
564 				  bool sstk, struct jit_ctx *ctx,
565 				  const u8 op) {
566 	emit_a32_alu_r(dst_lo, src_lo, dstk, sstk, ctx, is64, false, op);
567 	if (is64)
568 		emit_a32_alu_r(dst_hi, src_hi, dstk, sstk, ctx, is64, true, op);
569 	else
570 		emit_a32_mov_i(dst_hi, 0, dstk, ctx);
571 }
572 
573 /* dst = imm (4 bytes)*/
574 static inline void emit_a32_mov_r(const u8 dst, const u8 src,
575 				  bool dstk, bool sstk,
576 				  struct jit_ctx *ctx) {
577 	const u8 *tmp = bpf2a32[TMP_REG_1];
578 	u8 rt = sstk ? tmp[0] : src;
579 
580 	if (sstk)
581 		emit(ARM_LDR_I(tmp[0], ARM_SP, STACK_VAR(src)), ctx);
582 	if (dstk)
583 		emit(ARM_STR_I(rt, ARM_SP, STACK_VAR(dst)), ctx);
584 	else
585 		emit(ARM_MOV_R(dst, rt), ctx);
586 }
587 
588 /* dst = src */
589 static inline void emit_a32_mov_r64(const bool is64, const u8 dst[],
590 				  const u8 src[], bool dstk,
591 				  bool sstk, struct jit_ctx *ctx) {
592 	emit_a32_mov_r(dst_lo, src_lo, dstk, sstk, ctx);
593 	if (is64) {
594 		/* complete 8 byte move */
595 		emit_a32_mov_r(dst_hi, src_hi, dstk, sstk, ctx);
596 	} else {
597 		/* Zero out high 4 bytes */
598 		emit_a32_mov_i(dst_hi, 0, dstk, ctx);
599 	}
600 }
601 
602 /* Shift operations */
603 static inline void emit_a32_alu_i(const u8 dst, const u32 val, bool dstk,
604 				struct jit_ctx *ctx, const u8 op) {
605 	const u8 *tmp = bpf2a32[TMP_REG_1];
606 	u8 rd = dstk ? tmp[0] : dst;
607 
608 	if (dstk)
609 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst)), ctx);
610 
611 	/* Do shift operation */
612 	switch (op) {
613 	case BPF_LSH:
614 		emit(ARM_LSL_I(rd, rd, val), ctx);
615 		break;
616 	case BPF_RSH:
617 		emit(ARM_LSR_I(rd, rd, val), ctx);
618 		break;
619 	case BPF_NEG:
620 		emit(ARM_RSB_I(rd, rd, val), ctx);
621 		break;
622 	}
623 
624 	if (dstk)
625 		emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst)), ctx);
626 }
627 
628 /* dst = ~dst (64 bit) */
629 static inline void emit_a32_neg64(const u8 dst[], bool dstk,
630 				struct jit_ctx *ctx){
631 	const u8 *tmp = bpf2a32[TMP_REG_1];
632 	u8 rd = dstk ? tmp[1] : dst[1];
633 	u8 rm = dstk ? tmp[0] : dst[0];
634 
635 	/* Setup Operand */
636 	if (dstk) {
637 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
638 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
639 	}
640 
641 	/* Do Negate Operation */
642 	emit(ARM_RSBS_I(rd, rd, 0), ctx);
643 	emit(ARM_RSC_I(rm, rm, 0), ctx);
644 
645 	if (dstk) {
646 		emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
647 		emit(ARM_STR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
648 	}
649 }
650 
651 /* dst = dst << src */
652 static inline void emit_a32_lsh_r64(const u8 dst[], const u8 src[], bool dstk,
653 				    bool sstk, struct jit_ctx *ctx) {
654 	const u8 *tmp = bpf2a32[TMP_REG_1];
655 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
656 
657 	/* Setup Operands */
658 	u8 rt = sstk ? tmp2[1] : src_lo;
659 	u8 rd = dstk ? tmp[1] : dst_lo;
660 	u8 rm = dstk ? tmp[0] : dst_hi;
661 
662 	if (sstk)
663 		emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(src_lo)), ctx);
664 	if (dstk) {
665 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
666 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
667 	}
668 
669 	/* Do LSH operation */
670 	emit(ARM_SUB_I(ARM_IP, rt, 32), ctx);
671 	emit(ARM_RSB_I(tmp2[0], rt, 32), ctx);
672 	emit(ARM_MOV_SR(ARM_LR, rm, SRTYPE_ASL, rt), ctx);
673 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd, SRTYPE_ASL, ARM_IP), ctx);
674 	emit(ARM_ORR_SR(ARM_IP, ARM_LR, rd, SRTYPE_LSR, tmp2[0]), ctx);
675 	emit(ARM_MOV_SR(ARM_LR, rd, SRTYPE_ASL, rt), ctx);
676 
677 	if (dstk) {
678 		emit(ARM_STR_I(ARM_LR, ARM_SP, STACK_VAR(dst_lo)), ctx);
679 		emit(ARM_STR_I(ARM_IP, ARM_SP, STACK_VAR(dst_hi)), ctx);
680 	} else {
681 		emit(ARM_MOV_R(rd, ARM_LR), ctx);
682 		emit(ARM_MOV_R(rm, ARM_IP), ctx);
683 	}
684 }
685 
686 /* dst = dst >> src (signed)*/
687 static inline void emit_a32_arsh_r64(const u8 dst[], const u8 src[], bool dstk,
688 				    bool sstk, struct jit_ctx *ctx) {
689 	const u8 *tmp = bpf2a32[TMP_REG_1];
690 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
691 	/* Setup Operands */
692 	u8 rt = sstk ? tmp2[1] : src_lo;
693 	u8 rd = dstk ? tmp[1] : dst_lo;
694 	u8 rm = dstk ? tmp[0] : dst_hi;
695 
696 	if (sstk)
697 		emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(src_lo)), ctx);
698 	if (dstk) {
699 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
700 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
701 	}
702 
703 	/* Do the ARSH operation */
704 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
705 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
706 	emit(ARM_MOV_SR(ARM_LR, rd, SRTYPE_LSR, rt), ctx);
707 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rm, SRTYPE_ASL, ARM_IP), ctx);
708 	_emit(ARM_COND_MI, ARM_B(0), ctx);
709 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rm, SRTYPE_ASR, tmp2[0]), ctx);
710 	emit(ARM_MOV_SR(ARM_IP, rm, SRTYPE_ASR, rt), ctx);
711 	if (dstk) {
712 		emit(ARM_STR_I(ARM_LR, ARM_SP, STACK_VAR(dst_lo)), ctx);
713 		emit(ARM_STR_I(ARM_IP, ARM_SP, STACK_VAR(dst_hi)), ctx);
714 	} else {
715 		emit(ARM_MOV_R(rd, ARM_LR), ctx);
716 		emit(ARM_MOV_R(rm, ARM_IP), ctx);
717 	}
718 }
719 
720 /* dst = dst >> src */
721 static inline void emit_a32_lsr_r64(const u8 dst[], const u8 src[], bool dstk,
722 				     bool sstk, struct jit_ctx *ctx) {
723 	const u8 *tmp = bpf2a32[TMP_REG_1];
724 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
725 	/* Setup Operands */
726 	u8 rt = sstk ? tmp2[1] : src_lo;
727 	u8 rd = dstk ? tmp[1] : dst_lo;
728 	u8 rm = dstk ? tmp[0] : dst_hi;
729 
730 	if (sstk)
731 		emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(src_lo)), ctx);
732 	if (dstk) {
733 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
734 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
735 	}
736 
737 	/* Do LSH operation */
738 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
739 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
740 	emit(ARM_MOV_SR(ARM_LR, rd, SRTYPE_LSR, rt), ctx);
741 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rm, SRTYPE_ASL, ARM_IP), ctx);
742 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rm, SRTYPE_LSR, tmp2[0]), ctx);
743 	emit(ARM_MOV_SR(ARM_IP, rm, SRTYPE_LSR, rt), ctx);
744 	if (dstk) {
745 		emit(ARM_STR_I(ARM_LR, ARM_SP, STACK_VAR(dst_lo)), ctx);
746 		emit(ARM_STR_I(ARM_IP, ARM_SP, STACK_VAR(dst_hi)), ctx);
747 	} else {
748 		emit(ARM_MOV_R(rd, ARM_LR), ctx);
749 		emit(ARM_MOV_R(rm, ARM_IP), ctx);
750 	}
751 }
752 
753 /* dst = dst << val */
754 static inline void emit_a32_lsh_i64(const u8 dst[], bool dstk,
755 				     const u32 val, struct jit_ctx *ctx){
756 	const u8 *tmp = bpf2a32[TMP_REG_1];
757 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
758 	/* Setup operands */
759 	u8 rd = dstk ? tmp[1] : dst_lo;
760 	u8 rm = dstk ? tmp[0] : dst_hi;
761 
762 	if (dstk) {
763 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
764 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
765 	}
766 
767 	/* Do LSH operation */
768 	if (val < 32) {
769 		emit(ARM_MOV_SI(tmp2[0], rm, SRTYPE_ASL, val), ctx);
770 		emit(ARM_ORR_SI(rm, tmp2[0], rd, SRTYPE_LSR, 32 - val), ctx);
771 		emit(ARM_MOV_SI(rd, rd, SRTYPE_ASL, val), ctx);
772 	} else {
773 		if (val == 32)
774 			emit(ARM_MOV_R(rm, rd), ctx);
775 		else
776 			emit(ARM_MOV_SI(rm, rd, SRTYPE_ASL, val - 32), ctx);
777 		emit(ARM_EOR_R(rd, rd, rd), ctx);
778 	}
779 
780 	if (dstk) {
781 		emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
782 		emit(ARM_STR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
783 	}
784 }
785 
786 /* dst = dst >> val */
787 static inline void emit_a32_lsr_i64(const u8 dst[], bool dstk,
788 				    const u32 val, struct jit_ctx *ctx) {
789 	const u8 *tmp = bpf2a32[TMP_REG_1];
790 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
791 	/* Setup operands */
792 	u8 rd = dstk ? tmp[1] : dst_lo;
793 	u8 rm = dstk ? tmp[0] : dst_hi;
794 
795 	if (dstk) {
796 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
797 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
798 	}
799 
800 	/* Do LSR operation */
801 	if (val < 32) {
802 		emit(ARM_MOV_SI(tmp2[1], rd, SRTYPE_LSR, val), ctx);
803 		emit(ARM_ORR_SI(rd, tmp2[1], rm, SRTYPE_ASL, 32 - val), ctx);
804 		emit(ARM_MOV_SI(rm, rm, SRTYPE_LSR, val), ctx);
805 	} else if (val == 32) {
806 		emit(ARM_MOV_R(rd, rm), ctx);
807 		emit(ARM_MOV_I(rm, 0), ctx);
808 	} else {
809 		emit(ARM_MOV_SI(rd, rm, SRTYPE_LSR, val - 32), ctx);
810 		emit(ARM_MOV_I(rm, 0), ctx);
811 	}
812 
813 	if (dstk) {
814 		emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
815 		emit(ARM_STR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
816 	}
817 }
818 
819 /* dst = dst >> val (signed) */
820 static inline void emit_a32_arsh_i64(const u8 dst[], bool dstk,
821 				     const u32 val, struct jit_ctx *ctx){
822 	const u8 *tmp = bpf2a32[TMP_REG_1];
823 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
824 	 /* Setup operands */
825 	u8 rd = dstk ? tmp[1] : dst_lo;
826 	u8 rm = dstk ? tmp[0] : dst_hi;
827 
828 	if (dstk) {
829 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
830 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
831 	}
832 
833 	/* Do ARSH operation */
834 	if (val < 32) {
835 		emit(ARM_MOV_SI(tmp2[1], rd, SRTYPE_LSR, val), ctx);
836 		emit(ARM_ORR_SI(rd, tmp2[1], rm, SRTYPE_ASL, 32 - val), ctx);
837 		emit(ARM_MOV_SI(rm, rm, SRTYPE_ASR, val), ctx);
838 	} else if (val == 32) {
839 		emit(ARM_MOV_R(rd, rm), ctx);
840 		emit(ARM_MOV_SI(rm, rm, SRTYPE_ASR, 31), ctx);
841 	} else {
842 		emit(ARM_MOV_SI(rd, rm, SRTYPE_ASR, val - 32), ctx);
843 		emit(ARM_MOV_SI(rm, rm, SRTYPE_ASR, 31), ctx);
844 	}
845 
846 	if (dstk) {
847 		emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
848 		emit(ARM_STR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
849 	}
850 }
851 
852 static inline void emit_a32_mul_r64(const u8 dst[], const u8 src[], bool dstk,
853 				    bool sstk, struct jit_ctx *ctx) {
854 	const u8 *tmp = bpf2a32[TMP_REG_1];
855 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
856 	/* Setup operands for multiplication */
857 	u8 rd = dstk ? tmp[1] : dst_lo;
858 	u8 rm = dstk ? tmp[0] : dst_hi;
859 	u8 rt = sstk ? tmp2[1] : src_lo;
860 	u8 rn = sstk ? tmp2[0] : src_hi;
861 
862 	if (dstk) {
863 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
864 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
865 	}
866 	if (sstk) {
867 		emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(src_lo)), ctx);
868 		emit(ARM_LDR_I(rn, ARM_SP, STACK_VAR(src_hi)), ctx);
869 	}
870 
871 	/* Do Multiplication */
872 	emit(ARM_MUL(ARM_IP, rd, rn), ctx);
873 	emit(ARM_MUL(ARM_LR, rm, rt), ctx);
874 	emit(ARM_ADD_R(ARM_LR, ARM_IP, ARM_LR), ctx);
875 
876 	emit(ARM_UMULL(ARM_IP, rm, rd, rt), ctx);
877 	emit(ARM_ADD_R(rm, ARM_LR, rm), ctx);
878 	if (dstk) {
879 		emit(ARM_STR_I(ARM_IP, ARM_SP, STACK_VAR(dst_lo)), ctx);
880 		emit(ARM_STR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
881 	} else {
882 		emit(ARM_MOV_R(rd, ARM_IP), ctx);
883 	}
884 }
885 
886 /* *(size *)(dst + off) = src */
887 static inline void emit_str_r(const u8 dst, const u8 src, bool dstk,
888 			      const s32 off, struct jit_ctx *ctx, const u8 sz){
889 	const u8 *tmp = bpf2a32[TMP_REG_1];
890 	u8 rd = dstk ? tmp[1] : dst;
891 
892 	if (dstk)
893 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst)), ctx);
894 	if (off) {
895 		emit_a32_mov_i(tmp[0], off, false, ctx);
896 		emit(ARM_ADD_R(tmp[0], rd, tmp[0]), ctx);
897 		rd = tmp[0];
898 	}
899 	switch (sz) {
900 	case BPF_W:
901 		/* Store a Word */
902 		emit(ARM_STR_I(src, rd, 0), ctx);
903 		break;
904 	case BPF_H:
905 		/* Store a HalfWord */
906 		emit(ARM_STRH_I(src, rd, 0), ctx);
907 		break;
908 	case BPF_B:
909 		/* Store a Byte */
910 		emit(ARM_STRB_I(src, rd, 0), ctx);
911 		break;
912 	}
913 }
914 
915 /* dst = *(size*)(src + off) */
916 static inline void emit_ldx_r(const u8 dst[], const u8 src, bool dstk,
917 			      s32 off, struct jit_ctx *ctx, const u8 sz){
918 	const u8 *tmp = bpf2a32[TMP_REG_1];
919 	const u8 *rd = dstk ? tmp : dst;
920 	u8 rm = src;
921 	s32 off_max;
922 
923 	if (sz == BPF_H)
924 		off_max = 0xff;
925 	else
926 		off_max = 0xfff;
927 
928 	if (off < 0 || off > off_max) {
929 		emit_a32_mov_i(tmp[0], off, false, ctx);
930 		emit(ARM_ADD_R(tmp[0], tmp[0], src), ctx);
931 		rm = tmp[0];
932 		off = 0;
933 	} else if (rd[1] == rm) {
934 		emit(ARM_MOV_R(tmp[0], rm), ctx);
935 		rm = tmp[0];
936 	}
937 	switch (sz) {
938 	case BPF_B:
939 		/* Load a Byte */
940 		emit(ARM_LDRB_I(rd[1], rm, off), ctx);
941 		emit_a32_mov_i(dst[0], 0, dstk, ctx);
942 		break;
943 	case BPF_H:
944 		/* Load a HalfWord */
945 		emit(ARM_LDRH_I(rd[1], rm, off), ctx);
946 		emit_a32_mov_i(dst[0], 0, dstk, ctx);
947 		break;
948 	case BPF_W:
949 		/* Load a Word */
950 		emit(ARM_LDR_I(rd[1], rm, off), ctx);
951 		emit_a32_mov_i(dst[0], 0, dstk, ctx);
952 		break;
953 	case BPF_DW:
954 		/* Load a Double Word */
955 		emit(ARM_LDR_I(rd[1], rm, off), ctx);
956 		emit(ARM_LDR_I(rd[0], rm, off + 4), ctx);
957 		break;
958 	}
959 	if (dstk)
960 		emit(ARM_STR_I(rd[1], ARM_SP, STACK_VAR(dst[1])), ctx);
961 	if (dstk && sz == BPF_DW)
962 		emit(ARM_STR_I(rd[0], ARM_SP, STACK_VAR(dst[0])), ctx);
963 }
964 
965 /* Arithmatic Operation */
966 static inline void emit_ar_r(const u8 rd, const u8 rt, const u8 rm,
967 			     const u8 rn, struct jit_ctx *ctx, u8 op) {
968 	switch (op) {
969 	case BPF_JSET:
970 		emit(ARM_AND_R(ARM_IP, rt, rn), ctx);
971 		emit(ARM_AND_R(ARM_LR, rd, rm), ctx);
972 		emit(ARM_ORRS_R(ARM_IP, ARM_LR, ARM_IP), ctx);
973 		break;
974 	case BPF_JEQ:
975 	case BPF_JNE:
976 	case BPF_JGT:
977 	case BPF_JGE:
978 	case BPF_JLE:
979 	case BPF_JLT:
980 		emit(ARM_CMP_R(rd, rm), ctx);
981 		_emit(ARM_COND_EQ, ARM_CMP_R(rt, rn), ctx);
982 		break;
983 	case BPF_JSLE:
984 	case BPF_JSGT:
985 		emit(ARM_CMP_R(rn, rt), ctx);
986 		emit(ARM_SBCS_R(ARM_IP, rm, rd), ctx);
987 		break;
988 	case BPF_JSLT:
989 	case BPF_JSGE:
990 		emit(ARM_CMP_R(rt, rn), ctx);
991 		emit(ARM_SBCS_R(ARM_IP, rd, rm), ctx);
992 		break;
993 	}
994 }
995 
996 static int out_offset = -1; /* initialized on the first pass of build_body() */
997 static int emit_bpf_tail_call(struct jit_ctx *ctx)
998 {
999 
1000 	/* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
1001 	const u8 *r2 = bpf2a32[BPF_REG_2];
1002 	const u8 *r3 = bpf2a32[BPF_REG_3];
1003 	const u8 *tmp = bpf2a32[TMP_REG_1];
1004 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
1005 	const u8 *tcc = bpf2a32[TCALL_CNT];
1006 	const int idx0 = ctx->idx;
1007 #define cur_offset (ctx->idx - idx0)
1008 #define jmp_offset (out_offset - (cur_offset) - 2)
1009 	u32 off, lo, hi;
1010 
1011 	/* if (index >= array->map.max_entries)
1012 	 *	goto out;
1013 	 */
1014 	off = offsetof(struct bpf_array, map.max_entries);
1015 	/* array->map.max_entries */
1016 	emit_a32_mov_i(tmp[1], off, false, ctx);
1017 	emit(ARM_LDR_I(tmp2[1], ARM_SP, STACK_VAR(r2[1])), ctx);
1018 	emit(ARM_LDR_R(tmp[1], tmp2[1], tmp[1]), ctx);
1019 	/* index is 32-bit for arrays */
1020 	emit(ARM_LDR_I(tmp2[1], ARM_SP, STACK_VAR(r3[1])), ctx);
1021 	/* index >= array->map.max_entries */
1022 	emit(ARM_CMP_R(tmp2[1], tmp[1]), ctx);
1023 	_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
1024 
1025 	/* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
1026 	 *	goto out;
1027 	 * tail_call_cnt++;
1028 	 */
1029 	lo = (u32)MAX_TAIL_CALL_CNT;
1030 	hi = (u32)((u64)MAX_TAIL_CALL_CNT >> 32);
1031 	emit(ARM_LDR_I(tmp[1], ARM_SP, STACK_VAR(tcc[1])), ctx);
1032 	emit(ARM_LDR_I(tmp[0], ARM_SP, STACK_VAR(tcc[0])), ctx);
1033 	emit(ARM_CMP_I(tmp[0], hi), ctx);
1034 	_emit(ARM_COND_EQ, ARM_CMP_I(tmp[1], lo), ctx);
1035 	_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
1036 	emit(ARM_ADDS_I(tmp[1], tmp[1], 1), ctx);
1037 	emit(ARM_ADC_I(tmp[0], tmp[0], 0), ctx);
1038 	emit(ARM_STR_I(tmp[1], ARM_SP, STACK_VAR(tcc[1])), ctx);
1039 	emit(ARM_STR_I(tmp[0], ARM_SP, STACK_VAR(tcc[0])), ctx);
1040 
1041 	/* prog = array->ptrs[index]
1042 	 * if (prog == NULL)
1043 	 *	goto out;
1044 	 */
1045 	off = offsetof(struct bpf_array, ptrs);
1046 	emit_a32_mov_i(tmp[1], off, false, ctx);
1047 	emit(ARM_LDR_I(tmp2[1], ARM_SP, STACK_VAR(r2[1])), ctx);
1048 	emit(ARM_ADD_R(tmp[1], tmp2[1], tmp[1]), ctx);
1049 	emit(ARM_LDR_I(tmp2[1], ARM_SP, STACK_VAR(r3[1])), ctx);
1050 	emit(ARM_MOV_SI(tmp[0], tmp2[1], SRTYPE_ASL, 2), ctx);
1051 	emit(ARM_LDR_R(tmp[1], tmp[1], tmp[0]), ctx);
1052 	emit(ARM_CMP_I(tmp[1], 0), ctx);
1053 	_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1054 
1055 	/* goto *(prog->bpf_func + prologue_size); */
1056 	off = offsetof(struct bpf_prog, bpf_func);
1057 	emit_a32_mov_i(tmp2[1], off, false, ctx);
1058 	emit(ARM_LDR_R(tmp[1], tmp[1], tmp2[1]), ctx);
1059 	emit(ARM_ADD_I(tmp[1], tmp[1], ctx->prologue_bytes), ctx);
1060 	emit_bx_r(tmp[1], ctx);
1061 
1062 	/* out: */
1063 	if (out_offset == -1)
1064 		out_offset = cur_offset;
1065 	if (cur_offset != out_offset) {
1066 		pr_err_once("tail_call out_offset = %d, expected %d!\n",
1067 			    cur_offset, out_offset);
1068 		return -1;
1069 	}
1070 	return 0;
1071 #undef cur_offset
1072 #undef jmp_offset
1073 }
1074 
1075 /* 0xabcd => 0xcdab */
1076 static inline void emit_rev16(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1077 {
1078 #if __LINUX_ARM_ARCH__ < 6
1079 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
1080 
1081 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1082 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 8), ctx);
1083 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1084 	emit(ARM_ORR_SI(rd, tmp2[0], tmp2[1], SRTYPE_LSL, 8), ctx);
1085 #else /* ARMv6+ */
1086 	emit(ARM_REV16(rd, rn), ctx);
1087 #endif
1088 }
1089 
1090 /* 0xabcdefgh => 0xghefcdab */
1091 static inline void emit_rev32(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1092 {
1093 #if __LINUX_ARM_ARCH__ < 6
1094 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
1095 
1096 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1097 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 24), ctx);
1098 	emit(ARM_ORR_SI(ARM_IP, tmp2[0], tmp2[1], SRTYPE_LSL, 24), ctx);
1099 
1100 	emit(ARM_MOV_SI(tmp2[1], rn, SRTYPE_LSR, 8), ctx);
1101 	emit(ARM_AND_I(tmp2[1], tmp2[1], 0xff), ctx);
1102 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 16), ctx);
1103 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1104 	emit(ARM_MOV_SI(tmp2[0], tmp2[0], SRTYPE_LSL, 8), ctx);
1105 	emit(ARM_ORR_SI(tmp2[0], tmp2[0], tmp2[1], SRTYPE_LSL, 16), ctx);
1106 	emit(ARM_ORR_R(rd, ARM_IP, tmp2[0]), ctx);
1107 
1108 #else /* ARMv6+ */
1109 	emit(ARM_REV(rd, rn), ctx);
1110 #endif
1111 }
1112 
1113 // push the scratch stack register on top of the stack
1114 static inline void emit_push_r64(const u8 src[], const u8 shift,
1115 		struct jit_ctx *ctx)
1116 {
1117 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
1118 	u16 reg_set = 0;
1119 
1120 	emit(ARM_LDR_I(tmp2[1], ARM_SP, STACK_VAR(src[1]+shift)), ctx);
1121 	emit(ARM_LDR_I(tmp2[0], ARM_SP, STACK_VAR(src[0]+shift)), ctx);
1122 
1123 	reg_set = (1 << tmp2[1]) | (1 << tmp2[0]);
1124 	emit(ARM_PUSH(reg_set), ctx);
1125 }
1126 
1127 static void build_prologue(struct jit_ctx *ctx)
1128 {
1129 	const u8 r0 = bpf2a32[BPF_REG_0][1];
1130 	const u8 r2 = bpf2a32[BPF_REG_1][1];
1131 	const u8 r3 = bpf2a32[BPF_REG_1][0];
1132 	const u8 r4 = bpf2a32[BPF_REG_6][1];
1133 	const u8 fplo = bpf2a32[BPF_REG_FP][1];
1134 	const u8 fphi = bpf2a32[BPF_REG_FP][0];
1135 	const u8 *tcc = bpf2a32[TCALL_CNT];
1136 
1137 	/* Save callee saved registers. */
1138 #ifdef CONFIG_FRAME_POINTER
1139 	u16 reg_set = CALLEE_PUSH_MASK | 1 << ARM_IP | 1 << ARM_PC;
1140 	emit(ARM_MOV_R(ARM_IP, ARM_SP), ctx);
1141 	emit(ARM_PUSH(reg_set), ctx);
1142 	emit(ARM_SUB_I(ARM_FP, ARM_IP, 4), ctx);
1143 #else
1144 	emit(ARM_PUSH(CALLEE_PUSH_MASK), ctx);
1145 	emit(ARM_MOV_R(ARM_FP, ARM_SP), ctx);
1146 #endif
1147 	/* Save frame pointer for later */
1148 	emit(ARM_SUB_I(ARM_IP, ARM_SP, SCRATCH_SIZE), ctx);
1149 
1150 	ctx->stack_size = imm8m(STACK_SIZE);
1151 
1152 	/* Set up function call stack */
1153 	emit(ARM_SUB_I(ARM_SP, ARM_SP, ctx->stack_size), ctx);
1154 
1155 	/* Set up BPF prog stack base register */
1156 	emit_a32_mov_r(fplo, ARM_IP, true, false, ctx);
1157 	emit_a32_mov_i(fphi, 0, true, ctx);
1158 
1159 	/* mov r4, 0 */
1160 	emit(ARM_MOV_I(r4, 0), ctx);
1161 
1162 	/* Move BPF_CTX to BPF_R1 */
1163 	emit(ARM_MOV_R(r3, r4), ctx);
1164 	emit(ARM_MOV_R(r2, r0), ctx);
1165 	/* Initialize Tail Count */
1166 	emit(ARM_STR_I(r4, ARM_SP, STACK_VAR(tcc[0])), ctx);
1167 	emit(ARM_STR_I(r4, ARM_SP, STACK_VAR(tcc[1])), ctx);
1168 	/* end of prologue */
1169 }
1170 
1171 /* restore callee saved registers. */
1172 static void build_epilogue(struct jit_ctx *ctx)
1173 {
1174 #ifdef CONFIG_FRAME_POINTER
1175 	/* When using frame pointers, some additional registers need to
1176 	 * be loaded. */
1177 	u16 reg_set = CALLEE_POP_MASK | 1 << ARM_SP;
1178 	emit(ARM_SUB_I(ARM_SP, ARM_FP, hweight16(reg_set) * 4), ctx);
1179 	emit(ARM_LDM(ARM_SP, reg_set), ctx);
1180 #else
1181 	/* Restore callee saved registers. */
1182 	emit(ARM_MOV_R(ARM_SP, ARM_FP), ctx);
1183 	emit(ARM_POP(CALLEE_POP_MASK), ctx);
1184 #endif
1185 }
1186 
1187 /*
1188  * Convert an eBPF instruction to native instruction, i.e
1189  * JITs an eBPF instruction.
1190  * Returns :
1191  *	0  - Successfully JITed an 8-byte eBPF instruction
1192  *	>0 - Successfully JITed a 16-byte eBPF instruction
1193  *	<0 - Failed to JIT.
1194  */
1195 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx)
1196 {
1197 	const u8 code = insn->code;
1198 	const u8 *dst = bpf2a32[insn->dst_reg];
1199 	const u8 *src = bpf2a32[insn->src_reg];
1200 	const u8 *tmp = bpf2a32[TMP_REG_1];
1201 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
1202 	const s16 off = insn->off;
1203 	const s32 imm = insn->imm;
1204 	const int i = insn - ctx->prog->insnsi;
1205 	const bool is64 = BPF_CLASS(code) == BPF_ALU64;
1206 	const bool dstk = is_on_stack(insn->dst_reg);
1207 	const bool sstk = is_on_stack(insn->src_reg);
1208 	u8 rd, rt, rm, rn;
1209 	s32 jmp_offset;
1210 
1211 #define check_imm(bits, imm) do {				\
1212 	if ((((imm) > 0) && ((imm) >> (bits))) ||		\
1213 	    (((imm) < 0) && (~(imm) >> (bits)))) {		\
1214 		pr_info("[%2d] imm=%d(0x%x) out of range\n",	\
1215 			i, imm, imm);				\
1216 		return -EINVAL;					\
1217 	}							\
1218 } while (0)
1219 #define check_imm24(imm) check_imm(24, imm)
1220 
1221 	switch (code) {
1222 	/* ALU operations */
1223 
1224 	/* dst = src */
1225 	case BPF_ALU | BPF_MOV | BPF_K:
1226 	case BPF_ALU | BPF_MOV | BPF_X:
1227 	case BPF_ALU64 | BPF_MOV | BPF_K:
1228 	case BPF_ALU64 | BPF_MOV | BPF_X:
1229 		switch (BPF_SRC(code)) {
1230 		case BPF_X:
1231 			emit_a32_mov_r64(is64, dst, src, dstk, sstk, ctx);
1232 			break;
1233 		case BPF_K:
1234 			/* Sign-extend immediate value to destination reg */
1235 			emit_a32_mov_i64(is64, dst, imm, dstk, ctx);
1236 			break;
1237 		}
1238 		break;
1239 	/* dst = dst + src/imm */
1240 	/* dst = dst - src/imm */
1241 	/* dst = dst | src/imm */
1242 	/* dst = dst & src/imm */
1243 	/* dst = dst ^ src/imm */
1244 	/* dst = dst * src/imm */
1245 	/* dst = dst << src */
1246 	/* dst = dst >> src */
1247 	case BPF_ALU | BPF_ADD | BPF_K:
1248 	case BPF_ALU | BPF_ADD | BPF_X:
1249 	case BPF_ALU | BPF_SUB | BPF_K:
1250 	case BPF_ALU | BPF_SUB | BPF_X:
1251 	case BPF_ALU | BPF_OR | BPF_K:
1252 	case BPF_ALU | BPF_OR | BPF_X:
1253 	case BPF_ALU | BPF_AND | BPF_K:
1254 	case BPF_ALU | BPF_AND | BPF_X:
1255 	case BPF_ALU | BPF_XOR | BPF_K:
1256 	case BPF_ALU | BPF_XOR | BPF_X:
1257 	case BPF_ALU | BPF_MUL | BPF_K:
1258 	case BPF_ALU | BPF_MUL | BPF_X:
1259 	case BPF_ALU | BPF_LSH | BPF_X:
1260 	case BPF_ALU | BPF_RSH | BPF_X:
1261 	case BPF_ALU | BPF_ARSH | BPF_K:
1262 	case BPF_ALU | BPF_ARSH | BPF_X:
1263 	case BPF_ALU64 | BPF_ADD | BPF_K:
1264 	case BPF_ALU64 | BPF_ADD | BPF_X:
1265 	case BPF_ALU64 | BPF_SUB | BPF_K:
1266 	case BPF_ALU64 | BPF_SUB | BPF_X:
1267 	case BPF_ALU64 | BPF_OR | BPF_K:
1268 	case BPF_ALU64 | BPF_OR | BPF_X:
1269 	case BPF_ALU64 | BPF_AND | BPF_K:
1270 	case BPF_ALU64 | BPF_AND | BPF_X:
1271 	case BPF_ALU64 | BPF_XOR | BPF_K:
1272 	case BPF_ALU64 | BPF_XOR | BPF_X:
1273 		switch (BPF_SRC(code)) {
1274 		case BPF_X:
1275 			emit_a32_alu_r64(is64, dst, src, dstk, sstk,
1276 					 ctx, BPF_OP(code));
1277 			break;
1278 		case BPF_K:
1279 			/* Move immediate value to the temporary register
1280 			 * and then do the ALU operation on the temporary
1281 			 * register as this will sign-extend the immediate
1282 			 * value into temporary reg and then it would be
1283 			 * safe to do the operation on it.
1284 			 */
1285 			emit_a32_mov_i64(is64, tmp2, imm, false, ctx);
1286 			emit_a32_alu_r64(is64, dst, tmp2, dstk, false,
1287 					 ctx, BPF_OP(code));
1288 			break;
1289 		}
1290 		break;
1291 	/* dst = dst / src(imm) */
1292 	/* dst = dst % src(imm) */
1293 	case BPF_ALU | BPF_DIV | BPF_K:
1294 	case BPF_ALU | BPF_DIV | BPF_X:
1295 	case BPF_ALU | BPF_MOD | BPF_K:
1296 	case BPF_ALU | BPF_MOD | BPF_X:
1297 		rt = src_lo;
1298 		rd = dstk ? tmp2[1] : dst_lo;
1299 		if (dstk)
1300 			emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
1301 		switch (BPF_SRC(code)) {
1302 		case BPF_X:
1303 			rt = sstk ? tmp2[0] : rt;
1304 			if (sstk)
1305 				emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(src_lo)),
1306 				     ctx);
1307 			break;
1308 		case BPF_K:
1309 			rt = tmp2[0];
1310 			emit_a32_mov_i(rt, imm, false, ctx);
1311 			break;
1312 		}
1313 		emit_udivmod(rd, rd, rt, ctx, BPF_OP(code));
1314 		if (dstk)
1315 			emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
1316 		emit_a32_mov_i(dst_hi, 0, dstk, ctx);
1317 		break;
1318 	case BPF_ALU64 | BPF_DIV | BPF_K:
1319 	case BPF_ALU64 | BPF_DIV | BPF_X:
1320 	case BPF_ALU64 | BPF_MOD | BPF_K:
1321 	case BPF_ALU64 | BPF_MOD | BPF_X:
1322 		goto notyet;
1323 	/* dst = dst >> imm */
1324 	/* dst = dst << imm */
1325 	case BPF_ALU | BPF_RSH | BPF_K:
1326 	case BPF_ALU | BPF_LSH | BPF_K:
1327 		if (unlikely(imm > 31))
1328 			return -EINVAL;
1329 		if (imm)
1330 			emit_a32_alu_i(dst_lo, imm, dstk, ctx, BPF_OP(code));
1331 		emit_a32_mov_i(dst_hi, 0, dstk, ctx);
1332 		break;
1333 	/* dst = dst << imm */
1334 	case BPF_ALU64 | BPF_LSH | BPF_K:
1335 		if (unlikely(imm > 63))
1336 			return -EINVAL;
1337 		emit_a32_lsh_i64(dst, dstk, imm, ctx);
1338 		break;
1339 	/* dst = dst >> imm */
1340 	case BPF_ALU64 | BPF_RSH | BPF_K:
1341 		if (unlikely(imm > 63))
1342 			return -EINVAL;
1343 		emit_a32_lsr_i64(dst, dstk, imm, ctx);
1344 		break;
1345 	/* dst = dst << src */
1346 	case BPF_ALU64 | BPF_LSH | BPF_X:
1347 		emit_a32_lsh_r64(dst, src, dstk, sstk, ctx);
1348 		break;
1349 	/* dst = dst >> src */
1350 	case BPF_ALU64 | BPF_RSH | BPF_X:
1351 		emit_a32_lsr_r64(dst, src, dstk, sstk, ctx);
1352 		break;
1353 	/* dst = dst >> src (signed) */
1354 	case BPF_ALU64 | BPF_ARSH | BPF_X:
1355 		emit_a32_arsh_r64(dst, src, dstk, sstk, ctx);
1356 		break;
1357 	/* dst = dst >> imm (signed) */
1358 	case BPF_ALU64 | BPF_ARSH | BPF_K:
1359 		if (unlikely(imm > 63))
1360 			return -EINVAL;
1361 		emit_a32_arsh_i64(dst, dstk, imm, ctx);
1362 		break;
1363 	/* dst = ~dst */
1364 	case BPF_ALU | BPF_NEG:
1365 		emit_a32_alu_i(dst_lo, 0, dstk, ctx, BPF_OP(code));
1366 		emit_a32_mov_i(dst_hi, 0, dstk, ctx);
1367 		break;
1368 	/* dst = ~dst (64 bit) */
1369 	case BPF_ALU64 | BPF_NEG:
1370 		emit_a32_neg64(dst, dstk, ctx);
1371 		break;
1372 	/* dst = dst * src/imm */
1373 	case BPF_ALU64 | BPF_MUL | BPF_X:
1374 	case BPF_ALU64 | BPF_MUL | BPF_K:
1375 		switch (BPF_SRC(code)) {
1376 		case BPF_X:
1377 			emit_a32_mul_r64(dst, src, dstk, sstk, ctx);
1378 			break;
1379 		case BPF_K:
1380 			/* Move immediate value to the temporary register
1381 			 * and then do the multiplication on it as this
1382 			 * will sign-extend the immediate value into temp
1383 			 * reg then it would be safe to do the operation
1384 			 * on it.
1385 			 */
1386 			emit_a32_mov_i64(is64, tmp2, imm, false, ctx);
1387 			emit_a32_mul_r64(dst, tmp2, dstk, false, ctx);
1388 			break;
1389 		}
1390 		break;
1391 	/* dst = htole(dst) */
1392 	/* dst = htobe(dst) */
1393 	case BPF_ALU | BPF_END | BPF_FROM_LE:
1394 	case BPF_ALU | BPF_END | BPF_FROM_BE:
1395 		rd = dstk ? tmp[0] : dst_hi;
1396 		rt = dstk ? tmp[1] : dst_lo;
1397 		if (dstk) {
1398 			emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(dst_lo)), ctx);
1399 			emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_hi)), ctx);
1400 		}
1401 		if (BPF_SRC(code) == BPF_FROM_LE)
1402 			goto emit_bswap_uxt;
1403 		switch (imm) {
1404 		case 16:
1405 			emit_rev16(rt, rt, ctx);
1406 			goto emit_bswap_uxt;
1407 		case 32:
1408 			emit_rev32(rt, rt, ctx);
1409 			goto emit_bswap_uxt;
1410 		case 64:
1411 			emit_rev32(ARM_LR, rt, ctx);
1412 			emit_rev32(rt, rd, ctx);
1413 			emit(ARM_MOV_R(rd, ARM_LR), ctx);
1414 			break;
1415 		}
1416 		goto exit;
1417 emit_bswap_uxt:
1418 		switch (imm) {
1419 		case 16:
1420 			/* zero-extend 16 bits into 64 bits */
1421 #if __LINUX_ARM_ARCH__ < 6
1422 			emit_a32_mov_i(tmp2[1], 0xffff, false, ctx);
1423 			emit(ARM_AND_R(rt, rt, tmp2[1]), ctx);
1424 #else /* ARMv6+ */
1425 			emit(ARM_UXTH(rt, rt), ctx);
1426 #endif
1427 			emit(ARM_EOR_R(rd, rd, rd), ctx);
1428 			break;
1429 		case 32:
1430 			/* zero-extend 32 bits into 64 bits */
1431 			emit(ARM_EOR_R(rd, rd, rd), ctx);
1432 			break;
1433 		case 64:
1434 			/* nop */
1435 			break;
1436 		}
1437 exit:
1438 		if (dstk) {
1439 			emit(ARM_STR_I(rt, ARM_SP, STACK_VAR(dst_lo)), ctx);
1440 			emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_hi)), ctx);
1441 		}
1442 		break;
1443 	/* dst = imm64 */
1444 	case BPF_LD | BPF_IMM | BPF_DW:
1445 	{
1446 		const struct bpf_insn insn1 = insn[1];
1447 		u32 hi, lo = imm;
1448 
1449 		hi = insn1.imm;
1450 		emit_a32_mov_i(dst_lo, lo, dstk, ctx);
1451 		emit_a32_mov_i(dst_hi, hi, dstk, ctx);
1452 
1453 		return 1;
1454 	}
1455 	/* LDX: dst = *(size *)(src + off) */
1456 	case BPF_LDX | BPF_MEM | BPF_W:
1457 	case BPF_LDX | BPF_MEM | BPF_H:
1458 	case BPF_LDX | BPF_MEM | BPF_B:
1459 	case BPF_LDX | BPF_MEM | BPF_DW:
1460 		rn = sstk ? tmp2[1] : src_lo;
1461 		if (sstk)
1462 			emit(ARM_LDR_I(rn, ARM_SP, STACK_VAR(src_lo)), ctx);
1463 		emit_ldx_r(dst, rn, dstk, off, ctx, BPF_SIZE(code));
1464 		break;
1465 	/* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + imm)) */
1466 	case BPF_LD | BPF_ABS | BPF_W:
1467 	case BPF_LD | BPF_ABS | BPF_H:
1468 	case BPF_LD | BPF_ABS | BPF_B:
1469 	/* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + src + imm)) */
1470 	case BPF_LD | BPF_IND | BPF_W:
1471 	case BPF_LD | BPF_IND | BPF_H:
1472 	case BPF_LD | BPF_IND | BPF_B:
1473 	{
1474 		const u8 r4 = bpf2a32[BPF_REG_6][1]; /* r4 = ptr to sk_buff */
1475 		const u8 r0 = bpf2a32[BPF_REG_0][1]; /*r0: struct sk_buff *skb*/
1476 						     /* rtn value */
1477 		const u8 r1 = bpf2a32[BPF_REG_0][0]; /* r1: int k */
1478 		const u8 r2 = bpf2a32[BPF_REG_1][1]; /* r2: unsigned int size */
1479 		const u8 r3 = bpf2a32[BPF_REG_1][0]; /* r3: void *buffer */
1480 		const u8 r6 = bpf2a32[TMP_REG_1][1]; /* r6: void *(*func)(..) */
1481 		int size;
1482 
1483 		/* Setting up first argument */
1484 		emit(ARM_MOV_R(r0, r4), ctx);
1485 
1486 		/* Setting up second argument */
1487 		emit_a32_mov_i(r1, imm, false, ctx);
1488 		if (BPF_MODE(code) == BPF_IND)
1489 			emit_a32_alu_r(r1, src_lo, false, sstk, ctx,
1490 				       false, false, BPF_ADD);
1491 
1492 		/* Setting up third argument */
1493 		switch (BPF_SIZE(code)) {
1494 		case BPF_W:
1495 			size = 4;
1496 			break;
1497 		case BPF_H:
1498 			size = 2;
1499 			break;
1500 		case BPF_B:
1501 			size = 1;
1502 			break;
1503 		default:
1504 			return -EINVAL;
1505 		}
1506 		emit_a32_mov_i(r2, size, false, ctx);
1507 
1508 		/* Setting up fourth argument */
1509 		emit(ARM_ADD_I(r3, ARM_SP, imm8m(SKB_BUFFER)), ctx);
1510 
1511 		/* Setting up function pointer to call */
1512 		emit_a32_mov_i(r6, (unsigned int)bpf_load_pointer, false, ctx);
1513 		emit_blx_r(r6, ctx);
1514 
1515 		emit(ARM_EOR_R(r1, r1, r1), ctx);
1516 		/* Check if return address is NULL or not.
1517 		 * if NULL then jump to epilogue
1518 		 * else continue to load the value from retn address
1519 		 */
1520 		emit(ARM_CMP_I(r0, 0), ctx);
1521 		jmp_offset = epilogue_offset(ctx);
1522 		check_imm24(jmp_offset);
1523 		_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1524 
1525 		/* Load value from the address */
1526 		switch (BPF_SIZE(code)) {
1527 		case BPF_W:
1528 			emit(ARM_LDR_I(r0, r0, 0), ctx);
1529 			emit_rev32(r0, r0, ctx);
1530 			break;
1531 		case BPF_H:
1532 			emit(ARM_LDRH_I(r0, r0, 0), ctx);
1533 			emit_rev16(r0, r0, ctx);
1534 			break;
1535 		case BPF_B:
1536 			emit(ARM_LDRB_I(r0, r0, 0), ctx);
1537 			/* No need to reverse */
1538 			break;
1539 		}
1540 		break;
1541 	}
1542 	/* ST: *(size *)(dst + off) = imm */
1543 	case BPF_ST | BPF_MEM | BPF_W:
1544 	case BPF_ST | BPF_MEM | BPF_H:
1545 	case BPF_ST | BPF_MEM | BPF_B:
1546 	case BPF_ST | BPF_MEM | BPF_DW:
1547 		switch (BPF_SIZE(code)) {
1548 		case BPF_DW:
1549 			/* Sign-extend immediate value into temp reg */
1550 			emit_a32_mov_i64(true, tmp2, imm, false, ctx);
1551 			emit_str_r(dst_lo, tmp2[1], dstk, off, ctx, BPF_W);
1552 			emit_str_r(dst_lo, tmp2[0], dstk, off+4, ctx, BPF_W);
1553 			break;
1554 		case BPF_W:
1555 		case BPF_H:
1556 		case BPF_B:
1557 			emit_a32_mov_i(tmp2[1], imm, false, ctx);
1558 			emit_str_r(dst_lo, tmp2[1], dstk, off, ctx,
1559 				   BPF_SIZE(code));
1560 			break;
1561 		}
1562 		break;
1563 	/* STX XADD: lock *(u32 *)(dst + off) += src */
1564 	case BPF_STX | BPF_XADD | BPF_W:
1565 	/* STX XADD: lock *(u64 *)(dst + off) += src */
1566 	case BPF_STX | BPF_XADD | BPF_DW:
1567 		goto notyet;
1568 	/* STX: *(size *)(dst + off) = src */
1569 	case BPF_STX | BPF_MEM | BPF_W:
1570 	case BPF_STX | BPF_MEM | BPF_H:
1571 	case BPF_STX | BPF_MEM | BPF_B:
1572 	case BPF_STX | BPF_MEM | BPF_DW:
1573 	{
1574 		u8 sz = BPF_SIZE(code);
1575 
1576 		rn = sstk ? tmp2[1] : src_lo;
1577 		rm = sstk ? tmp2[0] : src_hi;
1578 		if (sstk) {
1579 			emit(ARM_LDR_I(rn, ARM_SP, STACK_VAR(src_lo)), ctx);
1580 			emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(src_hi)), ctx);
1581 		}
1582 
1583 		/* Store the value */
1584 		if (BPF_SIZE(code) == BPF_DW) {
1585 			emit_str_r(dst_lo, rn, dstk, off, ctx, BPF_W);
1586 			emit_str_r(dst_lo, rm, dstk, off+4, ctx, BPF_W);
1587 		} else {
1588 			emit_str_r(dst_lo, rn, dstk, off, ctx, sz);
1589 		}
1590 		break;
1591 	}
1592 	/* PC += off if dst == src */
1593 	/* PC += off if dst > src */
1594 	/* PC += off if dst >= src */
1595 	/* PC += off if dst < src */
1596 	/* PC += off if dst <= src */
1597 	/* PC += off if dst != src */
1598 	/* PC += off if dst > src (signed) */
1599 	/* PC += off if dst >= src (signed) */
1600 	/* PC += off if dst < src (signed) */
1601 	/* PC += off if dst <= src (signed) */
1602 	/* PC += off if dst & src */
1603 	case BPF_JMP | BPF_JEQ | BPF_X:
1604 	case BPF_JMP | BPF_JGT | BPF_X:
1605 	case BPF_JMP | BPF_JGE | BPF_X:
1606 	case BPF_JMP | BPF_JNE | BPF_X:
1607 	case BPF_JMP | BPF_JSGT | BPF_X:
1608 	case BPF_JMP | BPF_JSGE | BPF_X:
1609 	case BPF_JMP | BPF_JSET | BPF_X:
1610 	case BPF_JMP | BPF_JLE | BPF_X:
1611 	case BPF_JMP | BPF_JLT | BPF_X:
1612 	case BPF_JMP | BPF_JSLT | BPF_X:
1613 	case BPF_JMP | BPF_JSLE | BPF_X:
1614 		/* Setup source registers */
1615 		rm = sstk ? tmp2[0] : src_hi;
1616 		rn = sstk ? tmp2[1] : src_lo;
1617 		if (sstk) {
1618 			emit(ARM_LDR_I(rn, ARM_SP, STACK_VAR(src_lo)), ctx);
1619 			emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(src_hi)), ctx);
1620 		}
1621 		goto go_jmp;
1622 	/* PC += off if dst == imm */
1623 	/* PC += off if dst > imm */
1624 	/* PC += off if dst >= imm */
1625 	/* PC += off if dst < imm */
1626 	/* PC += off if dst <= imm */
1627 	/* PC += off if dst != imm */
1628 	/* PC += off if dst > imm (signed) */
1629 	/* PC += off if dst >= imm (signed) */
1630 	/* PC += off if dst < imm (signed) */
1631 	/* PC += off if dst <= imm (signed) */
1632 	/* PC += off if dst & imm */
1633 	case BPF_JMP | BPF_JEQ | BPF_K:
1634 	case BPF_JMP | BPF_JGT | BPF_K:
1635 	case BPF_JMP | BPF_JGE | BPF_K:
1636 	case BPF_JMP | BPF_JNE | BPF_K:
1637 	case BPF_JMP | BPF_JSGT | BPF_K:
1638 	case BPF_JMP | BPF_JSGE | BPF_K:
1639 	case BPF_JMP | BPF_JSET | BPF_K:
1640 	case BPF_JMP | BPF_JLT | BPF_K:
1641 	case BPF_JMP | BPF_JLE | BPF_K:
1642 	case BPF_JMP | BPF_JSLT | BPF_K:
1643 	case BPF_JMP | BPF_JSLE | BPF_K:
1644 		if (off == 0)
1645 			break;
1646 		rm = tmp2[0];
1647 		rn = tmp2[1];
1648 		/* Sign-extend immediate value */
1649 		emit_a32_mov_i64(true, tmp2, imm, false, ctx);
1650 go_jmp:
1651 		/* Setup destination register */
1652 		rd = dstk ? tmp[0] : dst_hi;
1653 		rt = dstk ? tmp[1] : dst_lo;
1654 		if (dstk) {
1655 			emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(dst_lo)), ctx);
1656 			emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_hi)), ctx);
1657 		}
1658 
1659 		/* Check for the condition */
1660 		emit_ar_r(rd, rt, rm, rn, ctx, BPF_OP(code));
1661 
1662 		/* Setup JUMP instruction */
1663 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1664 		switch (BPF_OP(code)) {
1665 		case BPF_JNE:
1666 		case BPF_JSET:
1667 			_emit(ARM_COND_NE, ARM_B(jmp_offset), ctx);
1668 			break;
1669 		case BPF_JEQ:
1670 			_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1671 			break;
1672 		case BPF_JGT:
1673 			_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
1674 			break;
1675 		case BPF_JGE:
1676 			_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
1677 			break;
1678 		case BPF_JSGT:
1679 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1680 			break;
1681 		case BPF_JSGE:
1682 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1683 			break;
1684 		case BPF_JLE:
1685 			_emit(ARM_COND_LS, ARM_B(jmp_offset), ctx);
1686 			break;
1687 		case BPF_JLT:
1688 			_emit(ARM_COND_CC, ARM_B(jmp_offset), ctx);
1689 			break;
1690 		case BPF_JSLT:
1691 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1692 			break;
1693 		case BPF_JSLE:
1694 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1695 			break;
1696 		}
1697 		break;
1698 	/* JMP OFF */
1699 	case BPF_JMP | BPF_JA:
1700 	{
1701 		if (off == 0)
1702 			break;
1703 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1704 		check_imm24(jmp_offset);
1705 		emit(ARM_B(jmp_offset), ctx);
1706 		break;
1707 	}
1708 	/* tail call */
1709 	case BPF_JMP | BPF_TAIL_CALL:
1710 		if (emit_bpf_tail_call(ctx))
1711 			return -EFAULT;
1712 		break;
1713 	/* function call */
1714 	case BPF_JMP | BPF_CALL:
1715 	{
1716 		const u8 *r0 = bpf2a32[BPF_REG_0];
1717 		const u8 *r1 = bpf2a32[BPF_REG_1];
1718 		const u8 *r2 = bpf2a32[BPF_REG_2];
1719 		const u8 *r3 = bpf2a32[BPF_REG_3];
1720 		const u8 *r4 = bpf2a32[BPF_REG_4];
1721 		const u8 *r5 = bpf2a32[BPF_REG_5];
1722 		const u32 func = (u32)__bpf_call_base + (u32)imm;
1723 
1724 		emit_a32_mov_r64(true, r0, r1, false, false, ctx);
1725 		emit_a32_mov_r64(true, r1, r2, false, true, ctx);
1726 		emit_push_r64(r5, 0, ctx);
1727 		emit_push_r64(r4, 8, ctx);
1728 		emit_push_r64(r3, 16, ctx);
1729 
1730 		emit_a32_mov_i(tmp[1], func, false, ctx);
1731 		emit_blx_r(tmp[1], ctx);
1732 
1733 		emit(ARM_ADD_I(ARM_SP, ARM_SP, imm8m(24)), ctx); // callee clean
1734 		break;
1735 	}
1736 	/* function return */
1737 	case BPF_JMP | BPF_EXIT:
1738 		/* Optimization: when last instruction is EXIT
1739 		 * simply fallthrough to epilogue.
1740 		 */
1741 		if (i == ctx->prog->len - 1)
1742 			break;
1743 		jmp_offset = epilogue_offset(ctx);
1744 		check_imm24(jmp_offset);
1745 		emit(ARM_B(jmp_offset), ctx);
1746 		break;
1747 notyet:
1748 		pr_info_once("*** NOT YET: opcode %02x ***\n", code);
1749 		return -EFAULT;
1750 	default:
1751 		pr_err_once("unknown opcode %02x\n", code);
1752 		return -EINVAL;
1753 	}
1754 
1755 	if (ctx->flags & FLAG_IMM_OVERFLOW)
1756 		/*
1757 		 * this instruction generated an overflow when
1758 		 * trying to access the literal pool, so
1759 		 * delegate this filter to the kernel interpreter.
1760 		 */
1761 		return -1;
1762 	return 0;
1763 }
1764 
1765 static int build_body(struct jit_ctx *ctx)
1766 {
1767 	const struct bpf_prog *prog = ctx->prog;
1768 	unsigned int i;
1769 
1770 	for (i = 0; i < prog->len; i++) {
1771 		const struct bpf_insn *insn = &(prog->insnsi[i]);
1772 		int ret;
1773 
1774 		ret = build_insn(insn, ctx);
1775 
1776 		/* It's used with loading the 64 bit immediate value. */
1777 		if (ret > 0) {
1778 			i++;
1779 			if (ctx->target == NULL)
1780 				ctx->offsets[i] = ctx->idx;
1781 			continue;
1782 		}
1783 
1784 		if (ctx->target == NULL)
1785 			ctx->offsets[i] = ctx->idx;
1786 
1787 		/* If unsuccesfull, return with error code */
1788 		if (ret)
1789 			return ret;
1790 	}
1791 	return 0;
1792 }
1793 
1794 static int validate_code(struct jit_ctx *ctx)
1795 {
1796 	int i;
1797 
1798 	for (i = 0; i < ctx->idx; i++) {
1799 		if (ctx->target[i] == __opcode_to_mem_arm(ARM_INST_UDF))
1800 			return -1;
1801 	}
1802 
1803 	return 0;
1804 }
1805 
1806 void bpf_jit_compile(struct bpf_prog *prog)
1807 {
1808 	/* Nothing to do here. We support Internal BPF. */
1809 }
1810 
1811 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1812 {
1813 	struct bpf_prog *tmp, *orig_prog = prog;
1814 	struct bpf_binary_header *header;
1815 	bool tmp_blinded = false;
1816 	struct jit_ctx ctx;
1817 	unsigned int tmp_idx;
1818 	unsigned int image_size;
1819 	u8 *image_ptr;
1820 
1821 	/* If BPF JIT was not enabled then we must fall back to
1822 	 * the interpreter.
1823 	 */
1824 	if (!bpf_jit_enable)
1825 		return orig_prog;
1826 
1827 	/* If constant blinding was enabled and we failed during blinding
1828 	 * then we must fall back to the interpreter. Otherwise, we save
1829 	 * the new JITed code.
1830 	 */
1831 	tmp = bpf_jit_blind_constants(prog);
1832 
1833 	if (IS_ERR(tmp))
1834 		return orig_prog;
1835 	if (tmp != prog) {
1836 		tmp_blinded = true;
1837 		prog = tmp;
1838 	}
1839 
1840 	memset(&ctx, 0, sizeof(ctx));
1841 	ctx.prog = prog;
1842 
1843 	/* Not able to allocate memory for offsets[] , then
1844 	 * we must fall back to the interpreter
1845 	 */
1846 	ctx.offsets = kcalloc(prog->len, sizeof(int), GFP_KERNEL);
1847 	if (ctx.offsets == NULL) {
1848 		prog = orig_prog;
1849 		goto out;
1850 	}
1851 
1852 	/* 1) fake pass to find in the length of the JITed code,
1853 	 * to compute ctx->offsets and other context variables
1854 	 * needed to compute final JITed code.
1855 	 * Also, calculate random starting pointer/start of JITed code
1856 	 * which is prefixed by random number of fault instructions.
1857 	 *
1858 	 * If the first pass fails then there is no chance of it
1859 	 * being successful in the second pass, so just fall back
1860 	 * to the interpreter.
1861 	 */
1862 	if (build_body(&ctx)) {
1863 		prog = orig_prog;
1864 		goto out_off;
1865 	}
1866 
1867 	tmp_idx = ctx.idx;
1868 	build_prologue(&ctx);
1869 	ctx.prologue_bytes = (ctx.idx - tmp_idx) * 4;
1870 
1871 	ctx.epilogue_offset = ctx.idx;
1872 
1873 #if __LINUX_ARM_ARCH__ < 7
1874 	tmp_idx = ctx.idx;
1875 	build_epilogue(&ctx);
1876 	ctx.epilogue_bytes = (ctx.idx - tmp_idx) * 4;
1877 
1878 	ctx.idx += ctx.imm_count;
1879 	if (ctx.imm_count) {
1880 		ctx.imms = kcalloc(ctx.imm_count, sizeof(u32), GFP_KERNEL);
1881 		if (ctx.imms == NULL) {
1882 			prog = orig_prog;
1883 			goto out_off;
1884 		}
1885 	}
1886 #else
1887 	/* there's nothing about the epilogue on ARMv7 */
1888 	build_epilogue(&ctx);
1889 #endif
1890 	/* Now we can get the actual image size of the JITed arm code.
1891 	 * Currently, we are not considering the THUMB-2 instructions
1892 	 * for jit, although it can decrease the size of the image.
1893 	 *
1894 	 * As each arm instruction is of length 32bit, we are translating
1895 	 * number of JITed intructions into the size required to store these
1896 	 * JITed code.
1897 	 */
1898 	image_size = sizeof(u32) * ctx.idx;
1899 
1900 	/* Now we know the size of the structure to make */
1901 	header = bpf_jit_binary_alloc(image_size, &image_ptr,
1902 				      sizeof(u32), jit_fill_hole);
1903 	/* Not able to allocate memory for the structure then
1904 	 * we must fall back to the interpretation
1905 	 */
1906 	if (header == NULL) {
1907 		prog = orig_prog;
1908 		goto out_imms;
1909 	}
1910 
1911 	/* 2.) Actual pass to generate final JIT code */
1912 	ctx.target = (u32 *) image_ptr;
1913 	ctx.idx = 0;
1914 
1915 	build_prologue(&ctx);
1916 
1917 	/* If building the body of the JITed code fails somehow,
1918 	 * we fall back to the interpretation.
1919 	 */
1920 	if (build_body(&ctx) < 0) {
1921 		image_ptr = NULL;
1922 		bpf_jit_binary_free(header);
1923 		prog = orig_prog;
1924 		goto out_imms;
1925 	}
1926 	build_epilogue(&ctx);
1927 
1928 	/* 3.) Extra pass to validate JITed Code */
1929 	if (validate_code(&ctx)) {
1930 		image_ptr = NULL;
1931 		bpf_jit_binary_free(header);
1932 		prog = orig_prog;
1933 		goto out_imms;
1934 	}
1935 	flush_icache_range((u32)header, (u32)(ctx.target + ctx.idx));
1936 
1937 	if (bpf_jit_enable > 1)
1938 		/* there are 2 passes here */
1939 		bpf_jit_dump(prog->len, image_size, 2, ctx.target);
1940 
1941 	set_memory_ro((unsigned long)header, header->pages);
1942 	prog->bpf_func = (void *)ctx.target;
1943 	prog->jited = 1;
1944 	prog->jited_len = image_size;
1945 
1946 out_imms:
1947 #if __LINUX_ARM_ARCH__ < 7
1948 	if (ctx.imm_count)
1949 		kfree(ctx.imms);
1950 #endif
1951 out_off:
1952 	kfree(ctx.offsets);
1953 out:
1954 	if (tmp_blinded)
1955 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
1956 					   tmp : orig_prog);
1957 	return prog;
1958 }
1959 
1960