xref: /openbmc/linux/arch/arm/net/bpf_jit_32.c (revision bbecb07f)
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 #define STACK_OFFSET(k)	(k)
31 #define TMP_REG_1	(MAX_BPF_JIT_REG + 0)	/* TEMP Register 1 */
32 #define TMP_REG_2	(MAX_BPF_JIT_REG + 1)	/* TEMP Register 2 */
33 #define TCALL_CNT	(MAX_BPF_JIT_REG + 2)	/* Tail Call Count */
34 
35 /* Flags used for JIT optimization */
36 #define SEEN_CALL	(1 << 0)
37 
38 #define FLAG_IMM_OVERFLOW	(1 << 0)
39 
40 /*
41  * Map eBPF registers to ARM 32bit registers or stack scratch space.
42  *
43  * 1. First argument is passed using the arm 32bit registers and rest of the
44  * arguments are passed on stack scratch space.
45  * 2. First callee-saved arugument is mapped to arm 32 bit registers and rest
46  * arguments are mapped to scratch space on stack.
47  * 3. We need two 64 bit temp registers to do complex operations on eBPF
48  * registers.
49  *
50  * As the eBPF registers are all 64 bit registers and arm has only 32 bit
51  * registers, we have to map each eBPF registers with two arm 32 bit regs or
52  * scratch memory space and we have to build eBPF 64 bit register from those.
53  *
54  */
55 static const u8 bpf2a32[][2] = {
56 	/* return value from in-kernel function, and exit value from eBPF */
57 	[BPF_REG_0] = {ARM_R1, ARM_R0},
58 	/* arguments from eBPF program to in-kernel function */
59 	[BPF_REG_1] = {ARM_R3, ARM_R2},
60 	/* Stored on stack scratch space */
61 	[BPF_REG_2] = {STACK_OFFSET(0), STACK_OFFSET(4)},
62 	[BPF_REG_3] = {STACK_OFFSET(8), STACK_OFFSET(12)},
63 	[BPF_REG_4] = {STACK_OFFSET(16), STACK_OFFSET(20)},
64 	[BPF_REG_5] = {STACK_OFFSET(24), STACK_OFFSET(28)},
65 	/* callee saved registers that in-kernel function will preserve */
66 	[BPF_REG_6] = {ARM_R5, ARM_R4},
67 	/* Stored on stack scratch space */
68 	[BPF_REG_7] = {STACK_OFFSET(32), STACK_OFFSET(36)},
69 	[BPF_REG_8] = {STACK_OFFSET(40), STACK_OFFSET(44)},
70 	[BPF_REG_9] = {STACK_OFFSET(48), STACK_OFFSET(52)},
71 	/* Read only Frame Pointer to access Stack */
72 	[BPF_REG_FP] = {STACK_OFFSET(56), STACK_OFFSET(60)},
73 	/* Temporary Register for internal BPF JIT, can be used
74 	 * for constant blindings and others.
75 	 */
76 	[TMP_REG_1] = {ARM_R7, ARM_R6},
77 	[TMP_REG_2] = {ARM_R10, ARM_R8},
78 	/* Tail call count. Stored on stack scratch space. */
79 	[TCALL_CNT] = {STACK_OFFSET(64), STACK_OFFSET(68)},
80 	/* temporary register for blinding constants.
81 	 * Stored on stack scratch space.
82 	 */
83 	[BPF_REG_AX] = {STACK_OFFSET(72), STACK_OFFSET(76)},
84 };
85 
86 #define	dst_lo	dst[1]
87 #define dst_hi	dst[0]
88 #define src_lo	src[1]
89 #define src_hi	src[0]
90 
91 /*
92  * JIT Context:
93  *
94  * prog			:	bpf_prog
95  * idx			:	index of current last JITed instruction.
96  * prologue_bytes	:	bytes used in prologue.
97  * epilogue_offset	:	offset of epilogue starting.
98  * seen			:	bit mask used for JIT optimization.
99  * offsets		:	array of eBPF instruction offsets in
100  *				JITed code.
101  * target		:	final JITed code.
102  * epilogue_bytes	:	no of bytes used in epilogue.
103  * imm_count		:	no of immediate counts used for global
104  *				variables.
105  * imms			:	array of global variable addresses.
106  */
107 
108 struct jit_ctx {
109 	const struct bpf_prog *prog;
110 	unsigned int idx;
111 	unsigned int prologue_bytes;
112 	unsigned int epilogue_offset;
113 	u32 seen;
114 	u32 flags;
115 	u32 *offsets;
116 	u32 *target;
117 	u32 stack_size;
118 #if __LINUX_ARM_ARCH__ < 7
119 	u16 epilogue_bytes;
120 	u16 imm_count;
121 	u32 *imms;
122 #endif
123 };
124 
125 /*
126  * Wrappers which handle both OABI and EABI and assures Thumb2 interworking
127  * (where the assembly routines like __aeabi_uidiv could cause problems).
128  */
129 static u32 jit_udiv32(u32 dividend, u32 divisor)
130 {
131 	return dividend / divisor;
132 }
133 
134 static u32 jit_mod32(u32 dividend, u32 divisor)
135 {
136 	return dividend % divisor;
137 }
138 
139 static inline void _emit(int cond, u32 inst, struct jit_ctx *ctx)
140 {
141 	inst |= (cond << 28);
142 	inst = __opcode_to_mem_arm(inst);
143 
144 	if (ctx->target != NULL)
145 		ctx->target[ctx->idx] = inst;
146 
147 	ctx->idx++;
148 }
149 
150 /*
151  * Emit an instruction that will be executed unconditionally.
152  */
153 static inline void emit(u32 inst, struct jit_ctx *ctx)
154 {
155 	_emit(ARM_COND_AL, inst, ctx);
156 }
157 
158 /*
159  * Checks if immediate value can be converted to imm12(12 bits) value.
160  */
161 static int16_t imm8m(u32 x)
162 {
163 	u32 rot;
164 
165 	for (rot = 0; rot < 16; rot++)
166 		if ((x & ~ror32(0xff, 2 * rot)) == 0)
167 			return rol32(x, 2 * rot) | (rot << 8);
168 	return -1;
169 }
170 
171 /*
172  * Initializes the JIT space with undefined instructions.
173  */
174 static void jit_fill_hole(void *area, unsigned int size)
175 {
176 	u32 *ptr;
177 	/* We are guaranteed to have aligned memory. */
178 	for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
179 		*ptr++ = __opcode_to_mem_arm(ARM_INST_UDF);
180 }
181 
182 /* Stack must be multiples of 16 Bytes */
183 #define STACK_ALIGN(sz) (((sz) + 3) & ~3)
184 
185 /* Stack space for BPF_REG_2, BPF_REG_3, BPF_REG_4,
186  * BPF_REG_5, BPF_REG_7, BPF_REG_8, BPF_REG_9,
187  * BPF_REG_FP and Tail call counts.
188  */
189 #define SCRATCH_SIZE 80
190 
191 /* total stack size used in JITed code */
192 #define _STACK_SIZE \
193 	(ctx->prog->aux->stack_depth + \
194 	 + SCRATCH_SIZE + \
195 	 + 4 /* extra for skb_copy_bits buffer */)
196 
197 #define STACK_SIZE STACK_ALIGN(_STACK_SIZE)
198 
199 /* Get the offset of eBPF REGISTERs stored on scratch space. */
200 #define STACK_VAR(off) (STACK_SIZE-off-4)
201 
202 /* Offset of skb_copy_bits buffer */
203 #define SKB_BUFFER STACK_VAR(SCRATCH_SIZE)
204 
205 #if __LINUX_ARM_ARCH__ < 7
206 
207 static u16 imm_offset(u32 k, struct jit_ctx *ctx)
208 {
209 	unsigned int i = 0, offset;
210 	u16 imm;
211 
212 	/* on the "fake" run we just count them (duplicates included) */
213 	if (ctx->target == NULL) {
214 		ctx->imm_count++;
215 		return 0;
216 	}
217 
218 	while ((i < ctx->imm_count) && ctx->imms[i]) {
219 		if (ctx->imms[i] == k)
220 			break;
221 		i++;
222 	}
223 
224 	if (ctx->imms[i] == 0)
225 		ctx->imms[i] = k;
226 
227 	/* constants go just after the epilogue */
228 	offset =  ctx->offsets[ctx->prog->len - 1] * 4;
229 	offset += ctx->prologue_bytes;
230 	offset += ctx->epilogue_bytes;
231 	offset += i * 4;
232 
233 	ctx->target[offset / 4] = k;
234 
235 	/* PC in ARM mode == address of the instruction + 8 */
236 	imm = offset - (8 + ctx->idx * 4);
237 
238 	if (imm & ~0xfff) {
239 		/*
240 		 * literal pool is too far, signal it into flags. we
241 		 * can only detect it on the second pass unfortunately.
242 		 */
243 		ctx->flags |= FLAG_IMM_OVERFLOW;
244 		return 0;
245 	}
246 
247 	return imm;
248 }
249 
250 #endif /* __LINUX_ARM_ARCH__ */
251 
252 static inline int bpf2a32_offset(int bpf_to, int bpf_from,
253 				 const struct jit_ctx *ctx) {
254 	int to, from;
255 
256 	if (ctx->target == NULL)
257 		return 0;
258 	to = ctx->offsets[bpf_to];
259 	from = ctx->offsets[bpf_from];
260 
261 	return to - from - 1;
262 }
263 
264 /*
265  * Move an immediate that's not an imm8m to a core register.
266  */
267 static inline void emit_mov_i_no8m(const u8 rd, u32 val, struct jit_ctx *ctx)
268 {
269 #if __LINUX_ARM_ARCH__ < 7
270 	emit(ARM_LDR_I(rd, ARM_PC, imm_offset(val, ctx)), ctx);
271 #else
272 	emit(ARM_MOVW(rd, val & 0xffff), ctx);
273 	if (val > 0xffff)
274 		emit(ARM_MOVT(rd, val >> 16), ctx);
275 #endif
276 }
277 
278 static inline void emit_mov_i(const u8 rd, u32 val, struct jit_ctx *ctx)
279 {
280 	int imm12 = imm8m(val);
281 
282 	if (imm12 >= 0)
283 		emit(ARM_MOV_I(rd, imm12), ctx);
284 	else
285 		emit_mov_i_no8m(rd, val, ctx);
286 }
287 
288 static inline void emit_blx_r(u8 tgt_reg, struct jit_ctx *ctx)
289 {
290 	ctx->seen |= SEEN_CALL;
291 #if __LINUX_ARM_ARCH__ < 5
292 	emit(ARM_MOV_R(ARM_LR, ARM_PC), ctx);
293 
294 	if (elf_hwcap & HWCAP_THUMB)
295 		emit(ARM_BX(tgt_reg), ctx);
296 	else
297 		emit(ARM_MOV_R(ARM_PC, tgt_reg), ctx);
298 #else
299 	emit(ARM_BLX_R(tgt_reg), ctx);
300 #endif
301 }
302 
303 static inline int epilogue_offset(const struct jit_ctx *ctx)
304 {
305 	int to, from;
306 	/* No need for 1st dummy run */
307 	if (ctx->target == NULL)
308 		return 0;
309 	to = ctx->epilogue_offset;
310 	from = ctx->idx;
311 
312 	return to - from - 2;
313 }
314 
315 static inline void emit_udivmod(u8 rd, u8 rm, u8 rn, struct jit_ctx *ctx, u8 op)
316 {
317 	const u8 *tmp = bpf2a32[TMP_REG_1];
318 	s32 jmp_offset;
319 
320 	/* checks if divisor is zero or not. If it is, then
321 	 * exit directly.
322 	 */
323 	emit(ARM_CMP_I(rn, 0), ctx);
324 	_emit(ARM_COND_EQ, ARM_MOV_I(ARM_R0, 0), ctx);
325 	jmp_offset = epilogue_offset(ctx);
326 	_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
327 #if __LINUX_ARM_ARCH__ == 7
328 	if (elf_hwcap & HWCAP_IDIVA) {
329 		if (op == BPF_DIV)
330 			emit(ARM_UDIV(rd, rm, rn), ctx);
331 		else {
332 			emit(ARM_UDIV(ARM_IP, rm, rn), ctx);
333 			emit(ARM_MLS(rd, rn, ARM_IP, rm), ctx);
334 		}
335 		return;
336 	}
337 #endif
338 
339 	/*
340 	 * For BPF_ALU | BPF_DIV | BPF_K instructions
341 	 * As ARM_R1 and ARM_R0 contains 1st argument of bpf
342 	 * function, we need to save it on caller side to save
343 	 * it from getting destroyed within callee.
344 	 * After the return from the callee, we restore ARM_R0
345 	 * ARM_R1.
346 	 */
347 	if (rn != ARM_R1) {
348 		emit(ARM_MOV_R(tmp[0], ARM_R1), ctx);
349 		emit(ARM_MOV_R(ARM_R1, rn), ctx);
350 	}
351 	if (rm != ARM_R0) {
352 		emit(ARM_MOV_R(tmp[1], ARM_R0), ctx);
353 		emit(ARM_MOV_R(ARM_R0, rm), ctx);
354 	}
355 
356 	/* Call appropriate function */
357 	ctx->seen |= SEEN_CALL;
358 	emit_mov_i(ARM_IP, op == BPF_DIV ?
359 		   (u32)jit_udiv32 : (u32)jit_mod32, ctx);
360 	emit_blx_r(ARM_IP, ctx);
361 
362 	/* Save return value */
363 	if (rd != ARM_R0)
364 		emit(ARM_MOV_R(rd, ARM_R0), ctx);
365 
366 	/* Restore ARM_R0 and ARM_R1 */
367 	if (rn != ARM_R1)
368 		emit(ARM_MOV_R(ARM_R1, tmp[0]), ctx);
369 	if (rm != ARM_R0)
370 		emit(ARM_MOV_R(ARM_R0, tmp[1]), ctx);
371 }
372 
373 /* Checks whether BPF register is on scratch stack space or not. */
374 static inline bool is_on_stack(u8 bpf_reg)
375 {
376 	static u8 stack_regs[] = {BPF_REG_AX, BPF_REG_3, BPF_REG_4, BPF_REG_5,
377 				BPF_REG_7, BPF_REG_8, BPF_REG_9, TCALL_CNT,
378 				BPF_REG_2, BPF_REG_FP};
379 	int i, reg_len = sizeof(stack_regs);
380 
381 	for (i = 0 ; i < reg_len ; i++) {
382 		if (bpf_reg == stack_regs[i])
383 			return true;
384 	}
385 	return false;
386 }
387 
388 static inline void emit_a32_mov_i(const u8 dst, const u32 val,
389 				  bool dstk, struct jit_ctx *ctx)
390 {
391 	const u8 *tmp = bpf2a32[TMP_REG_1];
392 
393 	if (dstk) {
394 		emit_mov_i(tmp[1], val, ctx);
395 		emit(ARM_STR_I(tmp[1], ARM_SP, STACK_VAR(dst)), ctx);
396 	} else {
397 		emit_mov_i(dst, val, ctx);
398 	}
399 }
400 
401 /* Sign extended move */
402 static inline void emit_a32_mov_i64(const bool is64, const u8 dst[],
403 				  const u32 val, bool dstk,
404 				  struct jit_ctx *ctx) {
405 	u32 hi = 0;
406 
407 	if (is64 && (val & (1<<31)))
408 		hi = (u32)~0;
409 	emit_a32_mov_i(dst_lo, val, dstk, ctx);
410 	emit_a32_mov_i(dst_hi, hi, dstk, ctx);
411 }
412 
413 static inline void emit_a32_add_r(const u8 dst, const u8 src,
414 			      const bool is64, const bool hi,
415 			      struct jit_ctx *ctx) {
416 	/* 64 bit :
417 	 *	adds dst_lo, dst_lo, src_lo
418 	 *	adc dst_hi, dst_hi, src_hi
419 	 * 32 bit :
420 	 *	add dst_lo, dst_lo, src_lo
421 	 */
422 	if (!hi && is64)
423 		emit(ARM_ADDS_R(dst, dst, src), ctx);
424 	else if (hi && is64)
425 		emit(ARM_ADC_R(dst, dst, src), ctx);
426 	else
427 		emit(ARM_ADD_R(dst, dst, src), ctx);
428 }
429 
430 static inline void emit_a32_sub_r(const u8 dst, const u8 src,
431 				  const bool is64, const bool hi,
432 				  struct jit_ctx *ctx) {
433 	/* 64 bit :
434 	 *	subs dst_lo, dst_lo, src_lo
435 	 *	sbc dst_hi, dst_hi, src_hi
436 	 * 32 bit :
437 	 *	sub dst_lo, dst_lo, src_lo
438 	 */
439 	if (!hi && is64)
440 		emit(ARM_SUBS_R(dst, dst, src), ctx);
441 	else if (hi && is64)
442 		emit(ARM_SBC_R(dst, dst, src), ctx);
443 	else
444 		emit(ARM_SUB_R(dst, dst, src), ctx);
445 }
446 
447 static inline void emit_alu_r(const u8 dst, const u8 src, const bool is64,
448 			      const bool hi, const u8 op, struct jit_ctx *ctx){
449 	switch (BPF_OP(op)) {
450 	/* dst = dst + src */
451 	case BPF_ADD:
452 		emit_a32_add_r(dst, src, is64, hi, ctx);
453 		break;
454 	/* dst = dst - src */
455 	case BPF_SUB:
456 		emit_a32_sub_r(dst, src, is64, hi, ctx);
457 		break;
458 	/* dst = dst | src */
459 	case BPF_OR:
460 		emit(ARM_ORR_R(dst, dst, src), ctx);
461 		break;
462 	/* dst = dst & src */
463 	case BPF_AND:
464 		emit(ARM_AND_R(dst, dst, src), ctx);
465 		break;
466 	/* dst = dst ^ src */
467 	case BPF_XOR:
468 		emit(ARM_EOR_R(dst, dst, src), ctx);
469 		break;
470 	/* dst = dst * src */
471 	case BPF_MUL:
472 		emit(ARM_MUL(dst, dst, src), ctx);
473 		break;
474 	/* dst = dst << src */
475 	case BPF_LSH:
476 		emit(ARM_LSL_R(dst, dst, src), ctx);
477 		break;
478 	/* dst = dst >> src */
479 	case BPF_RSH:
480 		emit(ARM_LSR_R(dst, dst, src), ctx);
481 		break;
482 	/* dst = dst >> src (signed)*/
483 	case BPF_ARSH:
484 		emit(ARM_MOV_SR(dst, dst, SRTYPE_ASR, src), ctx);
485 		break;
486 	}
487 }
488 
489 /* ALU operation (32 bit)
490  * dst = dst (op) src
491  */
492 static inline void emit_a32_alu_r(const u8 dst, const u8 src,
493 				  bool dstk, bool sstk,
494 				  struct jit_ctx *ctx, const bool is64,
495 				  const bool hi, const u8 op) {
496 	const u8 *tmp = bpf2a32[TMP_REG_1];
497 	u8 rn = sstk ? tmp[1] : src;
498 
499 	if (sstk)
500 		emit(ARM_LDR_I(rn, ARM_SP, STACK_VAR(src)), ctx);
501 
502 	/* ALU operation */
503 	if (dstk) {
504 		emit(ARM_LDR_I(tmp[0], ARM_SP, STACK_VAR(dst)), ctx);
505 		emit_alu_r(tmp[0], rn, is64, hi, op, ctx);
506 		emit(ARM_STR_I(tmp[0], ARM_SP, STACK_VAR(dst)), ctx);
507 	} else {
508 		emit_alu_r(dst, rn, is64, hi, op, ctx);
509 	}
510 }
511 
512 /* ALU operation (64 bit) */
513 static inline void emit_a32_alu_r64(const bool is64, const u8 dst[],
514 				  const u8 src[], bool dstk,
515 				  bool sstk, struct jit_ctx *ctx,
516 				  const u8 op) {
517 	emit_a32_alu_r(dst_lo, src_lo, dstk, sstk, ctx, is64, false, op);
518 	if (is64)
519 		emit_a32_alu_r(dst_hi, src_hi, dstk, sstk, ctx, is64, true, op);
520 	else
521 		emit_a32_mov_i(dst_hi, 0, dstk, ctx);
522 }
523 
524 /* dst = imm (4 bytes)*/
525 static inline void emit_a32_mov_r(const u8 dst, const u8 src,
526 				  bool dstk, bool sstk,
527 				  struct jit_ctx *ctx) {
528 	const u8 *tmp = bpf2a32[TMP_REG_1];
529 	u8 rt = sstk ? tmp[0] : src;
530 
531 	if (sstk)
532 		emit(ARM_LDR_I(tmp[0], ARM_SP, STACK_VAR(src)), ctx);
533 	if (dstk)
534 		emit(ARM_STR_I(rt, ARM_SP, STACK_VAR(dst)), ctx);
535 	else
536 		emit(ARM_MOV_R(dst, rt), ctx);
537 }
538 
539 /* dst = src */
540 static inline void emit_a32_mov_r64(const bool is64, const u8 dst[],
541 				  const u8 src[], bool dstk,
542 				  bool sstk, struct jit_ctx *ctx) {
543 	emit_a32_mov_r(dst_lo, src_lo, dstk, sstk, ctx);
544 	if (is64) {
545 		/* complete 8 byte move */
546 		emit_a32_mov_r(dst_hi, src_hi, dstk, sstk, ctx);
547 	} else {
548 		/* Zero out high 4 bytes */
549 		emit_a32_mov_i(dst_hi, 0, dstk, ctx);
550 	}
551 }
552 
553 /* Shift operations */
554 static inline void emit_a32_alu_i(const u8 dst, const u32 val, bool dstk,
555 				struct jit_ctx *ctx, const u8 op) {
556 	const u8 *tmp = bpf2a32[TMP_REG_1];
557 	u8 rd = dstk ? tmp[0] : dst;
558 
559 	if (dstk)
560 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst)), ctx);
561 
562 	/* Do shift operation */
563 	switch (op) {
564 	case BPF_LSH:
565 		emit(ARM_LSL_I(rd, rd, val), ctx);
566 		break;
567 	case BPF_RSH:
568 		emit(ARM_LSR_I(rd, rd, val), ctx);
569 		break;
570 	case BPF_NEG:
571 		emit(ARM_RSB_I(rd, rd, val), ctx);
572 		break;
573 	}
574 
575 	if (dstk)
576 		emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst)), ctx);
577 }
578 
579 /* dst = ~dst (64 bit) */
580 static inline void emit_a32_neg64(const u8 dst[], bool dstk,
581 				struct jit_ctx *ctx){
582 	const u8 *tmp = bpf2a32[TMP_REG_1];
583 	u8 rd = dstk ? tmp[1] : dst[1];
584 	u8 rm = dstk ? tmp[0] : dst[0];
585 
586 	/* Setup Operand */
587 	if (dstk) {
588 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
589 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
590 	}
591 
592 	/* Do Negate Operation */
593 	emit(ARM_RSBS_I(rd, rd, 0), ctx);
594 	emit(ARM_RSC_I(rm, rm, 0), ctx);
595 
596 	if (dstk) {
597 		emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
598 		emit(ARM_STR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
599 	}
600 }
601 
602 /* dst = dst << src */
603 static inline void emit_a32_lsh_r64(const u8 dst[], const u8 src[], bool dstk,
604 				    bool sstk, struct jit_ctx *ctx) {
605 	const u8 *tmp = bpf2a32[TMP_REG_1];
606 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
607 
608 	/* Setup Operands */
609 	u8 rt = sstk ? tmp2[1] : src_lo;
610 	u8 rd = dstk ? tmp[1] : dst_lo;
611 	u8 rm = dstk ? tmp[0] : dst_hi;
612 
613 	if (sstk)
614 		emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(src_lo)), ctx);
615 	if (dstk) {
616 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
617 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
618 	}
619 
620 	/* Do LSH operation */
621 	emit(ARM_SUB_I(ARM_IP, rt, 32), ctx);
622 	emit(ARM_RSB_I(tmp2[0], rt, 32), ctx);
623 	/* As we are using ARM_LR */
624 	ctx->seen |= SEEN_CALL;
625 	emit(ARM_MOV_SR(ARM_LR, rm, SRTYPE_ASL, rt), ctx);
626 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd, SRTYPE_ASL, ARM_IP), ctx);
627 	emit(ARM_ORR_SR(ARM_IP, ARM_LR, rd, SRTYPE_LSR, tmp2[0]), ctx);
628 	emit(ARM_MOV_SR(ARM_LR, rd, SRTYPE_ASL, rt), ctx);
629 
630 	if (dstk) {
631 		emit(ARM_STR_I(ARM_LR, ARM_SP, STACK_VAR(dst_lo)), ctx);
632 		emit(ARM_STR_I(ARM_IP, ARM_SP, STACK_VAR(dst_hi)), ctx);
633 	} else {
634 		emit(ARM_MOV_R(rd, ARM_LR), ctx);
635 		emit(ARM_MOV_R(rm, ARM_IP), ctx);
636 	}
637 }
638 
639 /* dst = dst >> src (signed)*/
640 static inline void emit_a32_arsh_r64(const u8 dst[], const u8 src[], bool dstk,
641 				    bool sstk, struct jit_ctx *ctx) {
642 	const u8 *tmp = bpf2a32[TMP_REG_1];
643 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
644 	/* Setup Operands */
645 	u8 rt = sstk ? tmp2[1] : src_lo;
646 	u8 rd = dstk ? tmp[1] : dst_lo;
647 	u8 rm = dstk ? tmp[0] : dst_hi;
648 
649 	if (sstk)
650 		emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(src_lo)), ctx);
651 	if (dstk) {
652 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
653 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
654 	}
655 
656 	/* Do the ARSH operation */
657 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
658 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
659 	/* As we are using ARM_LR */
660 	ctx->seen |= SEEN_CALL;
661 	emit(ARM_MOV_SR(ARM_LR, rd, SRTYPE_LSR, rt), ctx);
662 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rm, SRTYPE_ASL, ARM_IP), ctx);
663 	_emit(ARM_COND_MI, ARM_B(0), ctx);
664 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rm, SRTYPE_ASR, tmp2[0]), ctx);
665 	emit(ARM_MOV_SR(ARM_IP, rm, SRTYPE_ASR, rt), ctx);
666 	if (dstk) {
667 		emit(ARM_STR_I(ARM_LR, ARM_SP, STACK_VAR(dst_lo)), ctx);
668 		emit(ARM_STR_I(ARM_IP, ARM_SP, STACK_VAR(dst_hi)), ctx);
669 	} else {
670 		emit(ARM_MOV_R(rd, ARM_LR), ctx);
671 		emit(ARM_MOV_R(rm, ARM_IP), ctx);
672 	}
673 }
674 
675 /* dst = dst >> src */
676 static inline void emit_a32_lsr_r64(const u8 dst[], const u8 src[], bool dstk,
677 				     bool sstk, struct jit_ctx *ctx) {
678 	const u8 *tmp = bpf2a32[TMP_REG_1];
679 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
680 	/* Setup Operands */
681 	u8 rt = sstk ? tmp2[1] : src_lo;
682 	u8 rd = dstk ? tmp[1] : dst_lo;
683 	u8 rm = dstk ? tmp[0] : dst_hi;
684 
685 	if (sstk)
686 		emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(src_lo)), ctx);
687 	if (dstk) {
688 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
689 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
690 	}
691 
692 	/* Do LSH operation */
693 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
694 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
695 	/* As we are using ARM_LR */
696 	ctx->seen |= SEEN_CALL;
697 	emit(ARM_MOV_SR(ARM_LR, rd, SRTYPE_LSR, rt), ctx);
698 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rm, SRTYPE_ASL, ARM_IP), ctx);
699 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rm, SRTYPE_LSR, tmp2[0]), ctx);
700 	emit(ARM_MOV_SR(ARM_IP, rm, SRTYPE_LSR, rt), ctx);
701 	if (dstk) {
702 		emit(ARM_STR_I(ARM_LR, ARM_SP, STACK_VAR(dst_lo)), ctx);
703 		emit(ARM_STR_I(ARM_IP, ARM_SP, STACK_VAR(dst_hi)), ctx);
704 	} else {
705 		emit(ARM_MOV_R(rd, ARM_LR), ctx);
706 		emit(ARM_MOV_R(rm, ARM_IP), ctx);
707 	}
708 }
709 
710 /* dst = dst << val */
711 static inline void emit_a32_lsh_i64(const u8 dst[], bool dstk,
712 				     const u32 val, struct jit_ctx *ctx){
713 	const u8 *tmp = bpf2a32[TMP_REG_1];
714 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
715 	/* Setup operands */
716 	u8 rd = dstk ? tmp[1] : dst_lo;
717 	u8 rm = dstk ? tmp[0] : dst_hi;
718 
719 	if (dstk) {
720 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
721 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
722 	}
723 
724 	/* Do LSH operation */
725 	if (val < 32) {
726 		emit(ARM_MOV_SI(tmp2[0], rm, SRTYPE_ASL, val), ctx);
727 		emit(ARM_ORR_SI(rm, tmp2[0], rd, SRTYPE_LSR, 32 - val), ctx);
728 		emit(ARM_MOV_SI(rd, rd, SRTYPE_ASL, val), ctx);
729 	} else {
730 		if (val == 32)
731 			emit(ARM_MOV_R(rm, rd), ctx);
732 		else
733 			emit(ARM_MOV_SI(rm, rd, SRTYPE_ASL, val - 32), ctx);
734 		emit(ARM_EOR_R(rd, rd, rd), ctx);
735 	}
736 
737 	if (dstk) {
738 		emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
739 		emit(ARM_STR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
740 	}
741 }
742 
743 /* dst = dst >> val */
744 static inline void emit_a32_lsr_i64(const u8 dst[], bool dstk,
745 				    const u32 val, struct jit_ctx *ctx) {
746 	const u8 *tmp = bpf2a32[TMP_REG_1];
747 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
748 	/* Setup operands */
749 	u8 rd = dstk ? tmp[1] : dst_lo;
750 	u8 rm = dstk ? tmp[0] : dst_hi;
751 
752 	if (dstk) {
753 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
754 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
755 	}
756 
757 	/* Do LSR operation */
758 	if (val < 32) {
759 		emit(ARM_MOV_SI(tmp2[1], rd, SRTYPE_LSR, val), ctx);
760 		emit(ARM_ORR_SI(rd, tmp2[1], rm, SRTYPE_ASL, 32 - val), ctx);
761 		emit(ARM_MOV_SI(rm, rm, SRTYPE_LSR, val), ctx);
762 	} else if (val == 32) {
763 		emit(ARM_MOV_R(rd, rm), ctx);
764 		emit(ARM_MOV_I(rm, 0), ctx);
765 	} else {
766 		emit(ARM_MOV_SI(rd, rm, SRTYPE_LSR, val - 32), ctx);
767 		emit(ARM_MOV_I(rm, 0), ctx);
768 	}
769 
770 	if (dstk) {
771 		emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
772 		emit(ARM_STR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
773 	}
774 }
775 
776 /* dst = dst >> val (signed) */
777 static inline void emit_a32_arsh_i64(const u8 dst[], bool dstk,
778 				     const u32 val, struct jit_ctx *ctx){
779 	const u8 *tmp = bpf2a32[TMP_REG_1];
780 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
781 	 /* Setup operands */
782 	u8 rd = dstk ? tmp[1] : dst_lo;
783 	u8 rm = dstk ? tmp[0] : dst_hi;
784 
785 	if (dstk) {
786 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
787 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
788 	}
789 
790 	/* Do ARSH operation */
791 	if (val < 32) {
792 		emit(ARM_MOV_SI(tmp2[1], rd, SRTYPE_LSR, val), ctx);
793 		emit(ARM_ORR_SI(rd, tmp2[1], rm, SRTYPE_ASL, 32 - val), ctx);
794 		emit(ARM_MOV_SI(rm, rm, SRTYPE_ASR, val), ctx);
795 	} else if (val == 32) {
796 		emit(ARM_MOV_R(rd, rm), ctx);
797 		emit(ARM_MOV_SI(rm, rm, SRTYPE_ASR, 31), ctx);
798 	} else {
799 		emit(ARM_MOV_SI(rd, rm, SRTYPE_ASR, val - 32), ctx);
800 		emit(ARM_MOV_SI(rm, rm, SRTYPE_ASR, 31), ctx);
801 	}
802 
803 	if (dstk) {
804 		emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
805 		emit(ARM_STR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
806 	}
807 }
808 
809 static inline void emit_a32_mul_r64(const u8 dst[], const u8 src[], bool dstk,
810 				    bool sstk, struct jit_ctx *ctx) {
811 	const u8 *tmp = bpf2a32[TMP_REG_1];
812 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
813 	/* Setup operands for multiplication */
814 	u8 rd = dstk ? tmp[1] : dst_lo;
815 	u8 rm = dstk ? tmp[0] : dst_hi;
816 	u8 rt = sstk ? tmp2[1] : src_lo;
817 	u8 rn = sstk ? tmp2[0] : src_hi;
818 
819 	if (dstk) {
820 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
821 		emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
822 	}
823 	if (sstk) {
824 		emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(src_lo)), ctx);
825 		emit(ARM_LDR_I(rn, ARM_SP, STACK_VAR(src_hi)), ctx);
826 	}
827 
828 	/* Do Multiplication */
829 	emit(ARM_MUL(ARM_IP, rd, rn), ctx);
830 	emit(ARM_MUL(ARM_LR, rm, rt), ctx);
831 	/* As we are using ARM_LR */
832 	ctx->seen |= SEEN_CALL;
833 	emit(ARM_ADD_R(ARM_LR, ARM_IP, ARM_LR), ctx);
834 
835 	emit(ARM_UMULL(ARM_IP, rm, rd, rt), ctx);
836 	emit(ARM_ADD_R(rm, ARM_LR, rm), ctx);
837 	if (dstk) {
838 		emit(ARM_STR_I(ARM_IP, ARM_SP, STACK_VAR(dst_lo)), ctx);
839 		emit(ARM_STR_I(rm, ARM_SP, STACK_VAR(dst_hi)), ctx);
840 	} else {
841 		emit(ARM_MOV_R(rd, ARM_IP), ctx);
842 	}
843 }
844 
845 /* *(size *)(dst + off) = src */
846 static inline void emit_str_r(const u8 dst, const u8 src, bool dstk,
847 			      const s32 off, struct jit_ctx *ctx, const u8 sz){
848 	const u8 *tmp = bpf2a32[TMP_REG_1];
849 	u8 rd = dstk ? tmp[1] : dst;
850 
851 	if (dstk)
852 		emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst)), ctx);
853 	if (off) {
854 		emit_a32_mov_i(tmp[0], off, false, ctx);
855 		emit(ARM_ADD_R(tmp[0], rd, tmp[0]), ctx);
856 		rd = tmp[0];
857 	}
858 	switch (sz) {
859 	case BPF_W:
860 		/* Store a Word */
861 		emit(ARM_STR_I(src, rd, 0), ctx);
862 		break;
863 	case BPF_H:
864 		/* Store a HalfWord */
865 		emit(ARM_STRH_I(src, rd, 0), ctx);
866 		break;
867 	case BPF_B:
868 		/* Store a Byte */
869 		emit(ARM_STRB_I(src, rd, 0), ctx);
870 		break;
871 	}
872 }
873 
874 /* dst = *(size*)(src + off) */
875 static inline void emit_ldx_r(const u8 dst, const u8 src, bool dstk,
876 			      const s32 off, struct jit_ctx *ctx, const u8 sz){
877 	const u8 *tmp = bpf2a32[TMP_REG_1];
878 	u8 rd = dstk ? tmp[1] : dst;
879 	u8 rm = src;
880 
881 	if (off) {
882 		emit_a32_mov_i(tmp[0], off, false, ctx);
883 		emit(ARM_ADD_R(tmp[0], tmp[0], src), ctx);
884 		rm = tmp[0];
885 	}
886 	switch (sz) {
887 	case BPF_W:
888 		/* Load a Word */
889 		emit(ARM_LDR_I(rd, rm, 0), ctx);
890 		break;
891 	case BPF_H:
892 		/* Load a HalfWord */
893 		emit(ARM_LDRH_I(rd, rm, 0), ctx);
894 		break;
895 	case BPF_B:
896 		/* Load a Byte */
897 		emit(ARM_LDRB_I(rd, rm, 0), ctx);
898 		break;
899 	}
900 	if (dstk)
901 		emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst)), ctx);
902 }
903 
904 /* Arithmatic Operation */
905 static inline void emit_ar_r(const u8 rd, const u8 rt, const u8 rm,
906 			     const u8 rn, struct jit_ctx *ctx, u8 op) {
907 	switch (op) {
908 	case BPF_JSET:
909 		ctx->seen |= SEEN_CALL;
910 		emit(ARM_AND_R(ARM_IP, rt, rn), ctx);
911 		emit(ARM_AND_R(ARM_LR, rd, rm), ctx);
912 		emit(ARM_ORRS_R(ARM_IP, ARM_LR, ARM_IP), ctx);
913 		break;
914 	case BPF_JEQ:
915 	case BPF_JNE:
916 	case BPF_JGT:
917 	case BPF_JGE:
918 	case BPF_JLE:
919 	case BPF_JLT:
920 		emit(ARM_CMP_R(rd, rm), ctx);
921 		_emit(ARM_COND_EQ, ARM_CMP_R(rt, rn), ctx);
922 		break;
923 	case BPF_JSLE:
924 	case BPF_JSGT:
925 		emit(ARM_CMP_R(rn, rt), ctx);
926 		emit(ARM_SBCS_R(ARM_IP, rm, rd), ctx);
927 		break;
928 	case BPF_JSLT:
929 	case BPF_JSGE:
930 		emit(ARM_CMP_R(rt, rn), ctx);
931 		emit(ARM_SBCS_R(ARM_IP, rd, rm), ctx);
932 		break;
933 	}
934 }
935 
936 static int out_offset = -1; /* initialized on the first pass of build_body() */
937 static int emit_bpf_tail_call(struct jit_ctx *ctx)
938 {
939 
940 	/* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
941 	const u8 *r2 = bpf2a32[BPF_REG_2];
942 	const u8 *r3 = bpf2a32[BPF_REG_3];
943 	const u8 *tmp = bpf2a32[TMP_REG_1];
944 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
945 	const u8 *tcc = bpf2a32[TCALL_CNT];
946 	const int idx0 = ctx->idx;
947 #define cur_offset (ctx->idx - idx0)
948 #define jmp_offset (out_offset - (cur_offset))
949 	u32 off, lo, hi;
950 
951 	/* if (index >= array->map.max_entries)
952 	 *	goto out;
953 	 */
954 	off = offsetof(struct bpf_array, map.max_entries);
955 	/* array->map.max_entries */
956 	emit_a32_mov_i(tmp[1], off, false, ctx);
957 	emit(ARM_LDR_I(tmp2[1], ARM_SP, STACK_VAR(r2[1])), ctx);
958 	emit(ARM_LDR_R(tmp[1], tmp2[1], tmp[1]), ctx);
959 	/* index (64 bit) */
960 	emit(ARM_LDR_I(tmp2[1], ARM_SP, STACK_VAR(r3[1])), ctx);
961 	/* index >= array->map.max_entries */
962 	emit(ARM_CMP_R(tmp2[1], tmp[1]), ctx);
963 	_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
964 
965 	/* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
966 	 *	goto out;
967 	 * tail_call_cnt++;
968 	 */
969 	lo = (u32)MAX_TAIL_CALL_CNT;
970 	hi = (u32)((u64)MAX_TAIL_CALL_CNT >> 32);
971 	emit(ARM_LDR_I(tmp[1], ARM_SP, STACK_VAR(tcc[1])), ctx);
972 	emit(ARM_LDR_I(tmp[0], ARM_SP, STACK_VAR(tcc[0])), ctx);
973 	emit(ARM_CMP_I(tmp[0], hi), ctx);
974 	_emit(ARM_COND_EQ, ARM_CMP_I(tmp[1], lo), ctx);
975 	_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
976 	emit(ARM_ADDS_I(tmp[1], tmp[1], 1), ctx);
977 	emit(ARM_ADC_I(tmp[0], tmp[0], 0), ctx);
978 	emit(ARM_STR_I(tmp[1], ARM_SP, STACK_VAR(tcc[1])), ctx);
979 	emit(ARM_STR_I(tmp[0], ARM_SP, STACK_VAR(tcc[0])), ctx);
980 
981 	/* prog = array->ptrs[index]
982 	 * if (prog == NULL)
983 	 *	goto out;
984 	 */
985 	off = offsetof(struct bpf_array, ptrs);
986 	emit_a32_mov_i(tmp[1], off, false, ctx);
987 	emit(ARM_LDR_I(tmp2[1], ARM_SP, STACK_VAR(r2[1])), ctx);
988 	emit(ARM_ADD_R(tmp[1], tmp2[1], tmp[1]), ctx);
989 	emit(ARM_LDR_I(tmp2[1], ARM_SP, STACK_VAR(r3[1])), ctx);
990 	emit(ARM_MOV_SI(tmp[0], tmp2[1], SRTYPE_ASL, 2), ctx);
991 	emit(ARM_LDR_R(tmp[1], tmp[1], tmp[0]), ctx);
992 	emit(ARM_CMP_I(tmp[1], 0), ctx);
993 	_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
994 
995 	/* goto *(prog->bpf_func + prologue_size); */
996 	off = offsetof(struct bpf_prog, bpf_func);
997 	emit_a32_mov_i(tmp2[1], off, false, ctx);
998 	emit(ARM_LDR_R(tmp[1], tmp[1], tmp2[1]), ctx);
999 	emit(ARM_ADD_I(tmp[1], tmp[1], ctx->prologue_bytes), ctx);
1000 	emit(ARM_BX(tmp[1]), ctx);
1001 
1002 	/* out: */
1003 	if (out_offset == -1)
1004 		out_offset = cur_offset;
1005 	if (cur_offset != out_offset) {
1006 		pr_err_once("tail_call out_offset = %d, expected %d!\n",
1007 			    cur_offset, out_offset);
1008 		return -1;
1009 	}
1010 	return 0;
1011 #undef cur_offset
1012 #undef jmp_offset
1013 }
1014 
1015 /* 0xabcd => 0xcdab */
1016 static inline void emit_rev16(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1017 {
1018 #if __LINUX_ARM_ARCH__ < 6
1019 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
1020 
1021 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1022 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 8), ctx);
1023 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1024 	emit(ARM_ORR_SI(rd, tmp2[0], tmp2[1], SRTYPE_LSL, 8), ctx);
1025 #else /* ARMv6+ */
1026 	emit(ARM_REV16(rd, rn), ctx);
1027 #endif
1028 }
1029 
1030 /* 0xabcdefgh => 0xghefcdab */
1031 static inline void emit_rev32(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1032 {
1033 #if __LINUX_ARM_ARCH__ < 6
1034 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
1035 
1036 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1037 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 24), ctx);
1038 	emit(ARM_ORR_SI(ARM_IP, tmp2[0], tmp2[1], SRTYPE_LSL, 24), ctx);
1039 
1040 	emit(ARM_MOV_SI(tmp2[1], rn, SRTYPE_LSR, 8), ctx);
1041 	emit(ARM_AND_I(tmp2[1], tmp2[1], 0xff), ctx);
1042 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 16), ctx);
1043 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1044 	emit(ARM_MOV_SI(tmp2[0], tmp2[0], SRTYPE_LSL, 8), ctx);
1045 	emit(ARM_ORR_SI(tmp2[0], tmp2[0], tmp2[1], SRTYPE_LSL, 16), ctx);
1046 	emit(ARM_ORR_R(rd, ARM_IP, tmp2[0]), ctx);
1047 
1048 #else /* ARMv6+ */
1049 	emit(ARM_REV(rd, rn), ctx);
1050 #endif
1051 }
1052 
1053 // push the scratch stack register on top of the stack
1054 static inline void emit_push_r64(const u8 src[], const u8 shift,
1055 		struct jit_ctx *ctx)
1056 {
1057 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
1058 	u16 reg_set = 0;
1059 
1060 	emit(ARM_LDR_I(tmp2[1], ARM_SP, STACK_VAR(src[1]+shift)), ctx);
1061 	emit(ARM_LDR_I(tmp2[0], ARM_SP, STACK_VAR(src[0]+shift)), ctx);
1062 
1063 	reg_set = (1 << tmp2[1]) | (1 << tmp2[0]);
1064 	emit(ARM_PUSH(reg_set), ctx);
1065 }
1066 
1067 static void build_prologue(struct jit_ctx *ctx)
1068 {
1069 	const u8 r0 = bpf2a32[BPF_REG_0][1];
1070 	const u8 r2 = bpf2a32[BPF_REG_1][1];
1071 	const u8 r3 = bpf2a32[BPF_REG_1][0];
1072 	const u8 r4 = bpf2a32[BPF_REG_6][1];
1073 	const u8 r5 = bpf2a32[BPF_REG_6][0];
1074 	const u8 r6 = bpf2a32[TMP_REG_1][1];
1075 	const u8 r7 = bpf2a32[TMP_REG_1][0];
1076 	const u8 r8 = bpf2a32[TMP_REG_2][1];
1077 	const u8 r10 = bpf2a32[TMP_REG_2][0];
1078 	const u8 fplo = bpf2a32[BPF_REG_FP][1];
1079 	const u8 fphi = bpf2a32[BPF_REG_FP][0];
1080 	const u8 sp = ARM_SP;
1081 	const u8 *tcc = bpf2a32[TCALL_CNT];
1082 
1083 	u16 reg_set = 0;
1084 
1085 	/*
1086 	 * eBPF prog stack layout
1087 	 *
1088 	 *                         high
1089 	 * original ARM_SP =>     +-----+ eBPF prologue
1090 	 *                        |FP/LR|
1091 	 * current ARM_FP =>      +-----+
1092 	 *                        | ... | callee saved registers
1093 	 * eBPF fp register =>    +-----+ <= (BPF_FP)
1094 	 *                        | ... | eBPF JIT scratch space
1095 	 *                        |     | eBPF prog stack
1096 	 *                        +-----+
1097 	 *			  |RSVD | JIT scratchpad
1098 	 * current A64_SP =>      +-----+ <= (BPF_FP - STACK_SIZE)
1099 	 *                        |     |
1100 	 *                        | ... | Function call stack
1101 	 *                        |     |
1102 	 *                        +-----+
1103 	 *                          low
1104 	 */
1105 
1106 	/* Save callee saved registers. */
1107 	reg_set |= (1<<r4) | (1<<r5) | (1<<r6) | (1<<r7) | (1<<r8) | (1<<r10);
1108 #ifdef CONFIG_FRAME_POINTER
1109 	reg_set |= (1<<ARM_FP) | (1<<ARM_IP) | (1<<ARM_LR) | (1<<ARM_PC);
1110 	emit(ARM_MOV_R(ARM_IP, sp), ctx);
1111 	emit(ARM_PUSH(reg_set), ctx);
1112 	emit(ARM_SUB_I(ARM_FP, ARM_IP, 4), ctx);
1113 #else
1114 	/* Check if call instruction exists in BPF body */
1115 	if (ctx->seen & SEEN_CALL)
1116 		reg_set |= (1<<ARM_LR);
1117 	emit(ARM_PUSH(reg_set), ctx);
1118 #endif
1119 	/* Save frame pointer for later */
1120 	emit(ARM_SUB_I(ARM_IP, sp, SCRATCH_SIZE), ctx);
1121 
1122 	ctx->stack_size = imm8m(STACK_SIZE);
1123 
1124 	/* Set up function call stack */
1125 	emit(ARM_SUB_I(ARM_SP, ARM_SP, ctx->stack_size), ctx);
1126 
1127 	/* Set up BPF prog stack base register */
1128 	emit_a32_mov_r(fplo, ARM_IP, true, false, ctx);
1129 	emit_a32_mov_i(fphi, 0, true, ctx);
1130 
1131 	/* mov r4, 0 */
1132 	emit(ARM_MOV_I(r4, 0), ctx);
1133 
1134 	/* Move BPF_CTX to BPF_R1 */
1135 	emit(ARM_MOV_R(r3, r4), ctx);
1136 	emit(ARM_MOV_R(r2, r0), ctx);
1137 	/* Initialize Tail Count */
1138 	emit(ARM_STR_I(r4, ARM_SP, STACK_VAR(tcc[0])), ctx);
1139 	emit(ARM_STR_I(r4, ARM_SP, STACK_VAR(tcc[1])), ctx);
1140 	/* end of prologue */
1141 }
1142 
1143 static void build_epilogue(struct jit_ctx *ctx)
1144 {
1145 	const u8 r4 = bpf2a32[BPF_REG_6][1];
1146 	const u8 r5 = bpf2a32[BPF_REG_6][0];
1147 	const u8 r6 = bpf2a32[TMP_REG_1][1];
1148 	const u8 r7 = bpf2a32[TMP_REG_1][0];
1149 	const u8 r8 = bpf2a32[TMP_REG_2][1];
1150 	const u8 r10 = bpf2a32[TMP_REG_2][0];
1151 	u16 reg_set = 0;
1152 
1153 	/* unwind function call stack */
1154 	emit(ARM_ADD_I(ARM_SP, ARM_SP, ctx->stack_size), ctx);
1155 
1156 	/* restore callee saved registers. */
1157 	reg_set |= (1<<r4) | (1<<r5) | (1<<r6) | (1<<r7) | (1<<r8) | (1<<r10);
1158 #ifdef CONFIG_FRAME_POINTER
1159 	/* the first instruction of the prologue was: mov ip, sp */
1160 	reg_set |= (1<<ARM_FP) | (1<<ARM_SP) | (1<<ARM_PC);
1161 	emit(ARM_LDM(ARM_SP, reg_set), ctx);
1162 #else
1163 	if (ctx->seen & SEEN_CALL)
1164 		reg_set |= (1<<ARM_PC);
1165 	/* Restore callee saved registers. */
1166 	emit(ARM_POP(reg_set), ctx);
1167 	/* Return back to the callee function */
1168 	if (!(ctx->seen & SEEN_CALL))
1169 		emit(ARM_BX(ARM_LR), ctx);
1170 #endif
1171 }
1172 
1173 /*
1174  * Convert an eBPF instruction to native instruction, i.e
1175  * JITs an eBPF instruction.
1176  * Returns :
1177  *	0  - Successfully JITed an 8-byte eBPF instruction
1178  *	>0 - Successfully JITed a 16-byte eBPF instruction
1179  *	<0 - Failed to JIT.
1180  */
1181 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx)
1182 {
1183 	const u8 code = insn->code;
1184 	const u8 *dst = bpf2a32[insn->dst_reg];
1185 	const u8 *src = bpf2a32[insn->src_reg];
1186 	const u8 *tmp = bpf2a32[TMP_REG_1];
1187 	const u8 *tmp2 = bpf2a32[TMP_REG_2];
1188 	const s16 off = insn->off;
1189 	const s32 imm = insn->imm;
1190 	const int i = insn - ctx->prog->insnsi;
1191 	const bool is64 = BPF_CLASS(code) == BPF_ALU64;
1192 	const bool dstk = is_on_stack(insn->dst_reg);
1193 	const bool sstk = is_on_stack(insn->src_reg);
1194 	u8 rd, rt, rm, rn;
1195 	s32 jmp_offset;
1196 
1197 #define check_imm(bits, imm) do {				\
1198 	if ((((imm) > 0) && ((imm) >> (bits))) ||		\
1199 	    (((imm) < 0) && (~(imm) >> (bits)))) {		\
1200 		pr_info("[%2d] imm=%d(0x%x) out of range\n",	\
1201 			i, imm, imm);				\
1202 		return -EINVAL;					\
1203 	}							\
1204 } while (0)
1205 #define check_imm24(imm) check_imm(24, imm)
1206 
1207 	switch (code) {
1208 	/* ALU operations */
1209 
1210 	/* dst = src */
1211 	case BPF_ALU | BPF_MOV | BPF_K:
1212 	case BPF_ALU | BPF_MOV | BPF_X:
1213 	case BPF_ALU64 | BPF_MOV | BPF_K:
1214 	case BPF_ALU64 | BPF_MOV | BPF_X:
1215 		switch (BPF_SRC(code)) {
1216 		case BPF_X:
1217 			emit_a32_mov_r64(is64, dst, src, dstk, sstk, ctx);
1218 			break;
1219 		case BPF_K:
1220 			/* Sign-extend immediate value to destination reg */
1221 			emit_a32_mov_i64(is64, dst, imm, dstk, ctx);
1222 			break;
1223 		}
1224 		break;
1225 	/* dst = dst + src/imm */
1226 	/* dst = dst - src/imm */
1227 	/* dst = dst | src/imm */
1228 	/* dst = dst & src/imm */
1229 	/* dst = dst ^ src/imm */
1230 	/* dst = dst * src/imm */
1231 	/* dst = dst << src */
1232 	/* dst = dst >> src */
1233 	case BPF_ALU | BPF_ADD | BPF_K:
1234 	case BPF_ALU | BPF_ADD | BPF_X:
1235 	case BPF_ALU | BPF_SUB | BPF_K:
1236 	case BPF_ALU | BPF_SUB | BPF_X:
1237 	case BPF_ALU | BPF_OR | BPF_K:
1238 	case BPF_ALU | BPF_OR | BPF_X:
1239 	case BPF_ALU | BPF_AND | BPF_K:
1240 	case BPF_ALU | BPF_AND | BPF_X:
1241 	case BPF_ALU | BPF_XOR | BPF_K:
1242 	case BPF_ALU | BPF_XOR | BPF_X:
1243 	case BPF_ALU | BPF_MUL | BPF_K:
1244 	case BPF_ALU | BPF_MUL | BPF_X:
1245 	case BPF_ALU | BPF_LSH | BPF_X:
1246 	case BPF_ALU | BPF_RSH | BPF_X:
1247 	case BPF_ALU | BPF_ARSH | BPF_K:
1248 	case BPF_ALU | BPF_ARSH | BPF_X:
1249 	case BPF_ALU64 | BPF_ADD | BPF_K:
1250 	case BPF_ALU64 | BPF_ADD | BPF_X:
1251 	case BPF_ALU64 | BPF_SUB | BPF_K:
1252 	case BPF_ALU64 | BPF_SUB | BPF_X:
1253 	case BPF_ALU64 | BPF_OR | BPF_K:
1254 	case BPF_ALU64 | BPF_OR | BPF_X:
1255 	case BPF_ALU64 | BPF_AND | BPF_K:
1256 	case BPF_ALU64 | BPF_AND | BPF_X:
1257 	case BPF_ALU64 | BPF_XOR | BPF_K:
1258 	case BPF_ALU64 | BPF_XOR | BPF_X:
1259 		switch (BPF_SRC(code)) {
1260 		case BPF_X:
1261 			emit_a32_alu_r64(is64, dst, src, dstk, sstk,
1262 					 ctx, BPF_OP(code));
1263 			break;
1264 		case BPF_K:
1265 			/* Move immediate value to the temporary register
1266 			 * and then do the ALU operation on the temporary
1267 			 * register as this will sign-extend the immediate
1268 			 * value into temporary reg and then it would be
1269 			 * safe to do the operation on it.
1270 			 */
1271 			emit_a32_mov_i64(is64, tmp2, imm, false, ctx);
1272 			emit_a32_alu_r64(is64, dst, tmp2, dstk, false,
1273 					 ctx, BPF_OP(code));
1274 			break;
1275 		}
1276 		break;
1277 	/* dst = dst / src(imm) */
1278 	/* dst = dst % src(imm) */
1279 	case BPF_ALU | BPF_DIV | BPF_K:
1280 	case BPF_ALU | BPF_DIV | BPF_X:
1281 	case BPF_ALU | BPF_MOD | BPF_K:
1282 	case BPF_ALU | BPF_MOD | BPF_X:
1283 		rt = src_lo;
1284 		rd = dstk ? tmp2[1] : dst_lo;
1285 		if (dstk)
1286 			emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
1287 		switch (BPF_SRC(code)) {
1288 		case BPF_X:
1289 			rt = sstk ? tmp2[0] : rt;
1290 			if (sstk)
1291 				emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(src_lo)),
1292 				     ctx);
1293 			break;
1294 		case BPF_K:
1295 			rt = tmp2[0];
1296 			emit_a32_mov_i(rt, imm, false, ctx);
1297 			break;
1298 		}
1299 		emit_udivmod(rd, rd, rt, ctx, BPF_OP(code));
1300 		if (dstk)
1301 			emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_lo)), ctx);
1302 		emit_a32_mov_i(dst_hi, 0, dstk, ctx);
1303 		break;
1304 	case BPF_ALU64 | BPF_DIV | BPF_K:
1305 	case BPF_ALU64 | BPF_DIV | BPF_X:
1306 	case BPF_ALU64 | BPF_MOD | BPF_K:
1307 	case BPF_ALU64 | BPF_MOD | BPF_X:
1308 		goto notyet;
1309 	/* dst = dst >> imm */
1310 	/* dst = dst << imm */
1311 	case BPF_ALU | BPF_RSH | BPF_K:
1312 	case BPF_ALU | BPF_LSH | BPF_K:
1313 		if (unlikely(imm > 31))
1314 			return -EINVAL;
1315 		if (imm)
1316 			emit_a32_alu_i(dst_lo, imm, dstk, ctx, BPF_OP(code));
1317 		emit_a32_mov_i(dst_hi, 0, dstk, ctx);
1318 		break;
1319 	/* dst = dst << imm */
1320 	case BPF_ALU64 | BPF_LSH | BPF_K:
1321 		if (unlikely(imm > 63))
1322 			return -EINVAL;
1323 		emit_a32_lsh_i64(dst, dstk, imm, ctx);
1324 		break;
1325 	/* dst = dst >> imm */
1326 	case BPF_ALU64 | BPF_RSH | BPF_K:
1327 		if (unlikely(imm > 63))
1328 			return -EINVAL;
1329 		emit_a32_lsr_i64(dst, dstk, imm, ctx);
1330 		break;
1331 	/* dst = dst << src */
1332 	case BPF_ALU64 | BPF_LSH | BPF_X:
1333 		emit_a32_lsh_r64(dst, src, dstk, sstk, ctx);
1334 		break;
1335 	/* dst = dst >> src */
1336 	case BPF_ALU64 | BPF_RSH | BPF_X:
1337 		emit_a32_lsr_r64(dst, src, dstk, sstk, ctx);
1338 		break;
1339 	/* dst = dst >> src (signed) */
1340 	case BPF_ALU64 | BPF_ARSH | BPF_X:
1341 		emit_a32_arsh_r64(dst, src, dstk, sstk, ctx);
1342 		break;
1343 	/* dst = dst >> imm (signed) */
1344 	case BPF_ALU64 | BPF_ARSH | BPF_K:
1345 		if (unlikely(imm > 63))
1346 			return -EINVAL;
1347 		emit_a32_arsh_i64(dst, dstk, imm, ctx);
1348 		break;
1349 	/* dst = ~dst */
1350 	case BPF_ALU | BPF_NEG:
1351 		emit_a32_alu_i(dst_lo, 0, dstk, ctx, BPF_OP(code));
1352 		emit_a32_mov_i(dst_hi, 0, dstk, ctx);
1353 		break;
1354 	/* dst = ~dst (64 bit) */
1355 	case BPF_ALU64 | BPF_NEG:
1356 		emit_a32_neg64(dst, dstk, ctx);
1357 		break;
1358 	/* dst = dst * src/imm */
1359 	case BPF_ALU64 | BPF_MUL | BPF_X:
1360 	case BPF_ALU64 | BPF_MUL | BPF_K:
1361 		switch (BPF_SRC(code)) {
1362 		case BPF_X:
1363 			emit_a32_mul_r64(dst, src, dstk, sstk, ctx);
1364 			break;
1365 		case BPF_K:
1366 			/* Move immediate value to the temporary register
1367 			 * and then do the multiplication on it as this
1368 			 * will sign-extend the immediate value into temp
1369 			 * reg then it would be safe to do the operation
1370 			 * on it.
1371 			 */
1372 			emit_a32_mov_i64(is64, tmp2, imm, false, ctx);
1373 			emit_a32_mul_r64(dst, tmp2, dstk, false, ctx);
1374 			break;
1375 		}
1376 		break;
1377 	/* dst = htole(dst) */
1378 	/* dst = htobe(dst) */
1379 	case BPF_ALU | BPF_END | BPF_FROM_LE:
1380 	case BPF_ALU | BPF_END | BPF_FROM_BE:
1381 		rd = dstk ? tmp[0] : dst_hi;
1382 		rt = dstk ? tmp[1] : dst_lo;
1383 		if (dstk) {
1384 			emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(dst_lo)), ctx);
1385 			emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_hi)), ctx);
1386 		}
1387 		if (BPF_SRC(code) == BPF_FROM_LE)
1388 			goto emit_bswap_uxt;
1389 		switch (imm) {
1390 		case 16:
1391 			emit_rev16(rt, rt, ctx);
1392 			goto emit_bswap_uxt;
1393 		case 32:
1394 			emit_rev32(rt, rt, ctx);
1395 			goto emit_bswap_uxt;
1396 		case 64:
1397 			/* Because of the usage of ARM_LR */
1398 			ctx->seen |= SEEN_CALL;
1399 			emit_rev32(ARM_LR, rt, ctx);
1400 			emit_rev32(rt, rd, ctx);
1401 			emit(ARM_MOV_R(rd, ARM_LR), ctx);
1402 			break;
1403 		}
1404 		goto exit;
1405 emit_bswap_uxt:
1406 		switch (imm) {
1407 		case 16:
1408 			/* zero-extend 16 bits into 64 bits */
1409 #if __LINUX_ARM_ARCH__ < 6
1410 			emit_a32_mov_i(tmp2[1], 0xffff, false, ctx);
1411 			emit(ARM_AND_R(rt, rt, tmp2[1]), ctx);
1412 #else /* ARMv6+ */
1413 			emit(ARM_UXTH(rt, rt), ctx);
1414 #endif
1415 			emit(ARM_EOR_R(rd, rd, rd), ctx);
1416 			break;
1417 		case 32:
1418 			/* zero-extend 32 bits into 64 bits */
1419 			emit(ARM_EOR_R(rd, rd, rd), ctx);
1420 			break;
1421 		case 64:
1422 			/* nop */
1423 			break;
1424 		}
1425 exit:
1426 		if (dstk) {
1427 			emit(ARM_STR_I(rt, ARM_SP, STACK_VAR(dst_lo)), ctx);
1428 			emit(ARM_STR_I(rd, ARM_SP, STACK_VAR(dst_hi)), ctx);
1429 		}
1430 		break;
1431 	/* dst = imm64 */
1432 	case BPF_LD | BPF_IMM | BPF_DW:
1433 	{
1434 		const struct bpf_insn insn1 = insn[1];
1435 		u32 hi, lo = imm;
1436 
1437 		hi = insn1.imm;
1438 		emit_a32_mov_i(dst_lo, lo, dstk, ctx);
1439 		emit_a32_mov_i(dst_hi, hi, dstk, ctx);
1440 
1441 		return 1;
1442 	}
1443 	/* LDX: dst = *(size *)(src + off) */
1444 	case BPF_LDX | BPF_MEM | BPF_W:
1445 	case BPF_LDX | BPF_MEM | BPF_H:
1446 	case BPF_LDX | BPF_MEM | BPF_B:
1447 	case BPF_LDX | BPF_MEM | BPF_DW:
1448 		rn = sstk ? tmp2[1] : src_lo;
1449 		if (sstk)
1450 			emit(ARM_LDR_I(rn, ARM_SP, STACK_VAR(src_lo)), ctx);
1451 		switch (BPF_SIZE(code)) {
1452 		case BPF_W:
1453 			/* Load a Word */
1454 		case BPF_H:
1455 			/* Load a Half-Word */
1456 		case BPF_B:
1457 			/* Load a Byte */
1458 			emit_ldx_r(dst_lo, rn, dstk, off, ctx, BPF_SIZE(code));
1459 			emit_a32_mov_i(dst_hi, 0, dstk, ctx);
1460 			break;
1461 		case BPF_DW:
1462 			/* Load a double word */
1463 			emit_ldx_r(dst_lo, rn, dstk, off, ctx, BPF_W);
1464 			emit_ldx_r(dst_hi, rn, dstk, off+4, ctx, BPF_W);
1465 			break;
1466 		}
1467 		break;
1468 	/* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + imm)) */
1469 	case BPF_LD | BPF_ABS | BPF_W:
1470 	case BPF_LD | BPF_ABS | BPF_H:
1471 	case BPF_LD | BPF_ABS | BPF_B:
1472 	/* R0 = ntohx(*(size *)(((struct sk_buff *)R6)->data + src + imm)) */
1473 	case BPF_LD | BPF_IND | BPF_W:
1474 	case BPF_LD | BPF_IND | BPF_H:
1475 	case BPF_LD | BPF_IND | BPF_B:
1476 	{
1477 		const u8 r4 = bpf2a32[BPF_REG_6][1]; /* r4 = ptr to sk_buff */
1478 		const u8 r0 = bpf2a32[BPF_REG_0][1]; /*r0: struct sk_buff *skb*/
1479 						     /* rtn value */
1480 		const u8 r1 = bpf2a32[BPF_REG_0][0]; /* r1: int k */
1481 		const u8 r2 = bpf2a32[BPF_REG_1][1]; /* r2: unsigned int size */
1482 		const u8 r3 = bpf2a32[BPF_REG_1][0]; /* r3: void *buffer */
1483 		const u8 r6 = bpf2a32[TMP_REG_1][1]; /* r6: void *(*func)(..) */
1484 		int size;
1485 
1486 		/* Setting up first argument */
1487 		emit(ARM_MOV_R(r0, r4), ctx);
1488 
1489 		/* Setting up second argument */
1490 		emit_a32_mov_i(r1, imm, false, ctx);
1491 		if (BPF_MODE(code) == BPF_IND)
1492 			emit_a32_alu_r(r1, src_lo, false, sstk, ctx,
1493 				       false, false, BPF_ADD);
1494 
1495 		/* Setting up third argument */
1496 		switch (BPF_SIZE(code)) {
1497 		case BPF_W:
1498 			size = 4;
1499 			break;
1500 		case BPF_H:
1501 			size = 2;
1502 			break;
1503 		case BPF_B:
1504 			size = 1;
1505 			break;
1506 		default:
1507 			return -EINVAL;
1508 		}
1509 		emit_a32_mov_i(r2, size, false, ctx);
1510 
1511 		/* Setting up fourth argument */
1512 		emit(ARM_ADD_I(r3, ARM_SP, imm8m(SKB_BUFFER)), ctx);
1513 
1514 		/* Setting up function pointer to call */
1515 		emit_a32_mov_i(r6, (unsigned int)bpf_load_pointer, false, ctx);
1516 		emit_blx_r(r6, ctx);
1517 
1518 		emit(ARM_EOR_R(r1, r1, r1), ctx);
1519 		/* Check if return address is NULL or not.
1520 		 * if NULL then jump to epilogue
1521 		 * else continue to load the value from retn address
1522 		 */
1523 		emit(ARM_CMP_I(r0, 0), ctx);
1524 		jmp_offset = epilogue_offset(ctx);
1525 		check_imm24(jmp_offset);
1526 		_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1527 
1528 		/* Load value from the address */
1529 		switch (BPF_SIZE(code)) {
1530 		case BPF_W:
1531 			emit(ARM_LDR_I(r0, r0, 0), ctx);
1532 			emit_rev32(r0, r0, ctx);
1533 			break;
1534 		case BPF_H:
1535 			emit(ARM_LDRH_I(r0, r0, 0), ctx);
1536 			emit_rev16(r0, r0, ctx);
1537 			break;
1538 		case BPF_B:
1539 			emit(ARM_LDRB_I(r0, r0, 0), ctx);
1540 			/* No need to reverse */
1541 			break;
1542 		}
1543 		break;
1544 	}
1545 	/* ST: *(size *)(dst + off) = imm */
1546 	case BPF_ST | BPF_MEM | BPF_W:
1547 	case BPF_ST | BPF_MEM | BPF_H:
1548 	case BPF_ST | BPF_MEM | BPF_B:
1549 	case BPF_ST | BPF_MEM | BPF_DW:
1550 		switch (BPF_SIZE(code)) {
1551 		case BPF_DW:
1552 			/* Sign-extend immediate value into temp reg */
1553 			emit_a32_mov_i64(true, tmp2, imm, false, ctx);
1554 			emit_str_r(dst_lo, tmp2[1], dstk, off, ctx, BPF_W);
1555 			emit_str_r(dst_lo, tmp2[0], dstk, off+4, ctx, BPF_W);
1556 			break;
1557 		case BPF_W:
1558 		case BPF_H:
1559 		case BPF_B:
1560 			emit_a32_mov_i(tmp2[1], imm, false, ctx);
1561 			emit_str_r(dst_lo, tmp2[1], dstk, off, ctx,
1562 				   BPF_SIZE(code));
1563 			break;
1564 		}
1565 		break;
1566 	/* STX XADD: lock *(u32 *)(dst + off) += src */
1567 	case BPF_STX | BPF_XADD | BPF_W:
1568 	/* STX XADD: lock *(u64 *)(dst + off) += src */
1569 	case BPF_STX | BPF_XADD | BPF_DW:
1570 		goto notyet;
1571 	/* STX: *(size *)(dst + off) = src */
1572 	case BPF_STX | BPF_MEM | BPF_W:
1573 	case BPF_STX | BPF_MEM | BPF_H:
1574 	case BPF_STX | BPF_MEM | BPF_B:
1575 	case BPF_STX | BPF_MEM | BPF_DW:
1576 	{
1577 		u8 sz = BPF_SIZE(code);
1578 
1579 		rn = sstk ? tmp2[1] : src_lo;
1580 		rm = sstk ? tmp2[0] : src_hi;
1581 		if (sstk) {
1582 			emit(ARM_LDR_I(rn, ARM_SP, STACK_VAR(src_lo)), ctx);
1583 			emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(src_hi)), ctx);
1584 		}
1585 
1586 		/* Store the value */
1587 		if (BPF_SIZE(code) == BPF_DW) {
1588 			emit_str_r(dst_lo, rn, dstk, off, ctx, BPF_W);
1589 			emit_str_r(dst_lo, rm, dstk, off+4, ctx, BPF_W);
1590 		} else {
1591 			emit_str_r(dst_lo, rn, dstk, off, ctx, sz);
1592 		}
1593 		break;
1594 	}
1595 	/* PC += off if dst == src */
1596 	/* PC += off if dst > src */
1597 	/* PC += off if dst >= src */
1598 	/* PC += off if dst < src */
1599 	/* PC += off if dst <= src */
1600 	/* PC += off if dst != src */
1601 	/* PC += off if dst > src (signed) */
1602 	/* PC += off if dst >= src (signed) */
1603 	/* PC += off if dst < src (signed) */
1604 	/* PC += off if dst <= src (signed) */
1605 	/* PC += off if dst & src */
1606 	case BPF_JMP | BPF_JEQ | BPF_X:
1607 	case BPF_JMP | BPF_JGT | BPF_X:
1608 	case BPF_JMP | BPF_JGE | BPF_X:
1609 	case BPF_JMP | BPF_JNE | BPF_X:
1610 	case BPF_JMP | BPF_JSGT | BPF_X:
1611 	case BPF_JMP | BPF_JSGE | BPF_X:
1612 	case BPF_JMP | BPF_JSET | BPF_X:
1613 	case BPF_JMP | BPF_JLE | BPF_X:
1614 	case BPF_JMP | BPF_JLT | BPF_X:
1615 	case BPF_JMP | BPF_JSLT | BPF_X:
1616 	case BPF_JMP | BPF_JSLE | BPF_X:
1617 		/* Setup source registers */
1618 		rm = sstk ? tmp2[0] : src_hi;
1619 		rn = sstk ? tmp2[1] : src_lo;
1620 		if (sstk) {
1621 			emit(ARM_LDR_I(rn, ARM_SP, STACK_VAR(src_lo)), ctx);
1622 			emit(ARM_LDR_I(rm, ARM_SP, STACK_VAR(src_hi)), ctx);
1623 		}
1624 		goto go_jmp;
1625 	/* PC += off if dst == imm */
1626 	/* PC += off if dst > imm */
1627 	/* PC += off if dst >= imm */
1628 	/* PC += off if dst < imm */
1629 	/* PC += off if dst <= imm */
1630 	/* PC += off if dst != imm */
1631 	/* PC += off if dst > imm (signed) */
1632 	/* PC += off if dst >= imm (signed) */
1633 	/* PC += off if dst < imm (signed) */
1634 	/* PC += off if dst <= imm (signed) */
1635 	/* PC += off if dst & imm */
1636 	case BPF_JMP | BPF_JEQ | BPF_K:
1637 	case BPF_JMP | BPF_JGT | BPF_K:
1638 	case BPF_JMP | BPF_JGE | BPF_K:
1639 	case BPF_JMP | BPF_JNE | BPF_K:
1640 	case BPF_JMP | BPF_JSGT | BPF_K:
1641 	case BPF_JMP | BPF_JSGE | BPF_K:
1642 	case BPF_JMP | BPF_JSET | BPF_K:
1643 	case BPF_JMP | BPF_JLT | BPF_K:
1644 	case BPF_JMP | BPF_JLE | BPF_K:
1645 	case BPF_JMP | BPF_JSLT | BPF_K:
1646 	case BPF_JMP | BPF_JSLE | BPF_K:
1647 		if (off == 0)
1648 			break;
1649 		rm = tmp2[0];
1650 		rn = tmp2[1];
1651 		/* Sign-extend immediate value */
1652 		emit_a32_mov_i64(true, tmp2, imm, false, ctx);
1653 go_jmp:
1654 		/* Setup destination register */
1655 		rd = dstk ? tmp[0] : dst_hi;
1656 		rt = dstk ? tmp[1] : dst_lo;
1657 		if (dstk) {
1658 			emit(ARM_LDR_I(rt, ARM_SP, STACK_VAR(dst_lo)), ctx);
1659 			emit(ARM_LDR_I(rd, ARM_SP, STACK_VAR(dst_hi)), ctx);
1660 		}
1661 
1662 		/* Check for the condition */
1663 		emit_ar_r(rd, rt, rm, rn, ctx, BPF_OP(code));
1664 
1665 		/* Setup JUMP instruction */
1666 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1667 		switch (BPF_OP(code)) {
1668 		case BPF_JNE:
1669 		case BPF_JSET:
1670 			_emit(ARM_COND_NE, ARM_B(jmp_offset), ctx);
1671 			break;
1672 		case BPF_JEQ:
1673 			_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1674 			break;
1675 		case BPF_JGT:
1676 			_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
1677 			break;
1678 		case BPF_JGE:
1679 			_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
1680 			break;
1681 		case BPF_JSGT:
1682 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1683 			break;
1684 		case BPF_JSGE:
1685 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1686 			break;
1687 		case BPF_JLE:
1688 			_emit(ARM_COND_LS, ARM_B(jmp_offset), ctx);
1689 			break;
1690 		case BPF_JLT:
1691 			_emit(ARM_COND_CC, ARM_B(jmp_offset), ctx);
1692 			break;
1693 		case BPF_JSLT:
1694 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1695 			break;
1696 		case BPF_JSLE:
1697 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1698 			break;
1699 		}
1700 		break;
1701 	/* JMP OFF */
1702 	case BPF_JMP | BPF_JA:
1703 	{
1704 		if (off == 0)
1705 			break;
1706 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1707 		check_imm24(jmp_offset);
1708 		emit(ARM_B(jmp_offset), ctx);
1709 		break;
1710 	}
1711 	/* tail call */
1712 	case BPF_JMP | BPF_TAIL_CALL:
1713 		if (emit_bpf_tail_call(ctx))
1714 			return -EFAULT;
1715 		break;
1716 	/* function call */
1717 	case BPF_JMP | BPF_CALL:
1718 	{
1719 		const u8 *r0 = bpf2a32[BPF_REG_0];
1720 		const u8 *r1 = bpf2a32[BPF_REG_1];
1721 		const u8 *r2 = bpf2a32[BPF_REG_2];
1722 		const u8 *r3 = bpf2a32[BPF_REG_3];
1723 		const u8 *r4 = bpf2a32[BPF_REG_4];
1724 		const u8 *r5 = bpf2a32[BPF_REG_5];
1725 		const u32 func = (u32)__bpf_call_base + (u32)imm;
1726 
1727 		emit_a32_mov_r64(true, r0, r1, false, false, ctx);
1728 		emit_a32_mov_r64(true, r1, r2, false, true, ctx);
1729 		emit_push_r64(r5, 0, ctx);
1730 		emit_push_r64(r4, 8, ctx);
1731 		emit_push_r64(r3, 16, ctx);
1732 
1733 		emit_a32_mov_i(tmp[1], func, false, ctx);
1734 		emit_blx_r(tmp[1], ctx);
1735 
1736 		emit(ARM_ADD_I(ARM_SP, ARM_SP, imm8m(24)), ctx); // callee clean
1737 		break;
1738 	}
1739 	/* function return */
1740 	case BPF_JMP | BPF_EXIT:
1741 		/* Optimization: when last instruction is EXIT
1742 		 * simply fallthrough to epilogue.
1743 		 */
1744 		if (i == ctx->prog->len - 1)
1745 			break;
1746 		jmp_offset = epilogue_offset(ctx);
1747 		check_imm24(jmp_offset);
1748 		emit(ARM_B(jmp_offset), ctx);
1749 		break;
1750 notyet:
1751 		pr_info_once("*** NOT YET: opcode %02x ***\n", code);
1752 		return -EFAULT;
1753 	default:
1754 		pr_err_once("unknown opcode %02x\n", code);
1755 		return -EINVAL;
1756 	}
1757 
1758 	if (ctx->flags & FLAG_IMM_OVERFLOW)
1759 		/*
1760 		 * this instruction generated an overflow when
1761 		 * trying to access the literal pool, so
1762 		 * delegate this filter to the kernel interpreter.
1763 		 */
1764 		return -1;
1765 	return 0;
1766 }
1767 
1768 static int build_body(struct jit_ctx *ctx)
1769 {
1770 	const struct bpf_prog *prog = ctx->prog;
1771 	unsigned int i;
1772 
1773 	for (i = 0; i < prog->len; i++) {
1774 		const struct bpf_insn *insn = &(prog->insnsi[i]);
1775 		int ret;
1776 
1777 		ret = build_insn(insn, ctx);
1778 
1779 		/* It's used with loading the 64 bit immediate value. */
1780 		if (ret > 0) {
1781 			i++;
1782 			if (ctx->target == NULL)
1783 				ctx->offsets[i] = ctx->idx;
1784 			continue;
1785 		}
1786 
1787 		if (ctx->target == NULL)
1788 			ctx->offsets[i] = ctx->idx;
1789 
1790 		/* If unsuccesfull, return with error code */
1791 		if (ret)
1792 			return ret;
1793 	}
1794 	return 0;
1795 }
1796 
1797 static int validate_code(struct jit_ctx *ctx)
1798 {
1799 	int i;
1800 
1801 	for (i = 0; i < ctx->idx; i++) {
1802 		if (ctx->target[i] == __opcode_to_mem_arm(ARM_INST_UDF))
1803 			return -1;
1804 	}
1805 
1806 	return 0;
1807 }
1808 
1809 void bpf_jit_compile(struct bpf_prog *prog)
1810 {
1811 	/* Nothing to do here. We support Internal BPF. */
1812 }
1813 
1814 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1815 {
1816 	struct bpf_prog *tmp, *orig_prog = prog;
1817 	struct bpf_binary_header *header;
1818 	bool tmp_blinded = false;
1819 	struct jit_ctx ctx;
1820 	unsigned int tmp_idx;
1821 	unsigned int image_size;
1822 	u8 *image_ptr;
1823 
1824 	/* If BPF JIT was not enabled then we must fall back to
1825 	 * the interpreter.
1826 	 */
1827 	if (!bpf_jit_enable)
1828 		return orig_prog;
1829 
1830 	/* If constant blinding was enabled and we failed during blinding
1831 	 * then we must fall back to the interpreter. Otherwise, we save
1832 	 * the new JITed code.
1833 	 */
1834 	tmp = bpf_jit_blind_constants(prog);
1835 
1836 	if (IS_ERR(tmp))
1837 		return orig_prog;
1838 	if (tmp != prog) {
1839 		tmp_blinded = true;
1840 		prog = tmp;
1841 	}
1842 
1843 	memset(&ctx, 0, sizeof(ctx));
1844 	ctx.prog = prog;
1845 
1846 	/* Not able to allocate memory for offsets[] , then
1847 	 * we must fall back to the interpreter
1848 	 */
1849 	ctx.offsets = kcalloc(prog->len, sizeof(int), GFP_KERNEL);
1850 	if (ctx.offsets == NULL) {
1851 		prog = orig_prog;
1852 		goto out;
1853 	}
1854 
1855 	/* 1) fake pass to find in the length of the JITed code,
1856 	 * to compute ctx->offsets and other context variables
1857 	 * needed to compute final JITed code.
1858 	 * Also, calculate random starting pointer/start of JITed code
1859 	 * which is prefixed by random number of fault instructions.
1860 	 *
1861 	 * If the first pass fails then there is no chance of it
1862 	 * being successful in the second pass, so just fall back
1863 	 * to the interpreter.
1864 	 */
1865 	if (build_body(&ctx)) {
1866 		prog = orig_prog;
1867 		goto out_off;
1868 	}
1869 
1870 	tmp_idx = ctx.idx;
1871 	build_prologue(&ctx);
1872 	ctx.prologue_bytes = (ctx.idx - tmp_idx) * 4;
1873 
1874 	ctx.epilogue_offset = ctx.idx;
1875 
1876 #if __LINUX_ARM_ARCH__ < 7
1877 	tmp_idx = ctx.idx;
1878 	build_epilogue(&ctx);
1879 	ctx.epilogue_bytes = (ctx.idx - tmp_idx) * 4;
1880 
1881 	ctx.idx += ctx.imm_count;
1882 	if (ctx.imm_count) {
1883 		ctx.imms = kcalloc(ctx.imm_count, sizeof(u32), GFP_KERNEL);
1884 		if (ctx.imms == NULL) {
1885 			prog = orig_prog;
1886 			goto out_off;
1887 		}
1888 	}
1889 #else
1890 	/* there's nothing about the epilogue on ARMv7 */
1891 	build_epilogue(&ctx);
1892 #endif
1893 	/* Now we can get the actual image size of the JITed arm code.
1894 	 * Currently, we are not considering the THUMB-2 instructions
1895 	 * for jit, although it can decrease the size of the image.
1896 	 *
1897 	 * As each arm instruction is of length 32bit, we are translating
1898 	 * number of JITed intructions into the size required to store these
1899 	 * JITed code.
1900 	 */
1901 	image_size = sizeof(u32) * ctx.idx;
1902 
1903 	/* Now we know the size of the structure to make */
1904 	header = bpf_jit_binary_alloc(image_size, &image_ptr,
1905 				      sizeof(u32), jit_fill_hole);
1906 	/* Not able to allocate memory for the structure then
1907 	 * we must fall back to the interpretation
1908 	 */
1909 	if (header == NULL) {
1910 		prog = orig_prog;
1911 		goto out_imms;
1912 	}
1913 
1914 	/* 2.) Actual pass to generate final JIT code */
1915 	ctx.target = (u32 *) image_ptr;
1916 	ctx.idx = 0;
1917 
1918 	build_prologue(&ctx);
1919 
1920 	/* If building the body of the JITed code fails somehow,
1921 	 * we fall back to the interpretation.
1922 	 */
1923 	if (build_body(&ctx) < 0) {
1924 		image_ptr = NULL;
1925 		bpf_jit_binary_free(header);
1926 		prog = orig_prog;
1927 		goto out_imms;
1928 	}
1929 	build_epilogue(&ctx);
1930 
1931 	/* 3.) Extra pass to validate JITed Code */
1932 	if (validate_code(&ctx)) {
1933 		image_ptr = NULL;
1934 		bpf_jit_binary_free(header);
1935 		prog = orig_prog;
1936 		goto out_imms;
1937 	}
1938 	flush_icache_range((u32)header, (u32)(ctx.target + ctx.idx));
1939 
1940 	if (bpf_jit_enable > 1)
1941 		/* there are 2 passes here */
1942 		bpf_jit_dump(prog->len, image_size, 2, ctx.target);
1943 
1944 	set_memory_ro((unsigned long)header, header->pages);
1945 	prog->bpf_func = (void *)ctx.target;
1946 	prog->jited = 1;
1947 	prog->jited_len = image_size;
1948 
1949 out_imms:
1950 #if __LINUX_ARM_ARCH__ < 7
1951 	if (ctx.imm_count)
1952 		kfree(ctx.imms);
1953 #endif
1954 out_off:
1955 	kfree(ctx.offsets);
1956 out:
1957 	if (tmp_blinded)
1958 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
1959 					   tmp : orig_prog);
1960 	return prog;
1961 }
1962 
1963