xref: /linux/arch/arm/net/bpf_jit_32.c (revision 7fcfa9a2d9a7c1b428d61992c2deaa9e37a437b0)
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 #include <asm/system_info.h>
26 
27 #include "bpf_jit_32.h"
28 
29 /*
30  * eBPF prog stack layout:
31  *
32  *                         high
33  * original ARM_SP =>     +-----+
34  *                        |     | callee saved registers
35  *                        +-----+ <= (BPF_FP + SCRATCH_SIZE)
36  *                        | ... | eBPF JIT scratch space
37  * eBPF fp register =>    +-----+
38  *   (BPF_FP)             | ... | eBPF prog stack
39  *                        +-----+
40  *                        |RSVD | JIT scratchpad
41  * current ARM_SP =>      +-----+ <= (BPF_FP - STACK_SIZE + SCRATCH_SIZE)
42  *                        |     |
43  *                        | ... | Function call stack
44  *                        |     |
45  *                        +-----+
46  *                          low
47  *
48  * The callee saved registers depends on whether frame pointers are enabled.
49  * With frame pointers (to be compliant with the ABI):
50  *
51  *                              high
52  * original ARM_SP =>     +--------------+ \
53  *                        |      pc      | |
54  * current ARM_FP =>      +--------------+ } callee saved registers
55  *                        |r4-r9,fp,ip,lr| |
56  *                        +--------------+ /
57  *                              low
58  *
59  * Without frame pointers:
60  *
61  *                              high
62  * original ARM_SP =>     +--------------+
63  *                        |  r4-r9,fp,lr | callee saved registers
64  * current ARM_FP =>      +--------------+
65  *                              low
66  *
67  * When popping registers off the stack at the end of a BPF function, we
68  * reference them via the current ARM_FP register.
69  */
70 #define CALLEE_MASK	(1 << ARM_R4 | 1 << ARM_R5 | 1 << ARM_R6 | \
71 			 1 << ARM_R7 | 1 << ARM_R8 | 1 << ARM_R9 | \
72 			 1 << ARM_FP)
73 #define CALLEE_PUSH_MASK (CALLEE_MASK | 1 << ARM_LR)
74 #define CALLEE_POP_MASK  (CALLEE_MASK | 1 << ARM_PC)
75 
76 enum {
77 	/* Stack layout - these are offsets from (top of stack - 4) */
78 	BPF_R2_HI,
79 	BPF_R2_LO,
80 	BPF_R3_HI,
81 	BPF_R3_LO,
82 	BPF_R4_HI,
83 	BPF_R4_LO,
84 	BPF_R5_HI,
85 	BPF_R5_LO,
86 	BPF_R7_HI,
87 	BPF_R7_LO,
88 	BPF_R8_HI,
89 	BPF_R8_LO,
90 	BPF_R9_HI,
91 	BPF_R9_LO,
92 	BPF_FP_HI,
93 	BPF_FP_LO,
94 	BPF_TC_HI,
95 	BPF_TC_LO,
96 	BPF_AX_HI,
97 	BPF_AX_LO,
98 	/* Stack space for BPF_REG_2, BPF_REG_3, BPF_REG_4,
99 	 * BPF_REG_5, BPF_REG_7, BPF_REG_8, BPF_REG_9,
100 	 * BPF_REG_FP and Tail call counts.
101 	 */
102 	BPF_JIT_SCRATCH_REGS,
103 };
104 
105 /*
106  * Negative "register" values indicate the register is stored on the stack
107  * and are the offset from the top of the eBPF JIT scratch space.
108  */
109 #define STACK_OFFSET(k)	(-4 - (k) * 4)
110 #define SCRATCH_SIZE	(BPF_JIT_SCRATCH_REGS * 4)
111 
112 #ifdef CONFIG_FRAME_POINTER
113 #define EBPF_SCRATCH_TO_ARM_FP(x) ((x) - 4 * hweight16(CALLEE_PUSH_MASK) - 4)
114 #else
115 #define EBPF_SCRATCH_TO_ARM_FP(x) (x)
116 #endif
117 
118 #define TMP_REG_1	(MAX_BPF_JIT_REG + 0)	/* TEMP Register 1 */
119 #define TMP_REG_2	(MAX_BPF_JIT_REG + 1)	/* TEMP Register 2 */
120 #define TCALL_CNT	(MAX_BPF_JIT_REG + 2)	/* Tail Call Count */
121 
122 #define FLAG_IMM_OVERFLOW	(1 << 0)
123 
124 /*
125  * Map eBPF registers to ARM 32bit registers or stack scratch space.
126  *
127  * 1. First argument is passed using the arm 32bit registers and rest of the
128  * arguments are passed on stack scratch space.
129  * 2. First callee-saved argument is mapped to arm 32 bit registers and rest
130  * arguments are mapped to scratch space on stack.
131  * 3. We need two 64 bit temp registers to do complex operations on eBPF
132  * registers.
133  *
134  * As the eBPF registers are all 64 bit registers and arm has only 32 bit
135  * registers, we have to map each eBPF registers with two arm 32 bit regs or
136  * scratch memory space and we have to build eBPF 64 bit register from those.
137  *
138  */
139 static const s8 bpf2a32[][2] = {
140 	/* return value from in-kernel function, and exit value from eBPF */
141 	[BPF_REG_0] = {ARM_R1, ARM_R0},
142 	/* arguments from eBPF program to in-kernel function */
143 	[BPF_REG_1] = {ARM_R3, ARM_R2},
144 	/* Stored on stack scratch space */
145 	[BPF_REG_2] = {STACK_OFFSET(BPF_R2_HI), STACK_OFFSET(BPF_R2_LO)},
146 	[BPF_REG_3] = {STACK_OFFSET(BPF_R3_HI), STACK_OFFSET(BPF_R3_LO)},
147 	[BPF_REG_4] = {STACK_OFFSET(BPF_R4_HI), STACK_OFFSET(BPF_R4_LO)},
148 	[BPF_REG_5] = {STACK_OFFSET(BPF_R5_HI), STACK_OFFSET(BPF_R5_LO)},
149 	/* callee saved registers that in-kernel function will preserve */
150 	[BPF_REG_6] = {ARM_R5, ARM_R4},
151 	/* Stored on stack scratch space */
152 	[BPF_REG_7] = {STACK_OFFSET(BPF_R7_HI), STACK_OFFSET(BPF_R7_LO)},
153 	[BPF_REG_8] = {STACK_OFFSET(BPF_R8_HI), STACK_OFFSET(BPF_R8_LO)},
154 	[BPF_REG_9] = {STACK_OFFSET(BPF_R9_HI), STACK_OFFSET(BPF_R9_LO)},
155 	/* Read only Frame Pointer to access Stack */
156 	[BPF_REG_FP] = {STACK_OFFSET(BPF_FP_HI), STACK_OFFSET(BPF_FP_LO)},
157 	/* Temporary Register for internal BPF JIT, can be used
158 	 * for constant blindings and others.
159 	 */
160 	[TMP_REG_1] = {ARM_R7, ARM_R6},
161 	[TMP_REG_2] = {ARM_R9, ARM_R8},
162 	/* Tail call count. Stored on stack scratch space. */
163 	[TCALL_CNT] = {STACK_OFFSET(BPF_TC_HI), STACK_OFFSET(BPF_TC_LO)},
164 	/* temporary register for blinding constants.
165 	 * Stored on stack scratch space.
166 	 */
167 	[BPF_REG_AX] = {STACK_OFFSET(BPF_AX_HI), STACK_OFFSET(BPF_AX_LO)},
168 };
169 
170 #define	dst_lo	dst[1]
171 #define dst_hi	dst[0]
172 #define src_lo	src[1]
173 #define src_hi	src[0]
174 
175 /*
176  * JIT Context:
177  *
178  * prog			:	bpf_prog
179  * idx			:	index of current last JITed instruction.
180  * prologue_bytes	:	bytes used in prologue.
181  * epilogue_offset	:	offset of epilogue starting.
182  * offsets		:	array of eBPF instruction offsets in
183  *				JITed code.
184  * target		:	final JITed code.
185  * epilogue_bytes	:	no of bytes used in epilogue.
186  * imm_count		:	no of immediate counts used for global
187  *				variables.
188  * imms			:	array of global variable addresses.
189  */
190 
191 struct jit_ctx {
192 	const struct bpf_prog *prog;
193 	unsigned int idx;
194 	unsigned int prologue_bytes;
195 	unsigned int epilogue_offset;
196 	unsigned int cpu_architecture;
197 	u32 flags;
198 	u32 *offsets;
199 	u32 *target;
200 	u32 stack_size;
201 #if __LINUX_ARM_ARCH__ < 7
202 	u16 epilogue_bytes;
203 	u16 imm_count;
204 	u32 *imms;
205 #endif
206 };
207 
208 /*
209  * Wrappers which handle both OABI and EABI and assures Thumb2 interworking
210  * (where the assembly routines like __aeabi_uidiv could cause problems).
211  */
212 static u32 jit_udiv32(u32 dividend, u32 divisor)
213 {
214 	return dividend / divisor;
215 }
216 
217 static u32 jit_mod32(u32 dividend, u32 divisor)
218 {
219 	return dividend % divisor;
220 }
221 
222 static inline void _emit(int cond, u32 inst, struct jit_ctx *ctx)
223 {
224 	inst |= (cond << 28);
225 	inst = __opcode_to_mem_arm(inst);
226 
227 	if (ctx->target != NULL)
228 		ctx->target[ctx->idx] = inst;
229 
230 	ctx->idx++;
231 }
232 
233 /*
234  * Emit an instruction that will be executed unconditionally.
235  */
236 static inline void emit(u32 inst, struct jit_ctx *ctx)
237 {
238 	_emit(ARM_COND_AL, inst, ctx);
239 }
240 
241 /*
242  * This is rather horrid, but necessary to convert an integer constant
243  * to an immediate operand for the opcodes, and be able to detect at
244  * build time whether the constant can't be converted (iow, usable in
245  * BUILD_BUG_ON()).
246  */
247 #define imm12val(v, s) (rol32(v, (s)) | (s) << 7)
248 #define const_imm8m(x)					\
249 	({ int r;					\
250 	   u32 v = (x);					\
251 	   if (!(v & ~0x000000ff))			\
252 		r = imm12val(v, 0);			\
253 	   else if (!(v & ~0xc000003f))			\
254 		r = imm12val(v, 2);			\
255 	   else if (!(v & ~0xf000000f))			\
256 		r = imm12val(v, 4);			\
257 	   else if (!(v & ~0xfc000003))			\
258 		r = imm12val(v, 6);			\
259 	   else if (!(v & ~0xff000000))			\
260 		r = imm12val(v, 8);			\
261 	   else if (!(v & ~0x3fc00000))			\
262 		r = imm12val(v, 10);			\
263 	   else if (!(v & ~0x0ff00000))			\
264 		r = imm12val(v, 12);			\
265 	   else if (!(v & ~0x03fc0000))			\
266 		r = imm12val(v, 14);			\
267 	   else if (!(v & ~0x00ff0000))			\
268 		r = imm12val(v, 16);			\
269 	   else if (!(v & ~0x003fc000))			\
270 		r = imm12val(v, 18);			\
271 	   else if (!(v & ~0x000ff000))			\
272 		r = imm12val(v, 20);			\
273 	   else if (!(v & ~0x0003fc00))			\
274 		r = imm12val(v, 22);			\
275 	   else if (!(v & ~0x0000ff00))			\
276 		r = imm12val(v, 24);			\
277 	   else if (!(v & ~0x00003fc0))			\
278 		r = imm12val(v, 26);			\
279 	   else if (!(v & ~0x00000ff0))			\
280 		r = imm12val(v, 28);			\
281 	   else if (!(v & ~0x000003fc))			\
282 		r = imm12val(v, 30);			\
283 	   else						\
284 		r = -1;					\
285 	   r; })
286 
287 /*
288  * Checks if immediate value can be converted to imm12(12 bits) value.
289  */
290 static int imm8m(u32 x)
291 {
292 	u32 rot;
293 
294 	for (rot = 0; rot < 16; rot++)
295 		if ((x & ~ror32(0xff, 2 * rot)) == 0)
296 			return rol32(x, 2 * rot) | (rot << 8);
297 	return -1;
298 }
299 
300 #define imm8m(x) (__builtin_constant_p(x) ? const_imm8m(x) : imm8m(x))
301 
302 static u32 arm_bpf_ldst_imm12(u32 op, u8 rt, u8 rn, s16 imm12)
303 {
304 	op |= rt << 12 | rn << 16;
305 	if (imm12 >= 0)
306 		op |= ARM_INST_LDST__U;
307 	else
308 		imm12 = -imm12;
309 	return op | (imm12 & ARM_INST_LDST__IMM12);
310 }
311 
312 static u32 arm_bpf_ldst_imm8(u32 op, u8 rt, u8 rn, s16 imm8)
313 {
314 	op |= rt << 12 | rn << 16;
315 	if (imm8 >= 0)
316 		op |= ARM_INST_LDST__U;
317 	else
318 		imm8 = -imm8;
319 	return op | (imm8 & 0xf0) << 4 | (imm8 & 0x0f);
320 }
321 
322 #define ARM_LDR_I(rt, rn, off)	arm_bpf_ldst_imm12(ARM_INST_LDR_I, rt, rn, off)
323 #define ARM_LDRB_I(rt, rn, off)	arm_bpf_ldst_imm12(ARM_INST_LDRB_I, rt, rn, off)
324 #define ARM_LDRD_I(rt, rn, off)	arm_bpf_ldst_imm8(ARM_INST_LDRD_I, rt, rn, off)
325 #define ARM_LDRH_I(rt, rn, off)	arm_bpf_ldst_imm8(ARM_INST_LDRH_I, rt, rn, off)
326 
327 #define ARM_STR_I(rt, rn, off)	arm_bpf_ldst_imm12(ARM_INST_STR_I, rt, rn, off)
328 #define ARM_STRB_I(rt, rn, off)	arm_bpf_ldst_imm12(ARM_INST_STRB_I, rt, rn, off)
329 #define ARM_STRD_I(rt, rn, off)	arm_bpf_ldst_imm8(ARM_INST_STRD_I, rt, rn, off)
330 #define ARM_STRH_I(rt, rn, off)	arm_bpf_ldst_imm8(ARM_INST_STRH_I, rt, rn, off)
331 
332 /*
333  * Initializes the JIT space with undefined instructions.
334  */
335 static void jit_fill_hole(void *area, unsigned int size)
336 {
337 	u32 *ptr;
338 	/* We are guaranteed to have aligned memory. */
339 	for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
340 		*ptr++ = __opcode_to_mem_arm(ARM_INST_UDF);
341 }
342 
343 #if defined(CONFIG_AEABI) && (__LINUX_ARM_ARCH__ >= 5)
344 /* EABI requires the stack to be aligned to 64-bit boundaries */
345 #define STACK_ALIGNMENT	8
346 #else
347 /* Stack must be aligned to 32-bit boundaries */
348 #define STACK_ALIGNMENT	4
349 #endif
350 
351 /* total stack size used in JITed code */
352 #define _STACK_SIZE	(ctx->prog->aux->stack_depth + SCRATCH_SIZE)
353 #define STACK_SIZE	ALIGN(_STACK_SIZE, STACK_ALIGNMENT)
354 
355 #if __LINUX_ARM_ARCH__ < 7
356 
357 static u16 imm_offset(u32 k, struct jit_ctx *ctx)
358 {
359 	unsigned int i = 0, offset;
360 	u16 imm;
361 
362 	/* on the "fake" run we just count them (duplicates included) */
363 	if (ctx->target == NULL) {
364 		ctx->imm_count++;
365 		return 0;
366 	}
367 
368 	while ((i < ctx->imm_count) && ctx->imms[i]) {
369 		if (ctx->imms[i] == k)
370 			break;
371 		i++;
372 	}
373 
374 	if (ctx->imms[i] == 0)
375 		ctx->imms[i] = k;
376 
377 	/* constants go just after the epilogue */
378 	offset =  ctx->offsets[ctx->prog->len - 1] * 4;
379 	offset += ctx->prologue_bytes;
380 	offset += ctx->epilogue_bytes;
381 	offset += i * 4;
382 
383 	ctx->target[offset / 4] = k;
384 
385 	/* PC in ARM mode == address of the instruction + 8 */
386 	imm = offset - (8 + ctx->idx * 4);
387 
388 	if (imm & ~0xfff) {
389 		/*
390 		 * literal pool is too far, signal it into flags. we
391 		 * can only detect it on the second pass unfortunately.
392 		 */
393 		ctx->flags |= FLAG_IMM_OVERFLOW;
394 		return 0;
395 	}
396 
397 	return imm;
398 }
399 
400 #endif /* __LINUX_ARM_ARCH__ */
401 
402 static inline int bpf2a32_offset(int bpf_to, int bpf_from,
403 				 const struct jit_ctx *ctx) {
404 	int to, from;
405 
406 	if (ctx->target == NULL)
407 		return 0;
408 	to = ctx->offsets[bpf_to];
409 	from = ctx->offsets[bpf_from];
410 
411 	return to - from - 1;
412 }
413 
414 /*
415  * Move an immediate that's not an imm8m to a core register.
416  */
417 static inline void emit_mov_i_no8m(const u8 rd, u32 val, struct jit_ctx *ctx)
418 {
419 #if __LINUX_ARM_ARCH__ < 7
420 	emit(ARM_LDR_I(rd, ARM_PC, imm_offset(val, ctx)), ctx);
421 #else
422 	emit(ARM_MOVW(rd, val & 0xffff), ctx);
423 	if (val > 0xffff)
424 		emit(ARM_MOVT(rd, val >> 16), ctx);
425 #endif
426 }
427 
428 static inline void emit_mov_i(const u8 rd, u32 val, struct jit_ctx *ctx)
429 {
430 	int imm12 = imm8m(val);
431 
432 	if (imm12 >= 0)
433 		emit(ARM_MOV_I(rd, imm12), ctx);
434 	else
435 		emit_mov_i_no8m(rd, val, ctx);
436 }
437 
438 static void emit_bx_r(u8 tgt_reg, struct jit_ctx *ctx)
439 {
440 	if (elf_hwcap & HWCAP_THUMB)
441 		emit(ARM_BX(tgt_reg), ctx);
442 	else
443 		emit(ARM_MOV_R(ARM_PC, tgt_reg), ctx);
444 }
445 
446 static inline void emit_blx_r(u8 tgt_reg, struct jit_ctx *ctx)
447 {
448 #if __LINUX_ARM_ARCH__ < 5
449 	emit(ARM_MOV_R(ARM_LR, ARM_PC), ctx);
450 	emit_bx_r(tgt_reg, ctx);
451 #else
452 	emit(ARM_BLX_R(tgt_reg), ctx);
453 #endif
454 }
455 
456 static inline int epilogue_offset(const struct jit_ctx *ctx)
457 {
458 	int to, from;
459 	/* No need for 1st dummy run */
460 	if (ctx->target == NULL)
461 		return 0;
462 	to = ctx->epilogue_offset;
463 	from = ctx->idx;
464 
465 	return to - from - 2;
466 }
467 
468 static inline void emit_udivmod(u8 rd, u8 rm, u8 rn, struct jit_ctx *ctx, u8 op)
469 {
470 	const s8 *tmp = bpf2a32[TMP_REG_1];
471 
472 #if __LINUX_ARM_ARCH__ == 7
473 	if (elf_hwcap & HWCAP_IDIVA) {
474 		if (op == BPF_DIV)
475 			emit(ARM_UDIV(rd, rm, rn), ctx);
476 		else {
477 			emit(ARM_UDIV(ARM_IP, rm, rn), ctx);
478 			emit(ARM_MLS(rd, rn, ARM_IP, rm), ctx);
479 		}
480 		return;
481 	}
482 #endif
483 
484 	/*
485 	 * For BPF_ALU | BPF_DIV | BPF_K instructions
486 	 * As ARM_R1 and ARM_R0 contains 1st argument of bpf
487 	 * function, we need to save it on caller side to save
488 	 * it from getting destroyed within callee.
489 	 * After the return from the callee, we restore ARM_R0
490 	 * ARM_R1.
491 	 */
492 	if (rn != ARM_R1) {
493 		emit(ARM_MOV_R(tmp[0], ARM_R1), ctx);
494 		emit(ARM_MOV_R(ARM_R1, rn), ctx);
495 	}
496 	if (rm != ARM_R0) {
497 		emit(ARM_MOV_R(tmp[1], ARM_R0), ctx);
498 		emit(ARM_MOV_R(ARM_R0, rm), ctx);
499 	}
500 
501 	/* Call appropriate function */
502 	emit_mov_i(ARM_IP, op == BPF_DIV ?
503 		   (u32)jit_udiv32 : (u32)jit_mod32, ctx);
504 	emit_blx_r(ARM_IP, ctx);
505 
506 	/* Save return value */
507 	if (rd != ARM_R0)
508 		emit(ARM_MOV_R(rd, ARM_R0), ctx);
509 
510 	/* Restore ARM_R0 and ARM_R1 */
511 	if (rn != ARM_R1)
512 		emit(ARM_MOV_R(ARM_R1, tmp[0]), ctx);
513 	if (rm != ARM_R0)
514 		emit(ARM_MOV_R(ARM_R0, tmp[1]), ctx);
515 }
516 
517 /* Is the translated BPF register on stack? */
518 static bool is_stacked(s8 reg)
519 {
520 	return reg < 0;
521 }
522 
523 /* If a BPF register is on the stack (stk is true), load it to the
524  * supplied temporary register and return the temporary register
525  * for subsequent operations, otherwise just use the CPU register.
526  */
527 static s8 arm_bpf_get_reg32(s8 reg, s8 tmp, struct jit_ctx *ctx)
528 {
529 	if (is_stacked(reg)) {
530 		emit(ARM_LDR_I(tmp, ARM_FP, EBPF_SCRATCH_TO_ARM_FP(reg)), ctx);
531 		reg = tmp;
532 	}
533 	return reg;
534 }
535 
536 static const s8 *arm_bpf_get_reg64(const s8 *reg, const s8 *tmp,
537 				   struct jit_ctx *ctx)
538 {
539 	if (is_stacked(reg[1])) {
540 		if (__LINUX_ARM_ARCH__ >= 6 ||
541 		    ctx->cpu_architecture >= CPU_ARCH_ARMv5TE) {
542 			emit(ARM_LDRD_I(tmp[1], ARM_FP,
543 					EBPF_SCRATCH_TO_ARM_FP(reg[1])), ctx);
544 		} else {
545 			emit(ARM_LDR_I(tmp[1], ARM_FP,
546 				       EBPF_SCRATCH_TO_ARM_FP(reg[1])), ctx);
547 			emit(ARM_LDR_I(tmp[0], ARM_FP,
548 				       EBPF_SCRATCH_TO_ARM_FP(reg[0])), ctx);
549 		}
550 		reg = tmp;
551 	}
552 	return reg;
553 }
554 
555 /* If a BPF register is on the stack (stk is true), save the register
556  * back to the stack.  If the source register is not the same, then
557  * move it into the correct register.
558  */
559 static void arm_bpf_put_reg32(s8 reg, s8 src, struct jit_ctx *ctx)
560 {
561 	if (is_stacked(reg))
562 		emit(ARM_STR_I(src, ARM_FP, EBPF_SCRATCH_TO_ARM_FP(reg)), ctx);
563 	else if (reg != src)
564 		emit(ARM_MOV_R(reg, src), ctx);
565 }
566 
567 static void arm_bpf_put_reg64(const s8 *reg, const s8 *src,
568 			      struct jit_ctx *ctx)
569 {
570 	if (is_stacked(reg[1])) {
571 		if (__LINUX_ARM_ARCH__ >= 6 ||
572 		    ctx->cpu_architecture >= CPU_ARCH_ARMv5TE) {
573 			emit(ARM_STRD_I(src[1], ARM_FP,
574 				       EBPF_SCRATCH_TO_ARM_FP(reg[1])), ctx);
575 		} else {
576 			emit(ARM_STR_I(src[1], ARM_FP,
577 				       EBPF_SCRATCH_TO_ARM_FP(reg[1])), ctx);
578 			emit(ARM_STR_I(src[0], ARM_FP,
579 				       EBPF_SCRATCH_TO_ARM_FP(reg[0])), ctx);
580 		}
581 	} else {
582 		if (reg[1] != src[1])
583 			emit(ARM_MOV_R(reg[1], src[1]), ctx);
584 		if (reg[0] != src[0])
585 			emit(ARM_MOV_R(reg[0], src[0]), ctx);
586 	}
587 }
588 
589 static inline void emit_a32_mov_i(const s8 dst, const u32 val,
590 				  struct jit_ctx *ctx)
591 {
592 	const s8 *tmp = bpf2a32[TMP_REG_1];
593 
594 	if (is_stacked(dst)) {
595 		emit_mov_i(tmp[1], val, ctx);
596 		arm_bpf_put_reg32(dst, tmp[1], ctx);
597 	} else {
598 		emit_mov_i(dst, val, ctx);
599 	}
600 }
601 
602 static void emit_a32_mov_i64(const s8 dst[], u64 val, struct jit_ctx *ctx)
603 {
604 	const s8 *tmp = bpf2a32[TMP_REG_1];
605 	const s8 *rd = is_stacked(dst_lo) ? tmp : dst;
606 
607 	emit_mov_i(rd[1], (u32)val, ctx);
608 	emit_mov_i(rd[0], val >> 32, ctx);
609 
610 	arm_bpf_put_reg64(dst, rd, ctx);
611 }
612 
613 /* Sign extended move */
614 static inline void emit_a32_mov_se_i64(const bool is64, const s8 dst[],
615 				       const u32 val, struct jit_ctx *ctx) {
616 	u64 val64 = val;
617 
618 	if (is64 && (val & (1<<31)))
619 		val64 |= 0xffffffff00000000ULL;
620 	emit_a32_mov_i64(dst, val64, ctx);
621 }
622 
623 static inline void emit_a32_add_r(const u8 dst, const u8 src,
624 			      const bool is64, const bool hi,
625 			      struct jit_ctx *ctx) {
626 	/* 64 bit :
627 	 *	adds dst_lo, dst_lo, src_lo
628 	 *	adc dst_hi, dst_hi, src_hi
629 	 * 32 bit :
630 	 *	add dst_lo, dst_lo, src_lo
631 	 */
632 	if (!hi && is64)
633 		emit(ARM_ADDS_R(dst, dst, src), ctx);
634 	else if (hi && is64)
635 		emit(ARM_ADC_R(dst, dst, src), ctx);
636 	else
637 		emit(ARM_ADD_R(dst, dst, src), ctx);
638 }
639 
640 static inline void emit_a32_sub_r(const u8 dst, const u8 src,
641 				  const bool is64, const bool hi,
642 				  struct jit_ctx *ctx) {
643 	/* 64 bit :
644 	 *	subs dst_lo, dst_lo, src_lo
645 	 *	sbc dst_hi, dst_hi, src_hi
646 	 * 32 bit :
647 	 *	sub dst_lo, dst_lo, src_lo
648 	 */
649 	if (!hi && is64)
650 		emit(ARM_SUBS_R(dst, dst, src), ctx);
651 	else if (hi && is64)
652 		emit(ARM_SBC_R(dst, dst, src), ctx);
653 	else
654 		emit(ARM_SUB_R(dst, dst, src), ctx);
655 }
656 
657 static inline void emit_alu_r(const u8 dst, const u8 src, const bool is64,
658 			      const bool hi, const u8 op, struct jit_ctx *ctx){
659 	switch (BPF_OP(op)) {
660 	/* dst = dst + src */
661 	case BPF_ADD:
662 		emit_a32_add_r(dst, src, is64, hi, ctx);
663 		break;
664 	/* dst = dst - src */
665 	case BPF_SUB:
666 		emit_a32_sub_r(dst, src, is64, hi, ctx);
667 		break;
668 	/* dst = dst | src */
669 	case BPF_OR:
670 		emit(ARM_ORR_R(dst, dst, src), ctx);
671 		break;
672 	/* dst = dst & src */
673 	case BPF_AND:
674 		emit(ARM_AND_R(dst, dst, src), ctx);
675 		break;
676 	/* dst = dst ^ src */
677 	case BPF_XOR:
678 		emit(ARM_EOR_R(dst, dst, src), ctx);
679 		break;
680 	/* dst = dst * src */
681 	case BPF_MUL:
682 		emit(ARM_MUL(dst, dst, src), ctx);
683 		break;
684 	/* dst = dst << src */
685 	case BPF_LSH:
686 		emit(ARM_LSL_R(dst, dst, src), ctx);
687 		break;
688 	/* dst = dst >> src */
689 	case BPF_RSH:
690 		emit(ARM_LSR_R(dst, dst, src), ctx);
691 		break;
692 	/* dst = dst >> src (signed)*/
693 	case BPF_ARSH:
694 		emit(ARM_MOV_SR(dst, dst, SRTYPE_ASR, src), ctx);
695 		break;
696 	}
697 }
698 
699 /* ALU operation (32 bit)
700  * dst = dst (op) src
701  */
702 static inline void emit_a32_alu_r(const s8 dst, const s8 src,
703 				  struct jit_ctx *ctx, const bool is64,
704 				  const bool hi, const u8 op) {
705 	const s8 *tmp = bpf2a32[TMP_REG_1];
706 	s8 rn, rd;
707 
708 	rn = arm_bpf_get_reg32(src, tmp[1], ctx);
709 	rd = arm_bpf_get_reg32(dst, tmp[0], ctx);
710 	/* ALU operation */
711 	emit_alu_r(rd, rn, is64, hi, op, ctx);
712 	arm_bpf_put_reg32(dst, rd, ctx);
713 }
714 
715 /* ALU operation (64 bit) */
716 static inline void emit_a32_alu_r64(const bool is64, const s8 dst[],
717 				  const s8 src[], struct jit_ctx *ctx,
718 				  const u8 op) {
719 	const s8 *tmp = bpf2a32[TMP_REG_1];
720 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
721 	const s8 *rd;
722 
723 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
724 	if (is64) {
725 		const s8 *rs;
726 
727 		rs = arm_bpf_get_reg64(src, tmp2, ctx);
728 
729 		/* ALU operation */
730 		emit_alu_r(rd[1], rs[1], true, false, op, ctx);
731 		emit_alu_r(rd[0], rs[0], true, true, op, ctx);
732 	} else {
733 		s8 rs;
734 
735 		rs = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
736 
737 		/* ALU operation */
738 		emit_alu_r(rd[1], rs, true, false, op, ctx);
739 		emit_a32_mov_i(rd[0], 0, ctx);
740 	}
741 
742 	arm_bpf_put_reg64(dst, rd, ctx);
743 }
744 
745 /* dst = src (4 bytes)*/
746 static inline void emit_a32_mov_r(const s8 dst, const s8 src,
747 				  struct jit_ctx *ctx) {
748 	const s8 *tmp = bpf2a32[TMP_REG_1];
749 	s8 rt;
750 
751 	rt = arm_bpf_get_reg32(src, tmp[0], ctx);
752 	arm_bpf_put_reg32(dst, rt, ctx);
753 }
754 
755 /* dst = src */
756 static inline void emit_a32_mov_r64(const bool is64, const s8 dst[],
757 				  const s8 src[],
758 				  struct jit_ctx *ctx) {
759 	if (!is64) {
760 		emit_a32_mov_r(dst_lo, src_lo, ctx);
761 		/* Zero out high 4 bytes */
762 		emit_a32_mov_i(dst_hi, 0, ctx);
763 	} else if (__LINUX_ARM_ARCH__ < 6 &&
764 		   ctx->cpu_architecture < CPU_ARCH_ARMv5TE) {
765 		/* complete 8 byte move */
766 		emit_a32_mov_r(dst_lo, src_lo, ctx);
767 		emit_a32_mov_r(dst_hi, src_hi, ctx);
768 	} else if (is_stacked(src_lo) && is_stacked(dst_lo)) {
769 		const u8 *tmp = bpf2a32[TMP_REG_1];
770 
771 		emit(ARM_LDRD_I(tmp[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(src_lo)), ctx);
772 		emit(ARM_STRD_I(tmp[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(dst_lo)), ctx);
773 	} else if (is_stacked(src_lo)) {
774 		emit(ARM_LDRD_I(dst[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(src_lo)), ctx);
775 	} else if (is_stacked(dst_lo)) {
776 		emit(ARM_STRD_I(src[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(dst_lo)), ctx);
777 	} else {
778 		emit(ARM_MOV_R(dst[0], src[0]), ctx);
779 		emit(ARM_MOV_R(dst[1], src[1]), ctx);
780 	}
781 }
782 
783 /* Shift operations */
784 static inline void emit_a32_alu_i(const s8 dst, const u32 val,
785 				struct jit_ctx *ctx, const u8 op) {
786 	const s8 *tmp = bpf2a32[TMP_REG_1];
787 	s8 rd;
788 
789 	rd = arm_bpf_get_reg32(dst, tmp[0], ctx);
790 
791 	/* Do shift operation */
792 	switch (op) {
793 	case BPF_LSH:
794 		emit(ARM_LSL_I(rd, rd, val), ctx);
795 		break;
796 	case BPF_RSH:
797 		emit(ARM_LSR_I(rd, rd, val), ctx);
798 		break;
799 	case BPF_NEG:
800 		emit(ARM_RSB_I(rd, rd, val), ctx);
801 		break;
802 	}
803 
804 	arm_bpf_put_reg32(dst, rd, ctx);
805 }
806 
807 /* dst = ~dst (64 bit) */
808 static inline void emit_a32_neg64(const s8 dst[],
809 				struct jit_ctx *ctx){
810 	const s8 *tmp = bpf2a32[TMP_REG_1];
811 	const s8 *rd;
812 
813 	/* Setup Operand */
814 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
815 
816 	/* Do Negate Operation */
817 	emit(ARM_RSBS_I(rd[1], rd[1], 0), ctx);
818 	emit(ARM_RSC_I(rd[0], rd[0], 0), ctx);
819 
820 	arm_bpf_put_reg64(dst, rd, ctx);
821 }
822 
823 /* dst = dst << src */
824 static inline void emit_a32_lsh_r64(const s8 dst[], const s8 src[],
825 				    struct jit_ctx *ctx) {
826 	const s8 *tmp = bpf2a32[TMP_REG_1];
827 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
828 	const s8 *rd;
829 	s8 rt;
830 
831 	/* Setup Operands */
832 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
833 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
834 
835 	/* Do LSH operation */
836 	emit(ARM_SUB_I(ARM_IP, rt, 32), ctx);
837 	emit(ARM_RSB_I(tmp2[0], rt, 32), ctx);
838 	emit(ARM_MOV_SR(ARM_LR, rd[0], SRTYPE_ASL, rt), ctx);
839 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[1], SRTYPE_ASL, ARM_IP), ctx);
840 	emit(ARM_ORR_SR(ARM_IP, ARM_LR, rd[1], SRTYPE_LSR, tmp2[0]), ctx);
841 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_ASL, rt), ctx);
842 
843 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
844 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
845 }
846 
847 /* dst = dst >> src (signed)*/
848 static inline void emit_a32_arsh_r64(const s8 dst[], const s8 src[],
849 				     struct jit_ctx *ctx) {
850 	const s8 *tmp = bpf2a32[TMP_REG_1];
851 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
852 	const s8 *rd;
853 	s8 rt;
854 
855 	/* Setup Operands */
856 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
857 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
858 
859 	/* Do the ARSH operation */
860 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
861 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
862 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_LSR, rt), ctx);
863 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASL, ARM_IP), ctx);
864 	_emit(ARM_COND_MI, ARM_B(0), ctx);
865 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASR, tmp2[0]), ctx);
866 	emit(ARM_MOV_SR(ARM_IP, rd[0], SRTYPE_ASR, rt), ctx);
867 
868 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
869 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
870 }
871 
872 /* dst = dst >> src */
873 static inline void emit_a32_rsh_r64(const s8 dst[], const s8 src[],
874 				    struct jit_ctx *ctx) {
875 	const s8 *tmp = bpf2a32[TMP_REG_1];
876 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
877 	const s8 *rd;
878 	s8 rt;
879 
880 	/* Setup Operands */
881 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
882 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
883 
884 	/* Do RSH operation */
885 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
886 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
887 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_LSR, rt), ctx);
888 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASL, ARM_IP), ctx);
889 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_LSR, tmp2[0]), ctx);
890 	emit(ARM_MOV_SR(ARM_IP, rd[0], SRTYPE_LSR, rt), ctx);
891 
892 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
893 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
894 }
895 
896 /* dst = dst << val */
897 static inline void emit_a32_lsh_i64(const s8 dst[],
898 				    const u32 val, struct jit_ctx *ctx){
899 	const s8 *tmp = bpf2a32[TMP_REG_1];
900 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
901 	const s8 *rd;
902 
903 	/* Setup operands */
904 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
905 
906 	/* Do LSH operation */
907 	if (val < 32) {
908 		emit(ARM_MOV_SI(tmp2[0], rd[0], SRTYPE_ASL, val), ctx);
909 		emit(ARM_ORR_SI(rd[0], tmp2[0], rd[1], SRTYPE_LSR, 32 - val), ctx);
910 		emit(ARM_MOV_SI(rd[1], rd[1], SRTYPE_ASL, val), ctx);
911 	} else {
912 		if (val == 32)
913 			emit(ARM_MOV_R(rd[0], rd[1]), ctx);
914 		else
915 			emit(ARM_MOV_SI(rd[0], rd[1], SRTYPE_ASL, val - 32), ctx);
916 		emit(ARM_EOR_R(rd[1], rd[1], rd[1]), ctx);
917 	}
918 
919 	arm_bpf_put_reg64(dst, rd, ctx);
920 }
921 
922 /* dst = dst >> val */
923 static inline void emit_a32_rsh_i64(const s8 dst[],
924 				    const u32 val, struct jit_ctx *ctx) {
925 	const s8 *tmp = bpf2a32[TMP_REG_1];
926 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
927 	const s8 *rd;
928 
929 	/* Setup operands */
930 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
931 
932 	/* Do LSR operation */
933 	if (val < 32) {
934 		emit(ARM_MOV_SI(tmp2[1], rd[1], SRTYPE_LSR, val), ctx);
935 		emit(ARM_ORR_SI(rd[1], tmp2[1], rd[0], SRTYPE_ASL, 32 - val), ctx);
936 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_LSR, val), ctx);
937 	} else if (val == 32) {
938 		emit(ARM_MOV_R(rd[1], rd[0]), ctx);
939 		emit(ARM_MOV_I(rd[0], 0), ctx);
940 	} else {
941 		emit(ARM_MOV_SI(rd[1], rd[0], SRTYPE_LSR, val - 32), ctx);
942 		emit(ARM_MOV_I(rd[0], 0), ctx);
943 	}
944 
945 	arm_bpf_put_reg64(dst, rd, ctx);
946 }
947 
948 /* dst = dst >> val (signed) */
949 static inline void emit_a32_arsh_i64(const s8 dst[],
950 				     const u32 val, struct jit_ctx *ctx){
951 	const s8 *tmp = bpf2a32[TMP_REG_1];
952 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
953 	const s8 *rd;
954 
955 	/* Setup operands */
956 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
957 
958 	/* Do ARSH operation */
959 	if (val < 32) {
960 		emit(ARM_MOV_SI(tmp2[1], rd[1], SRTYPE_LSR, val), ctx);
961 		emit(ARM_ORR_SI(rd[1], tmp2[1], rd[0], SRTYPE_ASL, 32 - val), ctx);
962 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, val), ctx);
963 	} else if (val == 32) {
964 		emit(ARM_MOV_R(rd[1], rd[0]), ctx);
965 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, 31), ctx);
966 	} else {
967 		emit(ARM_MOV_SI(rd[1], rd[0], SRTYPE_ASR, val - 32), ctx);
968 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, 31), ctx);
969 	}
970 
971 	arm_bpf_put_reg64(dst, rd, ctx);
972 }
973 
974 static inline void emit_a32_mul_r64(const s8 dst[], const s8 src[],
975 				    struct jit_ctx *ctx) {
976 	const s8 *tmp = bpf2a32[TMP_REG_1];
977 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
978 	const s8 *rd, *rt;
979 
980 	/* Setup operands for multiplication */
981 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
982 	rt = arm_bpf_get_reg64(src, tmp2, ctx);
983 
984 	/* Do Multiplication */
985 	emit(ARM_MUL(ARM_IP, rd[1], rt[0]), ctx);
986 	emit(ARM_MUL(ARM_LR, rd[0], rt[1]), ctx);
987 	emit(ARM_ADD_R(ARM_LR, ARM_IP, ARM_LR), ctx);
988 
989 	emit(ARM_UMULL(ARM_IP, rd[0], rd[1], rt[1]), ctx);
990 	emit(ARM_ADD_R(rd[0], ARM_LR, rd[0]), ctx);
991 
992 	arm_bpf_put_reg32(dst_lo, ARM_IP, ctx);
993 	arm_bpf_put_reg32(dst_hi, rd[0], ctx);
994 }
995 
996 /* *(size *)(dst + off) = src */
997 static inline void emit_str_r(const s8 dst, const s8 src[],
998 			      s32 off, struct jit_ctx *ctx, const u8 sz){
999 	const s8 *tmp = bpf2a32[TMP_REG_1];
1000 	s32 off_max;
1001 	s8 rd;
1002 
1003 	rd = arm_bpf_get_reg32(dst, tmp[1], ctx);
1004 
1005 	if (sz == BPF_H)
1006 		off_max = 0xff;
1007 	else
1008 		off_max = 0xfff;
1009 
1010 	if (off < 0 || off > off_max) {
1011 		emit_a32_mov_i(tmp[0], off, ctx);
1012 		emit(ARM_ADD_R(tmp[0], tmp[0], rd), ctx);
1013 		rd = tmp[0];
1014 		off = 0;
1015 	}
1016 	switch (sz) {
1017 	case BPF_B:
1018 		/* Store a Byte */
1019 		emit(ARM_STRB_I(src_lo, rd, off), ctx);
1020 		break;
1021 	case BPF_H:
1022 		/* Store a HalfWord */
1023 		emit(ARM_STRH_I(src_lo, rd, off), ctx);
1024 		break;
1025 	case BPF_W:
1026 		/* Store a Word */
1027 		emit(ARM_STR_I(src_lo, rd, off), ctx);
1028 		break;
1029 	case BPF_DW:
1030 		/* Store a Double Word */
1031 		emit(ARM_STR_I(src_lo, rd, off), ctx);
1032 		emit(ARM_STR_I(src_hi, rd, off + 4), ctx);
1033 		break;
1034 	}
1035 }
1036 
1037 /* dst = *(size*)(src + off) */
1038 static inline void emit_ldx_r(const s8 dst[], const s8 src,
1039 			      s32 off, struct jit_ctx *ctx, const u8 sz){
1040 	const s8 *tmp = bpf2a32[TMP_REG_1];
1041 	const s8 *rd = is_stacked(dst_lo) ? tmp : dst;
1042 	s8 rm = src;
1043 	s32 off_max;
1044 
1045 	if (sz == BPF_H)
1046 		off_max = 0xff;
1047 	else
1048 		off_max = 0xfff;
1049 
1050 	if (off < 0 || off > off_max) {
1051 		emit_a32_mov_i(tmp[0], off, ctx);
1052 		emit(ARM_ADD_R(tmp[0], tmp[0], src), ctx);
1053 		rm = tmp[0];
1054 		off = 0;
1055 	} else if (rd[1] == rm) {
1056 		emit(ARM_MOV_R(tmp[0], rm), ctx);
1057 		rm = tmp[0];
1058 	}
1059 	switch (sz) {
1060 	case BPF_B:
1061 		/* Load a Byte */
1062 		emit(ARM_LDRB_I(rd[1], rm, off), ctx);
1063 		emit_a32_mov_i(rd[0], 0, ctx);
1064 		break;
1065 	case BPF_H:
1066 		/* Load a HalfWord */
1067 		emit(ARM_LDRH_I(rd[1], rm, off), ctx);
1068 		emit_a32_mov_i(rd[0], 0, ctx);
1069 		break;
1070 	case BPF_W:
1071 		/* Load a Word */
1072 		emit(ARM_LDR_I(rd[1], rm, off), ctx);
1073 		emit_a32_mov_i(rd[0], 0, ctx);
1074 		break;
1075 	case BPF_DW:
1076 		/* Load a Double Word */
1077 		emit(ARM_LDR_I(rd[1], rm, off), ctx);
1078 		emit(ARM_LDR_I(rd[0], rm, off + 4), ctx);
1079 		break;
1080 	}
1081 	arm_bpf_put_reg64(dst, rd, ctx);
1082 }
1083 
1084 /* Arithmatic Operation */
1085 static inline void emit_ar_r(const u8 rd, const u8 rt, const u8 rm,
1086 			     const u8 rn, struct jit_ctx *ctx, u8 op,
1087 			     bool is_jmp64) {
1088 	switch (op) {
1089 	case BPF_JSET:
1090 		if (is_jmp64) {
1091 			emit(ARM_AND_R(ARM_IP, rt, rn), ctx);
1092 			emit(ARM_AND_R(ARM_LR, rd, rm), ctx);
1093 			emit(ARM_ORRS_R(ARM_IP, ARM_LR, ARM_IP), ctx);
1094 		} else {
1095 			emit(ARM_ANDS_R(ARM_IP, rt, rn), ctx);
1096 		}
1097 		break;
1098 	case BPF_JEQ:
1099 	case BPF_JNE:
1100 	case BPF_JGT:
1101 	case BPF_JGE:
1102 	case BPF_JLE:
1103 	case BPF_JLT:
1104 		if (is_jmp64) {
1105 			emit(ARM_CMP_R(rd, rm), ctx);
1106 			/* Only compare low halve if high halve are equal. */
1107 			_emit(ARM_COND_EQ, ARM_CMP_R(rt, rn), ctx);
1108 		} else {
1109 			emit(ARM_CMP_R(rt, rn), ctx);
1110 		}
1111 		break;
1112 	case BPF_JSLE:
1113 	case BPF_JSGT:
1114 		emit(ARM_CMP_R(rn, rt), ctx);
1115 		if (is_jmp64)
1116 			emit(ARM_SBCS_R(ARM_IP, rm, rd), ctx);
1117 		break;
1118 	case BPF_JSLT:
1119 	case BPF_JSGE:
1120 		emit(ARM_CMP_R(rt, rn), ctx);
1121 		if (is_jmp64)
1122 			emit(ARM_SBCS_R(ARM_IP, rd, rm), ctx);
1123 		break;
1124 	}
1125 }
1126 
1127 static int out_offset = -1; /* initialized on the first pass of build_body() */
1128 static int emit_bpf_tail_call(struct jit_ctx *ctx)
1129 {
1130 
1131 	/* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
1132 	const s8 *r2 = bpf2a32[BPF_REG_2];
1133 	const s8 *r3 = bpf2a32[BPF_REG_3];
1134 	const s8 *tmp = bpf2a32[TMP_REG_1];
1135 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1136 	const s8 *tcc = bpf2a32[TCALL_CNT];
1137 	const s8 *tc;
1138 	const int idx0 = ctx->idx;
1139 #define cur_offset (ctx->idx - idx0)
1140 #define jmp_offset (out_offset - (cur_offset) - 2)
1141 	u32 lo, hi;
1142 	s8 r_array, r_index;
1143 	int off;
1144 
1145 	/* if (index >= array->map.max_entries)
1146 	 *	goto out;
1147 	 */
1148 	BUILD_BUG_ON(offsetof(struct bpf_array, map.max_entries) >
1149 		     ARM_INST_LDST__IMM12);
1150 	off = offsetof(struct bpf_array, map.max_entries);
1151 	r_array = arm_bpf_get_reg32(r2[1], tmp2[0], ctx);
1152 	/* index is 32-bit for arrays */
1153 	r_index = arm_bpf_get_reg32(r3[1], tmp2[1], ctx);
1154 	/* array->map.max_entries */
1155 	emit(ARM_LDR_I(tmp[1], r_array, off), ctx);
1156 	/* index >= array->map.max_entries */
1157 	emit(ARM_CMP_R(r_index, tmp[1]), ctx);
1158 	_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
1159 
1160 	/* tmp2[0] = array, tmp2[1] = index */
1161 
1162 	/* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
1163 	 *	goto out;
1164 	 * tail_call_cnt++;
1165 	 */
1166 	lo = (u32)MAX_TAIL_CALL_CNT;
1167 	hi = (u32)((u64)MAX_TAIL_CALL_CNT >> 32);
1168 	tc = arm_bpf_get_reg64(tcc, tmp, ctx);
1169 	emit(ARM_CMP_I(tc[0], hi), ctx);
1170 	_emit(ARM_COND_EQ, ARM_CMP_I(tc[1], lo), ctx);
1171 	_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
1172 	emit(ARM_ADDS_I(tc[1], tc[1], 1), ctx);
1173 	emit(ARM_ADC_I(tc[0], tc[0], 0), ctx);
1174 	arm_bpf_put_reg64(tcc, tmp, ctx);
1175 
1176 	/* prog = array->ptrs[index]
1177 	 * if (prog == NULL)
1178 	 *	goto out;
1179 	 */
1180 	BUILD_BUG_ON(imm8m(offsetof(struct bpf_array, ptrs)) < 0);
1181 	off = imm8m(offsetof(struct bpf_array, ptrs));
1182 	emit(ARM_ADD_I(tmp[1], r_array, off), ctx);
1183 	emit(ARM_LDR_R_SI(tmp[1], tmp[1], r_index, SRTYPE_ASL, 2), ctx);
1184 	emit(ARM_CMP_I(tmp[1], 0), ctx);
1185 	_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1186 
1187 	/* goto *(prog->bpf_func + prologue_size); */
1188 	BUILD_BUG_ON(offsetof(struct bpf_prog, bpf_func) >
1189 		     ARM_INST_LDST__IMM12);
1190 	off = offsetof(struct bpf_prog, bpf_func);
1191 	emit(ARM_LDR_I(tmp[1], tmp[1], off), ctx);
1192 	emit(ARM_ADD_I(tmp[1], tmp[1], ctx->prologue_bytes), ctx);
1193 	emit_bx_r(tmp[1], ctx);
1194 
1195 	/* out: */
1196 	if (out_offset == -1)
1197 		out_offset = cur_offset;
1198 	if (cur_offset != out_offset) {
1199 		pr_err_once("tail_call out_offset = %d, expected %d!\n",
1200 			    cur_offset, out_offset);
1201 		return -1;
1202 	}
1203 	return 0;
1204 #undef cur_offset
1205 #undef jmp_offset
1206 }
1207 
1208 /* 0xabcd => 0xcdab */
1209 static inline void emit_rev16(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1210 {
1211 #if __LINUX_ARM_ARCH__ < 6
1212 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1213 
1214 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1215 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 8), ctx);
1216 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1217 	emit(ARM_ORR_SI(rd, tmp2[0], tmp2[1], SRTYPE_LSL, 8), ctx);
1218 #else /* ARMv6+ */
1219 	emit(ARM_REV16(rd, rn), ctx);
1220 #endif
1221 }
1222 
1223 /* 0xabcdefgh => 0xghefcdab */
1224 static inline void emit_rev32(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1225 {
1226 #if __LINUX_ARM_ARCH__ < 6
1227 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1228 
1229 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1230 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 24), ctx);
1231 	emit(ARM_ORR_SI(ARM_IP, tmp2[0], tmp2[1], SRTYPE_LSL, 24), ctx);
1232 
1233 	emit(ARM_MOV_SI(tmp2[1], rn, SRTYPE_LSR, 8), ctx);
1234 	emit(ARM_AND_I(tmp2[1], tmp2[1], 0xff), ctx);
1235 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 16), ctx);
1236 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1237 	emit(ARM_MOV_SI(tmp2[0], tmp2[0], SRTYPE_LSL, 8), ctx);
1238 	emit(ARM_ORR_SI(tmp2[0], tmp2[0], tmp2[1], SRTYPE_LSL, 16), ctx);
1239 	emit(ARM_ORR_R(rd, ARM_IP, tmp2[0]), ctx);
1240 
1241 #else /* ARMv6+ */
1242 	emit(ARM_REV(rd, rn), ctx);
1243 #endif
1244 }
1245 
1246 // push the scratch stack register on top of the stack
1247 static inline void emit_push_r64(const s8 src[], struct jit_ctx *ctx)
1248 {
1249 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1250 	const s8 *rt;
1251 	u16 reg_set = 0;
1252 
1253 	rt = arm_bpf_get_reg64(src, tmp2, ctx);
1254 
1255 	reg_set = (1 << rt[1]) | (1 << rt[0]);
1256 	emit(ARM_PUSH(reg_set), ctx);
1257 }
1258 
1259 static void build_prologue(struct jit_ctx *ctx)
1260 {
1261 	const s8 r0 = bpf2a32[BPF_REG_0][1];
1262 	const s8 r2 = bpf2a32[BPF_REG_1][1];
1263 	const s8 r3 = bpf2a32[BPF_REG_1][0];
1264 	const s8 r4 = bpf2a32[BPF_REG_6][1];
1265 	const s8 fplo = bpf2a32[BPF_REG_FP][1];
1266 	const s8 fphi = bpf2a32[BPF_REG_FP][0];
1267 	const s8 *tcc = bpf2a32[TCALL_CNT];
1268 
1269 	/* Save callee saved registers. */
1270 #ifdef CONFIG_FRAME_POINTER
1271 	u16 reg_set = CALLEE_PUSH_MASK | 1 << ARM_IP | 1 << ARM_PC;
1272 	emit(ARM_MOV_R(ARM_IP, ARM_SP), ctx);
1273 	emit(ARM_PUSH(reg_set), ctx);
1274 	emit(ARM_SUB_I(ARM_FP, ARM_IP, 4), ctx);
1275 #else
1276 	emit(ARM_PUSH(CALLEE_PUSH_MASK), ctx);
1277 	emit(ARM_MOV_R(ARM_FP, ARM_SP), ctx);
1278 #endif
1279 	/* Save frame pointer for later */
1280 	emit(ARM_SUB_I(ARM_IP, ARM_SP, SCRATCH_SIZE), ctx);
1281 
1282 	ctx->stack_size = imm8m(STACK_SIZE);
1283 
1284 	/* Set up function call stack */
1285 	emit(ARM_SUB_I(ARM_SP, ARM_SP, ctx->stack_size), ctx);
1286 
1287 	/* Set up BPF prog stack base register */
1288 	emit_a32_mov_r(fplo, ARM_IP, ctx);
1289 	emit_a32_mov_i(fphi, 0, ctx);
1290 
1291 	/* mov r4, 0 */
1292 	emit(ARM_MOV_I(r4, 0), ctx);
1293 
1294 	/* Move BPF_CTX to BPF_R1 */
1295 	emit(ARM_MOV_R(r3, r4), ctx);
1296 	emit(ARM_MOV_R(r2, r0), ctx);
1297 	/* Initialize Tail Count */
1298 	emit(ARM_STR_I(r4, ARM_FP, EBPF_SCRATCH_TO_ARM_FP(tcc[0])), ctx);
1299 	emit(ARM_STR_I(r4, ARM_FP, EBPF_SCRATCH_TO_ARM_FP(tcc[1])), ctx);
1300 	/* end of prologue */
1301 }
1302 
1303 /* restore callee saved registers. */
1304 static void build_epilogue(struct jit_ctx *ctx)
1305 {
1306 #ifdef CONFIG_FRAME_POINTER
1307 	/* When using frame pointers, some additional registers need to
1308 	 * be loaded. */
1309 	u16 reg_set = CALLEE_POP_MASK | 1 << ARM_SP;
1310 	emit(ARM_SUB_I(ARM_SP, ARM_FP, hweight16(reg_set) * 4), ctx);
1311 	emit(ARM_LDM(ARM_SP, reg_set), ctx);
1312 #else
1313 	/* Restore callee saved registers. */
1314 	emit(ARM_MOV_R(ARM_SP, ARM_FP), ctx);
1315 	emit(ARM_POP(CALLEE_POP_MASK), ctx);
1316 #endif
1317 }
1318 
1319 /*
1320  * Convert an eBPF instruction to native instruction, i.e
1321  * JITs an eBPF instruction.
1322  * Returns :
1323  *	0  - Successfully JITed an 8-byte eBPF instruction
1324  *	>0 - Successfully JITed a 16-byte eBPF instruction
1325  *	<0 - Failed to JIT.
1326  */
1327 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx)
1328 {
1329 	const u8 code = insn->code;
1330 	const s8 *dst = bpf2a32[insn->dst_reg];
1331 	const s8 *src = bpf2a32[insn->src_reg];
1332 	const s8 *tmp = bpf2a32[TMP_REG_1];
1333 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1334 	const s16 off = insn->off;
1335 	const s32 imm = insn->imm;
1336 	const int i = insn - ctx->prog->insnsi;
1337 	const bool is64 = BPF_CLASS(code) == BPF_ALU64;
1338 	const s8 *rd, *rs;
1339 	s8 rd_lo, rt, rm, rn;
1340 	s32 jmp_offset;
1341 
1342 #define check_imm(bits, imm) do {				\
1343 	if ((imm) >= (1 << ((bits) - 1)) ||			\
1344 	    (imm) < -(1 << ((bits) - 1))) {			\
1345 		pr_info("[%2d] imm=%d(0x%x) out of range\n",	\
1346 			i, imm, imm);				\
1347 		return -EINVAL;					\
1348 	}							\
1349 } while (0)
1350 #define check_imm24(imm) check_imm(24, imm)
1351 
1352 	switch (code) {
1353 	/* ALU operations */
1354 
1355 	/* dst = src */
1356 	case BPF_ALU | BPF_MOV | BPF_K:
1357 	case BPF_ALU | BPF_MOV | BPF_X:
1358 	case BPF_ALU64 | BPF_MOV | BPF_K:
1359 	case BPF_ALU64 | BPF_MOV | BPF_X:
1360 		switch (BPF_SRC(code)) {
1361 		case BPF_X:
1362 			emit_a32_mov_r64(is64, dst, src, ctx);
1363 			break;
1364 		case BPF_K:
1365 			/* Sign-extend immediate value to destination reg */
1366 			emit_a32_mov_se_i64(is64, dst, imm, ctx);
1367 			break;
1368 		}
1369 		break;
1370 	/* dst = dst + src/imm */
1371 	/* dst = dst - src/imm */
1372 	/* dst = dst | src/imm */
1373 	/* dst = dst & src/imm */
1374 	/* dst = dst ^ src/imm */
1375 	/* dst = dst * src/imm */
1376 	/* dst = dst << src */
1377 	/* dst = dst >> src */
1378 	case BPF_ALU | BPF_ADD | BPF_K:
1379 	case BPF_ALU | BPF_ADD | BPF_X:
1380 	case BPF_ALU | BPF_SUB | BPF_K:
1381 	case BPF_ALU | BPF_SUB | BPF_X:
1382 	case BPF_ALU | BPF_OR | BPF_K:
1383 	case BPF_ALU | BPF_OR | BPF_X:
1384 	case BPF_ALU | BPF_AND | BPF_K:
1385 	case BPF_ALU | BPF_AND | BPF_X:
1386 	case BPF_ALU | BPF_XOR | BPF_K:
1387 	case BPF_ALU | BPF_XOR | BPF_X:
1388 	case BPF_ALU | BPF_MUL | BPF_K:
1389 	case BPF_ALU | BPF_MUL | BPF_X:
1390 	case BPF_ALU | BPF_LSH | BPF_X:
1391 	case BPF_ALU | BPF_RSH | BPF_X:
1392 	case BPF_ALU | BPF_ARSH | BPF_K:
1393 	case BPF_ALU | BPF_ARSH | BPF_X:
1394 	case BPF_ALU64 | BPF_ADD | BPF_K:
1395 	case BPF_ALU64 | BPF_ADD | BPF_X:
1396 	case BPF_ALU64 | BPF_SUB | BPF_K:
1397 	case BPF_ALU64 | BPF_SUB | BPF_X:
1398 	case BPF_ALU64 | BPF_OR | BPF_K:
1399 	case BPF_ALU64 | BPF_OR | BPF_X:
1400 	case BPF_ALU64 | BPF_AND | BPF_K:
1401 	case BPF_ALU64 | BPF_AND | BPF_X:
1402 	case BPF_ALU64 | BPF_XOR | BPF_K:
1403 	case BPF_ALU64 | BPF_XOR | BPF_X:
1404 		switch (BPF_SRC(code)) {
1405 		case BPF_X:
1406 			emit_a32_alu_r64(is64, dst, src, ctx, BPF_OP(code));
1407 			break;
1408 		case BPF_K:
1409 			/* Move immediate value to the temporary register
1410 			 * and then do the ALU operation on the temporary
1411 			 * register as this will sign-extend the immediate
1412 			 * value into temporary reg and then it would be
1413 			 * safe to do the operation on it.
1414 			 */
1415 			emit_a32_mov_se_i64(is64, tmp2, imm, ctx);
1416 			emit_a32_alu_r64(is64, dst, tmp2, ctx, BPF_OP(code));
1417 			break;
1418 		}
1419 		break;
1420 	/* dst = dst / src(imm) */
1421 	/* dst = dst % src(imm) */
1422 	case BPF_ALU | BPF_DIV | BPF_K:
1423 	case BPF_ALU | BPF_DIV | BPF_X:
1424 	case BPF_ALU | BPF_MOD | BPF_K:
1425 	case BPF_ALU | BPF_MOD | BPF_X:
1426 		rd_lo = arm_bpf_get_reg32(dst_lo, tmp2[1], ctx);
1427 		switch (BPF_SRC(code)) {
1428 		case BPF_X:
1429 			rt = arm_bpf_get_reg32(src_lo, tmp2[0], ctx);
1430 			break;
1431 		case BPF_K:
1432 			rt = tmp2[0];
1433 			emit_a32_mov_i(rt, imm, ctx);
1434 			break;
1435 		default:
1436 			rt = src_lo;
1437 			break;
1438 		}
1439 		emit_udivmod(rd_lo, rd_lo, rt, ctx, BPF_OP(code));
1440 		arm_bpf_put_reg32(dst_lo, rd_lo, ctx);
1441 		emit_a32_mov_i(dst_hi, 0, ctx);
1442 		break;
1443 	case BPF_ALU64 | BPF_DIV | BPF_K:
1444 	case BPF_ALU64 | BPF_DIV | BPF_X:
1445 	case BPF_ALU64 | BPF_MOD | BPF_K:
1446 	case BPF_ALU64 | BPF_MOD | BPF_X:
1447 		goto notyet;
1448 	/* dst = dst >> imm */
1449 	/* dst = dst << imm */
1450 	case BPF_ALU | BPF_RSH | BPF_K:
1451 	case BPF_ALU | BPF_LSH | BPF_K:
1452 		if (unlikely(imm > 31))
1453 			return -EINVAL;
1454 		if (imm)
1455 			emit_a32_alu_i(dst_lo, imm, ctx, BPF_OP(code));
1456 		emit_a32_mov_i(dst_hi, 0, ctx);
1457 		break;
1458 	/* dst = dst << imm */
1459 	case BPF_ALU64 | BPF_LSH | BPF_K:
1460 		if (unlikely(imm > 63))
1461 			return -EINVAL;
1462 		emit_a32_lsh_i64(dst, imm, ctx);
1463 		break;
1464 	/* dst = dst >> imm */
1465 	case BPF_ALU64 | BPF_RSH | BPF_K:
1466 		if (unlikely(imm > 63))
1467 			return -EINVAL;
1468 		emit_a32_rsh_i64(dst, imm, ctx);
1469 		break;
1470 	/* dst = dst << src */
1471 	case BPF_ALU64 | BPF_LSH | BPF_X:
1472 		emit_a32_lsh_r64(dst, src, ctx);
1473 		break;
1474 	/* dst = dst >> src */
1475 	case BPF_ALU64 | BPF_RSH | BPF_X:
1476 		emit_a32_rsh_r64(dst, src, ctx);
1477 		break;
1478 	/* dst = dst >> src (signed) */
1479 	case BPF_ALU64 | BPF_ARSH | BPF_X:
1480 		emit_a32_arsh_r64(dst, src, ctx);
1481 		break;
1482 	/* dst = dst >> imm (signed) */
1483 	case BPF_ALU64 | BPF_ARSH | BPF_K:
1484 		if (unlikely(imm > 63))
1485 			return -EINVAL;
1486 		emit_a32_arsh_i64(dst, imm, ctx);
1487 		break;
1488 	/* dst = ~dst */
1489 	case BPF_ALU | BPF_NEG:
1490 		emit_a32_alu_i(dst_lo, 0, ctx, BPF_OP(code));
1491 		emit_a32_mov_i(dst_hi, 0, ctx);
1492 		break;
1493 	/* dst = ~dst (64 bit) */
1494 	case BPF_ALU64 | BPF_NEG:
1495 		emit_a32_neg64(dst, ctx);
1496 		break;
1497 	/* dst = dst * src/imm */
1498 	case BPF_ALU64 | BPF_MUL | BPF_X:
1499 	case BPF_ALU64 | BPF_MUL | BPF_K:
1500 		switch (BPF_SRC(code)) {
1501 		case BPF_X:
1502 			emit_a32_mul_r64(dst, src, ctx);
1503 			break;
1504 		case BPF_K:
1505 			/* Move immediate value to the temporary register
1506 			 * and then do the multiplication on it as this
1507 			 * will sign-extend the immediate value into temp
1508 			 * reg then it would be safe to do the operation
1509 			 * on it.
1510 			 */
1511 			emit_a32_mov_se_i64(is64, tmp2, imm, ctx);
1512 			emit_a32_mul_r64(dst, tmp2, ctx);
1513 			break;
1514 		}
1515 		break;
1516 	/* dst = htole(dst) */
1517 	/* dst = htobe(dst) */
1518 	case BPF_ALU | BPF_END | BPF_FROM_LE:
1519 	case BPF_ALU | BPF_END | BPF_FROM_BE:
1520 		rd = arm_bpf_get_reg64(dst, tmp, ctx);
1521 		if (BPF_SRC(code) == BPF_FROM_LE)
1522 			goto emit_bswap_uxt;
1523 		switch (imm) {
1524 		case 16:
1525 			emit_rev16(rd[1], rd[1], ctx);
1526 			goto emit_bswap_uxt;
1527 		case 32:
1528 			emit_rev32(rd[1], rd[1], ctx);
1529 			goto emit_bswap_uxt;
1530 		case 64:
1531 			emit_rev32(ARM_LR, rd[1], ctx);
1532 			emit_rev32(rd[1], rd[0], ctx);
1533 			emit(ARM_MOV_R(rd[0], ARM_LR), ctx);
1534 			break;
1535 		}
1536 		goto exit;
1537 emit_bswap_uxt:
1538 		switch (imm) {
1539 		case 16:
1540 			/* zero-extend 16 bits into 64 bits */
1541 #if __LINUX_ARM_ARCH__ < 6
1542 			emit_a32_mov_i(tmp2[1], 0xffff, ctx);
1543 			emit(ARM_AND_R(rd[1], rd[1], tmp2[1]), ctx);
1544 #else /* ARMv6+ */
1545 			emit(ARM_UXTH(rd[1], rd[1]), ctx);
1546 #endif
1547 			emit(ARM_EOR_R(rd[0], rd[0], rd[0]), ctx);
1548 			break;
1549 		case 32:
1550 			/* zero-extend 32 bits into 64 bits */
1551 			emit(ARM_EOR_R(rd[0], rd[0], rd[0]), ctx);
1552 			break;
1553 		case 64:
1554 			/* nop */
1555 			break;
1556 		}
1557 exit:
1558 		arm_bpf_put_reg64(dst, rd, ctx);
1559 		break;
1560 	/* dst = imm64 */
1561 	case BPF_LD | BPF_IMM | BPF_DW:
1562 	{
1563 		u64 val = (u32)imm | (u64)insn[1].imm << 32;
1564 
1565 		emit_a32_mov_i64(dst, val, ctx);
1566 
1567 		return 1;
1568 	}
1569 	/* LDX: dst = *(size *)(src + off) */
1570 	case BPF_LDX | BPF_MEM | BPF_W:
1571 	case BPF_LDX | BPF_MEM | BPF_H:
1572 	case BPF_LDX | BPF_MEM | BPF_B:
1573 	case BPF_LDX | BPF_MEM | BPF_DW:
1574 		rn = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
1575 		emit_ldx_r(dst, rn, off, ctx, BPF_SIZE(code));
1576 		break;
1577 	/* ST: *(size *)(dst + off) = imm */
1578 	case BPF_ST | BPF_MEM | BPF_W:
1579 	case BPF_ST | BPF_MEM | BPF_H:
1580 	case BPF_ST | BPF_MEM | BPF_B:
1581 	case BPF_ST | BPF_MEM | BPF_DW:
1582 		switch (BPF_SIZE(code)) {
1583 		case BPF_DW:
1584 			/* Sign-extend immediate value into temp reg */
1585 			emit_a32_mov_se_i64(true, tmp2, imm, ctx);
1586 			break;
1587 		case BPF_W:
1588 		case BPF_H:
1589 		case BPF_B:
1590 			emit_a32_mov_i(tmp2[1], imm, ctx);
1591 			break;
1592 		}
1593 		emit_str_r(dst_lo, tmp2, off, ctx, BPF_SIZE(code));
1594 		break;
1595 	/* STX XADD: lock *(u32 *)(dst + off) += src */
1596 	case BPF_STX | BPF_XADD | BPF_W:
1597 	/* STX XADD: lock *(u64 *)(dst + off) += src */
1598 	case BPF_STX | BPF_XADD | BPF_DW:
1599 		goto notyet;
1600 	/* STX: *(size *)(dst + off) = src */
1601 	case BPF_STX | BPF_MEM | BPF_W:
1602 	case BPF_STX | BPF_MEM | BPF_H:
1603 	case BPF_STX | BPF_MEM | BPF_B:
1604 	case BPF_STX | BPF_MEM | BPF_DW:
1605 		rs = arm_bpf_get_reg64(src, tmp2, ctx);
1606 		emit_str_r(dst_lo, rs, off, ctx, BPF_SIZE(code));
1607 		break;
1608 	/* PC += off if dst == src */
1609 	/* PC += off if dst > src */
1610 	/* PC += off if dst >= src */
1611 	/* PC += off if dst < src */
1612 	/* PC += off if dst <= src */
1613 	/* PC += off if dst != src */
1614 	/* PC += off if dst > src (signed) */
1615 	/* PC += off if dst >= src (signed) */
1616 	/* PC += off if dst < src (signed) */
1617 	/* PC += off if dst <= src (signed) */
1618 	/* PC += off if dst & src */
1619 	case BPF_JMP | BPF_JEQ | BPF_X:
1620 	case BPF_JMP | BPF_JGT | BPF_X:
1621 	case BPF_JMP | BPF_JGE | BPF_X:
1622 	case BPF_JMP | BPF_JNE | BPF_X:
1623 	case BPF_JMP | BPF_JSGT | BPF_X:
1624 	case BPF_JMP | BPF_JSGE | BPF_X:
1625 	case BPF_JMP | BPF_JSET | BPF_X:
1626 	case BPF_JMP | BPF_JLE | BPF_X:
1627 	case BPF_JMP | BPF_JLT | BPF_X:
1628 	case BPF_JMP | BPF_JSLT | BPF_X:
1629 	case BPF_JMP | BPF_JSLE | BPF_X:
1630 	case BPF_JMP32 | BPF_JEQ | BPF_X:
1631 	case BPF_JMP32 | BPF_JGT | BPF_X:
1632 	case BPF_JMP32 | BPF_JGE | BPF_X:
1633 	case BPF_JMP32 | BPF_JNE | BPF_X:
1634 	case BPF_JMP32 | BPF_JSGT | BPF_X:
1635 	case BPF_JMP32 | BPF_JSGE | BPF_X:
1636 	case BPF_JMP32 | BPF_JSET | BPF_X:
1637 	case BPF_JMP32 | BPF_JLE | BPF_X:
1638 	case BPF_JMP32 | BPF_JLT | BPF_X:
1639 	case BPF_JMP32 | BPF_JSLT | BPF_X:
1640 	case BPF_JMP32 | BPF_JSLE | BPF_X:
1641 		/* Setup source registers */
1642 		rm = arm_bpf_get_reg32(src_hi, tmp2[0], ctx);
1643 		rn = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
1644 		goto go_jmp;
1645 	/* PC += off if dst == imm */
1646 	/* PC += off if dst > imm */
1647 	/* PC += off if dst >= imm */
1648 	/* PC += off if dst < imm */
1649 	/* PC += off if dst <= imm */
1650 	/* PC += off if dst != imm */
1651 	/* PC += off if dst > imm (signed) */
1652 	/* PC += off if dst >= imm (signed) */
1653 	/* PC += off if dst < imm (signed) */
1654 	/* PC += off if dst <= imm (signed) */
1655 	/* PC += off if dst & imm */
1656 	case BPF_JMP | BPF_JEQ | BPF_K:
1657 	case BPF_JMP | BPF_JGT | BPF_K:
1658 	case BPF_JMP | BPF_JGE | BPF_K:
1659 	case BPF_JMP | BPF_JNE | BPF_K:
1660 	case BPF_JMP | BPF_JSGT | BPF_K:
1661 	case BPF_JMP | BPF_JSGE | BPF_K:
1662 	case BPF_JMP | BPF_JSET | BPF_K:
1663 	case BPF_JMP | BPF_JLT | BPF_K:
1664 	case BPF_JMP | BPF_JLE | BPF_K:
1665 	case BPF_JMP | BPF_JSLT | BPF_K:
1666 	case BPF_JMP | BPF_JSLE | BPF_K:
1667 	case BPF_JMP32 | BPF_JEQ | BPF_K:
1668 	case BPF_JMP32 | BPF_JGT | BPF_K:
1669 	case BPF_JMP32 | BPF_JGE | BPF_K:
1670 	case BPF_JMP32 | BPF_JNE | BPF_K:
1671 	case BPF_JMP32 | BPF_JSGT | BPF_K:
1672 	case BPF_JMP32 | BPF_JSGE | BPF_K:
1673 	case BPF_JMP32 | BPF_JSET | BPF_K:
1674 	case BPF_JMP32 | BPF_JLT | BPF_K:
1675 	case BPF_JMP32 | BPF_JLE | BPF_K:
1676 	case BPF_JMP32 | BPF_JSLT | BPF_K:
1677 	case BPF_JMP32 | BPF_JSLE | BPF_K:
1678 		if (off == 0)
1679 			break;
1680 		rm = tmp2[0];
1681 		rn = tmp2[1];
1682 		/* Sign-extend immediate value */
1683 		emit_a32_mov_se_i64(true, tmp2, imm, ctx);
1684 go_jmp:
1685 		/* Setup destination register */
1686 		rd = arm_bpf_get_reg64(dst, tmp, ctx);
1687 
1688 		/* Check for the condition */
1689 		emit_ar_r(rd[0], rd[1], rm, rn, ctx, BPF_OP(code),
1690 			  BPF_CLASS(code) == BPF_JMP);
1691 
1692 		/* Setup JUMP instruction */
1693 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1694 		switch (BPF_OP(code)) {
1695 		case BPF_JNE:
1696 		case BPF_JSET:
1697 			_emit(ARM_COND_NE, ARM_B(jmp_offset), ctx);
1698 			break;
1699 		case BPF_JEQ:
1700 			_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1701 			break;
1702 		case BPF_JGT:
1703 			_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
1704 			break;
1705 		case BPF_JGE:
1706 			_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
1707 			break;
1708 		case BPF_JSGT:
1709 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1710 			break;
1711 		case BPF_JSGE:
1712 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1713 			break;
1714 		case BPF_JLE:
1715 			_emit(ARM_COND_LS, ARM_B(jmp_offset), ctx);
1716 			break;
1717 		case BPF_JLT:
1718 			_emit(ARM_COND_CC, ARM_B(jmp_offset), ctx);
1719 			break;
1720 		case BPF_JSLT:
1721 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1722 			break;
1723 		case BPF_JSLE:
1724 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1725 			break;
1726 		}
1727 		break;
1728 	/* JMP OFF */
1729 	case BPF_JMP | BPF_JA:
1730 	{
1731 		if (off == 0)
1732 			break;
1733 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1734 		check_imm24(jmp_offset);
1735 		emit(ARM_B(jmp_offset), ctx);
1736 		break;
1737 	}
1738 	/* tail call */
1739 	case BPF_JMP | BPF_TAIL_CALL:
1740 		if (emit_bpf_tail_call(ctx))
1741 			return -EFAULT;
1742 		break;
1743 	/* function call */
1744 	case BPF_JMP | BPF_CALL:
1745 	{
1746 		const s8 *r0 = bpf2a32[BPF_REG_0];
1747 		const s8 *r1 = bpf2a32[BPF_REG_1];
1748 		const s8 *r2 = bpf2a32[BPF_REG_2];
1749 		const s8 *r3 = bpf2a32[BPF_REG_3];
1750 		const s8 *r4 = bpf2a32[BPF_REG_4];
1751 		const s8 *r5 = bpf2a32[BPF_REG_5];
1752 		const u32 func = (u32)__bpf_call_base + (u32)imm;
1753 
1754 		emit_a32_mov_r64(true, r0, r1, ctx);
1755 		emit_a32_mov_r64(true, r1, r2, ctx);
1756 		emit_push_r64(r5, ctx);
1757 		emit_push_r64(r4, ctx);
1758 		emit_push_r64(r3, ctx);
1759 
1760 		emit_a32_mov_i(tmp[1], func, ctx);
1761 		emit_blx_r(tmp[1], ctx);
1762 
1763 		emit(ARM_ADD_I(ARM_SP, ARM_SP, imm8m(24)), ctx); // callee clean
1764 		break;
1765 	}
1766 	/* function return */
1767 	case BPF_JMP | BPF_EXIT:
1768 		/* Optimization: when last instruction is EXIT
1769 		 * simply fallthrough to epilogue.
1770 		 */
1771 		if (i == ctx->prog->len - 1)
1772 			break;
1773 		jmp_offset = epilogue_offset(ctx);
1774 		check_imm24(jmp_offset);
1775 		emit(ARM_B(jmp_offset), ctx);
1776 		break;
1777 notyet:
1778 		pr_info_once("*** NOT YET: opcode %02x ***\n", code);
1779 		return -EFAULT;
1780 	default:
1781 		pr_err_once("unknown opcode %02x\n", code);
1782 		return -EINVAL;
1783 	}
1784 
1785 	if (ctx->flags & FLAG_IMM_OVERFLOW)
1786 		/*
1787 		 * this instruction generated an overflow when
1788 		 * trying to access the literal pool, so
1789 		 * delegate this filter to the kernel interpreter.
1790 		 */
1791 		return -1;
1792 	return 0;
1793 }
1794 
1795 static int build_body(struct jit_ctx *ctx)
1796 {
1797 	const struct bpf_prog *prog = ctx->prog;
1798 	unsigned int i;
1799 
1800 	for (i = 0; i < prog->len; i++) {
1801 		const struct bpf_insn *insn = &(prog->insnsi[i]);
1802 		int ret;
1803 
1804 		ret = build_insn(insn, ctx);
1805 
1806 		/* It's used with loading the 64 bit immediate value. */
1807 		if (ret > 0) {
1808 			i++;
1809 			if (ctx->target == NULL)
1810 				ctx->offsets[i] = ctx->idx;
1811 			continue;
1812 		}
1813 
1814 		if (ctx->target == NULL)
1815 			ctx->offsets[i] = ctx->idx;
1816 
1817 		/* If unsuccesfull, return with error code */
1818 		if (ret)
1819 			return ret;
1820 	}
1821 	return 0;
1822 }
1823 
1824 static int validate_code(struct jit_ctx *ctx)
1825 {
1826 	int i;
1827 
1828 	for (i = 0; i < ctx->idx; i++) {
1829 		if (ctx->target[i] == __opcode_to_mem_arm(ARM_INST_UDF))
1830 			return -1;
1831 	}
1832 
1833 	return 0;
1834 }
1835 
1836 void bpf_jit_compile(struct bpf_prog *prog)
1837 {
1838 	/* Nothing to do here. We support Internal BPF. */
1839 }
1840 
1841 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1842 {
1843 	struct bpf_prog *tmp, *orig_prog = prog;
1844 	struct bpf_binary_header *header;
1845 	bool tmp_blinded = false;
1846 	struct jit_ctx ctx;
1847 	unsigned int tmp_idx;
1848 	unsigned int image_size;
1849 	u8 *image_ptr;
1850 
1851 	/* If BPF JIT was not enabled then we must fall back to
1852 	 * the interpreter.
1853 	 */
1854 	if (!prog->jit_requested)
1855 		return orig_prog;
1856 
1857 	/* If constant blinding was enabled and we failed during blinding
1858 	 * then we must fall back to the interpreter. Otherwise, we save
1859 	 * the new JITed code.
1860 	 */
1861 	tmp = bpf_jit_blind_constants(prog);
1862 
1863 	if (IS_ERR(tmp))
1864 		return orig_prog;
1865 	if (tmp != prog) {
1866 		tmp_blinded = true;
1867 		prog = tmp;
1868 	}
1869 
1870 	memset(&ctx, 0, sizeof(ctx));
1871 	ctx.prog = prog;
1872 	ctx.cpu_architecture = cpu_architecture();
1873 
1874 	/* Not able to allocate memory for offsets[] , then
1875 	 * we must fall back to the interpreter
1876 	 */
1877 	ctx.offsets = kcalloc(prog->len, sizeof(int), GFP_KERNEL);
1878 	if (ctx.offsets == NULL) {
1879 		prog = orig_prog;
1880 		goto out;
1881 	}
1882 
1883 	/* 1) fake pass to find in the length of the JITed code,
1884 	 * to compute ctx->offsets and other context variables
1885 	 * needed to compute final JITed code.
1886 	 * Also, calculate random starting pointer/start of JITed code
1887 	 * which is prefixed by random number of fault instructions.
1888 	 *
1889 	 * If the first pass fails then there is no chance of it
1890 	 * being successful in the second pass, so just fall back
1891 	 * to the interpreter.
1892 	 */
1893 	if (build_body(&ctx)) {
1894 		prog = orig_prog;
1895 		goto out_off;
1896 	}
1897 
1898 	tmp_idx = ctx.idx;
1899 	build_prologue(&ctx);
1900 	ctx.prologue_bytes = (ctx.idx - tmp_idx) * 4;
1901 
1902 	ctx.epilogue_offset = ctx.idx;
1903 
1904 #if __LINUX_ARM_ARCH__ < 7
1905 	tmp_idx = ctx.idx;
1906 	build_epilogue(&ctx);
1907 	ctx.epilogue_bytes = (ctx.idx - tmp_idx) * 4;
1908 
1909 	ctx.idx += ctx.imm_count;
1910 	if (ctx.imm_count) {
1911 		ctx.imms = kcalloc(ctx.imm_count, sizeof(u32), GFP_KERNEL);
1912 		if (ctx.imms == NULL) {
1913 			prog = orig_prog;
1914 			goto out_off;
1915 		}
1916 	}
1917 #else
1918 	/* there's nothing about the epilogue on ARMv7 */
1919 	build_epilogue(&ctx);
1920 #endif
1921 	/* Now we can get the actual image size of the JITed arm code.
1922 	 * Currently, we are not considering the THUMB-2 instructions
1923 	 * for jit, although it can decrease the size of the image.
1924 	 *
1925 	 * As each arm instruction is of length 32bit, we are translating
1926 	 * number of JITed intructions into the size required to store these
1927 	 * JITed code.
1928 	 */
1929 	image_size = sizeof(u32) * ctx.idx;
1930 
1931 	/* Now we know the size of the structure to make */
1932 	header = bpf_jit_binary_alloc(image_size, &image_ptr,
1933 				      sizeof(u32), jit_fill_hole);
1934 	/* Not able to allocate memory for the structure then
1935 	 * we must fall back to the interpretation
1936 	 */
1937 	if (header == NULL) {
1938 		prog = orig_prog;
1939 		goto out_imms;
1940 	}
1941 
1942 	/* 2.) Actual pass to generate final JIT code */
1943 	ctx.target = (u32 *) image_ptr;
1944 	ctx.idx = 0;
1945 
1946 	build_prologue(&ctx);
1947 
1948 	/* If building the body of the JITed code fails somehow,
1949 	 * we fall back to the interpretation.
1950 	 */
1951 	if (build_body(&ctx) < 0) {
1952 		image_ptr = NULL;
1953 		bpf_jit_binary_free(header);
1954 		prog = orig_prog;
1955 		goto out_imms;
1956 	}
1957 	build_epilogue(&ctx);
1958 
1959 	/* 3.) Extra pass to validate JITed Code */
1960 	if (validate_code(&ctx)) {
1961 		image_ptr = NULL;
1962 		bpf_jit_binary_free(header);
1963 		prog = orig_prog;
1964 		goto out_imms;
1965 	}
1966 	flush_icache_range((u32)header, (u32)(ctx.target + ctx.idx));
1967 
1968 	if (bpf_jit_enable > 1)
1969 		/* there are 2 passes here */
1970 		bpf_jit_dump(prog->len, image_size, 2, ctx.target);
1971 
1972 	bpf_jit_binary_lock_ro(header);
1973 	prog->bpf_func = (void *)ctx.target;
1974 	prog->jited = 1;
1975 	prog->jited_len = image_size;
1976 
1977 out_imms:
1978 #if __LINUX_ARM_ARCH__ < 7
1979 	if (ctx.imm_count)
1980 		kfree(ctx.imms);
1981 #endif
1982 out_off:
1983 	kfree(ctx.offsets);
1984 out:
1985 	if (tmp_blinded)
1986 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
1987 					   tmp : orig_prog);
1988 	return prog;
1989 }
1990 
1991