xref: /linux/arch/arm/net/bpf_jit_32.c (revision ca64d84e93762f4e587e040a44ad9f6089afc777)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Just-In-Time compiler for eBPF filters on 32bit ARM
4  *
5  * Copyright (c) 2017 Shubham Bansal <illusionist.neo@gmail.com>
6  * Copyright (c) 2011 Mircea Gherzan <mgherzan@gmail.com>
7  */
8 
9 #include <linux/bpf.h>
10 #include <linux/bitops.h>
11 #include <linux/compiler.h>
12 #include <linux/errno.h>
13 #include <linux/filter.h>
14 #include <linux/netdevice.h>
15 #include <linux/string.h>
16 #include <linux/slab.h>
17 #include <linux/if_vlan.h>
18 
19 #include <asm/cacheflush.h>
20 #include <asm/hwcap.h>
21 #include <asm/opcodes.h>
22 #include <asm/system_info.h>
23 
24 #include "bpf_jit_32.h"
25 
26 /*
27  * eBPF prog stack layout:
28  *
29  *                         high
30  * original ARM_SP =>     +-----+
31  *                        |     | callee saved registers
32  *                        +-----+ <= (BPF_FP + SCRATCH_SIZE)
33  *                        | ... | eBPF JIT scratch space
34  * eBPF fp register =>    +-----+
35  *   (BPF_FP)             | ... | eBPF prog stack
36  *                        +-----+
37  *                        |RSVD | JIT scratchpad
38  * current ARM_SP =>      +-----+ <= (BPF_FP - STACK_SIZE + SCRATCH_SIZE)
39  *                        |     |
40  *                        | ... | Function call stack
41  *                        |     |
42  *                        +-----+
43  *                          low
44  *
45  * The callee saved registers depends on whether frame pointers are enabled.
46  * With frame pointers (to be compliant with the ABI):
47  *
48  *                              high
49  * original ARM_SP =>     +--------------+ \
50  *                        |      pc      | |
51  * current ARM_FP =>      +--------------+ } callee saved registers
52  *                        |r4-r9,fp,ip,lr| |
53  *                        +--------------+ /
54  *                              low
55  *
56  * Without frame pointers:
57  *
58  *                              high
59  * original ARM_SP =>     +--------------+
60  *                        |  r4-r9,fp,lr | callee saved registers
61  * current ARM_FP =>      +--------------+
62  *                              low
63  *
64  * When popping registers off the stack at the end of a BPF function, we
65  * reference them via the current ARM_FP register.
66  */
67 #define CALLEE_MASK	(1 << ARM_R4 | 1 << ARM_R5 | 1 << ARM_R6 | \
68 			 1 << ARM_R7 | 1 << ARM_R8 | 1 << ARM_R9 | \
69 			 1 << ARM_FP)
70 #define CALLEE_PUSH_MASK (CALLEE_MASK | 1 << ARM_LR)
71 #define CALLEE_POP_MASK  (CALLEE_MASK | 1 << ARM_PC)
72 
73 enum {
74 	/* Stack layout - these are offsets from (top of stack - 4) */
75 	BPF_R2_HI,
76 	BPF_R2_LO,
77 	BPF_R3_HI,
78 	BPF_R3_LO,
79 	BPF_R4_HI,
80 	BPF_R4_LO,
81 	BPF_R5_HI,
82 	BPF_R5_LO,
83 	BPF_R7_HI,
84 	BPF_R7_LO,
85 	BPF_R8_HI,
86 	BPF_R8_LO,
87 	BPF_R9_HI,
88 	BPF_R9_LO,
89 	BPF_FP_HI,
90 	BPF_FP_LO,
91 	BPF_TC_HI,
92 	BPF_TC_LO,
93 	BPF_AX_HI,
94 	BPF_AX_LO,
95 	/* Stack space for BPF_REG_2, BPF_REG_3, BPF_REG_4,
96 	 * BPF_REG_5, BPF_REG_7, BPF_REG_8, BPF_REG_9,
97 	 * BPF_REG_FP and Tail call counts.
98 	 */
99 	BPF_JIT_SCRATCH_REGS,
100 };
101 
102 /*
103  * Negative "register" values indicate the register is stored on the stack
104  * and are the offset from the top of the eBPF JIT scratch space.
105  */
106 #define STACK_OFFSET(k)	(-4 - (k) * 4)
107 #define SCRATCH_SIZE	(BPF_JIT_SCRATCH_REGS * 4)
108 
109 #ifdef CONFIG_FRAME_POINTER
110 #define EBPF_SCRATCH_TO_ARM_FP(x) ((x) - 4 * hweight16(CALLEE_PUSH_MASK) - 4)
111 #else
112 #define EBPF_SCRATCH_TO_ARM_FP(x) (x)
113 #endif
114 
115 #define TMP_REG_1	(MAX_BPF_JIT_REG + 0)	/* TEMP Register 1 */
116 #define TMP_REG_2	(MAX_BPF_JIT_REG + 1)	/* TEMP Register 2 */
117 #define TCALL_CNT	(MAX_BPF_JIT_REG + 2)	/* Tail Call Count */
118 
119 #define FLAG_IMM_OVERFLOW	(1 << 0)
120 
121 /*
122  * Map eBPF registers to ARM 32bit registers or stack scratch space.
123  *
124  * 1. First argument is passed using the arm 32bit registers and rest of the
125  * arguments are passed on stack scratch space.
126  * 2. First callee-saved argument is mapped to arm 32 bit registers and rest
127  * arguments are mapped to scratch space on stack.
128  * 3. We need two 64 bit temp registers to do complex operations on eBPF
129  * registers.
130  *
131  * As the eBPF registers are all 64 bit registers and arm has only 32 bit
132  * registers, we have to map each eBPF registers with two arm 32 bit regs or
133  * scratch memory space and we have to build eBPF 64 bit register from those.
134  *
135  */
136 static const s8 bpf2a32[][2] = {
137 	/* return value from in-kernel function, and exit value from eBPF */
138 	[BPF_REG_0] = {ARM_R1, ARM_R0},
139 	/* arguments from eBPF program to in-kernel function */
140 	[BPF_REG_1] = {ARM_R3, ARM_R2},
141 	/* Stored on stack scratch space */
142 	[BPF_REG_2] = {STACK_OFFSET(BPF_R2_HI), STACK_OFFSET(BPF_R2_LO)},
143 	[BPF_REG_3] = {STACK_OFFSET(BPF_R3_HI), STACK_OFFSET(BPF_R3_LO)},
144 	[BPF_REG_4] = {STACK_OFFSET(BPF_R4_HI), STACK_OFFSET(BPF_R4_LO)},
145 	[BPF_REG_5] = {STACK_OFFSET(BPF_R5_HI), STACK_OFFSET(BPF_R5_LO)},
146 	/* callee saved registers that in-kernel function will preserve */
147 	[BPF_REG_6] = {ARM_R5, ARM_R4},
148 	/* Stored on stack scratch space */
149 	[BPF_REG_7] = {STACK_OFFSET(BPF_R7_HI), STACK_OFFSET(BPF_R7_LO)},
150 	[BPF_REG_8] = {STACK_OFFSET(BPF_R8_HI), STACK_OFFSET(BPF_R8_LO)},
151 	[BPF_REG_9] = {STACK_OFFSET(BPF_R9_HI), STACK_OFFSET(BPF_R9_LO)},
152 	/* Read only Frame Pointer to access Stack */
153 	[BPF_REG_FP] = {STACK_OFFSET(BPF_FP_HI), STACK_OFFSET(BPF_FP_LO)},
154 	/* Temporary Register for internal BPF JIT, can be used
155 	 * for constant blindings and others.
156 	 */
157 	[TMP_REG_1] = {ARM_R7, ARM_R6},
158 	[TMP_REG_2] = {ARM_R9, ARM_R8},
159 	/* Tail call count. Stored on stack scratch space. */
160 	[TCALL_CNT] = {STACK_OFFSET(BPF_TC_HI), STACK_OFFSET(BPF_TC_LO)},
161 	/* temporary register for blinding constants.
162 	 * Stored on stack scratch space.
163 	 */
164 	[BPF_REG_AX] = {STACK_OFFSET(BPF_AX_HI), STACK_OFFSET(BPF_AX_LO)},
165 };
166 
167 #define	dst_lo	dst[1]
168 #define dst_hi	dst[0]
169 #define src_lo	src[1]
170 #define src_hi	src[0]
171 
172 /*
173  * JIT Context:
174  *
175  * prog			:	bpf_prog
176  * idx			:	index of current last JITed instruction.
177  * prologue_bytes	:	bytes used in prologue.
178  * epilogue_offset	:	offset of epilogue starting.
179  * offsets		:	array of eBPF instruction offsets in
180  *				JITed code.
181  * target		:	final JITed code.
182  * epilogue_bytes	:	no of bytes used in epilogue.
183  * imm_count		:	no of immediate counts used for global
184  *				variables.
185  * imms			:	array of global variable addresses.
186  */
187 
188 struct jit_ctx {
189 	const struct bpf_prog *prog;
190 	unsigned int idx;
191 	unsigned int prologue_bytes;
192 	unsigned int epilogue_offset;
193 	unsigned int cpu_architecture;
194 	u32 flags;
195 	u32 *offsets;
196 	u32 *target;
197 	u32 stack_size;
198 #if __LINUX_ARM_ARCH__ < 7
199 	u16 epilogue_bytes;
200 	u16 imm_count;
201 	u32 *imms;
202 #endif
203 };
204 
205 /*
206  * Wrappers which handle both OABI and EABI and assures Thumb2 interworking
207  * (where the assembly routines like __aeabi_uidiv could cause problems).
208  */
209 static u32 jit_udiv32(u32 dividend, u32 divisor)
210 {
211 	return dividend / divisor;
212 }
213 
214 static u32 jit_mod32(u32 dividend, u32 divisor)
215 {
216 	return dividend % divisor;
217 }
218 
219 static inline void _emit(int cond, u32 inst, struct jit_ctx *ctx)
220 {
221 	inst |= (cond << 28);
222 	inst = __opcode_to_mem_arm(inst);
223 
224 	if (ctx->target != NULL)
225 		ctx->target[ctx->idx] = inst;
226 
227 	ctx->idx++;
228 }
229 
230 /*
231  * Emit an instruction that will be executed unconditionally.
232  */
233 static inline void emit(u32 inst, struct jit_ctx *ctx)
234 {
235 	_emit(ARM_COND_AL, inst, ctx);
236 }
237 
238 /*
239  * This is rather horrid, but necessary to convert an integer constant
240  * to an immediate operand for the opcodes, and be able to detect at
241  * build time whether the constant can't be converted (iow, usable in
242  * BUILD_BUG_ON()).
243  */
244 #define imm12val(v, s) (rol32(v, (s)) | (s) << 7)
245 #define const_imm8m(x)					\
246 	({ int r;					\
247 	   u32 v = (x);					\
248 	   if (!(v & ~0x000000ff))			\
249 		r = imm12val(v, 0);			\
250 	   else if (!(v & ~0xc000003f))			\
251 		r = imm12val(v, 2);			\
252 	   else if (!(v & ~0xf000000f))			\
253 		r = imm12val(v, 4);			\
254 	   else if (!(v & ~0xfc000003))			\
255 		r = imm12val(v, 6);			\
256 	   else if (!(v & ~0xff000000))			\
257 		r = imm12val(v, 8);			\
258 	   else if (!(v & ~0x3fc00000))			\
259 		r = imm12val(v, 10);			\
260 	   else if (!(v & ~0x0ff00000))			\
261 		r = imm12val(v, 12);			\
262 	   else if (!(v & ~0x03fc0000))			\
263 		r = imm12val(v, 14);			\
264 	   else if (!(v & ~0x00ff0000))			\
265 		r = imm12val(v, 16);			\
266 	   else if (!(v & ~0x003fc000))			\
267 		r = imm12val(v, 18);			\
268 	   else if (!(v & ~0x000ff000))			\
269 		r = imm12val(v, 20);			\
270 	   else if (!(v & ~0x0003fc00))			\
271 		r = imm12val(v, 22);			\
272 	   else if (!(v & ~0x0000ff00))			\
273 		r = imm12val(v, 24);			\
274 	   else if (!(v & ~0x00003fc0))			\
275 		r = imm12val(v, 26);			\
276 	   else if (!(v & ~0x00000ff0))			\
277 		r = imm12val(v, 28);			\
278 	   else if (!(v & ~0x000003fc))			\
279 		r = imm12val(v, 30);			\
280 	   else						\
281 		r = -1;					\
282 	   r; })
283 
284 /*
285  * Checks if immediate value can be converted to imm12(12 bits) value.
286  */
287 static int imm8m(u32 x)
288 {
289 	u32 rot;
290 
291 	for (rot = 0; rot < 16; rot++)
292 		if ((x & ~ror32(0xff, 2 * rot)) == 0)
293 			return rol32(x, 2 * rot) | (rot << 8);
294 	return -1;
295 }
296 
297 #define imm8m(x) (__builtin_constant_p(x) ? const_imm8m(x) : imm8m(x))
298 
299 static u32 arm_bpf_ldst_imm12(u32 op, u8 rt, u8 rn, s16 imm12)
300 {
301 	op |= rt << 12 | rn << 16;
302 	if (imm12 >= 0)
303 		op |= ARM_INST_LDST__U;
304 	else
305 		imm12 = -imm12;
306 	return op | (imm12 & ARM_INST_LDST__IMM12);
307 }
308 
309 static u32 arm_bpf_ldst_imm8(u32 op, u8 rt, u8 rn, s16 imm8)
310 {
311 	op |= rt << 12 | rn << 16;
312 	if (imm8 >= 0)
313 		op |= ARM_INST_LDST__U;
314 	else
315 		imm8 = -imm8;
316 	return op | (imm8 & 0xf0) << 4 | (imm8 & 0x0f);
317 }
318 
319 #define ARM_LDR_I(rt, rn, off)	arm_bpf_ldst_imm12(ARM_INST_LDR_I, rt, rn, off)
320 #define ARM_LDRB_I(rt, rn, off)	arm_bpf_ldst_imm12(ARM_INST_LDRB_I, rt, rn, off)
321 #define ARM_LDRD_I(rt, rn, off)	arm_bpf_ldst_imm8(ARM_INST_LDRD_I, rt, rn, off)
322 #define ARM_LDRH_I(rt, rn, off)	arm_bpf_ldst_imm8(ARM_INST_LDRH_I, rt, rn, off)
323 
324 #define ARM_STR_I(rt, rn, off)	arm_bpf_ldst_imm12(ARM_INST_STR_I, rt, rn, off)
325 #define ARM_STRB_I(rt, rn, off)	arm_bpf_ldst_imm12(ARM_INST_STRB_I, rt, rn, off)
326 #define ARM_STRD_I(rt, rn, off)	arm_bpf_ldst_imm8(ARM_INST_STRD_I, rt, rn, off)
327 #define ARM_STRH_I(rt, rn, off)	arm_bpf_ldst_imm8(ARM_INST_STRH_I, rt, rn, off)
328 
329 /*
330  * Initializes the JIT space with undefined instructions.
331  */
332 static void jit_fill_hole(void *area, unsigned int size)
333 {
334 	u32 *ptr;
335 	/* We are guaranteed to have aligned memory. */
336 	for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
337 		*ptr++ = __opcode_to_mem_arm(ARM_INST_UDF);
338 }
339 
340 #if defined(CONFIG_AEABI) && (__LINUX_ARM_ARCH__ >= 5)
341 /* EABI requires the stack to be aligned to 64-bit boundaries */
342 #define STACK_ALIGNMENT	8
343 #else
344 /* Stack must be aligned to 32-bit boundaries */
345 #define STACK_ALIGNMENT	4
346 #endif
347 
348 /* total stack size used in JITed code */
349 #define _STACK_SIZE	(ctx->prog->aux->stack_depth + SCRATCH_SIZE)
350 #define STACK_SIZE	ALIGN(_STACK_SIZE, STACK_ALIGNMENT)
351 
352 #if __LINUX_ARM_ARCH__ < 7
353 
354 static u16 imm_offset(u32 k, struct jit_ctx *ctx)
355 {
356 	unsigned int i = 0, offset;
357 	u16 imm;
358 
359 	/* on the "fake" run we just count them (duplicates included) */
360 	if (ctx->target == NULL) {
361 		ctx->imm_count++;
362 		return 0;
363 	}
364 
365 	while ((i < ctx->imm_count) && ctx->imms[i]) {
366 		if (ctx->imms[i] == k)
367 			break;
368 		i++;
369 	}
370 
371 	if (ctx->imms[i] == 0)
372 		ctx->imms[i] = k;
373 
374 	/* constants go just after the epilogue */
375 	offset =  ctx->offsets[ctx->prog->len - 1] * 4;
376 	offset += ctx->prologue_bytes;
377 	offset += ctx->epilogue_bytes;
378 	offset += i * 4;
379 
380 	ctx->target[offset / 4] = k;
381 
382 	/* PC in ARM mode == address of the instruction + 8 */
383 	imm = offset - (8 + ctx->idx * 4);
384 
385 	if (imm & ~0xfff) {
386 		/*
387 		 * literal pool is too far, signal it into flags. we
388 		 * can only detect it on the second pass unfortunately.
389 		 */
390 		ctx->flags |= FLAG_IMM_OVERFLOW;
391 		return 0;
392 	}
393 
394 	return imm;
395 }
396 
397 #endif /* __LINUX_ARM_ARCH__ */
398 
399 static inline int bpf2a32_offset(int bpf_to, int bpf_from,
400 				 const struct jit_ctx *ctx) {
401 	int to, from;
402 
403 	if (ctx->target == NULL)
404 		return 0;
405 	to = ctx->offsets[bpf_to];
406 	from = ctx->offsets[bpf_from];
407 
408 	return to - from - 1;
409 }
410 
411 /*
412  * Move an immediate that's not an imm8m to a core register.
413  */
414 static inline void emit_mov_i_no8m(const u8 rd, u32 val, struct jit_ctx *ctx)
415 {
416 #if __LINUX_ARM_ARCH__ < 7
417 	emit(ARM_LDR_I(rd, ARM_PC, imm_offset(val, ctx)), ctx);
418 #else
419 	emit(ARM_MOVW(rd, val & 0xffff), ctx);
420 	if (val > 0xffff)
421 		emit(ARM_MOVT(rd, val >> 16), ctx);
422 #endif
423 }
424 
425 static inline void emit_mov_i(const u8 rd, u32 val, struct jit_ctx *ctx)
426 {
427 	int imm12 = imm8m(val);
428 
429 	if (imm12 >= 0)
430 		emit(ARM_MOV_I(rd, imm12), ctx);
431 	else
432 		emit_mov_i_no8m(rd, val, ctx);
433 }
434 
435 static void emit_bx_r(u8 tgt_reg, struct jit_ctx *ctx)
436 {
437 	if (elf_hwcap & HWCAP_THUMB)
438 		emit(ARM_BX(tgt_reg), ctx);
439 	else
440 		emit(ARM_MOV_R(ARM_PC, tgt_reg), ctx);
441 }
442 
443 static inline void emit_blx_r(u8 tgt_reg, struct jit_ctx *ctx)
444 {
445 #if __LINUX_ARM_ARCH__ < 5
446 	emit(ARM_MOV_R(ARM_LR, ARM_PC), ctx);
447 	emit_bx_r(tgt_reg, ctx);
448 #else
449 	emit(ARM_BLX_R(tgt_reg), ctx);
450 #endif
451 }
452 
453 static inline int epilogue_offset(const struct jit_ctx *ctx)
454 {
455 	int to, from;
456 	/* No need for 1st dummy run */
457 	if (ctx->target == NULL)
458 		return 0;
459 	to = ctx->epilogue_offset;
460 	from = ctx->idx;
461 
462 	return to - from - 2;
463 }
464 
465 static inline void emit_udivmod(u8 rd, u8 rm, u8 rn, struct jit_ctx *ctx, u8 op)
466 {
467 	const s8 *tmp = bpf2a32[TMP_REG_1];
468 
469 #if __LINUX_ARM_ARCH__ == 7
470 	if (elf_hwcap & HWCAP_IDIVA) {
471 		if (op == BPF_DIV)
472 			emit(ARM_UDIV(rd, rm, rn), ctx);
473 		else {
474 			emit(ARM_UDIV(ARM_IP, rm, rn), ctx);
475 			emit(ARM_MLS(rd, rn, ARM_IP, rm), ctx);
476 		}
477 		return;
478 	}
479 #endif
480 
481 	/*
482 	 * For BPF_ALU | BPF_DIV | BPF_K instructions
483 	 * As ARM_R1 and ARM_R0 contains 1st argument of bpf
484 	 * function, we need to save it on caller side to save
485 	 * it from getting destroyed within callee.
486 	 * After the return from the callee, we restore ARM_R0
487 	 * ARM_R1.
488 	 */
489 	if (rn != ARM_R1) {
490 		emit(ARM_MOV_R(tmp[0], ARM_R1), ctx);
491 		emit(ARM_MOV_R(ARM_R1, rn), ctx);
492 	}
493 	if (rm != ARM_R0) {
494 		emit(ARM_MOV_R(tmp[1], ARM_R0), ctx);
495 		emit(ARM_MOV_R(ARM_R0, rm), ctx);
496 	}
497 
498 	/* Call appropriate function */
499 	emit_mov_i(ARM_IP, op == BPF_DIV ?
500 		   (u32)jit_udiv32 : (u32)jit_mod32, ctx);
501 	emit_blx_r(ARM_IP, ctx);
502 
503 	/* Save return value */
504 	if (rd != ARM_R0)
505 		emit(ARM_MOV_R(rd, ARM_R0), ctx);
506 
507 	/* Restore ARM_R0 and ARM_R1 */
508 	if (rn != ARM_R1)
509 		emit(ARM_MOV_R(ARM_R1, tmp[0]), ctx);
510 	if (rm != ARM_R0)
511 		emit(ARM_MOV_R(ARM_R0, tmp[1]), ctx);
512 }
513 
514 /* Is the translated BPF register on stack? */
515 static bool is_stacked(s8 reg)
516 {
517 	return reg < 0;
518 }
519 
520 /* If a BPF register is on the stack (stk is true), load it to the
521  * supplied temporary register and return the temporary register
522  * for subsequent operations, otherwise just use the CPU register.
523  */
524 static s8 arm_bpf_get_reg32(s8 reg, s8 tmp, struct jit_ctx *ctx)
525 {
526 	if (is_stacked(reg)) {
527 		emit(ARM_LDR_I(tmp, ARM_FP, EBPF_SCRATCH_TO_ARM_FP(reg)), ctx);
528 		reg = tmp;
529 	}
530 	return reg;
531 }
532 
533 static const s8 *arm_bpf_get_reg64(const s8 *reg, const s8 *tmp,
534 				   struct jit_ctx *ctx)
535 {
536 	if (is_stacked(reg[1])) {
537 		if (__LINUX_ARM_ARCH__ >= 6 ||
538 		    ctx->cpu_architecture >= CPU_ARCH_ARMv5TE) {
539 			emit(ARM_LDRD_I(tmp[1], ARM_FP,
540 					EBPF_SCRATCH_TO_ARM_FP(reg[1])), ctx);
541 		} else {
542 			emit(ARM_LDR_I(tmp[1], ARM_FP,
543 				       EBPF_SCRATCH_TO_ARM_FP(reg[1])), ctx);
544 			emit(ARM_LDR_I(tmp[0], ARM_FP,
545 				       EBPF_SCRATCH_TO_ARM_FP(reg[0])), ctx);
546 		}
547 		reg = tmp;
548 	}
549 	return reg;
550 }
551 
552 /* If a BPF register is on the stack (stk is true), save the register
553  * back to the stack.  If the source register is not the same, then
554  * move it into the correct register.
555  */
556 static void arm_bpf_put_reg32(s8 reg, s8 src, struct jit_ctx *ctx)
557 {
558 	if (is_stacked(reg))
559 		emit(ARM_STR_I(src, ARM_FP, EBPF_SCRATCH_TO_ARM_FP(reg)), ctx);
560 	else if (reg != src)
561 		emit(ARM_MOV_R(reg, src), ctx);
562 }
563 
564 static void arm_bpf_put_reg64(const s8 *reg, const s8 *src,
565 			      struct jit_ctx *ctx)
566 {
567 	if (is_stacked(reg[1])) {
568 		if (__LINUX_ARM_ARCH__ >= 6 ||
569 		    ctx->cpu_architecture >= CPU_ARCH_ARMv5TE) {
570 			emit(ARM_STRD_I(src[1], ARM_FP,
571 				       EBPF_SCRATCH_TO_ARM_FP(reg[1])), ctx);
572 		} else {
573 			emit(ARM_STR_I(src[1], ARM_FP,
574 				       EBPF_SCRATCH_TO_ARM_FP(reg[1])), ctx);
575 			emit(ARM_STR_I(src[0], ARM_FP,
576 				       EBPF_SCRATCH_TO_ARM_FP(reg[0])), ctx);
577 		}
578 	} else {
579 		if (reg[1] != src[1])
580 			emit(ARM_MOV_R(reg[1], src[1]), ctx);
581 		if (reg[0] != src[0])
582 			emit(ARM_MOV_R(reg[0], src[0]), ctx);
583 	}
584 }
585 
586 static inline void emit_a32_mov_i(const s8 dst, const u32 val,
587 				  struct jit_ctx *ctx)
588 {
589 	const s8 *tmp = bpf2a32[TMP_REG_1];
590 
591 	if (is_stacked(dst)) {
592 		emit_mov_i(tmp[1], val, ctx);
593 		arm_bpf_put_reg32(dst, tmp[1], ctx);
594 	} else {
595 		emit_mov_i(dst, val, ctx);
596 	}
597 }
598 
599 static void emit_a32_mov_i64(const s8 dst[], u64 val, struct jit_ctx *ctx)
600 {
601 	const s8 *tmp = bpf2a32[TMP_REG_1];
602 	const s8 *rd = is_stacked(dst_lo) ? tmp : dst;
603 
604 	emit_mov_i(rd[1], (u32)val, ctx);
605 	emit_mov_i(rd[0], val >> 32, ctx);
606 
607 	arm_bpf_put_reg64(dst, rd, ctx);
608 }
609 
610 /* Sign extended move */
611 static inline void emit_a32_mov_se_i64(const bool is64, const s8 dst[],
612 				       const u32 val, struct jit_ctx *ctx) {
613 	u64 val64 = val;
614 
615 	if (is64 && (val & (1<<31)))
616 		val64 |= 0xffffffff00000000ULL;
617 	emit_a32_mov_i64(dst, val64, ctx);
618 }
619 
620 static inline void emit_a32_add_r(const u8 dst, const u8 src,
621 			      const bool is64, const bool hi,
622 			      struct jit_ctx *ctx) {
623 	/* 64 bit :
624 	 *	adds dst_lo, dst_lo, src_lo
625 	 *	adc dst_hi, dst_hi, src_hi
626 	 * 32 bit :
627 	 *	add dst_lo, dst_lo, src_lo
628 	 */
629 	if (!hi && is64)
630 		emit(ARM_ADDS_R(dst, dst, src), ctx);
631 	else if (hi && is64)
632 		emit(ARM_ADC_R(dst, dst, src), ctx);
633 	else
634 		emit(ARM_ADD_R(dst, dst, src), ctx);
635 }
636 
637 static inline void emit_a32_sub_r(const u8 dst, const u8 src,
638 				  const bool is64, const bool hi,
639 				  struct jit_ctx *ctx) {
640 	/* 64 bit :
641 	 *	subs dst_lo, dst_lo, src_lo
642 	 *	sbc dst_hi, dst_hi, src_hi
643 	 * 32 bit :
644 	 *	sub dst_lo, dst_lo, src_lo
645 	 */
646 	if (!hi && is64)
647 		emit(ARM_SUBS_R(dst, dst, src), ctx);
648 	else if (hi && is64)
649 		emit(ARM_SBC_R(dst, dst, src), ctx);
650 	else
651 		emit(ARM_SUB_R(dst, dst, src), ctx);
652 }
653 
654 static inline void emit_alu_r(const u8 dst, const u8 src, const bool is64,
655 			      const bool hi, const u8 op, struct jit_ctx *ctx){
656 	switch (BPF_OP(op)) {
657 	/* dst = dst + src */
658 	case BPF_ADD:
659 		emit_a32_add_r(dst, src, is64, hi, ctx);
660 		break;
661 	/* dst = dst - src */
662 	case BPF_SUB:
663 		emit_a32_sub_r(dst, src, is64, hi, ctx);
664 		break;
665 	/* dst = dst | src */
666 	case BPF_OR:
667 		emit(ARM_ORR_R(dst, dst, src), ctx);
668 		break;
669 	/* dst = dst & src */
670 	case BPF_AND:
671 		emit(ARM_AND_R(dst, dst, src), ctx);
672 		break;
673 	/* dst = dst ^ src */
674 	case BPF_XOR:
675 		emit(ARM_EOR_R(dst, dst, src), ctx);
676 		break;
677 	/* dst = dst * src */
678 	case BPF_MUL:
679 		emit(ARM_MUL(dst, dst, src), ctx);
680 		break;
681 	/* dst = dst << src */
682 	case BPF_LSH:
683 		emit(ARM_LSL_R(dst, dst, src), ctx);
684 		break;
685 	/* dst = dst >> src */
686 	case BPF_RSH:
687 		emit(ARM_LSR_R(dst, dst, src), ctx);
688 		break;
689 	/* dst = dst >> src (signed)*/
690 	case BPF_ARSH:
691 		emit(ARM_MOV_SR(dst, dst, SRTYPE_ASR, src), ctx);
692 		break;
693 	}
694 }
695 
696 /* ALU operation (32 bit)
697  * dst = dst (op) src
698  */
699 static inline void emit_a32_alu_r(const s8 dst, const s8 src,
700 				  struct jit_ctx *ctx, const bool is64,
701 				  const bool hi, const u8 op) {
702 	const s8 *tmp = bpf2a32[TMP_REG_1];
703 	s8 rn, rd;
704 
705 	rn = arm_bpf_get_reg32(src, tmp[1], ctx);
706 	rd = arm_bpf_get_reg32(dst, tmp[0], ctx);
707 	/* ALU operation */
708 	emit_alu_r(rd, rn, is64, hi, op, ctx);
709 	arm_bpf_put_reg32(dst, rd, ctx);
710 }
711 
712 /* ALU operation (64 bit) */
713 static inline void emit_a32_alu_r64(const bool is64, const s8 dst[],
714 				  const s8 src[], struct jit_ctx *ctx,
715 				  const u8 op) {
716 	const s8 *tmp = bpf2a32[TMP_REG_1];
717 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
718 	const s8 *rd;
719 
720 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
721 	if (is64) {
722 		const s8 *rs;
723 
724 		rs = arm_bpf_get_reg64(src, tmp2, ctx);
725 
726 		/* ALU operation */
727 		emit_alu_r(rd[1], rs[1], true, false, op, ctx);
728 		emit_alu_r(rd[0], rs[0], true, true, op, ctx);
729 	} else {
730 		s8 rs;
731 
732 		rs = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
733 
734 		/* ALU operation */
735 		emit_alu_r(rd[1], rs, true, false, op, ctx);
736 		if (!ctx->prog->aux->verifier_zext)
737 			emit_a32_mov_i(rd[0], 0, ctx);
738 	}
739 
740 	arm_bpf_put_reg64(dst, rd, ctx);
741 }
742 
743 /* dst = src (4 bytes)*/
744 static inline void emit_a32_mov_r(const s8 dst, const s8 src,
745 				  struct jit_ctx *ctx) {
746 	const s8 *tmp = bpf2a32[TMP_REG_1];
747 	s8 rt;
748 
749 	rt = arm_bpf_get_reg32(src, tmp[0], ctx);
750 	arm_bpf_put_reg32(dst, rt, ctx);
751 }
752 
753 /* dst = src */
754 static inline void emit_a32_mov_r64(const bool is64, const s8 dst[],
755 				  const s8 src[],
756 				  struct jit_ctx *ctx) {
757 	if (!is64) {
758 		emit_a32_mov_r(dst_lo, src_lo, ctx);
759 		if (!ctx->prog->aux->verifier_zext)
760 			/* Zero out high 4 bytes */
761 			emit_a32_mov_i(dst_hi, 0, ctx);
762 	} else if (__LINUX_ARM_ARCH__ < 6 &&
763 		   ctx->cpu_architecture < CPU_ARCH_ARMv5TE) {
764 		/* complete 8 byte move */
765 		emit_a32_mov_r(dst_lo, src_lo, ctx);
766 		emit_a32_mov_r(dst_hi, src_hi, ctx);
767 	} else if (is_stacked(src_lo) && is_stacked(dst_lo)) {
768 		const u8 *tmp = bpf2a32[TMP_REG_1];
769 
770 		emit(ARM_LDRD_I(tmp[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(src_lo)), ctx);
771 		emit(ARM_STRD_I(tmp[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(dst_lo)), ctx);
772 	} else if (is_stacked(src_lo)) {
773 		emit(ARM_LDRD_I(dst[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(src_lo)), ctx);
774 	} else if (is_stacked(dst_lo)) {
775 		emit(ARM_STRD_I(src[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(dst_lo)), ctx);
776 	} else {
777 		emit(ARM_MOV_R(dst[0], src[0]), ctx);
778 		emit(ARM_MOV_R(dst[1], src[1]), ctx);
779 	}
780 }
781 
782 /* Shift operations */
783 static inline void emit_a32_alu_i(const s8 dst, const u32 val,
784 				struct jit_ctx *ctx, const u8 op) {
785 	const s8 *tmp = bpf2a32[TMP_REG_1];
786 	s8 rd;
787 
788 	rd = arm_bpf_get_reg32(dst, tmp[0], ctx);
789 
790 	/* Do shift operation */
791 	switch (op) {
792 	case BPF_LSH:
793 		emit(ARM_LSL_I(rd, rd, val), ctx);
794 		break;
795 	case BPF_RSH:
796 		emit(ARM_LSR_I(rd, rd, val), ctx);
797 		break;
798 	case BPF_NEG:
799 		emit(ARM_RSB_I(rd, rd, val), ctx);
800 		break;
801 	}
802 
803 	arm_bpf_put_reg32(dst, rd, ctx);
804 }
805 
806 /* dst = ~dst (64 bit) */
807 static inline void emit_a32_neg64(const s8 dst[],
808 				struct jit_ctx *ctx){
809 	const s8 *tmp = bpf2a32[TMP_REG_1];
810 	const s8 *rd;
811 
812 	/* Setup Operand */
813 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
814 
815 	/* Do Negate Operation */
816 	emit(ARM_RSBS_I(rd[1], rd[1], 0), ctx);
817 	emit(ARM_RSC_I(rd[0], rd[0], 0), ctx);
818 
819 	arm_bpf_put_reg64(dst, rd, ctx);
820 }
821 
822 /* dst = dst << src */
823 static inline void emit_a32_lsh_r64(const s8 dst[], const s8 src[],
824 				    struct jit_ctx *ctx) {
825 	const s8 *tmp = bpf2a32[TMP_REG_1];
826 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
827 	const s8 *rd;
828 	s8 rt;
829 
830 	/* Setup Operands */
831 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
832 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
833 
834 	/* Do LSH operation */
835 	emit(ARM_SUB_I(ARM_IP, rt, 32), ctx);
836 	emit(ARM_RSB_I(tmp2[0], rt, 32), ctx);
837 	emit(ARM_MOV_SR(ARM_LR, rd[0], SRTYPE_ASL, rt), ctx);
838 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[1], SRTYPE_ASL, ARM_IP), ctx);
839 	emit(ARM_ORR_SR(ARM_IP, ARM_LR, rd[1], SRTYPE_LSR, tmp2[0]), ctx);
840 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_ASL, rt), ctx);
841 
842 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
843 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
844 }
845 
846 /* dst = dst >> src (signed)*/
847 static inline void emit_a32_arsh_r64(const s8 dst[], const s8 src[],
848 				     struct jit_ctx *ctx) {
849 	const s8 *tmp = bpf2a32[TMP_REG_1];
850 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
851 	const s8 *rd;
852 	s8 rt;
853 
854 	/* Setup Operands */
855 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
856 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
857 
858 	/* Do the ARSH operation */
859 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
860 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
861 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_LSR, rt), ctx);
862 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASL, ARM_IP), ctx);
863 	_emit(ARM_COND_MI, ARM_B(0), ctx);
864 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASR, tmp2[0]), ctx);
865 	emit(ARM_MOV_SR(ARM_IP, rd[0], SRTYPE_ASR, rt), ctx);
866 
867 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
868 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
869 }
870 
871 /* dst = dst >> src */
872 static inline void emit_a32_rsh_r64(const s8 dst[], const s8 src[],
873 				    struct jit_ctx *ctx) {
874 	const s8 *tmp = bpf2a32[TMP_REG_1];
875 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
876 	const s8 *rd;
877 	s8 rt;
878 
879 	/* Setup Operands */
880 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
881 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
882 
883 	/* Do RSH operation */
884 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
885 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
886 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_LSR, rt), ctx);
887 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASL, ARM_IP), ctx);
888 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_LSR, tmp2[0]), ctx);
889 	emit(ARM_MOV_SR(ARM_IP, rd[0], SRTYPE_LSR, rt), ctx);
890 
891 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
892 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
893 }
894 
895 /* dst = dst << val */
896 static inline void emit_a32_lsh_i64(const s8 dst[],
897 				    const u32 val, struct jit_ctx *ctx){
898 	const s8 *tmp = bpf2a32[TMP_REG_1];
899 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
900 	const s8 *rd;
901 
902 	/* Setup operands */
903 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
904 
905 	/* Do LSH operation */
906 	if (val < 32) {
907 		emit(ARM_MOV_SI(tmp2[0], rd[0], SRTYPE_ASL, val), ctx);
908 		emit(ARM_ORR_SI(rd[0], tmp2[0], rd[1], SRTYPE_LSR, 32 - val), ctx);
909 		emit(ARM_MOV_SI(rd[1], rd[1], SRTYPE_ASL, val), ctx);
910 	} else {
911 		if (val == 32)
912 			emit(ARM_MOV_R(rd[0], rd[1]), ctx);
913 		else
914 			emit(ARM_MOV_SI(rd[0], rd[1], SRTYPE_ASL, val - 32), ctx);
915 		emit(ARM_EOR_R(rd[1], rd[1], rd[1]), ctx);
916 	}
917 
918 	arm_bpf_put_reg64(dst, rd, ctx);
919 }
920 
921 /* dst = dst >> val */
922 static inline void emit_a32_rsh_i64(const s8 dst[],
923 				    const u32 val, struct jit_ctx *ctx) {
924 	const s8 *tmp = bpf2a32[TMP_REG_1];
925 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
926 	const s8 *rd;
927 
928 	/* Setup operands */
929 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
930 
931 	/* Do LSR operation */
932 	if (val < 32) {
933 		emit(ARM_MOV_SI(tmp2[1], rd[1], SRTYPE_LSR, val), ctx);
934 		emit(ARM_ORR_SI(rd[1], tmp2[1], rd[0], SRTYPE_ASL, 32 - val), ctx);
935 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_LSR, val), ctx);
936 	} else if (val == 32) {
937 		emit(ARM_MOV_R(rd[1], rd[0]), ctx);
938 		emit(ARM_MOV_I(rd[0], 0), ctx);
939 	} else {
940 		emit(ARM_MOV_SI(rd[1], rd[0], SRTYPE_LSR, val - 32), ctx);
941 		emit(ARM_MOV_I(rd[0], 0), ctx);
942 	}
943 
944 	arm_bpf_put_reg64(dst, rd, ctx);
945 }
946 
947 /* dst = dst >> val (signed) */
948 static inline void emit_a32_arsh_i64(const s8 dst[],
949 				     const u32 val, struct jit_ctx *ctx){
950 	const s8 *tmp = bpf2a32[TMP_REG_1];
951 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
952 	const s8 *rd;
953 
954 	/* Setup operands */
955 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
956 
957 	/* Do ARSH operation */
958 	if (val < 32) {
959 		emit(ARM_MOV_SI(tmp2[1], rd[1], SRTYPE_LSR, val), ctx);
960 		emit(ARM_ORR_SI(rd[1], tmp2[1], rd[0], SRTYPE_ASL, 32 - val), ctx);
961 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, val), ctx);
962 	} else if (val == 32) {
963 		emit(ARM_MOV_R(rd[1], rd[0]), ctx);
964 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, 31), ctx);
965 	} else {
966 		emit(ARM_MOV_SI(rd[1], rd[0], SRTYPE_ASR, val - 32), ctx);
967 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, 31), ctx);
968 	}
969 
970 	arm_bpf_put_reg64(dst, rd, ctx);
971 }
972 
973 static inline void emit_a32_mul_r64(const s8 dst[], const s8 src[],
974 				    struct jit_ctx *ctx) {
975 	const s8 *tmp = bpf2a32[TMP_REG_1];
976 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
977 	const s8 *rd, *rt;
978 
979 	/* Setup operands for multiplication */
980 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
981 	rt = arm_bpf_get_reg64(src, tmp2, ctx);
982 
983 	/* Do Multiplication */
984 	emit(ARM_MUL(ARM_IP, rd[1], rt[0]), ctx);
985 	emit(ARM_MUL(ARM_LR, rd[0], rt[1]), ctx);
986 	emit(ARM_ADD_R(ARM_LR, ARM_IP, ARM_LR), ctx);
987 
988 	emit(ARM_UMULL(ARM_IP, rd[0], rd[1], rt[1]), ctx);
989 	emit(ARM_ADD_R(rd[0], ARM_LR, rd[0]), ctx);
990 
991 	arm_bpf_put_reg32(dst_lo, ARM_IP, ctx);
992 	arm_bpf_put_reg32(dst_hi, rd[0], ctx);
993 }
994 
995 /* *(size *)(dst + off) = src */
996 static inline void emit_str_r(const s8 dst, const s8 src[],
997 			      s32 off, struct jit_ctx *ctx, const u8 sz){
998 	const s8 *tmp = bpf2a32[TMP_REG_1];
999 	s32 off_max;
1000 	s8 rd;
1001 
1002 	rd = arm_bpf_get_reg32(dst, tmp[1], ctx);
1003 
1004 	if (sz == BPF_H)
1005 		off_max = 0xff;
1006 	else
1007 		off_max = 0xfff;
1008 
1009 	if (off < 0 || off > off_max) {
1010 		emit_a32_mov_i(tmp[0], off, ctx);
1011 		emit(ARM_ADD_R(tmp[0], tmp[0], rd), ctx);
1012 		rd = tmp[0];
1013 		off = 0;
1014 	}
1015 	switch (sz) {
1016 	case BPF_B:
1017 		/* Store a Byte */
1018 		emit(ARM_STRB_I(src_lo, rd, off), ctx);
1019 		break;
1020 	case BPF_H:
1021 		/* Store a HalfWord */
1022 		emit(ARM_STRH_I(src_lo, rd, off), ctx);
1023 		break;
1024 	case BPF_W:
1025 		/* Store a Word */
1026 		emit(ARM_STR_I(src_lo, rd, off), ctx);
1027 		break;
1028 	case BPF_DW:
1029 		/* Store a Double Word */
1030 		emit(ARM_STR_I(src_lo, rd, off), ctx);
1031 		emit(ARM_STR_I(src_hi, rd, off + 4), ctx);
1032 		break;
1033 	}
1034 }
1035 
1036 /* dst = *(size*)(src + off) */
1037 static inline void emit_ldx_r(const s8 dst[], const s8 src,
1038 			      s32 off, struct jit_ctx *ctx, const u8 sz){
1039 	const s8 *tmp = bpf2a32[TMP_REG_1];
1040 	const s8 *rd = is_stacked(dst_lo) ? tmp : dst;
1041 	s8 rm = src;
1042 	s32 off_max;
1043 
1044 	if (sz == BPF_H)
1045 		off_max = 0xff;
1046 	else
1047 		off_max = 0xfff;
1048 
1049 	if (off < 0 || off > off_max) {
1050 		emit_a32_mov_i(tmp[0], off, ctx);
1051 		emit(ARM_ADD_R(tmp[0], tmp[0], src), ctx);
1052 		rm = tmp[0];
1053 		off = 0;
1054 	} else if (rd[1] == rm) {
1055 		emit(ARM_MOV_R(tmp[0], rm), ctx);
1056 		rm = tmp[0];
1057 	}
1058 	switch (sz) {
1059 	case BPF_B:
1060 		/* Load a Byte */
1061 		emit(ARM_LDRB_I(rd[1], rm, off), ctx);
1062 		if (!ctx->prog->aux->verifier_zext)
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 		if (!ctx->prog->aux->verifier_zext)
1069 			emit_a32_mov_i(rd[0], 0, ctx);
1070 		break;
1071 	case BPF_W:
1072 		/* Load a Word */
1073 		emit(ARM_LDR_I(rd[1], rm, off), ctx);
1074 		if (!ctx->prog->aux->verifier_zext)
1075 			emit_a32_mov_i(rd[0], 0, ctx);
1076 		break;
1077 	case BPF_DW:
1078 		/* Load a Double Word */
1079 		emit(ARM_LDR_I(rd[1], rm, off), ctx);
1080 		emit(ARM_LDR_I(rd[0], rm, off + 4), ctx);
1081 		break;
1082 	}
1083 	arm_bpf_put_reg64(dst, rd, ctx);
1084 }
1085 
1086 /* Arithmatic Operation */
1087 static inline void emit_ar_r(const u8 rd, const u8 rt, const u8 rm,
1088 			     const u8 rn, struct jit_ctx *ctx, u8 op,
1089 			     bool is_jmp64) {
1090 	switch (op) {
1091 	case BPF_JSET:
1092 		if (is_jmp64) {
1093 			emit(ARM_AND_R(ARM_IP, rt, rn), ctx);
1094 			emit(ARM_AND_R(ARM_LR, rd, rm), ctx);
1095 			emit(ARM_ORRS_R(ARM_IP, ARM_LR, ARM_IP), ctx);
1096 		} else {
1097 			emit(ARM_ANDS_R(ARM_IP, rt, rn), ctx);
1098 		}
1099 		break;
1100 	case BPF_JEQ:
1101 	case BPF_JNE:
1102 	case BPF_JGT:
1103 	case BPF_JGE:
1104 	case BPF_JLE:
1105 	case BPF_JLT:
1106 		if (is_jmp64) {
1107 			emit(ARM_CMP_R(rd, rm), ctx);
1108 			/* Only compare low halve if high halve are equal. */
1109 			_emit(ARM_COND_EQ, ARM_CMP_R(rt, rn), ctx);
1110 		} else {
1111 			emit(ARM_CMP_R(rt, rn), ctx);
1112 		}
1113 		break;
1114 	case BPF_JSLE:
1115 	case BPF_JSGT:
1116 		emit(ARM_CMP_R(rn, rt), ctx);
1117 		if (is_jmp64)
1118 			emit(ARM_SBCS_R(ARM_IP, rm, rd), ctx);
1119 		break;
1120 	case BPF_JSLT:
1121 	case BPF_JSGE:
1122 		emit(ARM_CMP_R(rt, rn), ctx);
1123 		if (is_jmp64)
1124 			emit(ARM_SBCS_R(ARM_IP, rd, rm), ctx);
1125 		break;
1126 	}
1127 }
1128 
1129 static int out_offset = -1; /* initialized on the first pass of build_body() */
1130 static int emit_bpf_tail_call(struct jit_ctx *ctx)
1131 {
1132 
1133 	/* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
1134 	const s8 *r2 = bpf2a32[BPF_REG_2];
1135 	const s8 *r3 = bpf2a32[BPF_REG_3];
1136 	const s8 *tmp = bpf2a32[TMP_REG_1];
1137 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1138 	const s8 *tcc = bpf2a32[TCALL_CNT];
1139 	const s8 *tc;
1140 	const int idx0 = ctx->idx;
1141 #define cur_offset (ctx->idx - idx0)
1142 #define jmp_offset (out_offset - (cur_offset) - 2)
1143 	u32 lo, hi;
1144 	s8 r_array, r_index;
1145 	int off;
1146 
1147 	/* if (index >= array->map.max_entries)
1148 	 *	goto out;
1149 	 */
1150 	BUILD_BUG_ON(offsetof(struct bpf_array, map.max_entries) >
1151 		     ARM_INST_LDST__IMM12);
1152 	off = offsetof(struct bpf_array, map.max_entries);
1153 	r_array = arm_bpf_get_reg32(r2[1], tmp2[0], ctx);
1154 	/* index is 32-bit for arrays */
1155 	r_index = arm_bpf_get_reg32(r3[1], tmp2[1], ctx);
1156 	/* array->map.max_entries */
1157 	emit(ARM_LDR_I(tmp[1], r_array, off), ctx);
1158 	/* index >= array->map.max_entries */
1159 	emit(ARM_CMP_R(r_index, tmp[1]), ctx);
1160 	_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
1161 
1162 	/* tmp2[0] = array, tmp2[1] = index */
1163 
1164 	/* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
1165 	 *	goto out;
1166 	 * tail_call_cnt++;
1167 	 */
1168 	lo = (u32)MAX_TAIL_CALL_CNT;
1169 	hi = (u32)((u64)MAX_TAIL_CALL_CNT >> 32);
1170 	tc = arm_bpf_get_reg64(tcc, tmp, ctx);
1171 	emit(ARM_CMP_I(tc[0], hi), ctx);
1172 	_emit(ARM_COND_EQ, ARM_CMP_I(tc[1], lo), ctx);
1173 	_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
1174 	emit(ARM_ADDS_I(tc[1], tc[1], 1), ctx);
1175 	emit(ARM_ADC_I(tc[0], tc[0], 0), ctx);
1176 	arm_bpf_put_reg64(tcc, tmp, ctx);
1177 
1178 	/* prog = array->ptrs[index]
1179 	 * if (prog == NULL)
1180 	 *	goto out;
1181 	 */
1182 	BUILD_BUG_ON(imm8m(offsetof(struct bpf_array, ptrs)) < 0);
1183 	off = imm8m(offsetof(struct bpf_array, ptrs));
1184 	emit(ARM_ADD_I(tmp[1], r_array, off), ctx);
1185 	emit(ARM_LDR_R_SI(tmp[1], tmp[1], r_index, SRTYPE_ASL, 2), ctx);
1186 	emit(ARM_CMP_I(tmp[1], 0), ctx);
1187 	_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1188 
1189 	/* goto *(prog->bpf_func + prologue_size); */
1190 	BUILD_BUG_ON(offsetof(struct bpf_prog, bpf_func) >
1191 		     ARM_INST_LDST__IMM12);
1192 	off = offsetof(struct bpf_prog, bpf_func);
1193 	emit(ARM_LDR_I(tmp[1], tmp[1], off), ctx);
1194 	emit(ARM_ADD_I(tmp[1], tmp[1], ctx->prologue_bytes), ctx);
1195 	emit_bx_r(tmp[1], ctx);
1196 
1197 	/* out: */
1198 	if (out_offset == -1)
1199 		out_offset = cur_offset;
1200 	if (cur_offset != out_offset) {
1201 		pr_err_once("tail_call out_offset = %d, expected %d!\n",
1202 			    cur_offset, out_offset);
1203 		return -1;
1204 	}
1205 	return 0;
1206 #undef cur_offset
1207 #undef jmp_offset
1208 }
1209 
1210 /* 0xabcd => 0xcdab */
1211 static inline void emit_rev16(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1212 {
1213 #if __LINUX_ARM_ARCH__ < 6
1214 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1215 
1216 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1217 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 8), ctx);
1218 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1219 	emit(ARM_ORR_SI(rd, tmp2[0], tmp2[1], SRTYPE_LSL, 8), ctx);
1220 #else /* ARMv6+ */
1221 	emit(ARM_REV16(rd, rn), ctx);
1222 #endif
1223 }
1224 
1225 /* 0xabcdefgh => 0xghefcdab */
1226 static inline void emit_rev32(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1227 {
1228 #if __LINUX_ARM_ARCH__ < 6
1229 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1230 
1231 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1232 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 24), ctx);
1233 	emit(ARM_ORR_SI(ARM_IP, tmp2[0], tmp2[1], SRTYPE_LSL, 24), ctx);
1234 
1235 	emit(ARM_MOV_SI(tmp2[1], rn, SRTYPE_LSR, 8), ctx);
1236 	emit(ARM_AND_I(tmp2[1], tmp2[1], 0xff), ctx);
1237 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 16), ctx);
1238 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1239 	emit(ARM_MOV_SI(tmp2[0], tmp2[0], SRTYPE_LSL, 8), ctx);
1240 	emit(ARM_ORR_SI(tmp2[0], tmp2[0], tmp2[1], SRTYPE_LSL, 16), ctx);
1241 	emit(ARM_ORR_R(rd, ARM_IP, tmp2[0]), ctx);
1242 
1243 #else /* ARMv6+ */
1244 	emit(ARM_REV(rd, rn), ctx);
1245 #endif
1246 }
1247 
1248 // push the scratch stack register on top of the stack
1249 static inline void emit_push_r64(const s8 src[], struct jit_ctx *ctx)
1250 {
1251 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1252 	const s8 *rt;
1253 	u16 reg_set = 0;
1254 
1255 	rt = arm_bpf_get_reg64(src, tmp2, ctx);
1256 
1257 	reg_set = (1 << rt[1]) | (1 << rt[0]);
1258 	emit(ARM_PUSH(reg_set), ctx);
1259 }
1260 
1261 static void build_prologue(struct jit_ctx *ctx)
1262 {
1263 	const s8 arm_r0 = bpf2a32[BPF_REG_0][1];
1264 	const s8 *bpf_r1 = bpf2a32[BPF_REG_1];
1265 	const s8 *bpf_fp = bpf2a32[BPF_REG_FP];
1266 	const s8 *tcc = bpf2a32[TCALL_CNT];
1267 
1268 	/* Save callee saved registers. */
1269 #ifdef CONFIG_FRAME_POINTER
1270 	u16 reg_set = CALLEE_PUSH_MASK | 1 << ARM_IP | 1 << ARM_PC;
1271 	emit(ARM_MOV_R(ARM_IP, ARM_SP), ctx);
1272 	emit(ARM_PUSH(reg_set), ctx);
1273 	emit(ARM_SUB_I(ARM_FP, ARM_IP, 4), ctx);
1274 #else
1275 	emit(ARM_PUSH(CALLEE_PUSH_MASK), ctx);
1276 	emit(ARM_MOV_R(ARM_FP, ARM_SP), ctx);
1277 #endif
1278 	/* mov r3, #0 */
1279 	/* sub r2, sp, #SCRATCH_SIZE */
1280 	emit(ARM_MOV_I(bpf_r1[0], 0), ctx);
1281 	emit(ARM_SUB_I(bpf_r1[1], ARM_SP, SCRATCH_SIZE), ctx);
1282 
1283 	ctx->stack_size = imm8m(STACK_SIZE);
1284 
1285 	/* Set up function call stack */
1286 	emit(ARM_SUB_I(ARM_SP, ARM_SP, ctx->stack_size), ctx);
1287 
1288 	/* Set up BPF prog stack base register */
1289 	emit_a32_mov_r64(true, bpf_fp, bpf_r1, ctx);
1290 
1291 	/* Initialize Tail Count */
1292 	emit(ARM_MOV_I(bpf_r1[1], 0), ctx);
1293 	emit_a32_mov_r64(true, tcc, bpf_r1, ctx);
1294 
1295 	/* Move BPF_CTX to BPF_R1 */
1296 	emit(ARM_MOV_R(bpf_r1[1], arm_r0), ctx);
1297 
1298 	/* end of prologue */
1299 }
1300 
1301 /* restore callee saved registers. */
1302 static void build_epilogue(struct jit_ctx *ctx)
1303 {
1304 #ifdef CONFIG_FRAME_POINTER
1305 	/* When using frame pointers, some additional registers need to
1306 	 * be loaded. */
1307 	u16 reg_set = CALLEE_POP_MASK | 1 << ARM_SP;
1308 	emit(ARM_SUB_I(ARM_SP, ARM_FP, hweight16(reg_set) * 4), ctx);
1309 	emit(ARM_LDM(ARM_SP, reg_set), ctx);
1310 #else
1311 	/* Restore callee saved registers. */
1312 	emit(ARM_MOV_R(ARM_SP, ARM_FP), ctx);
1313 	emit(ARM_POP(CALLEE_POP_MASK), ctx);
1314 #endif
1315 }
1316 
1317 /*
1318  * Convert an eBPF instruction to native instruction, i.e
1319  * JITs an eBPF instruction.
1320  * Returns :
1321  *	0  - Successfully JITed an 8-byte eBPF instruction
1322  *	>0 - Successfully JITed a 16-byte eBPF instruction
1323  *	<0 - Failed to JIT.
1324  */
1325 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx)
1326 {
1327 	const u8 code = insn->code;
1328 	const s8 *dst = bpf2a32[insn->dst_reg];
1329 	const s8 *src = bpf2a32[insn->src_reg];
1330 	const s8 *tmp = bpf2a32[TMP_REG_1];
1331 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1332 	const s16 off = insn->off;
1333 	const s32 imm = insn->imm;
1334 	const int i = insn - ctx->prog->insnsi;
1335 	const bool is64 = BPF_CLASS(code) == BPF_ALU64;
1336 	const s8 *rd, *rs;
1337 	s8 rd_lo, rt, rm, rn;
1338 	s32 jmp_offset;
1339 
1340 #define check_imm(bits, imm) do {				\
1341 	if ((imm) >= (1 << ((bits) - 1)) ||			\
1342 	    (imm) < -(1 << ((bits) - 1))) {			\
1343 		pr_info("[%2d] imm=%d(0x%x) out of range\n",	\
1344 			i, imm, imm);				\
1345 		return -EINVAL;					\
1346 	}							\
1347 } while (0)
1348 #define check_imm24(imm) check_imm(24, imm)
1349 
1350 	switch (code) {
1351 	/* ALU operations */
1352 
1353 	/* dst = src */
1354 	case BPF_ALU | BPF_MOV | BPF_K:
1355 	case BPF_ALU | BPF_MOV | BPF_X:
1356 	case BPF_ALU64 | BPF_MOV | BPF_K:
1357 	case BPF_ALU64 | BPF_MOV | BPF_X:
1358 		switch (BPF_SRC(code)) {
1359 		case BPF_X:
1360 			if (imm == 1) {
1361 				/* Special mov32 for zext */
1362 				emit_a32_mov_i(dst_hi, 0, ctx);
1363 				break;
1364 			}
1365 			emit_a32_mov_r64(is64, dst, src, ctx);
1366 			break;
1367 		case BPF_K:
1368 			/* Sign-extend immediate value to destination reg */
1369 			emit_a32_mov_se_i64(is64, dst, imm, ctx);
1370 			break;
1371 		}
1372 		break;
1373 	/* dst = dst + src/imm */
1374 	/* dst = dst - src/imm */
1375 	/* dst = dst | src/imm */
1376 	/* dst = dst & src/imm */
1377 	/* dst = dst ^ src/imm */
1378 	/* dst = dst * src/imm */
1379 	/* dst = dst << src */
1380 	/* dst = dst >> src */
1381 	case BPF_ALU | BPF_ADD | BPF_K:
1382 	case BPF_ALU | BPF_ADD | BPF_X:
1383 	case BPF_ALU | BPF_SUB | BPF_K:
1384 	case BPF_ALU | BPF_SUB | BPF_X:
1385 	case BPF_ALU | BPF_OR | BPF_K:
1386 	case BPF_ALU | BPF_OR | BPF_X:
1387 	case BPF_ALU | BPF_AND | BPF_K:
1388 	case BPF_ALU | BPF_AND | BPF_X:
1389 	case BPF_ALU | BPF_XOR | BPF_K:
1390 	case BPF_ALU | BPF_XOR | BPF_X:
1391 	case BPF_ALU | BPF_MUL | BPF_K:
1392 	case BPF_ALU | BPF_MUL | BPF_X:
1393 	case BPF_ALU | BPF_LSH | BPF_X:
1394 	case BPF_ALU | BPF_RSH | BPF_X:
1395 	case BPF_ALU | BPF_ARSH | BPF_K:
1396 	case BPF_ALU | BPF_ARSH | BPF_X:
1397 	case BPF_ALU64 | BPF_ADD | BPF_K:
1398 	case BPF_ALU64 | BPF_ADD | BPF_X:
1399 	case BPF_ALU64 | BPF_SUB | BPF_K:
1400 	case BPF_ALU64 | BPF_SUB | BPF_X:
1401 	case BPF_ALU64 | BPF_OR | BPF_K:
1402 	case BPF_ALU64 | BPF_OR | BPF_X:
1403 	case BPF_ALU64 | BPF_AND | BPF_K:
1404 	case BPF_ALU64 | BPF_AND | BPF_X:
1405 	case BPF_ALU64 | BPF_XOR | BPF_K:
1406 	case BPF_ALU64 | BPF_XOR | BPF_X:
1407 		switch (BPF_SRC(code)) {
1408 		case BPF_X:
1409 			emit_a32_alu_r64(is64, dst, src, ctx, BPF_OP(code));
1410 			break;
1411 		case BPF_K:
1412 			/* Move immediate value to the temporary register
1413 			 * and then do the ALU operation on the temporary
1414 			 * register as this will sign-extend the immediate
1415 			 * value into temporary reg and then it would be
1416 			 * safe to do the operation on it.
1417 			 */
1418 			emit_a32_mov_se_i64(is64, tmp2, imm, ctx);
1419 			emit_a32_alu_r64(is64, dst, tmp2, ctx, BPF_OP(code));
1420 			break;
1421 		}
1422 		break;
1423 	/* dst = dst / src(imm) */
1424 	/* dst = dst % src(imm) */
1425 	case BPF_ALU | BPF_DIV | BPF_K:
1426 	case BPF_ALU | BPF_DIV | BPF_X:
1427 	case BPF_ALU | BPF_MOD | BPF_K:
1428 	case BPF_ALU | BPF_MOD | BPF_X:
1429 		rd_lo = arm_bpf_get_reg32(dst_lo, tmp2[1], ctx);
1430 		switch (BPF_SRC(code)) {
1431 		case BPF_X:
1432 			rt = arm_bpf_get_reg32(src_lo, tmp2[0], ctx);
1433 			break;
1434 		case BPF_K:
1435 			rt = tmp2[0];
1436 			emit_a32_mov_i(rt, imm, ctx);
1437 			break;
1438 		default:
1439 			rt = src_lo;
1440 			break;
1441 		}
1442 		emit_udivmod(rd_lo, rd_lo, rt, ctx, BPF_OP(code));
1443 		arm_bpf_put_reg32(dst_lo, rd_lo, ctx);
1444 		if (!ctx->prog->aux->verifier_zext)
1445 			emit_a32_mov_i(dst_hi, 0, ctx);
1446 		break;
1447 	case BPF_ALU64 | BPF_DIV | BPF_K:
1448 	case BPF_ALU64 | BPF_DIV | BPF_X:
1449 	case BPF_ALU64 | BPF_MOD | BPF_K:
1450 	case BPF_ALU64 | BPF_MOD | BPF_X:
1451 		goto notyet;
1452 	/* dst = dst >> imm */
1453 	/* dst = dst << imm */
1454 	case BPF_ALU | BPF_RSH | BPF_K:
1455 	case BPF_ALU | BPF_LSH | BPF_K:
1456 		if (unlikely(imm > 31))
1457 			return -EINVAL;
1458 		if (imm)
1459 			emit_a32_alu_i(dst_lo, imm, ctx, BPF_OP(code));
1460 		if (!ctx->prog->aux->verifier_zext)
1461 			emit_a32_mov_i(dst_hi, 0, ctx);
1462 		break;
1463 	/* dst = dst << imm */
1464 	case BPF_ALU64 | BPF_LSH | BPF_K:
1465 		if (unlikely(imm > 63))
1466 			return -EINVAL;
1467 		emit_a32_lsh_i64(dst, imm, ctx);
1468 		break;
1469 	/* dst = dst >> imm */
1470 	case BPF_ALU64 | BPF_RSH | BPF_K:
1471 		if (unlikely(imm > 63))
1472 			return -EINVAL;
1473 		emit_a32_rsh_i64(dst, imm, ctx);
1474 		break;
1475 	/* dst = dst << src */
1476 	case BPF_ALU64 | BPF_LSH | BPF_X:
1477 		emit_a32_lsh_r64(dst, src, ctx);
1478 		break;
1479 	/* dst = dst >> src */
1480 	case BPF_ALU64 | BPF_RSH | BPF_X:
1481 		emit_a32_rsh_r64(dst, src, ctx);
1482 		break;
1483 	/* dst = dst >> src (signed) */
1484 	case BPF_ALU64 | BPF_ARSH | BPF_X:
1485 		emit_a32_arsh_r64(dst, src, ctx);
1486 		break;
1487 	/* dst = dst >> imm (signed) */
1488 	case BPF_ALU64 | BPF_ARSH | BPF_K:
1489 		if (unlikely(imm > 63))
1490 			return -EINVAL;
1491 		emit_a32_arsh_i64(dst, imm, ctx);
1492 		break;
1493 	/* dst = ~dst */
1494 	case BPF_ALU | BPF_NEG:
1495 		emit_a32_alu_i(dst_lo, 0, ctx, BPF_OP(code));
1496 		if (!ctx->prog->aux->verifier_zext)
1497 			emit_a32_mov_i(dst_hi, 0, ctx);
1498 		break;
1499 	/* dst = ~dst (64 bit) */
1500 	case BPF_ALU64 | BPF_NEG:
1501 		emit_a32_neg64(dst, ctx);
1502 		break;
1503 	/* dst = dst * src/imm */
1504 	case BPF_ALU64 | BPF_MUL | BPF_X:
1505 	case BPF_ALU64 | BPF_MUL | BPF_K:
1506 		switch (BPF_SRC(code)) {
1507 		case BPF_X:
1508 			emit_a32_mul_r64(dst, src, ctx);
1509 			break;
1510 		case BPF_K:
1511 			/* Move immediate value to the temporary register
1512 			 * and then do the multiplication on it as this
1513 			 * will sign-extend the immediate value into temp
1514 			 * reg then it would be safe to do the operation
1515 			 * on it.
1516 			 */
1517 			emit_a32_mov_se_i64(is64, tmp2, imm, ctx);
1518 			emit_a32_mul_r64(dst, tmp2, ctx);
1519 			break;
1520 		}
1521 		break;
1522 	/* dst = htole(dst) */
1523 	/* dst = htobe(dst) */
1524 	case BPF_ALU | BPF_END | BPF_FROM_LE:
1525 	case BPF_ALU | BPF_END | BPF_FROM_BE:
1526 		rd = arm_bpf_get_reg64(dst, tmp, ctx);
1527 		if (BPF_SRC(code) == BPF_FROM_LE)
1528 			goto emit_bswap_uxt;
1529 		switch (imm) {
1530 		case 16:
1531 			emit_rev16(rd[1], rd[1], ctx);
1532 			goto emit_bswap_uxt;
1533 		case 32:
1534 			emit_rev32(rd[1], rd[1], ctx);
1535 			goto emit_bswap_uxt;
1536 		case 64:
1537 			emit_rev32(ARM_LR, rd[1], ctx);
1538 			emit_rev32(rd[1], rd[0], ctx);
1539 			emit(ARM_MOV_R(rd[0], ARM_LR), ctx);
1540 			break;
1541 		}
1542 		goto exit;
1543 emit_bswap_uxt:
1544 		switch (imm) {
1545 		case 16:
1546 			/* zero-extend 16 bits into 64 bits */
1547 #if __LINUX_ARM_ARCH__ < 6
1548 			emit_a32_mov_i(tmp2[1], 0xffff, ctx);
1549 			emit(ARM_AND_R(rd[1], rd[1], tmp2[1]), ctx);
1550 #else /* ARMv6+ */
1551 			emit(ARM_UXTH(rd[1], rd[1]), ctx);
1552 #endif
1553 			if (!ctx->prog->aux->verifier_zext)
1554 				emit(ARM_EOR_R(rd[0], rd[0], rd[0]), ctx);
1555 			break;
1556 		case 32:
1557 			/* zero-extend 32 bits into 64 bits */
1558 			if (!ctx->prog->aux->verifier_zext)
1559 				emit(ARM_EOR_R(rd[0], rd[0], rd[0]), ctx);
1560 			break;
1561 		case 64:
1562 			/* nop */
1563 			break;
1564 		}
1565 exit:
1566 		arm_bpf_put_reg64(dst, rd, ctx);
1567 		break;
1568 	/* dst = imm64 */
1569 	case BPF_LD | BPF_IMM | BPF_DW:
1570 	{
1571 		u64 val = (u32)imm | (u64)insn[1].imm << 32;
1572 
1573 		emit_a32_mov_i64(dst, val, ctx);
1574 
1575 		return 1;
1576 	}
1577 	/* LDX: dst = *(size *)(src + off) */
1578 	case BPF_LDX | BPF_MEM | BPF_W:
1579 	case BPF_LDX | BPF_MEM | BPF_H:
1580 	case BPF_LDX | BPF_MEM | BPF_B:
1581 	case BPF_LDX | BPF_MEM | BPF_DW:
1582 		rn = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
1583 		emit_ldx_r(dst, rn, off, ctx, BPF_SIZE(code));
1584 		break;
1585 	/* ST: *(size *)(dst + off) = imm */
1586 	case BPF_ST | BPF_MEM | BPF_W:
1587 	case BPF_ST | BPF_MEM | BPF_H:
1588 	case BPF_ST | BPF_MEM | BPF_B:
1589 	case BPF_ST | BPF_MEM | BPF_DW:
1590 		switch (BPF_SIZE(code)) {
1591 		case BPF_DW:
1592 			/* Sign-extend immediate value into temp reg */
1593 			emit_a32_mov_se_i64(true, tmp2, imm, ctx);
1594 			break;
1595 		case BPF_W:
1596 		case BPF_H:
1597 		case BPF_B:
1598 			emit_a32_mov_i(tmp2[1], imm, ctx);
1599 			break;
1600 		}
1601 		emit_str_r(dst_lo, tmp2, off, ctx, BPF_SIZE(code));
1602 		break;
1603 	/* STX XADD: lock *(u32 *)(dst + off) += src */
1604 	case BPF_STX | BPF_XADD | BPF_W:
1605 	/* STX XADD: lock *(u64 *)(dst + off) += src */
1606 	case BPF_STX | BPF_XADD | BPF_DW:
1607 		goto notyet;
1608 	/* STX: *(size *)(dst + off) = src */
1609 	case BPF_STX | BPF_MEM | BPF_W:
1610 	case BPF_STX | BPF_MEM | BPF_H:
1611 	case BPF_STX | BPF_MEM | BPF_B:
1612 	case BPF_STX | BPF_MEM | BPF_DW:
1613 		rs = arm_bpf_get_reg64(src, tmp2, ctx);
1614 		emit_str_r(dst_lo, rs, off, ctx, BPF_SIZE(code));
1615 		break;
1616 	/* PC += off if dst == src */
1617 	/* PC += off if dst > src */
1618 	/* PC += off if dst >= src */
1619 	/* PC += off if dst < src */
1620 	/* PC += off if dst <= src */
1621 	/* PC += off if dst != src */
1622 	/* PC += off if dst > src (signed) */
1623 	/* PC += off if dst >= src (signed) */
1624 	/* PC += off if dst < src (signed) */
1625 	/* PC += off if dst <= src (signed) */
1626 	/* PC += off if dst & src */
1627 	case BPF_JMP | BPF_JEQ | BPF_X:
1628 	case BPF_JMP | BPF_JGT | BPF_X:
1629 	case BPF_JMP | BPF_JGE | BPF_X:
1630 	case BPF_JMP | BPF_JNE | BPF_X:
1631 	case BPF_JMP | BPF_JSGT | BPF_X:
1632 	case BPF_JMP | BPF_JSGE | BPF_X:
1633 	case BPF_JMP | BPF_JSET | BPF_X:
1634 	case BPF_JMP | BPF_JLE | BPF_X:
1635 	case BPF_JMP | BPF_JLT | BPF_X:
1636 	case BPF_JMP | BPF_JSLT | BPF_X:
1637 	case BPF_JMP | BPF_JSLE | BPF_X:
1638 	case BPF_JMP32 | BPF_JEQ | BPF_X:
1639 	case BPF_JMP32 | BPF_JGT | BPF_X:
1640 	case BPF_JMP32 | BPF_JGE | BPF_X:
1641 	case BPF_JMP32 | BPF_JNE | BPF_X:
1642 	case BPF_JMP32 | BPF_JSGT | BPF_X:
1643 	case BPF_JMP32 | BPF_JSGE | BPF_X:
1644 	case BPF_JMP32 | BPF_JSET | BPF_X:
1645 	case BPF_JMP32 | BPF_JLE | BPF_X:
1646 	case BPF_JMP32 | BPF_JLT | BPF_X:
1647 	case BPF_JMP32 | BPF_JSLT | BPF_X:
1648 	case BPF_JMP32 | BPF_JSLE | BPF_X:
1649 		/* Setup source registers */
1650 		rm = arm_bpf_get_reg32(src_hi, tmp2[0], ctx);
1651 		rn = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
1652 		goto go_jmp;
1653 	/* PC += off if dst == imm */
1654 	/* PC += off if dst > imm */
1655 	/* PC += off if dst >= imm */
1656 	/* PC += off if dst < imm */
1657 	/* PC += off if dst <= imm */
1658 	/* PC += off if dst != imm */
1659 	/* PC += off if dst > imm (signed) */
1660 	/* PC += off if dst >= imm (signed) */
1661 	/* PC += off if dst < imm (signed) */
1662 	/* PC += off if dst <= imm (signed) */
1663 	/* PC += off if dst & imm */
1664 	case BPF_JMP | BPF_JEQ | BPF_K:
1665 	case BPF_JMP | BPF_JGT | BPF_K:
1666 	case BPF_JMP | BPF_JGE | BPF_K:
1667 	case BPF_JMP | BPF_JNE | BPF_K:
1668 	case BPF_JMP | BPF_JSGT | BPF_K:
1669 	case BPF_JMP | BPF_JSGE | BPF_K:
1670 	case BPF_JMP | BPF_JSET | BPF_K:
1671 	case BPF_JMP | BPF_JLT | BPF_K:
1672 	case BPF_JMP | BPF_JLE | BPF_K:
1673 	case BPF_JMP | BPF_JSLT | BPF_K:
1674 	case BPF_JMP | BPF_JSLE | BPF_K:
1675 	case BPF_JMP32 | BPF_JEQ | BPF_K:
1676 	case BPF_JMP32 | BPF_JGT | BPF_K:
1677 	case BPF_JMP32 | BPF_JGE | BPF_K:
1678 	case BPF_JMP32 | BPF_JNE | BPF_K:
1679 	case BPF_JMP32 | BPF_JSGT | BPF_K:
1680 	case BPF_JMP32 | BPF_JSGE | BPF_K:
1681 	case BPF_JMP32 | BPF_JSET | BPF_K:
1682 	case BPF_JMP32 | BPF_JLT | BPF_K:
1683 	case BPF_JMP32 | BPF_JLE | BPF_K:
1684 	case BPF_JMP32 | BPF_JSLT | BPF_K:
1685 	case BPF_JMP32 | BPF_JSLE | BPF_K:
1686 		if (off == 0)
1687 			break;
1688 		rm = tmp2[0];
1689 		rn = tmp2[1];
1690 		/* Sign-extend immediate value */
1691 		emit_a32_mov_se_i64(true, tmp2, imm, ctx);
1692 go_jmp:
1693 		/* Setup destination register */
1694 		rd = arm_bpf_get_reg64(dst, tmp, ctx);
1695 
1696 		/* Check for the condition */
1697 		emit_ar_r(rd[0], rd[1], rm, rn, ctx, BPF_OP(code),
1698 			  BPF_CLASS(code) == BPF_JMP);
1699 
1700 		/* Setup JUMP instruction */
1701 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1702 		switch (BPF_OP(code)) {
1703 		case BPF_JNE:
1704 		case BPF_JSET:
1705 			_emit(ARM_COND_NE, ARM_B(jmp_offset), ctx);
1706 			break;
1707 		case BPF_JEQ:
1708 			_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1709 			break;
1710 		case BPF_JGT:
1711 			_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
1712 			break;
1713 		case BPF_JGE:
1714 			_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
1715 			break;
1716 		case BPF_JSGT:
1717 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1718 			break;
1719 		case BPF_JSGE:
1720 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1721 			break;
1722 		case BPF_JLE:
1723 			_emit(ARM_COND_LS, ARM_B(jmp_offset), ctx);
1724 			break;
1725 		case BPF_JLT:
1726 			_emit(ARM_COND_CC, ARM_B(jmp_offset), ctx);
1727 			break;
1728 		case BPF_JSLT:
1729 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1730 			break;
1731 		case BPF_JSLE:
1732 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1733 			break;
1734 		}
1735 		break;
1736 	/* JMP OFF */
1737 	case BPF_JMP | BPF_JA:
1738 	{
1739 		if (off == 0)
1740 			break;
1741 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1742 		check_imm24(jmp_offset);
1743 		emit(ARM_B(jmp_offset), ctx);
1744 		break;
1745 	}
1746 	/* tail call */
1747 	case BPF_JMP | BPF_TAIL_CALL:
1748 		if (emit_bpf_tail_call(ctx))
1749 			return -EFAULT;
1750 		break;
1751 	/* function call */
1752 	case BPF_JMP | BPF_CALL:
1753 	{
1754 		const s8 *r0 = bpf2a32[BPF_REG_0];
1755 		const s8 *r1 = bpf2a32[BPF_REG_1];
1756 		const s8 *r2 = bpf2a32[BPF_REG_2];
1757 		const s8 *r3 = bpf2a32[BPF_REG_3];
1758 		const s8 *r4 = bpf2a32[BPF_REG_4];
1759 		const s8 *r5 = bpf2a32[BPF_REG_5];
1760 		const u32 func = (u32)__bpf_call_base + (u32)imm;
1761 
1762 		emit_a32_mov_r64(true, r0, r1, ctx);
1763 		emit_a32_mov_r64(true, r1, r2, ctx);
1764 		emit_push_r64(r5, ctx);
1765 		emit_push_r64(r4, ctx);
1766 		emit_push_r64(r3, ctx);
1767 
1768 		emit_a32_mov_i(tmp[1], func, ctx);
1769 		emit_blx_r(tmp[1], ctx);
1770 
1771 		emit(ARM_ADD_I(ARM_SP, ARM_SP, imm8m(24)), ctx); // callee clean
1772 		break;
1773 	}
1774 	/* function return */
1775 	case BPF_JMP | BPF_EXIT:
1776 		/* Optimization: when last instruction is EXIT
1777 		 * simply fallthrough to epilogue.
1778 		 */
1779 		if (i == ctx->prog->len - 1)
1780 			break;
1781 		jmp_offset = epilogue_offset(ctx);
1782 		check_imm24(jmp_offset);
1783 		emit(ARM_B(jmp_offset), ctx);
1784 		break;
1785 notyet:
1786 		pr_info_once("*** NOT YET: opcode %02x ***\n", code);
1787 		return -EFAULT;
1788 	default:
1789 		pr_err_once("unknown opcode %02x\n", code);
1790 		return -EINVAL;
1791 	}
1792 
1793 	if (ctx->flags & FLAG_IMM_OVERFLOW)
1794 		/*
1795 		 * this instruction generated an overflow when
1796 		 * trying to access the literal pool, so
1797 		 * delegate this filter to the kernel interpreter.
1798 		 */
1799 		return -1;
1800 	return 0;
1801 }
1802 
1803 static int build_body(struct jit_ctx *ctx)
1804 {
1805 	const struct bpf_prog *prog = ctx->prog;
1806 	unsigned int i;
1807 
1808 	for (i = 0; i < prog->len; i++) {
1809 		const struct bpf_insn *insn = &(prog->insnsi[i]);
1810 		int ret;
1811 
1812 		ret = build_insn(insn, ctx);
1813 
1814 		/* It's used with loading the 64 bit immediate value. */
1815 		if (ret > 0) {
1816 			i++;
1817 			if (ctx->target == NULL)
1818 				ctx->offsets[i] = ctx->idx;
1819 			continue;
1820 		}
1821 
1822 		if (ctx->target == NULL)
1823 			ctx->offsets[i] = ctx->idx;
1824 
1825 		/* If unsuccesfull, return with error code */
1826 		if (ret)
1827 			return ret;
1828 	}
1829 	return 0;
1830 }
1831 
1832 static int validate_code(struct jit_ctx *ctx)
1833 {
1834 	int i;
1835 
1836 	for (i = 0; i < ctx->idx; i++) {
1837 		if (ctx->target[i] == __opcode_to_mem_arm(ARM_INST_UDF))
1838 			return -1;
1839 	}
1840 
1841 	return 0;
1842 }
1843 
1844 void bpf_jit_compile(struct bpf_prog *prog)
1845 {
1846 	/* Nothing to do here. We support Internal BPF. */
1847 }
1848 
1849 bool bpf_jit_needs_zext(void)
1850 {
1851 	return true;
1852 }
1853 
1854 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1855 {
1856 	struct bpf_prog *tmp, *orig_prog = prog;
1857 	struct bpf_binary_header *header;
1858 	bool tmp_blinded = false;
1859 	struct jit_ctx ctx;
1860 	unsigned int tmp_idx;
1861 	unsigned int image_size;
1862 	u8 *image_ptr;
1863 
1864 	/* If BPF JIT was not enabled then we must fall back to
1865 	 * the interpreter.
1866 	 */
1867 	if (!prog->jit_requested)
1868 		return orig_prog;
1869 
1870 	/* If constant blinding was enabled and we failed during blinding
1871 	 * then we must fall back to the interpreter. Otherwise, we save
1872 	 * the new JITed code.
1873 	 */
1874 	tmp = bpf_jit_blind_constants(prog);
1875 
1876 	if (IS_ERR(tmp))
1877 		return orig_prog;
1878 	if (tmp != prog) {
1879 		tmp_blinded = true;
1880 		prog = tmp;
1881 	}
1882 
1883 	memset(&ctx, 0, sizeof(ctx));
1884 	ctx.prog = prog;
1885 	ctx.cpu_architecture = cpu_architecture();
1886 
1887 	/* Not able to allocate memory for offsets[] , then
1888 	 * we must fall back to the interpreter
1889 	 */
1890 	ctx.offsets = kcalloc(prog->len, sizeof(int), GFP_KERNEL);
1891 	if (ctx.offsets == NULL) {
1892 		prog = orig_prog;
1893 		goto out;
1894 	}
1895 
1896 	/* 1) fake pass to find in the length of the JITed code,
1897 	 * to compute ctx->offsets and other context variables
1898 	 * needed to compute final JITed code.
1899 	 * Also, calculate random starting pointer/start of JITed code
1900 	 * which is prefixed by random number of fault instructions.
1901 	 *
1902 	 * If the first pass fails then there is no chance of it
1903 	 * being successful in the second pass, so just fall back
1904 	 * to the interpreter.
1905 	 */
1906 	if (build_body(&ctx)) {
1907 		prog = orig_prog;
1908 		goto out_off;
1909 	}
1910 
1911 	tmp_idx = ctx.idx;
1912 	build_prologue(&ctx);
1913 	ctx.prologue_bytes = (ctx.idx - tmp_idx) * 4;
1914 
1915 	ctx.epilogue_offset = ctx.idx;
1916 
1917 #if __LINUX_ARM_ARCH__ < 7
1918 	tmp_idx = ctx.idx;
1919 	build_epilogue(&ctx);
1920 	ctx.epilogue_bytes = (ctx.idx - tmp_idx) * 4;
1921 
1922 	ctx.idx += ctx.imm_count;
1923 	if (ctx.imm_count) {
1924 		ctx.imms = kcalloc(ctx.imm_count, sizeof(u32), GFP_KERNEL);
1925 		if (ctx.imms == NULL) {
1926 			prog = orig_prog;
1927 			goto out_off;
1928 		}
1929 	}
1930 #else
1931 	/* there's nothing about the epilogue on ARMv7 */
1932 	build_epilogue(&ctx);
1933 #endif
1934 	/* Now we can get the actual image size of the JITed arm code.
1935 	 * Currently, we are not considering the THUMB-2 instructions
1936 	 * for jit, although it can decrease the size of the image.
1937 	 *
1938 	 * As each arm instruction is of length 32bit, we are translating
1939 	 * number of JITed intructions into the size required to store these
1940 	 * JITed code.
1941 	 */
1942 	image_size = sizeof(u32) * ctx.idx;
1943 
1944 	/* Now we know the size of the structure to make */
1945 	header = bpf_jit_binary_alloc(image_size, &image_ptr,
1946 				      sizeof(u32), jit_fill_hole);
1947 	/* Not able to allocate memory for the structure then
1948 	 * we must fall back to the interpretation
1949 	 */
1950 	if (header == NULL) {
1951 		prog = orig_prog;
1952 		goto out_imms;
1953 	}
1954 
1955 	/* 2.) Actual pass to generate final JIT code */
1956 	ctx.target = (u32 *) image_ptr;
1957 	ctx.idx = 0;
1958 
1959 	build_prologue(&ctx);
1960 
1961 	/* If building the body of the JITed code fails somehow,
1962 	 * we fall back to the interpretation.
1963 	 */
1964 	if (build_body(&ctx) < 0) {
1965 		image_ptr = NULL;
1966 		bpf_jit_binary_free(header);
1967 		prog = orig_prog;
1968 		goto out_imms;
1969 	}
1970 	build_epilogue(&ctx);
1971 
1972 	/* 3.) Extra pass to validate JITed Code */
1973 	if (validate_code(&ctx)) {
1974 		image_ptr = NULL;
1975 		bpf_jit_binary_free(header);
1976 		prog = orig_prog;
1977 		goto out_imms;
1978 	}
1979 	flush_icache_range((u32)header, (u32)(ctx.target + ctx.idx));
1980 
1981 	if (bpf_jit_enable > 1)
1982 		/* there are 2 passes here */
1983 		bpf_jit_dump(prog->len, image_size, 2, ctx.target);
1984 
1985 	bpf_jit_binary_lock_ro(header);
1986 	prog->bpf_func = (void *)ctx.target;
1987 	prog->jited = 1;
1988 	prog->jited_len = image_size;
1989 
1990 out_imms:
1991 #if __LINUX_ARM_ARCH__ < 7
1992 	if (ctx.imm_count)
1993 		kfree(ctx.imms);
1994 #endif
1995 out_off:
1996 	kfree(ctx.offsets);
1997 out:
1998 	if (tmp_blinded)
1999 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2000 					   tmp : orig_prog);
2001 	return prog;
2002 }
2003 
2004