xref: /linux/arch/arm/net/bpf_jit_32.c (revision cea0f76a483d1270ac6f6513964e3e75193dda48)
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_ARSH:
799 		emit(ARM_ASR_I(rd, rd, val), ctx);
800 		break;
801 	case BPF_NEG:
802 		emit(ARM_RSB_I(rd, rd, val), ctx);
803 		break;
804 	}
805 
806 	arm_bpf_put_reg32(dst, rd, ctx);
807 }
808 
809 /* dst = ~dst (64 bit) */
810 static inline void emit_a32_neg64(const s8 dst[],
811 				struct jit_ctx *ctx){
812 	const s8 *tmp = bpf2a32[TMP_REG_1];
813 	const s8 *rd;
814 
815 	/* Setup Operand */
816 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
817 
818 	/* Do Negate Operation */
819 	emit(ARM_RSBS_I(rd[1], rd[1], 0), ctx);
820 	emit(ARM_RSC_I(rd[0], rd[0], 0), ctx);
821 
822 	arm_bpf_put_reg64(dst, rd, ctx);
823 }
824 
825 /* dst = dst << src */
826 static inline void emit_a32_lsh_r64(const s8 dst[], const s8 src[],
827 				    struct jit_ctx *ctx) {
828 	const s8 *tmp = bpf2a32[TMP_REG_1];
829 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
830 	const s8 *rd;
831 	s8 rt;
832 
833 	/* Setup Operands */
834 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
835 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
836 
837 	/* Do LSH operation */
838 	emit(ARM_SUB_I(ARM_IP, rt, 32), ctx);
839 	emit(ARM_RSB_I(tmp2[0], rt, 32), ctx);
840 	emit(ARM_MOV_SR(ARM_LR, rd[0], SRTYPE_ASL, rt), ctx);
841 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[1], SRTYPE_ASL, ARM_IP), ctx);
842 	emit(ARM_ORR_SR(ARM_IP, ARM_LR, rd[1], SRTYPE_LSR, tmp2[0]), ctx);
843 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_ASL, rt), ctx);
844 
845 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
846 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
847 }
848 
849 /* dst = dst >> src (signed)*/
850 static inline void emit_a32_arsh_r64(const s8 dst[], const s8 src[],
851 				     struct jit_ctx *ctx) {
852 	const s8 *tmp = bpf2a32[TMP_REG_1];
853 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
854 	const s8 *rd;
855 	s8 rt;
856 
857 	/* Setup Operands */
858 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
859 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
860 
861 	/* Do the ARSH operation */
862 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
863 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
864 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_LSR, rt), ctx);
865 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASL, ARM_IP), ctx);
866 	_emit(ARM_COND_PL,
867 	      ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASR, tmp2[0]), ctx);
868 	emit(ARM_MOV_SR(ARM_IP, rd[0], SRTYPE_ASR, rt), ctx);
869 
870 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
871 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
872 }
873 
874 /* dst = dst >> src */
875 static inline void emit_a32_rsh_r64(const s8 dst[], const s8 src[],
876 				    struct jit_ctx *ctx) {
877 	const s8 *tmp = bpf2a32[TMP_REG_1];
878 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
879 	const s8 *rd;
880 	s8 rt;
881 
882 	/* Setup Operands */
883 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
884 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
885 
886 	/* Do RSH operation */
887 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
888 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
889 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_LSR, rt), ctx);
890 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASL, ARM_IP), ctx);
891 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_LSR, tmp2[0]), ctx);
892 	emit(ARM_MOV_SR(ARM_IP, rd[0], SRTYPE_LSR, rt), ctx);
893 
894 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
895 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
896 }
897 
898 /* dst = dst << val */
899 static inline void emit_a32_lsh_i64(const s8 dst[],
900 				    const u32 val, struct jit_ctx *ctx){
901 	const s8 *tmp = bpf2a32[TMP_REG_1];
902 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
903 	const s8 *rd;
904 
905 	/* Setup operands */
906 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
907 
908 	/* Do LSH operation */
909 	if (val < 32) {
910 		emit(ARM_MOV_SI(tmp2[0], rd[0], SRTYPE_ASL, val), ctx);
911 		emit(ARM_ORR_SI(rd[0], tmp2[0], rd[1], SRTYPE_LSR, 32 - val), ctx);
912 		emit(ARM_MOV_SI(rd[1], rd[1], SRTYPE_ASL, val), ctx);
913 	} else {
914 		if (val == 32)
915 			emit(ARM_MOV_R(rd[0], rd[1]), ctx);
916 		else
917 			emit(ARM_MOV_SI(rd[0], rd[1], SRTYPE_ASL, val - 32), ctx);
918 		emit(ARM_EOR_R(rd[1], rd[1], rd[1]), ctx);
919 	}
920 
921 	arm_bpf_put_reg64(dst, rd, ctx);
922 }
923 
924 /* dst = dst >> val */
925 static inline void emit_a32_rsh_i64(const s8 dst[],
926 				    const u32 val, struct jit_ctx *ctx) {
927 	const s8 *tmp = bpf2a32[TMP_REG_1];
928 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
929 	const s8 *rd;
930 
931 	/* Setup operands */
932 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
933 
934 	/* Do LSR operation */
935 	if (val == 0) {
936 		/* An immediate value of 0 encodes a shift amount of 32
937 		 * for LSR. To shift by 0, don't do anything.
938 		 */
939 	} else if (val < 32) {
940 		emit(ARM_MOV_SI(tmp2[1], rd[1], SRTYPE_LSR, val), ctx);
941 		emit(ARM_ORR_SI(rd[1], tmp2[1], rd[0], SRTYPE_ASL, 32 - val), ctx);
942 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_LSR, val), ctx);
943 	} else if (val == 32) {
944 		emit(ARM_MOV_R(rd[1], rd[0]), ctx);
945 		emit(ARM_MOV_I(rd[0], 0), ctx);
946 	} else {
947 		emit(ARM_MOV_SI(rd[1], rd[0], SRTYPE_LSR, val - 32), ctx);
948 		emit(ARM_MOV_I(rd[0], 0), ctx);
949 	}
950 
951 	arm_bpf_put_reg64(dst, rd, ctx);
952 }
953 
954 /* dst = dst >> val (signed) */
955 static inline void emit_a32_arsh_i64(const s8 dst[],
956 				     const u32 val, struct jit_ctx *ctx){
957 	const s8 *tmp = bpf2a32[TMP_REG_1];
958 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
959 	const s8 *rd;
960 
961 	/* Setup operands */
962 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
963 
964 	/* Do ARSH operation */
965 	if (val == 0) {
966 		/* An immediate value of 0 encodes a shift amount of 32
967 		 * for ASR. To shift by 0, don't do anything.
968 		 */
969 	} else if (val < 32) {
970 		emit(ARM_MOV_SI(tmp2[1], rd[1], SRTYPE_LSR, val), ctx);
971 		emit(ARM_ORR_SI(rd[1], tmp2[1], rd[0], SRTYPE_ASL, 32 - val), ctx);
972 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, val), ctx);
973 	} else if (val == 32) {
974 		emit(ARM_MOV_R(rd[1], rd[0]), ctx);
975 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, 31), ctx);
976 	} else {
977 		emit(ARM_MOV_SI(rd[1], rd[0], SRTYPE_ASR, val - 32), ctx);
978 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, 31), ctx);
979 	}
980 
981 	arm_bpf_put_reg64(dst, rd, ctx);
982 }
983 
984 static inline void emit_a32_mul_r64(const s8 dst[], const s8 src[],
985 				    struct jit_ctx *ctx) {
986 	const s8 *tmp = bpf2a32[TMP_REG_1];
987 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
988 	const s8 *rd, *rt;
989 
990 	/* Setup operands for multiplication */
991 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
992 	rt = arm_bpf_get_reg64(src, tmp2, ctx);
993 
994 	/* Do Multiplication */
995 	emit(ARM_MUL(ARM_IP, rd[1], rt[0]), ctx);
996 	emit(ARM_MUL(ARM_LR, rd[0], rt[1]), ctx);
997 	emit(ARM_ADD_R(ARM_LR, ARM_IP, ARM_LR), ctx);
998 
999 	emit(ARM_UMULL(ARM_IP, rd[0], rd[1], rt[1]), ctx);
1000 	emit(ARM_ADD_R(rd[0], ARM_LR, rd[0]), ctx);
1001 
1002 	arm_bpf_put_reg32(dst_lo, ARM_IP, ctx);
1003 	arm_bpf_put_reg32(dst_hi, rd[0], ctx);
1004 }
1005 
1006 static bool is_ldst_imm(s16 off, const u8 size)
1007 {
1008 	s16 off_max = 0;
1009 
1010 	switch (size) {
1011 	case BPF_B:
1012 	case BPF_W:
1013 		off_max = 0xfff;
1014 		break;
1015 	case BPF_H:
1016 		off_max = 0xff;
1017 		break;
1018 	case BPF_DW:
1019 		/* Need to make sure off+4 does not overflow. */
1020 		off_max = 0xfff - 4;
1021 		break;
1022 	}
1023 	return -off_max <= off && off <= off_max;
1024 }
1025 
1026 /* *(size *)(dst + off) = src */
1027 static inline void emit_str_r(const s8 dst, const s8 src[],
1028 			      s16 off, struct jit_ctx *ctx, const u8 sz){
1029 	const s8 *tmp = bpf2a32[TMP_REG_1];
1030 	s8 rd;
1031 
1032 	rd = arm_bpf_get_reg32(dst, tmp[1], ctx);
1033 
1034 	if (!is_ldst_imm(off, sz)) {
1035 		emit_a32_mov_i(tmp[0], off, ctx);
1036 		emit(ARM_ADD_R(tmp[0], tmp[0], rd), ctx);
1037 		rd = tmp[0];
1038 		off = 0;
1039 	}
1040 	switch (sz) {
1041 	case BPF_B:
1042 		/* Store a Byte */
1043 		emit(ARM_STRB_I(src_lo, rd, off), ctx);
1044 		break;
1045 	case BPF_H:
1046 		/* Store a HalfWord */
1047 		emit(ARM_STRH_I(src_lo, rd, off), ctx);
1048 		break;
1049 	case BPF_W:
1050 		/* Store a Word */
1051 		emit(ARM_STR_I(src_lo, rd, off), ctx);
1052 		break;
1053 	case BPF_DW:
1054 		/* Store a Double Word */
1055 		emit(ARM_STR_I(src_lo, rd, off), ctx);
1056 		emit(ARM_STR_I(src_hi, rd, off + 4), ctx);
1057 		break;
1058 	}
1059 }
1060 
1061 /* dst = *(size*)(src + off) */
1062 static inline void emit_ldx_r(const s8 dst[], const s8 src,
1063 			      s16 off, struct jit_ctx *ctx, const u8 sz){
1064 	const s8 *tmp = bpf2a32[TMP_REG_1];
1065 	const s8 *rd = is_stacked(dst_lo) ? tmp : dst;
1066 	s8 rm = src;
1067 
1068 	if (!is_ldst_imm(off, sz)) {
1069 		emit_a32_mov_i(tmp[0], off, ctx);
1070 		emit(ARM_ADD_R(tmp[0], tmp[0], src), ctx);
1071 		rm = tmp[0];
1072 		off = 0;
1073 	} else if (rd[1] == rm) {
1074 		emit(ARM_MOV_R(tmp[0], rm), ctx);
1075 		rm = tmp[0];
1076 	}
1077 	switch (sz) {
1078 	case BPF_B:
1079 		/* Load a Byte */
1080 		emit(ARM_LDRB_I(rd[1], rm, off), ctx);
1081 		if (!ctx->prog->aux->verifier_zext)
1082 			emit_a32_mov_i(rd[0], 0, ctx);
1083 		break;
1084 	case BPF_H:
1085 		/* Load a HalfWord */
1086 		emit(ARM_LDRH_I(rd[1], rm, off), ctx);
1087 		if (!ctx->prog->aux->verifier_zext)
1088 			emit_a32_mov_i(rd[0], 0, ctx);
1089 		break;
1090 	case BPF_W:
1091 		/* Load a Word */
1092 		emit(ARM_LDR_I(rd[1], rm, off), ctx);
1093 		if (!ctx->prog->aux->verifier_zext)
1094 			emit_a32_mov_i(rd[0], 0, ctx);
1095 		break;
1096 	case BPF_DW:
1097 		/* Load a Double Word */
1098 		emit(ARM_LDR_I(rd[1], rm, off), ctx);
1099 		emit(ARM_LDR_I(rd[0], rm, off + 4), ctx);
1100 		break;
1101 	}
1102 	arm_bpf_put_reg64(dst, rd, ctx);
1103 }
1104 
1105 /* Arithmatic Operation */
1106 static inline void emit_ar_r(const u8 rd, const u8 rt, const u8 rm,
1107 			     const u8 rn, struct jit_ctx *ctx, u8 op,
1108 			     bool is_jmp64) {
1109 	switch (op) {
1110 	case BPF_JSET:
1111 		if (is_jmp64) {
1112 			emit(ARM_AND_R(ARM_IP, rt, rn), ctx);
1113 			emit(ARM_AND_R(ARM_LR, rd, rm), ctx);
1114 			emit(ARM_ORRS_R(ARM_IP, ARM_LR, ARM_IP), ctx);
1115 		} else {
1116 			emit(ARM_ANDS_R(ARM_IP, rt, rn), ctx);
1117 		}
1118 		break;
1119 	case BPF_JEQ:
1120 	case BPF_JNE:
1121 	case BPF_JGT:
1122 	case BPF_JGE:
1123 	case BPF_JLE:
1124 	case BPF_JLT:
1125 		if (is_jmp64) {
1126 			emit(ARM_CMP_R(rd, rm), ctx);
1127 			/* Only compare low halve if high halve are equal. */
1128 			_emit(ARM_COND_EQ, ARM_CMP_R(rt, rn), ctx);
1129 		} else {
1130 			emit(ARM_CMP_R(rt, rn), ctx);
1131 		}
1132 		break;
1133 	case BPF_JSLE:
1134 	case BPF_JSGT:
1135 		emit(ARM_CMP_R(rn, rt), ctx);
1136 		if (is_jmp64)
1137 			emit(ARM_SBCS_R(ARM_IP, rm, rd), ctx);
1138 		break;
1139 	case BPF_JSLT:
1140 	case BPF_JSGE:
1141 		emit(ARM_CMP_R(rt, rn), ctx);
1142 		if (is_jmp64)
1143 			emit(ARM_SBCS_R(ARM_IP, rd, rm), ctx);
1144 		break;
1145 	}
1146 }
1147 
1148 static int out_offset = -1; /* initialized on the first pass of build_body() */
1149 static int emit_bpf_tail_call(struct jit_ctx *ctx)
1150 {
1151 
1152 	/* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
1153 	const s8 *r2 = bpf2a32[BPF_REG_2];
1154 	const s8 *r3 = bpf2a32[BPF_REG_3];
1155 	const s8 *tmp = bpf2a32[TMP_REG_1];
1156 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1157 	const s8 *tcc = bpf2a32[TCALL_CNT];
1158 	const s8 *tc;
1159 	const int idx0 = ctx->idx;
1160 #define cur_offset (ctx->idx - idx0)
1161 #define jmp_offset (out_offset - (cur_offset) - 2)
1162 	u32 lo, hi;
1163 	s8 r_array, r_index;
1164 	int off;
1165 
1166 	/* if (index >= array->map.max_entries)
1167 	 *	goto out;
1168 	 */
1169 	BUILD_BUG_ON(offsetof(struct bpf_array, map.max_entries) >
1170 		     ARM_INST_LDST__IMM12);
1171 	off = offsetof(struct bpf_array, map.max_entries);
1172 	r_array = arm_bpf_get_reg32(r2[1], tmp2[0], ctx);
1173 	/* index is 32-bit for arrays */
1174 	r_index = arm_bpf_get_reg32(r3[1], tmp2[1], ctx);
1175 	/* array->map.max_entries */
1176 	emit(ARM_LDR_I(tmp[1], r_array, off), ctx);
1177 	/* index >= array->map.max_entries */
1178 	emit(ARM_CMP_R(r_index, tmp[1]), ctx);
1179 	_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
1180 
1181 	/* tmp2[0] = array, tmp2[1] = index */
1182 
1183 	/* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
1184 	 *	goto out;
1185 	 * tail_call_cnt++;
1186 	 */
1187 	lo = (u32)MAX_TAIL_CALL_CNT;
1188 	hi = (u32)((u64)MAX_TAIL_CALL_CNT >> 32);
1189 	tc = arm_bpf_get_reg64(tcc, tmp, ctx);
1190 	emit(ARM_CMP_I(tc[0], hi), ctx);
1191 	_emit(ARM_COND_EQ, ARM_CMP_I(tc[1], lo), ctx);
1192 	_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
1193 	emit(ARM_ADDS_I(tc[1], tc[1], 1), ctx);
1194 	emit(ARM_ADC_I(tc[0], tc[0], 0), ctx);
1195 	arm_bpf_put_reg64(tcc, tmp, ctx);
1196 
1197 	/* prog = array->ptrs[index]
1198 	 * if (prog == NULL)
1199 	 *	goto out;
1200 	 */
1201 	BUILD_BUG_ON(imm8m(offsetof(struct bpf_array, ptrs)) < 0);
1202 	off = imm8m(offsetof(struct bpf_array, ptrs));
1203 	emit(ARM_ADD_I(tmp[1], r_array, off), ctx);
1204 	emit(ARM_LDR_R_SI(tmp[1], tmp[1], r_index, SRTYPE_ASL, 2), ctx);
1205 	emit(ARM_CMP_I(tmp[1], 0), ctx);
1206 	_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1207 
1208 	/* goto *(prog->bpf_func + prologue_size); */
1209 	BUILD_BUG_ON(offsetof(struct bpf_prog, bpf_func) >
1210 		     ARM_INST_LDST__IMM12);
1211 	off = offsetof(struct bpf_prog, bpf_func);
1212 	emit(ARM_LDR_I(tmp[1], tmp[1], off), ctx);
1213 	emit(ARM_ADD_I(tmp[1], tmp[1], ctx->prologue_bytes), ctx);
1214 	emit_bx_r(tmp[1], ctx);
1215 
1216 	/* out: */
1217 	if (out_offset == -1)
1218 		out_offset = cur_offset;
1219 	if (cur_offset != out_offset) {
1220 		pr_err_once("tail_call out_offset = %d, expected %d!\n",
1221 			    cur_offset, out_offset);
1222 		return -1;
1223 	}
1224 	return 0;
1225 #undef cur_offset
1226 #undef jmp_offset
1227 }
1228 
1229 /* 0xabcd => 0xcdab */
1230 static inline void emit_rev16(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1231 {
1232 #if __LINUX_ARM_ARCH__ < 6
1233 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1234 
1235 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1236 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 8), ctx);
1237 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1238 	emit(ARM_ORR_SI(rd, tmp2[0], tmp2[1], SRTYPE_LSL, 8), ctx);
1239 #else /* ARMv6+ */
1240 	emit(ARM_REV16(rd, rn), ctx);
1241 #endif
1242 }
1243 
1244 /* 0xabcdefgh => 0xghefcdab */
1245 static inline void emit_rev32(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1246 {
1247 #if __LINUX_ARM_ARCH__ < 6
1248 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1249 
1250 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1251 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 24), ctx);
1252 	emit(ARM_ORR_SI(ARM_IP, tmp2[0], tmp2[1], SRTYPE_LSL, 24), ctx);
1253 
1254 	emit(ARM_MOV_SI(tmp2[1], rn, SRTYPE_LSR, 8), ctx);
1255 	emit(ARM_AND_I(tmp2[1], tmp2[1], 0xff), ctx);
1256 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 16), ctx);
1257 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1258 	emit(ARM_MOV_SI(tmp2[0], tmp2[0], SRTYPE_LSL, 8), ctx);
1259 	emit(ARM_ORR_SI(tmp2[0], tmp2[0], tmp2[1], SRTYPE_LSL, 16), ctx);
1260 	emit(ARM_ORR_R(rd, ARM_IP, tmp2[0]), ctx);
1261 
1262 #else /* ARMv6+ */
1263 	emit(ARM_REV(rd, rn), ctx);
1264 #endif
1265 }
1266 
1267 // push the scratch stack register on top of the stack
1268 static inline void emit_push_r64(const s8 src[], struct jit_ctx *ctx)
1269 {
1270 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1271 	const s8 *rt;
1272 	u16 reg_set = 0;
1273 
1274 	rt = arm_bpf_get_reg64(src, tmp2, ctx);
1275 
1276 	reg_set = (1 << rt[1]) | (1 << rt[0]);
1277 	emit(ARM_PUSH(reg_set), ctx);
1278 }
1279 
1280 static void build_prologue(struct jit_ctx *ctx)
1281 {
1282 	const s8 arm_r0 = bpf2a32[BPF_REG_0][1];
1283 	const s8 *bpf_r1 = bpf2a32[BPF_REG_1];
1284 	const s8 *bpf_fp = bpf2a32[BPF_REG_FP];
1285 	const s8 *tcc = bpf2a32[TCALL_CNT];
1286 
1287 	/* Save callee saved registers. */
1288 #ifdef CONFIG_FRAME_POINTER
1289 	u16 reg_set = CALLEE_PUSH_MASK | 1 << ARM_IP | 1 << ARM_PC;
1290 	emit(ARM_MOV_R(ARM_IP, ARM_SP), ctx);
1291 	emit(ARM_PUSH(reg_set), ctx);
1292 	emit(ARM_SUB_I(ARM_FP, ARM_IP, 4), ctx);
1293 #else
1294 	emit(ARM_PUSH(CALLEE_PUSH_MASK), ctx);
1295 	emit(ARM_MOV_R(ARM_FP, ARM_SP), ctx);
1296 #endif
1297 	/* mov r3, #0 */
1298 	/* sub r2, sp, #SCRATCH_SIZE */
1299 	emit(ARM_MOV_I(bpf_r1[0], 0), ctx);
1300 	emit(ARM_SUB_I(bpf_r1[1], ARM_SP, SCRATCH_SIZE), ctx);
1301 
1302 	ctx->stack_size = imm8m(STACK_SIZE);
1303 
1304 	/* Set up function call stack */
1305 	emit(ARM_SUB_I(ARM_SP, ARM_SP, ctx->stack_size), ctx);
1306 
1307 	/* Set up BPF prog stack base register */
1308 	emit_a32_mov_r64(true, bpf_fp, bpf_r1, ctx);
1309 
1310 	/* Initialize Tail Count */
1311 	emit(ARM_MOV_I(bpf_r1[1], 0), ctx);
1312 	emit_a32_mov_r64(true, tcc, bpf_r1, ctx);
1313 
1314 	/* Move BPF_CTX to BPF_R1 */
1315 	emit(ARM_MOV_R(bpf_r1[1], arm_r0), ctx);
1316 
1317 	/* end of prologue */
1318 }
1319 
1320 /* restore callee saved registers. */
1321 static void build_epilogue(struct jit_ctx *ctx)
1322 {
1323 #ifdef CONFIG_FRAME_POINTER
1324 	/* When using frame pointers, some additional registers need to
1325 	 * be loaded. */
1326 	u16 reg_set = CALLEE_POP_MASK | 1 << ARM_SP;
1327 	emit(ARM_SUB_I(ARM_SP, ARM_FP, hweight16(reg_set) * 4), ctx);
1328 	emit(ARM_LDM(ARM_SP, reg_set), ctx);
1329 #else
1330 	/* Restore callee saved registers. */
1331 	emit(ARM_MOV_R(ARM_SP, ARM_FP), ctx);
1332 	emit(ARM_POP(CALLEE_POP_MASK), ctx);
1333 #endif
1334 }
1335 
1336 /*
1337  * Convert an eBPF instruction to native instruction, i.e
1338  * JITs an eBPF instruction.
1339  * Returns :
1340  *	0  - Successfully JITed an 8-byte eBPF instruction
1341  *	>0 - Successfully JITed a 16-byte eBPF instruction
1342  *	<0 - Failed to JIT.
1343  */
1344 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx)
1345 {
1346 	const u8 code = insn->code;
1347 	const s8 *dst = bpf2a32[insn->dst_reg];
1348 	const s8 *src = bpf2a32[insn->src_reg];
1349 	const s8 *tmp = bpf2a32[TMP_REG_1];
1350 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1351 	const s16 off = insn->off;
1352 	const s32 imm = insn->imm;
1353 	const int i = insn - ctx->prog->insnsi;
1354 	const bool is64 = BPF_CLASS(code) == BPF_ALU64;
1355 	const s8 *rd, *rs;
1356 	s8 rd_lo, rt, rm, rn;
1357 	s32 jmp_offset;
1358 
1359 #define check_imm(bits, imm) do {				\
1360 	if ((imm) >= (1 << ((bits) - 1)) ||			\
1361 	    (imm) < -(1 << ((bits) - 1))) {			\
1362 		pr_info("[%2d] imm=%d(0x%x) out of range\n",	\
1363 			i, imm, imm);				\
1364 		return -EINVAL;					\
1365 	}							\
1366 } while (0)
1367 #define check_imm24(imm) check_imm(24, imm)
1368 
1369 	switch (code) {
1370 	/* ALU operations */
1371 
1372 	/* dst = src */
1373 	case BPF_ALU | BPF_MOV | BPF_K:
1374 	case BPF_ALU | BPF_MOV | BPF_X:
1375 	case BPF_ALU64 | BPF_MOV | BPF_K:
1376 	case BPF_ALU64 | BPF_MOV | BPF_X:
1377 		switch (BPF_SRC(code)) {
1378 		case BPF_X:
1379 			if (imm == 1) {
1380 				/* Special mov32 for zext */
1381 				emit_a32_mov_i(dst_hi, 0, ctx);
1382 				break;
1383 			}
1384 			emit_a32_mov_r64(is64, dst, src, ctx);
1385 			break;
1386 		case BPF_K:
1387 			/* Sign-extend immediate value to destination reg */
1388 			emit_a32_mov_se_i64(is64, dst, imm, ctx);
1389 			break;
1390 		}
1391 		break;
1392 	/* dst = dst + src/imm */
1393 	/* dst = dst - src/imm */
1394 	/* dst = dst | src/imm */
1395 	/* dst = dst & src/imm */
1396 	/* dst = dst ^ src/imm */
1397 	/* dst = dst * src/imm */
1398 	/* dst = dst << src */
1399 	/* dst = dst >> src */
1400 	case BPF_ALU | BPF_ADD | BPF_K:
1401 	case BPF_ALU | BPF_ADD | BPF_X:
1402 	case BPF_ALU | BPF_SUB | BPF_K:
1403 	case BPF_ALU | BPF_SUB | BPF_X:
1404 	case BPF_ALU | BPF_OR | BPF_K:
1405 	case BPF_ALU | BPF_OR | BPF_X:
1406 	case BPF_ALU | BPF_AND | BPF_K:
1407 	case BPF_ALU | BPF_AND | BPF_X:
1408 	case BPF_ALU | BPF_XOR | BPF_K:
1409 	case BPF_ALU | BPF_XOR | BPF_X:
1410 	case BPF_ALU | BPF_MUL | BPF_K:
1411 	case BPF_ALU | BPF_MUL | BPF_X:
1412 	case BPF_ALU | BPF_LSH | BPF_X:
1413 	case BPF_ALU | BPF_RSH | BPF_X:
1414 	case BPF_ALU | BPF_ARSH | BPF_X:
1415 	case BPF_ALU64 | BPF_ADD | BPF_K:
1416 	case BPF_ALU64 | BPF_ADD | BPF_X:
1417 	case BPF_ALU64 | BPF_SUB | BPF_K:
1418 	case BPF_ALU64 | BPF_SUB | BPF_X:
1419 	case BPF_ALU64 | BPF_OR | BPF_K:
1420 	case BPF_ALU64 | BPF_OR | BPF_X:
1421 	case BPF_ALU64 | BPF_AND | BPF_K:
1422 	case BPF_ALU64 | BPF_AND | BPF_X:
1423 	case BPF_ALU64 | BPF_XOR | BPF_K:
1424 	case BPF_ALU64 | BPF_XOR | BPF_X:
1425 		switch (BPF_SRC(code)) {
1426 		case BPF_X:
1427 			emit_a32_alu_r64(is64, dst, src, ctx, BPF_OP(code));
1428 			break;
1429 		case BPF_K:
1430 			/* Move immediate value to the temporary register
1431 			 * and then do the ALU operation on the temporary
1432 			 * register as this will sign-extend the immediate
1433 			 * value into temporary reg and then it would be
1434 			 * safe to do the operation on it.
1435 			 */
1436 			emit_a32_mov_se_i64(is64, tmp2, imm, ctx);
1437 			emit_a32_alu_r64(is64, dst, tmp2, ctx, BPF_OP(code));
1438 			break;
1439 		}
1440 		break;
1441 	/* dst = dst / src(imm) */
1442 	/* dst = dst % src(imm) */
1443 	case BPF_ALU | BPF_DIV | BPF_K:
1444 	case BPF_ALU | BPF_DIV | BPF_X:
1445 	case BPF_ALU | BPF_MOD | BPF_K:
1446 	case BPF_ALU | BPF_MOD | BPF_X:
1447 		rd_lo = arm_bpf_get_reg32(dst_lo, tmp2[1], ctx);
1448 		switch (BPF_SRC(code)) {
1449 		case BPF_X:
1450 			rt = arm_bpf_get_reg32(src_lo, tmp2[0], ctx);
1451 			break;
1452 		case BPF_K:
1453 			rt = tmp2[0];
1454 			emit_a32_mov_i(rt, imm, ctx);
1455 			break;
1456 		default:
1457 			rt = src_lo;
1458 			break;
1459 		}
1460 		emit_udivmod(rd_lo, rd_lo, rt, ctx, BPF_OP(code));
1461 		arm_bpf_put_reg32(dst_lo, rd_lo, ctx);
1462 		if (!ctx->prog->aux->verifier_zext)
1463 			emit_a32_mov_i(dst_hi, 0, ctx);
1464 		break;
1465 	case BPF_ALU64 | BPF_DIV | BPF_K:
1466 	case BPF_ALU64 | BPF_DIV | BPF_X:
1467 	case BPF_ALU64 | BPF_MOD | BPF_K:
1468 	case BPF_ALU64 | BPF_MOD | BPF_X:
1469 		goto notyet;
1470 	/* dst = dst << imm */
1471 	/* dst = dst >> imm */
1472 	/* dst = dst >> imm (signed) */
1473 	case BPF_ALU | BPF_LSH | BPF_K:
1474 	case BPF_ALU | BPF_RSH | BPF_K:
1475 	case BPF_ALU | BPF_ARSH | BPF_K:
1476 		if (unlikely(imm > 31))
1477 			return -EINVAL;
1478 		if (imm)
1479 			emit_a32_alu_i(dst_lo, imm, ctx, BPF_OP(code));
1480 		if (!ctx->prog->aux->verifier_zext)
1481 			emit_a32_mov_i(dst_hi, 0, ctx);
1482 		break;
1483 	/* dst = dst << imm */
1484 	case BPF_ALU64 | BPF_LSH | BPF_K:
1485 		if (unlikely(imm > 63))
1486 			return -EINVAL;
1487 		emit_a32_lsh_i64(dst, imm, ctx);
1488 		break;
1489 	/* dst = dst >> imm */
1490 	case BPF_ALU64 | BPF_RSH | BPF_K:
1491 		if (unlikely(imm > 63))
1492 			return -EINVAL;
1493 		emit_a32_rsh_i64(dst, imm, ctx);
1494 		break;
1495 	/* dst = dst << src */
1496 	case BPF_ALU64 | BPF_LSH | BPF_X:
1497 		emit_a32_lsh_r64(dst, src, ctx);
1498 		break;
1499 	/* dst = dst >> src */
1500 	case BPF_ALU64 | BPF_RSH | BPF_X:
1501 		emit_a32_rsh_r64(dst, src, ctx);
1502 		break;
1503 	/* dst = dst >> src (signed) */
1504 	case BPF_ALU64 | BPF_ARSH | BPF_X:
1505 		emit_a32_arsh_r64(dst, src, ctx);
1506 		break;
1507 	/* dst = dst >> imm (signed) */
1508 	case BPF_ALU64 | BPF_ARSH | BPF_K:
1509 		if (unlikely(imm > 63))
1510 			return -EINVAL;
1511 		emit_a32_arsh_i64(dst, imm, ctx);
1512 		break;
1513 	/* dst = ~dst */
1514 	case BPF_ALU | BPF_NEG:
1515 		emit_a32_alu_i(dst_lo, 0, ctx, BPF_OP(code));
1516 		if (!ctx->prog->aux->verifier_zext)
1517 			emit_a32_mov_i(dst_hi, 0, ctx);
1518 		break;
1519 	/* dst = ~dst (64 bit) */
1520 	case BPF_ALU64 | BPF_NEG:
1521 		emit_a32_neg64(dst, ctx);
1522 		break;
1523 	/* dst = dst * src/imm */
1524 	case BPF_ALU64 | BPF_MUL | BPF_X:
1525 	case BPF_ALU64 | BPF_MUL | BPF_K:
1526 		switch (BPF_SRC(code)) {
1527 		case BPF_X:
1528 			emit_a32_mul_r64(dst, src, ctx);
1529 			break;
1530 		case BPF_K:
1531 			/* Move immediate value to the temporary register
1532 			 * and then do the multiplication on it as this
1533 			 * will sign-extend the immediate value into temp
1534 			 * reg then it would be safe to do the operation
1535 			 * on it.
1536 			 */
1537 			emit_a32_mov_se_i64(is64, tmp2, imm, ctx);
1538 			emit_a32_mul_r64(dst, tmp2, ctx);
1539 			break;
1540 		}
1541 		break;
1542 	/* dst = htole(dst) */
1543 	/* dst = htobe(dst) */
1544 	case BPF_ALU | BPF_END | BPF_FROM_LE:
1545 	case BPF_ALU | BPF_END | BPF_FROM_BE:
1546 		rd = arm_bpf_get_reg64(dst, tmp, ctx);
1547 		if (BPF_SRC(code) == BPF_FROM_LE)
1548 			goto emit_bswap_uxt;
1549 		switch (imm) {
1550 		case 16:
1551 			emit_rev16(rd[1], rd[1], ctx);
1552 			goto emit_bswap_uxt;
1553 		case 32:
1554 			emit_rev32(rd[1], rd[1], ctx);
1555 			goto emit_bswap_uxt;
1556 		case 64:
1557 			emit_rev32(ARM_LR, rd[1], ctx);
1558 			emit_rev32(rd[1], rd[0], ctx);
1559 			emit(ARM_MOV_R(rd[0], ARM_LR), ctx);
1560 			break;
1561 		}
1562 		goto exit;
1563 emit_bswap_uxt:
1564 		switch (imm) {
1565 		case 16:
1566 			/* zero-extend 16 bits into 64 bits */
1567 #if __LINUX_ARM_ARCH__ < 6
1568 			emit_a32_mov_i(tmp2[1], 0xffff, ctx);
1569 			emit(ARM_AND_R(rd[1], rd[1], tmp2[1]), ctx);
1570 #else /* ARMv6+ */
1571 			emit(ARM_UXTH(rd[1], rd[1]), ctx);
1572 #endif
1573 			if (!ctx->prog->aux->verifier_zext)
1574 				emit(ARM_EOR_R(rd[0], rd[0], rd[0]), ctx);
1575 			break;
1576 		case 32:
1577 			/* zero-extend 32 bits into 64 bits */
1578 			if (!ctx->prog->aux->verifier_zext)
1579 				emit(ARM_EOR_R(rd[0], rd[0], rd[0]), ctx);
1580 			break;
1581 		case 64:
1582 			/* nop */
1583 			break;
1584 		}
1585 exit:
1586 		arm_bpf_put_reg64(dst, rd, ctx);
1587 		break;
1588 	/* dst = imm64 */
1589 	case BPF_LD | BPF_IMM | BPF_DW:
1590 	{
1591 		u64 val = (u32)imm | (u64)insn[1].imm << 32;
1592 
1593 		emit_a32_mov_i64(dst, val, ctx);
1594 
1595 		return 1;
1596 	}
1597 	/* LDX: dst = *(size *)(src + off) */
1598 	case BPF_LDX | BPF_MEM | BPF_W:
1599 	case BPF_LDX | BPF_MEM | BPF_H:
1600 	case BPF_LDX | BPF_MEM | BPF_B:
1601 	case BPF_LDX | BPF_MEM | BPF_DW:
1602 		rn = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
1603 		emit_ldx_r(dst, rn, off, ctx, BPF_SIZE(code));
1604 		break;
1605 	/* ST: *(size *)(dst + off) = imm */
1606 	case BPF_ST | BPF_MEM | BPF_W:
1607 	case BPF_ST | BPF_MEM | BPF_H:
1608 	case BPF_ST | BPF_MEM | BPF_B:
1609 	case BPF_ST | BPF_MEM | BPF_DW:
1610 		switch (BPF_SIZE(code)) {
1611 		case BPF_DW:
1612 			/* Sign-extend immediate value into temp reg */
1613 			emit_a32_mov_se_i64(true, tmp2, imm, ctx);
1614 			break;
1615 		case BPF_W:
1616 		case BPF_H:
1617 		case BPF_B:
1618 			emit_a32_mov_i(tmp2[1], imm, ctx);
1619 			break;
1620 		}
1621 		emit_str_r(dst_lo, tmp2, off, ctx, BPF_SIZE(code));
1622 		break;
1623 	/* STX XADD: lock *(u32 *)(dst + off) += src */
1624 	case BPF_STX | BPF_XADD | BPF_W:
1625 	/* STX XADD: lock *(u64 *)(dst + off) += src */
1626 	case BPF_STX | BPF_XADD | BPF_DW:
1627 		goto notyet;
1628 	/* STX: *(size *)(dst + off) = src */
1629 	case BPF_STX | BPF_MEM | BPF_W:
1630 	case BPF_STX | BPF_MEM | BPF_H:
1631 	case BPF_STX | BPF_MEM | BPF_B:
1632 	case BPF_STX | BPF_MEM | BPF_DW:
1633 		rs = arm_bpf_get_reg64(src, tmp2, ctx);
1634 		emit_str_r(dst_lo, rs, off, ctx, BPF_SIZE(code));
1635 		break;
1636 	/* PC += off if dst == src */
1637 	/* PC += off if dst > src */
1638 	/* PC += off if dst >= src */
1639 	/* PC += off if dst < src */
1640 	/* PC += off if dst <= src */
1641 	/* PC += off if dst != src */
1642 	/* PC += off if dst > src (signed) */
1643 	/* PC += off if dst >= src (signed) */
1644 	/* PC += off if dst < src (signed) */
1645 	/* PC += off if dst <= src (signed) */
1646 	/* PC += off if dst & src */
1647 	case BPF_JMP | BPF_JEQ | BPF_X:
1648 	case BPF_JMP | BPF_JGT | BPF_X:
1649 	case BPF_JMP | BPF_JGE | BPF_X:
1650 	case BPF_JMP | BPF_JNE | BPF_X:
1651 	case BPF_JMP | BPF_JSGT | BPF_X:
1652 	case BPF_JMP | BPF_JSGE | BPF_X:
1653 	case BPF_JMP | BPF_JSET | BPF_X:
1654 	case BPF_JMP | BPF_JLE | BPF_X:
1655 	case BPF_JMP | BPF_JLT | BPF_X:
1656 	case BPF_JMP | BPF_JSLT | BPF_X:
1657 	case BPF_JMP | BPF_JSLE | BPF_X:
1658 	case BPF_JMP32 | BPF_JEQ | BPF_X:
1659 	case BPF_JMP32 | BPF_JGT | BPF_X:
1660 	case BPF_JMP32 | BPF_JGE | BPF_X:
1661 	case BPF_JMP32 | BPF_JNE | BPF_X:
1662 	case BPF_JMP32 | BPF_JSGT | BPF_X:
1663 	case BPF_JMP32 | BPF_JSGE | BPF_X:
1664 	case BPF_JMP32 | BPF_JSET | BPF_X:
1665 	case BPF_JMP32 | BPF_JLE | BPF_X:
1666 	case BPF_JMP32 | BPF_JLT | BPF_X:
1667 	case BPF_JMP32 | BPF_JSLT | BPF_X:
1668 	case BPF_JMP32 | BPF_JSLE | BPF_X:
1669 		/* Setup source registers */
1670 		rm = arm_bpf_get_reg32(src_hi, tmp2[0], ctx);
1671 		rn = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
1672 		goto go_jmp;
1673 	/* PC += off if dst == imm */
1674 	/* PC += off if dst > imm */
1675 	/* PC += off if dst >= imm */
1676 	/* PC += off if dst < imm */
1677 	/* PC += off if dst <= imm */
1678 	/* PC += off if dst != imm */
1679 	/* PC += off if dst > imm (signed) */
1680 	/* PC += off if dst >= imm (signed) */
1681 	/* PC += off if dst < imm (signed) */
1682 	/* PC += off if dst <= imm (signed) */
1683 	/* PC += off if dst & imm */
1684 	case BPF_JMP | BPF_JEQ | BPF_K:
1685 	case BPF_JMP | BPF_JGT | BPF_K:
1686 	case BPF_JMP | BPF_JGE | BPF_K:
1687 	case BPF_JMP | BPF_JNE | BPF_K:
1688 	case BPF_JMP | BPF_JSGT | BPF_K:
1689 	case BPF_JMP | BPF_JSGE | BPF_K:
1690 	case BPF_JMP | BPF_JSET | BPF_K:
1691 	case BPF_JMP | BPF_JLT | BPF_K:
1692 	case BPF_JMP | BPF_JLE | BPF_K:
1693 	case BPF_JMP | BPF_JSLT | BPF_K:
1694 	case BPF_JMP | BPF_JSLE | BPF_K:
1695 	case BPF_JMP32 | BPF_JEQ | BPF_K:
1696 	case BPF_JMP32 | BPF_JGT | BPF_K:
1697 	case BPF_JMP32 | BPF_JGE | BPF_K:
1698 	case BPF_JMP32 | BPF_JNE | BPF_K:
1699 	case BPF_JMP32 | BPF_JSGT | BPF_K:
1700 	case BPF_JMP32 | BPF_JSGE | BPF_K:
1701 	case BPF_JMP32 | BPF_JSET | BPF_K:
1702 	case BPF_JMP32 | BPF_JLT | BPF_K:
1703 	case BPF_JMP32 | BPF_JLE | BPF_K:
1704 	case BPF_JMP32 | BPF_JSLT | BPF_K:
1705 	case BPF_JMP32 | BPF_JSLE | BPF_K:
1706 		if (off == 0)
1707 			break;
1708 		rm = tmp2[0];
1709 		rn = tmp2[1];
1710 		/* Sign-extend immediate value */
1711 		emit_a32_mov_se_i64(true, tmp2, imm, ctx);
1712 go_jmp:
1713 		/* Setup destination register */
1714 		rd = arm_bpf_get_reg64(dst, tmp, ctx);
1715 
1716 		/* Check for the condition */
1717 		emit_ar_r(rd[0], rd[1], rm, rn, ctx, BPF_OP(code),
1718 			  BPF_CLASS(code) == BPF_JMP);
1719 
1720 		/* Setup JUMP instruction */
1721 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1722 		switch (BPF_OP(code)) {
1723 		case BPF_JNE:
1724 		case BPF_JSET:
1725 			_emit(ARM_COND_NE, ARM_B(jmp_offset), ctx);
1726 			break;
1727 		case BPF_JEQ:
1728 			_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1729 			break;
1730 		case BPF_JGT:
1731 			_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
1732 			break;
1733 		case BPF_JGE:
1734 			_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
1735 			break;
1736 		case BPF_JSGT:
1737 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1738 			break;
1739 		case BPF_JSGE:
1740 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1741 			break;
1742 		case BPF_JLE:
1743 			_emit(ARM_COND_LS, ARM_B(jmp_offset), ctx);
1744 			break;
1745 		case BPF_JLT:
1746 			_emit(ARM_COND_CC, ARM_B(jmp_offset), ctx);
1747 			break;
1748 		case BPF_JSLT:
1749 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1750 			break;
1751 		case BPF_JSLE:
1752 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1753 			break;
1754 		}
1755 		break;
1756 	/* JMP OFF */
1757 	case BPF_JMP | BPF_JA:
1758 	{
1759 		if (off == 0)
1760 			break;
1761 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1762 		check_imm24(jmp_offset);
1763 		emit(ARM_B(jmp_offset), ctx);
1764 		break;
1765 	}
1766 	/* tail call */
1767 	case BPF_JMP | BPF_TAIL_CALL:
1768 		if (emit_bpf_tail_call(ctx))
1769 			return -EFAULT;
1770 		break;
1771 	/* function call */
1772 	case BPF_JMP | BPF_CALL:
1773 	{
1774 		const s8 *r0 = bpf2a32[BPF_REG_0];
1775 		const s8 *r1 = bpf2a32[BPF_REG_1];
1776 		const s8 *r2 = bpf2a32[BPF_REG_2];
1777 		const s8 *r3 = bpf2a32[BPF_REG_3];
1778 		const s8 *r4 = bpf2a32[BPF_REG_4];
1779 		const s8 *r5 = bpf2a32[BPF_REG_5];
1780 		const u32 func = (u32)__bpf_call_base + (u32)imm;
1781 
1782 		emit_a32_mov_r64(true, r0, r1, ctx);
1783 		emit_a32_mov_r64(true, r1, r2, ctx);
1784 		emit_push_r64(r5, ctx);
1785 		emit_push_r64(r4, ctx);
1786 		emit_push_r64(r3, ctx);
1787 
1788 		emit_a32_mov_i(tmp[1], func, ctx);
1789 		emit_blx_r(tmp[1], ctx);
1790 
1791 		emit(ARM_ADD_I(ARM_SP, ARM_SP, imm8m(24)), ctx); // callee clean
1792 		break;
1793 	}
1794 	/* function return */
1795 	case BPF_JMP | BPF_EXIT:
1796 		/* Optimization: when last instruction is EXIT
1797 		 * simply fallthrough to epilogue.
1798 		 */
1799 		if (i == ctx->prog->len - 1)
1800 			break;
1801 		jmp_offset = epilogue_offset(ctx);
1802 		check_imm24(jmp_offset);
1803 		emit(ARM_B(jmp_offset), ctx);
1804 		break;
1805 notyet:
1806 		pr_info_once("*** NOT YET: opcode %02x ***\n", code);
1807 		return -EFAULT;
1808 	default:
1809 		pr_err_once("unknown opcode %02x\n", code);
1810 		return -EINVAL;
1811 	}
1812 
1813 	if (ctx->flags & FLAG_IMM_OVERFLOW)
1814 		/*
1815 		 * this instruction generated an overflow when
1816 		 * trying to access the literal pool, so
1817 		 * delegate this filter to the kernel interpreter.
1818 		 */
1819 		return -1;
1820 	return 0;
1821 }
1822 
1823 static int build_body(struct jit_ctx *ctx)
1824 {
1825 	const struct bpf_prog *prog = ctx->prog;
1826 	unsigned int i;
1827 
1828 	for (i = 0; i < prog->len; i++) {
1829 		const struct bpf_insn *insn = &(prog->insnsi[i]);
1830 		int ret;
1831 
1832 		ret = build_insn(insn, ctx);
1833 
1834 		/* It's used with loading the 64 bit immediate value. */
1835 		if (ret > 0) {
1836 			i++;
1837 			if (ctx->target == NULL)
1838 				ctx->offsets[i] = ctx->idx;
1839 			continue;
1840 		}
1841 
1842 		if (ctx->target == NULL)
1843 			ctx->offsets[i] = ctx->idx;
1844 
1845 		/* If unsuccesfull, return with error code */
1846 		if (ret)
1847 			return ret;
1848 	}
1849 	return 0;
1850 }
1851 
1852 static int validate_code(struct jit_ctx *ctx)
1853 {
1854 	int i;
1855 
1856 	for (i = 0; i < ctx->idx; i++) {
1857 		if (ctx->target[i] == __opcode_to_mem_arm(ARM_INST_UDF))
1858 			return -1;
1859 	}
1860 
1861 	return 0;
1862 }
1863 
1864 void bpf_jit_compile(struct bpf_prog *prog)
1865 {
1866 	/* Nothing to do here. We support Internal BPF. */
1867 }
1868 
1869 bool bpf_jit_needs_zext(void)
1870 {
1871 	return true;
1872 }
1873 
1874 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1875 {
1876 	struct bpf_prog *tmp, *orig_prog = prog;
1877 	struct bpf_binary_header *header;
1878 	bool tmp_blinded = false;
1879 	struct jit_ctx ctx;
1880 	unsigned int tmp_idx;
1881 	unsigned int image_size;
1882 	u8 *image_ptr;
1883 
1884 	/* If BPF JIT was not enabled then we must fall back to
1885 	 * the interpreter.
1886 	 */
1887 	if (!prog->jit_requested)
1888 		return orig_prog;
1889 
1890 	/* If constant blinding was enabled and we failed during blinding
1891 	 * then we must fall back to the interpreter. Otherwise, we save
1892 	 * the new JITed code.
1893 	 */
1894 	tmp = bpf_jit_blind_constants(prog);
1895 
1896 	if (IS_ERR(tmp))
1897 		return orig_prog;
1898 	if (tmp != prog) {
1899 		tmp_blinded = true;
1900 		prog = tmp;
1901 	}
1902 
1903 	memset(&ctx, 0, sizeof(ctx));
1904 	ctx.prog = prog;
1905 	ctx.cpu_architecture = cpu_architecture();
1906 
1907 	/* Not able to allocate memory for offsets[] , then
1908 	 * we must fall back to the interpreter
1909 	 */
1910 	ctx.offsets = kcalloc(prog->len, sizeof(int), GFP_KERNEL);
1911 	if (ctx.offsets == NULL) {
1912 		prog = orig_prog;
1913 		goto out;
1914 	}
1915 
1916 	/* 1) fake pass to find in the length of the JITed code,
1917 	 * to compute ctx->offsets and other context variables
1918 	 * needed to compute final JITed code.
1919 	 * Also, calculate random starting pointer/start of JITed code
1920 	 * which is prefixed by random number of fault instructions.
1921 	 *
1922 	 * If the first pass fails then there is no chance of it
1923 	 * being successful in the second pass, so just fall back
1924 	 * to the interpreter.
1925 	 */
1926 	if (build_body(&ctx)) {
1927 		prog = orig_prog;
1928 		goto out_off;
1929 	}
1930 
1931 	tmp_idx = ctx.idx;
1932 	build_prologue(&ctx);
1933 	ctx.prologue_bytes = (ctx.idx - tmp_idx) * 4;
1934 
1935 	ctx.epilogue_offset = ctx.idx;
1936 
1937 #if __LINUX_ARM_ARCH__ < 7
1938 	tmp_idx = ctx.idx;
1939 	build_epilogue(&ctx);
1940 	ctx.epilogue_bytes = (ctx.idx - tmp_idx) * 4;
1941 
1942 	ctx.idx += ctx.imm_count;
1943 	if (ctx.imm_count) {
1944 		ctx.imms = kcalloc(ctx.imm_count, sizeof(u32), GFP_KERNEL);
1945 		if (ctx.imms == NULL) {
1946 			prog = orig_prog;
1947 			goto out_off;
1948 		}
1949 	}
1950 #else
1951 	/* there's nothing about the epilogue on ARMv7 */
1952 	build_epilogue(&ctx);
1953 #endif
1954 	/* Now we can get the actual image size of the JITed arm code.
1955 	 * Currently, we are not considering the THUMB-2 instructions
1956 	 * for jit, although it can decrease the size of the image.
1957 	 *
1958 	 * As each arm instruction is of length 32bit, we are translating
1959 	 * number of JITed intructions into the size required to store these
1960 	 * JITed code.
1961 	 */
1962 	image_size = sizeof(u32) * ctx.idx;
1963 
1964 	/* Now we know the size of the structure to make */
1965 	header = bpf_jit_binary_alloc(image_size, &image_ptr,
1966 				      sizeof(u32), jit_fill_hole);
1967 	/* Not able to allocate memory for the structure then
1968 	 * we must fall back to the interpretation
1969 	 */
1970 	if (header == NULL) {
1971 		prog = orig_prog;
1972 		goto out_imms;
1973 	}
1974 
1975 	/* 2.) Actual pass to generate final JIT code */
1976 	ctx.target = (u32 *) image_ptr;
1977 	ctx.idx = 0;
1978 
1979 	build_prologue(&ctx);
1980 
1981 	/* If building the body of the JITed code fails somehow,
1982 	 * we fall back to the interpretation.
1983 	 */
1984 	if (build_body(&ctx) < 0) {
1985 		image_ptr = NULL;
1986 		bpf_jit_binary_free(header);
1987 		prog = orig_prog;
1988 		goto out_imms;
1989 	}
1990 	build_epilogue(&ctx);
1991 
1992 	/* 3.) Extra pass to validate JITed Code */
1993 	if (validate_code(&ctx)) {
1994 		image_ptr = NULL;
1995 		bpf_jit_binary_free(header);
1996 		prog = orig_prog;
1997 		goto out_imms;
1998 	}
1999 	flush_icache_range((u32)header, (u32)(ctx.target + ctx.idx));
2000 
2001 	if (bpf_jit_enable > 1)
2002 		/* there are 2 passes here */
2003 		bpf_jit_dump(prog->len, image_size, 2, ctx.target);
2004 
2005 	bpf_jit_binary_lock_ro(header);
2006 	prog->bpf_func = (void *)ctx.target;
2007 	prog->jited = 1;
2008 	prog->jited_len = image_size;
2009 
2010 out_imms:
2011 #if __LINUX_ARM_ARCH__ < 7
2012 	if (ctx.imm_count)
2013 		kfree(ctx.imms);
2014 #endif
2015 out_off:
2016 	kfree(ctx.offsets);
2017 out:
2018 	if (tmp_blinded)
2019 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
2020 					   tmp : orig_prog);
2021 	return prog;
2022 }
2023 
2024