xref: /linux/arch/arm/net/bpf_jit_32.c (revision ebf68996de0ab250c5d520eb2291ab65643e9a1e)
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 		emit_a32_mov_i(rd[0], 0, ctx);
737 	}
738 
739 	arm_bpf_put_reg64(dst, rd, ctx);
740 }
741 
742 /* dst = src (4 bytes)*/
743 static inline void emit_a32_mov_r(const s8 dst, const s8 src,
744 				  struct jit_ctx *ctx) {
745 	const s8 *tmp = bpf2a32[TMP_REG_1];
746 	s8 rt;
747 
748 	rt = arm_bpf_get_reg32(src, tmp[0], ctx);
749 	arm_bpf_put_reg32(dst, rt, ctx);
750 }
751 
752 /* dst = src */
753 static inline void emit_a32_mov_r64(const bool is64, const s8 dst[],
754 				  const s8 src[],
755 				  struct jit_ctx *ctx) {
756 	if (!is64) {
757 		emit_a32_mov_r(dst_lo, src_lo, ctx);
758 		/* Zero out high 4 bytes */
759 		emit_a32_mov_i(dst_hi, 0, ctx);
760 	} else if (__LINUX_ARM_ARCH__ < 6 &&
761 		   ctx->cpu_architecture < CPU_ARCH_ARMv5TE) {
762 		/* complete 8 byte move */
763 		emit_a32_mov_r(dst_lo, src_lo, ctx);
764 		emit_a32_mov_r(dst_hi, src_hi, ctx);
765 	} else if (is_stacked(src_lo) && is_stacked(dst_lo)) {
766 		const u8 *tmp = bpf2a32[TMP_REG_1];
767 
768 		emit(ARM_LDRD_I(tmp[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(src_lo)), ctx);
769 		emit(ARM_STRD_I(tmp[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(dst_lo)), ctx);
770 	} else if (is_stacked(src_lo)) {
771 		emit(ARM_LDRD_I(dst[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(src_lo)), ctx);
772 	} else if (is_stacked(dst_lo)) {
773 		emit(ARM_STRD_I(src[1], ARM_FP, EBPF_SCRATCH_TO_ARM_FP(dst_lo)), ctx);
774 	} else {
775 		emit(ARM_MOV_R(dst[0], src[0]), ctx);
776 		emit(ARM_MOV_R(dst[1], src[1]), ctx);
777 	}
778 }
779 
780 /* Shift operations */
781 static inline void emit_a32_alu_i(const s8 dst, const u32 val,
782 				struct jit_ctx *ctx, const u8 op) {
783 	const s8 *tmp = bpf2a32[TMP_REG_1];
784 	s8 rd;
785 
786 	rd = arm_bpf_get_reg32(dst, tmp[0], ctx);
787 
788 	/* Do shift operation */
789 	switch (op) {
790 	case BPF_LSH:
791 		emit(ARM_LSL_I(rd, rd, val), ctx);
792 		break;
793 	case BPF_RSH:
794 		emit(ARM_LSR_I(rd, rd, val), ctx);
795 		break;
796 	case BPF_NEG:
797 		emit(ARM_RSB_I(rd, rd, val), ctx);
798 		break;
799 	}
800 
801 	arm_bpf_put_reg32(dst, rd, ctx);
802 }
803 
804 /* dst = ~dst (64 bit) */
805 static inline void emit_a32_neg64(const s8 dst[],
806 				struct jit_ctx *ctx){
807 	const s8 *tmp = bpf2a32[TMP_REG_1];
808 	const s8 *rd;
809 
810 	/* Setup Operand */
811 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
812 
813 	/* Do Negate Operation */
814 	emit(ARM_RSBS_I(rd[1], rd[1], 0), ctx);
815 	emit(ARM_RSC_I(rd[0], rd[0], 0), ctx);
816 
817 	arm_bpf_put_reg64(dst, rd, ctx);
818 }
819 
820 /* dst = dst << src */
821 static inline void emit_a32_lsh_r64(const s8 dst[], const s8 src[],
822 				    struct jit_ctx *ctx) {
823 	const s8 *tmp = bpf2a32[TMP_REG_1];
824 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
825 	const s8 *rd;
826 	s8 rt;
827 
828 	/* Setup Operands */
829 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
830 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
831 
832 	/* Do LSH operation */
833 	emit(ARM_SUB_I(ARM_IP, rt, 32), ctx);
834 	emit(ARM_RSB_I(tmp2[0], rt, 32), ctx);
835 	emit(ARM_MOV_SR(ARM_LR, rd[0], SRTYPE_ASL, rt), ctx);
836 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[1], SRTYPE_ASL, ARM_IP), ctx);
837 	emit(ARM_ORR_SR(ARM_IP, ARM_LR, rd[1], SRTYPE_LSR, tmp2[0]), ctx);
838 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_ASL, rt), ctx);
839 
840 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
841 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
842 }
843 
844 /* dst = dst >> src (signed)*/
845 static inline void emit_a32_arsh_r64(const s8 dst[], const s8 src[],
846 				     struct jit_ctx *ctx) {
847 	const s8 *tmp = bpf2a32[TMP_REG_1];
848 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
849 	const s8 *rd;
850 	s8 rt;
851 
852 	/* Setup Operands */
853 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
854 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
855 
856 	/* Do the ARSH operation */
857 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
858 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
859 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_LSR, rt), ctx);
860 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASL, ARM_IP), ctx);
861 	_emit(ARM_COND_MI, ARM_B(0), ctx);
862 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASR, tmp2[0]), ctx);
863 	emit(ARM_MOV_SR(ARM_IP, rd[0], SRTYPE_ASR, rt), ctx);
864 
865 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
866 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
867 }
868 
869 /* dst = dst >> src */
870 static inline void emit_a32_rsh_r64(const s8 dst[], const s8 src[],
871 				    struct jit_ctx *ctx) {
872 	const s8 *tmp = bpf2a32[TMP_REG_1];
873 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
874 	const s8 *rd;
875 	s8 rt;
876 
877 	/* Setup Operands */
878 	rt = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
879 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
880 
881 	/* Do RSH operation */
882 	emit(ARM_RSB_I(ARM_IP, rt, 32), ctx);
883 	emit(ARM_SUBS_I(tmp2[0], rt, 32), ctx);
884 	emit(ARM_MOV_SR(ARM_LR, rd[1], SRTYPE_LSR, rt), ctx);
885 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_ASL, ARM_IP), ctx);
886 	emit(ARM_ORR_SR(ARM_LR, ARM_LR, rd[0], SRTYPE_LSR, tmp2[0]), ctx);
887 	emit(ARM_MOV_SR(ARM_IP, rd[0], SRTYPE_LSR, rt), ctx);
888 
889 	arm_bpf_put_reg32(dst_lo, ARM_LR, ctx);
890 	arm_bpf_put_reg32(dst_hi, ARM_IP, ctx);
891 }
892 
893 /* dst = dst << val */
894 static inline void emit_a32_lsh_i64(const s8 dst[],
895 				    const u32 val, struct jit_ctx *ctx){
896 	const s8 *tmp = bpf2a32[TMP_REG_1];
897 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
898 	const s8 *rd;
899 
900 	/* Setup operands */
901 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
902 
903 	/* Do LSH operation */
904 	if (val < 32) {
905 		emit(ARM_MOV_SI(tmp2[0], rd[0], SRTYPE_ASL, val), ctx);
906 		emit(ARM_ORR_SI(rd[0], tmp2[0], rd[1], SRTYPE_LSR, 32 - val), ctx);
907 		emit(ARM_MOV_SI(rd[1], rd[1], SRTYPE_ASL, val), ctx);
908 	} else {
909 		if (val == 32)
910 			emit(ARM_MOV_R(rd[0], rd[1]), ctx);
911 		else
912 			emit(ARM_MOV_SI(rd[0], rd[1], SRTYPE_ASL, val - 32), ctx);
913 		emit(ARM_EOR_R(rd[1], rd[1], rd[1]), ctx);
914 	}
915 
916 	arm_bpf_put_reg64(dst, rd, ctx);
917 }
918 
919 /* dst = dst >> val */
920 static inline void emit_a32_rsh_i64(const s8 dst[],
921 				    const u32 val, struct jit_ctx *ctx) {
922 	const s8 *tmp = bpf2a32[TMP_REG_1];
923 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
924 	const s8 *rd;
925 
926 	/* Setup operands */
927 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
928 
929 	/* Do LSR operation */
930 	if (val < 32) {
931 		emit(ARM_MOV_SI(tmp2[1], rd[1], SRTYPE_LSR, val), ctx);
932 		emit(ARM_ORR_SI(rd[1], tmp2[1], rd[0], SRTYPE_ASL, 32 - val), ctx);
933 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_LSR, val), ctx);
934 	} else if (val == 32) {
935 		emit(ARM_MOV_R(rd[1], rd[0]), ctx);
936 		emit(ARM_MOV_I(rd[0], 0), ctx);
937 	} else {
938 		emit(ARM_MOV_SI(rd[1], rd[0], SRTYPE_LSR, val - 32), ctx);
939 		emit(ARM_MOV_I(rd[0], 0), ctx);
940 	}
941 
942 	arm_bpf_put_reg64(dst, rd, ctx);
943 }
944 
945 /* dst = dst >> val (signed) */
946 static inline void emit_a32_arsh_i64(const s8 dst[],
947 				     const u32 val, struct jit_ctx *ctx){
948 	const s8 *tmp = bpf2a32[TMP_REG_1];
949 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
950 	const s8 *rd;
951 
952 	/* Setup operands */
953 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
954 
955 	/* Do ARSH operation */
956 	if (val < 32) {
957 		emit(ARM_MOV_SI(tmp2[1], rd[1], SRTYPE_LSR, val), ctx);
958 		emit(ARM_ORR_SI(rd[1], tmp2[1], rd[0], SRTYPE_ASL, 32 - val), ctx);
959 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, val), ctx);
960 	} else if (val == 32) {
961 		emit(ARM_MOV_R(rd[1], rd[0]), ctx);
962 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, 31), ctx);
963 	} else {
964 		emit(ARM_MOV_SI(rd[1], rd[0], SRTYPE_ASR, val - 32), ctx);
965 		emit(ARM_MOV_SI(rd[0], rd[0], SRTYPE_ASR, 31), ctx);
966 	}
967 
968 	arm_bpf_put_reg64(dst, rd, ctx);
969 }
970 
971 static inline void emit_a32_mul_r64(const s8 dst[], const s8 src[],
972 				    struct jit_ctx *ctx) {
973 	const s8 *tmp = bpf2a32[TMP_REG_1];
974 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
975 	const s8 *rd, *rt;
976 
977 	/* Setup operands for multiplication */
978 	rd = arm_bpf_get_reg64(dst, tmp, ctx);
979 	rt = arm_bpf_get_reg64(src, tmp2, ctx);
980 
981 	/* Do Multiplication */
982 	emit(ARM_MUL(ARM_IP, rd[1], rt[0]), ctx);
983 	emit(ARM_MUL(ARM_LR, rd[0], rt[1]), ctx);
984 	emit(ARM_ADD_R(ARM_LR, ARM_IP, ARM_LR), ctx);
985 
986 	emit(ARM_UMULL(ARM_IP, rd[0], rd[1], rt[1]), ctx);
987 	emit(ARM_ADD_R(rd[0], ARM_LR, rd[0]), ctx);
988 
989 	arm_bpf_put_reg32(dst_lo, ARM_IP, ctx);
990 	arm_bpf_put_reg32(dst_hi, rd[0], ctx);
991 }
992 
993 /* *(size *)(dst + off) = src */
994 static inline void emit_str_r(const s8 dst, const s8 src[],
995 			      s32 off, struct jit_ctx *ctx, const u8 sz){
996 	const s8 *tmp = bpf2a32[TMP_REG_1];
997 	s32 off_max;
998 	s8 rd;
999 
1000 	rd = arm_bpf_get_reg32(dst, tmp[1], ctx);
1001 
1002 	if (sz == BPF_H)
1003 		off_max = 0xff;
1004 	else
1005 		off_max = 0xfff;
1006 
1007 	if (off < 0 || off > off_max) {
1008 		emit_a32_mov_i(tmp[0], off, ctx);
1009 		emit(ARM_ADD_R(tmp[0], tmp[0], rd), ctx);
1010 		rd = tmp[0];
1011 		off = 0;
1012 	}
1013 	switch (sz) {
1014 	case BPF_B:
1015 		/* Store a Byte */
1016 		emit(ARM_STRB_I(src_lo, rd, off), ctx);
1017 		break;
1018 	case BPF_H:
1019 		/* Store a HalfWord */
1020 		emit(ARM_STRH_I(src_lo, rd, off), ctx);
1021 		break;
1022 	case BPF_W:
1023 		/* Store a Word */
1024 		emit(ARM_STR_I(src_lo, rd, off), ctx);
1025 		break;
1026 	case BPF_DW:
1027 		/* Store a Double Word */
1028 		emit(ARM_STR_I(src_lo, rd, off), ctx);
1029 		emit(ARM_STR_I(src_hi, rd, off + 4), ctx);
1030 		break;
1031 	}
1032 }
1033 
1034 /* dst = *(size*)(src + off) */
1035 static inline void emit_ldx_r(const s8 dst[], const s8 src,
1036 			      s32 off, struct jit_ctx *ctx, const u8 sz){
1037 	const s8 *tmp = bpf2a32[TMP_REG_1];
1038 	const s8 *rd = is_stacked(dst_lo) ? tmp : dst;
1039 	s8 rm = src;
1040 	s32 off_max;
1041 
1042 	if (sz == BPF_H)
1043 		off_max = 0xff;
1044 	else
1045 		off_max = 0xfff;
1046 
1047 	if (off < 0 || off > off_max) {
1048 		emit_a32_mov_i(tmp[0], off, ctx);
1049 		emit(ARM_ADD_R(tmp[0], tmp[0], src), ctx);
1050 		rm = tmp[0];
1051 		off = 0;
1052 	} else if (rd[1] == rm) {
1053 		emit(ARM_MOV_R(tmp[0], rm), ctx);
1054 		rm = tmp[0];
1055 	}
1056 	switch (sz) {
1057 	case BPF_B:
1058 		/* Load a Byte */
1059 		emit(ARM_LDRB_I(rd[1], rm, off), ctx);
1060 		emit_a32_mov_i(rd[0], 0, ctx);
1061 		break;
1062 	case BPF_H:
1063 		/* Load a HalfWord */
1064 		emit(ARM_LDRH_I(rd[1], rm, off), ctx);
1065 		emit_a32_mov_i(rd[0], 0, ctx);
1066 		break;
1067 	case BPF_W:
1068 		/* Load a Word */
1069 		emit(ARM_LDR_I(rd[1], rm, off), ctx);
1070 		emit_a32_mov_i(rd[0], 0, ctx);
1071 		break;
1072 	case BPF_DW:
1073 		/* Load a Double Word */
1074 		emit(ARM_LDR_I(rd[1], rm, off), ctx);
1075 		emit(ARM_LDR_I(rd[0], rm, off + 4), ctx);
1076 		break;
1077 	}
1078 	arm_bpf_put_reg64(dst, rd, ctx);
1079 }
1080 
1081 /* Arithmatic Operation */
1082 static inline void emit_ar_r(const u8 rd, const u8 rt, const u8 rm,
1083 			     const u8 rn, struct jit_ctx *ctx, u8 op,
1084 			     bool is_jmp64) {
1085 	switch (op) {
1086 	case BPF_JSET:
1087 		if (is_jmp64) {
1088 			emit(ARM_AND_R(ARM_IP, rt, rn), ctx);
1089 			emit(ARM_AND_R(ARM_LR, rd, rm), ctx);
1090 			emit(ARM_ORRS_R(ARM_IP, ARM_LR, ARM_IP), ctx);
1091 		} else {
1092 			emit(ARM_ANDS_R(ARM_IP, rt, rn), ctx);
1093 		}
1094 		break;
1095 	case BPF_JEQ:
1096 	case BPF_JNE:
1097 	case BPF_JGT:
1098 	case BPF_JGE:
1099 	case BPF_JLE:
1100 	case BPF_JLT:
1101 		if (is_jmp64) {
1102 			emit(ARM_CMP_R(rd, rm), ctx);
1103 			/* Only compare low halve if high halve are equal. */
1104 			_emit(ARM_COND_EQ, ARM_CMP_R(rt, rn), ctx);
1105 		} else {
1106 			emit(ARM_CMP_R(rt, rn), ctx);
1107 		}
1108 		break;
1109 	case BPF_JSLE:
1110 	case BPF_JSGT:
1111 		emit(ARM_CMP_R(rn, rt), ctx);
1112 		if (is_jmp64)
1113 			emit(ARM_SBCS_R(ARM_IP, rm, rd), ctx);
1114 		break;
1115 	case BPF_JSLT:
1116 	case BPF_JSGE:
1117 		emit(ARM_CMP_R(rt, rn), ctx);
1118 		if (is_jmp64)
1119 			emit(ARM_SBCS_R(ARM_IP, rd, rm), ctx);
1120 		break;
1121 	}
1122 }
1123 
1124 static int out_offset = -1; /* initialized on the first pass of build_body() */
1125 static int emit_bpf_tail_call(struct jit_ctx *ctx)
1126 {
1127 
1128 	/* bpf_tail_call(void *prog_ctx, struct bpf_array *array, u64 index) */
1129 	const s8 *r2 = bpf2a32[BPF_REG_2];
1130 	const s8 *r3 = bpf2a32[BPF_REG_3];
1131 	const s8 *tmp = bpf2a32[TMP_REG_1];
1132 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1133 	const s8 *tcc = bpf2a32[TCALL_CNT];
1134 	const s8 *tc;
1135 	const int idx0 = ctx->idx;
1136 #define cur_offset (ctx->idx - idx0)
1137 #define jmp_offset (out_offset - (cur_offset) - 2)
1138 	u32 lo, hi;
1139 	s8 r_array, r_index;
1140 	int off;
1141 
1142 	/* if (index >= array->map.max_entries)
1143 	 *	goto out;
1144 	 */
1145 	BUILD_BUG_ON(offsetof(struct bpf_array, map.max_entries) >
1146 		     ARM_INST_LDST__IMM12);
1147 	off = offsetof(struct bpf_array, map.max_entries);
1148 	r_array = arm_bpf_get_reg32(r2[1], tmp2[0], ctx);
1149 	/* index is 32-bit for arrays */
1150 	r_index = arm_bpf_get_reg32(r3[1], tmp2[1], ctx);
1151 	/* array->map.max_entries */
1152 	emit(ARM_LDR_I(tmp[1], r_array, off), ctx);
1153 	/* index >= array->map.max_entries */
1154 	emit(ARM_CMP_R(r_index, tmp[1]), ctx);
1155 	_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
1156 
1157 	/* tmp2[0] = array, tmp2[1] = index */
1158 
1159 	/* if (tail_call_cnt > MAX_TAIL_CALL_CNT)
1160 	 *	goto out;
1161 	 * tail_call_cnt++;
1162 	 */
1163 	lo = (u32)MAX_TAIL_CALL_CNT;
1164 	hi = (u32)((u64)MAX_TAIL_CALL_CNT >> 32);
1165 	tc = arm_bpf_get_reg64(tcc, tmp, ctx);
1166 	emit(ARM_CMP_I(tc[0], hi), ctx);
1167 	_emit(ARM_COND_EQ, ARM_CMP_I(tc[1], lo), ctx);
1168 	_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
1169 	emit(ARM_ADDS_I(tc[1], tc[1], 1), ctx);
1170 	emit(ARM_ADC_I(tc[0], tc[0], 0), ctx);
1171 	arm_bpf_put_reg64(tcc, tmp, ctx);
1172 
1173 	/* prog = array->ptrs[index]
1174 	 * if (prog == NULL)
1175 	 *	goto out;
1176 	 */
1177 	BUILD_BUG_ON(imm8m(offsetof(struct bpf_array, ptrs)) < 0);
1178 	off = imm8m(offsetof(struct bpf_array, ptrs));
1179 	emit(ARM_ADD_I(tmp[1], r_array, off), ctx);
1180 	emit(ARM_LDR_R_SI(tmp[1], tmp[1], r_index, SRTYPE_ASL, 2), ctx);
1181 	emit(ARM_CMP_I(tmp[1], 0), ctx);
1182 	_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1183 
1184 	/* goto *(prog->bpf_func + prologue_size); */
1185 	BUILD_BUG_ON(offsetof(struct bpf_prog, bpf_func) >
1186 		     ARM_INST_LDST__IMM12);
1187 	off = offsetof(struct bpf_prog, bpf_func);
1188 	emit(ARM_LDR_I(tmp[1], tmp[1], off), ctx);
1189 	emit(ARM_ADD_I(tmp[1], tmp[1], ctx->prologue_bytes), ctx);
1190 	emit_bx_r(tmp[1], ctx);
1191 
1192 	/* out: */
1193 	if (out_offset == -1)
1194 		out_offset = cur_offset;
1195 	if (cur_offset != out_offset) {
1196 		pr_err_once("tail_call out_offset = %d, expected %d!\n",
1197 			    cur_offset, out_offset);
1198 		return -1;
1199 	}
1200 	return 0;
1201 #undef cur_offset
1202 #undef jmp_offset
1203 }
1204 
1205 /* 0xabcd => 0xcdab */
1206 static inline void emit_rev16(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1207 {
1208 #if __LINUX_ARM_ARCH__ < 6
1209 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1210 
1211 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1212 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 8), ctx);
1213 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1214 	emit(ARM_ORR_SI(rd, tmp2[0], tmp2[1], SRTYPE_LSL, 8), ctx);
1215 #else /* ARMv6+ */
1216 	emit(ARM_REV16(rd, rn), ctx);
1217 #endif
1218 }
1219 
1220 /* 0xabcdefgh => 0xghefcdab */
1221 static inline void emit_rev32(const u8 rd, const u8 rn, struct jit_ctx *ctx)
1222 {
1223 #if __LINUX_ARM_ARCH__ < 6
1224 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1225 
1226 	emit(ARM_AND_I(tmp2[1], rn, 0xff), ctx);
1227 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 24), ctx);
1228 	emit(ARM_ORR_SI(ARM_IP, tmp2[0], tmp2[1], SRTYPE_LSL, 24), ctx);
1229 
1230 	emit(ARM_MOV_SI(tmp2[1], rn, SRTYPE_LSR, 8), ctx);
1231 	emit(ARM_AND_I(tmp2[1], tmp2[1], 0xff), ctx);
1232 	emit(ARM_MOV_SI(tmp2[0], rn, SRTYPE_LSR, 16), ctx);
1233 	emit(ARM_AND_I(tmp2[0], tmp2[0], 0xff), ctx);
1234 	emit(ARM_MOV_SI(tmp2[0], tmp2[0], SRTYPE_LSL, 8), ctx);
1235 	emit(ARM_ORR_SI(tmp2[0], tmp2[0], tmp2[1], SRTYPE_LSL, 16), ctx);
1236 	emit(ARM_ORR_R(rd, ARM_IP, tmp2[0]), ctx);
1237 
1238 #else /* ARMv6+ */
1239 	emit(ARM_REV(rd, rn), ctx);
1240 #endif
1241 }
1242 
1243 // push the scratch stack register on top of the stack
1244 static inline void emit_push_r64(const s8 src[], struct jit_ctx *ctx)
1245 {
1246 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1247 	const s8 *rt;
1248 	u16 reg_set = 0;
1249 
1250 	rt = arm_bpf_get_reg64(src, tmp2, ctx);
1251 
1252 	reg_set = (1 << rt[1]) | (1 << rt[0]);
1253 	emit(ARM_PUSH(reg_set), ctx);
1254 }
1255 
1256 static void build_prologue(struct jit_ctx *ctx)
1257 {
1258 	const s8 r0 = bpf2a32[BPF_REG_0][1];
1259 	const s8 r2 = bpf2a32[BPF_REG_1][1];
1260 	const s8 r3 = bpf2a32[BPF_REG_1][0];
1261 	const s8 r4 = bpf2a32[BPF_REG_6][1];
1262 	const s8 fplo = bpf2a32[BPF_REG_FP][1];
1263 	const s8 fphi = bpf2a32[BPF_REG_FP][0];
1264 	const s8 *tcc = bpf2a32[TCALL_CNT];
1265 
1266 	/* Save callee saved registers. */
1267 #ifdef CONFIG_FRAME_POINTER
1268 	u16 reg_set = CALLEE_PUSH_MASK | 1 << ARM_IP | 1 << ARM_PC;
1269 	emit(ARM_MOV_R(ARM_IP, ARM_SP), ctx);
1270 	emit(ARM_PUSH(reg_set), ctx);
1271 	emit(ARM_SUB_I(ARM_FP, ARM_IP, 4), ctx);
1272 #else
1273 	emit(ARM_PUSH(CALLEE_PUSH_MASK), ctx);
1274 	emit(ARM_MOV_R(ARM_FP, ARM_SP), ctx);
1275 #endif
1276 	/* Save frame pointer for later */
1277 	emit(ARM_SUB_I(ARM_IP, ARM_SP, SCRATCH_SIZE), ctx);
1278 
1279 	ctx->stack_size = imm8m(STACK_SIZE);
1280 
1281 	/* Set up function call stack */
1282 	emit(ARM_SUB_I(ARM_SP, ARM_SP, ctx->stack_size), ctx);
1283 
1284 	/* Set up BPF prog stack base register */
1285 	emit_a32_mov_r(fplo, ARM_IP, ctx);
1286 	emit_a32_mov_i(fphi, 0, ctx);
1287 
1288 	/* mov r4, 0 */
1289 	emit(ARM_MOV_I(r4, 0), ctx);
1290 
1291 	/* Move BPF_CTX to BPF_R1 */
1292 	emit(ARM_MOV_R(r3, r4), ctx);
1293 	emit(ARM_MOV_R(r2, r0), ctx);
1294 	/* Initialize Tail Count */
1295 	emit(ARM_STR_I(r4, ARM_FP, EBPF_SCRATCH_TO_ARM_FP(tcc[0])), ctx);
1296 	emit(ARM_STR_I(r4, ARM_FP, EBPF_SCRATCH_TO_ARM_FP(tcc[1])), ctx);
1297 	/* end of prologue */
1298 }
1299 
1300 /* restore callee saved registers. */
1301 static void build_epilogue(struct jit_ctx *ctx)
1302 {
1303 #ifdef CONFIG_FRAME_POINTER
1304 	/* When using frame pointers, some additional registers need to
1305 	 * be loaded. */
1306 	u16 reg_set = CALLEE_POP_MASK | 1 << ARM_SP;
1307 	emit(ARM_SUB_I(ARM_SP, ARM_FP, hweight16(reg_set) * 4), ctx);
1308 	emit(ARM_LDM(ARM_SP, reg_set), ctx);
1309 #else
1310 	/* Restore callee saved registers. */
1311 	emit(ARM_MOV_R(ARM_SP, ARM_FP), ctx);
1312 	emit(ARM_POP(CALLEE_POP_MASK), ctx);
1313 #endif
1314 }
1315 
1316 /*
1317  * Convert an eBPF instruction to native instruction, i.e
1318  * JITs an eBPF instruction.
1319  * Returns :
1320  *	0  - Successfully JITed an 8-byte eBPF instruction
1321  *	>0 - Successfully JITed a 16-byte eBPF instruction
1322  *	<0 - Failed to JIT.
1323  */
1324 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx)
1325 {
1326 	const u8 code = insn->code;
1327 	const s8 *dst = bpf2a32[insn->dst_reg];
1328 	const s8 *src = bpf2a32[insn->src_reg];
1329 	const s8 *tmp = bpf2a32[TMP_REG_1];
1330 	const s8 *tmp2 = bpf2a32[TMP_REG_2];
1331 	const s16 off = insn->off;
1332 	const s32 imm = insn->imm;
1333 	const int i = insn - ctx->prog->insnsi;
1334 	const bool is64 = BPF_CLASS(code) == BPF_ALU64;
1335 	const s8 *rd, *rs;
1336 	s8 rd_lo, rt, rm, rn;
1337 	s32 jmp_offset;
1338 
1339 #define check_imm(bits, imm) do {				\
1340 	if ((imm) >= (1 << ((bits) - 1)) ||			\
1341 	    (imm) < -(1 << ((bits) - 1))) {			\
1342 		pr_info("[%2d] imm=%d(0x%x) out of range\n",	\
1343 			i, imm, imm);				\
1344 		return -EINVAL;					\
1345 	}							\
1346 } while (0)
1347 #define check_imm24(imm) check_imm(24, imm)
1348 
1349 	switch (code) {
1350 	/* ALU operations */
1351 
1352 	/* dst = src */
1353 	case BPF_ALU | BPF_MOV | BPF_K:
1354 	case BPF_ALU | BPF_MOV | BPF_X:
1355 	case BPF_ALU64 | BPF_MOV | BPF_K:
1356 	case BPF_ALU64 | BPF_MOV | BPF_X:
1357 		switch (BPF_SRC(code)) {
1358 		case BPF_X:
1359 			emit_a32_mov_r64(is64, dst, src, ctx);
1360 			break;
1361 		case BPF_K:
1362 			/* Sign-extend immediate value to destination reg */
1363 			emit_a32_mov_se_i64(is64, dst, imm, ctx);
1364 			break;
1365 		}
1366 		break;
1367 	/* dst = dst + src/imm */
1368 	/* dst = dst - src/imm */
1369 	/* dst = dst | src/imm */
1370 	/* dst = dst & src/imm */
1371 	/* dst = dst ^ src/imm */
1372 	/* dst = dst * src/imm */
1373 	/* dst = dst << src */
1374 	/* dst = dst >> src */
1375 	case BPF_ALU | BPF_ADD | BPF_K:
1376 	case BPF_ALU | BPF_ADD | BPF_X:
1377 	case BPF_ALU | BPF_SUB | BPF_K:
1378 	case BPF_ALU | BPF_SUB | BPF_X:
1379 	case BPF_ALU | BPF_OR | BPF_K:
1380 	case BPF_ALU | BPF_OR | BPF_X:
1381 	case BPF_ALU | BPF_AND | BPF_K:
1382 	case BPF_ALU | BPF_AND | BPF_X:
1383 	case BPF_ALU | BPF_XOR | BPF_K:
1384 	case BPF_ALU | BPF_XOR | BPF_X:
1385 	case BPF_ALU | BPF_MUL | BPF_K:
1386 	case BPF_ALU | BPF_MUL | BPF_X:
1387 	case BPF_ALU | BPF_LSH | BPF_X:
1388 	case BPF_ALU | BPF_RSH | BPF_X:
1389 	case BPF_ALU | BPF_ARSH | BPF_K:
1390 	case BPF_ALU | BPF_ARSH | BPF_X:
1391 	case BPF_ALU64 | BPF_ADD | BPF_K:
1392 	case BPF_ALU64 | BPF_ADD | BPF_X:
1393 	case BPF_ALU64 | BPF_SUB | BPF_K:
1394 	case BPF_ALU64 | BPF_SUB | BPF_X:
1395 	case BPF_ALU64 | BPF_OR | BPF_K:
1396 	case BPF_ALU64 | BPF_OR | BPF_X:
1397 	case BPF_ALU64 | BPF_AND | BPF_K:
1398 	case BPF_ALU64 | BPF_AND | BPF_X:
1399 	case BPF_ALU64 | BPF_XOR | BPF_K:
1400 	case BPF_ALU64 | BPF_XOR | BPF_X:
1401 		switch (BPF_SRC(code)) {
1402 		case BPF_X:
1403 			emit_a32_alu_r64(is64, dst, src, ctx, BPF_OP(code));
1404 			break;
1405 		case BPF_K:
1406 			/* Move immediate value to the temporary register
1407 			 * and then do the ALU operation on the temporary
1408 			 * register as this will sign-extend the immediate
1409 			 * value into temporary reg and then it would be
1410 			 * safe to do the operation on it.
1411 			 */
1412 			emit_a32_mov_se_i64(is64, tmp2, imm, ctx);
1413 			emit_a32_alu_r64(is64, dst, tmp2, ctx, BPF_OP(code));
1414 			break;
1415 		}
1416 		break;
1417 	/* dst = dst / src(imm) */
1418 	/* dst = dst % src(imm) */
1419 	case BPF_ALU | BPF_DIV | BPF_K:
1420 	case BPF_ALU | BPF_DIV | BPF_X:
1421 	case BPF_ALU | BPF_MOD | BPF_K:
1422 	case BPF_ALU | BPF_MOD | BPF_X:
1423 		rd_lo = arm_bpf_get_reg32(dst_lo, tmp2[1], ctx);
1424 		switch (BPF_SRC(code)) {
1425 		case BPF_X:
1426 			rt = arm_bpf_get_reg32(src_lo, tmp2[0], ctx);
1427 			break;
1428 		case BPF_K:
1429 			rt = tmp2[0];
1430 			emit_a32_mov_i(rt, imm, ctx);
1431 			break;
1432 		default:
1433 			rt = src_lo;
1434 			break;
1435 		}
1436 		emit_udivmod(rd_lo, rd_lo, rt, ctx, BPF_OP(code));
1437 		arm_bpf_put_reg32(dst_lo, rd_lo, ctx);
1438 		emit_a32_mov_i(dst_hi, 0, ctx);
1439 		break;
1440 	case BPF_ALU64 | BPF_DIV | BPF_K:
1441 	case BPF_ALU64 | BPF_DIV | BPF_X:
1442 	case BPF_ALU64 | BPF_MOD | BPF_K:
1443 	case BPF_ALU64 | BPF_MOD | BPF_X:
1444 		goto notyet;
1445 	/* dst = dst >> imm */
1446 	/* dst = dst << imm */
1447 	case BPF_ALU | BPF_RSH | BPF_K:
1448 	case BPF_ALU | BPF_LSH | BPF_K:
1449 		if (unlikely(imm > 31))
1450 			return -EINVAL;
1451 		if (imm)
1452 			emit_a32_alu_i(dst_lo, imm, ctx, BPF_OP(code));
1453 		emit_a32_mov_i(dst_hi, 0, ctx);
1454 		break;
1455 	/* dst = dst << imm */
1456 	case BPF_ALU64 | BPF_LSH | BPF_K:
1457 		if (unlikely(imm > 63))
1458 			return -EINVAL;
1459 		emit_a32_lsh_i64(dst, imm, ctx);
1460 		break;
1461 	/* dst = dst >> imm */
1462 	case BPF_ALU64 | BPF_RSH | BPF_K:
1463 		if (unlikely(imm > 63))
1464 			return -EINVAL;
1465 		emit_a32_rsh_i64(dst, imm, ctx);
1466 		break;
1467 	/* dst = dst << src */
1468 	case BPF_ALU64 | BPF_LSH | BPF_X:
1469 		emit_a32_lsh_r64(dst, src, ctx);
1470 		break;
1471 	/* dst = dst >> src */
1472 	case BPF_ALU64 | BPF_RSH | BPF_X:
1473 		emit_a32_rsh_r64(dst, src, ctx);
1474 		break;
1475 	/* dst = dst >> src (signed) */
1476 	case BPF_ALU64 | BPF_ARSH | BPF_X:
1477 		emit_a32_arsh_r64(dst, src, ctx);
1478 		break;
1479 	/* dst = dst >> imm (signed) */
1480 	case BPF_ALU64 | BPF_ARSH | BPF_K:
1481 		if (unlikely(imm > 63))
1482 			return -EINVAL;
1483 		emit_a32_arsh_i64(dst, imm, ctx);
1484 		break;
1485 	/* dst = ~dst */
1486 	case BPF_ALU | BPF_NEG:
1487 		emit_a32_alu_i(dst_lo, 0, ctx, BPF_OP(code));
1488 		emit_a32_mov_i(dst_hi, 0, ctx);
1489 		break;
1490 	/* dst = ~dst (64 bit) */
1491 	case BPF_ALU64 | BPF_NEG:
1492 		emit_a32_neg64(dst, ctx);
1493 		break;
1494 	/* dst = dst * src/imm */
1495 	case BPF_ALU64 | BPF_MUL | BPF_X:
1496 	case BPF_ALU64 | BPF_MUL | BPF_K:
1497 		switch (BPF_SRC(code)) {
1498 		case BPF_X:
1499 			emit_a32_mul_r64(dst, src, ctx);
1500 			break;
1501 		case BPF_K:
1502 			/* Move immediate value to the temporary register
1503 			 * and then do the multiplication on it as this
1504 			 * will sign-extend the immediate value into temp
1505 			 * reg then it would be safe to do the operation
1506 			 * on it.
1507 			 */
1508 			emit_a32_mov_se_i64(is64, tmp2, imm, ctx);
1509 			emit_a32_mul_r64(dst, tmp2, ctx);
1510 			break;
1511 		}
1512 		break;
1513 	/* dst = htole(dst) */
1514 	/* dst = htobe(dst) */
1515 	case BPF_ALU | BPF_END | BPF_FROM_LE:
1516 	case BPF_ALU | BPF_END | BPF_FROM_BE:
1517 		rd = arm_bpf_get_reg64(dst, tmp, ctx);
1518 		if (BPF_SRC(code) == BPF_FROM_LE)
1519 			goto emit_bswap_uxt;
1520 		switch (imm) {
1521 		case 16:
1522 			emit_rev16(rd[1], rd[1], ctx);
1523 			goto emit_bswap_uxt;
1524 		case 32:
1525 			emit_rev32(rd[1], rd[1], ctx);
1526 			goto emit_bswap_uxt;
1527 		case 64:
1528 			emit_rev32(ARM_LR, rd[1], ctx);
1529 			emit_rev32(rd[1], rd[0], ctx);
1530 			emit(ARM_MOV_R(rd[0], ARM_LR), ctx);
1531 			break;
1532 		}
1533 		goto exit;
1534 emit_bswap_uxt:
1535 		switch (imm) {
1536 		case 16:
1537 			/* zero-extend 16 bits into 64 bits */
1538 #if __LINUX_ARM_ARCH__ < 6
1539 			emit_a32_mov_i(tmp2[1], 0xffff, ctx);
1540 			emit(ARM_AND_R(rd[1], rd[1], tmp2[1]), ctx);
1541 #else /* ARMv6+ */
1542 			emit(ARM_UXTH(rd[1], rd[1]), ctx);
1543 #endif
1544 			emit(ARM_EOR_R(rd[0], rd[0], rd[0]), ctx);
1545 			break;
1546 		case 32:
1547 			/* zero-extend 32 bits into 64 bits */
1548 			emit(ARM_EOR_R(rd[0], rd[0], rd[0]), ctx);
1549 			break;
1550 		case 64:
1551 			/* nop */
1552 			break;
1553 		}
1554 exit:
1555 		arm_bpf_put_reg64(dst, rd, ctx);
1556 		break;
1557 	/* dst = imm64 */
1558 	case BPF_LD | BPF_IMM | BPF_DW:
1559 	{
1560 		u64 val = (u32)imm | (u64)insn[1].imm << 32;
1561 
1562 		emit_a32_mov_i64(dst, val, ctx);
1563 
1564 		return 1;
1565 	}
1566 	/* LDX: dst = *(size *)(src + off) */
1567 	case BPF_LDX | BPF_MEM | BPF_W:
1568 	case BPF_LDX | BPF_MEM | BPF_H:
1569 	case BPF_LDX | BPF_MEM | BPF_B:
1570 	case BPF_LDX | BPF_MEM | BPF_DW:
1571 		rn = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
1572 		emit_ldx_r(dst, rn, off, ctx, BPF_SIZE(code));
1573 		break;
1574 	/* ST: *(size *)(dst + off) = imm */
1575 	case BPF_ST | BPF_MEM | BPF_W:
1576 	case BPF_ST | BPF_MEM | BPF_H:
1577 	case BPF_ST | BPF_MEM | BPF_B:
1578 	case BPF_ST | BPF_MEM | BPF_DW:
1579 		switch (BPF_SIZE(code)) {
1580 		case BPF_DW:
1581 			/* Sign-extend immediate value into temp reg */
1582 			emit_a32_mov_se_i64(true, tmp2, imm, ctx);
1583 			break;
1584 		case BPF_W:
1585 		case BPF_H:
1586 		case BPF_B:
1587 			emit_a32_mov_i(tmp2[1], imm, ctx);
1588 			break;
1589 		}
1590 		emit_str_r(dst_lo, tmp2, off, ctx, BPF_SIZE(code));
1591 		break;
1592 	/* STX XADD: lock *(u32 *)(dst + off) += src */
1593 	case BPF_STX | BPF_XADD | BPF_W:
1594 	/* STX XADD: lock *(u64 *)(dst + off) += src */
1595 	case BPF_STX | BPF_XADD | BPF_DW:
1596 		goto notyet;
1597 	/* STX: *(size *)(dst + off) = src */
1598 	case BPF_STX | BPF_MEM | BPF_W:
1599 	case BPF_STX | BPF_MEM | BPF_H:
1600 	case BPF_STX | BPF_MEM | BPF_B:
1601 	case BPF_STX | BPF_MEM | BPF_DW:
1602 		rs = arm_bpf_get_reg64(src, tmp2, ctx);
1603 		emit_str_r(dst_lo, rs, off, ctx, BPF_SIZE(code));
1604 		break;
1605 	/* PC += off if dst == src */
1606 	/* PC += off if dst > src */
1607 	/* PC += off if dst >= src */
1608 	/* PC += off if dst < src */
1609 	/* PC += off if dst <= src */
1610 	/* PC += off if dst != src */
1611 	/* PC += off if dst > src (signed) */
1612 	/* PC += off if dst >= src (signed) */
1613 	/* PC += off if dst < src (signed) */
1614 	/* PC += off if dst <= src (signed) */
1615 	/* PC += off if dst & src */
1616 	case BPF_JMP | BPF_JEQ | BPF_X:
1617 	case BPF_JMP | BPF_JGT | BPF_X:
1618 	case BPF_JMP | BPF_JGE | BPF_X:
1619 	case BPF_JMP | BPF_JNE | BPF_X:
1620 	case BPF_JMP | BPF_JSGT | BPF_X:
1621 	case BPF_JMP | BPF_JSGE | BPF_X:
1622 	case BPF_JMP | BPF_JSET | BPF_X:
1623 	case BPF_JMP | BPF_JLE | BPF_X:
1624 	case BPF_JMP | BPF_JLT | BPF_X:
1625 	case BPF_JMP | BPF_JSLT | BPF_X:
1626 	case BPF_JMP | BPF_JSLE | BPF_X:
1627 	case BPF_JMP32 | BPF_JEQ | BPF_X:
1628 	case BPF_JMP32 | BPF_JGT | BPF_X:
1629 	case BPF_JMP32 | BPF_JGE | BPF_X:
1630 	case BPF_JMP32 | BPF_JNE | BPF_X:
1631 	case BPF_JMP32 | BPF_JSGT | BPF_X:
1632 	case BPF_JMP32 | BPF_JSGE | BPF_X:
1633 	case BPF_JMP32 | BPF_JSET | BPF_X:
1634 	case BPF_JMP32 | BPF_JLE | BPF_X:
1635 	case BPF_JMP32 | BPF_JLT | BPF_X:
1636 	case BPF_JMP32 | BPF_JSLT | BPF_X:
1637 	case BPF_JMP32 | BPF_JSLE | BPF_X:
1638 		/* Setup source registers */
1639 		rm = arm_bpf_get_reg32(src_hi, tmp2[0], ctx);
1640 		rn = arm_bpf_get_reg32(src_lo, tmp2[1], ctx);
1641 		goto go_jmp;
1642 	/* PC += off if dst == imm */
1643 	/* PC += off if dst > imm */
1644 	/* PC += off if dst >= imm */
1645 	/* PC += off if dst < imm */
1646 	/* PC += off if dst <= imm */
1647 	/* PC += off if dst != imm */
1648 	/* PC += off if dst > imm (signed) */
1649 	/* PC += off if dst >= imm (signed) */
1650 	/* PC += off if dst < imm (signed) */
1651 	/* PC += off if dst <= imm (signed) */
1652 	/* PC += off if dst & imm */
1653 	case BPF_JMP | BPF_JEQ | BPF_K:
1654 	case BPF_JMP | BPF_JGT | BPF_K:
1655 	case BPF_JMP | BPF_JGE | BPF_K:
1656 	case BPF_JMP | BPF_JNE | BPF_K:
1657 	case BPF_JMP | BPF_JSGT | BPF_K:
1658 	case BPF_JMP | BPF_JSGE | BPF_K:
1659 	case BPF_JMP | BPF_JSET | BPF_K:
1660 	case BPF_JMP | BPF_JLT | BPF_K:
1661 	case BPF_JMP | BPF_JLE | BPF_K:
1662 	case BPF_JMP | BPF_JSLT | BPF_K:
1663 	case BPF_JMP | BPF_JSLE | BPF_K:
1664 	case BPF_JMP32 | BPF_JEQ | BPF_K:
1665 	case BPF_JMP32 | BPF_JGT | BPF_K:
1666 	case BPF_JMP32 | BPF_JGE | BPF_K:
1667 	case BPF_JMP32 | BPF_JNE | BPF_K:
1668 	case BPF_JMP32 | BPF_JSGT | BPF_K:
1669 	case BPF_JMP32 | BPF_JSGE | BPF_K:
1670 	case BPF_JMP32 | BPF_JSET | BPF_K:
1671 	case BPF_JMP32 | BPF_JLT | BPF_K:
1672 	case BPF_JMP32 | BPF_JLE | BPF_K:
1673 	case BPF_JMP32 | BPF_JSLT | BPF_K:
1674 	case BPF_JMP32 | BPF_JSLE | BPF_K:
1675 		if (off == 0)
1676 			break;
1677 		rm = tmp2[0];
1678 		rn = tmp2[1];
1679 		/* Sign-extend immediate value */
1680 		emit_a32_mov_se_i64(true, tmp2, imm, ctx);
1681 go_jmp:
1682 		/* Setup destination register */
1683 		rd = arm_bpf_get_reg64(dst, tmp, ctx);
1684 
1685 		/* Check for the condition */
1686 		emit_ar_r(rd[0], rd[1], rm, rn, ctx, BPF_OP(code),
1687 			  BPF_CLASS(code) == BPF_JMP);
1688 
1689 		/* Setup JUMP instruction */
1690 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1691 		switch (BPF_OP(code)) {
1692 		case BPF_JNE:
1693 		case BPF_JSET:
1694 			_emit(ARM_COND_NE, ARM_B(jmp_offset), ctx);
1695 			break;
1696 		case BPF_JEQ:
1697 			_emit(ARM_COND_EQ, ARM_B(jmp_offset), ctx);
1698 			break;
1699 		case BPF_JGT:
1700 			_emit(ARM_COND_HI, ARM_B(jmp_offset), ctx);
1701 			break;
1702 		case BPF_JGE:
1703 			_emit(ARM_COND_CS, ARM_B(jmp_offset), ctx);
1704 			break;
1705 		case BPF_JSGT:
1706 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1707 			break;
1708 		case BPF_JSGE:
1709 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1710 			break;
1711 		case BPF_JLE:
1712 			_emit(ARM_COND_LS, ARM_B(jmp_offset), ctx);
1713 			break;
1714 		case BPF_JLT:
1715 			_emit(ARM_COND_CC, ARM_B(jmp_offset), ctx);
1716 			break;
1717 		case BPF_JSLT:
1718 			_emit(ARM_COND_LT, ARM_B(jmp_offset), ctx);
1719 			break;
1720 		case BPF_JSLE:
1721 			_emit(ARM_COND_GE, ARM_B(jmp_offset), ctx);
1722 			break;
1723 		}
1724 		break;
1725 	/* JMP OFF */
1726 	case BPF_JMP | BPF_JA:
1727 	{
1728 		if (off == 0)
1729 			break;
1730 		jmp_offset = bpf2a32_offset(i+off, i, ctx);
1731 		check_imm24(jmp_offset);
1732 		emit(ARM_B(jmp_offset), ctx);
1733 		break;
1734 	}
1735 	/* tail call */
1736 	case BPF_JMP | BPF_TAIL_CALL:
1737 		if (emit_bpf_tail_call(ctx))
1738 			return -EFAULT;
1739 		break;
1740 	/* function call */
1741 	case BPF_JMP | BPF_CALL:
1742 	{
1743 		const s8 *r0 = bpf2a32[BPF_REG_0];
1744 		const s8 *r1 = bpf2a32[BPF_REG_1];
1745 		const s8 *r2 = bpf2a32[BPF_REG_2];
1746 		const s8 *r3 = bpf2a32[BPF_REG_3];
1747 		const s8 *r4 = bpf2a32[BPF_REG_4];
1748 		const s8 *r5 = bpf2a32[BPF_REG_5];
1749 		const u32 func = (u32)__bpf_call_base + (u32)imm;
1750 
1751 		emit_a32_mov_r64(true, r0, r1, ctx);
1752 		emit_a32_mov_r64(true, r1, r2, ctx);
1753 		emit_push_r64(r5, ctx);
1754 		emit_push_r64(r4, ctx);
1755 		emit_push_r64(r3, ctx);
1756 
1757 		emit_a32_mov_i(tmp[1], func, ctx);
1758 		emit_blx_r(tmp[1], ctx);
1759 
1760 		emit(ARM_ADD_I(ARM_SP, ARM_SP, imm8m(24)), ctx); // callee clean
1761 		break;
1762 	}
1763 	/* function return */
1764 	case BPF_JMP | BPF_EXIT:
1765 		/* Optimization: when last instruction is EXIT
1766 		 * simply fallthrough to epilogue.
1767 		 */
1768 		if (i == ctx->prog->len - 1)
1769 			break;
1770 		jmp_offset = epilogue_offset(ctx);
1771 		check_imm24(jmp_offset);
1772 		emit(ARM_B(jmp_offset), ctx);
1773 		break;
1774 notyet:
1775 		pr_info_once("*** NOT YET: opcode %02x ***\n", code);
1776 		return -EFAULT;
1777 	default:
1778 		pr_err_once("unknown opcode %02x\n", code);
1779 		return -EINVAL;
1780 	}
1781 
1782 	if (ctx->flags & FLAG_IMM_OVERFLOW)
1783 		/*
1784 		 * this instruction generated an overflow when
1785 		 * trying to access the literal pool, so
1786 		 * delegate this filter to the kernel interpreter.
1787 		 */
1788 		return -1;
1789 	return 0;
1790 }
1791 
1792 static int build_body(struct jit_ctx *ctx)
1793 {
1794 	const struct bpf_prog *prog = ctx->prog;
1795 	unsigned int i;
1796 
1797 	for (i = 0; i < prog->len; i++) {
1798 		const struct bpf_insn *insn = &(prog->insnsi[i]);
1799 		int ret;
1800 
1801 		ret = build_insn(insn, ctx);
1802 
1803 		/* It's used with loading the 64 bit immediate value. */
1804 		if (ret > 0) {
1805 			i++;
1806 			if (ctx->target == NULL)
1807 				ctx->offsets[i] = ctx->idx;
1808 			continue;
1809 		}
1810 
1811 		if (ctx->target == NULL)
1812 			ctx->offsets[i] = ctx->idx;
1813 
1814 		/* If unsuccesfull, return with error code */
1815 		if (ret)
1816 			return ret;
1817 	}
1818 	return 0;
1819 }
1820 
1821 static int validate_code(struct jit_ctx *ctx)
1822 {
1823 	int i;
1824 
1825 	for (i = 0; i < ctx->idx; i++) {
1826 		if (ctx->target[i] == __opcode_to_mem_arm(ARM_INST_UDF))
1827 			return -1;
1828 	}
1829 
1830 	return 0;
1831 }
1832 
1833 void bpf_jit_compile(struct bpf_prog *prog)
1834 {
1835 	/* Nothing to do here. We support Internal BPF. */
1836 }
1837 
1838 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1839 {
1840 	struct bpf_prog *tmp, *orig_prog = prog;
1841 	struct bpf_binary_header *header;
1842 	bool tmp_blinded = false;
1843 	struct jit_ctx ctx;
1844 	unsigned int tmp_idx;
1845 	unsigned int image_size;
1846 	u8 *image_ptr;
1847 
1848 	/* If BPF JIT was not enabled then we must fall back to
1849 	 * the interpreter.
1850 	 */
1851 	if (!prog->jit_requested)
1852 		return orig_prog;
1853 
1854 	/* If constant blinding was enabled and we failed during blinding
1855 	 * then we must fall back to the interpreter. Otherwise, we save
1856 	 * the new JITed code.
1857 	 */
1858 	tmp = bpf_jit_blind_constants(prog);
1859 
1860 	if (IS_ERR(tmp))
1861 		return orig_prog;
1862 	if (tmp != prog) {
1863 		tmp_blinded = true;
1864 		prog = tmp;
1865 	}
1866 
1867 	memset(&ctx, 0, sizeof(ctx));
1868 	ctx.prog = prog;
1869 	ctx.cpu_architecture = cpu_architecture();
1870 
1871 	/* Not able to allocate memory for offsets[] , then
1872 	 * we must fall back to the interpreter
1873 	 */
1874 	ctx.offsets = kcalloc(prog->len, sizeof(int), GFP_KERNEL);
1875 	if (ctx.offsets == NULL) {
1876 		prog = orig_prog;
1877 		goto out;
1878 	}
1879 
1880 	/* 1) fake pass to find in the length of the JITed code,
1881 	 * to compute ctx->offsets and other context variables
1882 	 * needed to compute final JITed code.
1883 	 * Also, calculate random starting pointer/start of JITed code
1884 	 * which is prefixed by random number of fault instructions.
1885 	 *
1886 	 * If the first pass fails then there is no chance of it
1887 	 * being successful in the second pass, so just fall back
1888 	 * to the interpreter.
1889 	 */
1890 	if (build_body(&ctx)) {
1891 		prog = orig_prog;
1892 		goto out_off;
1893 	}
1894 
1895 	tmp_idx = ctx.idx;
1896 	build_prologue(&ctx);
1897 	ctx.prologue_bytes = (ctx.idx - tmp_idx) * 4;
1898 
1899 	ctx.epilogue_offset = ctx.idx;
1900 
1901 #if __LINUX_ARM_ARCH__ < 7
1902 	tmp_idx = ctx.idx;
1903 	build_epilogue(&ctx);
1904 	ctx.epilogue_bytes = (ctx.idx - tmp_idx) * 4;
1905 
1906 	ctx.idx += ctx.imm_count;
1907 	if (ctx.imm_count) {
1908 		ctx.imms = kcalloc(ctx.imm_count, sizeof(u32), GFP_KERNEL);
1909 		if (ctx.imms == NULL) {
1910 			prog = orig_prog;
1911 			goto out_off;
1912 		}
1913 	}
1914 #else
1915 	/* there's nothing about the epilogue on ARMv7 */
1916 	build_epilogue(&ctx);
1917 #endif
1918 	/* Now we can get the actual image size of the JITed arm code.
1919 	 * Currently, we are not considering the THUMB-2 instructions
1920 	 * for jit, although it can decrease the size of the image.
1921 	 *
1922 	 * As each arm instruction is of length 32bit, we are translating
1923 	 * number of JITed intructions into the size required to store these
1924 	 * JITed code.
1925 	 */
1926 	image_size = sizeof(u32) * ctx.idx;
1927 
1928 	/* Now we know the size of the structure to make */
1929 	header = bpf_jit_binary_alloc(image_size, &image_ptr,
1930 				      sizeof(u32), jit_fill_hole);
1931 	/* Not able to allocate memory for the structure then
1932 	 * we must fall back to the interpretation
1933 	 */
1934 	if (header == NULL) {
1935 		prog = orig_prog;
1936 		goto out_imms;
1937 	}
1938 
1939 	/* 2.) Actual pass to generate final JIT code */
1940 	ctx.target = (u32 *) image_ptr;
1941 	ctx.idx = 0;
1942 
1943 	build_prologue(&ctx);
1944 
1945 	/* If building the body of the JITed code fails somehow,
1946 	 * we fall back to the interpretation.
1947 	 */
1948 	if (build_body(&ctx) < 0) {
1949 		image_ptr = NULL;
1950 		bpf_jit_binary_free(header);
1951 		prog = orig_prog;
1952 		goto out_imms;
1953 	}
1954 	build_epilogue(&ctx);
1955 
1956 	/* 3.) Extra pass to validate JITed Code */
1957 	if (validate_code(&ctx)) {
1958 		image_ptr = NULL;
1959 		bpf_jit_binary_free(header);
1960 		prog = orig_prog;
1961 		goto out_imms;
1962 	}
1963 	flush_icache_range((u32)header, (u32)(ctx.target + ctx.idx));
1964 
1965 	if (bpf_jit_enable > 1)
1966 		/* there are 2 passes here */
1967 		bpf_jit_dump(prog->len, image_size, 2, ctx.target);
1968 
1969 	bpf_jit_binary_lock_ro(header);
1970 	prog->bpf_func = (void *)ctx.target;
1971 	prog->jited = 1;
1972 	prog->jited_len = image_size;
1973 
1974 out_imms:
1975 #if __LINUX_ARM_ARCH__ < 7
1976 	if (ctx.imm_count)
1977 		kfree(ctx.imms);
1978 #endif
1979 out_off:
1980 	kfree(ctx.offsets);
1981 out:
1982 	if (tmp_blinded)
1983 		bpf_jit_prog_release_other(prog, prog == orig_prog ?
1984 					   tmp : orig_prog);
1985 	return prog;
1986 }
1987 
1988