xref: /openbmc/linux/arch/arm/net/bpf_jit_32.c (revision 0f4630f3)
1 /*
2  * Just-In-Time compiler for BPF filters on 32bit ARM
3  *
4  * Copyright (c) 2011 Mircea Gherzan <mgherzan@gmail.com>
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms of the GNU General Public License as published by the
8  * Free Software Foundation; version 2 of the License.
9  */
10 
11 #include <linux/bitops.h>
12 #include <linux/compiler.h>
13 #include <linux/errno.h>
14 #include <linux/filter.h>
15 #include <linux/netdevice.h>
16 #include <linux/string.h>
17 #include <linux/slab.h>
18 #include <linux/if_vlan.h>
19 
20 #include <asm/cacheflush.h>
21 #include <asm/hwcap.h>
22 #include <asm/opcodes.h>
23 
24 #include "bpf_jit_32.h"
25 
26 /*
27  * ABI:
28  *
29  * r0	scratch register
30  * r4	BPF register A
31  * r5	BPF register X
32  * r6	pointer to the skb
33  * r7	skb->data
34  * r8	skb_headlen(skb)
35  */
36 
37 #define r_scratch	ARM_R0
38 /* r1-r3 are (also) used for the unaligned loads on the non-ARMv7 slowpath */
39 #define r_off		ARM_R1
40 #define r_A		ARM_R4
41 #define r_X		ARM_R5
42 #define r_skb		ARM_R6
43 #define r_skb_data	ARM_R7
44 #define r_skb_hl	ARM_R8
45 
46 #define SCRATCH_SP_OFFSET	0
47 #define SCRATCH_OFF(k)		(SCRATCH_SP_OFFSET + 4 * (k))
48 
49 #define SEEN_MEM		((1 << BPF_MEMWORDS) - 1)
50 #define SEEN_MEM_WORD(k)	(1 << (k))
51 #define SEEN_X			(1 << BPF_MEMWORDS)
52 #define SEEN_CALL		(1 << (BPF_MEMWORDS + 1))
53 #define SEEN_SKB		(1 << (BPF_MEMWORDS + 2))
54 #define SEEN_DATA		(1 << (BPF_MEMWORDS + 3))
55 
56 #define FLAG_NEED_X_RESET	(1 << 0)
57 #define FLAG_IMM_OVERFLOW	(1 << 1)
58 
59 struct jit_ctx {
60 	const struct bpf_prog *skf;
61 	unsigned idx;
62 	unsigned prologue_bytes;
63 	int ret0_fp_idx;
64 	u32 seen;
65 	u32 flags;
66 	u32 *offsets;
67 	u32 *target;
68 #if __LINUX_ARM_ARCH__ < 7
69 	u16 epilogue_bytes;
70 	u16 imm_count;
71 	u32 *imms;
72 #endif
73 };
74 
75 int bpf_jit_enable __read_mostly;
76 
77 static inline int call_neg_helper(struct sk_buff *skb, int offset, void *ret,
78 		      unsigned int size)
79 {
80 	void *ptr = bpf_internal_load_pointer_neg_helper(skb, offset, size);
81 
82 	if (!ptr)
83 		return -EFAULT;
84 	memcpy(ret, ptr, size);
85 	return 0;
86 }
87 
88 static u64 jit_get_skb_b(struct sk_buff *skb, int offset)
89 {
90 	u8 ret;
91 	int err;
92 
93 	if (offset < 0)
94 		err = call_neg_helper(skb, offset, &ret, 1);
95 	else
96 		err = skb_copy_bits(skb, offset, &ret, 1);
97 
98 	return (u64)err << 32 | ret;
99 }
100 
101 static u64 jit_get_skb_h(struct sk_buff *skb, int offset)
102 {
103 	u16 ret;
104 	int err;
105 
106 	if (offset < 0)
107 		err = call_neg_helper(skb, offset, &ret, 2);
108 	else
109 		err = skb_copy_bits(skb, offset, &ret, 2);
110 
111 	return (u64)err << 32 | ntohs(ret);
112 }
113 
114 static u64 jit_get_skb_w(struct sk_buff *skb, int offset)
115 {
116 	u32 ret;
117 	int err;
118 
119 	if (offset < 0)
120 		err = call_neg_helper(skb, offset, &ret, 4);
121 	else
122 		err = skb_copy_bits(skb, offset, &ret, 4);
123 
124 	return (u64)err << 32 | ntohl(ret);
125 }
126 
127 /*
128  * Wrappers which handle both OABI and EABI and assures Thumb2 interworking
129  * (where the assembly routines like __aeabi_uidiv could cause problems).
130  */
131 static u32 jit_udiv(u32 dividend, u32 divisor)
132 {
133 	return dividend / divisor;
134 }
135 
136 static u32 jit_mod(u32 dividend, u32 divisor)
137 {
138 	return dividend % divisor;
139 }
140 
141 static inline void _emit(int cond, u32 inst, struct jit_ctx *ctx)
142 {
143 	inst |= (cond << 28);
144 	inst = __opcode_to_mem_arm(inst);
145 
146 	if (ctx->target != NULL)
147 		ctx->target[ctx->idx] = inst;
148 
149 	ctx->idx++;
150 }
151 
152 /*
153  * Emit an instruction that will be executed unconditionally.
154  */
155 static inline void emit(u32 inst, struct jit_ctx *ctx)
156 {
157 	_emit(ARM_COND_AL, inst, ctx);
158 }
159 
160 static u16 saved_regs(struct jit_ctx *ctx)
161 {
162 	u16 ret = 0;
163 
164 	if ((ctx->skf->len > 1) ||
165 	    (ctx->skf->insns[0].code == (BPF_RET | BPF_A)))
166 		ret |= 1 << r_A;
167 
168 #ifdef CONFIG_FRAME_POINTER
169 	ret |= (1 << ARM_FP) | (1 << ARM_IP) | (1 << ARM_LR) | (1 << ARM_PC);
170 #else
171 	if (ctx->seen & SEEN_CALL)
172 		ret |= 1 << ARM_LR;
173 #endif
174 	if (ctx->seen & (SEEN_DATA | SEEN_SKB))
175 		ret |= 1 << r_skb;
176 	if (ctx->seen & SEEN_DATA)
177 		ret |= (1 << r_skb_data) | (1 << r_skb_hl);
178 	if (ctx->seen & SEEN_X)
179 		ret |= 1 << r_X;
180 
181 	return ret;
182 }
183 
184 static inline int mem_words_used(struct jit_ctx *ctx)
185 {
186 	/* yes, we do waste some stack space IF there are "holes" in the set" */
187 	return fls(ctx->seen & SEEN_MEM);
188 }
189 
190 static inline bool is_load_to_a(u16 inst)
191 {
192 	switch (inst) {
193 	case BPF_LD | BPF_W | BPF_LEN:
194 	case BPF_LD | BPF_W | BPF_ABS:
195 	case BPF_LD | BPF_H | BPF_ABS:
196 	case BPF_LD | BPF_B | BPF_ABS:
197 		return true;
198 	default:
199 		return false;
200 	}
201 }
202 
203 static void jit_fill_hole(void *area, unsigned int size)
204 {
205 	u32 *ptr;
206 	/* We are guaranteed to have aligned memory. */
207 	for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
208 		*ptr++ = __opcode_to_mem_arm(ARM_INST_UDF);
209 }
210 
211 static void build_prologue(struct jit_ctx *ctx)
212 {
213 	u16 reg_set = saved_regs(ctx);
214 	u16 first_inst = ctx->skf->insns[0].code;
215 	u16 off;
216 
217 #ifdef CONFIG_FRAME_POINTER
218 	emit(ARM_MOV_R(ARM_IP, ARM_SP), ctx);
219 	emit(ARM_PUSH(reg_set), ctx);
220 	emit(ARM_SUB_I(ARM_FP, ARM_IP, 4), ctx);
221 #else
222 	if (reg_set)
223 		emit(ARM_PUSH(reg_set), ctx);
224 #endif
225 
226 	if (ctx->seen & (SEEN_DATA | SEEN_SKB))
227 		emit(ARM_MOV_R(r_skb, ARM_R0), ctx);
228 
229 	if (ctx->seen & SEEN_DATA) {
230 		off = offsetof(struct sk_buff, data);
231 		emit(ARM_LDR_I(r_skb_data, r_skb, off), ctx);
232 		/* headlen = len - data_len */
233 		off = offsetof(struct sk_buff, len);
234 		emit(ARM_LDR_I(r_skb_hl, r_skb, off), ctx);
235 		off = offsetof(struct sk_buff, data_len);
236 		emit(ARM_LDR_I(r_scratch, r_skb, off), ctx);
237 		emit(ARM_SUB_R(r_skb_hl, r_skb_hl, r_scratch), ctx);
238 	}
239 
240 	if (ctx->flags & FLAG_NEED_X_RESET)
241 		emit(ARM_MOV_I(r_X, 0), ctx);
242 
243 	/* do not leak kernel data to userspace */
244 	if ((first_inst != (BPF_RET | BPF_K)) && !(is_load_to_a(first_inst)))
245 		emit(ARM_MOV_I(r_A, 0), ctx);
246 
247 	/* stack space for the BPF_MEM words */
248 	if (ctx->seen & SEEN_MEM)
249 		emit(ARM_SUB_I(ARM_SP, ARM_SP, mem_words_used(ctx) * 4), ctx);
250 }
251 
252 static void build_epilogue(struct jit_ctx *ctx)
253 {
254 	u16 reg_set = saved_regs(ctx);
255 
256 	if (ctx->seen & SEEN_MEM)
257 		emit(ARM_ADD_I(ARM_SP, ARM_SP, mem_words_used(ctx) * 4), ctx);
258 
259 	reg_set &= ~(1 << ARM_LR);
260 
261 #ifdef CONFIG_FRAME_POINTER
262 	/* the first instruction of the prologue was: mov ip, sp */
263 	reg_set &= ~(1 << ARM_IP);
264 	reg_set |= (1 << ARM_SP);
265 	emit(ARM_LDM(ARM_SP, reg_set), ctx);
266 #else
267 	if (reg_set) {
268 		if (ctx->seen & SEEN_CALL)
269 			reg_set |= 1 << ARM_PC;
270 		emit(ARM_POP(reg_set), ctx);
271 	}
272 
273 	if (!(ctx->seen & SEEN_CALL))
274 		emit(ARM_BX(ARM_LR), ctx);
275 #endif
276 }
277 
278 static int16_t imm8m(u32 x)
279 {
280 	u32 rot;
281 
282 	for (rot = 0; rot < 16; rot++)
283 		if ((x & ~ror32(0xff, 2 * rot)) == 0)
284 			return rol32(x, 2 * rot) | (rot << 8);
285 
286 	return -1;
287 }
288 
289 #if __LINUX_ARM_ARCH__ < 7
290 
291 static u16 imm_offset(u32 k, struct jit_ctx *ctx)
292 {
293 	unsigned i = 0, offset;
294 	u16 imm;
295 
296 	/* on the "fake" run we just count them (duplicates included) */
297 	if (ctx->target == NULL) {
298 		ctx->imm_count++;
299 		return 0;
300 	}
301 
302 	while ((i < ctx->imm_count) && ctx->imms[i]) {
303 		if (ctx->imms[i] == k)
304 			break;
305 		i++;
306 	}
307 
308 	if (ctx->imms[i] == 0)
309 		ctx->imms[i] = k;
310 
311 	/* constants go just after the epilogue */
312 	offset =  ctx->offsets[ctx->skf->len];
313 	offset += ctx->prologue_bytes;
314 	offset += ctx->epilogue_bytes;
315 	offset += i * 4;
316 
317 	ctx->target[offset / 4] = k;
318 
319 	/* PC in ARM mode == address of the instruction + 8 */
320 	imm = offset - (8 + ctx->idx * 4);
321 
322 	if (imm & ~0xfff) {
323 		/*
324 		 * literal pool is too far, signal it into flags. we
325 		 * can only detect it on the second pass unfortunately.
326 		 */
327 		ctx->flags |= FLAG_IMM_OVERFLOW;
328 		return 0;
329 	}
330 
331 	return imm;
332 }
333 
334 #endif /* __LINUX_ARM_ARCH__ */
335 
336 /*
337  * Move an immediate that's not an imm8m to a core register.
338  */
339 static inline void emit_mov_i_no8m(int rd, u32 val, struct jit_ctx *ctx)
340 {
341 #if __LINUX_ARM_ARCH__ < 7
342 	emit(ARM_LDR_I(rd, ARM_PC, imm_offset(val, ctx)), ctx);
343 #else
344 	emit(ARM_MOVW(rd, val & 0xffff), ctx);
345 	if (val > 0xffff)
346 		emit(ARM_MOVT(rd, val >> 16), ctx);
347 #endif
348 }
349 
350 static inline void emit_mov_i(int rd, u32 val, struct jit_ctx *ctx)
351 {
352 	int imm12 = imm8m(val);
353 
354 	if (imm12 >= 0)
355 		emit(ARM_MOV_I(rd, imm12), ctx);
356 	else
357 		emit_mov_i_no8m(rd, val, ctx);
358 }
359 
360 #if __LINUX_ARM_ARCH__ < 6
361 
362 static void emit_load_be32(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx)
363 {
364 	_emit(cond, ARM_LDRB_I(ARM_R3, r_addr, 1), ctx);
365 	_emit(cond, ARM_LDRB_I(ARM_R1, r_addr, 0), ctx);
366 	_emit(cond, ARM_LDRB_I(ARM_R2, r_addr, 3), ctx);
367 	_emit(cond, ARM_LSL_I(ARM_R3, ARM_R3, 16), ctx);
368 	_emit(cond, ARM_LDRB_I(ARM_R0, r_addr, 2), ctx);
369 	_emit(cond, ARM_ORR_S(ARM_R3, ARM_R3, ARM_R1, SRTYPE_LSL, 24), ctx);
370 	_emit(cond, ARM_ORR_R(ARM_R3, ARM_R3, ARM_R2), ctx);
371 	_emit(cond, ARM_ORR_S(r_res, ARM_R3, ARM_R0, SRTYPE_LSL, 8), ctx);
372 }
373 
374 static void emit_load_be16(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx)
375 {
376 	_emit(cond, ARM_LDRB_I(ARM_R1, r_addr, 0), ctx);
377 	_emit(cond, ARM_LDRB_I(ARM_R2, r_addr, 1), ctx);
378 	_emit(cond, ARM_ORR_S(r_res, ARM_R2, ARM_R1, SRTYPE_LSL, 8), ctx);
379 }
380 
381 static inline void emit_swap16(u8 r_dst, u8 r_src, struct jit_ctx *ctx)
382 {
383 	/* r_dst = (r_src << 8) | (r_src >> 8) */
384 	emit(ARM_LSL_I(ARM_R1, r_src, 8), ctx);
385 	emit(ARM_ORR_S(r_dst, ARM_R1, r_src, SRTYPE_LSR, 8), ctx);
386 
387 	/*
388 	 * we need to mask out the bits set in r_dst[23:16] due to
389 	 * the first shift instruction.
390 	 *
391 	 * note that 0x8ff is the encoded immediate 0x00ff0000.
392 	 */
393 	emit(ARM_BIC_I(r_dst, r_dst, 0x8ff), ctx);
394 }
395 
396 #else  /* ARMv6+ */
397 
398 static void emit_load_be32(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx)
399 {
400 	_emit(cond, ARM_LDR_I(r_res, r_addr, 0), ctx);
401 #ifdef __LITTLE_ENDIAN
402 	_emit(cond, ARM_REV(r_res, r_res), ctx);
403 #endif
404 }
405 
406 static void emit_load_be16(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx)
407 {
408 	_emit(cond, ARM_LDRH_I(r_res, r_addr, 0), ctx);
409 #ifdef __LITTLE_ENDIAN
410 	_emit(cond, ARM_REV16(r_res, r_res), ctx);
411 #endif
412 }
413 
414 static inline void emit_swap16(u8 r_dst __maybe_unused,
415 			       u8 r_src __maybe_unused,
416 			       struct jit_ctx *ctx __maybe_unused)
417 {
418 #ifdef __LITTLE_ENDIAN
419 	emit(ARM_REV16(r_dst, r_src), ctx);
420 #endif
421 }
422 
423 #endif /* __LINUX_ARM_ARCH__ < 6 */
424 
425 
426 /* Compute the immediate value for a PC-relative branch. */
427 static inline u32 b_imm(unsigned tgt, struct jit_ctx *ctx)
428 {
429 	u32 imm;
430 
431 	if (ctx->target == NULL)
432 		return 0;
433 	/*
434 	 * BPF allows only forward jumps and the offset of the target is
435 	 * still the one computed during the first pass.
436 	 */
437 	imm  = ctx->offsets[tgt] + ctx->prologue_bytes - (ctx->idx * 4 + 8);
438 
439 	return imm >> 2;
440 }
441 
442 #define OP_IMM3(op, r1, r2, imm_val, ctx)				\
443 	do {								\
444 		imm12 = imm8m(imm_val);					\
445 		if (imm12 < 0) {					\
446 			emit_mov_i_no8m(r_scratch, imm_val, ctx);	\
447 			emit(op ## _R((r1), (r2), r_scratch), ctx);	\
448 		} else {						\
449 			emit(op ## _I((r1), (r2), imm12), ctx);		\
450 		}							\
451 	} while (0)
452 
453 static inline void emit_err_ret(u8 cond, struct jit_ctx *ctx)
454 {
455 	if (ctx->ret0_fp_idx >= 0) {
456 		_emit(cond, ARM_B(b_imm(ctx->ret0_fp_idx, ctx)), ctx);
457 		/* NOP to keep the size constant between passes */
458 		emit(ARM_MOV_R(ARM_R0, ARM_R0), ctx);
459 	} else {
460 		_emit(cond, ARM_MOV_I(ARM_R0, 0), ctx);
461 		_emit(cond, ARM_B(b_imm(ctx->skf->len, ctx)), ctx);
462 	}
463 }
464 
465 static inline void emit_blx_r(u8 tgt_reg, struct jit_ctx *ctx)
466 {
467 #if __LINUX_ARM_ARCH__ < 5
468 	emit(ARM_MOV_R(ARM_LR, ARM_PC), ctx);
469 
470 	if (elf_hwcap & HWCAP_THUMB)
471 		emit(ARM_BX(tgt_reg), ctx);
472 	else
473 		emit(ARM_MOV_R(ARM_PC, tgt_reg), ctx);
474 #else
475 	emit(ARM_BLX_R(tgt_reg), ctx);
476 #endif
477 }
478 
479 static inline void emit_udivmod(u8 rd, u8 rm, u8 rn, struct jit_ctx *ctx,
480 				int bpf_op)
481 {
482 #if __LINUX_ARM_ARCH__ == 7
483 	if (elf_hwcap & HWCAP_IDIVA) {
484 		if (bpf_op == BPF_DIV)
485 			emit(ARM_UDIV(rd, rm, rn), ctx);
486 		else {
487 			emit(ARM_UDIV(ARM_R3, rm, rn), ctx);
488 			emit(ARM_MLS(rd, rn, ARM_R3, rm), ctx);
489 		}
490 		return;
491 	}
492 #endif
493 
494 	/*
495 	 * For BPF_ALU | BPF_DIV | BPF_K instructions, rm is ARM_R4
496 	 * (r_A) and rn is ARM_R0 (r_scratch) so load rn first into
497 	 * ARM_R1 to avoid accidentally overwriting ARM_R0 with rm
498 	 * before using it as a source for ARM_R1.
499 	 *
500 	 * For BPF_ALU | BPF_DIV | BPF_X rm is ARM_R4 (r_A) and rn is
501 	 * ARM_R5 (r_X) so there is no particular register overlap
502 	 * issues.
503 	 */
504 	if (rn != ARM_R1)
505 		emit(ARM_MOV_R(ARM_R1, rn), ctx);
506 	if (rm != ARM_R0)
507 		emit(ARM_MOV_R(ARM_R0, rm), ctx);
508 
509 	ctx->seen |= SEEN_CALL;
510 	emit_mov_i(ARM_R3, bpf_op == BPF_DIV ? (u32)jit_udiv : (u32)jit_mod,
511 		   ctx);
512 	emit_blx_r(ARM_R3, ctx);
513 
514 	if (rd != ARM_R0)
515 		emit(ARM_MOV_R(rd, ARM_R0), ctx);
516 }
517 
518 static inline void update_on_xread(struct jit_ctx *ctx)
519 {
520 	if (!(ctx->seen & SEEN_X))
521 		ctx->flags |= FLAG_NEED_X_RESET;
522 
523 	ctx->seen |= SEEN_X;
524 }
525 
526 static int build_body(struct jit_ctx *ctx)
527 {
528 	void *load_func[] = {jit_get_skb_b, jit_get_skb_h, jit_get_skb_w};
529 	const struct bpf_prog *prog = ctx->skf;
530 	const struct sock_filter *inst;
531 	unsigned i, load_order, off, condt;
532 	int imm12;
533 	u32 k;
534 
535 	for (i = 0; i < prog->len; i++) {
536 		u16 code;
537 
538 		inst = &(prog->insns[i]);
539 		/* K as an immediate value operand */
540 		k = inst->k;
541 		code = bpf_anc_helper(inst);
542 
543 		/* compute offsets only in the fake pass */
544 		if (ctx->target == NULL)
545 			ctx->offsets[i] = ctx->idx * 4;
546 
547 		switch (code) {
548 		case BPF_LD | BPF_IMM:
549 			emit_mov_i(r_A, k, ctx);
550 			break;
551 		case BPF_LD | BPF_W | BPF_LEN:
552 			ctx->seen |= SEEN_SKB;
553 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
554 			emit(ARM_LDR_I(r_A, r_skb,
555 				       offsetof(struct sk_buff, len)), ctx);
556 			break;
557 		case BPF_LD | BPF_MEM:
558 			/* A = scratch[k] */
559 			ctx->seen |= SEEN_MEM_WORD(k);
560 			emit(ARM_LDR_I(r_A, ARM_SP, SCRATCH_OFF(k)), ctx);
561 			break;
562 		case BPF_LD | BPF_W | BPF_ABS:
563 			load_order = 2;
564 			goto load;
565 		case BPF_LD | BPF_H | BPF_ABS:
566 			load_order = 1;
567 			goto load;
568 		case BPF_LD | BPF_B | BPF_ABS:
569 			load_order = 0;
570 load:
571 			emit_mov_i(r_off, k, ctx);
572 load_common:
573 			ctx->seen |= SEEN_DATA | SEEN_CALL;
574 
575 			if (load_order > 0) {
576 				emit(ARM_SUB_I(r_scratch, r_skb_hl,
577 					       1 << load_order), ctx);
578 				emit(ARM_CMP_R(r_scratch, r_off), ctx);
579 				condt = ARM_COND_GE;
580 			} else {
581 				emit(ARM_CMP_R(r_skb_hl, r_off), ctx);
582 				condt = ARM_COND_HI;
583 			}
584 
585 			/*
586 			 * test for negative offset, only if we are
587 			 * currently scheduled to take the fast
588 			 * path. this will update the flags so that
589 			 * the slowpath instruction are ignored if the
590 			 * offset is negative.
591 			 *
592 			 * for loard_order == 0 the HI condition will
593 			 * make loads at offset 0 take the slow path too.
594 			 */
595 			_emit(condt, ARM_CMP_I(r_off, 0), ctx);
596 
597 			_emit(condt, ARM_ADD_R(r_scratch, r_off, r_skb_data),
598 			      ctx);
599 
600 			if (load_order == 0)
601 				_emit(condt, ARM_LDRB_I(r_A, r_scratch, 0),
602 				      ctx);
603 			else if (load_order == 1)
604 				emit_load_be16(condt, r_A, r_scratch, ctx);
605 			else if (load_order == 2)
606 				emit_load_be32(condt, r_A, r_scratch, ctx);
607 
608 			_emit(condt, ARM_B(b_imm(i + 1, ctx)), ctx);
609 
610 			/* the slowpath */
611 			emit_mov_i(ARM_R3, (u32)load_func[load_order], ctx);
612 			emit(ARM_MOV_R(ARM_R0, r_skb), ctx);
613 			/* the offset is already in R1 */
614 			emit_blx_r(ARM_R3, ctx);
615 			/* check the result of skb_copy_bits */
616 			emit(ARM_CMP_I(ARM_R1, 0), ctx);
617 			emit_err_ret(ARM_COND_NE, ctx);
618 			emit(ARM_MOV_R(r_A, ARM_R0), ctx);
619 			break;
620 		case BPF_LD | BPF_W | BPF_IND:
621 			load_order = 2;
622 			goto load_ind;
623 		case BPF_LD | BPF_H | BPF_IND:
624 			load_order = 1;
625 			goto load_ind;
626 		case BPF_LD | BPF_B | BPF_IND:
627 			load_order = 0;
628 load_ind:
629 			update_on_xread(ctx);
630 			OP_IMM3(ARM_ADD, r_off, r_X, k, ctx);
631 			goto load_common;
632 		case BPF_LDX | BPF_IMM:
633 			ctx->seen |= SEEN_X;
634 			emit_mov_i(r_X, k, ctx);
635 			break;
636 		case BPF_LDX | BPF_W | BPF_LEN:
637 			ctx->seen |= SEEN_X | SEEN_SKB;
638 			emit(ARM_LDR_I(r_X, r_skb,
639 				       offsetof(struct sk_buff, len)), ctx);
640 			break;
641 		case BPF_LDX | BPF_MEM:
642 			ctx->seen |= SEEN_X | SEEN_MEM_WORD(k);
643 			emit(ARM_LDR_I(r_X, ARM_SP, SCRATCH_OFF(k)), ctx);
644 			break;
645 		case BPF_LDX | BPF_B | BPF_MSH:
646 			/* x = ((*(frame + k)) & 0xf) << 2; */
647 			ctx->seen |= SEEN_X | SEEN_DATA | SEEN_CALL;
648 			/* the interpreter should deal with the negative K */
649 			if ((int)k < 0)
650 				return -1;
651 			/* offset in r1: we might have to take the slow path */
652 			emit_mov_i(r_off, k, ctx);
653 			emit(ARM_CMP_R(r_skb_hl, r_off), ctx);
654 
655 			/* load in r0: common with the slowpath */
656 			_emit(ARM_COND_HI, ARM_LDRB_R(ARM_R0, r_skb_data,
657 						      ARM_R1), ctx);
658 			/*
659 			 * emit_mov_i() might generate one or two instructions,
660 			 * the same holds for emit_blx_r()
661 			 */
662 			_emit(ARM_COND_HI, ARM_B(b_imm(i + 1, ctx) - 2), ctx);
663 
664 			emit(ARM_MOV_R(ARM_R0, r_skb), ctx);
665 			/* r_off is r1 */
666 			emit_mov_i(ARM_R3, (u32)jit_get_skb_b, ctx);
667 			emit_blx_r(ARM_R3, ctx);
668 			/* check the return value of skb_copy_bits */
669 			emit(ARM_CMP_I(ARM_R1, 0), ctx);
670 			emit_err_ret(ARM_COND_NE, ctx);
671 
672 			emit(ARM_AND_I(r_X, ARM_R0, 0x00f), ctx);
673 			emit(ARM_LSL_I(r_X, r_X, 2), ctx);
674 			break;
675 		case BPF_ST:
676 			ctx->seen |= SEEN_MEM_WORD(k);
677 			emit(ARM_STR_I(r_A, ARM_SP, SCRATCH_OFF(k)), ctx);
678 			break;
679 		case BPF_STX:
680 			update_on_xread(ctx);
681 			ctx->seen |= SEEN_MEM_WORD(k);
682 			emit(ARM_STR_I(r_X, ARM_SP, SCRATCH_OFF(k)), ctx);
683 			break;
684 		case BPF_ALU | BPF_ADD | BPF_K:
685 			/* A += K */
686 			OP_IMM3(ARM_ADD, r_A, r_A, k, ctx);
687 			break;
688 		case BPF_ALU | BPF_ADD | BPF_X:
689 			update_on_xread(ctx);
690 			emit(ARM_ADD_R(r_A, r_A, r_X), ctx);
691 			break;
692 		case BPF_ALU | BPF_SUB | BPF_K:
693 			/* A -= K */
694 			OP_IMM3(ARM_SUB, r_A, r_A, k, ctx);
695 			break;
696 		case BPF_ALU | BPF_SUB | BPF_X:
697 			update_on_xread(ctx);
698 			emit(ARM_SUB_R(r_A, r_A, r_X), ctx);
699 			break;
700 		case BPF_ALU | BPF_MUL | BPF_K:
701 			/* A *= K */
702 			emit_mov_i(r_scratch, k, ctx);
703 			emit(ARM_MUL(r_A, r_A, r_scratch), ctx);
704 			break;
705 		case BPF_ALU | BPF_MUL | BPF_X:
706 			update_on_xread(ctx);
707 			emit(ARM_MUL(r_A, r_A, r_X), ctx);
708 			break;
709 		case BPF_ALU | BPF_DIV | BPF_K:
710 			if (k == 1)
711 				break;
712 			emit_mov_i(r_scratch, k, ctx);
713 			emit_udivmod(r_A, r_A, r_scratch, ctx, BPF_DIV);
714 			break;
715 		case BPF_ALU | BPF_DIV | BPF_X:
716 			update_on_xread(ctx);
717 			emit(ARM_CMP_I(r_X, 0), ctx);
718 			emit_err_ret(ARM_COND_EQ, ctx);
719 			emit_udivmod(r_A, r_A, r_X, ctx, BPF_DIV);
720 			break;
721 		case BPF_ALU | BPF_MOD | BPF_K:
722 			if (k == 1) {
723 				emit_mov_i(r_A, 0, ctx);
724 				break;
725 			}
726 			emit_mov_i(r_scratch, k, ctx);
727 			emit_udivmod(r_A, r_A, r_scratch, ctx, BPF_MOD);
728 			break;
729 		case BPF_ALU | BPF_MOD | BPF_X:
730 			update_on_xread(ctx);
731 			emit(ARM_CMP_I(r_X, 0), ctx);
732 			emit_err_ret(ARM_COND_EQ, ctx);
733 			emit_udivmod(r_A, r_A, r_X, ctx, BPF_MOD);
734 			break;
735 		case BPF_ALU | BPF_OR | BPF_K:
736 			/* A |= K */
737 			OP_IMM3(ARM_ORR, r_A, r_A, k, ctx);
738 			break;
739 		case BPF_ALU | BPF_OR | BPF_X:
740 			update_on_xread(ctx);
741 			emit(ARM_ORR_R(r_A, r_A, r_X), ctx);
742 			break;
743 		case BPF_ALU | BPF_XOR | BPF_K:
744 			/* A ^= K; */
745 			OP_IMM3(ARM_EOR, r_A, r_A, k, ctx);
746 			break;
747 		case BPF_ANC | SKF_AD_ALU_XOR_X:
748 		case BPF_ALU | BPF_XOR | BPF_X:
749 			/* A ^= X */
750 			update_on_xread(ctx);
751 			emit(ARM_EOR_R(r_A, r_A, r_X), ctx);
752 			break;
753 		case BPF_ALU | BPF_AND | BPF_K:
754 			/* A &= K */
755 			OP_IMM3(ARM_AND, r_A, r_A, k, ctx);
756 			break;
757 		case BPF_ALU | BPF_AND | BPF_X:
758 			update_on_xread(ctx);
759 			emit(ARM_AND_R(r_A, r_A, r_X), ctx);
760 			break;
761 		case BPF_ALU | BPF_LSH | BPF_K:
762 			if (unlikely(k > 31))
763 				return -1;
764 			emit(ARM_LSL_I(r_A, r_A, k), ctx);
765 			break;
766 		case BPF_ALU | BPF_LSH | BPF_X:
767 			update_on_xread(ctx);
768 			emit(ARM_LSL_R(r_A, r_A, r_X), ctx);
769 			break;
770 		case BPF_ALU | BPF_RSH | BPF_K:
771 			if (unlikely(k > 31))
772 				return -1;
773 			emit(ARM_LSR_I(r_A, r_A, k), ctx);
774 			break;
775 		case BPF_ALU | BPF_RSH | BPF_X:
776 			update_on_xread(ctx);
777 			emit(ARM_LSR_R(r_A, r_A, r_X), ctx);
778 			break;
779 		case BPF_ALU | BPF_NEG:
780 			/* A = -A */
781 			emit(ARM_RSB_I(r_A, r_A, 0), ctx);
782 			break;
783 		case BPF_JMP | BPF_JA:
784 			/* pc += K */
785 			emit(ARM_B(b_imm(i + k + 1, ctx)), ctx);
786 			break;
787 		case BPF_JMP | BPF_JEQ | BPF_K:
788 			/* pc += (A == K) ? pc->jt : pc->jf */
789 			condt  = ARM_COND_EQ;
790 			goto cmp_imm;
791 		case BPF_JMP | BPF_JGT | BPF_K:
792 			/* pc += (A > K) ? pc->jt : pc->jf */
793 			condt  = ARM_COND_HI;
794 			goto cmp_imm;
795 		case BPF_JMP | BPF_JGE | BPF_K:
796 			/* pc += (A >= K) ? pc->jt : pc->jf */
797 			condt  = ARM_COND_HS;
798 cmp_imm:
799 			imm12 = imm8m(k);
800 			if (imm12 < 0) {
801 				emit_mov_i_no8m(r_scratch, k, ctx);
802 				emit(ARM_CMP_R(r_A, r_scratch), ctx);
803 			} else {
804 				emit(ARM_CMP_I(r_A, imm12), ctx);
805 			}
806 cond_jump:
807 			if (inst->jt)
808 				_emit(condt, ARM_B(b_imm(i + inst->jt + 1,
809 						   ctx)), ctx);
810 			if (inst->jf)
811 				_emit(condt ^ 1, ARM_B(b_imm(i + inst->jf + 1,
812 							     ctx)), ctx);
813 			break;
814 		case BPF_JMP | BPF_JEQ | BPF_X:
815 			/* pc += (A == X) ? pc->jt : pc->jf */
816 			condt   = ARM_COND_EQ;
817 			goto cmp_x;
818 		case BPF_JMP | BPF_JGT | BPF_X:
819 			/* pc += (A > X) ? pc->jt : pc->jf */
820 			condt   = ARM_COND_HI;
821 			goto cmp_x;
822 		case BPF_JMP | BPF_JGE | BPF_X:
823 			/* pc += (A >= X) ? pc->jt : pc->jf */
824 			condt   = ARM_COND_CS;
825 cmp_x:
826 			update_on_xread(ctx);
827 			emit(ARM_CMP_R(r_A, r_X), ctx);
828 			goto cond_jump;
829 		case BPF_JMP | BPF_JSET | BPF_K:
830 			/* pc += (A & K) ? pc->jt : pc->jf */
831 			condt  = ARM_COND_NE;
832 			/* not set iff all zeroes iff Z==1 iff EQ */
833 
834 			imm12 = imm8m(k);
835 			if (imm12 < 0) {
836 				emit_mov_i_no8m(r_scratch, k, ctx);
837 				emit(ARM_TST_R(r_A, r_scratch), ctx);
838 			} else {
839 				emit(ARM_TST_I(r_A, imm12), ctx);
840 			}
841 			goto cond_jump;
842 		case BPF_JMP | BPF_JSET | BPF_X:
843 			/* pc += (A & X) ? pc->jt : pc->jf */
844 			update_on_xread(ctx);
845 			condt  = ARM_COND_NE;
846 			emit(ARM_TST_R(r_A, r_X), ctx);
847 			goto cond_jump;
848 		case BPF_RET | BPF_A:
849 			emit(ARM_MOV_R(ARM_R0, r_A), ctx);
850 			goto b_epilogue;
851 		case BPF_RET | BPF_K:
852 			if ((k == 0) && (ctx->ret0_fp_idx < 0))
853 				ctx->ret0_fp_idx = i;
854 			emit_mov_i(ARM_R0, k, ctx);
855 b_epilogue:
856 			if (i != ctx->skf->len - 1)
857 				emit(ARM_B(b_imm(prog->len, ctx)), ctx);
858 			break;
859 		case BPF_MISC | BPF_TAX:
860 			/* X = A */
861 			ctx->seen |= SEEN_X;
862 			emit(ARM_MOV_R(r_X, r_A), ctx);
863 			break;
864 		case BPF_MISC | BPF_TXA:
865 			/* A = X */
866 			update_on_xread(ctx);
867 			emit(ARM_MOV_R(r_A, r_X), ctx);
868 			break;
869 		case BPF_ANC | SKF_AD_PROTOCOL:
870 			/* A = ntohs(skb->protocol) */
871 			ctx->seen |= SEEN_SKB;
872 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
873 						  protocol) != 2);
874 			off = offsetof(struct sk_buff, protocol);
875 			emit(ARM_LDRH_I(r_scratch, r_skb, off), ctx);
876 			emit_swap16(r_A, r_scratch, ctx);
877 			break;
878 		case BPF_ANC | SKF_AD_CPU:
879 			/* r_scratch = current_thread_info() */
880 			OP_IMM3(ARM_BIC, r_scratch, ARM_SP, THREAD_SIZE - 1, ctx);
881 			/* A = current_thread_info()->cpu */
882 			BUILD_BUG_ON(FIELD_SIZEOF(struct thread_info, cpu) != 4);
883 			off = offsetof(struct thread_info, cpu);
884 			emit(ARM_LDR_I(r_A, r_scratch, off), ctx);
885 			break;
886 		case BPF_ANC | SKF_AD_IFINDEX:
887 		case BPF_ANC | SKF_AD_HATYPE:
888 			/* A = skb->dev->ifindex */
889 			/* A = skb->dev->type */
890 			ctx->seen |= SEEN_SKB;
891 			off = offsetof(struct sk_buff, dev);
892 			emit(ARM_LDR_I(r_scratch, r_skb, off), ctx);
893 
894 			emit(ARM_CMP_I(r_scratch, 0), ctx);
895 			emit_err_ret(ARM_COND_EQ, ctx);
896 
897 			BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
898 						  ifindex) != 4);
899 			BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
900 						  type) != 2);
901 
902 			if (code == (BPF_ANC | SKF_AD_IFINDEX)) {
903 				off = offsetof(struct net_device, ifindex);
904 				emit(ARM_LDR_I(r_A, r_scratch, off), ctx);
905 			} else {
906 				/*
907 				 * offset of field "type" in "struct
908 				 * net_device" is above what can be
909 				 * used in the ldrh rd, [rn, #imm]
910 				 * instruction, so load the offset in
911 				 * a register and use ldrh rd, [rn, rm]
912 				 */
913 				off = offsetof(struct net_device, type);
914 				emit_mov_i(ARM_R3, off, ctx);
915 				emit(ARM_LDRH_R(r_A, r_scratch, ARM_R3), ctx);
916 			}
917 			break;
918 		case BPF_ANC | SKF_AD_MARK:
919 			ctx->seen |= SEEN_SKB;
920 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
921 			off = offsetof(struct sk_buff, mark);
922 			emit(ARM_LDR_I(r_A, r_skb, off), ctx);
923 			break;
924 		case BPF_ANC | SKF_AD_RXHASH:
925 			ctx->seen |= SEEN_SKB;
926 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
927 			off = offsetof(struct sk_buff, hash);
928 			emit(ARM_LDR_I(r_A, r_skb, off), ctx);
929 			break;
930 		case BPF_ANC | SKF_AD_VLAN_TAG:
931 		case BPF_ANC | SKF_AD_VLAN_TAG_PRESENT:
932 			ctx->seen |= SEEN_SKB;
933 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2);
934 			off = offsetof(struct sk_buff, vlan_tci);
935 			emit(ARM_LDRH_I(r_A, r_skb, off), ctx);
936 			if (code == (BPF_ANC | SKF_AD_VLAN_TAG))
937 				OP_IMM3(ARM_AND, r_A, r_A, ~VLAN_TAG_PRESENT, ctx);
938 			else {
939 				OP_IMM3(ARM_LSR, r_A, r_A, 12, ctx);
940 				OP_IMM3(ARM_AND, r_A, r_A, 0x1, ctx);
941 			}
942 			break;
943 		case BPF_ANC | SKF_AD_PKTTYPE:
944 			ctx->seen |= SEEN_SKB;
945 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
946 						  __pkt_type_offset[0]) != 1);
947 			off = PKT_TYPE_OFFSET();
948 			emit(ARM_LDRB_I(r_A, r_skb, off), ctx);
949 			emit(ARM_AND_I(r_A, r_A, PKT_TYPE_MAX), ctx);
950 #ifdef __BIG_ENDIAN_BITFIELD
951 			emit(ARM_LSR_I(r_A, r_A, 5), ctx);
952 #endif
953 			break;
954 		case BPF_ANC | SKF_AD_QUEUE:
955 			ctx->seen |= SEEN_SKB;
956 			BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
957 						  queue_mapping) != 2);
958 			BUILD_BUG_ON(offsetof(struct sk_buff,
959 					      queue_mapping) > 0xff);
960 			off = offsetof(struct sk_buff, queue_mapping);
961 			emit(ARM_LDRH_I(r_A, r_skb, off), ctx);
962 			break;
963 		case BPF_ANC | SKF_AD_PAY_OFFSET:
964 			ctx->seen |= SEEN_SKB | SEEN_CALL;
965 
966 			emit(ARM_MOV_R(ARM_R0, r_skb), ctx);
967 			emit_mov_i(ARM_R3, (unsigned int)skb_get_poff, ctx);
968 			emit_blx_r(ARM_R3, ctx);
969 			emit(ARM_MOV_R(r_A, ARM_R0), ctx);
970 			break;
971 		case BPF_LDX | BPF_W | BPF_ABS:
972 			/*
973 			 * load a 32bit word from struct seccomp_data.
974 			 * seccomp_check_filter() will already have checked
975 			 * that k is 32bit aligned and lies within the
976 			 * struct seccomp_data.
977 			 */
978 			ctx->seen |= SEEN_SKB;
979 			emit(ARM_LDR_I(r_A, r_skb, k), ctx);
980 			break;
981 		default:
982 			return -1;
983 		}
984 
985 		if (ctx->flags & FLAG_IMM_OVERFLOW)
986 			/*
987 			 * this instruction generated an overflow when
988 			 * trying to access the literal pool, so
989 			 * delegate this filter to the kernel interpreter.
990 			 */
991 			return -1;
992 	}
993 
994 	/* compute offsets only during the first pass */
995 	if (ctx->target == NULL)
996 		ctx->offsets[i] = ctx->idx * 4;
997 
998 	return 0;
999 }
1000 
1001 
1002 void bpf_jit_compile(struct bpf_prog *fp)
1003 {
1004 	struct bpf_binary_header *header;
1005 	struct jit_ctx ctx;
1006 	unsigned tmp_idx;
1007 	unsigned alloc_size;
1008 	u8 *target_ptr;
1009 
1010 	if (!bpf_jit_enable)
1011 		return;
1012 
1013 	memset(&ctx, 0, sizeof(ctx));
1014 	ctx.skf		= fp;
1015 	ctx.ret0_fp_idx = -1;
1016 
1017 	ctx.offsets = kzalloc(4 * (ctx.skf->len + 1), GFP_KERNEL);
1018 	if (ctx.offsets == NULL)
1019 		return;
1020 
1021 	/* fake pass to fill in the ctx->seen */
1022 	if (unlikely(build_body(&ctx)))
1023 		goto out;
1024 
1025 	tmp_idx = ctx.idx;
1026 	build_prologue(&ctx);
1027 	ctx.prologue_bytes = (ctx.idx - tmp_idx) * 4;
1028 
1029 #if __LINUX_ARM_ARCH__ < 7
1030 	tmp_idx = ctx.idx;
1031 	build_epilogue(&ctx);
1032 	ctx.epilogue_bytes = (ctx.idx - tmp_idx) * 4;
1033 
1034 	ctx.idx += ctx.imm_count;
1035 	if (ctx.imm_count) {
1036 		ctx.imms = kzalloc(4 * ctx.imm_count, GFP_KERNEL);
1037 		if (ctx.imms == NULL)
1038 			goto out;
1039 	}
1040 #else
1041 	/* there's nothing after the epilogue on ARMv7 */
1042 	build_epilogue(&ctx);
1043 #endif
1044 	alloc_size = 4 * ctx.idx;
1045 	header = bpf_jit_binary_alloc(alloc_size, &target_ptr,
1046 				      4, jit_fill_hole);
1047 	if (header == NULL)
1048 		goto out;
1049 
1050 	ctx.target = (u32 *) target_ptr;
1051 	ctx.idx = 0;
1052 
1053 	build_prologue(&ctx);
1054 	if (build_body(&ctx) < 0) {
1055 #if __LINUX_ARM_ARCH__ < 7
1056 		if (ctx.imm_count)
1057 			kfree(ctx.imms);
1058 #endif
1059 		bpf_jit_binary_free(header);
1060 		goto out;
1061 	}
1062 	build_epilogue(&ctx);
1063 
1064 	flush_icache_range((u32)header, (u32)(ctx.target + ctx.idx));
1065 
1066 #if __LINUX_ARM_ARCH__ < 7
1067 	if (ctx.imm_count)
1068 		kfree(ctx.imms);
1069 #endif
1070 
1071 	if (bpf_jit_enable > 1)
1072 		/* there are 2 passes here */
1073 		bpf_jit_dump(fp->len, alloc_size, 2, ctx.target);
1074 
1075 	set_memory_ro((unsigned long)header, header->pages);
1076 	fp->bpf_func = (void *)ctx.target;
1077 	fp->jited = 1;
1078 out:
1079 	kfree(ctx.offsets);
1080 	return;
1081 }
1082 
1083 void bpf_jit_free(struct bpf_prog *fp)
1084 {
1085 	unsigned long addr = (unsigned long)fp->bpf_func & PAGE_MASK;
1086 	struct bpf_binary_header *header = (void *)addr;
1087 
1088 	if (!fp->jited)
1089 		goto free_filter;
1090 
1091 	set_memory_rw(addr, header->pages);
1092 	bpf_jit_binary_free(header);
1093 
1094 free_filter:
1095 	bpf_prog_unlock_free(fp);
1096 }
1097