Commit | Line | Data |
---|---|---|
ddecdfce MG |
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/moduleloader.h> | |
16 | #include <linux/netdevice.h> | |
17 | #include <linux/string.h> | |
18 | #include <linux/slab.h> | |
bf0098f2 | 19 | #include <linux/if_vlan.h> |
ddecdfce MG |
20 | #include <asm/cacheflush.h> |
21 | #include <asm/hwcap.h> | |
22 | ||
23 | #include "bpf_jit_32.h" | |
24 | ||
25 | /* | |
26 | * ABI: | |
27 | * | |
28 | * r0 scratch register | |
29 | * r4 BPF register A | |
30 | * r5 BPF register X | |
31 | * r6 pointer to the skb | |
32 | * r7 skb->data | |
33 | * r8 skb_headlen(skb) | |
34 | */ | |
35 | ||
36 | #define r_scratch ARM_R0 | |
37 | /* r1-r3 are (also) used for the unaligned loads on the non-ARMv7 slowpath */ | |
38 | #define r_off ARM_R1 | |
39 | #define r_A ARM_R4 | |
40 | #define r_X ARM_R5 | |
41 | #define r_skb ARM_R6 | |
42 | #define r_skb_data ARM_R7 | |
43 | #define r_skb_hl ARM_R8 | |
44 | ||
45 | #define SCRATCH_SP_OFFSET 0 | |
fe15f3f1 | 46 | #define SCRATCH_OFF(k) (SCRATCH_SP_OFFSET + 4 * (k)) |
ddecdfce MG |
47 | |
48 | #define SEEN_MEM ((1 << BPF_MEMWORDS) - 1) | |
49 | #define SEEN_MEM_WORD(k) (1 << (k)) | |
50 | #define SEEN_X (1 << BPF_MEMWORDS) | |
51 | #define SEEN_CALL (1 << (BPF_MEMWORDS + 1)) | |
52 | #define SEEN_SKB (1 << (BPF_MEMWORDS + 2)) | |
53 | #define SEEN_DATA (1 << (BPF_MEMWORDS + 3)) | |
54 | ||
55 | #define FLAG_NEED_X_RESET (1 << 0) | |
56 | ||
57 | struct jit_ctx { | |
58 | const struct sk_filter *skf; | |
59 | unsigned idx; | |
60 | unsigned prologue_bytes; | |
61 | int ret0_fp_idx; | |
62 | u32 seen; | |
63 | u32 flags; | |
64 | u32 *offsets; | |
65 | u32 *target; | |
66 | #if __LINUX_ARM_ARCH__ < 7 | |
67 | u16 epilogue_bytes; | |
68 | u16 imm_count; | |
69 | u32 *imms; | |
70 | #endif | |
71 | }; | |
72 | ||
73 | int bpf_jit_enable __read_mostly; | |
74 | ||
75 | static u64 jit_get_skb_b(struct sk_buff *skb, unsigned offset) | |
76 | { | |
77 | u8 ret; | |
78 | int err; | |
79 | ||
80 | err = skb_copy_bits(skb, offset, &ret, 1); | |
81 | ||
82 | return (u64)err << 32 | ret; | |
83 | } | |
84 | ||
85 | static u64 jit_get_skb_h(struct sk_buff *skb, unsigned offset) | |
86 | { | |
87 | u16 ret; | |
88 | int err; | |
89 | ||
90 | err = skb_copy_bits(skb, offset, &ret, 2); | |
91 | ||
92 | return (u64)err << 32 | ntohs(ret); | |
93 | } | |
94 | ||
95 | static u64 jit_get_skb_w(struct sk_buff *skb, unsigned offset) | |
96 | { | |
97 | u32 ret; | |
98 | int err; | |
99 | ||
100 | err = skb_copy_bits(skb, offset, &ret, 4); | |
101 | ||
102 | return (u64)err << 32 | ntohl(ret); | |
103 | } | |
104 | ||
105 | /* | |
106 | * Wrapper that handles both OABI and EABI and assures Thumb2 interworking | |
107 | * (where the assembly routines like __aeabi_uidiv could cause problems). | |
108 | */ | |
109 | static u32 jit_udiv(u32 dividend, u32 divisor) | |
110 | { | |
111 | return dividend / divisor; | |
112 | } | |
113 | ||
114 | static inline void _emit(int cond, u32 inst, struct jit_ctx *ctx) | |
115 | { | |
116 | if (ctx->target != NULL) | |
117 | ctx->target[ctx->idx] = inst | (cond << 28); | |
118 | ||
119 | ctx->idx++; | |
120 | } | |
121 | ||
122 | /* | |
123 | * Emit an instruction that will be executed unconditionally. | |
124 | */ | |
125 | static inline void emit(u32 inst, struct jit_ctx *ctx) | |
126 | { | |
127 | _emit(ARM_COND_AL, inst, ctx); | |
128 | } | |
129 | ||
130 | static u16 saved_regs(struct jit_ctx *ctx) | |
131 | { | |
132 | u16 ret = 0; | |
133 | ||
134 | if ((ctx->skf->len > 1) || | |
135 | (ctx->skf->insns[0].code == BPF_S_RET_A)) | |
136 | ret |= 1 << r_A; | |
137 | ||
138 | #ifdef CONFIG_FRAME_POINTER | |
139 | ret |= (1 << ARM_FP) | (1 << ARM_IP) | (1 << ARM_LR) | (1 << ARM_PC); | |
140 | #else | |
141 | if (ctx->seen & SEEN_CALL) | |
142 | ret |= 1 << ARM_LR; | |
143 | #endif | |
144 | if (ctx->seen & (SEEN_DATA | SEEN_SKB)) | |
145 | ret |= 1 << r_skb; | |
146 | if (ctx->seen & SEEN_DATA) | |
147 | ret |= (1 << r_skb_data) | (1 << r_skb_hl); | |
148 | if (ctx->seen & SEEN_X) | |
149 | ret |= 1 << r_X; | |
150 | ||
151 | return ret; | |
152 | } | |
153 | ||
154 | static inline int mem_words_used(struct jit_ctx *ctx) | |
155 | { | |
156 | /* yes, we do waste some stack space IF there are "holes" in the set" */ | |
157 | return fls(ctx->seen & SEEN_MEM); | |
158 | } | |
159 | ||
160 | static inline bool is_load_to_a(u16 inst) | |
161 | { | |
162 | switch (inst) { | |
163 | case BPF_S_LD_W_LEN: | |
164 | case BPF_S_LD_W_ABS: | |
165 | case BPF_S_LD_H_ABS: | |
166 | case BPF_S_LD_B_ABS: | |
167 | case BPF_S_ANC_CPU: | |
168 | case BPF_S_ANC_IFINDEX: | |
169 | case BPF_S_ANC_MARK: | |
170 | case BPF_S_ANC_PROTOCOL: | |
171 | case BPF_S_ANC_RXHASH: | |
bf0098f2 DB |
172 | case BPF_S_ANC_VLAN_TAG: |
173 | case BPF_S_ANC_VLAN_TAG_PRESENT: | |
ddecdfce MG |
174 | case BPF_S_ANC_QUEUE: |
175 | return true; | |
176 | default: | |
177 | return false; | |
178 | } | |
179 | } | |
180 | ||
181 | static void build_prologue(struct jit_ctx *ctx) | |
182 | { | |
183 | u16 reg_set = saved_regs(ctx); | |
184 | u16 first_inst = ctx->skf->insns[0].code; | |
185 | u16 off; | |
186 | ||
187 | #ifdef CONFIG_FRAME_POINTER | |
188 | emit(ARM_MOV_R(ARM_IP, ARM_SP), ctx); | |
189 | emit(ARM_PUSH(reg_set), ctx); | |
190 | emit(ARM_SUB_I(ARM_FP, ARM_IP, 4), ctx); | |
191 | #else | |
192 | if (reg_set) | |
193 | emit(ARM_PUSH(reg_set), ctx); | |
194 | #endif | |
195 | ||
196 | if (ctx->seen & (SEEN_DATA | SEEN_SKB)) | |
197 | emit(ARM_MOV_R(r_skb, ARM_R0), ctx); | |
198 | ||
199 | if (ctx->seen & SEEN_DATA) { | |
200 | off = offsetof(struct sk_buff, data); | |
201 | emit(ARM_LDR_I(r_skb_data, r_skb, off), ctx); | |
202 | /* headlen = len - data_len */ | |
203 | off = offsetof(struct sk_buff, len); | |
204 | emit(ARM_LDR_I(r_skb_hl, r_skb, off), ctx); | |
205 | off = offsetof(struct sk_buff, data_len); | |
206 | emit(ARM_LDR_I(r_scratch, r_skb, off), ctx); | |
207 | emit(ARM_SUB_R(r_skb_hl, r_skb_hl, r_scratch), ctx); | |
208 | } | |
209 | ||
210 | if (ctx->flags & FLAG_NEED_X_RESET) | |
211 | emit(ARM_MOV_I(r_X, 0), ctx); | |
212 | ||
213 | /* do not leak kernel data to userspace */ | |
214 | if ((first_inst != BPF_S_RET_K) && !(is_load_to_a(first_inst))) | |
215 | emit(ARM_MOV_I(r_A, 0), ctx); | |
216 | ||
217 | /* stack space for the BPF_MEM words */ | |
218 | if (ctx->seen & SEEN_MEM) | |
219 | emit(ARM_SUB_I(ARM_SP, ARM_SP, mem_words_used(ctx) * 4), ctx); | |
220 | } | |
221 | ||
222 | static void build_epilogue(struct jit_ctx *ctx) | |
223 | { | |
224 | u16 reg_set = saved_regs(ctx); | |
225 | ||
226 | if (ctx->seen & SEEN_MEM) | |
227 | emit(ARM_ADD_I(ARM_SP, ARM_SP, mem_words_used(ctx) * 4), ctx); | |
228 | ||
229 | reg_set &= ~(1 << ARM_LR); | |
230 | ||
231 | #ifdef CONFIG_FRAME_POINTER | |
232 | /* the first instruction of the prologue was: mov ip, sp */ | |
233 | reg_set &= ~(1 << ARM_IP); | |
234 | reg_set |= (1 << ARM_SP); | |
235 | emit(ARM_LDM(ARM_SP, reg_set), ctx); | |
236 | #else | |
237 | if (reg_set) { | |
238 | if (ctx->seen & SEEN_CALL) | |
239 | reg_set |= 1 << ARM_PC; | |
240 | emit(ARM_POP(reg_set), ctx); | |
241 | } | |
242 | ||
243 | if (!(ctx->seen & SEEN_CALL)) | |
244 | emit(ARM_BX(ARM_LR), ctx); | |
245 | #endif | |
246 | } | |
247 | ||
248 | static int16_t imm8m(u32 x) | |
249 | { | |
250 | u32 rot; | |
251 | ||
252 | for (rot = 0; rot < 16; rot++) | |
253 | if ((x & ~ror32(0xff, 2 * rot)) == 0) | |
254 | return rol32(x, 2 * rot) | (rot << 8); | |
255 | ||
256 | return -1; | |
257 | } | |
258 | ||
259 | #if __LINUX_ARM_ARCH__ < 7 | |
260 | ||
261 | static u16 imm_offset(u32 k, struct jit_ctx *ctx) | |
262 | { | |
263 | unsigned i = 0, offset; | |
264 | u16 imm; | |
265 | ||
266 | /* on the "fake" run we just count them (duplicates included) */ | |
267 | if (ctx->target == NULL) { | |
268 | ctx->imm_count++; | |
269 | return 0; | |
270 | } | |
271 | ||
272 | while ((i < ctx->imm_count) && ctx->imms[i]) { | |
273 | if (ctx->imms[i] == k) | |
274 | break; | |
275 | i++; | |
276 | } | |
277 | ||
278 | if (ctx->imms[i] == 0) | |
279 | ctx->imms[i] = k; | |
280 | ||
281 | /* constants go just after the epilogue */ | |
282 | offset = ctx->offsets[ctx->skf->len]; | |
283 | offset += ctx->prologue_bytes; | |
284 | offset += ctx->epilogue_bytes; | |
285 | offset += i * 4; | |
286 | ||
287 | ctx->target[offset / 4] = k; | |
288 | ||
289 | /* PC in ARM mode == address of the instruction + 8 */ | |
290 | imm = offset - (8 + ctx->idx * 4); | |
291 | ||
292 | return imm; | |
293 | } | |
294 | ||
295 | #endif /* __LINUX_ARM_ARCH__ */ | |
296 | ||
297 | /* | |
298 | * Move an immediate that's not an imm8m to a core register. | |
299 | */ | |
300 | static inline void emit_mov_i_no8m(int rd, u32 val, struct jit_ctx *ctx) | |
301 | { | |
302 | #if __LINUX_ARM_ARCH__ < 7 | |
303 | emit(ARM_LDR_I(rd, ARM_PC, imm_offset(val, ctx)), ctx); | |
304 | #else | |
305 | emit(ARM_MOVW(rd, val & 0xffff), ctx); | |
306 | if (val > 0xffff) | |
307 | emit(ARM_MOVT(rd, val >> 16), ctx); | |
308 | #endif | |
309 | } | |
310 | ||
311 | static inline void emit_mov_i(int rd, u32 val, struct jit_ctx *ctx) | |
312 | { | |
313 | int imm12 = imm8m(val); | |
314 | ||
315 | if (imm12 >= 0) | |
316 | emit(ARM_MOV_I(rd, imm12), ctx); | |
317 | else | |
318 | emit_mov_i_no8m(rd, val, ctx); | |
319 | } | |
320 | ||
321 | #if __LINUX_ARM_ARCH__ < 6 | |
322 | ||
323 | static void emit_load_be32(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx) | |
324 | { | |
325 | _emit(cond, ARM_LDRB_I(ARM_R3, r_addr, 1), ctx); | |
326 | _emit(cond, ARM_LDRB_I(ARM_R1, r_addr, 0), ctx); | |
327 | _emit(cond, ARM_LDRB_I(ARM_R2, r_addr, 3), ctx); | |
328 | _emit(cond, ARM_LSL_I(ARM_R3, ARM_R3, 16), ctx); | |
329 | _emit(cond, ARM_LDRB_I(ARM_R0, r_addr, 2), ctx); | |
330 | _emit(cond, ARM_ORR_S(ARM_R3, ARM_R3, ARM_R1, SRTYPE_LSL, 24), ctx); | |
331 | _emit(cond, ARM_ORR_R(ARM_R3, ARM_R3, ARM_R2), ctx); | |
332 | _emit(cond, ARM_ORR_S(r_res, ARM_R3, ARM_R0, SRTYPE_LSL, 8), ctx); | |
333 | } | |
334 | ||
335 | static void emit_load_be16(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx) | |
336 | { | |
337 | _emit(cond, ARM_LDRB_I(ARM_R1, r_addr, 0), ctx); | |
338 | _emit(cond, ARM_LDRB_I(ARM_R2, r_addr, 1), ctx); | |
339 | _emit(cond, ARM_ORR_S(r_res, ARM_R2, ARM_R1, SRTYPE_LSL, 8), ctx); | |
340 | } | |
341 | ||
342 | static inline void emit_swap16(u8 r_dst, u8 r_src, struct jit_ctx *ctx) | |
343 | { | |
344 | emit(ARM_LSL_R(ARM_R1, r_src, 8), ctx); | |
345 | emit(ARM_ORR_S(r_dst, ARM_R1, r_src, SRTYPE_LSL, 8), ctx); | |
346 | emit(ARM_LSL_I(r_dst, r_dst, 8), ctx); | |
347 | emit(ARM_LSL_R(r_dst, r_dst, 8), ctx); | |
348 | } | |
349 | ||
350 | #else /* ARMv6+ */ | |
351 | ||
352 | static void emit_load_be32(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx) | |
353 | { | |
354 | _emit(cond, ARM_LDR_I(r_res, r_addr, 0), ctx); | |
355 | #ifdef __LITTLE_ENDIAN | |
356 | _emit(cond, ARM_REV(r_res, r_res), ctx); | |
357 | #endif | |
358 | } | |
359 | ||
360 | static void emit_load_be16(u8 cond, u8 r_res, u8 r_addr, struct jit_ctx *ctx) | |
361 | { | |
362 | _emit(cond, ARM_LDRH_I(r_res, r_addr, 0), ctx); | |
363 | #ifdef __LITTLE_ENDIAN | |
364 | _emit(cond, ARM_REV16(r_res, r_res), ctx); | |
365 | #endif | |
366 | } | |
367 | ||
368 | static inline void emit_swap16(u8 r_dst __maybe_unused, | |
369 | u8 r_src __maybe_unused, | |
370 | struct jit_ctx *ctx __maybe_unused) | |
371 | { | |
372 | #ifdef __LITTLE_ENDIAN | |
373 | emit(ARM_REV16(r_dst, r_src), ctx); | |
374 | #endif | |
375 | } | |
376 | ||
377 | #endif /* __LINUX_ARM_ARCH__ < 6 */ | |
378 | ||
379 | ||
380 | /* Compute the immediate value for a PC-relative branch. */ | |
381 | static inline u32 b_imm(unsigned tgt, struct jit_ctx *ctx) | |
382 | { | |
383 | u32 imm; | |
384 | ||
385 | if (ctx->target == NULL) | |
386 | return 0; | |
387 | /* | |
388 | * BPF allows only forward jumps and the offset of the target is | |
389 | * still the one computed during the first pass. | |
390 | */ | |
391 | imm = ctx->offsets[tgt] + ctx->prologue_bytes - (ctx->idx * 4 + 8); | |
392 | ||
393 | return imm >> 2; | |
394 | } | |
395 | ||
396 | #define OP_IMM3(op, r1, r2, imm_val, ctx) \ | |
397 | do { \ | |
398 | imm12 = imm8m(imm_val); \ | |
399 | if (imm12 < 0) { \ | |
400 | emit_mov_i_no8m(r_scratch, imm_val, ctx); \ | |
401 | emit(op ## _R((r1), (r2), r_scratch), ctx); \ | |
402 | } else { \ | |
403 | emit(op ## _I((r1), (r2), imm12), ctx); \ | |
404 | } \ | |
405 | } while (0) | |
406 | ||
407 | static inline void emit_err_ret(u8 cond, struct jit_ctx *ctx) | |
408 | { | |
409 | if (ctx->ret0_fp_idx >= 0) { | |
410 | _emit(cond, ARM_B(b_imm(ctx->ret0_fp_idx, ctx)), ctx); | |
411 | /* NOP to keep the size constant between passes */ | |
412 | emit(ARM_MOV_R(ARM_R0, ARM_R0), ctx); | |
413 | } else { | |
414 | _emit(cond, ARM_MOV_I(ARM_R0, 0), ctx); | |
415 | _emit(cond, ARM_B(b_imm(ctx->skf->len, ctx)), ctx); | |
416 | } | |
417 | } | |
418 | ||
419 | static inline void emit_blx_r(u8 tgt_reg, struct jit_ctx *ctx) | |
420 | { | |
421 | #if __LINUX_ARM_ARCH__ < 5 | |
422 | emit(ARM_MOV_R(ARM_LR, ARM_PC), ctx); | |
423 | ||
424 | if (elf_hwcap & HWCAP_THUMB) | |
425 | emit(ARM_BX(tgt_reg), ctx); | |
426 | else | |
427 | emit(ARM_MOV_R(ARM_PC, tgt_reg), ctx); | |
428 | #else | |
429 | emit(ARM_BLX_R(tgt_reg), ctx); | |
430 | #endif | |
431 | } | |
432 | ||
433 | static inline void emit_udiv(u8 rd, u8 rm, u8 rn, struct jit_ctx *ctx) | |
434 | { | |
435 | #if __LINUX_ARM_ARCH__ == 7 | |
436 | if (elf_hwcap & HWCAP_IDIVA) { | |
437 | emit(ARM_UDIV(rd, rm, rn), ctx); | |
438 | return; | |
439 | } | |
440 | #endif | |
441 | if (rm != ARM_R0) | |
442 | emit(ARM_MOV_R(ARM_R0, rm), ctx); | |
443 | if (rn != ARM_R1) | |
444 | emit(ARM_MOV_R(ARM_R1, rn), ctx); | |
445 | ||
446 | ctx->seen |= SEEN_CALL; | |
447 | emit_mov_i(ARM_R3, (u32)jit_udiv, ctx); | |
448 | emit_blx_r(ARM_R3, ctx); | |
449 | ||
450 | if (rd != ARM_R0) | |
451 | emit(ARM_MOV_R(rd, ARM_R0), ctx); | |
452 | } | |
453 | ||
454 | static inline void update_on_xread(struct jit_ctx *ctx) | |
455 | { | |
456 | if (!(ctx->seen & SEEN_X)) | |
457 | ctx->flags |= FLAG_NEED_X_RESET; | |
458 | ||
459 | ctx->seen |= SEEN_X; | |
460 | } | |
461 | ||
462 | static int build_body(struct jit_ctx *ctx) | |
463 | { | |
464 | void *load_func[] = {jit_get_skb_b, jit_get_skb_h, jit_get_skb_w}; | |
465 | const struct sk_filter *prog = ctx->skf; | |
466 | const struct sock_filter *inst; | |
467 | unsigned i, load_order, off, condt; | |
468 | int imm12; | |
469 | u32 k; | |
470 | ||
471 | for (i = 0; i < prog->len; i++) { | |
472 | inst = &(prog->insns[i]); | |
473 | /* K as an immediate value operand */ | |
474 | k = inst->k; | |
475 | ||
476 | /* compute offsets only in the fake pass */ | |
477 | if (ctx->target == NULL) | |
478 | ctx->offsets[i] = ctx->idx * 4; | |
479 | ||
480 | switch (inst->code) { | |
481 | case BPF_S_LD_IMM: | |
482 | emit_mov_i(r_A, k, ctx); | |
483 | break; | |
484 | case BPF_S_LD_W_LEN: | |
485 | ctx->seen |= SEEN_SKB; | |
486 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4); | |
487 | emit(ARM_LDR_I(r_A, r_skb, | |
488 | offsetof(struct sk_buff, len)), ctx); | |
489 | break; | |
490 | case BPF_S_LD_MEM: | |
491 | /* A = scratch[k] */ | |
492 | ctx->seen |= SEEN_MEM_WORD(k); | |
493 | emit(ARM_LDR_I(r_A, ARM_SP, SCRATCH_OFF(k)), ctx); | |
494 | break; | |
495 | case BPF_S_LD_W_ABS: | |
496 | load_order = 2; | |
497 | goto load; | |
498 | case BPF_S_LD_H_ABS: | |
499 | load_order = 1; | |
500 | goto load; | |
501 | case BPF_S_LD_B_ABS: | |
502 | load_order = 0; | |
503 | load: | |
504 | /* the interpreter will deal with the negative K */ | |
505 | if ((int)k < 0) | |
506 | return -ENOTSUPP; | |
507 | emit_mov_i(r_off, k, ctx); | |
508 | load_common: | |
509 | ctx->seen |= SEEN_DATA | SEEN_CALL; | |
510 | ||
511 | if (load_order > 0) { | |
512 | emit(ARM_SUB_I(r_scratch, r_skb_hl, | |
513 | 1 << load_order), ctx); | |
514 | emit(ARM_CMP_R(r_scratch, r_off), ctx); | |
515 | condt = ARM_COND_HS; | |
516 | } else { | |
517 | emit(ARM_CMP_R(r_skb_hl, r_off), ctx); | |
518 | condt = ARM_COND_HI; | |
519 | } | |
520 | ||
521 | _emit(condt, ARM_ADD_R(r_scratch, r_off, r_skb_data), | |
522 | ctx); | |
523 | ||
524 | if (load_order == 0) | |
525 | _emit(condt, ARM_LDRB_I(r_A, r_scratch, 0), | |
526 | ctx); | |
527 | else if (load_order == 1) | |
528 | emit_load_be16(condt, r_A, r_scratch, ctx); | |
529 | else if (load_order == 2) | |
530 | emit_load_be32(condt, r_A, r_scratch, ctx); | |
531 | ||
532 | _emit(condt, ARM_B(b_imm(i + 1, ctx)), ctx); | |
533 | ||
534 | /* the slowpath */ | |
535 | emit_mov_i(ARM_R3, (u32)load_func[load_order], ctx); | |
536 | emit(ARM_MOV_R(ARM_R0, r_skb), ctx); | |
537 | /* the offset is already in R1 */ | |
538 | emit_blx_r(ARM_R3, ctx); | |
539 | /* check the result of skb_copy_bits */ | |
540 | emit(ARM_CMP_I(ARM_R1, 0), ctx); | |
541 | emit_err_ret(ARM_COND_NE, ctx); | |
542 | emit(ARM_MOV_R(r_A, ARM_R0), ctx); | |
543 | break; | |
544 | case BPF_S_LD_W_IND: | |
545 | load_order = 2; | |
546 | goto load_ind; | |
547 | case BPF_S_LD_H_IND: | |
548 | load_order = 1; | |
549 | goto load_ind; | |
550 | case BPF_S_LD_B_IND: | |
551 | load_order = 0; | |
552 | load_ind: | |
553 | OP_IMM3(ARM_ADD, r_off, r_X, k, ctx); | |
554 | goto load_common; | |
555 | case BPF_S_LDX_IMM: | |
556 | ctx->seen |= SEEN_X; | |
557 | emit_mov_i(r_X, k, ctx); | |
558 | break; | |
559 | case BPF_S_LDX_W_LEN: | |
560 | ctx->seen |= SEEN_X | SEEN_SKB; | |
561 | emit(ARM_LDR_I(r_X, r_skb, | |
562 | offsetof(struct sk_buff, len)), ctx); | |
563 | break; | |
564 | case BPF_S_LDX_MEM: | |
565 | ctx->seen |= SEEN_X | SEEN_MEM_WORD(k); | |
566 | emit(ARM_LDR_I(r_X, ARM_SP, SCRATCH_OFF(k)), ctx); | |
567 | break; | |
568 | case BPF_S_LDX_B_MSH: | |
569 | /* x = ((*(frame + k)) & 0xf) << 2; */ | |
570 | ctx->seen |= SEEN_X | SEEN_DATA | SEEN_CALL; | |
571 | /* the interpreter should deal with the negative K */ | |
572 | if (k < 0) | |
573 | return -1; | |
574 | /* offset in r1: we might have to take the slow path */ | |
575 | emit_mov_i(r_off, k, ctx); | |
576 | emit(ARM_CMP_R(r_skb_hl, r_off), ctx); | |
577 | ||
578 | /* load in r0: common with the slowpath */ | |
579 | _emit(ARM_COND_HI, ARM_LDRB_R(ARM_R0, r_skb_data, | |
580 | ARM_R1), ctx); | |
581 | /* | |
582 | * emit_mov_i() might generate one or two instructions, | |
583 | * the same holds for emit_blx_r() | |
584 | */ | |
585 | _emit(ARM_COND_HI, ARM_B(b_imm(i + 1, ctx) - 2), ctx); | |
586 | ||
587 | emit(ARM_MOV_R(ARM_R0, r_skb), ctx); | |
588 | /* r_off is r1 */ | |
589 | emit_mov_i(ARM_R3, (u32)jit_get_skb_b, ctx); | |
590 | emit_blx_r(ARM_R3, ctx); | |
591 | /* check the return value of skb_copy_bits */ | |
592 | emit(ARM_CMP_I(ARM_R1, 0), ctx); | |
593 | emit_err_ret(ARM_COND_NE, ctx); | |
594 | ||
595 | emit(ARM_AND_I(r_X, ARM_R0, 0x00f), ctx); | |
596 | emit(ARM_LSL_I(r_X, r_X, 2), ctx); | |
597 | break; | |
598 | case BPF_S_ST: | |
599 | ctx->seen |= SEEN_MEM_WORD(k); | |
600 | emit(ARM_STR_I(r_A, ARM_SP, SCRATCH_OFF(k)), ctx); | |
601 | break; | |
602 | case BPF_S_STX: | |
603 | update_on_xread(ctx); | |
604 | ctx->seen |= SEEN_MEM_WORD(k); | |
605 | emit(ARM_STR_I(r_X, ARM_SP, SCRATCH_OFF(k)), ctx); | |
606 | break; | |
607 | case BPF_S_ALU_ADD_K: | |
608 | /* A += K */ | |
609 | OP_IMM3(ARM_ADD, r_A, r_A, k, ctx); | |
610 | break; | |
611 | case BPF_S_ALU_ADD_X: | |
612 | update_on_xread(ctx); | |
613 | emit(ARM_ADD_R(r_A, r_A, r_X), ctx); | |
614 | break; | |
615 | case BPF_S_ALU_SUB_K: | |
616 | /* A -= K */ | |
617 | OP_IMM3(ARM_SUB, r_A, r_A, k, ctx); | |
618 | break; | |
619 | case BPF_S_ALU_SUB_X: | |
620 | update_on_xread(ctx); | |
621 | emit(ARM_SUB_R(r_A, r_A, r_X), ctx); | |
622 | break; | |
623 | case BPF_S_ALU_MUL_K: | |
624 | /* A *= K */ | |
625 | emit_mov_i(r_scratch, k, ctx); | |
626 | emit(ARM_MUL(r_A, r_A, r_scratch), ctx); | |
627 | break; | |
628 | case BPF_S_ALU_MUL_X: | |
629 | update_on_xread(ctx); | |
630 | emit(ARM_MUL(r_A, r_A, r_X), ctx); | |
631 | break; | |
632 | case BPF_S_ALU_DIV_K: | |
633 | /* current k == reciprocal_value(userspace k) */ | |
634 | emit_mov_i(r_scratch, k, ctx); | |
635 | /* A = top 32 bits of the product */ | |
636 | emit(ARM_UMULL(r_scratch, r_A, r_A, r_scratch), ctx); | |
637 | break; | |
638 | case BPF_S_ALU_DIV_X: | |
639 | update_on_xread(ctx); | |
640 | emit(ARM_CMP_I(r_X, 0), ctx); | |
641 | emit_err_ret(ARM_COND_EQ, ctx); | |
642 | emit_udiv(r_A, r_A, r_X, ctx); | |
643 | break; | |
644 | case BPF_S_ALU_OR_K: | |
645 | /* A |= K */ | |
646 | OP_IMM3(ARM_ORR, r_A, r_A, k, ctx); | |
647 | break; | |
648 | case BPF_S_ALU_OR_X: | |
649 | update_on_xread(ctx); | |
650 | emit(ARM_ORR_R(r_A, r_A, r_X), ctx); | |
651 | break; | |
3cbe2041 DB |
652 | case BPF_S_ALU_XOR_K: |
653 | /* A ^= K; */ | |
654 | OP_IMM3(ARM_EOR, r_A, r_A, k, ctx); | |
655 | break; | |
656 | case BPF_S_ANC_ALU_XOR_X: | |
657 | case BPF_S_ALU_XOR_X: | |
658 | /* A ^= X */ | |
659 | update_on_xread(ctx); | |
660 | emit(ARM_EOR_R(r_A, r_A, r_X), ctx); | |
661 | break; | |
ddecdfce MG |
662 | case BPF_S_ALU_AND_K: |
663 | /* A &= K */ | |
664 | OP_IMM3(ARM_AND, r_A, r_A, k, ctx); | |
665 | break; | |
666 | case BPF_S_ALU_AND_X: | |
667 | update_on_xread(ctx); | |
668 | emit(ARM_AND_R(r_A, r_A, r_X), ctx); | |
669 | break; | |
670 | case BPF_S_ALU_LSH_K: | |
671 | if (unlikely(k > 31)) | |
672 | return -1; | |
673 | emit(ARM_LSL_I(r_A, r_A, k), ctx); | |
674 | break; | |
675 | case BPF_S_ALU_LSH_X: | |
676 | update_on_xread(ctx); | |
677 | emit(ARM_LSL_R(r_A, r_A, r_X), ctx); | |
678 | break; | |
679 | case BPF_S_ALU_RSH_K: | |
680 | if (unlikely(k > 31)) | |
681 | return -1; | |
682 | emit(ARM_LSR_I(r_A, r_A, k), ctx); | |
683 | break; | |
684 | case BPF_S_ALU_RSH_X: | |
685 | update_on_xread(ctx); | |
686 | emit(ARM_LSR_R(r_A, r_A, r_X), ctx); | |
687 | break; | |
688 | case BPF_S_ALU_NEG: | |
689 | /* A = -A */ | |
690 | emit(ARM_RSB_I(r_A, r_A, 0), ctx); | |
691 | break; | |
692 | case BPF_S_JMP_JA: | |
693 | /* pc += K */ | |
694 | emit(ARM_B(b_imm(i + k + 1, ctx)), ctx); | |
695 | break; | |
696 | case BPF_S_JMP_JEQ_K: | |
697 | /* pc += (A == K) ? pc->jt : pc->jf */ | |
698 | condt = ARM_COND_EQ; | |
699 | goto cmp_imm; | |
700 | case BPF_S_JMP_JGT_K: | |
701 | /* pc += (A > K) ? pc->jt : pc->jf */ | |
702 | condt = ARM_COND_HI; | |
703 | goto cmp_imm; | |
704 | case BPF_S_JMP_JGE_K: | |
705 | /* pc += (A >= K) ? pc->jt : pc->jf */ | |
706 | condt = ARM_COND_HS; | |
707 | cmp_imm: | |
708 | imm12 = imm8m(k); | |
709 | if (imm12 < 0) { | |
710 | emit_mov_i_no8m(r_scratch, k, ctx); | |
711 | emit(ARM_CMP_R(r_A, r_scratch), ctx); | |
712 | } else { | |
713 | emit(ARM_CMP_I(r_A, imm12), ctx); | |
714 | } | |
715 | cond_jump: | |
716 | if (inst->jt) | |
717 | _emit(condt, ARM_B(b_imm(i + inst->jt + 1, | |
718 | ctx)), ctx); | |
719 | if (inst->jf) | |
720 | _emit(condt ^ 1, ARM_B(b_imm(i + inst->jf + 1, | |
721 | ctx)), ctx); | |
722 | break; | |
723 | case BPF_S_JMP_JEQ_X: | |
724 | /* pc += (A == X) ? pc->jt : pc->jf */ | |
725 | condt = ARM_COND_EQ; | |
726 | goto cmp_x; | |
727 | case BPF_S_JMP_JGT_X: | |
728 | /* pc += (A > X) ? pc->jt : pc->jf */ | |
729 | condt = ARM_COND_HI; | |
730 | goto cmp_x; | |
731 | case BPF_S_JMP_JGE_X: | |
732 | /* pc += (A >= X) ? pc->jt : pc->jf */ | |
733 | condt = ARM_COND_CS; | |
734 | cmp_x: | |
735 | update_on_xread(ctx); | |
736 | emit(ARM_CMP_R(r_A, r_X), ctx); | |
737 | goto cond_jump; | |
738 | case BPF_S_JMP_JSET_K: | |
739 | /* pc += (A & K) ? pc->jt : pc->jf */ | |
740 | condt = ARM_COND_NE; | |
741 | /* not set iff all zeroes iff Z==1 iff EQ */ | |
742 | ||
743 | imm12 = imm8m(k); | |
744 | if (imm12 < 0) { | |
745 | emit_mov_i_no8m(r_scratch, k, ctx); | |
746 | emit(ARM_TST_R(r_A, r_scratch), ctx); | |
747 | } else { | |
748 | emit(ARM_TST_I(r_A, imm12), ctx); | |
749 | } | |
750 | goto cond_jump; | |
751 | case BPF_S_JMP_JSET_X: | |
752 | /* pc += (A & X) ? pc->jt : pc->jf */ | |
753 | update_on_xread(ctx); | |
754 | condt = ARM_COND_NE; | |
755 | emit(ARM_TST_R(r_A, r_X), ctx); | |
756 | goto cond_jump; | |
757 | case BPF_S_RET_A: | |
758 | emit(ARM_MOV_R(ARM_R0, r_A), ctx); | |
759 | goto b_epilogue; | |
760 | case BPF_S_RET_K: | |
761 | if ((k == 0) && (ctx->ret0_fp_idx < 0)) | |
762 | ctx->ret0_fp_idx = i; | |
763 | emit_mov_i(ARM_R0, k, ctx); | |
764 | b_epilogue: | |
765 | if (i != ctx->skf->len - 1) | |
766 | emit(ARM_B(b_imm(prog->len, ctx)), ctx); | |
767 | break; | |
768 | case BPF_S_MISC_TAX: | |
769 | /* X = A */ | |
770 | ctx->seen |= SEEN_X; | |
771 | emit(ARM_MOV_R(r_X, r_A), ctx); | |
772 | break; | |
773 | case BPF_S_MISC_TXA: | |
774 | /* A = X */ | |
775 | update_on_xread(ctx); | |
776 | emit(ARM_MOV_R(r_A, r_X), ctx); | |
777 | break; | |
778 | case BPF_S_ANC_PROTOCOL: | |
779 | /* A = ntohs(skb->protocol) */ | |
780 | ctx->seen |= SEEN_SKB; | |
781 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, | |
782 | protocol) != 2); | |
783 | off = offsetof(struct sk_buff, protocol); | |
784 | emit(ARM_LDRH_I(r_scratch, r_skb, off), ctx); | |
785 | emit_swap16(r_A, r_scratch, ctx); | |
786 | break; | |
787 | case BPF_S_ANC_CPU: | |
788 | /* r_scratch = current_thread_info() */ | |
789 | OP_IMM3(ARM_BIC, r_scratch, ARM_SP, THREAD_SIZE - 1, ctx); | |
790 | /* A = current_thread_info()->cpu */ | |
791 | BUILD_BUG_ON(FIELD_SIZEOF(struct thread_info, cpu) != 4); | |
792 | off = offsetof(struct thread_info, cpu); | |
793 | emit(ARM_LDR_I(r_A, r_scratch, off), ctx); | |
794 | break; | |
795 | case BPF_S_ANC_IFINDEX: | |
796 | /* A = skb->dev->ifindex */ | |
797 | ctx->seen |= SEEN_SKB; | |
798 | off = offsetof(struct sk_buff, dev); | |
799 | emit(ARM_LDR_I(r_scratch, r_skb, off), ctx); | |
800 | ||
801 | emit(ARM_CMP_I(r_scratch, 0), ctx); | |
802 | emit_err_ret(ARM_COND_EQ, ctx); | |
803 | ||
804 | BUILD_BUG_ON(FIELD_SIZEOF(struct net_device, | |
805 | ifindex) != 4); | |
806 | off = offsetof(struct net_device, ifindex); | |
807 | emit(ARM_LDR_I(r_A, r_scratch, off), ctx); | |
808 | break; | |
809 | case BPF_S_ANC_MARK: | |
810 | ctx->seen |= SEEN_SKB; | |
811 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4); | |
812 | off = offsetof(struct sk_buff, mark); | |
813 | emit(ARM_LDR_I(r_A, r_skb, off), ctx); | |
814 | break; | |
815 | case BPF_S_ANC_RXHASH: | |
816 | ctx->seen |= SEEN_SKB; | |
817 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, rxhash) != 4); | |
818 | off = offsetof(struct sk_buff, rxhash); | |
819 | emit(ARM_LDR_I(r_A, r_skb, off), ctx); | |
820 | break; | |
bf0098f2 DB |
821 | case BPF_S_ANC_VLAN_TAG: |
822 | case BPF_S_ANC_VLAN_TAG_PRESENT: | |
823 | ctx->seen |= SEEN_SKB; | |
824 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, vlan_tci) != 2); | |
825 | off = offsetof(struct sk_buff, vlan_tci); | |
826 | emit(ARM_LDRH_I(r_A, r_skb, off), ctx); | |
827 | if (inst->code == BPF_S_ANC_VLAN_TAG) | |
828 | OP_IMM3(ARM_AND, r_A, r_A, VLAN_VID_MASK, ctx); | |
829 | else | |
830 | OP_IMM3(ARM_AND, r_A, r_A, VLAN_TAG_PRESENT, ctx); | |
831 | break; | |
ddecdfce MG |
832 | case BPF_S_ANC_QUEUE: |
833 | ctx->seen |= SEEN_SKB; | |
834 | BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, | |
835 | queue_mapping) != 2); | |
836 | BUILD_BUG_ON(offsetof(struct sk_buff, | |
837 | queue_mapping) > 0xff); | |
838 | off = offsetof(struct sk_buff, queue_mapping); | |
839 | emit(ARM_LDRH_I(r_A, r_skb, off), ctx); | |
840 | break; | |
841 | default: | |
842 | return -1; | |
843 | } | |
844 | } | |
845 | ||
846 | /* compute offsets only during the first pass */ | |
847 | if (ctx->target == NULL) | |
848 | ctx->offsets[i] = ctx->idx * 4; | |
849 | ||
850 | return 0; | |
851 | } | |
852 | ||
853 | ||
854 | void bpf_jit_compile(struct sk_filter *fp) | |
855 | { | |
856 | struct jit_ctx ctx; | |
857 | unsigned tmp_idx; | |
858 | unsigned alloc_size; | |
859 | ||
860 | if (!bpf_jit_enable) | |
861 | return; | |
862 | ||
863 | memset(&ctx, 0, sizeof(ctx)); | |
864 | ctx.skf = fp; | |
865 | ctx.ret0_fp_idx = -1; | |
866 | ||
89c2e009 | 867 | ctx.offsets = kzalloc(4 * (ctx.skf->len + 1), GFP_KERNEL); |
ddecdfce MG |
868 | if (ctx.offsets == NULL) |
869 | return; | |
870 | ||
871 | /* fake pass to fill in the ctx->seen */ | |
872 | if (unlikely(build_body(&ctx))) | |
873 | goto out; | |
874 | ||
875 | tmp_idx = ctx.idx; | |
876 | build_prologue(&ctx); | |
877 | ctx.prologue_bytes = (ctx.idx - tmp_idx) * 4; | |
878 | ||
879 | #if __LINUX_ARM_ARCH__ < 7 | |
880 | tmp_idx = ctx.idx; | |
881 | build_epilogue(&ctx); | |
882 | ctx.epilogue_bytes = (ctx.idx - tmp_idx) * 4; | |
883 | ||
884 | ctx.idx += ctx.imm_count; | |
885 | if (ctx.imm_count) { | |
89c2e009 | 886 | ctx.imms = kzalloc(4 * ctx.imm_count, GFP_KERNEL); |
ddecdfce MG |
887 | if (ctx.imms == NULL) |
888 | goto out; | |
889 | } | |
890 | #else | |
891 | /* there's nothing after the epilogue on ARMv7 */ | |
892 | build_epilogue(&ctx); | |
893 | #endif | |
894 | ||
895 | alloc_size = 4 * ctx.idx; | |
896 | ctx.target = module_alloc(max(sizeof(struct work_struct), | |
897 | alloc_size)); | |
898 | if (unlikely(ctx.target == NULL)) | |
899 | goto out; | |
900 | ||
901 | ctx.idx = 0; | |
902 | build_prologue(&ctx); | |
903 | build_body(&ctx); | |
904 | build_epilogue(&ctx); | |
905 | ||
906 | flush_icache_range((u32)ctx.target, (u32)(ctx.target + ctx.idx)); | |
907 | ||
908 | #if __LINUX_ARM_ARCH__ < 7 | |
909 | if (ctx.imm_count) | |
910 | kfree(ctx.imms); | |
911 | #endif | |
912 | ||
913 | if (bpf_jit_enable > 1) | |
914 | print_hex_dump(KERN_INFO, "BPF JIT code: ", | |
915 | DUMP_PREFIX_ADDRESS, 16, 4, ctx.target, | |
916 | alloc_size, false); | |
917 | ||
918 | fp->bpf_func = (void *)ctx.target; | |
919 | out: | |
920 | kfree(ctx.offsets); | |
921 | return; | |
922 | } | |
923 | ||
924 | static void bpf_jit_free_worker(struct work_struct *work) | |
925 | { | |
926 | module_free(NULL, work); | |
927 | } | |
928 | ||
929 | void bpf_jit_free(struct sk_filter *fp) | |
930 | { | |
931 | struct work_struct *work; | |
932 | ||
933 | if (fp->bpf_func != sk_run_filter) { | |
934 | work = (struct work_struct *)fp->bpf_func; | |
935 | ||
936 | INIT_WORK(work, bpf_jit_free_worker); | |
937 | schedule_work(work); | |
938 | } | |
939 | } |