mm: hugetlb: fix hugepage memory leak caused by wrong reserve count
[linux/fpc-iii.git] / arch / mips / net / bpf_jit.c
blob77cb27309db27f781ec9862037074d39e928401a
1 /*
2 * Just-In-Time compiler for BPF filters on MIPS
4 * Copyright (c) 2014 Imagination Technologies Ltd.
5 * Author: Markos Chandras <markos.chandras@imgtec.com>
7 * This program is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License as published by the
9 * Free Software Foundation; version 2 of the License.
12 #include <linux/bitops.h>
13 #include <linux/compiler.h>
14 #include <linux/errno.h>
15 #include <linux/filter.h>
16 #include <linux/if_vlan.h>
17 #include <linux/kconfig.h>
18 #include <linux/moduleloader.h>
19 #include <linux/netdevice.h>
20 #include <linux/string.h>
21 #include <linux/slab.h>
22 #include <linux/types.h>
23 #include <asm/asm.h>
24 #include <asm/bitops.h>
25 #include <asm/cacheflush.h>
26 #include <asm/cpu-features.h>
27 #include <asm/uasm.h>
29 #include "bpf_jit.h"
31 /* ABI
32 * r_skb_hl SKB header length
33 * r_data SKB data pointer
34 * r_off Offset
35 * r_A BPF register A
36 * r_X BPF register X
37 * r_skb *skb
38 * r_M *scratch memory
39 * r_skb_len SKB length
41 * On entry (*bpf_func)(*skb, *filter)
42 * a0 = MIPS_R_A0 = skb;
43 * a1 = MIPS_R_A1 = filter;
45 * Stack
46 * ...
47 * M[15]
48 * M[14]
49 * M[13]
50 * ...
51 * M[0] <-- r_M
52 * saved reg k-1
53 * saved reg k-2
54 * ...
55 * saved reg 0 <-- r_sp
56 * <no argument area>
58 * Packet layout
60 * <--------------------- len ------------------------>
61 * <--skb-len(r_skb_hl)-->< ----- skb->data_len ------>
62 * ----------------------------------------------------
63 * | skb->data |
64 * ----------------------------------------------------
67 #define ptr typeof(unsigned long)
69 #define SCRATCH_OFF(k) (4 * (k))
71 /* JIT flags */
72 #define SEEN_CALL (1 << BPF_MEMWORDS)
73 #define SEEN_SREG_SFT (BPF_MEMWORDS + 1)
74 #define SEEN_SREG_BASE (1 << SEEN_SREG_SFT)
75 #define SEEN_SREG(x) (SEEN_SREG_BASE << (x))
76 #define SEEN_OFF SEEN_SREG(2)
77 #define SEEN_A SEEN_SREG(3)
78 #define SEEN_X SEEN_SREG(4)
79 #define SEEN_SKB SEEN_SREG(5)
80 #define SEEN_MEM SEEN_SREG(6)
81 /* SEEN_SK_DATA also implies skb_hl an skb_len */
82 #define SEEN_SKB_DATA (SEEN_SREG(7) | SEEN_SREG(1) | SEEN_SREG(0))
84 /* Arguments used by JIT */
85 #define ARGS_USED_BY_JIT 2 /* only applicable to 64-bit */
87 #define SBIT(x) (1 << (x)) /* Signed version of BIT() */
89 /**
90 * struct jit_ctx - JIT context
91 * @skf: The sk_filter
92 * @prologue_bytes: Number of bytes for prologue
93 * @idx: Instruction index
94 * @flags: JIT flags
95 * @offsets: Instruction offsets
96 * @target: Memory location for the compiled filter
98 struct jit_ctx {
99 const struct bpf_prog *skf;
100 unsigned int prologue_bytes;
101 u32 idx;
102 u32 flags;
103 u32 *offsets;
104 u32 *target;
108 static inline int optimize_div(u32 *k)
110 /* power of 2 divides can be implemented with right shift */
111 if (!(*k & (*k-1))) {
112 *k = ilog2(*k);
113 return 1;
116 return 0;
119 static inline void emit_jit_reg_move(ptr dst, ptr src, struct jit_ctx *ctx);
121 /* Simply emit the instruction if the JIT memory space has been allocated */
122 #define emit_instr(ctx, func, ...) \
123 do { \
124 if ((ctx)->target != NULL) { \
125 u32 *p = &(ctx)->target[ctx->idx]; \
126 uasm_i_##func(&p, ##__VA_ARGS__); \
128 (ctx)->idx++; \
129 } while (0)
132 * Similar to emit_instr but it must be used when we need to emit
133 * 32-bit or 64-bit instructions
135 #define emit_long_instr(ctx, func, ...) \
136 do { \
137 if ((ctx)->target != NULL) { \
138 u32 *p = &(ctx)->target[ctx->idx]; \
139 UASM_i_##func(&p, ##__VA_ARGS__); \
141 (ctx)->idx++; \
142 } while (0)
144 /* Determine if immediate is within the 16-bit signed range */
145 static inline bool is_range16(s32 imm)
147 return !(imm >= SBIT(15) || imm < -SBIT(15));
150 static inline void emit_addu(unsigned int dst, unsigned int src1,
151 unsigned int src2, struct jit_ctx *ctx)
153 emit_instr(ctx, addu, dst, src1, src2);
156 static inline void emit_nop(struct jit_ctx *ctx)
158 emit_instr(ctx, nop);
161 /* Load a u32 immediate to a register */
162 static inline void emit_load_imm(unsigned int dst, u32 imm, struct jit_ctx *ctx)
164 if (ctx->target != NULL) {
165 /* addiu can only handle s16 */
166 if (!is_range16(imm)) {
167 u32 *p = &ctx->target[ctx->idx];
168 uasm_i_lui(&p, r_tmp_imm, (s32)imm >> 16);
169 p = &ctx->target[ctx->idx + 1];
170 uasm_i_ori(&p, dst, r_tmp_imm, imm & 0xffff);
171 } else {
172 u32 *p = &ctx->target[ctx->idx];
173 uasm_i_addiu(&p, dst, r_zero, imm);
176 ctx->idx++;
178 if (!is_range16(imm))
179 ctx->idx++;
182 static inline void emit_or(unsigned int dst, unsigned int src1,
183 unsigned int src2, struct jit_ctx *ctx)
185 emit_instr(ctx, or, dst, src1, src2);
188 static inline void emit_ori(unsigned int dst, unsigned src, u32 imm,
189 struct jit_ctx *ctx)
191 if (imm >= BIT(16)) {
192 emit_load_imm(r_tmp, imm, ctx);
193 emit_or(dst, src, r_tmp, ctx);
194 } else {
195 emit_instr(ctx, ori, dst, src, imm);
199 static inline void emit_daddiu(unsigned int dst, unsigned int src,
200 int imm, struct jit_ctx *ctx)
203 * Only used for stack, so the imm is relatively small
204 * and it fits in 15-bits
206 emit_instr(ctx, daddiu, dst, src, imm);
209 static inline void emit_addiu(unsigned int dst, unsigned int src,
210 u32 imm, struct jit_ctx *ctx)
212 if (!is_range16(imm)) {
213 emit_load_imm(r_tmp, imm, ctx);
214 emit_addu(dst, r_tmp, src, ctx);
215 } else {
216 emit_instr(ctx, addiu, dst, src, imm);
220 static inline void emit_and(unsigned int dst, unsigned int src1,
221 unsigned int src2, struct jit_ctx *ctx)
223 emit_instr(ctx, and, dst, src1, src2);
226 static inline void emit_andi(unsigned int dst, unsigned int src,
227 u32 imm, struct jit_ctx *ctx)
229 /* If imm does not fit in u16 then load it to register */
230 if (imm >= BIT(16)) {
231 emit_load_imm(r_tmp, imm, ctx);
232 emit_and(dst, src, r_tmp, ctx);
233 } else {
234 emit_instr(ctx, andi, dst, src, imm);
238 static inline void emit_xor(unsigned int dst, unsigned int src1,
239 unsigned int src2, struct jit_ctx *ctx)
241 emit_instr(ctx, xor, dst, src1, src2);
244 static inline void emit_xori(ptr dst, ptr src, u32 imm, struct jit_ctx *ctx)
246 /* If imm does not fit in u16 then load it to register */
247 if (imm >= BIT(16)) {
248 emit_load_imm(r_tmp, imm, ctx);
249 emit_xor(dst, src, r_tmp, ctx);
250 } else {
251 emit_instr(ctx, xori, dst, src, imm);
255 static inline void emit_stack_offset(int offset, struct jit_ctx *ctx)
257 emit_long_instr(ctx, ADDIU, r_sp, r_sp, offset);
260 static inline void emit_subu(unsigned int dst, unsigned int src1,
261 unsigned int src2, struct jit_ctx *ctx)
263 emit_instr(ctx, subu, dst, src1, src2);
266 static inline void emit_neg(unsigned int reg, struct jit_ctx *ctx)
268 emit_subu(reg, r_zero, reg, ctx);
271 static inline void emit_sllv(unsigned int dst, unsigned int src,
272 unsigned int sa, struct jit_ctx *ctx)
274 emit_instr(ctx, sllv, dst, src, sa);
277 static inline void emit_sll(unsigned int dst, unsigned int src,
278 unsigned int sa, struct jit_ctx *ctx)
280 /* sa is 5-bits long */
281 if (sa >= BIT(5))
282 /* Shifting >= 32 results in zero */
283 emit_jit_reg_move(dst, r_zero, ctx);
284 else
285 emit_instr(ctx, sll, dst, src, sa);
288 static inline void emit_srlv(unsigned int dst, unsigned int src,
289 unsigned int sa, struct jit_ctx *ctx)
291 emit_instr(ctx, srlv, dst, src, sa);
294 static inline void emit_srl(unsigned int dst, unsigned int src,
295 unsigned int sa, struct jit_ctx *ctx)
297 /* sa is 5-bits long */
298 if (sa >= BIT(5))
299 /* Shifting >= 32 results in zero */
300 emit_jit_reg_move(dst, r_zero, ctx);
301 else
302 emit_instr(ctx, srl, dst, src, sa);
305 static inline void emit_slt(unsigned int dst, unsigned int src1,
306 unsigned int src2, struct jit_ctx *ctx)
308 emit_instr(ctx, slt, dst, src1, src2);
311 static inline void emit_sltu(unsigned int dst, unsigned int src1,
312 unsigned int src2, struct jit_ctx *ctx)
314 emit_instr(ctx, sltu, dst, src1, src2);
317 static inline void emit_sltiu(unsigned dst, unsigned int src,
318 unsigned int imm, struct jit_ctx *ctx)
320 /* 16 bit immediate */
321 if (!is_range16((s32)imm)) {
322 emit_load_imm(r_tmp, imm, ctx);
323 emit_sltu(dst, src, r_tmp, ctx);
324 } else {
325 emit_instr(ctx, sltiu, dst, src, imm);
330 /* Store register on the stack */
331 static inline void emit_store_stack_reg(ptr reg, ptr base,
332 unsigned int offset,
333 struct jit_ctx *ctx)
335 emit_long_instr(ctx, SW, reg, offset, base);
338 static inline void emit_store(ptr reg, ptr base, unsigned int offset,
339 struct jit_ctx *ctx)
341 emit_instr(ctx, sw, reg, offset, base);
344 static inline void emit_load_stack_reg(ptr reg, ptr base,
345 unsigned int offset,
346 struct jit_ctx *ctx)
348 emit_long_instr(ctx, LW, reg, offset, base);
351 static inline void emit_load(unsigned int reg, unsigned int base,
352 unsigned int offset, struct jit_ctx *ctx)
354 emit_instr(ctx, lw, reg, offset, base);
357 static inline void emit_load_byte(unsigned int reg, unsigned int base,
358 unsigned int offset, struct jit_ctx *ctx)
360 emit_instr(ctx, lb, reg, offset, base);
363 static inline void emit_half_load(unsigned int reg, unsigned int base,
364 unsigned int offset, struct jit_ctx *ctx)
366 emit_instr(ctx, lh, reg, offset, base);
369 static inline void emit_mul(unsigned int dst, unsigned int src1,
370 unsigned int src2, struct jit_ctx *ctx)
372 emit_instr(ctx, mul, dst, src1, src2);
375 static inline void emit_div(unsigned int dst, unsigned int src,
376 struct jit_ctx *ctx)
378 if (ctx->target != NULL) {
379 u32 *p = &ctx->target[ctx->idx];
380 uasm_i_divu(&p, dst, src);
381 p = &ctx->target[ctx->idx + 1];
382 uasm_i_mflo(&p, dst);
384 ctx->idx += 2; /* 2 insts */
387 static inline void emit_mod(unsigned int dst, unsigned int src,
388 struct jit_ctx *ctx)
390 if (ctx->target != NULL) {
391 u32 *p = &ctx->target[ctx->idx];
392 uasm_i_divu(&p, dst, src);
393 p = &ctx->target[ctx->idx + 1];
394 uasm_i_mfhi(&p, dst);
396 ctx->idx += 2; /* 2 insts */
399 static inline void emit_dsll(unsigned int dst, unsigned int src,
400 unsigned int sa, struct jit_ctx *ctx)
402 emit_instr(ctx, dsll, dst, src, sa);
405 static inline void emit_dsrl32(unsigned int dst, unsigned int src,
406 unsigned int sa, struct jit_ctx *ctx)
408 emit_instr(ctx, dsrl32, dst, src, sa);
411 static inline void emit_wsbh(unsigned int dst, unsigned int src,
412 struct jit_ctx *ctx)
414 emit_instr(ctx, wsbh, dst, src);
417 /* load pointer to register */
418 static inline void emit_load_ptr(unsigned int dst, unsigned int src,
419 int imm, struct jit_ctx *ctx)
421 /* src contains the base addr of the 32/64-pointer */
422 emit_long_instr(ctx, LW, dst, imm, src);
425 /* load a function pointer to register */
426 static inline void emit_load_func(unsigned int reg, ptr imm,
427 struct jit_ctx *ctx)
429 if (config_enabled(CONFIG_64BIT)) {
430 /* At this point imm is always 64-bit */
431 emit_load_imm(r_tmp, (u64)imm >> 32, ctx);
432 emit_dsll(r_tmp_imm, r_tmp, 16, ctx); /* left shift by 16 */
433 emit_ori(r_tmp, r_tmp_imm, (imm >> 16) & 0xffff, ctx);
434 emit_dsll(r_tmp_imm, r_tmp, 16, ctx); /* left shift by 16 */
435 emit_ori(reg, r_tmp_imm, imm & 0xffff, ctx);
436 } else {
437 emit_load_imm(reg, imm, ctx);
441 /* Move to real MIPS register */
442 static inline void emit_reg_move(ptr dst, ptr src, struct jit_ctx *ctx)
444 emit_long_instr(ctx, ADDU, dst, src, r_zero);
447 /* Move to JIT (32-bit) register */
448 static inline void emit_jit_reg_move(ptr dst, ptr src, struct jit_ctx *ctx)
450 emit_addu(dst, src, r_zero, ctx);
453 /* Compute the immediate value for PC-relative branches. */
454 static inline u32 b_imm(unsigned int tgt, struct jit_ctx *ctx)
456 if (ctx->target == NULL)
457 return 0;
460 * We want a pc-relative branch. We only do forward branches
461 * so tgt is always after pc. tgt is the instruction offset
462 * we want to jump to.
464 * Branch on MIPS:
465 * I: target_offset <- sign_extend(offset)
466 * I+1: PC += target_offset (delay slot)
468 * ctx->idx currently points to the branch instruction
469 * but the offset is added to the delay slot so we need
470 * to subtract 4.
472 return ctx->offsets[tgt] -
473 (ctx->idx * 4 - ctx->prologue_bytes) - 4;
476 static inline void emit_bcond(int cond, unsigned int reg1, unsigned int reg2,
477 unsigned int imm, struct jit_ctx *ctx)
479 if (ctx->target != NULL) {
480 u32 *p = &ctx->target[ctx->idx];
482 switch (cond) {
483 case MIPS_COND_EQ:
484 uasm_i_beq(&p, reg1, reg2, imm);
485 break;
486 case MIPS_COND_NE:
487 uasm_i_bne(&p, reg1, reg2, imm);
488 break;
489 case MIPS_COND_ALL:
490 uasm_i_b(&p, imm);
491 break;
492 default:
493 pr_warn("%s: Unhandled branch conditional: %d\n",
494 __func__, cond);
497 ctx->idx++;
500 static inline void emit_b(unsigned int imm, struct jit_ctx *ctx)
502 emit_bcond(MIPS_COND_ALL, r_zero, r_zero, imm, ctx);
505 static inline void emit_jalr(unsigned int link, unsigned int reg,
506 struct jit_ctx *ctx)
508 emit_instr(ctx, jalr, link, reg);
511 static inline void emit_jr(unsigned int reg, struct jit_ctx *ctx)
513 emit_instr(ctx, jr, reg);
516 static inline u16 align_sp(unsigned int num)
518 /* Double word alignment for 32-bit, quadword for 64-bit */
519 unsigned int align = config_enabled(CONFIG_64BIT) ? 16 : 8;
520 num = (num + (align - 1)) & -align;
521 return num;
524 static bool is_load_to_a(u16 inst)
526 switch (inst) {
527 case BPF_LD | BPF_W | BPF_LEN:
528 case BPF_LD | BPF_W | BPF_ABS:
529 case BPF_LD | BPF_H | BPF_ABS:
530 case BPF_LD | BPF_B | BPF_ABS:
531 return true;
532 default:
533 return false;
537 static void save_bpf_jit_regs(struct jit_ctx *ctx, unsigned offset)
539 int i = 0, real_off = 0;
540 u32 sflags, tmp_flags;
542 /* Adjust the stack pointer */
543 emit_stack_offset(-align_sp(offset), ctx);
545 tmp_flags = sflags = ctx->flags >> SEEN_SREG_SFT;
546 /* sflags is essentially a bitmap */
547 while (tmp_flags) {
548 if ((sflags >> i) & 0x1) {
549 emit_store_stack_reg(MIPS_R_S0 + i, r_sp, real_off,
550 ctx);
551 real_off += SZREG;
553 i++;
554 tmp_flags >>= 1;
557 /* save return address */
558 if (ctx->flags & SEEN_CALL) {
559 emit_store_stack_reg(r_ra, r_sp, real_off, ctx);
560 real_off += SZREG;
563 /* Setup r_M leaving the alignment gap if necessary */
564 if (ctx->flags & SEEN_MEM) {
565 if (real_off % (SZREG * 2))
566 real_off += SZREG;
567 emit_long_instr(ctx, ADDIU, r_M, r_sp, real_off);
571 static void restore_bpf_jit_regs(struct jit_ctx *ctx,
572 unsigned int offset)
574 int i, real_off = 0;
575 u32 sflags, tmp_flags;
577 tmp_flags = sflags = ctx->flags >> SEEN_SREG_SFT;
578 /* sflags is a bitmap */
579 i = 0;
580 while (tmp_flags) {
581 if ((sflags >> i) & 0x1) {
582 emit_load_stack_reg(MIPS_R_S0 + i, r_sp, real_off,
583 ctx);
584 real_off += SZREG;
586 i++;
587 tmp_flags >>= 1;
590 /* restore return address */
591 if (ctx->flags & SEEN_CALL)
592 emit_load_stack_reg(r_ra, r_sp, real_off, ctx);
594 /* Restore the sp and discard the scrach memory */
595 emit_stack_offset(align_sp(offset), ctx);
598 static unsigned int get_stack_depth(struct jit_ctx *ctx)
600 int sp_off = 0;
603 /* How may s* regs do we need to preserved? */
604 sp_off += hweight32(ctx->flags >> SEEN_SREG_SFT) * SZREG;
606 if (ctx->flags & SEEN_MEM)
607 sp_off += 4 * BPF_MEMWORDS; /* BPF_MEMWORDS are 32-bit */
609 if (ctx->flags & SEEN_CALL)
610 sp_off += SZREG; /* Space for our ra register */
612 return sp_off;
615 static void build_prologue(struct jit_ctx *ctx)
617 u16 first_inst = ctx->skf->insns[0].code;
618 int sp_off;
620 /* Calculate the total offset for the stack pointer */
621 sp_off = get_stack_depth(ctx);
622 save_bpf_jit_regs(ctx, sp_off);
624 if (ctx->flags & SEEN_SKB)
625 emit_reg_move(r_skb, MIPS_R_A0, ctx);
627 if (ctx->flags & SEEN_SKB_DATA) {
628 /* Load packet length */
629 emit_load(r_skb_len, r_skb, offsetof(struct sk_buff, len),
630 ctx);
631 emit_load(r_tmp, r_skb, offsetof(struct sk_buff, data_len),
632 ctx);
633 /* Load the data pointer */
634 emit_load_ptr(r_skb_data, r_skb,
635 offsetof(struct sk_buff, data), ctx);
636 /* Load the header length */
637 emit_subu(r_skb_hl, r_skb_len, r_tmp, ctx);
640 if (ctx->flags & SEEN_X)
641 emit_jit_reg_move(r_X, r_zero, ctx);
643 /* Do not leak kernel data to userspace */
644 if ((first_inst != (BPF_RET | BPF_K)) && !(is_load_to_a(first_inst)))
645 emit_jit_reg_move(r_A, r_zero, ctx);
648 static void build_epilogue(struct jit_ctx *ctx)
650 unsigned int sp_off;
652 /* Calculate the total offset for the stack pointer */
654 sp_off = get_stack_depth(ctx);
655 restore_bpf_jit_regs(ctx, sp_off);
657 /* Return */
658 emit_jr(r_ra, ctx);
659 emit_nop(ctx);
662 #define CHOOSE_LOAD_FUNC(K, func) \
663 ((int)K < 0 ? ((int)K >= SKF_LL_OFF ? func##_negative : func) : \
664 func##_positive)
666 static int build_body(struct jit_ctx *ctx)
668 const struct bpf_prog *prog = ctx->skf;
669 const struct sock_filter *inst;
670 unsigned int i, off, condt;
671 u32 k, b_off __maybe_unused;
672 u8 (*sk_load_func)(unsigned long *skb, int offset);
674 for (i = 0; i < prog->len; i++) {
675 u16 code;
677 inst = &(prog->insns[i]);
678 pr_debug("%s: code->0x%02x, jt->0x%x, jf->0x%x, k->0x%x\n",
679 __func__, inst->code, inst->jt, inst->jf, inst->k);
680 k = inst->k;
681 code = bpf_anc_helper(inst);
683 if (ctx->target == NULL)
684 ctx->offsets[i] = ctx->idx * 4;
686 switch (code) {
687 case BPF_LD | BPF_IMM:
688 /* A <- k ==> li r_A, k */
689 ctx->flags |= SEEN_A;
690 emit_load_imm(r_A, k, ctx);
691 break;
692 case BPF_LD | BPF_W | BPF_LEN:
693 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, len) != 4);
694 /* A <- len ==> lw r_A, offset(skb) */
695 ctx->flags |= SEEN_SKB | SEEN_A;
696 off = offsetof(struct sk_buff, len);
697 emit_load(r_A, r_skb, off, ctx);
698 break;
699 case BPF_LD | BPF_MEM:
700 /* A <- M[k] ==> lw r_A, offset(M) */
701 ctx->flags |= SEEN_MEM | SEEN_A;
702 emit_load(r_A, r_M, SCRATCH_OFF(k), ctx);
703 break;
704 case BPF_LD | BPF_W | BPF_ABS:
705 /* A <- P[k:4] */
706 sk_load_func = CHOOSE_LOAD_FUNC(k, sk_load_word);
707 goto load;
708 case BPF_LD | BPF_H | BPF_ABS:
709 /* A <- P[k:2] */
710 sk_load_func = CHOOSE_LOAD_FUNC(k, sk_load_half);
711 goto load;
712 case BPF_LD | BPF_B | BPF_ABS:
713 /* A <- P[k:1] */
714 sk_load_func = CHOOSE_LOAD_FUNC(k, sk_load_byte);
715 load:
716 emit_load_imm(r_off, k, ctx);
717 load_common:
718 ctx->flags |= SEEN_CALL | SEEN_OFF |
719 SEEN_SKB | SEEN_A | SEEN_SKB_DATA;
721 emit_load_func(r_s0, (ptr)sk_load_func, ctx);
722 emit_reg_move(MIPS_R_A0, r_skb, ctx);
723 emit_jalr(MIPS_R_RA, r_s0, ctx);
724 /* Load second argument to delay slot */
725 emit_reg_move(MIPS_R_A1, r_off, ctx);
726 /* Check the error value */
727 emit_bcond(MIPS_COND_EQ, r_ret, 0, b_imm(i + 1, ctx),
728 ctx);
729 /* Load return register on DS for failures */
730 emit_reg_move(r_ret, r_zero, ctx);
731 /* Return with error */
732 emit_b(b_imm(prog->len, ctx), ctx);
733 emit_nop(ctx);
734 break;
735 case BPF_LD | BPF_W | BPF_IND:
736 /* A <- P[X + k:4] */
737 sk_load_func = sk_load_word;
738 goto load_ind;
739 case BPF_LD | BPF_H | BPF_IND:
740 /* A <- P[X + k:2] */
741 sk_load_func = sk_load_half;
742 goto load_ind;
743 case BPF_LD | BPF_B | BPF_IND:
744 /* A <- P[X + k:1] */
745 sk_load_func = sk_load_byte;
746 load_ind:
747 ctx->flags |= SEEN_OFF | SEEN_X;
748 emit_addiu(r_off, r_X, k, ctx);
749 goto load_common;
750 case BPF_LDX | BPF_IMM:
751 /* X <- k */
752 ctx->flags |= SEEN_X;
753 emit_load_imm(r_X, k, ctx);
754 break;
755 case BPF_LDX | BPF_MEM:
756 /* X <- M[k] */
757 ctx->flags |= SEEN_X | SEEN_MEM;
758 emit_load(r_X, r_M, SCRATCH_OFF(k), ctx);
759 break;
760 case BPF_LDX | BPF_W | BPF_LEN:
761 /* X <- len */
762 ctx->flags |= SEEN_X | SEEN_SKB;
763 off = offsetof(struct sk_buff, len);
764 emit_load(r_X, r_skb, off, ctx);
765 break;
766 case BPF_LDX | BPF_B | BPF_MSH:
767 /* X <- 4 * (P[k:1] & 0xf) */
768 ctx->flags |= SEEN_X | SEEN_CALL | SEEN_SKB;
769 /* Load offset to a1 */
770 emit_load_func(r_s0, (ptr)sk_load_byte, ctx);
772 * This may emit two instructions so it may not fit
773 * in the delay slot. So use a0 in the delay slot.
775 emit_load_imm(MIPS_R_A1, k, ctx);
776 emit_jalr(MIPS_R_RA, r_s0, ctx);
777 emit_reg_move(MIPS_R_A0, r_skb, ctx); /* delay slot */
778 /* Check the error value */
779 emit_bcond(MIPS_COND_NE, r_ret, 0,
780 b_imm(prog->len, ctx), ctx);
781 emit_reg_move(r_ret, r_zero, ctx);
782 /* We are good */
783 /* X <- P[1:K] & 0xf */
784 emit_andi(r_X, r_A, 0xf, ctx);
785 /* X << 2 */
786 emit_b(b_imm(i + 1, ctx), ctx);
787 emit_sll(r_X, r_X, 2, ctx); /* delay slot */
788 break;
789 case BPF_ST:
790 /* M[k] <- A */
791 ctx->flags |= SEEN_MEM | SEEN_A;
792 emit_store(r_A, r_M, SCRATCH_OFF(k), ctx);
793 break;
794 case BPF_STX:
795 /* M[k] <- X */
796 ctx->flags |= SEEN_MEM | SEEN_X;
797 emit_store(r_X, r_M, SCRATCH_OFF(k), ctx);
798 break;
799 case BPF_ALU | BPF_ADD | BPF_K:
800 /* A += K */
801 ctx->flags |= SEEN_A;
802 emit_addiu(r_A, r_A, k, ctx);
803 break;
804 case BPF_ALU | BPF_ADD | BPF_X:
805 /* A += X */
806 ctx->flags |= SEEN_A | SEEN_X;
807 emit_addu(r_A, r_A, r_X, ctx);
808 break;
809 case BPF_ALU | BPF_SUB | BPF_K:
810 /* A -= K */
811 ctx->flags |= SEEN_A;
812 emit_addiu(r_A, r_A, -k, ctx);
813 break;
814 case BPF_ALU | BPF_SUB | BPF_X:
815 /* A -= X */
816 ctx->flags |= SEEN_A | SEEN_X;
817 emit_subu(r_A, r_A, r_X, ctx);
818 break;
819 case BPF_ALU | BPF_MUL | BPF_K:
820 /* A *= K */
821 /* Load K to scratch register before MUL */
822 ctx->flags |= SEEN_A;
823 emit_load_imm(r_s0, k, ctx);
824 emit_mul(r_A, r_A, r_s0, ctx);
825 break;
826 case BPF_ALU | BPF_MUL | BPF_X:
827 /* A *= X */
828 ctx->flags |= SEEN_A | SEEN_X;
829 emit_mul(r_A, r_A, r_X, ctx);
830 break;
831 case BPF_ALU | BPF_DIV | BPF_K:
832 /* A /= k */
833 if (k == 1)
834 break;
835 if (optimize_div(&k)) {
836 ctx->flags |= SEEN_A;
837 emit_srl(r_A, r_A, k, ctx);
838 break;
840 ctx->flags |= SEEN_A;
841 emit_load_imm(r_s0, k, ctx);
842 emit_div(r_A, r_s0, ctx);
843 break;
844 case BPF_ALU | BPF_MOD | BPF_K:
845 /* A %= k */
846 if (k == 1) {
847 ctx->flags |= SEEN_A;
848 emit_jit_reg_move(r_A, r_zero, ctx);
849 } else {
850 ctx->flags |= SEEN_A;
851 emit_load_imm(r_s0, k, ctx);
852 emit_mod(r_A, r_s0, ctx);
854 break;
855 case BPF_ALU | BPF_DIV | BPF_X:
856 /* A /= X */
857 ctx->flags |= SEEN_X | SEEN_A;
858 /* Check if r_X is zero */
859 emit_bcond(MIPS_COND_EQ, r_X, r_zero,
860 b_imm(prog->len, ctx), ctx);
861 emit_load_imm(r_ret, 0, ctx); /* delay slot */
862 emit_div(r_A, r_X, ctx);
863 break;
864 case BPF_ALU | BPF_MOD | BPF_X:
865 /* A %= X */
866 ctx->flags |= SEEN_X | SEEN_A;
867 /* Check if r_X is zero */
868 emit_bcond(MIPS_COND_EQ, r_X, r_zero,
869 b_imm(prog->len, ctx), ctx);
870 emit_load_imm(r_ret, 0, ctx); /* delay slot */
871 emit_mod(r_A, r_X, ctx);
872 break;
873 case BPF_ALU | BPF_OR | BPF_K:
874 /* A |= K */
875 ctx->flags |= SEEN_A;
876 emit_ori(r_A, r_A, k, ctx);
877 break;
878 case BPF_ALU | BPF_OR | BPF_X:
879 /* A |= X */
880 ctx->flags |= SEEN_A;
881 emit_ori(r_A, r_A, r_X, ctx);
882 break;
883 case BPF_ALU | BPF_XOR | BPF_K:
884 /* A ^= k */
885 ctx->flags |= SEEN_A;
886 emit_xori(r_A, r_A, k, ctx);
887 break;
888 case BPF_ANC | SKF_AD_ALU_XOR_X:
889 case BPF_ALU | BPF_XOR | BPF_X:
890 /* A ^= X */
891 ctx->flags |= SEEN_A;
892 emit_xor(r_A, r_A, r_X, ctx);
893 break;
894 case BPF_ALU | BPF_AND | BPF_K:
895 /* A &= K */
896 ctx->flags |= SEEN_A;
897 emit_andi(r_A, r_A, k, ctx);
898 break;
899 case BPF_ALU | BPF_AND | BPF_X:
900 /* A &= X */
901 ctx->flags |= SEEN_A | SEEN_X;
902 emit_and(r_A, r_A, r_X, ctx);
903 break;
904 case BPF_ALU | BPF_LSH | BPF_K:
905 /* A <<= K */
906 ctx->flags |= SEEN_A;
907 emit_sll(r_A, r_A, k, ctx);
908 break;
909 case BPF_ALU | BPF_LSH | BPF_X:
910 /* A <<= X */
911 ctx->flags |= SEEN_A | SEEN_X;
912 emit_sllv(r_A, r_A, r_X, ctx);
913 break;
914 case BPF_ALU | BPF_RSH | BPF_K:
915 /* A >>= K */
916 ctx->flags |= SEEN_A;
917 emit_srl(r_A, r_A, k, ctx);
918 break;
919 case BPF_ALU | BPF_RSH | BPF_X:
920 ctx->flags |= SEEN_A | SEEN_X;
921 emit_srlv(r_A, r_A, r_X, ctx);
922 break;
923 case BPF_ALU | BPF_NEG:
924 /* A = -A */
925 ctx->flags |= SEEN_A;
926 emit_neg(r_A, ctx);
927 break;
928 case BPF_JMP | BPF_JA:
929 /* pc += K */
930 emit_b(b_imm(i + k + 1, ctx), ctx);
931 emit_nop(ctx);
932 break;
933 case BPF_JMP | BPF_JEQ | BPF_K:
934 /* pc += ( A == K ) ? pc->jt : pc->jf */
935 condt = MIPS_COND_EQ | MIPS_COND_K;
936 goto jmp_cmp;
937 case BPF_JMP | BPF_JEQ | BPF_X:
938 ctx->flags |= SEEN_X;
939 /* pc += ( A == X ) ? pc->jt : pc->jf */
940 condt = MIPS_COND_EQ | MIPS_COND_X;
941 goto jmp_cmp;
942 case BPF_JMP | BPF_JGE | BPF_K:
943 /* pc += ( A >= K ) ? pc->jt : pc->jf */
944 condt = MIPS_COND_GE | MIPS_COND_K;
945 goto jmp_cmp;
946 case BPF_JMP | BPF_JGE | BPF_X:
947 ctx->flags |= SEEN_X;
948 /* pc += ( A >= X ) ? pc->jt : pc->jf */
949 condt = MIPS_COND_GE | MIPS_COND_X;
950 goto jmp_cmp;
951 case BPF_JMP | BPF_JGT | BPF_K:
952 /* pc += ( A > K ) ? pc->jt : pc->jf */
953 condt = MIPS_COND_GT | MIPS_COND_K;
954 goto jmp_cmp;
955 case BPF_JMP | BPF_JGT | BPF_X:
956 ctx->flags |= SEEN_X;
957 /* pc += ( A > X ) ? pc->jt : pc->jf */
958 condt = MIPS_COND_GT | MIPS_COND_X;
959 jmp_cmp:
960 /* Greater or Equal */
961 if ((condt & MIPS_COND_GE) ||
962 (condt & MIPS_COND_GT)) {
963 if (condt & MIPS_COND_K) { /* K */
964 ctx->flags |= SEEN_A;
965 emit_sltiu(r_s0, r_A, k, ctx);
966 } else { /* X */
967 ctx->flags |= SEEN_A |
968 SEEN_X;
969 emit_sltu(r_s0, r_A, r_X, ctx);
971 /* A < (K|X) ? r_scrach = 1 */
972 b_off = b_imm(i + inst->jf + 1, ctx);
973 emit_bcond(MIPS_COND_NE, r_s0, r_zero, b_off,
974 ctx);
975 emit_nop(ctx);
976 /* A > (K|X) ? scratch = 0 */
977 if (condt & MIPS_COND_GT) {
978 /* Checking for equality */
979 ctx->flags |= SEEN_A | SEEN_X;
980 if (condt & MIPS_COND_K)
981 emit_load_imm(r_s0, k, ctx);
982 else
983 emit_jit_reg_move(r_s0, r_X,
984 ctx);
985 b_off = b_imm(i + inst->jf + 1, ctx);
986 emit_bcond(MIPS_COND_EQ, r_A, r_s0,
987 b_off, ctx);
988 emit_nop(ctx);
989 /* Finally, A > K|X */
990 b_off = b_imm(i + inst->jt + 1, ctx);
991 emit_b(b_off, ctx);
992 emit_nop(ctx);
993 } else {
994 /* A >= (K|X) so jump */
995 b_off = b_imm(i + inst->jt + 1, ctx);
996 emit_b(b_off, ctx);
997 emit_nop(ctx);
999 } else {
1000 /* A == K|X */
1001 if (condt & MIPS_COND_K) { /* K */
1002 ctx->flags |= SEEN_A;
1003 emit_load_imm(r_s0, k, ctx);
1004 /* jump true */
1005 b_off = b_imm(i + inst->jt + 1, ctx);
1006 emit_bcond(MIPS_COND_EQ, r_A, r_s0,
1007 b_off, ctx);
1008 emit_nop(ctx);
1009 /* jump false */
1010 b_off = b_imm(i + inst->jf + 1,
1011 ctx);
1012 emit_bcond(MIPS_COND_NE, r_A, r_s0,
1013 b_off, ctx);
1014 emit_nop(ctx);
1015 } else { /* X */
1016 /* jump true */
1017 ctx->flags |= SEEN_A | SEEN_X;
1018 b_off = b_imm(i + inst->jt + 1,
1019 ctx);
1020 emit_bcond(MIPS_COND_EQ, r_A, r_X,
1021 b_off, ctx);
1022 emit_nop(ctx);
1023 /* jump false */
1024 b_off = b_imm(i + inst->jf + 1, ctx);
1025 emit_bcond(MIPS_COND_NE, r_A, r_X,
1026 b_off, ctx);
1027 emit_nop(ctx);
1030 break;
1031 case BPF_JMP | BPF_JSET | BPF_K:
1032 ctx->flags |= SEEN_A;
1033 /* pc += (A & K) ? pc -> jt : pc -> jf */
1034 emit_load_imm(r_s1, k, ctx);
1035 emit_and(r_s0, r_A, r_s1, ctx);
1036 /* jump true */
1037 b_off = b_imm(i + inst->jt + 1, ctx);
1038 emit_bcond(MIPS_COND_NE, r_s0, r_zero, b_off, ctx);
1039 emit_nop(ctx);
1040 /* jump false */
1041 b_off = b_imm(i + inst->jf + 1, ctx);
1042 emit_b(b_off, ctx);
1043 emit_nop(ctx);
1044 break;
1045 case BPF_JMP | BPF_JSET | BPF_X:
1046 ctx->flags |= SEEN_X | SEEN_A;
1047 /* pc += (A & X) ? pc -> jt : pc -> jf */
1048 emit_and(r_s0, r_A, r_X, ctx);
1049 /* jump true */
1050 b_off = b_imm(i + inst->jt + 1, ctx);
1051 emit_bcond(MIPS_COND_NE, r_s0, r_zero, b_off, ctx);
1052 emit_nop(ctx);
1053 /* jump false */
1054 b_off = b_imm(i + inst->jf + 1, ctx);
1055 emit_b(b_off, ctx);
1056 emit_nop(ctx);
1057 break;
1058 case BPF_RET | BPF_A:
1059 ctx->flags |= SEEN_A;
1060 if (i != prog->len - 1)
1062 * If this is not the last instruction
1063 * then jump to the epilogue
1065 emit_b(b_imm(prog->len, ctx), ctx);
1066 emit_reg_move(r_ret, r_A, ctx); /* delay slot */
1067 break;
1068 case BPF_RET | BPF_K:
1070 * It can emit two instructions so it does not fit on
1071 * the delay slot.
1073 emit_load_imm(r_ret, k, ctx);
1074 if (i != prog->len - 1) {
1076 * If this is not the last instruction
1077 * then jump to the epilogue
1079 emit_b(b_imm(prog->len, ctx), ctx);
1080 emit_nop(ctx);
1082 break;
1083 case BPF_MISC | BPF_TAX:
1084 /* X = A */
1085 ctx->flags |= SEEN_X | SEEN_A;
1086 emit_jit_reg_move(r_X, r_A, ctx);
1087 break;
1088 case BPF_MISC | BPF_TXA:
1089 /* A = X */
1090 ctx->flags |= SEEN_A | SEEN_X;
1091 emit_jit_reg_move(r_A, r_X, ctx);
1092 break;
1093 /* AUX */
1094 case BPF_ANC | SKF_AD_PROTOCOL:
1095 /* A = ntohs(skb->protocol */
1096 ctx->flags |= SEEN_SKB | SEEN_OFF | SEEN_A;
1097 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
1098 protocol) != 2);
1099 off = offsetof(struct sk_buff, protocol);
1100 emit_half_load(r_A, r_skb, off, ctx);
1101 #ifdef CONFIG_CPU_LITTLE_ENDIAN
1102 /* This needs little endian fixup */
1103 if (cpu_has_wsbh) {
1104 /* R2 and later have the wsbh instruction */
1105 emit_wsbh(r_A, r_A, ctx);
1106 } else {
1107 /* Get first byte */
1108 emit_andi(r_tmp_imm, r_A, 0xff, ctx);
1109 /* Shift it */
1110 emit_sll(r_tmp, r_tmp_imm, 8, ctx);
1111 /* Get second byte */
1112 emit_srl(r_tmp_imm, r_A, 8, ctx);
1113 emit_andi(r_tmp_imm, r_tmp_imm, 0xff, ctx);
1114 /* Put everyting together in r_A */
1115 emit_or(r_A, r_tmp, r_tmp_imm, ctx);
1117 #endif
1118 break;
1119 case BPF_ANC | SKF_AD_CPU:
1120 ctx->flags |= SEEN_A | SEEN_OFF;
1121 /* A = current_thread_info()->cpu */
1122 BUILD_BUG_ON(FIELD_SIZEOF(struct thread_info,
1123 cpu) != 4);
1124 off = offsetof(struct thread_info, cpu);
1125 /* $28/gp points to the thread_info struct */
1126 emit_load(r_A, 28, off, ctx);
1127 break;
1128 case BPF_ANC | SKF_AD_IFINDEX:
1129 /* A = skb->dev->ifindex */
1130 ctx->flags |= SEEN_SKB | SEEN_A;
1131 off = offsetof(struct sk_buff, dev);
1132 /* Load *dev pointer */
1133 emit_load_ptr(r_s0, r_skb, off, ctx);
1134 /* error (0) in the delay slot */
1135 emit_bcond(MIPS_COND_EQ, r_s0, r_zero,
1136 b_imm(prog->len, ctx), ctx);
1137 emit_reg_move(r_ret, r_zero, ctx);
1138 BUILD_BUG_ON(FIELD_SIZEOF(struct net_device,
1139 ifindex) != 4);
1140 off = offsetof(struct net_device, ifindex);
1141 emit_load(r_A, r_s0, off, ctx);
1142 break;
1143 case BPF_ANC | SKF_AD_MARK:
1144 ctx->flags |= SEEN_SKB | SEEN_A;
1145 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, mark) != 4);
1146 off = offsetof(struct sk_buff, mark);
1147 emit_load(r_A, r_skb, off, ctx);
1148 break;
1149 case BPF_ANC | SKF_AD_RXHASH:
1150 ctx->flags |= SEEN_SKB | SEEN_A;
1151 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff, hash) != 4);
1152 off = offsetof(struct sk_buff, hash);
1153 emit_load(r_A, r_skb, off, ctx);
1154 break;
1155 case BPF_ANC | SKF_AD_VLAN_TAG:
1156 case BPF_ANC | SKF_AD_VLAN_TAG_PRESENT:
1157 ctx->flags |= SEEN_SKB | SEEN_A;
1158 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
1159 vlan_tci) != 2);
1160 off = offsetof(struct sk_buff, vlan_tci);
1161 emit_half_load(r_s0, r_skb, off, ctx);
1162 if (code == (BPF_ANC | SKF_AD_VLAN_TAG)) {
1163 emit_andi(r_A, r_s0, (u16)~VLAN_TAG_PRESENT, ctx);
1164 } else {
1165 emit_andi(r_A, r_s0, VLAN_TAG_PRESENT, ctx);
1166 /* return 1 if present */
1167 emit_sltu(r_A, r_zero, r_A, ctx);
1169 break;
1170 case BPF_ANC | SKF_AD_PKTTYPE:
1171 ctx->flags |= SEEN_SKB;
1173 emit_load_byte(r_tmp, r_skb, PKT_TYPE_OFFSET(), ctx);
1174 /* Keep only the last 3 bits */
1175 emit_andi(r_A, r_tmp, PKT_TYPE_MAX, ctx);
1176 #ifdef __BIG_ENDIAN_BITFIELD
1177 /* Get the actual packet type to the lower 3 bits */
1178 emit_srl(r_A, r_A, 5, ctx);
1179 #endif
1180 break;
1181 case BPF_ANC | SKF_AD_QUEUE:
1182 ctx->flags |= SEEN_SKB | SEEN_A;
1183 BUILD_BUG_ON(FIELD_SIZEOF(struct sk_buff,
1184 queue_mapping) != 2);
1185 BUILD_BUG_ON(offsetof(struct sk_buff,
1186 queue_mapping) > 0xff);
1187 off = offsetof(struct sk_buff, queue_mapping);
1188 emit_half_load(r_A, r_skb, off, ctx);
1189 break;
1190 default:
1191 pr_debug("%s: Unhandled opcode: 0x%02x\n", __FILE__,
1192 inst->code);
1193 return -1;
1197 /* compute offsets only during the first pass */
1198 if (ctx->target == NULL)
1199 ctx->offsets[i] = ctx->idx * 4;
1201 return 0;
1204 int bpf_jit_enable __read_mostly;
1206 void bpf_jit_compile(struct bpf_prog *fp)
1208 struct jit_ctx ctx;
1209 unsigned int alloc_size, tmp_idx;
1211 if (!bpf_jit_enable)
1212 return;
1214 memset(&ctx, 0, sizeof(ctx));
1216 ctx.offsets = kcalloc(fp->len, sizeof(*ctx.offsets), GFP_KERNEL);
1217 if (ctx.offsets == NULL)
1218 return;
1220 ctx.skf = fp;
1222 if (build_body(&ctx))
1223 goto out;
1225 tmp_idx = ctx.idx;
1226 build_prologue(&ctx);
1227 ctx.prologue_bytes = (ctx.idx - tmp_idx) * 4;
1228 /* just to complete the ctx.idx count */
1229 build_epilogue(&ctx);
1231 alloc_size = 4 * ctx.idx;
1232 ctx.target = module_alloc(alloc_size);
1233 if (ctx.target == NULL)
1234 goto out;
1236 /* Clean it */
1237 memset(ctx.target, 0, alloc_size);
1239 ctx.idx = 0;
1241 /* Generate the actual JIT code */
1242 build_prologue(&ctx);
1243 build_body(&ctx);
1244 build_epilogue(&ctx);
1246 /* Update the icache */
1247 flush_icache_range((ptr)ctx.target, (ptr)(ctx.target + ctx.idx));
1249 if (bpf_jit_enable > 1)
1250 /* Dump JIT code */
1251 bpf_jit_dump(fp->len, alloc_size, 2, ctx.target);
1253 fp->bpf_func = (void *)ctx.target;
1254 fp->jited = 1;
1256 out:
1257 kfree(ctx.offsets);
1260 void bpf_jit_free(struct bpf_prog *fp)
1262 if (fp->jited)
1263 module_memfree(fp->bpf_func);
1265 bpf_prog_unlock_free(fp);