1 // SPDX-License-Identifier: GPL-2.0-only
3 * Just-In-Time compiler for eBPF filters on MIPS
5 * Copyright (c) 2017 Cavium, Inc.
9 * Copyright (c) 2014 Imagination Technologies Ltd.
10 * Author: Markos Chandras <markos.chandras@imgtec.com>
13 #include <linux/bitops.h>
14 #include <linux/errno.h>
15 #include <linux/filter.h>
16 #include <linux/bpf.h>
17 #include <linux/slab.h>
18 #include <asm/bitops.h>
19 #include <asm/byteorder.h>
20 #include <asm/cacheflush.h>
21 #include <asm/cpu-features.h>
22 #include <asm/isa-rev.h>
25 /* Registers used by JIT */
28 #define MIPS_R_V0 2 /* BPF_R0 */
30 #define MIPS_R_A0 4 /* BPF_R1 */
31 #define MIPS_R_A1 5 /* BPF_R2 */
32 #define MIPS_R_A2 6 /* BPF_R3 */
33 #define MIPS_R_A3 7 /* BPF_R4 */
34 #define MIPS_R_A4 8 /* BPF_R5 */
35 #define MIPS_R_T4 12 /* BPF_AX */
39 #define MIPS_R_S0 16 /* BPF_R6 */
40 #define MIPS_R_S1 17 /* BPF_R7 */
41 #define MIPS_R_S2 18 /* BPF_R8 */
42 #define MIPS_R_S3 19 /* BPF_R9 */
43 #define MIPS_R_S4 20 /* BPF_TCC */
53 #define EBPF_SAVE_S0 BIT(0)
54 #define EBPF_SAVE_S1 BIT(1)
55 #define EBPF_SAVE_S2 BIT(2)
56 #define EBPF_SAVE_S3 BIT(3)
57 #define EBPF_SAVE_S4 BIT(4)
58 #define EBPF_SAVE_RA BIT(5)
59 #define EBPF_SEEN_FP BIT(6)
60 #define EBPF_SEEN_TC BIT(7)
61 #define EBPF_TCC_IN_V1 BIT(8)
64 * For the mips64 ISA, we need to track the value range or type for
65 * each JIT register. The BPF machine requires zero extended 32-bit
66 * values, but the mips64 ISA requires sign extended 32-bit values.
67 * At each point in the BPF program we track the state of every
68 * register so that we can zero extend or sign extend as the BPF
74 /* not known to be 32-bit compatible. */
76 /* 32-bit compatible, no truncation needed for 64-bit ops. */
78 /* 32-bit compatible, need truncation for 64-bit ops. */
80 /* 32-bit no sign/zero extension needed. */
85 * high bit of offsets indicates if long branch conversion done at
88 #define OFFSETS_B_CONV BIT(31)
91 * struct jit_ctx - JIT context
93 * @stack_size: eBPF stack size
94 * @idx: Instruction index
96 * @offsets: Instruction offsets
97 * @target: Memory location for the compiled filter
98 * @reg_val_types Packed enum reg_val_type for each register.
101 const struct bpf_prog *skf;
108 unsigned int long_b_conversion:1;
109 unsigned int gen_b_offsets:1;
110 unsigned int use_bbit_insns:1;
113 static void set_reg_val_type(u64 *rvt, int reg, enum reg_val_type type)
115 *rvt &= ~(7ull << (reg * 3));
116 *rvt |= ((u64)type << (reg * 3));
119 static enum reg_val_type get_reg_val_type(const struct jit_ctx *ctx,
122 return (ctx->reg_val_types[index] >> (reg * 3)) & 7;
125 /* Simply emit the instruction if the JIT memory space has been allocated */
126 #define emit_instr_long(ctx, func64, func32, ...) \
128 if ((ctx)->target != NULL) { \
129 u32 *p = &(ctx)->target[ctx->idx]; \
130 if (IS_ENABLED(CONFIG_64BIT)) \
131 uasm_i_##func64(&p, ##__VA_ARGS__); \
133 uasm_i_##func32(&p, ##__VA_ARGS__); \
138 #define emit_instr(ctx, func, ...) \
139 emit_instr_long(ctx, func, func, ##__VA_ARGS__)
141 static unsigned int j_target(struct jit_ctx *ctx, int target_idx)
143 unsigned long target_va, base_va;
149 base_va = (unsigned long)ctx->target;
150 target_va = base_va + (ctx->offsets[target_idx] & ~OFFSETS_B_CONV);
152 if ((base_va & ~0x0ffffffful) != (target_va & ~0x0ffffffful))
153 return (unsigned int)-1;
154 r = target_va & 0x0ffffffful;
158 /* Compute the immediate value for PC-relative branches. */
159 static u32 b_imm(unsigned int tgt, struct jit_ctx *ctx)
161 if (!ctx->gen_b_offsets)
165 * We want a pc-relative branch. tgt is the instruction offset
166 * we want to jump to.
169 * I: target_offset <- sign_extend(offset)
170 * I+1: PC += target_offset (delay slot)
172 * ctx->idx currently points to the branch instruction
173 * but the offset is added to the delay slot so we need
176 return (ctx->offsets[tgt] & ~OFFSETS_B_CONV) -
180 enum which_ebpf_reg {
188 * For eBPF, the register mapping naturally falls out of the
189 * requirements of eBPF and the MIPS n64 ABI. We don't maintain a
190 * separate frame pointer, so BPF_REG_10 relative accesses are
191 * adjusted to be $sp relative.
193 static int ebpf_to_mips_reg(struct jit_ctx *ctx,
194 const struct bpf_insn *insn,
195 enum which_ebpf_reg w)
197 int ebpf_reg = (w == src_reg || w == src_reg_no_fp) ?
198 insn->src_reg : insn->dst_reg;
214 ctx->flags |= EBPF_SAVE_S0;
217 ctx->flags |= EBPF_SAVE_S1;
220 ctx->flags |= EBPF_SAVE_S2;
223 ctx->flags |= EBPF_SAVE_S3;
226 if (w == dst_reg || w == src_reg_no_fp)
228 ctx->flags |= EBPF_SEEN_FP;
230 * Needs special handling, return something that
231 * cannot be clobbered just in case.
238 WARN(1, "Illegal bpf reg: %d\n", ebpf_reg);
243 * eBPF stack frame will be something like:
245 * Entry $sp ------> +--------------------------------+
247 * +--------------------------------+
249 * +--------------------------------+
251 * +--------------------------------+
253 * +--------------------------------+
255 * +--------------------------------+
257 * +--------------------------------+
258 * | tmp-storage (if $ra saved) |
259 * $sp + tmp_offset --> +--------------------------------+ <--BPF_REG_10
260 * | BPF_REG_10 relative storage |
261 * | MAX_BPF_STACK (optional) |
265 * $sp --------> +--------------------------------+
267 * If BPF_REG_10 is never referenced, then the MAX_BPF_STACK sized
268 * area is not allocated.
270 static int gen_int_prologue(struct jit_ctx *ctx)
272 int stack_adjust = 0;
276 if (ctx->flags & EBPF_SAVE_RA)
278 * If RA we are doing a function call and may need
279 * extra 8-byte tmp area.
281 stack_adjust += 2 * sizeof(long);
282 if (ctx->flags & EBPF_SAVE_S0)
283 stack_adjust += sizeof(long);
284 if (ctx->flags & EBPF_SAVE_S1)
285 stack_adjust += sizeof(long);
286 if (ctx->flags & EBPF_SAVE_S2)
287 stack_adjust += sizeof(long);
288 if (ctx->flags & EBPF_SAVE_S3)
289 stack_adjust += sizeof(long);
290 if (ctx->flags & EBPF_SAVE_S4)
291 stack_adjust += sizeof(long);
293 BUILD_BUG_ON(MAX_BPF_STACK & 7);
294 locals_size = (ctx->flags & EBPF_SEEN_FP) ? MAX_BPF_STACK : 0;
296 stack_adjust += locals_size;
298 ctx->stack_size = stack_adjust;
301 * First instruction initializes the tail call count (TCC).
302 * On tail call we skip this instruction, and the TCC is
303 * passed in $v1 from the caller.
305 emit_instr(ctx, addiu, MIPS_R_V1, MIPS_R_ZERO, MAX_TAIL_CALL_CNT);
307 emit_instr_long(ctx, daddiu, addiu,
308 MIPS_R_SP, MIPS_R_SP, -stack_adjust);
312 store_offset = stack_adjust - sizeof(long);
314 if (ctx->flags & EBPF_SAVE_RA) {
315 emit_instr_long(ctx, sd, sw,
316 MIPS_R_RA, store_offset, MIPS_R_SP);
317 store_offset -= sizeof(long);
319 if (ctx->flags & EBPF_SAVE_S0) {
320 emit_instr_long(ctx, sd, sw,
321 MIPS_R_S0, store_offset, MIPS_R_SP);
322 store_offset -= sizeof(long);
324 if (ctx->flags & EBPF_SAVE_S1) {
325 emit_instr_long(ctx, sd, sw,
326 MIPS_R_S1, store_offset, MIPS_R_SP);
327 store_offset -= sizeof(long);
329 if (ctx->flags & EBPF_SAVE_S2) {
330 emit_instr_long(ctx, sd, sw,
331 MIPS_R_S2, store_offset, MIPS_R_SP);
332 store_offset -= sizeof(long);
334 if (ctx->flags & EBPF_SAVE_S3) {
335 emit_instr_long(ctx, sd, sw,
336 MIPS_R_S3, store_offset, MIPS_R_SP);
337 store_offset -= sizeof(long);
339 if (ctx->flags & EBPF_SAVE_S4) {
340 emit_instr_long(ctx, sd, sw,
341 MIPS_R_S4, store_offset, MIPS_R_SP);
342 store_offset -= sizeof(long);
345 if ((ctx->flags & EBPF_SEEN_TC) && !(ctx->flags & EBPF_TCC_IN_V1))
346 emit_instr_long(ctx, daddu, addu,
347 MIPS_R_S4, MIPS_R_V1, MIPS_R_ZERO);
352 static int build_int_epilogue(struct jit_ctx *ctx, int dest_reg)
354 const struct bpf_prog *prog = ctx->skf;
355 int stack_adjust = ctx->stack_size;
356 int store_offset = stack_adjust - sizeof(long);
357 enum reg_val_type td;
360 if (dest_reg == MIPS_R_RA) {
361 /* Don't let zero extended value escape. */
362 td = get_reg_val_type(ctx, prog->len, BPF_REG_0);
364 emit_instr(ctx, sll, r0, r0, 0);
367 if (ctx->flags & EBPF_SAVE_RA) {
368 emit_instr_long(ctx, ld, lw,
369 MIPS_R_RA, store_offset, MIPS_R_SP);
370 store_offset -= sizeof(long);
372 if (ctx->flags & EBPF_SAVE_S0) {
373 emit_instr_long(ctx, ld, lw,
374 MIPS_R_S0, store_offset, MIPS_R_SP);
375 store_offset -= sizeof(long);
377 if (ctx->flags & EBPF_SAVE_S1) {
378 emit_instr_long(ctx, ld, lw,
379 MIPS_R_S1, store_offset, MIPS_R_SP);
380 store_offset -= sizeof(long);
382 if (ctx->flags & EBPF_SAVE_S2) {
383 emit_instr_long(ctx, ld, lw,
384 MIPS_R_S2, store_offset, MIPS_R_SP);
385 store_offset -= sizeof(long);
387 if (ctx->flags & EBPF_SAVE_S3) {
388 emit_instr_long(ctx, ld, lw,
389 MIPS_R_S3, store_offset, MIPS_R_SP);
390 store_offset -= sizeof(long);
392 if (ctx->flags & EBPF_SAVE_S4) {
393 emit_instr_long(ctx, ld, lw,
394 MIPS_R_S4, store_offset, MIPS_R_SP);
395 store_offset -= sizeof(long);
397 emit_instr(ctx, jr, dest_reg);
400 emit_instr_long(ctx, daddiu, addiu,
401 MIPS_R_SP, MIPS_R_SP, stack_adjust);
403 emit_instr(ctx, nop);
408 static void gen_imm_to_reg(const struct bpf_insn *insn, int reg,
411 if (insn->imm >= S16_MIN && insn->imm <= S16_MAX) {
412 emit_instr(ctx, addiu, reg, MIPS_R_ZERO, insn->imm);
414 int lower = (s16)(insn->imm & 0xffff);
415 int upper = insn->imm - lower;
417 emit_instr(ctx, lui, reg, upper >> 16);
418 emit_instr(ctx, addiu, reg, reg, lower);
422 static int gen_imm_insn(const struct bpf_insn *insn, struct jit_ctx *ctx,
425 int upper_bound, lower_bound;
426 int dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
431 switch (BPF_OP(insn->code)) {
434 upper_bound = S16_MAX;
435 lower_bound = S16_MIN;
438 upper_bound = -(int)S16_MIN;
439 lower_bound = -(int)S16_MAX;
444 upper_bound = 0xffff;
450 /* Shift amounts are truncated, no need for bounds */
451 upper_bound = S32_MAX;
452 lower_bound = S32_MIN;
459 * Immediate move clobbers the register, so no sign/zero
462 if (BPF_CLASS(insn->code) == BPF_ALU64 &&
463 BPF_OP(insn->code) != BPF_MOV &&
464 get_reg_val_type(ctx, idx, insn->dst_reg) == REG_32BIT)
465 emit_instr(ctx, dinsu, dst, MIPS_R_ZERO, 32, 32);
466 /* BPF_ALU | BPF_LSH doesn't need separate sign extension */
467 if (BPF_CLASS(insn->code) == BPF_ALU &&
468 BPF_OP(insn->code) != BPF_LSH &&
469 BPF_OP(insn->code) != BPF_MOV &&
470 get_reg_val_type(ctx, idx, insn->dst_reg) != REG_32BIT)
471 emit_instr(ctx, sll, dst, dst, 0);
473 if (insn->imm >= lower_bound && insn->imm <= upper_bound) {
474 /* single insn immediate case */
475 switch (BPF_OP(insn->code) | BPF_CLASS(insn->code)) {
476 case BPF_ALU64 | BPF_MOV:
477 emit_instr(ctx, daddiu, dst, MIPS_R_ZERO, insn->imm);
479 case BPF_ALU64 | BPF_AND:
480 case BPF_ALU | BPF_AND:
481 emit_instr(ctx, andi, dst, dst, insn->imm);
483 case BPF_ALU64 | BPF_OR:
484 case BPF_ALU | BPF_OR:
485 emit_instr(ctx, ori, dst, dst, insn->imm);
487 case BPF_ALU64 | BPF_XOR:
488 case BPF_ALU | BPF_XOR:
489 emit_instr(ctx, xori, dst, dst, insn->imm);
491 case BPF_ALU64 | BPF_ADD:
492 emit_instr(ctx, daddiu, dst, dst, insn->imm);
494 case BPF_ALU64 | BPF_SUB:
495 emit_instr(ctx, daddiu, dst, dst, -insn->imm);
497 case BPF_ALU64 | BPF_RSH:
498 emit_instr(ctx, dsrl_safe, dst, dst, insn->imm & 0x3f);
500 case BPF_ALU | BPF_RSH:
501 emit_instr(ctx, srl, dst, dst, insn->imm & 0x1f);
503 case BPF_ALU64 | BPF_LSH:
504 emit_instr(ctx, dsll_safe, dst, dst, insn->imm & 0x3f);
506 case BPF_ALU | BPF_LSH:
507 emit_instr(ctx, sll, dst, dst, insn->imm & 0x1f);
509 case BPF_ALU64 | BPF_ARSH:
510 emit_instr(ctx, dsra_safe, dst, dst, insn->imm & 0x3f);
512 case BPF_ALU | BPF_ARSH:
513 emit_instr(ctx, sra, dst, dst, insn->imm & 0x1f);
515 case BPF_ALU | BPF_MOV:
516 emit_instr(ctx, addiu, dst, MIPS_R_ZERO, insn->imm);
518 case BPF_ALU | BPF_ADD:
519 emit_instr(ctx, addiu, dst, dst, insn->imm);
521 case BPF_ALU | BPF_SUB:
522 emit_instr(ctx, addiu, dst, dst, -insn->imm);
528 /* multi insn immediate case */
529 if (BPF_OP(insn->code) == BPF_MOV) {
530 gen_imm_to_reg(insn, dst, ctx);
532 gen_imm_to_reg(insn, MIPS_R_AT, ctx);
533 switch (BPF_OP(insn->code) | BPF_CLASS(insn->code)) {
534 case BPF_ALU64 | BPF_AND:
535 case BPF_ALU | BPF_AND:
536 emit_instr(ctx, and, dst, dst, MIPS_R_AT);
538 case BPF_ALU64 | BPF_OR:
539 case BPF_ALU | BPF_OR:
540 emit_instr(ctx, or, dst, dst, MIPS_R_AT);
542 case BPF_ALU64 | BPF_XOR:
543 case BPF_ALU | BPF_XOR:
544 emit_instr(ctx, xor, dst, dst, MIPS_R_AT);
546 case BPF_ALU64 | BPF_ADD:
547 emit_instr(ctx, daddu, dst, dst, MIPS_R_AT);
549 case BPF_ALU64 | BPF_SUB:
550 emit_instr(ctx, dsubu, dst, dst, MIPS_R_AT);
552 case BPF_ALU | BPF_ADD:
553 emit_instr(ctx, addu, dst, dst, MIPS_R_AT);
555 case BPF_ALU | BPF_SUB:
556 emit_instr(ctx, subu, dst, dst, MIPS_R_AT);
567 static void emit_const_to_reg(struct jit_ctx *ctx, int dst, u64 value)
569 if (value >= 0xffffffffffff8000ull || value < 0x8000ull) {
570 emit_instr(ctx, daddiu, dst, MIPS_R_ZERO, (int)value);
571 } else if (value >= 0xffffffff80000000ull ||
572 (value < 0x80000000 && value > 0xffff)) {
573 emit_instr(ctx, lui, dst, (s32)(s16)(value >> 16));
574 emit_instr(ctx, ori, dst, dst, (unsigned int)(value & 0xffff));
577 bool seen_part = false;
578 int needed_shift = 0;
580 for (i = 0; i < 4; i++) {
581 u64 part = (value >> (16 * (3 - i))) & 0xffff;
583 if (seen_part && needed_shift > 0 && (part || i == 3)) {
584 emit_instr(ctx, dsll_safe, dst, dst, needed_shift);
588 if (i == 0 || (!seen_part && i < 3 && part < 0x8000)) {
589 emit_instr(ctx, lui, dst, (s32)(s16)part);
592 emit_instr(ctx, ori, dst,
593 seen_part ? dst : MIPS_R_ZERO,
604 static int emit_bpf_tail_call(struct jit_ctx *ctx, int this_idx)
608 ctx->flags |= EBPF_SEEN_TC;
610 * if (index >= array->map.max_entries)
613 off = offsetof(struct bpf_array, map.max_entries);
614 emit_instr(ctx, lwu, MIPS_R_T5, off, MIPS_R_A1);
615 emit_instr(ctx, sltu, MIPS_R_AT, MIPS_R_T5, MIPS_R_A2);
616 b_off = b_imm(this_idx + 1, ctx);
617 emit_instr(ctx, bne, MIPS_R_AT, MIPS_R_ZERO, b_off);
623 emit_instr(ctx, daddiu, MIPS_R_T5,
624 (ctx->flags & EBPF_TCC_IN_V1) ? MIPS_R_V1 : MIPS_R_S4, -1);
625 b_off = b_imm(this_idx + 1, ctx);
626 emit_instr(ctx, bltz, MIPS_R_T5, b_off);
628 * prog = array->ptrs[index];
633 emit_instr(ctx, dsll, MIPS_R_T8, MIPS_R_A2, 3);
634 emit_instr(ctx, daddu, MIPS_R_T8, MIPS_R_T8, MIPS_R_A1);
635 off = offsetof(struct bpf_array, ptrs);
636 emit_instr(ctx, ld, MIPS_R_AT, off, MIPS_R_T8);
637 b_off = b_imm(this_idx + 1, ctx);
638 emit_instr(ctx, beq, MIPS_R_AT, MIPS_R_ZERO, b_off);
640 emit_instr(ctx, nop);
642 /* goto *(prog->bpf_func + 4); */
643 off = offsetof(struct bpf_prog, bpf_func);
644 emit_instr(ctx, ld, MIPS_R_T9, off, MIPS_R_AT);
645 /* All systems are go... propagate TCC */
646 emit_instr(ctx, daddu, MIPS_R_V1, MIPS_R_T5, MIPS_R_ZERO);
647 /* Skip first instruction (TCC initialization) */
648 emit_instr(ctx, daddiu, MIPS_R_T9, MIPS_R_T9, 4);
649 return build_int_epilogue(ctx, MIPS_R_T9);
652 static bool is_bad_offset(int b_off)
654 return b_off > 0x1ffff || b_off < -0x20000;
657 /* Returns the number of insn slots consumed. */
658 static int build_one_insn(const struct bpf_insn *insn, struct jit_ctx *ctx,
659 int this_idx, int exit_idx)
661 int src, dst, r, td, ts, mem_off, b_off;
662 bool need_swap, did_move, cmp_eq;
663 unsigned int target = 0;
666 int bpf_op = BPF_OP(insn->code);
668 if (IS_ENABLED(CONFIG_32BIT) && ((BPF_CLASS(insn->code) == BPF_ALU64)
669 || (bpf_op == BPF_DW)))
672 switch (insn->code) {
673 case BPF_ALU64 | BPF_ADD | BPF_K: /* ALU64_IMM */
674 case BPF_ALU64 | BPF_SUB | BPF_K: /* ALU64_IMM */
675 case BPF_ALU64 | BPF_OR | BPF_K: /* ALU64_IMM */
676 case BPF_ALU64 | BPF_AND | BPF_K: /* ALU64_IMM */
677 case BPF_ALU64 | BPF_LSH | BPF_K: /* ALU64_IMM */
678 case BPF_ALU64 | BPF_RSH | BPF_K: /* ALU64_IMM */
679 case BPF_ALU64 | BPF_XOR | BPF_K: /* ALU64_IMM */
680 case BPF_ALU64 | BPF_ARSH | BPF_K: /* ALU64_IMM */
681 case BPF_ALU64 | BPF_MOV | BPF_K: /* ALU64_IMM */
682 case BPF_ALU | BPF_MOV | BPF_K: /* ALU32_IMM */
683 case BPF_ALU | BPF_ADD | BPF_K: /* ALU32_IMM */
684 case BPF_ALU | BPF_SUB | BPF_K: /* ALU32_IMM */
685 case BPF_ALU | BPF_OR | BPF_K: /* ALU64_IMM */
686 case BPF_ALU | BPF_AND | BPF_K: /* ALU64_IMM */
687 case BPF_ALU | BPF_LSH | BPF_K: /* ALU64_IMM */
688 case BPF_ALU | BPF_RSH | BPF_K: /* ALU64_IMM */
689 case BPF_ALU | BPF_XOR | BPF_K: /* ALU64_IMM */
690 case BPF_ALU | BPF_ARSH | BPF_K: /* ALU64_IMM */
691 r = gen_imm_insn(insn, ctx, this_idx);
695 case BPF_ALU64 | BPF_MUL | BPF_K: /* ALU64_IMM */
696 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
699 if (get_reg_val_type(ctx, this_idx, insn->dst_reg) == REG_32BIT)
700 emit_instr(ctx, dinsu, dst, MIPS_R_ZERO, 32, 32);
701 if (insn->imm == 1) /* Mult by 1 is a nop */
703 gen_imm_to_reg(insn, MIPS_R_AT, ctx);
704 if (MIPS_ISA_REV >= 6) {
705 emit_instr(ctx, dmulu, dst, dst, MIPS_R_AT);
707 emit_instr(ctx, dmultu, MIPS_R_AT, dst);
708 emit_instr(ctx, mflo, dst);
711 case BPF_ALU64 | BPF_NEG | BPF_K: /* ALU64_IMM */
712 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
715 if (get_reg_val_type(ctx, this_idx, insn->dst_reg) == REG_32BIT)
716 emit_instr(ctx, dinsu, dst, MIPS_R_ZERO, 32, 32);
717 emit_instr(ctx, dsubu, dst, MIPS_R_ZERO, dst);
719 case BPF_ALU | BPF_MUL | BPF_K: /* ALU_IMM */
720 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
723 td = get_reg_val_type(ctx, this_idx, insn->dst_reg);
724 if (td == REG_64BIT) {
726 emit_instr(ctx, sll, dst, dst, 0);
728 if (insn->imm == 1) /* Mult by 1 is a nop */
730 gen_imm_to_reg(insn, MIPS_R_AT, ctx);
731 if (MIPS_ISA_REV >= 6) {
732 emit_instr(ctx, mulu, dst, dst, MIPS_R_AT);
734 emit_instr(ctx, multu, dst, MIPS_R_AT);
735 emit_instr(ctx, mflo, dst);
738 case BPF_ALU | BPF_NEG | BPF_K: /* ALU_IMM */
739 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
742 td = get_reg_val_type(ctx, this_idx, insn->dst_reg);
743 if (td == REG_64BIT) {
745 emit_instr(ctx, sll, dst, dst, 0);
747 emit_instr(ctx, subu, dst, MIPS_R_ZERO, dst);
749 case BPF_ALU | BPF_DIV | BPF_K: /* ALU_IMM */
750 case BPF_ALU | BPF_MOD | BPF_K: /* ALU_IMM */
753 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
756 td = get_reg_val_type(ctx, this_idx, insn->dst_reg);
759 emit_instr(ctx, sll, dst, dst, 0);
760 if (insn->imm == 1) {
761 /* div by 1 is a nop, mod by 1 is zero */
762 if (bpf_op == BPF_MOD)
763 emit_instr(ctx, addu, dst, MIPS_R_ZERO, MIPS_R_ZERO);
766 gen_imm_to_reg(insn, MIPS_R_AT, ctx);
767 if (MIPS_ISA_REV >= 6) {
768 if (bpf_op == BPF_DIV)
769 emit_instr(ctx, divu_r6, dst, dst, MIPS_R_AT);
771 emit_instr(ctx, modu, dst, dst, MIPS_R_AT);
774 emit_instr(ctx, divu, dst, MIPS_R_AT);
775 if (bpf_op == BPF_DIV)
776 emit_instr(ctx, mflo, dst);
778 emit_instr(ctx, mfhi, dst);
780 case BPF_ALU64 | BPF_DIV | BPF_K: /* ALU_IMM */
781 case BPF_ALU64 | BPF_MOD | BPF_K: /* ALU_IMM */
784 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
787 if (get_reg_val_type(ctx, this_idx, insn->dst_reg) == REG_32BIT)
788 emit_instr(ctx, dinsu, dst, MIPS_R_ZERO, 32, 32);
789 if (insn->imm == 1) {
790 /* div by 1 is a nop, mod by 1 is zero */
791 if (bpf_op == BPF_MOD)
792 emit_instr(ctx, addu, dst, MIPS_R_ZERO, MIPS_R_ZERO);
795 gen_imm_to_reg(insn, MIPS_R_AT, ctx);
796 if (MIPS_ISA_REV >= 6) {
797 if (bpf_op == BPF_DIV)
798 emit_instr(ctx, ddivu_r6, dst, dst, MIPS_R_AT);
800 emit_instr(ctx, modu, dst, dst, MIPS_R_AT);
803 emit_instr(ctx, ddivu, dst, MIPS_R_AT);
804 if (bpf_op == BPF_DIV)
805 emit_instr(ctx, mflo, dst);
807 emit_instr(ctx, mfhi, dst);
809 case BPF_ALU64 | BPF_MOV | BPF_X: /* ALU64_REG */
810 case BPF_ALU64 | BPF_ADD | BPF_X: /* ALU64_REG */
811 case BPF_ALU64 | BPF_SUB | BPF_X: /* ALU64_REG */
812 case BPF_ALU64 | BPF_XOR | BPF_X: /* ALU64_REG */
813 case BPF_ALU64 | BPF_OR | BPF_X: /* ALU64_REG */
814 case BPF_ALU64 | BPF_AND | BPF_X: /* ALU64_REG */
815 case BPF_ALU64 | BPF_MUL | BPF_X: /* ALU64_REG */
816 case BPF_ALU64 | BPF_DIV | BPF_X: /* ALU64_REG */
817 case BPF_ALU64 | BPF_MOD | BPF_X: /* ALU64_REG */
818 case BPF_ALU64 | BPF_LSH | BPF_X: /* ALU64_REG */
819 case BPF_ALU64 | BPF_RSH | BPF_X: /* ALU64_REG */
820 case BPF_ALU64 | BPF_ARSH | BPF_X: /* ALU64_REG */
821 src = ebpf_to_mips_reg(ctx, insn, src_reg);
822 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
823 if (src < 0 || dst < 0)
825 if (get_reg_val_type(ctx, this_idx, insn->dst_reg) == REG_32BIT)
826 emit_instr(ctx, dinsu, dst, MIPS_R_ZERO, 32, 32);
828 if (insn->src_reg == BPF_REG_10) {
829 if (bpf_op == BPF_MOV) {
830 emit_instr(ctx, daddiu, dst, MIPS_R_SP, MAX_BPF_STACK);
833 emit_instr(ctx, daddiu, MIPS_R_AT, MIPS_R_SP, MAX_BPF_STACK);
836 } else if (get_reg_val_type(ctx, this_idx, insn->src_reg) == REG_32BIT) {
837 int tmp_reg = MIPS_R_AT;
839 if (bpf_op == BPF_MOV) {
843 emit_instr(ctx, daddu, tmp_reg, src, MIPS_R_ZERO);
844 emit_instr(ctx, dinsu, tmp_reg, MIPS_R_ZERO, 32, 32);
850 emit_instr(ctx, daddu, dst, src, MIPS_R_ZERO);
853 emit_instr(ctx, daddu, dst, dst, src);
856 emit_instr(ctx, dsubu, dst, dst, src);
859 emit_instr(ctx, xor, dst, dst, src);
862 emit_instr(ctx, or, dst, dst, src);
865 emit_instr(ctx, and, dst, dst, src);
868 if (MIPS_ISA_REV >= 6) {
869 emit_instr(ctx, dmulu, dst, dst, src);
871 emit_instr(ctx, dmultu, dst, src);
872 emit_instr(ctx, mflo, dst);
877 if (MIPS_ISA_REV >= 6) {
878 if (bpf_op == BPF_DIV)
879 emit_instr(ctx, ddivu_r6,
882 emit_instr(ctx, modu, dst, dst, src);
885 emit_instr(ctx, ddivu, dst, src);
886 if (bpf_op == BPF_DIV)
887 emit_instr(ctx, mflo, dst);
889 emit_instr(ctx, mfhi, dst);
892 emit_instr(ctx, dsllv, dst, dst, src);
895 emit_instr(ctx, dsrlv, dst, dst, src);
898 emit_instr(ctx, dsrav, dst, dst, src);
901 pr_err("ALU64_REG NOT HANDLED\n");
905 case BPF_ALU | BPF_MOV | BPF_X: /* ALU_REG */
906 case BPF_ALU | BPF_ADD | BPF_X: /* ALU_REG */
907 case BPF_ALU | BPF_SUB | BPF_X: /* ALU_REG */
908 case BPF_ALU | BPF_XOR | BPF_X: /* ALU_REG */
909 case BPF_ALU | BPF_OR | BPF_X: /* ALU_REG */
910 case BPF_ALU | BPF_AND | BPF_X: /* ALU_REG */
911 case BPF_ALU | BPF_MUL | BPF_X: /* ALU_REG */
912 case BPF_ALU | BPF_DIV | BPF_X: /* ALU_REG */
913 case BPF_ALU | BPF_MOD | BPF_X: /* ALU_REG */
914 case BPF_ALU | BPF_LSH | BPF_X: /* ALU_REG */
915 case BPF_ALU | BPF_RSH | BPF_X: /* ALU_REG */
916 case BPF_ALU | BPF_ARSH | BPF_X: /* ALU_REG */
917 src = ebpf_to_mips_reg(ctx, insn, src_reg_no_fp);
918 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
919 if (src < 0 || dst < 0)
921 td = get_reg_val_type(ctx, this_idx, insn->dst_reg);
922 if (td == REG_64BIT) {
924 emit_instr(ctx, sll, dst, dst, 0);
927 ts = get_reg_val_type(ctx, this_idx, insn->src_reg);
928 if (ts == REG_64BIT) {
929 int tmp_reg = MIPS_R_AT;
931 if (bpf_op == BPF_MOV) {
936 emit_instr(ctx, sll, tmp_reg, src, 0);
942 emit_instr(ctx, addu, dst, src, MIPS_R_ZERO);
945 emit_instr(ctx, addu, dst, dst, src);
948 emit_instr(ctx, subu, dst, dst, src);
951 emit_instr(ctx, xor, dst, dst, src);
954 emit_instr(ctx, or, dst, dst, src);
957 emit_instr(ctx, and, dst, dst, src);
960 emit_instr(ctx, mul, dst, dst, src);
964 if (MIPS_ISA_REV >= 6) {
965 if (bpf_op == BPF_DIV)
966 emit_instr(ctx, divu_r6, dst, dst, src);
968 emit_instr(ctx, modu, dst, dst, src);
971 emit_instr(ctx, divu, dst, src);
972 if (bpf_op == BPF_DIV)
973 emit_instr(ctx, mflo, dst);
975 emit_instr(ctx, mfhi, dst);
978 emit_instr(ctx, sllv, dst, dst, src);
981 emit_instr(ctx, srlv, dst, dst, src);
984 emit_instr(ctx, srav, dst, dst, src);
987 pr_err("ALU_REG NOT HANDLED\n");
991 case BPF_JMP | BPF_EXIT:
992 if (this_idx + 1 < exit_idx) {
993 b_off = b_imm(exit_idx, ctx);
994 if (is_bad_offset(b_off))
996 emit_instr(ctx, beq, MIPS_R_ZERO, MIPS_R_ZERO, b_off);
997 emit_instr(ctx, nop);
1000 case BPF_JMP | BPF_JEQ | BPF_K: /* JMP_IMM */
1001 case BPF_JMP | BPF_JNE | BPF_K: /* JMP_IMM */
1002 cmp_eq = (bpf_op == BPF_JEQ);
1003 dst = ebpf_to_mips_reg(ctx, insn, dst_reg_fp_ok);
1006 if (insn->imm == 0) {
1009 gen_imm_to_reg(insn, MIPS_R_AT, ctx);
1013 case BPF_JMP | BPF_JEQ | BPF_X: /* JMP_REG */
1014 case BPF_JMP | BPF_JNE | BPF_X:
1015 case BPF_JMP | BPF_JSLT | BPF_X:
1016 case BPF_JMP | BPF_JSLE | BPF_X:
1017 case BPF_JMP | BPF_JSGT | BPF_X:
1018 case BPF_JMP | BPF_JSGE | BPF_X:
1019 case BPF_JMP | BPF_JLT | BPF_X:
1020 case BPF_JMP | BPF_JLE | BPF_X:
1021 case BPF_JMP | BPF_JGT | BPF_X:
1022 case BPF_JMP | BPF_JGE | BPF_X:
1023 case BPF_JMP | BPF_JSET | BPF_X:
1024 src = ebpf_to_mips_reg(ctx, insn, src_reg_no_fp);
1025 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
1026 if (src < 0 || dst < 0)
1028 td = get_reg_val_type(ctx, this_idx, insn->dst_reg);
1029 ts = get_reg_val_type(ctx, this_idx, insn->src_reg);
1030 if (td == REG_32BIT && ts != REG_32BIT) {
1031 emit_instr(ctx, sll, MIPS_R_AT, src, 0);
1033 } else if (ts == REG_32BIT && td != REG_32BIT) {
1034 emit_instr(ctx, sll, MIPS_R_AT, dst, 0);
1037 if (bpf_op == BPF_JSET) {
1038 emit_instr(ctx, and, MIPS_R_AT, dst, src);
1042 } else if (bpf_op == BPF_JSGT || bpf_op == BPF_JSLE) {
1043 emit_instr(ctx, dsubu, MIPS_R_AT, dst, src);
1044 if ((insn + 1)->code == (BPF_JMP | BPF_EXIT) && insn->off == 1) {
1045 b_off = b_imm(exit_idx, ctx);
1046 if (is_bad_offset(b_off))
1048 if (bpf_op == BPF_JSGT)
1049 emit_instr(ctx, blez, MIPS_R_AT, b_off);
1051 emit_instr(ctx, bgtz, MIPS_R_AT, b_off);
1052 emit_instr(ctx, nop);
1053 return 2; /* We consumed the exit. */
1055 b_off = b_imm(this_idx + insn->off + 1, ctx);
1056 if (is_bad_offset(b_off))
1058 if (bpf_op == BPF_JSGT)
1059 emit_instr(ctx, bgtz, MIPS_R_AT, b_off);
1061 emit_instr(ctx, blez, MIPS_R_AT, b_off);
1062 emit_instr(ctx, nop);
1064 } else if (bpf_op == BPF_JSGE || bpf_op == BPF_JSLT) {
1065 emit_instr(ctx, slt, MIPS_R_AT, dst, src);
1066 cmp_eq = bpf_op == BPF_JSGE;
1069 } else if (bpf_op == BPF_JGT || bpf_op == BPF_JLE) {
1070 /* dst or src could be AT */
1071 emit_instr(ctx, dsubu, MIPS_R_T8, dst, src);
1072 emit_instr(ctx, sltu, MIPS_R_AT, dst, src);
1073 /* SP known to be non-zero, movz becomes boolean not */
1074 if (MIPS_ISA_REV >= 6) {
1075 emit_instr(ctx, seleqz, MIPS_R_T9,
1076 MIPS_R_SP, MIPS_R_T8);
1078 emit_instr(ctx, movz, MIPS_R_T9,
1079 MIPS_R_SP, MIPS_R_T8);
1080 emit_instr(ctx, movn, MIPS_R_T9,
1081 MIPS_R_ZERO, MIPS_R_T8);
1083 emit_instr(ctx, or, MIPS_R_AT, MIPS_R_T9, MIPS_R_AT);
1084 cmp_eq = bpf_op == BPF_JGT;
1087 } else if (bpf_op == BPF_JGE || bpf_op == BPF_JLT) {
1088 emit_instr(ctx, sltu, MIPS_R_AT, dst, src);
1089 cmp_eq = bpf_op == BPF_JGE;
1092 } else { /* JNE/JEQ case */
1093 cmp_eq = (bpf_op == BPF_JEQ);
1097 * If the next insn is EXIT and we are jumping arround
1098 * only it, invert the sense of the compare and
1099 * conditionally jump to the exit. Poor man's branch
1102 if ((insn + 1)->code == (BPF_JMP | BPF_EXIT) && insn->off == 1) {
1103 b_off = b_imm(exit_idx, ctx);
1104 if (is_bad_offset(b_off)) {
1105 target = j_target(ctx, exit_idx);
1106 if (target == (unsigned int)-1)
1110 if (!(ctx->offsets[this_idx] & OFFSETS_B_CONV)) {
1111 ctx->offsets[this_idx] |= OFFSETS_B_CONV;
1112 ctx->long_b_conversion = 1;
1117 emit_instr(ctx, bne, dst, src, b_off);
1119 emit_instr(ctx, beq, dst, src, b_off);
1120 emit_instr(ctx, nop);
1121 if (ctx->offsets[this_idx] & OFFSETS_B_CONV) {
1122 emit_instr(ctx, j, target);
1123 emit_instr(ctx, nop);
1125 return 2; /* We consumed the exit. */
1127 b_off = b_imm(this_idx + insn->off + 1, ctx);
1128 if (is_bad_offset(b_off)) {
1129 target = j_target(ctx, this_idx + insn->off + 1);
1130 if (target == (unsigned int)-1)
1134 if (!(ctx->offsets[this_idx] & OFFSETS_B_CONV)) {
1135 ctx->offsets[this_idx] |= OFFSETS_B_CONV;
1136 ctx->long_b_conversion = 1;
1141 emit_instr(ctx, beq, dst, src, b_off);
1143 emit_instr(ctx, bne, dst, src, b_off);
1144 emit_instr(ctx, nop);
1145 if (ctx->offsets[this_idx] & OFFSETS_B_CONV) {
1146 emit_instr(ctx, j, target);
1147 emit_instr(ctx, nop);
1150 case BPF_JMP | BPF_JSGT | BPF_K: /* JMP_IMM */
1151 case BPF_JMP | BPF_JSGE | BPF_K: /* JMP_IMM */
1152 case BPF_JMP | BPF_JSLT | BPF_K: /* JMP_IMM */
1153 case BPF_JMP | BPF_JSLE | BPF_K: /* JMP_IMM */
1154 cmp_eq = (bpf_op == BPF_JSGE);
1155 dst = ebpf_to_mips_reg(ctx, insn, dst_reg_fp_ok);
1159 if (insn->imm == 0) {
1160 if ((insn + 1)->code == (BPF_JMP | BPF_EXIT) && insn->off == 1) {
1161 b_off = b_imm(exit_idx, ctx);
1162 if (is_bad_offset(b_off))
1166 emit_instr(ctx, blez, dst, b_off);
1169 emit_instr(ctx, bltz, dst, b_off);
1172 emit_instr(ctx, bgez, dst, b_off);
1175 emit_instr(ctx, bgtz, dst, b_off);
1178 emit_instr(ctx, nop);
1179 return 2; /* We consumed the exit. */
1181 b_off = b_imm(this_idx + insn->off + 1, ctx);
1182 if (is_bad_offset(b_off))
1186 emit_instr(ctx, bgtz, dst, b_off);
1189 emit_instr(ctx, bgez, dst, b_off);
1192 emit_instr(ctx, bltz, dst, b_off);
1195 emit_instr(ctx, blez, dst, b_off);
1198 emit_instr(ctx, nop);
1202 * only "LT" compare available, so we must use imm + 1
1203 * to generate "GT" and imm -1 to generate LE
1205 if (bpf_op == BPF_JSGT)
1206 t64s = insn->imm + 1;
1207 else if (bpf_op == BPF_JSLE)
1208 t64s = insn->imm + 1;
1212 cmp_eq = bpf_op == BPF_JSGT || bpf_op == BPF_JSGE;
1213 if (t64s >= S16_MIN && t64s <= S16_MAX) {
1214 emit_instr(ctx, slti, MIPS_R_AT, dst, (int)t64s);
1219 emit_const_to_reg(ctx, MIPS_R_AT, (u64)t64s);
1220 emit_instr(ctx, slt, MIPS_R_AT, dst, MIPS_R_AT);
1225 case BPF_JMP | BPF_JGT | BPF_K:
1226 case BPF_JMP | BPF_JGE | BPF_K:
1227 case BPF_JMP | BPF_JLT | BPF_K:
1228 case BPF_JMP | BPF_JLE | BPF_K:
1229 cmp_eq = (bpf_op == BPF_JGE);
1230 dst = ebpf_to_mips_reg(ctx, insn, dst_reg_fp_ok);
1234 * only "LT" compare available, so we must use imm + 1
1235 * to generate "GT" and imm -1 to generate LE
1237 if (bpf_op == BPF_JGT)
1238 t64s = (u64)(u32)(insn->imm) + 1;
1239 else if (bpf_op == BPF_JLE)
1240 t64s = (u64)(u32)(insn->imm) + 1;
1242 t64s = (u64)(u32)(insn->imm);
1244 cmp_eq = bpf_op == BPF_JGT || bpf_op == BPF_JGE;
1246 emit_const_to_reg(ctx, MIPS_R_AT, (u64)t64s);
1247 emit_instr(ctx, sltu, MIPS_R_AT, dst, MIPS_R_AT);
1252 case BPF_JMP | BPF_JSET | BPF_K: /* JMP_IMM */
1253 dst = ebpf_to_mips_reg(ctx, insn, dst_reg_fp_ok);
1257 if (ctx->use_bbit_insns && hweight32((u32)insn->imm) == 1) {
1258 if ((insn + 1)->code == (BPF_JMP | BPF_EXIT) && insn->off == 1) {
1259 b_off = b_imm(exit_idx, ctx);
1260 if (is_bad_offset(b_off))
1262 emit_instr(ctx, bbit0, dst, ffs((u32)insn->imm) - 1, b_off);
1263 emit_instr(ctx, nop);
1264 return 2; /* We consumed the exit. */
1266 b_off = b_imm(this_idx + insn->off + 1, ctx);
1267 if (is_bad_offset(b_off))
1269 emit_instr(ctx, bbit1, dst, ffs((u32)insn->imm) - 1, b_off);
1270 emit_instr(ctx, nop);
1273 t64 = (u32)insn->imm;
1274 emit_const_to_reg(ctx, MIPS_R_AT, t64);
1275 emit_instr(ctx, and, MIPS_R_AT, dst, MIPS_R_AT);
1281 case BPF_JMP | BPF_JA:
1283 * Prefer relative branch for easier debugging, but
1284 * fall back if needed.
1286 b_off = b_imm(this_idx + insn->off + 1, ctx);
1287 if (is_bad_offset(b_off)) {
1288 target = j_target(ctx, this_idx + insn->off + 1);
1289 if (target == (unsigned int)-1)
1291 emit_instr(ctx, j, target);
1293 emit_instr(ctx, b, b_off);
1295 emit_instr(ctx, nop);
1297 case BPF_LD | BPF_DW | BPF_IMM:
1298 if (insn->src_reg != 0)
1300 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
1303 t64 = ((u64)(u32)insn->imm) | ((u64)(insn + 1)->imm << 32);
1304 emit_const_to_reg(ctx, dst, t64);
1305 return 2; /* Double slot insn */
1307 case BPF_JMP | BPF_CALL:
1308 ctx->flags |= EBPF_SAVE_RA;
1309 t64s = (s64)insn->imm + (long)__bpf_call_base;
1310 emit_const_to_reg(ctx, MIPS_R_T9, (u64)t64s);
1311 emit_instr(ctx, jalr, MIPS_R_RA, MIPS_R_T9);
1313 emit_instr(ctx, nop);
1316 case BPF_JMP | BPF_TAIL_CALL:
1317 if (emit_bpf_tail_call(ctx, this_idx))
1321 case BPF_ALU | BPF_END | BPF_FROM_BE:
1322 case BPF_ALU | BPF_END | BPF_FROM_LE:
1323 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
1326 td = get_reg_val_type(ctx, this_idx, insn->dst_reg);
1327 if (insn->imm == 64 && td == REG_32BIT)
1328 emit_instr(ctx, dinsu, dst, MIPS_R_ZERO, 32, 32);
1330 if (insn->imm != 64 && td == REG_64BIT) {
1332 emit_instr(ctx, sll, dst, dst, 0);
1336 need_swap = (BPF_SRC(insn->code) == BPF_FROM_LE);
1338 need_swap = (BPF_SRC(insn->code) == BPF_FROM_BE);
1340 if (insn->imm == 16) {
1342 emit_instr(ctx, wsbh, dst, dst);
1343 emit_instr(ctx, andi, dst, dst, 0xffff);
1344 } else if (insn->imm == 32) {
1346 emit_instr(ctx, wsbh, dst, dst);
1347 emit_instr(ctx, rotr, dst, dst, 16);
1349 } else { /* 64-bit*/
1351 emit_instr(ctx, dsbh, dst, dst);
1352 emit_instr(ctx, dshd, dst, dst);
1357 case BPF_ST | BPF_B | BPF_MEM:
1358 case BPF_ST | BPF_H | BPF_MEM:
1359 case BPF_ST | BPF_W | BPF_MEM:
1360 case BPF_ST | BPF_DW | BPF_MEM:
1361 if (insn->dst_reg == BPF_REG_10) {
1362 ctx->flags |= EBPF_SEEN_FP;
1364 mem_off = insn->off + MAX_BPF_STACK;
1366 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
1369 mem_off = insn->off;
1371 gen_imm_to_reg(insn, MIPS_R_AT, ctx);
1372 switch (BPF_SIZE(insn->code)) {
1374 emit_instr(ctx, sb, MIPS_R_AT, mem_off, dst);
1377 emit_instr(ctx, sh, MIPS_R_AT, mem_off, dst);
1380 emit_instr(ctx, sw, MIPS_R_AT, mem_off, dst);
1383 emit_instr(ctx, sd, MIPS_R_AT, mem_off, dst);
1388 case BPF_LDX | BPF_B | BPF_MEM:
1389 case BPF_LDX | BPF_H | BPF_MEM:
1390 case BPF_LDX | BPF_W | BPF_MEM:
1391 case BPF_LDX | BPF_DW | BPF_MEM:
1392 if (insn->src_reg == BPF_REG_10) {
1393 ctx->flags |= EBPF_SEEN_FP;
1395 mem_off = insn->off + MAX_BPF_STACK;
1397 src = ebpf_to_mips_reg(ctx, insn, src_reg_no_fp);
1400 mem_off = insn->off;
1402 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
1405 switch (BPF_SIZE(insn->code)) {
1407 emit_instr(ctx, lbu, dst, mem_off, src);
1410 emit_instr(ctx, lhu, dst, mem_off, src);
1413 emit_instr(ctx, lw, dst, mem_off, src);
1416 emit_instr(ctx, ld, dst, mem_off, src);
1421 case BPF_STX | BPF_B | BPF_MEM:
1422 case BPF_STX | BPF_H | BPF_MEM:
1423 case BPF_STX | BPF_W | BPF_MEM:
1424 case BPF_STX | BPF_DW | BPF_MEM:
1425 case BPF_STX | BPF_W | BPF_XADD:
1426 case BPF_STX | BPF_DW | BPF_XADD:
1427 if (insn->dst_reg == BPF_REG_10) {
1428 ctx->flags |= EBPF_SEEN_FP;
1430 mem_off = insn->off + MAX_BPF_STACK;
1432 dst = ebpf_to_mips_reg(ctx, insn, dst_reg);
1435 mem_off = insn->off;
1437 src = ebpf_to_mips_reg(ctx, insn, src_reg_no_fp);
1440 if (BPF_MODE(insn->code) == BPF_XADD) {
1442 * If mem_off does not fit within the 9 bit ll/sc
1443 * instruction immediate field, use a temp reg.
1445 if (MIPS_ISA_REV >= 6 &&
1446 (mem_off >= BIT(8) || mem_off < -BIT(8))) {
1447 emit_instr(ctx, daddiu, MIPS_R_T6,
1452 switch (BPF_SIZE(insn->code)) {
1454 if (get_reg_val_type(ctx, this_idx, insn->src_reg) == REG_32BIT) {
1455 emit_instr(ctx, sll, MIPS_R_AT, src, 0);
1458 emit_instr(ctx, ll, MIPS_R_T8, mem_off, dst);
1459 emit_instr(ctx, addu, MIPS_R_T8, MIPS_R_T8, src);
1460 emit_instr(ctx, sc, MIPS_R_T8, mem_off, dst);
1462 * On failure back up to LL (-4
1463 * instructions of 4 bytes each
1465 emit_instr(ctx, beq, MIPS_R_T8, MIPS_R_ZERO, -4 * 4);
1466 emit_instr(ctx, nop);
1469 if (get_reg_val_type(ctx, this_idx, insn->src_reg) == REG_32BIT) {
1470 emit_instr(ctx, daddu, MIPS_R_AT, src, MIPS_R_ZERO);
1471 emit_instr(ctx, dinsu, MIPS_R_AT, MIPS_R_ZERO, 32, 32);
1474 emit_instr(ctx, lld, MIPS_R_T8, mem_off, dst);
1475 emit_instr(ctx, daddu, MIPS_R_T8, MIPS_R_T8, src);
1476 emit_instr(ctx, scd, MIPS_R_T8, mem_off, dst);
1477 emit_instr(ctx, beq, MIPS_R_T8, MIPS_R_ZERO, -4 * 4);
1478 emit_instr(ctx, nop);
1481 } else { /* BPF_MEM */
1482 switch (BPF_SIZE(insn->code)) {
1484 emit_instr(ctx, sb, src, mem_off, dst);
1487 emit_instr(ctx, sh, src, mem_off, dst);
1490 emit_instr(ctx, sw, src, mem_off, dst);
1493 if (get_reg_val_type(ctx, this_idx, insn->src_reg) == REG_32BIT) {
1494 emit_instr(ctx, daddu, MIPS_R_AT, src, MIPS_R_ZERO);
1495 emit_instr(ctx, dinsu, MIPS_R_AT, MIPS_R_ZERO, 32, 32);
1498 emit_instr(ctx, sd, src, mem_off, dst);
1505 pr_err("NOT HANDLED %d - (%02x)\n",
1506 this_idx, (unsigned int)insn->code);
1512 #define RVT_VISITED_MASK 0xc000000000000000ull
1513 #define RVT_FALL_THROUGH 0x4000000000000000ull
1514 #define RVT_BRANCH_TAKEN 0x8000000000000000ull
1515 #define RVT_DONE (RVT_FALL_THROUGH | RVT_BRANCH_TAKEN)
1517 static int build_int_body(struct jit_ctx *ctx)
1519 const struct bpf_prog *prog = ctx->skf;
1520 const struct bpf_insn *insn;
1523 for (i = 0; i < prog->len; ) {
1524 insn = prog->insnsi + i;
1525 if ((ctx->reg_val_types[i] & RVT_VISITED_MASK) == 0) {
1526 /* dead instruction, don't emit it. */
1531 if (ctx->target == NULL)
1532 ctx->offsets[i] = (ctx->offsets[i] & OFFSETS_B_CONV) | (ctx->idx * 4);
1534 r = build_one_insn(insn, ctx, i, prog->len);
1539 /* epilogue offset */
1540 if (ctx->target == NULL)
1541 ctx->offsets[i] = ctx->idx * 4;
1544 * All exits have an offset of the epilogue, some offsets may
1545 * not have been set due to banch-around threading, so set
1548 if (ctx->target == NULL)
1549 for (i = 0; i < prog->len; i++) {
1550 insn = prog->insnsi + i;
1551 if (insn->code == (BPF_JMP | BPF_EXIT))
1552 ctx->offsets[i] = ctx->idx * 4;
1557 /* return the last idx processed, or negative for error */
1558 static int reg_val_propagate_range(struct jit_ctx *ctx, u64 initial_rvt,
1559 int start_idx, bool follow_taken)
1561 const struct bpf_prog *prog = ctx->skf;
1562 const struct bpf_insn *insn;
1563 u64 exit_rvt = initial_rvt;
1564 u64 *rvt = ctx->reg_val_types;
1568 for (idx = start_idx; idx < prog->len; idx++) {
1569 rvt[idx] = (rvt[idx] & RVT_VISITED_MASK) | exit_rvt;
1570 insn = prog->insnsi + idx;
1571 switch (BPF_CLASS(insn->code)) {
1573 switch (BPF_OP(insn->code)) {
1585 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT);
1588 if (BPF_SRC(insn->code)) {
1589 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT);
1591 /* IMM to REG move*/
1593 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT_POS);
1595 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT);
1599 if (insn->imm == 64)
1600 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_64BIT);
1601 else if (insn->imm == 32)
1602 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT);
1603 else /* insn->imm == 16 */
1604 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT_POS);
1607 rvt[idx] |= RVT_DONE;
1610 switch (BPF_OP(insn->code)) {
1612 if (BPF_SRC(insn->code)) {
1613 /* REG to REG move*/
1614 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_64BIT);
1616 /* IMM to REG move*/
1618 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT_POS);
1620 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_64BIT_32BIT);
1624 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_64BIT);
1626 rvt[idx] |= RVT_DONE;
1629 switch (BPF_SIZE(insn->code)) {
1631 if (BPF_MODE(insn->code) == BPF_IMM) {
1634 val = (s64)((u32)insn->imm | ((u64)(insn + 1)->imm << 32));
1635 if (val > 0 && val <= S32_MAX)
1636 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT_POS);
1637 else if (val >= S32_MIN && val <= S32_MAX)
1638 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_64BIT_32BIT);
1640 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_64BIT);
1641 rvt[idx] |= RVT_DONE;
1644 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_64BIT);
1649 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT_POS);
1652 if (BPF_MODE(insn->code) == BPF_IMM)
1653 set_reg_val_type(&exit_rvt, insn->dst_reg,
1654 insn->imm >= 0 ? REG_32BIT_POS : REG_32BIT);
1656 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT);
1659 rvt[idx] |= RVT_DONE;
1662 switch (BPF_SIZE(insn->code)) {
1664 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_64BIT);
1668 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT_POS);
1671 set_reg_val_type(&exit_rvt, insn->dst_reg, REG_32BIT);
1674 rvt[idx] |= RVT_DONE;
1677 switch (BPF_OP(insn->code)) {
1679 rvt[idx] = RVT_DONE | exit_rvt;
1680 rvt[prog->len] = exit_rvt;
1683 rvt[idx] |= RVT_DONE;
1698 rvt[idx] |= RVT_BRANCH_TAKEN;
1700 follow_taken = false;
1702 rvt[idx] |= RVT_FALL_THROUGH;
1706 set_reg_val_type(&exit_rvt, BPF_REG_0, REG_64BIT);
1707 /* Upon call return, argument registers are clobbered. */
1708 for (reg = BPF_REG_0; reg <= BPF_REG_5; reg++)
1709 set_reg_val_type(&exit_rvt, reg, REG_64BIT);
1711 rvt[idx] |= RVT_DONE;
1714 WARN(1, "Unhandled BPF_JMP case.\n");
1715 rvt[idx] |= RVT_DONE;
1720 rvt[idx] |= RVT_DONE;
1728 * Track the value range (i.e. 32-bit vs. 64-bit) of each register at
1729 * each eBPF insn. This allows unneeded sign and zero extension
1730 * operations to be omitted.
1732 * Doesn't handle yet confluence of control paths with conflicting
1733 * ranges, but it is good enough for most sane code.
1735 static int reg_val_propagate(struct jit_ctx *ctx)
1737 const struct bpf_prog *prog = ctx->skf;
1743 * 11 registers * 3 bits/reg leaves top bits free for other
1744 * uses. Bit-62..63 used to see if we have visited an insn.
1748 /* Upon entry, argument registers are 64-bit. */
1749 for (reg = BPF_REG_1; reg <= BPF_REG_5; reg++)
1750 set_reg_val_type(&exit_rvt, reg, REG_64BIT);
1753 * First follow all conditional branches on the fall-through
1754 * edge of control flow..
1756 reg_val_propagate_range(ctx, exit_rvt, 0, false);
1759 * Then repeatedly find the first conditional branch where
1760 * both edges of control flow have not been taken, and follow
1761 * the branch taken edge. We will end up restarting the
1762 * search once per conditional branch insn.
1764 for (i = 0; i < prog->len; i++) {
1765 u64 rvt = ctx->reg_val_types[i];
1767 if ((rvt & RVT_VISITED_MASK) == RVT_DONE ||
1768 (rvt & RVT_VISITED_MASK) == 0)
1770 if ((rvt & RVT_VISITED_MASK) == RVT_FALL_THROUGH) {
1771 reg_val_propagate_range(ctx, rvt & ~RVT_VISITED_MASK, i, true);
1772 } else { /* RVT_BRANCH_TAKEN */
1773 WARN(1, "Unexpected RVT_BRANCH_TAKEN case.\n");
1774 reg_val_propagate_range(ctx, rvt & ~RVT_VISITED_MASK, i, false);
1776 goto restart_search;
1779 * Eventually all conditional branches have been followed on
1780 * both branches and we are done. Any insn that has not been
1781 * visited at this point is dead.
1787 static void jit_fill_hole(void *area, unsigned int size)
1791 /* We are guaranteed to have aligned memory. */
1792 for (p = area; size >= sizeof(u32); size -= sizeof(u32))
1793 uasm_i_break(&p, BRK_BUG); /* Increments p */
1796 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1798 struct bpf_prog *orig_prog = prog;
1799 bool tmp_blinded = false;
1800 struct bpf_prog *tmp;
1801 struct bpf_binary_header *header = NULL;
1803 unsigned int image_size;
1806 if (!prog->jit_requested || MIPS_ISA_REV < 2)
1809 tmp = bpf_jit_blind_constants(prog);
1810 /* If blinding was requested and we failed during blinding,
1811 * we must fall back to the interpreter.
1820 memset(&ctx, 0, sizeof(ctx));
1823 switch (current_cpu_type()) {
1824 case CPU_CAVIUM_OCTEON:
1825 case CPU_CAVIUM_OCTEON_PLUS:
1826 case CPU_CAVIUM_OCTEON2:
1827 case CPU_CAVIUM_OCTEON3:
1828 ctx.use_bbit_insns = 1;
1831 ctx.use_bbit_insns = 0;
1835 ctx.offsets = kcalloc(prog->len + 1, sizeof(*ctx.offsets), GFP_KERNEL);
1836 if (ctx.offsets == NULL)
1839 ctx.reg_val_types = kcalloc(prog->len + 1, sizeof(*ctx.reg_val_types), GFP_KERNEL);
1840 if (ctx.reg_val_types == NULL)
1845 if (reg_val_propagate(&ctx))
1849 * First pass discovers used resources and instruction offsets
1850 * assuming short branches are used.
1852 if (build_int_body(&ctx))
1856 * If no calls are made (EBPF_SAVE_RA), then tail call count
1857 * in $v1, else we must save in n$s4.
1859 if (ctx.flags & EBPF_SEEN_TC) {
1860 if (ctx.flags & EBPF_SAVE_RA)
1861 ctx.flags |= EBPF_SAVE_S4;
1863 ctx.flags |= EBPF_TCC_IN_V1;
1867 * Second pass generates offsets, if any branches are out of
1868 * range a jump-around long sequence is generated, and we have
1869 * to try again from the beginning to generate the new
1870 * offsets. This is done until no additional conversions are
1875 ctx.gen_b_offsets = 1;
1876 ctx.long_b_conversion = 0;
1877 if (gen_int_prologue(&ctx))
1879 if (build_int_body(&ctx))
1881 if (build_int_epilogue(&ctx, MIPS_R_RA))
1883 } while (ctx.long_b_conversion);
1885 image_size = 4 * ctx.idx;
1887 header = bpf_jit_binary_alloc(image_size, &image_ptr,
1888 sizeof(u32), jit_fill_hole);
1892 ctx.target = (u32 *)image_ptr;
1894 /* Third pass generates the code */
1896 if (gen_int_prologue(&ctx))
1898 if (build_int_body(&ctx))
1900 if (build_int_epilogue(&ctx, MIPS_R_RA))
1903 /* Update the icache */
1904 flush_icache_range((unsigned long)ctx.target,
1905 (unsigned long)&ctx.target[ctx.idx]);
1907 if (bpf_jit_enable > 1)
1909 bpf_jit_dump(prog->len, image_size, 2, ctx.target);
1911 bpf_jit_binary_lock_ro(header);
1912 prog->bpf_func = (void *)ctx.target;
1914 prog->jited_len = image_size;
1917 bpf_jit_prog_release_other(prog, prog == orig_prog ?
1920 kfree(ctx.reg_val_types);
1927 bpf_jit_binary_free(header);