1 // SPDX-License-Identifier: GPL-2.0-only
3 * BPF JIT compiler for LoongArch
5 * Copyright (C) 2022 Loongson Technology Corporation Limited
9 #define REG_TCC LOONGARCH_GPR_A6
10 #define TCC_SAVED LOONGARCH_GPR_S5
12 #define SAVE_RA BIT(0)
13 #define SAVE_TCC BIT(1)
15 static const int regmap[] = {
16 /* return value from in-kernel function, and exit value for eBPF program */
17 [BPF_REG_0] = LOONGARCH_GPR_A5,
18 /* arguments from eBPF program to in-kernel function */
19 [BPF_REG_1] = LOONGARCH_GPR_A0,
20 [BPF_REG_2] = LOONGARCH_GPR_A1,
21 [BPF_REG_3] = LOONGARCH_GPR_A2,
22 [BPF_REG_4] = LOONGARCH_GPR_A3,
23 [BPF_REG_5] = LOONGARCH_GPR_A4,
24 /* callee saved registers that in-kernel function will preserve */
25 [BPF_REG_6] = LOONGARCH_GPR_S0,
26 [BPF_REG_7] = LOONGARCH_GPR_S1,
27 [BPF_REG_8] = LOONGARCH_GPR_S2,
28 [BPF_REG_9] = LOONGARCH_GPR_S3,
29 /* read-only frame pointer to access stack */
30 [BPF_REG_FP] = LOONGARCH_GPR_S4,
31 /* temporary register for blinding constants */
32 [BPF_REG_AX] = LOONGARCH_GPR_T0,
35 static void mark_call(struct jit_ctx *ctx)
37 ctx->flags |= SAVE_RA;
40 static void mark_tail_call(struct jit_ctx *ctx)
42 ctx->flags |= SAVE_TCC;
45 static bool seen_call(struct jit_ctx *ctx)
47 return (ctx->flags & SAVE_RA);
50 static bool seen_tail_call(struct jit_ctx *ctx)
52 return (ctx->flags & SAVE_TCC);
55 static u8 tail_call_reg(struct jit_ctx *ctx)
64 * eBPF prog stack layout:
67 * original $sp ------------> +-------------------------+ <--LOONGARCH_GPR_FP
69 * +-------------------------+
71 * +-------------------------+
73 * +-------------------------+
75 * +-------------------------+
77 * +-------------------------+
79 * +-------------------------+
81 * +-------------------------+
83 * +-------------------------+ <--BPF_REG_FP
84 * | prog->aux->stack_depth |
86 * current $sp -------------> +-------------------------+
89 static void build_prologue(struct jit_ctx *ctx)
91 int stack_adjust = 0, store_offset, bpf_stack_adjust;
93 bpf_stack_adjust = round_up(ctx->prog->aux->stack_depth, 16);
95 /* To store ra, fp, s0, s1, s2, s3, s4 and s5. */
96 stack_adjust += sizeof(long) * 8;
98 stack_adjust = round_up(stack_adjust, 16);
99 stack_adjust += bpf_stack_adjust;
102 * First instruction initializes the tail call count (TCC).
103 * On tail call we skip this instruction, and the TCC is
104 * passed in REG_TCC from the caller.
106 emit_insn(ctx, addid, REG_TCC, LOONGARCH_GPR_ZERO, MAX_TAIL_CALL_CNT);
108 emit_insn(ctx, addid, LOONGARCH_GPR_SP, LOONGARCH_GPR_SP, -stack_adjust);
110 store_offset = stack_adjust - sizeof(long);
111 emit_insn(ctx, std, LOONGARCH_GPR_RA, LOONGARCH_GPR_SP, store_offset);
113 store_offset -= sizeof(long);
114 emit_insn(ctx, std, LOONGARCH_GPR_FP, LOONGARCH_GPR_SP, store_offset);
116 store_offset -= sizeof(long);
117 emit_insn(ctx, std, LOONGARCH_GPR_S0, LOONGARCH_GPR_SP, store_offset);
119 store_offset -= sizeof(long);
120 emit_insn(ctx, std, LOONGARCH_GPR_S1, LOONGARCH_GPR_SP, store_offset);
122 store_offset -= sizeof(long);
123 emit_insn(ctx, std, LOONGARCH_GPR_S2, LOONGARCH_GPR_SP, store_offset);
125 store_offset -= sizeof(long);
126 emit_insn(ctx, std, LOONGARCH_GPR_S3, LOONGARCH_GPR_SP, store_offset);
128 store_offset -= sizeof(long);
129 emit_insn(ctx, std, LOONGARCH_GPR_S4, LOONGARCH_GPR_SP, store_offset);
131 store_offset -= sizeof(long);
132 emit_insn(ctx, std, LOONGARCH_GPR_S5, LOONGARCH_GPR_SP, store_offset);
134 emit_insn(ctx, addid, LOONGARCH_GPR_FP, LOONGARCH_GPR_SP, stack_adjust);
136 if (bpf_stack_adjust)
137 emit_insn(ctx, addid, regmap[BPF_REG_FP], LOONGARCH_GPR_SP, bpf_stack_adjust);
140 * Program contains calls and tail calls, so REG_TCC need
141 * to be saved across calls.
143 if (seen_tail_call(ctx) && seen_call(ctx))
144 move_reg(ctx, TCC_SAVED, REG_TCC);
146 ctx->stack_size = stack_adjust;
149 static void __build_epilogue(struct jit_ctx *ctx, bool is_tail_call)
151 int stack_adjust = ctx->stack_size;
154 load_offset = stack_adjust - sizeof(long);
155 emit_insn(ctx, ldd, LOONGARCH_GPR_RA, LOONGARCH_GPR_SP, load_offset);
157 load_offset -= sizeof(long);
158 emit_insn(ctx, ldd, LOONGARCH_GPR_FP, LOONGARCH_GPR_SP, load_offset);
160 load_offset -= sizeof(long);
161 emit_insn(ctx, ldd, LOONGARCH_GPR_S0, LOONGARCH_GPR_SP, load_offset);
163 load_offset -= sizeof(long);
164 emit_insn(ctx, ldd, LOONGARCH_GPR_S1, LOONGARCH_GPR_SP, load_offset);
166 load_offset -= sizeof(long);
167 emit_insn(ctx, ldd, LOONGARCH_GPR_S2, LOONGARCH_GPR_SP, load_offset);
169 load_offset -= sizeof(long);
170 emit_insn(ctx, ldd, LOONGARCH_GPR_S3, LOONGARCH_GPR_SP, load_offset);
172 load_offset -= sizeof(long);
173 emit_insn(ctx, ldd, LOONGARCH_GPR_S4, LOONGARCH_GPR_SP, load_offset);
175 load_offset -= sizeof(long);
176 emit_insn(ctx, ldd, LOONGARCH_GPR_S5, LOONGARCH_GPR_SP, load_offset);
178 emit_insn(ctx, addid, LOONGARCH_GPR_SP, LOONGARCH_GPR_SP, stack_adjust);
181 /* Set return value */
182 move_reg(ctx, LOONGARCH_GPR_A0, regmap[BPF_REG_0]);
183 /* Return to the caller */
184 emit_insn(ctx, jirl, LOONGARCH_GPR_RA, LOONGARCH_GPR_ZERO, 0);
187 * Call the next bpf prog and skip the first instruction
188 * of TCC initialization.
190 emit_insn(ctx, jirl, LOONGARCH_GPR_T3, LOONGARCH_GPR_ZERO, 1);
194 static void build_epilogue(struct jit_ctx *ctx)
196 __build_epilogue(ctx, false);
199 bool bpf_jit_supports_kfunc_call(void)
204 /* initialized on the first pass of build_body() */
205 static int out_offset = -1;
206 static int emit_bpf_tail_call(struct jit_ctx *ctx)
209 u8 tcc = tail_call_reg(ctx);
210 u8 a1 = LOONGARCH_GPR_A1;
211 u8 a2 = LOONGARCH_GPR_A2;
212 u8 t1 = LOONGARCH_GPR_T1;
213 u8 t2 = LOONGARCH_GPR_T2;
214 u8 t3 = LOONGARCH_GPR_T3;
215 const int idx0 = ctx->idx;
217 #define cur_offset (ctx->idx - idx0)
218 #define jmp_offset (out_offset - (cur_offset))
225 * if (index >= array->map.max_entries)
228 off = offsetof(struct bpf_array, map.max_entries);
229 emit_insn(ctx, ldwu, t1, a1, off);
230 /* bgeu $a2, $t1, jmp_offset */
231 if (emit_tailcall_jmp(ctx, BPF_JGE, a2, t1, jmp_offset) < 0)
238 emit_insn(ctx, addid, REG_TCC, tcc, -1);
239 if (emit_tailcall_jmp(ctx, BPF_JSLT, REG_TCC, LOONGARCH_GPR_ZERO, jmp_offset) < 0)
243 * prog = array->ptrs[index];
247 emit_insn(ctx, alsld, t2, a2, a1, 2);
248 off = offsetof(struct bpf_array, ptrs);
249 emit_insn(ctx, ldd, t2, t2, off);
250 /* beq $t2, $zero, jmp_offset */
251 if (emit_tailcall_jmp(ctx, BPF_JEQ, t2, LOONGARCH_GPR_ZERO, jmp_offset) < 0)
254 /* goto *(prog->bpf_func + 4); */
255 off = offsetof(struct bpf_prog, bpf_func);
256 emit_insn(ctx, ldd, t3, t2, off);
257 __build_epilogue(ctx, true);
260 if (out_offset == -1)
261 out_offset = cur_offset;
262 if (cur_offset != out_offset) {
263 pr_err_once("tail_call out_offset = %d, expected %d!\n",
264 cur_offset, out_offset);
271 pr_info_once("tail_call: jump too far\n");
277 static void emit_atomic(const struct bpf_insn *insn, struct jit_ctx *ctx)
279 const u8 t1 = LOONGARCH_GPR_T1;
280 const u8 t2 = LOONGARCH_GPR_T2;
281 const u8 t3 = LOONGARCH_GPR_T3;
282 const u8 r0 = regmap[BPF_REG_0];
283 const u8 src = regmap[insn->src_reg];
284 const u8 dst = regmap[insn->dst_reg];
285 const s16 off = insn->off;
286 const s32 imm = insn->imm;
287 const bool isdw = BPF_SIZE(insn->code) == BPF_DW;
289 move_imm(ctx, t1, off, false);
290 emit_insn(ctx, addd, t1, dst, t1);
291 move_reg(ctx, t3, src);
294 /* lock *(size *)(dst + off) <op>= src */
297 emit_insn(ctx, amaddd, t2, t1, src);
299 emit_insn(ctx, amaddw, t2, t1, src);
303 emit_insn(ctx, amandd, t2, t1, src);
305 emit_insn(ctx, amandw, t2, t1, src);
309 emit_insn(ctx, amord, t2, t1, src);
311 emit_insn(ctx, amorw, t2, t1, src);
315 emit_insn(ctx, amxord, t2, t1, src);
317 emit_insn(ctx, amxorw, t2, t1, src);
319 /* src = atomic_fetch_<op>(dst + off, src) */
320 case BPF_ADD | BPF_FETCH:
322 emit_insn(ctx, amaddd, src, t1, t3);
324 emit_insn(ctx, amaddw, src, t1, t3);
325 emit_zext_32(ctx, src, true);
328 case BPF_AND | BPF_FETCH:
330 emit_insn(ctx, amandd, src, t1, t3);
332 emit_insn(ctx, amandw, src, t1, t3);
333 emit_zext_32(ctx, src, true);
336 case BPF_OR | BPF_FETCH:
338 emit_insn(ctx, amord, src, t1, t3);
340 emit_insn(ctx, amorw, src, t1, t3);
341 emit_zext_32(ctx, src, true);
344 case BPF_XOR | BPF_FETCH:
346 emit_insn(ctx, amxord, src, t1, t3);
348 emit_insn(ctx, amxorw, src, t1, t3);
349 emit_zext_32(ctx, src, true);
352 /* src = atomic_xchg(dst + off, src); */
355 emit_insn(ctx, amswapd, src, t1, t3);
357 emit_insn(ctx, amswapw, src, t1, t3);
358 emit_zext_32(ctx, src, true);
361 /* r0 = atomic_cmpxchg(dst + off, r0, src); */
363 move_reg(ctx, t2, r0);
365 emit_insn(ctx, lld, r0, t1, 0);
366 emit_insn(ctx, bne, t2, r0, 4);
367 move_reg(ctx, t3, src);
368 emit_insn(ctx, scd, t3, t1, 0);
369 emit_insn(ctx, beq, t3, LOONGARCH_GPR_ZERO, -4);
371 emit_insn(ctx, llw, r0, t1, 0);
372 emit_zext_32(ctx, t2, true);
373 emit_zext_32(ctx, r0, true);
374 emit_insn(ctx, bne, t2, r0, 4);
375 move_reg(ctx, t3, src);
376 emit_insn(ctx, scw, t3, t1, 0);
377 emit_insn(ctx, beq, t3, LOONGARCH_GPR_ZERO, -6);
378 emit_zext_32(ctx, r0, true);
384 static bool is_signed_bpf_cond(u8 cond)
386 return cond == BPF_JSGT || cond == BPF_JSLT ||
387 cond == BPF_JSGE || cond == BPF_JSLE;
390 #define BPF_FIXUP_REG_MASK GENMASK(31, 27)
391 #define BPF_FIXUP_OFFSET_MASK GENMASK(26, 0)
393 bool ex_handler_bpf(const struct exception_table_entry *ex,
394 struct pt_regs *regs)
396 int dst_reg = FIELD_GET(BPF_FIXUP_REG_MASK, ex->fixup);
397 off_t offset = FIELD_GET(BPF_FIXUP_OFFSET_MASK, ex->fixup);
399 regs->regs[dst_reg] = 0;
400 regs->csr_era = (unsigned long)&ex->fixup - offset;
405 /* For accesses to BTF pointers, add an entry to the exception table */
406 static int add_exception_handler(const struct bpf_insn *insn,
412 struct exception_table_entry *ex;
414 if (!ctx->image || !ctx->prog->aux->extable || BPF_MODE(insn->code) != BPF_PROBE_MEM)
417 if (WARN_ON_ONCE(ctx->num_exentries >= ctx->prog->aux->num_exentries))
420 ex = &ctx->prog->aux->extable[ctx->num_exentries];
421 pc = (unsigned long)&ctx->image[ctx->idx - 1];
423 offset = pc - (long)&ex->insn;
424 if (WARN_ON_ONCE(offset >= 0 || offset < INT_MIN))
430 * Since the extable follows the program, the fixup offset is always
431 * negative and limited to BPF_JIT_REGION_SIZE. Store a positive value
432 * to keep things simple, and put the destination register in the upper
433 * bits. We don't need to worry about buildtime or runtime sort
434 * modifying the upper bits because the table is already sorted, and
435 * isn't part of the main exception table.
437 offset = (long)&ex->fixup - (pc + LOONGARCH_INSN_SIZE);
438 if (!FIELD_FIT(BPF_FIXUP_OFFSET_MASK, offset))
441 ex->type = EX_TYPE_BPF;
442 ex->fixup = FIELD_PREP(BPF_FIXUP_OFFSET_MASK, offset) | FIELD_PREP(BPF_FIXUP_REG_MASK, dst_reg);
444 ctx->num_exentries++;
449 static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx, bool extra_pass)
453 bool func_addr_fixed;
454 int i = insn - ctx->prog->insnsi;
456 const u8 code = insn->code;
457 const u8 cond = BPF_OP(code);
458 const u8 t1 = LOONGARCH_GPR_T1;
459 const u8 t2 = LOONGARCH_GPR_T2;
460 const u8 src = regmap[insn->src_reg];
461 const u8 dst = regmap[insn->dst_reg];
462 const s16 off = insn->off;
463 const s32 imm = insn->imm;
464 const bool is32 = BPF_CLASS(insn->code) == BPF_ALU || BPF_CLASS(insn->code) == BPF_JMP32;
468 case BPF_ALU | BPF_MOV | BPF_X:
469 case BPF_ALU64 | BPF_MOV | BPF_X:
470 move_reg(ctx, dst, src);
471 emit_zext_32(ctx, dst, is32);
475 case BPF_ALU | BPF_MOV | BPF_K:
476 case BPF_ALU64 | BPF_MOV | BPF_K:
477 move_imm(ctx, dst, imm, is32);
480 /* dst = dst + src */
481 case BPF_ALU | BPF_ADD | BPF_X:
482 case BPF_ALU64 | BPF_ADD | BPF_X:
483 emit_insn(ctx, addd, dst, dst, src);
484 emit_zext_32(ctx, dst, is32);
487 /* dst = dst + imm */
488 case BPF_ALU | BPF_ADD | BPF_K:
489 case BPF_ALU64 | BPF_ADD | BPF_K:
490 if (is_signed_imm12(imm)) {
491 emit_insn(ctx, addid, dst, dst, imm);
493 move_imm(ctx, t1, imm, is32);
494 emit_insn(ctx, addd, dst, dst, t1);
496 emit_zext_32(ctx, dst, is32);
499 /* dst = dst - src */
500 case BPF_ALU | BPF_SUB | BPF_X:
501 case BPF_ALU64 | BPF_SUB | BPF_X:
502 emit_insn(ctx, subd, dst, dst, src);
503 emit_zext_32(ctx, dst, is32);
506 /* dst = dst - imm */
507 case BPF_ALU | BPF_SUB | BPF_K:
508 case BPF_ALU64 | BPF_SUB | BPF_K:
509 if (is_signed_imm12(-imm)) {
510 emit_insn(ctx, addid, dst, dst, -imm);
512 move_imm(ctx, t1, imm, is32);
513 emit_insn(ctx, subd, dst, dst, t1);
515 emit_zext_32(ctx, dst, is32);
518 /* dst = dst * src */
519 case BPF_ALU | BPF_MUL | BPF_X:
520 case BPF_ALU64 | BPF_MUL | BPF_X:
521 emit_insn(ctx, muld, dst, dst, src);
522 emit_zext_32(ctx, dst, is32);
525 /* dst = dst * imm */
526 case BPF_ALU | BPF_MUL | BPF_K:
527 case BPF_ALU64 | BPF_MUL | BPF_K:
528 move_imm(ctx, t1, imm, is32);
529 emit_insn(ctx, muld, dst, dst, t1);
530 emit_zext_32(ctx, dst, is32);
533 /* dst = dst / src */
534 case BPF_ALU | BPF_DIV | BPF_X:
535 case BPF_ALU64 | BPF_DIV | BPF_X:
536 emit_zext_32(ctx, dst, is32);
537 move_reg(ctx, t1, src);
538 emit_zext_32(ctx, t1, is32);
539 emit_insn(ctx, divdu, dst, dst, t1);
540 emit_zext_32(ctx, dst, is32);
543 /* dst = dst / imm */
544 case BPF_ALU | BPF_DIV | BPF_K:
545 case BPF_ALU64 | BPF_DIV | BPF_K:
546 move_imm(ctx, t1, imm, is32);
547 emit_zext_32(ctx, dst, is32);
548 emit_insn(ctx, divdu, dst, dst, t1);
549 emit_zext_32(ctx, dst, is32);
552 /* dst = dst % src */
553 case BPF_ALU | BPF_MOD | BPF_X:
554 case BPF_ALU64 | BPF_MOD | BPF_X:
555 emit_zext_32(ctx, dst, is32);
556 move_reg(ctx, t1, src);
557 emit_zext_32(ctx, t1, is32);
558 emit_insn(ctx, moddu, dst, dst, t1);
559 emit_zext_32(ctx, dst, is32);
562 /* dst = dst % imm */
563 case BPF_ALU | BPF_MOD | BPF_K:
564 case BPF_ALU64 | BPF_MOD | BPF_K:
565 move_imm(ctx, t1, imm, is32);
566 emit_zext_32(ctx, dst, is32);
567 emit_insn(ctx, moddu, dst, dst, t1);
568 emit_zext_32(ctx, dst, is32);
572 case BPF_ALU | BPF_NEG:
573 case BPF_ALU64 | BPF_NEG:
574 move_imm(ctx, t1, imm, is32);
575 emit_insn(ctx, subd, dst, LOONGARCH_GPR_ZERO, dst);
576 emit_zext_32(ctx, dst, is32);
579 /* dst = dst & src */
580 case BPF_ALU | BPF_AND | BPF_X:
581 case BPF_ALU64 | BPF_AND | BPF_X:
582 emit_insn(ctx, and, dst, dst, src);
583 emit_zext_32(ctx, dst, is32);
586 /* dst = dst & imm */
587 case BPF_ALU | BPF_AND | BPF_K:
588 case BPF_ALU64 | BPF_AND | BPF_K:
589 if (is_unsigned_imm12(imm)) {
590 emit_insn(ctx, andi, dst, dst, imm);
592 move_imm(ctx, t1, imm, is32);
593 emit_insn(ctx, and, dst, dst, t1);
595 emit_zext_32(ctx, dst, is32);
598 /* dst = dst | src */
599 case BPF_ALU | BPF_OR | BPF_X:
600 case BPF_ALU64 | BPF_OR | BPF_X:
601 emit_insn(ctx, or, dst, dst, src);
602 emit_zext_32(ctx, dst, is32);
605 /* dst = dst | imm */
606 case BPF_ALU | BPF_OR | BPF_K:
607 case BPF_ALU64 | BPF_OR | BPF_K:
608 if (is_unsigned_imm12(imm)) {
609 emit_insn(ctx, ori, dst, dst, imm);
611 move_imm(ctx, t1, imm, is32);
612 emit_insn(ctx, or, dst, dst, t1);
614 emit_zext_32(ctx, dst, is32);
617 /* dst = dst ^ src */
618 case BPF_ALU | BPF_XOR | BPF_X:
619 case BPF_ALU64 | BPF_XOR | BPF_X:
620 emit_insn(ctx, xor, dst, dst, src);
621 emit_zext_32(ctx, dst, is32);
624 /* dst = dst ^ imm */
625 case BPF_ALU | BPF_XOR | BPF_K:
626 case BPF_ALU64 | BPF_XOR | BPF_K:
627 if (is_unsigned_imm12(imm)) {
628 emit_insn(ctx, xori, dst, dst, imm);
630 move_imm(ctx, t1, imm, is32);
631 emit_insn(ctx, xor, dst, dst, t1);
633 emit_zext_32(ctx, dst, is32);
636 /* dst = dst << src (logical) */
637 case BPF_ALU | BPF_LSH | BPF_X:
638 emit_insn(ctx, sllw, dst, dst, src);
639 emit_zext_32(ctx, dst, is32);
642 case BPF_ALU64 | BPF_LSH | BPF_X:
643 emit_insn(ctx, slld, dst, dst, src);
646 /* dst = dst << imm (logical) */
647 case BPF_ALU | BPF_LSH | BPF_K:
648 emit_insn(ctx, slliw, dst, dst, imm);
649 emit_zext_32(ctx, dst, is32);
652 case BPF_ALU64 | BPF_LSH | BPF_K:
653 emit_insn(ctx, sllid, dst, dst, imm);
656 /* dst = dst >> src (logical) */
657 case BPF_ALU | BPF_RSH | BPF_X:
658 emit_insn(ctx, srlw, dst, dst, src);
659 emit_zext_32(ctx, dst, is32);
662 case BPF_ALU64 | BPF_RSH | BPF_X:
663 emit_insn(ctx, srld, dst, dst, src);
666 /* dst = dst >> imm (logical) */
667 case BPF_ALU | BPF_RSH | BPF_K:
668 emit_insn(ctx, srliw, dst, dst, imm);
669 emit_zext_32(ctx, dst, is32);
672 case BPF_ALU64 | BPF_RSH | BPF_K:
673 emit_insn(ctx, srlid, dst, dst, imm);
676 /* dst = dst >> src (arithmetic) */
677 case BPF_ALU | BPF_ARSH | BPF_X:
678 emit_insn(ctx, sraw, dst, dst, src);
679 emit_zext_32(ctx, dst, is32);
682 case BPF_ALU64 | BPF_ARSH | BPF_X:
683 emit_insn(ctx, srad, dst, dst, src);
686 /* dst = dst >> imm (arithmetic) */
687 case BPF_ALU | BPF_ARSH | BPF_K:
688 emit_insn(ctx, sraiw, dst, dst, imm);
689 emit_zext_32(ctx, dst, is32);
692 case BPF_ALU64 | BPF_ARSH | BPF_K:
693 emit_insn(ctx, sraid, dst, dst, imm);
696 /* dst = BSWAP##imm(dst) */
697 case BPF_ALU | BPF_END | BPF_FROM_LE:
700 /* zero-extend 16 bits into 64 bits */
701 emit_insn(ctx, bstrpickd, dst, dst, 15, 0);
704 /* zero-extend 32 bits into 64 bits */
705 emit_zext_32(ctx, dst, is32);
713 case BPF_ALU | BPF_END | BPF_FROM_BE:
716 emit_insn(ctx, revb2h, dst, dst);
717 /* zero-extend 16 bits into 64 bits */
718 emit_insn(ctx, bstrpickd, dst, dst, 15, 0);
721 emit_insn(ctx, revb2w, dst, dst);
722 /* zero-extend 32 bits into 64 bits */
723 emit_zext_32(ctx, dst, is32);
726 emit_insn(ctx, revbd, dst, dst);
731 /* PC += off if dst cond src */
732 case BPF_JMP | BPF_JEQ | BPF_X:
733 case BPF_JMP | BPF_JNE | BPF_X:
734 case BPF_JMP | BPF_JGT | BPF_X:
735 case BPF_JMP | BPF_JGE | BPF_X:
736 case BPF_JMP | BPF_JLT | BPF_X:
737 case BPF_JMP | BPF_JLE | BPF_X:
738 case BPF_JMP | BPF_JSGT | BPF_X:
739 case BPF_JMP | BPF_JSGE | BPF_X:
740 case BPF_JMP | BPF_JSLT | BPF_X:
741 case BPF_JMP | BPF_JSLE | BPF_X:
742 case BPF_JMP32 | BPF_JEQ | BPF_X:
743 case BPF_JMP32 | BPF_JNE | BPF_X:
744 case BPF_JMP32 | BPF_JGT | BPF_X:
745 case BPF_JMP32 | BPF_JGE | BPF_X:
746 case BPF_JMP32 | BPF_JLT | BPF_X:
747 case BPF_JMP32 | BPF_JLE | BPF_X:
748 case BPF_JMP32 | BPF_JSGT | BPF_X:
749 case BPF_JMP32 | BPF_JSGE | BPF_X:
750 case BPF_JMP32 | BPF_JSLT | BPF_X:
751 case BPF_JMP32 | BPF_JSLE | BPF_X:
752 jmp_offset = bpf2la_offset(i, off, ctx);
753 move_reg(ctx, t1, dst);
754 move_reg(ctx, t2, src);
755 if (is_signed_bpf_cond(BPF_OP(code))) {
756 emit_sext_32(ctx, t1, is32);
757 emit_sext_32(ctx, t2, is32);
759 emit_zext_32(ctx, t1, is32);
760 emit_zext_32(ctx, t2, is32);
762 if (emit_cond_jmp(ctx, cond, t1, t2, jmp_offset) < 0)
766 /* PC += off if dst cond imm */
767 case BPF_JMP | BPF_JEQ | BPF_K:
768 case BPF_JMP | BPF_JNE | BPF_K:
769 case BPF_JMP | BPF_JGT | BPF_K:
770 case BPF_JMP | BPF_JGE | BPF_K:
771 case BPF_JMP | BPF_JLT | BPF_K:
772 case BPF_JMP | BPF_JLE | BPF_K:
773 case BPF_JMP | BPF_JSGT | BPF_K:
774 case BPF_JMP | BPF_JSGE | BPF_K:
775 case BPF_JMP | BPF_JSLT | BPF_K:
776 case BPF_JMP | BPF_JSLE | BPF_K:
777 case BPF_JMP32 | BPF_JEQ | BPF_K:
778 case BPF_JMP32 | BPF_JNE | BPF_K:
779 case BPF_JMP32 | BPF_JGT | BPF_K:
780 case BPF_JMP32 | BPF_JGE | BPF_K:
781 case BPF_JMP32 | BPF_JLT | BPF_K:
782 case BPF_JMP32 | BPF_JLE | BPF_K:
783 case BPF_JMP32 | BPF_JSGT | BPF_K:
784 case BPF_JMP32 | BPF_JSGE | BPF_K:
785 case BPF_JMP32 | BPF_JSLT | BPF_K:
786 case BPF_JMP32 | BPF_JSLE | BPF_K:
787 jmp_offset = bpf2la_offset(i, off, ctx);
789 move_imm(ctx, t1, imm, false);
792 /* If imm is 0, simply use zero register. */
793 tm = LOONGARCH_GPR_ZERO;
795 move_reg(ctx, t2, dst);
796 if (is_signed_bpf_cond(BPF_OP(code))) {
797 emit_sext_32(ctx, tm, is32);
798 emit_sext_32(ctx, t2, is32);
800 emit_zext_32(ctx, tm, is32);
801 emit_zext_32(ctx, t2, is32);
803 if (emit_cond_jmp(ctx, cond, t2, tm, jmp_offset) < 0)
807 /* PC += off if dst & src */
808 case BPF_JMP | BPF_JSET | BPF_X:
809 case BPF_JMP32 | BPF_JSET | BPF_X:
810 jmp_offset = bpf2la_offset(i, off, ctx);
811 emit_insn(ctx, and, t1, dst, src);
812 emit_zext_32(ctx, t1, is32);
813 if (emit_cond_jmp(ctx, cond, t1, LOONGARCH_GPR_ZERO, jmp_offset) < 0)
817 /* PC += off if dst & imm */
818 case BPF_JMP | BPF_JSET | BPF_K:
819 case BPF_JMP32 | BPF_JSET | BPF_K:
820 jmp_offset = bpf2la_offset(i, off, ctx);
821 move_imm(ctx, t1, imm, is32);
822 emit_insn(ctx, and, t1, dst, t1);
823 emit_zext_32(ctx, t1, is32);
824 if (emit_cond_jmp(ctx, cond, t1, LOONGARCH_GPR_ZERO, jmp_offset) < 0)
829 case BPF_JMP | BPF_JA:
830 jmp_offset = bpf2la_offset(i, off, ctx);
831 if (emit_uncond_jmp(ctx, jmp_offset) < 0)
836 case BPF_JMP | BPF_CALL:
838 ret = bpf_jit_get_func_addr(ctx->prog, insn, extra_pass,
839 &func_addr, &func_addr_fixed);
843 move_addr(ctx, t1, func_addr);
844 emit_insn(ctx, jirl, t1, LOONGARCH_GPR_RA, 0);
845 move_reg(ctx, regmap[BPF_REG_0], LOONGARCH_GPR_A0);
849 case BPF_JMP | BPF_TAIL_CALL:
851 if (emit_bpf_tail_call(ctx) < 0)
855 /* function return */
856 case BPF_JMP | BPF_EXIT:
857 if (i == ctx->prog->len - 1)
860 jmp_offset = epilogue_offset(ctx);
861 if (emit_uncond_jmp(ctx, jmp_offset) < 0)
866 case BPF_LD | BPF_IMM | BPF_DW:
868 const u64 imm64 = (u64)(insn + 1)->imm << 32 | (u32)insn->imm;
870 move_imm(ctx, dst, imm64, is32);
874 /* dst = *(size *)(src + off) */
875 case BPF_LDX | BPF_MEM | BPF_B:
876 case BPF_LDX | BPF_MEM | BPF_H:
877 case BPF_LDX | BPF_MEM | BPF_W:
878 case BPF_LDX | BPF_MEM | BPF_DW:
879 case BPF_LDX | BPF_PROBE_MEM | BPF_DW:
880 case BPF_LDX | BPF_PROBE_MEM | BPF_W:
881 case BPF_LDX | BPF_PROBE_MEM | BPF_H:
882 case BPF_LDX | BPF_PROBE_MEM | BPF_B:
883 switch (BPF_SIZE(code)) {
885 if (is_signed_imm12(off)) {
886 emit_insn(ctx, ldbu, dst, src, off);
888 move_imm(ctx, t1, off, is32);
889 emit_insn(ctx, ldxbu, dst, src, t1);
893 if (is_signed_imm12(off)) {
894 emit_insn(ctx, ldhu, dst, src, off);
896 move_imm(ctx, t1, off, is32);
897 emit_insn(ctx, ldxhu, dst, src, t1);
901 if (is_signed_imm12(off)) {
902 emit_insn(ctx, ldwu, dst, src, off);
903 } else if (is_signed_imm14(off)) {
904 emit_insn(ctx, ldptrw, dst, src, off);
906 move_imm(ctx, t1, off, is32);
907 emit_insn(ctx, ldxwu, dst, src, t1);
911 move_imm(ctx, t1, off, is32);
912 emit_insn(ctx, ldxd, dst, src, t1);
916 ret = add_exception_handler(insn, ctx, dst);
921 /* *(size *)(dst + off) = imm */
922 case BPF_ST | BPF_MEM | BPF_B:
923 case BPF_ST | BPF_MEM | BPF_H:
924 case BPF_ST | BPF_MEM | BPF_W:
925 case BPF_ST | BPF_MEM | BPF_DW:
926 switch (BPF_SIZE(code)) {
928 move_imm(ctx, t1, imm, is32);
929 if (is_signed_imm12(off)) {
930 emit_insn(ctx, stb, t1, dst, off);
932 move_imm(ctx, t2, off, is32);
933 emit_insn(ctx, stxb, t1, dst, t2);
937 move_imm(ctx, t1, imm, is32);
938 if (is_signed_imm12(off)) {
939 emit_insn(ctx, sth, t1, dst, off);
941 move_imm(ctx, t2, off, is32);
942 emit_insn(ctx, stxh, t1, dst, t2);
946 move_imm(ctx, t1, imm, is32);
947 if (is_signed_imm12(off)) {
948 emit_insn(ctx, stw, t1, dst, off);
949 } else if (is_signed_imm14(off)) {
950 emit_insn(ctx, stptrw, t1, dst, off);
952 move_imm(ctx, t2, off, is32);
953 emit_insn(ctx, stxw, t1, dst, t2);
957 move_imm(ctx, t1, imm, is32);
958 if (is_signed_imm12(off)) {
959 emit_insn(ctx, std, t1, dst, off);
960 } else if (is_signed_imm14(off)) {
961 emit_insn(ctx, stptrd, t1, dst, off);
963 move_imm(ctx, t2, off, is32);
964 emit_insn(ctx, stxd, t1, dst, t2);
970 /* *(size *)(dst + off) = src */
971 case BPF_STX | BPF_MEM | BPF_B:
972 case BPF_STX | BPF_MEM | BPF_H:
973 case BPF_STX | BPF_MEM | BPF_W:
974 case BPF_STX | BPF_MEM | BPF_DW:
975 switch (BPF_SIZE(code)) {
977 if (is_signed_imm12(off)) {
978 emit_insn(ctx, stb, src, dst, off);
980 move_imm(ctx, t1, off, is32);
981 emit_insn(ctx, stxb, src, dst, t1);
985 if (is_signed_imm12(off)) {
986 emit_insn(ctx, sth, src, dst, off);
988 move_imm(ctx, t1, off, is32);
989 emit_insn(ctx, stxh, src, dst, t1);
993 if (is_signed_imm12(off)) {
994 emit_insn(ctx, stw, src, dst, off);
995 } else if (is_signed_imm14(off)) {
996 emit_insn(ctx, stptrw, src, dst, off);
998 move_imm(ctx, t1, off, is32);
999 emit_insn(ctx, stxw, src, dst, t1);
1003 if (is_signed_imm12(off)) {
1004 emit_insn(ctx, std, src, dst, off);
1005 } else if (is_signed_imm14(off)) {
1006 emit_insn(ctx, stptrd, src, dst, off);
1008 move_imm(ctx, t1, off, is32);
1009 emit_insn(ctx, stxd, src, dst, t1);
1015 case BPF_STX | BPF_ATOMIC | BPF_W:
1016 case BPF_STX | BPF_ATOMIC | BPF_DW:
1017 emit_atomic(insn, ctx);
1020 /* Speculation barrier */
1021 case BPF_ST | BPF_NOSPEC:
1025 pr_err("bpf_jit: unknown opcode %02x\n", code);
1032 pr_info_once("bpf_jit: opcode %02x, jump too far\n", code);
1036 static int build_body(struct jit_ctx *ctx, bool extra_pass)
1039 const struct bpf_prog *prog = ctx->prog;
1041 for (i = 0; i < prog->len; i++) {
1042 const struct bpf_insn *insn = &prog->insnsi[i];
1045 if (ctx->image == NULL)
1046 ctx->offset[i] = ctx->idx;
1048 ret = build_insn(insn, ctx, extra_pass);
1051 if (ctx->image == NULL)
1052 ctx->offset[i] = ctx->idx;
1059 if (ctx->image == NULL)
1060 ctx->offset[i] = ctx->idx;
1065 /* Fill space with break instructions */
1066 static void jit_fill_hole(void *area, unsigned int size)
1070 /* We are guaranteed to have aligned memory */
1071 for (ptr = area; size >= sizeof(u32); size -= sizeof(u32))
1072 *ptr++ = INSN_BREAK;
1075 static int validate_code(struct jit_ctx *ctx)
1078 union loongarch_instruction insn;
1080 for (i = 0; i < ctx->idx; i++) {
1081 insn = ctx->image[i];
1082 /* Check INSN_BREAK */
1083 if (insn.word == INSN_BREAK)
1087 if (WARN_ON_ONCE(ctx->num_exentries != ctx->prog->aux->num_exentries))
1093 struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog)
1095 bool tmp_blinded = false, extra_pass = false;
1097 int image_size, prog_size, extable_size;
1099 struct jit_data *jit_data;
1100 struct bpf_binary_header *header;
1101 struct bpf_prog *tmp, *orig_prog = prog;
1104 * If BPF JIT was not enabled then we must fall back to
1107 if (!prog->jit_requested)
1110 tmp = bpf_jit_blind_constants(prog);
1112 * If blinding was requested and we failed during blinding,
1113 * we must fall back to the interpreter. Otherwise, we save
1114 * the new JITed code.
1124 jit_data = prog->aux->jit_data;
1126 jit_data = kzalloc(sizeof(*jit_data), GFP_KERNEL);
1131 prog->aux->jit_data = jit_data;
1133 if (jit_data->ctx.offset) {
1134 ctx = jit_data->ctx;
1135 image_ptr = jit_data->image;
1136 header = jit_data->header;
1138 prog_size = sizeof(u32) * ctx.idx;
1142 memset(&ctx, 0, sizeof(ctx));
1145 ctx.offset = kvcalloc(prog->len + 1, sizeof(u32), GFP_KERNEL);
1146 if (ctx.offset == NULL) {
1151 /* 1. Initial fake pass to compute ctx->idx and set ctx->flags */
1152 build_prologue(&ctx);
1153 if (build_body(&ctx, extra_pass)) {
1157 ctx.epilogue_offset = ctx.idx;
1158 build_epilogue(&ctx);
1160 extable_size = prog->aux->num_exentries * sizeof(struct exception_table_entry);
1162 /* Now we know the actual image size.
1163 * As each LoongArch instruction is of length 32bit,
1164 * we are translating number of JITed intructions into
1165 * the size required to store these JITed code.
1167 prog_size = sizeof(u32) * ctx.idx;
1168 image_size = prog_size + extable_size;
1169 /* Now we know the size of the structure to make */
1170 header = bpf_jit_binary_alloc(image_size, &image_ptr,
1171 sizeof(u32), jit_fill_hole);
1172 if (header == NULL) {
1177 /* 2. Now, the actual pass to generate final JIT code */
1178 ctx.image = (union loongarch_instruction *)image_ptr;
1180 prog->aux->extable = (void *)image_ptr + prog_size;
1184 ctx.num_exentries = 0;
1186 build_prologue(&ctx);
1187 if (build_body(&ctx, extra_pass)) {
1188 bpf_jit_binary_free(header);
1192 build_epilogue(&ctx);
1194 /* 3. Extra pass to validate JITed code */
1195 if (validate_code(&ctx)) {
1196 bpf_jit_binary_free(header);
1201 /* And we're done */
1202 if (bpf_jit_enable > 1)
1203 bpf_jit_dump(prog->len, prog_size, 2, ctx.image);
1205 /* Update the icache */
1206 flush_icache_range((unsigned long)header, (unsigned long)(ctx.image + ctx.idx));
1208 if (!prog->is_func || extra_pass) {
1209 if (extra_pass && ctx.idx != jit_data->ctx.idx) {
1210 pr_err_once("multi-func JIT bug %d != %d\n",
1211 ctx.idx, jit_data->ctx.idx);
1212 bpf_jit_binary_free(header);
1213 prog->bpf_func = NULL;
1215 prog->jited_len = 0;
1218 bpf_jit_binary_lock_ro(header);
1220 jit_data->ctx = ctx;
1221 jit_data->image = image_ptr;
1222 jit_data->header = header;
1225 prog->jited_len = prog_size;
1226 prog->bpf_func = (void *)ctx.image;
1228 if (!prog->is_func || extra_pass) {
1231 /* offset[prog->len] is the size of program */
1232 for (i = 0; i <= prog->len; i++)
1233 ctx.offset[i] *= LOONGARCH_INSN_SIZE;
1234 bpf_prog_fill_jited_linfo(prog, ctx.offset + 1);
1239 prog->aux->jit_data = NULL;
1244 bpf_jit_prog_release_other(prog, prog == orig_prog ? tmp : orig_prog);
1251 /* Indicate the JIT backend supports mixing bpf2bpf and tailcalls. */
1252 bool bpf_jit_supports_subprog_tailcalls(void)