Message ID | 20191216091343.23260-3-bjorn.topel@gmail.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | riscv: BPF JIT fix, optimizations and far jumps support | expand |
On Mon, 16 Dec 2019 01:13:36 PST (-0800), Bjorn Topel wrote: > This commit adds branch relaxation to the BPF JIT, and with that > support for far (offset greater than 12b) branching. Interesting. We don't actually relax these in the binutils linker, but instead just do it staticly at assembly time... > The branch relaxation requires more than two passes to converge. For > most programs it is three passes, but for larger programs it can be > more. ... and that's why :). In binutils we just worst-case the link-time relaxations when doing assembler relaxation, which proves to be good enough. C code doesn't branch outside a function, so most branches end up fairly short anyway. > Reviewed-by: Luke Nelson <lukenels@cs.washington.edu> > Cc: Xi Wang <xi.wang@gmail.com> > Signed-off-by: Björn Töpel <bjorn.topel@gmail.com> > --- > arch/riscv/net/bpf_jit_comp.c | 352 ++++++++++++++++++---------------- > 1 file changed, 188 insertions(+), 164 deletions(-) > > diff --git a/arch/riscv/net/bpf_jit_comp.c b/arch/riscv/net/bpf_jit_comp.c > index 1606ebd49666..e599458a9bcd 100644 > --- a/arch/riscv/net/bpf_jit_comp.c > +++ b/arch/riscv/net/bpf_jit_comp.c > @@ -461,6 +461,11 @@ static u32 rv_amoadd_d(u8 rd, u8 rs2, u8 rs1, u8 aq, u8 rl) > return rv_amo_insn(0, aq, rl, rs2, rs1, 3, rd, 0x2f); > } > > +static u32 rv_auipc(u8 rd, u32 imm31_12) > +{ > + return rv_u_insn(imm31_12, rd, 0x17); > +} > + > static bool is_12b_int(s64 val) > { > return -(1 << 11) <= val && val < (1 << 11); > @@ -484,7 +489,7 @@ static bool is_32b_int(s64 val) > static int is_12b_check(int off, int insn) > { > if (!is_12b_int(off)) { > - pr_err("bpf-jit: insn=%d offset=%d not supported yet!\n", > + pr_err("bpf-jit: insn=%d 12b < offset=%d not supported yet!\n", > insn, (int)off); > return -1; > } > @@ -494,7 +499,7 @@ static int is_12b_check(int off, int insn) > static int is_13b_check(int off, int insn) > { > if (!is_13b_int(off)) { > - pr_err("bpf-jit: insn=%d offset=%d not supported yet!\n", > + pr_err("bpf-jit: insn=%d 13b < offset=%d not supported yet!\n", > insn, (int)off); > return -1; > } > @@ -504,7 +509,7 @@ static int is_13b_check(int off, int insn) > static int is_21b_check(int off, int insn) > { > if (!is_21b_int(off)) { > - pr_err("bpf-jit: insn=%d offset=%d not supported yet!\n", > + pr_err("bpf-jit: insn=%d 21b < offset=%d not supported yet!\n", > insn, (int)off); > return -1; > } > @@ -550,10 +555,13 @@ static void emit_imm(u8 rd, s64 val, struct rv_jit_context *ctx) > emit(rv_addi(rd, rd, lower), ctx); > } > > -static int rv_offset(int bpf_to, int bpf_from, struct rv_jit_context *ctx) > +static int rv_offset(int insn, int off, struct rv_jit_context *ctx) > { > - int from = ctx->offset[bpf_from] - 1, to = ctx->offset[bpf_to]; > + int from, to; > > + off++; /* BPF branch is from PC+1, RV is from PC */ > + from = (insn > 0) ? ctx->offset[insn - 1] : 0; > + to = (insn + off > 0) ? ctx->offset[insn + off - 1] : 0; > return (to - from) << 2; > } > > @@ -606,6 +614,109 @@ static void __build_epilogue(u8 reg, struct rv_jit_context *ctx) > emit(rv_jalr(RV_REG_ZERO, reg, 0), ctx); > } > > +/* return -1 or inverted cond */ > +static int invert_bpf_cond(u8 cond) > +{ > + switch (cond) { > + case BPF_JEQ: > + return BPF_JNE; > + case BPF_JGT: > + return BPF_JLE; > + case BPF_JLT: > + return BPF_JGE; > + case BPF_JGE: > + return BPF_JLT; > + case BPF_JLE: > + return BPF_JGT; > + case BPF_JNE: > + return BPF_JEQ; > + case BPF_JSGT: > + return BPF_JSLE; > + case BPF_JSLT: > + return BPF_JSGE; > + case BPF_JSGE: > + return BPF_JSLT; > + case BPF_JSLE: > + return BPF_JSGT; > + } > + return -1; > +} > + > +static void emit_bcc(u8 cond, u8 rd, u8 rs, int rvoff, > + struct rv_jit_context *ctx) > +{ > + switch (cond) { > + case BPF_JEQ: > + emit(rv_beq(rd, rs, rvoff >> 1), ctx); > + return; > + case BPF_JGT: > + emit(rv_bltu(rs, rd, rvoff >> 1), ctx); > + return; > + case BPF_JLT: > + emit(rv_bltu(rd, rs, rvoff >> 1), ctx); > + return; > + case BPF_JGE: > + emit(rv_bgeu(rd, rs, rvoff >> 1), ctx); > + return; > + case BPF_JLE: > + emit(rv_bgeu(rs, rd, rvoff >> 1), ctx); > + return; > + case BPF_JNE: > + emit(rv_bne(rd, rs, rvoff >> 1), ctx); > + return; > + case BPF_JSGT: > + emit(rv_blt(rs, rd, rvoff >> 1), ctx); > + return; > + case BPF_JSLT: > + emit(rv_blt(rd, rs, rvoff >> 1), ctx); > + return; > + case BPF_JSGE: > + emit(rv_bge(rd, rs, rvoff >> 1), ctx); > + return; > + case BPF_JSLE: > + emit(rv_bge(rs, rd, rvoff >> 1), ctx); > + } > +} > + > +static void emit_branch(u8 cond, u8 rd, u8 rs, int rvoff, > + struct rv_jit_context *ctx) > +{ > + s64 upper, lower; > + > + if (is_13b_int(rvoff)) { > + emit_bcc(cond, rd, rs, rvoff, ctx); > + return; > + } > + > + /* Adjust for jal */ > + rvoff -= 4; > + > + /* Transform, e.g.: > + * bne rd,rs,foo > + * to > + * beq rd,rs,<.L1> > + * (auipc foo) > + * jal(r) foo > + * .L1 > + */ > + cond = invert_bpf_cond(cond); > + if (is_21b_int(rvoff)) { > + emit_bcc(cond, rd, rs, 8, ctx); > + emit(rv_jal(RV_REG_ZERO, rvoff >> 1), ctx); > + return; > + } > + > + /* 32b No need for an additional rvoff adjustment, since we > + * get that from the auipc at PC', where PC = PC' + 4. > + */ > + upper = (rvoff + (1 << 11)) >> 12; > + lower = rvoff & 0xfff; > + > + emit_bcc(cond, rd, rs, 12, ctx); > + emit(rv_auipc(RV_REG_T1, upper), ctx); > + emit(rv_jalr(RV_REG_ZERO, RV_REG_T1, lower), ctx); > +} > + > static void emit_zext_32(u8 reg, struct rv_jit_context *ctx) > { > emit(rv_slli(reg, reg, 32), ctx); > @@ -693,13 +804,6 @@ static void init_regs(u8 *rd, u8 *rs, const struct bpf_insn *insn, > *rs = bpf_to_rv_reg(insn->src_reg, ctx); > } > > -static int rv_offset_check(int *rvoff, s16 off, int insn, > - struct rv_jit_context *ctx) > -{ > - *rvoff = rv_offset(insn + off, insn, ctx); > - return is_13b_check(*rvoff, insn); > -} > - > static void emit_zext_32_rd_rs(u8 *rd, u8 *rs, struct rv_jit_context *ctx) > { > emit(rv_addi(RV_REG_T2, *rd, 0), ctx); > @@ -732,13 +836,19 @@ static void emit_sext_32_rd(u8 *rd, struct rv_jit_context *ctx) > *rd = RV_REG_T2; > } > > +static bool is_signed_bpf_cond(u8 cond) > +{ > + return cond == BPF_JSGT || cond == BPF_JSLT || > + cond == BPF_JSGE || cond == BPF_JSLE; > +} > + > static int emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx, > bool extra_pass) > { > bool is64 = BPF_CLASS(insn->code) == BPF_ALU64 || > BPF_CLASS(insn->code) == BPF_JMP; > + int s, e, rvoff, i = insn - ctx->prog->insnsi; > struct bpf_prog_aux *aux = ctx->prog->aux; > - int rvoff, i = insn - ctx->prog->insnsi; > u8 rd = -1, rs = -1, code = insn->code; > s16 off = insn->off; > s32 imm = insn->imm; > @@ -1006,7 +1116,7 @@ static int emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx, > > /* JUMP off */ > case BPF_JMP | BPF_JA: > - rvoff = rv_offset(i + off, i, ctx); > + rvoff = rv_offset(i, off, ctx); > if (!is_21b_int(rvoff)) { > pr_err("bpf-jit: insn=%d offset=%d not supported yet!\n", > i, rvoff); > @@ -1019,194 +1129,96 @@ static int emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx, > /* IF (dst COND src) JUMP off */ > case BPF_JMP | BPF_JEQ | BPF_X: > case BPF_JMP32 | BPF_JEQ | BPF_X: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - if (!is64) > - emit_zext_32_rd_rs(&rd, &rs, ctx); > - emit(rv_beq(rd, rs, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JGT | BPF_X: > case BPF_JMP32 | BPF_JGT | BPF_X: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - if (!is64) > - emit_zext_32_rd_rs(&rd, &rs, ctx); > - emit(rv_bltu(rs, rd, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JLT | BPF_X: > case BPF_JMP32 | BPF_JLT | BPF_X: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - if (!is64) > - emit_zext_32_rd_rs(&rd, &rs, ctx); > - emit(rv_bltu(rd, rs, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JGE | BPF_X: > case BPF_JMP32 | BPF_JGE | BPF_X: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - if (!is64) > - emit_zext_32_rd_rs(&rd, &rs, ctx); > - emit(rv_bgeu(rd, rs, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JLE | BPF_X: > case BPF_JMP32 | BPF_JLE | BPF_X: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - if (!is64) > - emit_zext_32_rd_rs(&rd, &rs, ctx); > - emit(rv_bgeu(rs, rd, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JNE | BPF_X: > case BPF_JMP32 | BPF_JNE | BPF_X: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - if (!is64) > - emit_zext_32_rd_rs(&rd, &rs, ctx); > - emit(rv_bne(rd, rs, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JSGT | BPF_X: > case BPF_JMP32 | BPF_JSGT | BPF_X: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - if (!is64) > - emit_sext_32_rd_rs(&rd, &rs, ctx); > - emit(rv_blt(rs, rd, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JSLT | BPF_X: > case BPF_JMP32 | BPF_JSLT | BPF_X: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - if (!is64) > - emit_sext_32_rd_rs(&rd, &rs, ctx); > - emit(rv_blt(rd, rs, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JSGE | BPF_X: > case BPF_JMP32 | BPF_JSGE | BPF_X: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - if (!is64) > - emit_sext_32_rd_rs(&rd, &rs, ctx); > - emit(rv_bge(rd, rs, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JSLE | BPF_X: > case BPF_JMP32 | BPF_JSLE | BPF_X: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - if (!is64) > - emit_sext_32_rd_rs(&rd, &rs, ctx); > - emit(rv_bge(rs, rd, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JSET | BPF_X: > case BPF_JMP32 | BPF_JSET | BPF_X: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - if (!is64) > - emit_zext_32_rd_rs(&rd, &rs, ctx); > - emit(rv_and(RV_REG_T1, rd, rs), ctx); > - emit(rv_bne(RV_REG_T1, RV_REG_ZERO, rvoff >> 1), ctx); > + rvoff = rv_offset(i, off, ctx); > + if (!is64) { > + s = ctx->ninsns; > + if (is_signed_bpf_cond(BPF_OP(code))) > + emit_sext_32_rd_rs(&rd, &rs, ctx); > + else > + emit_zext_32_rd_rs(&rd, &rs, ctx); > + e = ctx->ninsns; > + > + /* Adjust for extra insns */ > + rvoff -= (e - s) << 2; > + } > + > + if (BPF_OP(code) == BPF_JSET) { > + /* Adjust for and */ > + rvoff -= 4; > + emit(rv_and(RV_REG_T1, rd, rs), ctx); > + emit_branch(BPF_JNE, RV_REG_T1, RV_REG_ZERO, rvoff, > + ctx); > + } else { > + emit_branch(BPF_OP(code), rd, rs, rvoff, ctx); > + } > break; > > /* IF (dst COND imm) JUMP off */ > case BPF_JMP | BPF_JEQ | BPF_K: > case BPF_JMP32 | BPF_JEQ | BPF_K: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - emit_imm(RV_REG_T1, imm, ctx); > - if (!is64) > - emit_zext_32_rd_t1(&rd, ctx); > - emit(rv_beq(rd, RV_REG_T1, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JGT | BPF_K: > case BPF_JMP32 | BPF_JGT | BPF_K: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - emit_imm(RV_REG_T1, imm, ctx); > - if (!is64) > - emit_zext_32_rd_t1(&rd, ctx); > - emit(rv_bltu(RV_REG_T1, rd, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JLT | BPF_K: > case BPF_JMP32 | BPF_JLT | BPF_K: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - emit_imm(RV_REG_T1, imm, ctx); > - if (!is64) > - emit_zext_32_rd_t1(&rd, ctx); > - emit(rv_bltu(rd, RV_REG_T1, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JGE | BPF_K: > case BPF_JMP32 | BPF_JGE | BPF_K: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - emit_imm(RV_REG_T1, imm, ctx); > - if (!is64) > - emit_zext_32_rd_t1(&rd, ctx); > - emit(rv_bgeu(rd, RV_REG_T1, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JLE | BPF_K: > case BPF_JMP32 | BPF_JLE | BPF_K: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - emit_imm(RV_REG_T1, imm, ctx); > - if (!is64) > - emit_zext_32_rd_t1(&rd, ctx); > - emit(rv_bgeu(RV_REG_T1, rd, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JNE | BPF_K: > case BPF_JMP32 | BPF_JNE | BPF_K: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - emit_imm(RV_REG_T1, imm, ctx); > - if (!is64) > - emit_zext_32_rd_t1(&rd, ctx); > - emit(rv_bne(rd, RV_REG_T1, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JSGT | BPF_K: > case BPF_JMP32 | BPF_JSGT | BPF_K: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - emit_imm(RV_REG_T1, imm, ctx); > - if (!is64) > - emit_sext_32_rd(&rd, ctx); > - emit(rv_blt(RV_REG_T1, rd, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JSLT | BPF_K: > case BPF_JMP32 | BPF_JSLT | BPF_K: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - emit_imm(RV_REG_T1, imm, ctx); > - if (!is64) > - emit_sext_32_rd(&rd, ctx); > - emit(rv_blt(rd, RV_REG_T1, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JSGE | BPF_K: > case BPF_JMP32 | BPF_JSGE | BPF_K: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - emit_imm(RV_REG_T1, imm, ctx); > - if (!is64) > - emit_sext_32_rd(&rd, ctx); > - emit(rv_bge(rd, RV_REG_T1, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JSLE | BPF_K: > case BPF_JMP32 | BPF_JSLE | BPF_K: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > - emit_imm(RV_REG_T1, imm, ctx); > - if (!is64) > - emit_sext_32_rd(&rd, ctx); > - emit(rv_bge(RV_REG_T1, rd, rvoff >> 1), ctx); > - break; > case BPF_JMP | BPF_JSET | BPF_K: > case BPF_JMP32 | BPF_JSET | BPF_K: > - if (rv_offset_check(&rvoff, off, i, ctx)) > - return -1; > + rvoff = rv_offset(i, off, ctx); > + s = ctx->ninsns; > emit_imm(RV_REG_T1, imm, ctx); > - if (!is64) > - emit_zext_32_rd_t1(&rd, ctx); > - emit(rv_and(RV_REG_T1, rd, RV_REG_T1), ctx); > - emit(rv_bne(RV_REG_T1, RV_REG_ZERO, rvoff >> 1), ctx); > + if (!is64) { > + if (is_signed_bpf_cond(BPF_OP(code))) > + emit_sext_32_rd(&rd, ctx); > + else > + emit_zext_32_rd_t1(&rd, ctx); > + } > + e = ctx->ninsns; > + > + /* Adjust for extra insns */ > + rvoff -= (e - s) << 2; > + > + if (BPF_OP(code) == BPF_JSET) { > + /* Adjust for and */ > + rvoff -= 4; > + emit(rv_and(RV_REG_T1, rd, RV_REG_T1), ctx); > + emit_branch(BPF_JNE, RV_REG_T1, RV_REG_ZERO, rvoff, > + ctx); > + } else { > + emit_branch(BPF_OP(code), rd, RV_REG_T1, rvoff, ctx); > + } > break; > > /* function call */ > @@ -1557,6 +1569,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) > { > bool tmp_blinded = false, extra_pass = false; > struct bpf_prog *tmp, *orig_prog = prog; > + int pass = 0, prev_ninsns = 0, i; > struct rv_jit_data *jit_data; > struct rv_jit_context *ctx; > unsigned int image_size; > @@ -1596,15 +1609,25 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) > prog = orig_prog; > goto out_offset; > } > + for (i = 0; i < prog->len; i++) { > + prev_ninsns += 32; > + ctx->offset[i] = prev_ninsns; > + } It feels like the first-order implementation is the same as binutils here: the first round is worst cased, after which things can be more exact. We're only doing one pass in binutils because most of the relaxation happens in the linker, but this approach seems reasonable to me. I'd be interested in seeing some benchmarks, as it may be worth relaxing these in the binutils linker as well -- I can certainly come up with contrived test cases that aren't relaxed, but I'm not sure how common this is. My only worry is that that invariant should be more explicit. Specifically, I'm thinking that every time offset is updated there should be some sort of assertion that the offset is shrinking. This is enforced structurally in the binutils code because we only generate code once and then move it around, but since you're generating code every time it'd be easy for a bug to sneak in as the JIT gets more complicated. Since most of the branches should be forward, you'll probably end up with way fewer iterations if you do the optimization passes backwards. > - /* First pass generates the ctx->offset, but does not emit an image. */ > - if (build_body(ctx, extra_pass)) { > - prog = orig_prog; > - goto out_offset; > + for (i = 0; i < 16; i++) { > + pass++; > + ctx->ninsns = 0; > + if (build_body(ctx, extra_pass)) { > + prog = orig_prog; > + goto out_offset; Isn't this returning a broken program if build_body() errors out the first time through? > + } > + build_prologue(ctx); > + ctx->epilogue_offset = ctx->ninsns; > + build_epilogue(ctx); > + if (ctx->ninsns == prev_ninsns) > + break; > + prev_ninsns = ctx->ninsns; IDK how important the performance of the JIT is, but you could probably get away with skipping an iteration by keeping track of some simple metric that determines if it would be possible to > } > - build_prologue(ctx); > - ctx->epilogue_offset = ctx->ninsns; > - build_epilogue(ctx); > > /* Allocate image, now that we know the size. */ > image_size = sizeof(u32) * ctx->ninsns; > @@ -1619,6 +1642,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) > /* Second, real pass, that acutally emits the image. */ > ctx->insns = (u32 *)jit_data->image; > skip_init_ctx: > + pass++; > ctx->ninsns = 0; > > build_prologue(ctx); > @@ -1630,7 +1654,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) > build_epilogue(ctx); > > if (bpf_jit_enable > 1) > - bpf_jit_dump(prog->len, image_size, 2, ctx->insns); > + bpf_jit_dump(prog->len, image_size, pass, ctx->insns); > > prog->bpf_func = (void *)ctx->insns; > prog->jited = 1;
Back from the holidays; Sorry about the delayed reply. On Mon, 23 Dec 2019 at 19:03, Palmer Dabbelt <palmerdabbelt@google.com> wrote: > > On Mon, 16 Dec 2019 01:13:36 PST (-0800), Bjorn Topel wrote: > > This commit adds branch relaxation to the BPF JIT, and with that [...] > > @@ -1557,6 +1569,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) > > { > > bool tmp_blinded = false, extra_pass = false; > > struct bpf_prog *tmp, *orig_prog = prog; > > + int pass = 0, prev_ninsns = 0, i; > > struct rv_jit_data *jit_data; > > struct rv_jit_context *ctx; > > unsigned int image_size; > > @@ -1596,15 +1609,25 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) > > prog = orig_prog; > > goto out_offset; > > } > > + for (i = 0; i < prog->len; i++) { > > + prev_ninsns += 32; > > + ctx->offset[i] = prev_ninsns; > > + } > > It feels like the first-order implementation is the same as binutils here: the > first round is worst cased, after which things can be more exact. We're only > doing one pass in binutils because most of the relaxation happens in the > linker, but this approach seems reasonable to me. I'd be interested in seeing > some benchmarks, as it may be worth relaxing these in the binutils linker as > well -- I can certainly come up with contrived test cases that aren't relaxed, > but I'm not sure how common this is. > Ah, interesting! Let me try to pull out some branch relaxation statistics/benchmarks for the BPF selftests. > My only worry is that that invariant should be more explicit. Specifically, > I'm thinking that every time offset is updated there should be some sort of > assertion that the offset is shrinking. This is enforced structurally in the > binutils code because we only generate code once and then move it around, but > since you're generating code every time it'd be easy for a bug to sneak in as > the JIT gets more complicated. > Hmm, yes. Maybe use a checksum for the program in addition to the length invariant, and converge condition would then be prev_len == len && prev_crc == crc? > Since most of the branches should be forward, you'll probably end up with way > fewer iterations if you do the optimization passes backwards. > Good idea! > > - /* First pass generates the ctx->offset, but does not emit an image. */ > > - if (build_body(ctx, extra_pass)) { > > - prog = orig_prog; > > - goto out_offset; > > + for (i = 0; i < 16; i++) { > > + pass++; > > + ctx->ninsns = 0; > > + if (build_body(ctx, extra_pass)) { > > + prog = orig_prog; > > + goto out_offset; > > Isn't this returning a broken program if build_body() errors out the first time > through? > Hmm, care to elaborate? I don't see how? > > + } > > + build_prologue(ctx); > > + ctx->epilogue_offset = ctx->ninsns; > > + build_epilogue(ctx); > > + if (ctx->ninsns == prev_ninsns) > > + break; > > + prev_ninsns = ctx->ninsns; > > IDK how important the performance of the JIT is, but you could probably get > away with skipping an iteration by keeping track of some simple metric that > determines if it would be possible to > ...to? Given that the programs are getting larger, performance of the JIT is important. So, any means the number of passes can be reduced is a good thing! Thanks for the review/suggestions! Björn
On Tue, 07 Jan 2020 00:13:56 PST (-0800), Bjorn Topel wrote: > Back from the holidays; Sorry about the delayed reply. > > On Mon, 23 Dec 2019 at 19:03, Palmer Dabbelt <palmerdabbelt@google.com> wrote: >> >> On Mon, 16 Dec 2019 01:13:36 PST (-0800), Bjorn Topel wrote: >> > This commit adds branch relaxation to the BPF JIT, and with that > [...] >> > @@ -1557,6 +1569,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) >> > { >> > bool tmp_blinded = false, extra_pass = false; >> > struct bpf_prog *tmp, *orig_prog = prog; >> > + int pass = 0, prev_ninsns = 0, i; >> > struct rv_jit_data *jit_data; >> > struct rv_jit_context *ctx; >> > unsigned int image_size; >> > @@ -1596,15 +1609,25 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) >> > prog = orig_prog; >> > goto out_offset; >> > } >> > + for (i = 0; i < prog->len; i++) { >> > + prev_ninsns += 32; >> > + ctx->offset[i] = prev_ninsns; >> > + } >> >> It feels like the first-order implementation is the same as binutils here: the >> first round is worst cased, after which things can be more exact. We're only >> doing one pass in binutils because most of the relaxation happens in the >> linker, but this approach seems reasonable to me. I'd be interested in seeing >> some benchmarks, as it may be worth relaxing these in the binutils linker as >> well -- I can certainly come up with contrived test cases that aren't relaxed, >> but I'm not sure how common this is. >> > > Ah, interesting! Let me try to pull out some branch relaxation > statistics/benchmarks for the BPF selftests. > >> My only worry is that that invariant should be more explicit. Specifically, >> I'm thinking that every time offset is updated there should be some sort of >> assertion that the offset is shrinking. This is enforced structurally in the >> binutils code because we only generate code once and then move it around, but >> since you're generating code every time it'd be easy for a bug to sneak in as >> the JIT gets more complicated. >> > > Hmm, yes. Maybe use a checksum for the program in addition to the > length invariant, and converge condition would then be prev_len == len > && prev_crc == crc? That would work, but it breaks my unfinished optimization below. I was thinking something more like "every time offset[i] is updated, check that it gets smaller and otherwise barf". >> Since most of the branches should be forward, you'll probably end up with way >> fewer iterations if you do the optimization passes backwards. >> > > Good idea! > >> > - /* First pass generates the ctx->offset, but does not emit an image. */ >> > - if (build_body(ctx, extra_pass)) { >> > - prog = orig_prog; >> > - goto out_offset; >> > + for (i = 0; i < 16; i++) { >> > + pass++; >> > + ctx->ninsns = 0; >> > + if (build_body(ctx, extra_pass)) { >> > + prog = orig_prog; >> > + goto out_offset; >> >> Isn't this returning a broken program if build_body() errors out the first time >> through? >> > > Hmm, care to elaborate? I don't see how? Ya, I don't either any more. Hopefully I just got confused between prog and ctx... >> > + } >> > + build_prologue(ctx); >> > + ctx->epilogue_offset = ctx->ninsns; >> > + build_epilogue(ctx); >> > + if (ctx->ninsns == prev_ninsns) >> > + break; >> > + prev_ninsns = ctx->ninsns; >> >> IDK how important the performance of the JIT is, but you could probably get >> away with skipping an iteration by keeping track of some simple metric that >> determines if it would be possible to >> > > ...to? Given that the programs are getting larger, performance of the > JIT is important. So, any means the number of passes can be reduced is > a good thing! I guess I meant to say "determines if it would be possible to make any modifications next time". I was thinking something along the lines of: * as you run through the program, keep track of the shortest branch distance * if you didn't remove enough bytes to make that branch cross a relaxation boundary, then you know that next time you won't be able to do any useful work You're already computing all the branch lengths, so it's just an extra min(). Since we're assuming a small number of passes (after reversing the relaxation direction), you'll probably save more work avoiding the extra pass than it'll take to compute the extra information. I guess some sort of benchmark would give a real answer, but it certainly smells like a good idea ;) > > > Thanks for the review/suggestions! > Björn
diff --git a/arch/riscv/net/bpf_jit_comp.c b/arch/riscv/net/bpf_jit_comp.c index 1606ebd49666..e599458a9bcd 100644 --- a/arch/riscv/net/bpf_jit_comp.c +++ b/arch/riscv/net/bpf_jit_comp.c @@ -461,6 +461,11 @@ static u32 rv_amoadd_d(u8 rd, u8 rs2, u8 rs1, u8 aq, u8 rl) return rv_amo_insn(0, aq, rl, rs2, rs1, 3, rd, 0x2f); } +static u32 rv_auipc(u8 rd, u32 imm31_12) +{ + return rv_u_insn(imm31_12, rd, 0x17); +} + static bool is_12b_int(s64 val) { return -(1 << 11) <= val && val < (1 << 11); @@ -484,7 +489,7 @@ static bool is_32b_int(s64 val) static int is_12b_check(int off, int insn) { if (!is_12b_int(off)) { - pr_err("bpf-jit: insn=%d offset=%d not supported yet!\n", + pr_err("bpf-jit: insn=%d 12b < offset=%d not supported yet!\n", insn, (int)off); return -1; } @@ -494,7 +499,7 @@ static int is_12b_check(int off, int insn) static int is_13b_check(int off, int insn) { if (!is_13b_int(off)) { - pr_err("bpf-jit: insn=%d offset=%d not supported yet!\n", + pr_err("bpf-jit: insn=%d 13b < offset=%d not supported yet!\n", insn, (int)off); return -1; } @@ -504,7 +509,7 @@ static int is_13b_check(int off, int insn) static int is_21b_check(int off, int insn) { if (!is_21b_int(off)) { - pr_err("bpf-jit: insn=%d offset=%d not supported yet!\n", + pr_err("bpf-jit: insn=%d 21b < offset=%d not supported yet!\n", insn, (int)off); return -1; } @@ -550,10 +555,13 @@ static void emit_imm(u8 rd, s64 val, struct rv_jit_context *ctx) emit(rv_addi(rd, rd, lower), ctx); } -static int rv_offset(int bpf_to, int bpf_from, struct rv_jit_context *ctx) +static int rv_offset(int insn, int off, struct rv_jit_context *ctx) { - int from = ctx->offset[bpf_from] - 1, to = ctx->offset[bpf_to]; + int from, to; + off++; /* BPF branch is from PC+1, RV is from PC */ + from = (insn > 0) ? ctx->offset[insn - 1] : 0; + to = (insn + off > 0) ? ctx->offset[insn + off - 1] : 0; return (to - from) << 2; } @@ -606,6 +614,109 @@ static void __build_epilogue(u8 reg, struct rv_jit_context *ctx) emit(rv_jalr(RV_REG_ZERO, reg, 0), ctx); } +/* return -1 or inverted cond */ +static int invert_bpf_cond(u8 cond) +{ + switch (cond) { + case BPF_JEQ: + return BPF_JNE; + case BPF_JGT: + return BPF_JLE; + case BPF_JLT: + return BPF_JGE; + case BPF_JGE: + return BPF_JLT; + case BPF_JLE: + return BPF_JGT; + case BPF_JNE: + return BPF_JEQ; + case BPF_JSGT: + return BPF_JSLE; + case BPF_JSLT: + return BPF_JSGE; + case BPF_JSGE: + return BPF_JSLT; + case BPF_JSLE: + return BPF_JSGT; + } + return -1; +} + +static void emit_bcc(u8 cond, u8 rd, u8 rs, int rvoff, + struct rv_jit_context *ctx) +{ + switch (cond) { + case BPF_JEQ: + emit(rv_beq(rd, rs, rvoff >> 1), ctx); + return; + case BPF_JGT: + emit(rv_bltu(rs, rd, rvoff >> 1), ctx); + return; + case BPF_JLT: + emit(rv_bltu(rd, rs, rvoff >> 1), ctx); + return; + case BPF_JGE: + emit(rv_bgeu(rd, rs, rvoff >> 1), ctx); + return; + case BPF_JLE: + emit(rv_bgeu(rs, rd, rvoff >> 1), ctx); + return; + case BPF_JNE: + emit(rv_bne(rd, rs, rvoff >> 1), ctx); + return; + case BPF_JSGT: + emit(rv_blt(rs, rd, rvoff >> 1), ctx); + return; + case BPF_JSLT: + emit(rv_blt(rd, rs, rvoff >> 1), ctx); + return; + case BPF_JSGE: + emit(rv_bge(rd, rs, rvoff >> 1), ctx); + return; + case BPF_JSLE: + emit(rv_bge(rs, rd, rvoff >> 1), ctx); + } +} + +static void emit_branch(u8 cond, u8 rd, u8 rs, int rvoff, + struct rv_jit_context *ctx) +{ + s64 upper, lower; + + if (is_13b_int(rvoff)) { + emit_bcc(cond, rd, rs, rvoff, ctx); + return; + } + + /* Adjust for jal */ + rvoff -= 4; + + /* Transform, e.g.: + * bne rd,rs,foo + * to + * beq rd,rs,<.L1> + * (auipc foo) + * jal(r) foo + * .L1 + */ + cond = invert_bpf_cond(cond); + if (is_21b_int(rvoff)) { + emit_bcc(cond, rd, rs, 8, ctx); + emit(rv_jal(RV_REG_ZERO, rvoff >> 1), ctx); + return; + } + + /* 32b No need for an additional rvoff adjustment, since we + * get that from the auipc at PC', where PC = PC' + 4. + */ + upper = (rvoff + (1 << 11)) >> 12; + lower = rvoff & 0xfff; + + emit_bcc(cond, rd, rs, 12, ctx); + emit(rv_auipc(RV_REG_T1, upper), ctx); + emit(rv_jalr(RV_REG_ZERO, RV_REG_T1, lower), ctx); +} + static void emit_zext_32(u8 reg, struct rv_jit_context *ctx) { emit(rv_slli(reg, reg, 32), ctx); @@ -693,13 +804,6 @@ static void init_regs(u8 *rd, u8 *rs, const struct bpf_insn *insn, *rs = bpf_to_rv_reg(insn->src_reg, ctx); } -static int rv_offset_check(int *rvoff, s16 off, int insn, - struct rv_jit_context *ctx) -{ - *rvoff = rv_offset(insn + off, insn, ctx); - return is_13b_check(*rvoff, insn); -} - static void emit_zext_32_rd_rs(u8 *rd, u8 *rs, struct rv_jit_context *ctx) { emit(rv_addi(RV_REG_T2, *rd, 0), ctx); @@ -732,13 +836,19 @@ static void emit_sext_32_rd(u8 *rd, struct rv_jit_context *ctx) *rd = RV_REG_T2; } +static bool is_signed_bpf_cond(u8 cond) +{ + return cond == BPF_JSGT || cond == BPF_JSLT || + cond == BPF_JSGE || cond == BPF_JSLE; +} + static int emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx, bool extra_pass) { bool is64 = BPF_CLASS(insn->code) == BPF_ALU64 || BPF_CLASS(insn->code) == BPF_JMP; + int s, e, rvoff, i = insn - ctx->prog->insnsi; struct bpf_prog_aux *aux = ctx->prog->aux; - int rvoff, i = insn - ctx->prog->insnsi; u8 rd = -1, rs = -1, code = insn->code; s16 off = insn->off; s32 imm = insn->imm; @@ -1006,7 +1116,7 @@ static int emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx, /* JUMP off */ case BPF_JMP | BPF_JA: - rvoff = rv_offset(i + off, i, ctx); + rvoff = rv_offset(i, off, ctx); if (!is_21b_int(rvoff)) { pr_err("bpf-jit: insn=%d offset=%d not supported yet!\n", i, rvoff); @@ -1019,194 +1129,96 @@ static int emit_insn(const struct bpf_insn *insn, struct rv_jit_context *ctx, /* IF (dst COND src) JUMP off */ case BPF_JMP | BPF_JEQ | BPF_X: case BPF_JMP32 | BPF_JEQ | BPF_X: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - if (!is64) - emit_zext_32_rd_rs(&rd, &rs, ctx); - emit(rv_beq(rd, rs, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JGT | BPF_X: case BPF_JMP32 | BPF_JGT | BPF_X: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - if (!is64) - emit_zext_32_rd_rs(&rd, &rs, ctx); - emit(rv_bltu(rs, rd, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JLT | BPF_X: case BPF_JMP32 | BPF_JLT | BPF_X: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - if (!is64) - emit_zext_32_rd_rs(&rd, &rs, ctx); - emit(rv_bltu(rd, rs, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JGE | BPF_X: case BPF_JMP32 | BPF_JGE | BPF_X: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - if (!is64) - emit_zext_32_rd_rs(&rd, &rs, ctx); - emit(rv_bgeu(rd, rs, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JLE | BPF_X: case BPF_JMP32 | BPF_JLE | BPF_X: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - if (!is64) - emit_zext_32_rd_rs(&rd, &rs, ctx); - emit(rv_bgeu(rs, rd, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JNE | BPF_X: case BPF_JMP32 | BPF_JNE | BPF_X: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - if (!is64) - emit_zext_32_rd_rs(&rd, &rs, ctx); - emit(rv_bne(rd, rs, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JSGT | BPF_X: case BPF_JMP32 | BPF_JSGT | BPF_X: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - if (!is64) - emit_sext_32_rd_rs(&rd, &rs, ctx); - emit(rv_blt(rs, rd, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JSLT | BPF_X: case BPF_JMP32 | BPF_JSLT | BPF_X: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - if (!is64) - emit_sext_32_rd_rs(&rd, &rs, ctx); - emit(rv_blt(rd, rs, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JSGE | BPF_X: case BPF_JMP32 | BPF_JSGE | BPF_X: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - if (!is64) - emit_sext_32_rd_rs(&rd, &rs, ctx); - emit(rv_bge(rd, rs, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JSLE | BPF_X: case BPF_JMP32 | BPF_JSLE | BPF_X: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - if (!is64) - emit_sext_32_rd_rs(&rd, &rs, ctx); - emit(rv_bge(rs, rd, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JSET | BPF_X: case BPF_JMP32 | BPF_JSET | BPF_X: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - if (!is64) - emit_zext_32_rd_rs(&rd, &rs, ctx); - emit(rv_and(RV_REG_T1, rd, rs), ctx); - emit(rv_bne(RV_REG_T1, RV_REG_ZERO, rvoff >> 1), ctx); + rvoff = rv_offset(i, off, ctx); + if (!is64) { + s = ctx->ninsns; + if (is_signed_bpf_cond(BPF_OP(code))) + emit_sext_32_rd_rs(&rd, &rs, ctx); + else + emit_zext_32_rd_rs(&rd, &rs, ctx); + e = ctx->ninsns; + + /* Adjust for extra insns */ + rvoff -= (e - s) << 2; + } + + if (BPF_OP(code) == BPF_JSET) { + /* Adjust for and */ + rvoff -= 4; + emit(rv_and(RV_REG_T1, rd, rs), ctx); + emit_branch(BPF_JNE, RV_REG_T1, RV_REG_ZERO, rvoff, + ctx); + } else { + emit_branch(BPF_OP(code), rd, rs, rvoff, ctx); + } break; /* IF (dst COND imm) JUMP off */ case BPF_JMP | BPF_JEQ | BPF_K: case BPF_JMP32 | BPF_JEQ | BPF_K: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - emit_imm(RV_REG_T1, imm, ctx); - if (!is64) - emit_zext_32_rd_t1(&rd, ctx); - emit(rv_beq(rd, RV_REG_T1, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JGT | BPF_K: case BPF_JMP32 | BPF_JGT | BPF_K: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - emit_imm(RV_REG_T1, imm, ctx); - if (!is64) - emit_zext_32_rd_t1(&rd, ctx); - emit(rv_bltu(RV_REG_T1, rd, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JLT | BPF_K: case BPF_JMP32 | BPF_JLT | BPF_K: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - emit_imm(RV_REG_T1, imm, ctx); - if (!is64) - emit_zext_32_rd_t1(&rd, ctx); - emit(rv_bltu(rd, RV_REG_T1, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JGE | BPF_K: case BPF_JMP32 | BPF_JGE | BPF_K: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - emit_imm(RV_REG_T1, imm, ctx); - if (!is64) - emit_zext_32_rd_t1(&rd, ctx); - emit(rv_bgeu(rd, RV_REG_T1, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JLE | BPF_K: case BPF_JMP32 | BPF_JLE | BPF_K: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - emit_imm(RV_REG_T1, imm, ctx); - if (!is64) - emit_zext_32_rd_t1(&rd, ctx); - emit(rv_bgeu(RV_REG_T1, rd, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JNE | BPF_K: case BPF_JMP32 | BPF_JNE | BPF_K: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - emit_imm(RV_REG_T1, imm, ctx); - if (!is64) - emit_zext_32_rd_t1(&rd, ctx); - emit(rv_bne(rd, RV_REG_T1, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JSGT | BPF_K: case BPF_JMP32 | BPF_JSGT | BPF_K: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - emit_imm(RV_REG_T1, imm, ctx); - if (!is64) - emit_sext_32_rd(&rd, ctx); - emit(rv_blt(RV_REG_T1, rd, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JSLT | BPF_K: case BPF_JMP32 | BPF_JSLT | BPF_K: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - emit_imm(RV_REG_T1, imm, ctx); - if (!is64) - emit_sext_32_rd(&rd, ctx); - emit(rv_blt(rd, RV_REG_T1, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JSGE | BPF_K: case BPF_JMP32 | BPF_JSGE | BPF_K: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - emit_imm(RV_REG_T1, imm, ctx); - if (!is64) - emit_sext_32_rd(&rd, ctx); - emit(rv_bge(rd, RV_REG_T1, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JSLE | BPF_K: case BPF_JMP32 | BPF_JSLE | BPF_K: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; - emit_imm(RV_REG_T1, imm, ctx); - if (!is64) - emit_sext_32_rd(&rd, ctx); - emit(rv_bge(RV_REG_T1, rd, rvoff >> 1), ctx); - break; case BPF_JMP | BPF_JSET | BPF_K: case BPF_JMP32 | BPF_JSET | BPF_K: - if (rv_offset_check(&rvoff, off, i, ctx)) - return -1; + rvoff = rv_offset(i, off, ctx); + s = ctx->ninsns; emit_imm(RV_REG_T1, imm, ctx); - if (!is64) - emit_zext_32_rd_t1(&rd, ctx); - emit(rv_and(RV_REG_T1, rd, RV_REG_T1), ctx); - emit(rv_bne(RV_REG_T1, RV_REG_ZERO, rvoff >> 1), ctx); + if (!is64) { + if (is_signed_bpf_cond(BPF_OP(code))) + emit_sext_32_rd(&rd, ctx); + else + emit_zext_32_rd_t1(&rd, ctx); + } + e = ctx->ninsns; + + /* Adjust for extra insns */ + rvoff -= (e - s) << 2; + + if (BPF_OP(code) == BPF_JSET) { + /* Adjust for and */ + rvoff -= 4; + emit(rv_and(RV_REG_T1, rd, RV_REG_T1), ctx); + emit_branch(BPF_JNE, RV_REG_T1, RV_REG_ZERO, rvoff, + ctx); + } else { + emit_branch(BPF_OP(code), rd, RV_REG_T1, rvoff, ctx); + } break; /* function call */ @@ -1557,6 +1569,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) { bool tmp_blinded = false, extra_pass = false; struct bpf_prog *tmp, *orig_prog = prog; + int pass = 0, prev_ninsns = 0, i; struct rv_jit_data *jit_data; struct rv_jit_context *ctx; unsigned int image_size; @@ -1596,15 +1609,25 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) prog = orig_prog; goto out_offset; } + for (i = 0; i < prog->len; i++) { + prev_ninsns += 32; + ctx->offset[i] = prev_ninsns; + } - /* First pass generates the ctx->offset, but does not emit an image. */ - if (build_body(ctx, extra_pass)) { - prog = orig_prog; - goto out_offset; + for (i = 0; i < 16; i++) { + pass++; + ctx->ninsns = 0; + if (build_body(ctx, extra_pass)) { + prog = orig_prog; + goto out_offset; + } + build_prologue(ctx); + ctx->epilogue_offset = ctx->ninsns; + build_epilogue(ctx); + if (ctx->ninsns == prev_ninsns) + break; + prev_ninsns = ctx->ninsns; } - build_prologue(ctx); - ctx->epilogue_offset = ctx->ninsns; - build_epilogue(ctx); /* Allocate image, now that we know the size. */ image_size = sizeof(u32) * ctx->ninsns; @@ -1619,6 +1642,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) /* Second, real pass, that acutally emits the image. */ ctx->insns = (u32 *)jit_data->image; skip_init_ctx: + pass++; ctx->ninsns = 0; build_prologue(ctx); @@ -1630,7 +1654,7 @@ struct bpf_prog *bpf_int_jit_compile(struct bpf_prog *prog) build_epilogue(ctx); if (bpf_jit_enable > 1) - bpf_jit_dump(prog->len, image_size, 2, ctx->insns); + bpf_jit_dump(prog->len, image_size, pass, ctx->insns); prog->bpf_func = (void *)ctx->insns; prog->jited = 1;