Re: [RFC PATCH bpf-next] RV32G eBPF JIT

[Jiong Wang]

Luke Nelson writes:

> From: Luke Nelson <luke.r.nels@xxxxxxxxx>
>
> This is an eBPF JIT for RV32G, adapted from the JIT for RV64G.
> Any feedback would be greatly appreciated.
>
> It passes 359 out of 378 tests in test_bpf.ko. The failing tests are
> features that are not supported right now:
>   - ALU64 DIV/MOD:
>       These require loops to emulate on 32-bit hardware,
>       and are not supported on other 32-bit JITs like
>       ARM32.
>   - BPF_XADD | BPF_DW:
>       RV32G does not have atomic instructions for operating
>       on double words. This is similar to ARM32.
>   - Tail calls:
>       I'm working on adding support for these now, but couldn't
>       find any test cases that use them. What's the best way
>       of testing tail call code?
>   - Far branches
>       These are not supported in RV64G either.
>
> There are two main changes required for this to work compared to the
> RV64 JIT.
>
> First, eBPF registers are 64-bit, while RV32G registers are 32-bit.
> I take an approach similar to ARM32: most BPF registers map directly to
> 2 RISC-V registers, while some reside in stack scratch space and must
> be saved / restored when used.
>
> Second, many 64-bit ALU operations do not trivially map to 32-bit
> operations. Operations that move bits between high and low words, such
> as ADD, LSH, MUL, and others must emulate the 64-bit behavior in terms
> of 32-bit instructions.
>
> Signed-off-by: Luke Nelson <luke.r.nels@xxxxxxxxx>
> Cc: Xi Wang <xi.wang@xxxxxxxxx>
> ---
>  arch/riscv/Kconfig              |    2 +-
>  arch/riscv/net/Makefile         |    7 +-
>  arch/riscv/net/bpf_jit_comp32.c | 1460 +++++++++++++++++++++++++++++++
>  3 files changed, 1467 insertions(+), 2 deletions(-)
>  create mode 100644 arch/riscv/net/bpf_jit_comp32.c
>
<snip>
> +static void rv32_bpf_put_reg32(const s8 *reg, const s8 *src,
> +			       struct rv_jit_context *ctx)
> +{
> +	if (is_stacked(reg[1])) {
> +		emit(rv_sw(RV_REG_FP, reg[1], src[1]), ctx);
> +		emit(rv_sw(RV_REG_FP, reg[0], RV_REG_ZERO), ctx);
> +	} else {
> +		emit(rv_addi(reg[0], RV_REG_ZERO, 0), ctx);
> +	}
> +}
> +

Looks to me 32-bit optimization is not enabled.

If you define bpf_jit_needs_zext to return true

  bool bpf_jit_needs_zext(void)
  {
        return true;
  }

Then you don't need to zero high 32-bit when writing 32-bit sub-register
and you just need to implement the explicit zero extension insn which is a
special variant of BPF_MOV. This can save quite a few instructions. RV64
and arches like arm has implemented this, please search
"aux->verifier_zext".

And there is a doc for this optimization:

https://git.kernel.org/pub/scm/linux/kernel/git/bpf/bpf-next.git/tree/Documentation/bpf/bpf_design_QA.rst#n168

Regards,
Jiong