Re: [dpdk-dev] [PATCH 5/5] bpf: x86 JIT support for packet data loadinstructions

["Morten Brørup" <mb@smartsharesystems.com>]

> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Konstantin Ananyev
> Sent: Monday, May 18, 2020 5:53 PM
> 
> Make x86 JIT to generate native code for
> (BPF_ABS | <size> | BPF_LD) and (BPF_IND | <size> | BPF_LD)
> instructions.
> 
> Signed-off-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
> ---
>  lib/librte_bpf/bpf_jit_x86.c  | 181 ++++++++++++++++++++++++++++++++++
>  lib/librte_bpf/bpf_validate.c |  17 +++-
>  2 files changed, 197 insertions(+), 1 deletion(-)
> 
> diff --git a/lib/librte_bpf/bpf_jit_x86.c b/lib/librte_bpf/bpf_jit_x86.c
> index f70cd6be5..8a7ad45b3 100644
> --- a/lib/librte_bpf/bpf_jit_x86.c
> +++ b/lib/librte_bpf/bpf_jit_x86.c
> @@ -87,6 +87,14 @@ enum {
>  	REG_TMP1 = R10,
>  };
> 
> +/* LD_ABS/LD_IMM offsets */
> +enum {
> +	LDMB_FSP_OFS, /* fast-path */
> +	LDMB_SLP_OFS, /* slow-path */
> +	LDMB_FIN_OFS, /* final part */
> +	LDMB_OFS_NUM
> +};
> +
>  /*
>   * callee saved registers list.
>   * keep RBP as the last one.
> @@ -100,6 +108,9 @@ struct bpf_jit_state {
>  		uint32_t num;
>  		int32_t off;
>  	} exit;
> +	struct {
> +		uint32_t stack_ofs;
> +	} ldmb;
>  	uint32_t reguse;
>  	int32_t *off;
>  	uint8_t *ins;
> @@ -1024,6 +1035,166 @@ emit_div(struct bpf_jit_state *st, uint32_t op,
> uint32_t sreg, uint32_t dreg,
>  		emit_mov_reg(st, EBPF_ALU64 | EBPF_MOV | BPF_X, REG_TMP1, RDX);
>  }
> 
> +/*
> + * helper function, used by emit_ld_mbuf().
> + * generates code for 'fast_path':
> + * calculate load offset and check is it inside first packet segment.
> + */
> +static void
> +emit_ldmb_fast_path(struct bpf_jit_state *st, const uint32_t
> rg[EBPF_REG_7],
> +	uint32_t sreg, uint32_t mode, uint32_t sz, uint32_t imm,
> +	const int32_t ofs[LDMB_OFS_NUM])
> +{
> +	/* make R2 contain *off* value */
> +
> +	if (sreg != rg[EBPF_REG_2]) {
> +		emit_mov_imm(st, EBPF_ALU64 | EBPF_MOV | BPF_K,
> +			rg[EBPF_REG_2], imm);
> +		if (mode == BPF_IND)
> +			emit_alu_reg(st, EBPF_ALU64 | BPF_ADD | BPF_X,
> +				sreg, rg[EBPF_REG_2]);
> +	} else
> +		/* BPF_IND with sreg == R2 */
> +		emit_alu_imm(st, EBPF_ALU64 | BPF_ADD | BPF_K,
> +			rg[EBPF_REG_2], imm);
> +
> +	/* R3 = mbuf->data_len */
> +	emit_ld_reg(st, BPF_LDX | BPF_MEM | BPF_H,
> +		rg[EBPF_REG_6], rg[EBPF_REG_3],
> +		offsetof(struct rte_mbuf, data_len));
> +
> +	/* R3 = R3 - R2 */
> +	emit_alu_reg(st, EBPF_ALU64 | BPF_SUB | BPF_X,
> +		rg[EBPF_REG_2], rg[EBPF_REG_3]);
> +
> +	/* JSLE R3, <sz> <slow_path> */
> +	emit_cmp_imm(st, EBPF_ALU64, rg[EBPF_REG_3], sz);
> +	emit_abs_jcc(st, BPF_JMP | EBPF_JSLE | BPF_K, ofs[LDMB_SLP_OFS]);
> +
> +	/* R3 = mbuf->data_off */
> +	emit_ld_reg(st, BPF_LDX | BPF_MEM | BPF_H,
> +		rg[EBPF_REG_6], rg[EBPF_REG_3],
> +		offsetof(struct rte_mbuf, data_off));
> +
> +	/* R0 = mbuf->buf_addr */
> +	emit_ld_reg(st, BPF_LDX | BPF_MEM | EBPF_DW,
> +		rg[EBPF_REG_6], rg[EBPF_REG_0],
> +		offsetof(struct rte_mbuf, buf_addr));
> +
> +	/* R0 = R0 + R3 */
> +	emit_alu_reg(st, EBPF_ALU64 | BPF_ADD | BPF_X,
> +		rg[EBPF_REG_3], rg[EBPF_REG_0]);
> +
> +	/* R0 = R0 + R2 */
> +	emit_alu_reg(st, EBPF_ALU64 | BPF_ADD | BPF_X,
> +		rg[EBPF_REG_2], rg[EBPF_REG_0]);
> +
> +	/* JMP <fin_part> */
> +	emit_abs_jmp(st, ofs[LDMB_FIN_OFS]);
> +}
> +
> +/*
> + * helper function, used by emit_ld_mbuf().
> + * generates code for 'slow_path':
> + * call __rte_pktmbuf_read() and check return value.
> + */
> +static void
> +emit_ldmb_slow_path(struct bpf_jit_state *st, const uint32_t
> rg[EBPF_REG_7],
> +	uint32_t sz)
> +{
> +	/* make R3 contain *len* value (1/2/4) */
> +
> +	emit_mov_imm(st, EBPF_ALU64 | EBPF_MOV | BPF_K, rg[EBPF_REG_3], sz);
> +
> +	/* make R4 contain (RBP - ldmb.stack_ofs) */
> +
> +	emit_mov_reg(st, EBPF_ALU64 | EBPF_MOV | BPF_X, RBP, rg[EBPF_REG_4]);
> +	emit_alu_imm(st, EBPF_ALU64 | BPF_SUB | BPF_K, rg[EBPF_REG_4],
> +		st->ldmb.stack_ofs);
> +
> +	/* make R1 contain mbuf ptr */
> +
> +	emit_mov_reg(st, EBPF_ALU64 | EBPF_MOV | BPF_X,
> +		rg[EBPF_REG_6], rg[EBPF_REG_1]);
> +
> +	/* call rte_pktmbuf_read */
> +	emit_call(st, (uintptr_t)__rte_pktmbuf_read);
> +
> +	/* check that return value (R0) is not zero */
> +	emit_tst_reg(st, EBPF_ALU64, rg[EBPF_REG_0], rg[EBPF_REG_0]);
> +	emit_abs_jcc(st, BPF_JMP | BPF_JEQ | BPF_K, st->exit.off);
> +}
> +
> +/*
> + * helper function, used by emit_ld_mbuf().
> + * generates final part of code for BPF_ABS/BPF_IND load:
> + * perform data load and endianness conversion.
> + * expects dreg to contain valid data pointer.
> + */
> +static void
> +emit_ldmb_fin(struct bpf_jit_state *st, uint32_t dreg, uint32_t opsz,
> +	uint32_t sz)
> +{
> +	emit_ld_reg(st, BPF_LDX | BPF_MEM | opsz, dreg, dreg, 0);
> +	if (sz != sizeof(uint8_t))
> +		emit_be2le(st, dreg, sz * CHAR_BIT);
> +}
> +
> +/*
> + * emit code for BPF_ABS/BPF_IND load.
> + * generates the following construction:
> + * fast_path:
> + *   off = ins->sreg + ins->imm
> + *   if (off + ins->opsz < mbuf->data_len)
> + *      goto slow_path;

I am not an eBPF expert, but I am not sure this is correct.

I think it should be > instead of <. Also, it looks like you actually emit:
if (mbuf->data_len - off <= ins->opsz)
   goto slow_path;

Could you please double check that both the comment and the emitted code has the comparison turning the correct way.

> + *   ptr = mbuf->buf_addr + bmf->data_off + off;
> + *   goto fin_part;
> + * slow_path:
> + *   typeof(ins->opsz) buf; //allocate space on the stack
> + *   ptr = __rte_pktmbuf_read(mbuf, off, ins->opsz, &buf);
> + *   if (ptr == NULL)
> + *      goto exit_label;
> + * fin_part:
> + *   res = *(typeof(ins->opsz))ptr;
> + *   res = bswap(res);
> + */
> +static void
> +emit_ld_mbuf(struct bpf_jit_state *st, uint32_t op, uint32_t sreg,
> uint32_t imm)
> +{
> +	uint32_t i, mode, opsz, sz;
> +	uint32_t rg[EBPF_REG_7];
> +	int32_t ofs[LDMB_OFS_NUM];
> +
> +	mode = BPF_MODE(op);
> +	opsz = BPF_SIZE(op);
> +	sz = bpf_size(opsz);
> +
> +	for (i = 0; i != RTE_DIM(rg); i++)
> +		rg[i] = ebpf2x86[i];
> +
> +	/* fill with fake offsets */
> +	for (i = 0; i != RTE_DIM(ofs); i++)
> +		ofs[i] = st->sz + INT8_MAX;
> +
> +	/* dry run first to calculate jump offsets */
> +
> +	ofs[LDMB_FSP_OFS] = st->sz;
> +	emit_ldmb_fast_path(st, rg, sreg, mode, sz, imm, ofs);
> +	ofs[LDMB_SLP_OFS] = st->sz;
> +	emit_ldmb_slow_path(st, rg, sz);
> +	ofs[LDMB_FIN_OFS] = st->sz;
> +	emit_ldmb_fin(st, rg[EBPF_REG_0], opsz, sz);
> +
> +	RTE_VERIFY(ofs[LDMB_FIN_OFS] - ofs[LDMB_FSP_OFS] <= INT8_MAX);
> +
> +	/* reset dry-run code and do a proper run */
> +
> +	st->sz = ofs[LDMB_FSP_OFS];
> +	emit_ldmb_fast_path(st, rg, sreg, mode, sz, imm, ofs);
> +	emit_ldmb_slow_path(st, rg, sz);
> +	emit_ldmb_fin(st, rg[EBPF_REG_0], opsz, sz);
> +}
> +
>  static void
>  emit_prolog(struct bpf_jit_state *st, int32_t stack_size)
>  {
> @@ -1121,6 +1292,7 @@ emit(struct bpf_jit_state *st, const struct rte_bpf
> *bpf)
>  	/* reset state fields */
>  	st->sz = 0;
>  	st->exit.num = 0;
> +	st->ldmb.stack_ofs = bpf->stack_sz;
> 
>  	emit_prolog(st, bpf->stack_sz);
> 
> @@ -1240,6 +1412,15 @@ emit(struct bpf_jit_state *st, const struct rte_bpf
> *bpf)
>  			emit_ld_imm64(st, dr, ins[0].imm, ins[1].imm);
>  			i++;
>  			break;
> +		/* load absolute/indirect instructions */
> +		case (BPF_LD | BPF_ABS | BPF_B):
> +		case (BPF_LD | BPF_ABS | BPF_H):
> +		case (BPF_LD | BPF_ABS | BPF_W):
> +		case (BPF_LD | BPF_IND | BPF_B):
> +		case (BPF_LD | BPF_IND | BPF_H):
> +		case (BPF_LD | BPF_IND | BPF_W):
> +			emit_ld_mbuf(st, op, sr, ins->imm);
> +			break;
>  		/* store instructions */
>  		case (BPF_STX | BPF_MEM | BPF_B):
>  		case (BPF_STX | BPF_MEM | BPF_H):
> diff --git a/lib/librte_bpf/bpf_validate.c b/lib/librte_bpf/bpf_validate.c
> index fecdda0e1..9214f1503 100644
> --- a/lib/librte_bpf/bpf_validate.c
> +++ b/lib/librte_bpf/bpf_validate.c
> @@ -70,6 +70,7 @@ struct bpf_verifier {
>  	uint64_t stack_sz;
>  	uint32_t nb_nodes;
>  	uint32_t nb_jcc_nodes;
> +	uint32_t nb_ldmb_nodes;
>  	uint32_t node_colour[MAX_NODE_COLOUR];
>  	uint32_t edge_type[MAX_EDGE_TYPE];
>  	struct bpf_eval_state *evst;
> @@ -2020,6 +2021,14 @@ validate(struct bpf_verifier *bvf)
>  			rc |= add_edge(bvf, node, i + 2);
>  			i++;
>  			break;
> +		case (BPF_LD | BPF_ABS | BPF_B):
> +		case (BPF_LD | BPF_ABS | BPF_H):
> +		case (BPF_LD | BPF_ABS | BPF_W):
> +		case (BPF_LD | BPF_IND | BPF_B):
> +		case (BPF_LD | BPF_IND | BPF_H):
> +		case (BPF_LD | BPF_IND | BPF_W):
> +			bvf->nb_ldmb_nodes++;
> +			/* fallthrough */
>  		default:
>  			rc |= add_edge(bvf, node, i + 1);
>  			break;
> @@ -2320,8 +2329,14 @@ bpf_validate(struct rte_bpf *bpf)
>  	free(bvf.in);
> 
>  	/* copy collected info */
> -	if (rc == 0)
> +	if (rc == 0) {
>  		bpf->stack_sz = bvf.stack_sz;
> 
> +		/* for LD_ABS/LD_IND, we'll need extra space on the stack */
> +		if (bvf.nb_ldmb_nodes != 0)
> +			bpf->stack_sz = RTE_ALIGN_CEIL(bpf->stack_sz +
> +				sizeof(uint64_t), sizeof(uint64_t));
> +	}
> +
>  	return rc;
>  }
> --
> 2.17.1
>