Re: [PATCH bpf-next 1/4] bpf: add internal-only per-CPU LDX instructions

[Stanislav Fomichev <sdf@google.com>]

On 03/29, Andrii Nakryiko wrote:
> Add BPF instructions for working with per-CPU data. These instructions
> are internal-only and users are not allowed to use them directly. They
> will only be used for internal inlining optimizations for now.
> 
> Two different instructions are added. One, with BPF_MEM_PERCPU opcode,
> performs memory dereferencing of a per-CPU "address" (which is actually
> an offset). This one is useful when inlined logic needs to load data
> stored in per-CPU storage (bpf_get_smp_processor_id() is one such
> example).
> 
> Another, with BPF_ADDR_PERCPU opcode, performs a resolution of a per-CPU
> address (offset) stored in a register. This one is useful anywhere where
> per-CPU data is not read, but rather is returned to user as just
> absolute raw memory pointer (useful in bpf_map_lookup_elem() helper
> inlinings, for example).
> 
> BPF disassembler is also taught to recognize them to support dumping
> final BPF assembly code (non-JIT'ed version).
> 
> Add arch-specific way for BPF JITs to mark support for this instructions.
> 
> This patch also adds support for these instructions in x86-64 BPF JIT.
> 
> Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
> ---
>  arch/x86/net/bpf_jit_comp.c | 29 +++++++++++++++++++++++++++++
>  include/linux/filter.h      | 27 +++++++++++++++++++++++++++
>  kernel/bpf/core.c           |  5 +++++
>  kernel/bpf/disasm.c         | 33 ++++++++++++++++++++++++++-------
>  4 files changed, 87 insertions(+), 7 deletions(-)
> 
> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
> index 3b639d6f2f54..610bbedaae70 100644
> --- a/arch/x86/net/bpf_jit_comp.c
> +++ b/arch/x86/net/bpf_jit_comp.c
> @@ -1910,6 +1910,30 @@ st:			if (is_imm8(insn->off))
>  			}
>  			break;
>  
> +		/* internal-only per-cpu zero-extending memory load */
> +		case BPF_LDX | BPF_MEM_PERCPU | BPF_B:
> +		case BPF_LDX | BPF_MEM_PERCPU | BPF_H:
> +		case BPF_LDX | BPF_MEM_PERCPU | BPF_W:
> +		case BPF_LDX | BPF_MEM_PERCPU | BPF_DW:
> +			insn_off = insn->off;
> +			EMIT1(0x65); /* gs segment modifier */
> +			emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
> +			break;
> +
> +		/* internal-only load-effective-address-of per-cpu offset */
> +		case BPF_LDX | BPF_ADDR_PERCPU | BPF_DW: {
> +			u32 off = (u32)(void *)&this_cpu_off;
> +
> +			/* mov <dst>, <src> (if necessary) */
> +			EMIT_mov(dst_reg, src_reg);
> +
> +			/* add <dst>, gs:[<off>] */
> +			EMIT2(0x65, add_1mod(0x48, dst_reg));
> +			EMIT3(0x03, add_1reg(0x04, dst_reg), 0x25);
> +			EMIT(off, 4);
> +
> +			break;
> +		}
>  		case BPF_STX | BPF_ATOMIC | BPF_W:
>  		case BPF_STX | BPF_ATOMIC | BPF_DW:
>  			if (insn->imm == (BPF_AND | BPF_FETCH) ||
> @@ -3365,6 +3389,11 @@ bool bpf_jit_supports_subprog_tailcalls(void)
>  	return true;
>  }
>  
> +bool bpf_jit_supports_percpu_insns(void)
> +{
> +	return true;
> +}
> +
>  void bpf_jit_free(struct bpf_prog *prog)
>  {
>  	if (prog->jited) {
> diff --git a/include/linux/filter.h b/include/linux/filter.h
> index 44934b968b57..85ffaa238bc1 100644
> --- a/include/linux/filter.h
> +++ b/include/linux/filter.h
> @@ -75,6 +75,14 @@ struct ctl_table_header;
>  /* unused opcode to mark special load instruction. Same as BPF_MSH */
>  #define BPF_PROBE_MEM32	0xa0
>  
> +/* unused opcode to mark special zero-extending per-cpu load instruction. */
> +#define BPF_MEM_PERCPU	0xc0
> +
> +/* unused opcode to mark special load-effective-address-of instruction for
> + * a given per-CPU offset
> + */
> +#define BPF_ADDR_PERCPU	0xe0
> +
>  /* unused opcode to mark call to interpreter with arguments */
>  #define BPF_CALL_ARGS	0xe0
>  
> @@ -318,6 +326,24 @@ static inline bool insn_is_cast_user(const struct bpf_insn *insn)
>  		.off   = OFF,					\
>  		.imm   = 0 })
>  
> +/* Per-CPU zero-extending memory load (internal-only) */
> +#define BPF_LDX_MEM_PERCPU(SIZE, DST, SRC, OFF)			\
> +	((struct bpf_insn) {					\
> +		.code  = BPF_LDX | BPF_SIZE(SIZE) | BPF_MEM_PERCPU,\
> +		.dst_reg = DST,					\
> +		.src_reg = SRC,					\
> +		.off   = OFF,					\
> +		.imm   = 0 })
> +

[..]

> +/* Load effective address of a given per-CPU offset */

nit: mark this one as internal only as well in the comment?

(the change overall looks awesome, looking forward to trying it out)