Re: [PATCH bpf-next v5 07/11] bpf: Add instructions for atomic_[cmp]xchg

[Yonghong Song]

On 12/15/20 4:18 AM, Brendan Jackman wrote:
> This adds two atomic opcodes, both of which include the BPF_FETCH
> flag. XCHG without the BPF_FETCH flag would naturally encode
> atomic_set. This is not supported because it would be of limited
> value to userspace (it doesn't imply any barriers). CMPXCHG without
> BPF_FETCH woulud be an atomic compare-and-write. We don't have such
> an operation in the kernel so it isn't provided to BPF either.
>
> There are two significant design decisions made for the CMPXCHG
> instruction:
>
>   - To solve the issue that this operation fundamentally has 3
>     operands, but we only have two register fields. Therefore the
>     operand we compare against (the kernel's API calls it 'old') is
>     hard-coded to be R0. x86 has similar design (and A64 doesn't
>     have this problem).
>
>     A potential alternative might be to encode the other operand's
>     register number in the immediate field.
>
>   - The kernel's atomic_cmpxchg returns the old value, while the C11
>     userspace APIs return a boolean indicating the comparison
>     result. Which should BPF do? A64 returns the old value. x86 returns
>     the old value in the hard-coded register (and also sets a
>     flag). That means return-old-value is easier to JIT, so that's
>     what we use.
>
> Signed-off-by: Brendan Jackman <jackmanb@xxxxxxxxxx>

Ack with a minor comment below.

Acked-by: Yonghong Song <yhs@xxxxxx>

> ---
>   arch/x86/net/bpf_jit_comp.c    |  8 ++++++++
>   include/linux/filter.h         |  2 ++
>   include/uapi/linux/bpf.h       |  4 +++-
>   kernel/bpf/core.c              | 20 ++++++++++++++++++++
>   kernel/bpf/disasm.c            | 15 +++++++++++++++
>   kernel/bpf/verifier.c          | 19 +++++++++++++++++--
>   tools/include/linux/filter.h   |  2 ++
>   tools/include/uapi/linux/bpf.h |  4 +++-
>   8 files changed, 70 insertions(+), 4 deletions(-)
>
> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
> index eea7d8b0bb12..308241187582 100644
> --- a/arch/x86/net/bpf_jit_comp.c
> +++ b/arch/x86/net/bpf_jit_comp.c
> @@ -815,6 +815,14 @@ static int emit_atomic(u8 **pprog, u8 atomic_op,
>   		/* src_reg = atomic_fetch_add(dst_reg + off, src_reg); */
>   		EMIT2(0x0F, 0xC1);
>   		break;
> +	case BPF_XCHG:
> +		/* src_reg = atomic_xchg(dst_reg + off, src_reg); */
> +		EMIT1(0x87);
> +		break;
> +	case BPF_CMPXCHG:
> +		/* r0 = atomic_cmpxchg(dst_reg + off, r0, src_reg); */
> +		EMIT2(0x0F, 0xB1);
> +		break;
>   	default:
>   		pr_err("bpf_jit: unknown atomic opcode %02x\n", atomic_op);
>   		return -EFAULT;
> diff --git a/include/linux/filter.h b/include/linux/filter.h
> index c3e87a63e0b8..16e0ba5e8937 100644
> --- a/include/linux/filter.h
> +++ b/include/linux/filter.h
> @@ -265,6 +265,8 @@ static inline bool insn_is_zext(const struct bpf_insn *insn)
>    *
>    *   BPF_ADD                  *(uint *) (dst_reg + off16) += src_reg
>    *   BPF_ADD | BPF_FETCH      src_reg = atomic_fetch_add(dst_reg + off16, src_reg);
> + *   BPF_XCHG                 src_reg = atomic_xchg(dst_reg + off16, src_reg)
> + *   BPF_CMPXCHG              r0 = atomic_cmpxchg(dst_reg + off16, r0, src_reg)
>    */
>   
>   #define BPF_ATOMIC_OP(SIZE, OP, DST, SRC, OFF)			\
> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> index 760ae333a5ed..538b95472c8f 100644
> --- a/include/uapi/linux/bpf.h
> +++ b/include/uapi/linux/bpf.h
> @@ -45,7 +45,9 @@
>   #define BPF_EXIT	0x90	/* function return */
>   
>   /* atomic op type fields (stored in immediate) */
> -#define BPF_FETCH	0x01	/* fetch previous value into src reg */
> +#define BPF_XCHG	(0xe0 | BPF_FETCH)	/* atomic exchange */
> +#define BPF_CMPXCHG	(0xf0 | BPF_FETCH)	/* atomic compare-and-write */
> +#define BPF_FETCH	0x01	/* not an opcode on its own, used to build others */

Although the above code works fine, I would suggest to put
BPF_FETCH definition before BPF_XCHG and BPF_CMPXCHG, which
makes more sense intuitively.

>   
>   /* Register numbers */
>   enum {
> diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
[...]		\
> diff --git a/tools/include/uapi/linux/bpf.h b/tools/include/uapi/linux/bpf.h
> index 760ae333a5ed..538b95472c8f 100644
> --- a/tools/include/uapi/linux/bpf.h
> +++ b/tools/include/uapi/linux/bpf.h
> @@ -45,7 +45,9 @@
>   #define BPF_EXIT	0x90	/* function return */
>   
>   /* atomic op type fields (stored in immediate) */
> -#define BPF_FETCH	0x01	/* fetch previous value into src reg */
> +#define BPF_XCHG	(0xe0 | BPF_FETCH)	/* atomic exchange */
> +#define BPF_CMPXCHG	(0xf0 | BPF_FETCH)	/* atomic compare-and-write */
> +#define BPF_FETCH	0x01	/* not an opcode on its own, used to build others */

same here.

>   
>   /* Register numbers */
>   enum {