Re: [PATCH bpf-next v2 4/9] bpf: Introduce load-acquire and store-release instructions

[Peilin Ye]

On Fri, Feb 14, 2025 at 07:04:13PM -0800, Alexei Starovoitov wrote:
> > > > How about:
> > > > #define BPF_LOAD_ACQ 2
> > > > #define BPF_STORE_REL 3
> > > >
> > > > and only use them with BPF_MOV like
> > > >
> > > > imm = BPF_MOV | BPF_LOAD_ACQ - is actual load acquire
> > > > imm = BPF_MOV | BPF_STORE_REL - release
> >
> > Based on everything discussed, should we proceed with the above
> > suggestion?  Specifically:
> >
> >   #define BPF_LD_ST     BPF_MOV /* 0xb0 */
> 
> The aliasing still bothers me.
> I hated doing it when going from cBPF to eBPF,
> but we only had 8-bit to work with.
> Here we have 32-bit.
> Aliases make disassemblers trickier, since value no longer
> translates to string as-is. It depends on the context.
> There is probably no use for BPF_MOV operation in atomic
> context, but by reusing BPF_ADD, BPF_XOR, etc in atomic
> we signed up ourselves for all of alu ops.

I see.

> That's why BPF_XCHG and BPF_CMPXCHG are outside
> of alu op range.
> 
> So my preference is to do:
> #define BPF_LOAD_ACQ 0x100
> #define BPF_STORE_REL 0x110
> #define BPF_CMPWAIT_RELAXED   0x120
>
> and keep growing it.
> We burn the first nibble, but so be it.

Got it!  In instruction-set.rst I'll make it clear that imm<0-3> must be
zero for load_acq/store_rel.

Cheers,
Peilin Ye