Re: [PATCH bpf-next v2 01/11] LoongArch: BPF: Fix tail call count pointer offset for arena programs

[Tiezhu Yang]

On 2026/7/7 下午5:43, George Guo wrote:
> Hi Tiezhu,
> Thanks for reviewing!
>
> > This introduces a runtime helper to dynamically calculate the stack
> > offset. If another optional slot is introduced in the future, this
> > helper will need further adjustments.
>
> Fair -- that's exactly the tradeoff. This patch was written to be the
> minimal, obviously-correct fix on top of the existing layout (arena slot
> above tcc/tcc_ptr) so it could go to the bpf tree as a standalone,
> backportable fix. It does mean any future optional slot needs another
> "if (...)" in the helper, as you say.
>
> > Please see the following simple and long-term solution, just move
> > the arena slot below TCC context slots, this ensures a constant
> > offset, reducing complex runtime calculations and eliminating extra
> > instructions.
> >
> >    LoongArch: BPF: Optimize redundant TCC loads in epilogue
> >    https://lore.kernel.org/loongarch/20260630022705.11536-2-yangtiezhu@xxxxxxxxxxx/
> >
> >    LoongArch: BPF: Move arena register slot below TCC context
> >    https://lore.kernel.org/loongarch/20260630022705.11536-3-yangtiezhu@xxxxxxxxxxx/
>
> I read through both -- the reordering is a nice fix: with the arena slot
> moved below tcc_ptr instead of above it, tcc_ptr's offset from $fp stays
> a fixed 80 regardless of arena, so the original hardcoded macro is
> correct again everywhere and the runtime branch goes away entirely.
> Agreed that's the better long-term shape.
>
> One thing to work out: this series (patch 5, exceptions/bpf_throw) also
> needs an optional extra slot for exception programs, using the same
> mechanism as arena (it currently piggybacks on the same
> `ctx->arena_vm_start || is_exception_prog` check your patch 2/4 would
> replace). If your reordering lands, I'd extend it the same way --
> moving the exception-program slot below tcc_ptr too, rather than adding
> a runtime check back in.
>
> Where does your series stand? I'd rather rebase this patch series on
> top of your fixed-offset approach once it's in bpf-next than have both
> land and immediately conflict. Let me know what's the best way to
> sequence these.

I am currently addressing the comments from the CI bot and
will send out v3 shortly.

Once this series is upstream, you can rebase your patches
on top of it.

Thanks,
Tiezhu