On Fri, Aug 30, 2024 at 5:11 AM Leon Hwang <leon.hwang@xxxxxxxxx> wrote:
>
>
>
> On 2024/8/30 18:00, Xu Kuohai wrote:
> > On 8/30/2024 5:08 PM, Leon Hwang wrote:
> >>
> >>
> >> On 30/8/24 15:37, Xu Kuohai wrote:
> >>> On 8/27/2024 10:23 AM, Leon Hwang wrote:
> >>>>
> >>
>
> [...]
>
> >>
> >> This approach is really cool!
> >>
> >> I want an alike approach on x86. But I failed. Because, on x86, it's an
> >> indirect call to "call *rdx", aka "bpf_func(ctx, insnsi)".
> >>
> >> Let us imagine the arch_run_bpf() on x86:
> >>
> >> unsigned int __naked arch_run_bpf(const void *ctx, const struct bpf_insn
> >> *insnsi, bpf_func_t bpf_func)
> >> {
> >> asm (
> >> "pushq %rbp\n\t"
> >> "movq %rsp, %rbp\n\t"
> >> "xor %rax, %rax\n\t"
> >> "pushq %rax\n\t"
> >> "movq %rsp, %rax\n\t"
> >> "callq *%rdx\n\t"
> >> "leave\n\t"
> >> "ret\n\t"
> >> );
> >> }
> >>
> >> If we can change "callq *%rdx" to a direct call, it'll be really
> >> wonderful to resolve this tailcall issue on x86.
> >>
> >
> > Right, so we need static call here, perhaps we can create a custom
> > static call trampoline to setup tail call counter.
> >
> >> How to introduce arch_bpf_run() for all JIT backends?
> >>
> >
> > Seems we can not avoid arch specific code. One approach could be
> > to define a default __weak function to call bpf_func directly,
> > and let each arch to provide its own overridden implementation.
> >
>
> Hi Xu Kuohai,
>
> Can you send a separate patch to fix this issue on arm64?
>
> After you fixing it, I'll send the patch to fix it on x64.
Hold on.
We're disabling freplace+tail_call in the verifier.
No need to change any JITs.
