Re: [PATCH bpf-next v6 1/2] bpf, x86: inline bpf_get_current_task() for x86_64

[Menglong Dong <menglong.dong@linux.dev>]

On 2026/1/21 11:10 Alexei Starovoitov <alexei.starovoitov@gmail.com> write:
> On Tue, Jan 20, 2026 at 5:58 PM Menglong Dong <menglong.dong@linux.dev> wrote:
> >
> > On 2026/1/21 09:23 Andrii Nakryiko <andrii.nakryiko@gmail.com> write:
> > > On Mon, Jan 19, 2026 at 11:06 PM Menglong Dong <menglong8.dong@gmail.com> wrote:
> > > >
> > > > Inline bpf_get_current_task() and bpf_get_current_task_btf() for x86_64
> > > > to obtain better performance.
> > > >
> > > > Signed-off-by: Menglong Dong <dongml2@chinatelecom.cn>
> > > > Acked-by: Eduard Zingerman <eddyz87@gmail.com>
> > > > ---
> > > > v5:
> > > > - don't support the !CONFIG_SMP case
> > > >
> > > > v4:
> > > > - handle the !CONFIG_SMP case
> > > >
> > > > v3:
> > > > - implement it in the verifier with BPF_MOV64_PERCPU_REG() instead of in
> > > >   x86_64 JIT.
> > > > ---
> > > >  kernel/bpf/verifier.c | 22 ++++++++++++++++++++++
> > > >  1 file changed, 22 insertions(+)
> > > >
> > > > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > > > index 9de0ec0c3ed9..c4e2ffadfb1f 100644
> > > > --- a/kernel/bpf/verifier.c
> > > > +++ b/kernel/bpf/verifier.c
> > > > @@ -17739,6 +17739,10 @@ static bool verifier_inlines_helper_call(struct bpf_verifier_env *env, s32 imm)
> > > >         switch (imm) {
> > > >  #ifdef CONFIG_X86_64
> > > >         case BPF_FUNC_get_smp_processor_id:
> > > > +#ifdef CONFIG_SMP
> > > > +       case BPF_FUNC_get_current_task_btf:
> > > > +       case BPF_FUNC_get_current_task:
> > > > +#endif
> > >
> > > Does this have to be x86-64 specific inlining? With verifier inlining
> > > and per_cpu instruction support it should theoretically work across
> > > all architectures that do support per-cpu instruction, no?
> > >
> > > Eduard pointed out [0] to me for why we have that x86-64 specific
> > > check. But looking at do_misc_fixups(), we have that early
> > > bpf_jit_inlines_helper_call(insn->imm)) check, so if some JIT has more
> > > performant inlining implementation, we will just do that.
> > >
> > > So it seems like we can just drop all that x86-64 specific logic and
> > > claim all three of these functions as inlinable, no?
> > >
> > > And even more. We can drop rather confusing
> > > verifier_inlines_helper_call() that duplicates the decision of which
> > > helpers can be inlined or not, and have:
> >
> > The verifier_inlines_helper_call() is confusing, but I think we can't
> > remove the x86-64 checking. For example, some architecture
> > don't support BPF_FUNC_get_current_task both in
> > bpf_jit_inlines_helper_call() and verifier_inlines_helper_call(), which
> > means it can't be inline.
> >
> > >
> > > if (env->prog->jit_requested && bpf_jit_supports_percpu_insn() {
> > >     switch (insn->imm) {
> > >     case BPF_FUNC_get_smp_processor_id:
> > >         ...
> > >         break;
> > >     case BPF_FUNC_get_current_task_btf:
> > >     case BPF_FUNC_get_current_task_btf:
> > >         ...
> > >         break;
> > >     default:
> > > }
> > >
> > > And the decision about inlining will live in one place.
> > >
> > > Or am I missing some complications?
> >
> > As Alexei said, the implement of "current" is architecture specific,
> > and the per-cpu variable "current_task" only exist on x86_64.
> >
> > >
> > > And with all that, should we mark get_current_task and
> > > get_current_task_btf as __bpf_fastcall?
> >
> > I think it make sense, and the I saw bpf_get_smp_processor_id does
> > such operation:
> >
> > const struct bpf_func_proto bpf_get_smp_processor_id_proto = {
> >         [...]
> >         .allow_fastcall = true,
> > };
> >
> > PS: I'm a little confused about the fast call. We inline many helper,
> > but it seems that bpf_get_smp_processor_id is the only one that
> > use the "allow_fastcall". Why? I'd better study harder.
> 
> It's
> static __bpf_fastcall __u32 (* const bpf_get_smp_processor_id)(void) =
> (void *) 8;
> 
> and
> #define __bpf_fastcall __attribute__((bpf_fastcall))

Ah, I see. It seems that the bpf_doc.py does the trick.

> 
> which makes LLVM use more registers at the callsite (less spill/fill).
> 
> Looking at the patch again. I think it's fine as-is.
> fastcall can be a follow up.

Okay!

Thanks!
Menglong Dong