Re: [RFC PATCH bpf-next 6/7] sched_ext: Make SCX use BPF capabilities

[Alexei Starovoitov]

On Thu, Jan 16, 2025 at 11:47 AM Juntong Deng <juntong.deng@xxxxxxxxxxx> wrote:
>
> This patch modifies SCX to use BPF capabilities.
>
> Make all SCX kfuncs register to BPF capabilities instead of
> BPF_PROG_TYPE_STRUCT_OPS.
>
> Add bpf_scx_bpf_capabilities_adjust as bpf_capabilities_adjust
> callback function.
>
> Signed-off-by: Juntong Deng <juntong.deng@xxxxxxxxxxx>
> ---
>  kernel/sched/ext.c | 74 ++++++++++++++++++++++++++++++++++++++--------
>  1 file changed, 62 insertions(+), 12 deletions(-)
>
> diff --git a/kernel/sched/ext.c b/kernel/sched/ext.c
> index 7fff1d045477..53cc7c3ed80b 100644
> --- a/kernel/sched/ext.c
> +++ b/kernel/sched/ext.c
> @@ -5765,10 +5765,66 @@ bpf_scx_get_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
>         }
>  }

'capabilities' name doesn't fit.
The word already has its meaning in the kernel.
It cannot be reused for a different purpose.

> +static int bpf_scx_bpf_capabilities_adjust(unsigned long *bpf_capabilities,
> +                                          u32 context_info, bool enter)
> +{
> +       if (enter) {
> +               switch (context_info) {
> +               case offsetof(struct sched_ext_ops, select_cpu):
> +                       ENABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_SELECT_CPU);
> +                       ENABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_ENQUEUE);
> +                       break;
> +               case offsetof(struct sched_ext_ops, enqueue):
> +                       ENABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_ENQUEUE);
> +                       break;
> +               case offsetof(struct sched_ext_ops, dispatch):
> +                       ENABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_DISPATCH);
> +                       break;
> +               case offsetof(struct sched_ext_ops, running):
> +               case offsetof(struct sched_ext_ops, stopping):
> +               case offsetof(struct sched_ext_ops, enable):
> +                       ENABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_REST);
> +                       break;
> +               case offsetof(struct sched_ext_ops, init):
> +               case offsetof(struct sched_ext_ops, exit):
> +                       ENABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_UNLOCKED);
> +                       break;
> +               default:
> +                       return -EINVAL;
> +               }
> +       } else {
> +               switch (context_info) {
> +               case offsetof(struct sched_ext_ops, select_cpu):
> +                       DISABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_SELECT_CPU);
> +                       DISABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_ENQUEUE);
> +                       break;
> +               case offsetof(struct sched_ext_ops, enqueue):
> +                       DISABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_ENQUEUE);
> +                       break;
> +               case offsetof(struct sched_ext_ops, dispatch):
> +                       DISABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_DISPATCH);
> +                       break;
> +               case offsetof(struct sched_ext_ops, running):
> +               case offsetof(struct sched_ext_ops, stopping):
> +               case offsetof(struct sched_ext_ops, enable):
> +                       DISABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_REST);
> +                       break;
> +               case offsetof(struct sched_ext_ops, init):
> +               case offsetof(struct sched_ext_ops, exit):
> +                       DISABLE_BPF_CAPABILITY(bpf_capabilities, BPF_CAP_SCX_KF_UNLOCKED);
> +                       break;
> +               default:
> +                       return -EINVAL;
> +               }
> +       }
> +       return 0;
> +}

and this callback defeats the whole point of u32 bitmask.

In earlier patch
env->context_info = __btf_member_bit_offset(t, member) / 8; // moff

is also wrong.
The context_info name is too generic and misleading.
and 'env' isn't a right place to save moff.

Let's try to implement what was discussed earlier:

1
After successful check_struct_ops_btf_id() save moff in
prog->aux->attach_st_ops_member_off.

2
Add .filter callback to sched-ext kfunc registration path and
let it allow/deny kfuncs based on st_ops attach point.

3
Remove scx_kf_allow() and current->scx.kf_mask.

That will be a nice perf win and will prove that
this approach works end-to-end.