Re: [PATCH v2] bpf: restore the ebpf program ID for BPF_AUDIT_UNLOAD and PERF_BPF_EVENT_PROG_UNLOAD

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Fri, Dec 23, 2022 at 5:49 PM Stanislav Fomichev <sdf@google.com> wrote:
get_func_ip() */
> > -                               tstamp_type_access:1; /* Accessed __sk_buff->tstamp_type */
> > +                               tstamp_type_access:1, /* Accessed __sk_buff->tstamp_type */
> > +                               valid_id:1; /* Is bpf_prog::aux::__id valid? */
> >         enum bpf_prog_type      type;           /* Type of BPF program */
> >         enum bpf_attach_type    expected_attach_type; /* For some prog types */
> >         u32                     len;            /* Number of filter blocks */
> > @@ -1688,6 +1689,12 @@ void bpf_prog_inc(struct bpf_prog *prog);
> >  struct bpf_prog * __must_check bpf_prog_inc_not_zero(struct bpf_prog *prog);
> >  void bpf_prog_put(struct bpf_prog *prog);
> >
> > +static inline u32 bpf_prog_get_id(const struct bpf_prog *prog)
> > +{
> > +       if (WARN(!prog->valid_id, "Attempting to use an invalid eBPF program"))
> > +               return 0;
> > +       return prog->aux->__id;
> > +}
>
> I'm still missing why we need to have this WARN and have a check at all.
> IIUC, we're actually too eager in resetting the id to 0, and need to
> keep that stale id around at least for perf/audit.
> Why not have a flag only to protect against double-idr_remove
> bpf_prog_free_id and keep the rest as is?
> Which places are we concerned about that used to report id=0 but now
> would report stale id?

What double-idr_remove are you concerned about?
bpf_prog_by_id() is doing bpf_prog_inc_not_zero
while __bpf_prog_put just dropped it to zero.

Maybe just move bpf_prog_free_id() into bpf_prog_put_deferred()
after perf_event_bpf_event and bpf_audit_prog ?
Probably can remove the obsolete do_idr_lock bool flag as
separate patch?

Much simpler fix and no code churn.
Both valid_id and saved_id approaches have flaws.