[PATCH net v4] bpf, net: Check cgroup_bpf_enabled() only once in do_sock_getsockopt()

[Alexei Starovoitov]

On Thu, Aug 29, 2024 at 5:45 AM Tze-nan Wu (吳澤南)
<Tze-nan.Wu at mediatek.com> wrote:
>
> On Fri, 2024-08-23 at 19:04 -0700, Stanislav Fomichev wrote:
> >
> > External email : Please do not click links or open attachments until
> > you have verified the sender or the content.
> >  On 08/22, Alexei Starovoitov wrote:
> > > On Thu, Aug 22, 2024 at 12:02 AM Tze-nan Wu (吳澤南)
> > > <Tze-nan.Wu at mediatek.com> wrote:
> > > >
> > > >
> > > > BTW, If this should be handled in kernel, modification shown
> > below
> > > > could fix the issue without breaking the "static_branch" usage in
> > both
> > > > macros:
> > > >
> > > >
> > > > +++ /include/linux/bpf-cgroup.h:
> > > >     -#define BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN(optlen)
> > > >     +#define BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN(optlen, compat)
> > > >      ({
> > > >             int __ret = 0;
> > > >             if (cgroup_bpf_enabled(CGROUP_GETSOCKOPT))
> > > >                 copy_from_sockptr(&__ret, optlen, sizeof(int));
> > > >      +      else
> > > >      +          *compat = true;
> > > >             __ret;
> > > >      })
> > > >
> > > >     #define BPF_CGROUP_RUN_PROG_GETSOCKOPT(sock, level, optname,
> > > > optval, optlen, max_optlen, retval)
> > > >      ({
> > > >          int __ret = retval;
> > > >     -    if (cgroup_bpf_enabled(CGROUP_GETSOCKOPT) &&
> > > >     -        cgroup_bpf_sock_enabled(sock, CGROUP_GETSOCKOPT))
> > > >     +    if (cgroup_bpf_sock_enabled(sock, CGROUP_GETSOCKOPT))
> > > >              if (!(sock)->sk_prot->bpf_bypass_getsockopt ||
> > > >                ...
> > > >
> > > >   +++ /net/socket.c:
> > > >     int do_sock_getsockopt(struct socket *sock, bool compat, int
> > level,
> > > >      {
> > > >         ...
> > > >         ...
> > > >     +     /* The meaning of `compat` variable could be changed
> > here
> > > >     +      * to indicate if cgroup_bpf_enabled(CGROUP_SOCK_OPS)
> > is
> > > > false.
> > > >     +      */
> > > >         if (!compat)
> > > >     -       max_optlen =
> > BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN(optlen);
> > > >     +       max_optlen = BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN(optlen,
> > > > &compat);
> > >
> > > This is better, but it's still quite a hack. Let's not override it.
> > > We can have another bool, but the question:
> > > do we really need BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN  ?
> > > copy_from_sockptr(&__ret, optlen, sizeof(int));
> > > should be fast enough to do it unconditionally.
> > > What are we saving here?
> > >
> > > Stan ?
> >
> > Agreed, most likely nobody would notice :-)
>
> Sorry for my late reply, just have the mailer fixed.
>
> If it is feasible to make the `copy_from_sockptr` unconditionally,
> should I submit a new patch that resolve the issue by removing
> `BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN`? Patch A shown as below.
>
>   +++ /net/socket.c:
>    int do_sock_getsockopt(...)
>    {
>   -     int max_optlen __maybe_unused;
>   +     int max_optlen __maybe_unused = 0;
>         const struct proto_ops *ops;
>         int err;
>   ...
>   ...
>         if (!compat) <== wonder if we should keep the condition here?
>   -         max_optlen = BPF_CGROUP_GETSOCKOPT_MAX_OPTLEN(optlen);
>   +         copy_from_sockptr(&max_optlen, optlen, sizeof(int));

This one.
And delete the macro from bpf-cgroup.h