Re: [net-next v5 1/4] net: af_unix: enable custom setsockopt for all socket types

[Jori Koolstra]

> Op 14-07-2026 11:19 CEST schreef Kuniyuki Iwashima <kuniyu@xxxxxxxxxx>:
> 
>  
> On Sun, Jul 12, 2026 at 9:29 PM Jori Koolstra <jkoolstra@xxxxxxxxx> wrote:
> >
> > unix_setsockopt() and the SOCK_CUSTOM_SOCKOPT flag were only wired up
> > for SOCK_STREAM (introduced along with the stream-only SO_INQ).
> > Consequently custom AF_UNIX options are unreachable on SOCK_DGRAM and
> > SOCK_SEQPACKET: those setsockopt() calls bypass unix_setsockopt() and
> > fall through to the generic sock_setsockopt(), failing with
> > -ENOPROTOOPT.
> >
> > Set SOCK_CUSTOM_SOCKOPT for every AF_UNIX socket type in unix_create(), and
> > also for accepted sockets in unix_accept() (reachable for stream and
> > seqpacket).
> >
> > This is a prerequisite for making SO_RIGHTS_NOTRUNC settable on all AF_UNIX
> > socket types.
> >
> > Signed-off-by: Jori Koolstra <jkoolstra@xxxxxxxxx>
> > ---
> >  net/unix/af_unix.c | 10 ++++++----
> >  1 file changed, 6 insertions(+), 4 deletions(-)
> >
> > diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
> > index f7a9d55eee8a..3d256255085b 100644
> > --- a/net/unix/af_unix.c
> > +++ b/net/unix/af_unix.c
> > @@ -949,7 +949,7 @@ static int unix_setsockopt(struct socket *sock, int level, int optname,
> >         switch (optname) {
> >         case SO_INQ:
> >                 if (sk->sk_type != SOCK_STREAM)
> > -                       return -EINVAL;
> > +                       return -ENOPROTOOPT;
> >
> >                 if (val > 1 || val < 0)
> >                         return -EINVAL;
> > @@ -1005,6 +1005,7 @@ static const struct proto_ops unix_dgram_ops = {
> >  #endif
> >         .listen =       sock_no_listen,
> >         .shutdown =     unix_shutdown,
> > +       .setsockopt =   unix_setsockopt,
> >         .sendmsg =      unix_dgram_sendmsg,
> >         .read_skb =     unix_read_skb,
> >         .recvmsg =      unix_dgram_recvmsg,
> > @@ -1029,6 +1030,7 @@ static const struct proto_ops unix_seqpacket_ops = {
> >  #endif
> >         .listen =       unix_listen,
> >         .shutdown =     unix_shutdown,
> > +       .setsockopt =   unix_setsockopt,
> >         .sendmsg =      unix_seqpacket_sendmsg,
> >         .recvmsg =      unix_seqpacket_recvmsg,
> >         .mmap =         sock_no_mmap,
> > @@ -1142,9 +1144,10 @@ static int unix_create(struct net *net, struct socket *sock, int protocol,
> >         if (protocol && protocol != PF_UNIX)
> >                 return -EPROTONOSUPPORT;
> >
> > +       set_bit(SOCK_CUSTOM_SOCKOPT, &sock->flags);
> 
> Please move this to unix_create1(), then we don't need the chunk below.
> 
> 

I don't understand how this helps. The accept() path gets its struct socket from
sock_alloc() not via unix_create() afaict. Also, unix_create1() is called with
sock == NULL on connect(), so now you have to guard against that too.

But maybe I misunderstand you?

I do agree with you other comment, accept() should probably inherit the truncate
behavior from the listen() socket.

Thanks,
Jori.