Re: [RFC] connectat()/bindat() or an alternative design

["John Ericson" <mail@johnericson.me>]

On Wed, Jul 1, 2026, at 8:32 PM, Cong Wang wrote:
> Hm? Why not just setsockopt()? Something like:
>
> struct unix_lookup_ctx {
>     int dirfd;
>     __u64 resolve_flags;   /* RESOLVE_BENEATH, RESOLVE_NO_SYMLINKS, etc. */
>     __u64 op_flags;        /* maybe future */
> };
>
> setsockopt(fd, SOL_UNIX, UNIX_NEXT_LOOKUP,
>            &ctx, sizeof(ctx));
>
> bind(fd, (struct sockaddr *)&addr, addrlen);
> /* or connect(fd, ...) */
>
>
> Zero new syscall is needed.

If this supports the `AT_EMPTY_PATH` no filesystem at all use-case
(which is what I chose to highlight in the previous summary email), I'll
take it --- it hits my desiderata. But implementation-side, I don't
relish the thought of even more state for socket setup --- since we need
to persist the lookup context until the bind.

If setsockopt is the "ioctl of sockets" --- a convenient way to de facto
create syscalls without doing so de jure --- I would think it would be
better to just make it do the bind in one go so there is no state:

    struct unix_bind_anon_ctx {
        int *connect_fd;       /* fd used to connect to this socket
                                  written to this */
        __u64 op_flags;        /* future usage */
    };

    setsockopt(fd, SOL_UNIX, UNIX_BIND_ANON,
               &ctx, sizeof(ctx));

Regular `linkat` can be used to place the "connect fd" in the filesystem
for regular file-based connecting (e.g. to cope with long paths), in
addition to how the "connect fd" can directly be used in the
`AT_EMPTY_PATH` case.

and for the connecting side:

    struct unix_connect_at_ctx {
        int dirfd;
        const char *path;
        __u64 resolve_flags;   /* RESOLVE_BENEATH, RESOLVE_NO_SYMLINKS, etc. */
        __u64 op_flags;        /* SOMETHING_EMPTY_PATH, other future usage */
    };

    setsockopt(fd, SOL_UNIX, UNIX_CONNECT_AT,
               &ctx, sizeof(ctx));

This is just what I had in my previous `summary email`, but with the new
system calls twisted into `setsockopt` variations instead. (I forgot to
mention the `linkat` usage in that email, oops, but it was always
intended to be part of the plan.)

Cheers,

John