svn commit: r361360 - head/sys/dev/hyperv/hvsock

[Kyle Evans]

On Fri, May 22, 2020 at 4:17 AM Wei Hu <whu at freebsd.org> wrote:
>
> Author: whu
> Date: Fri May 22 09:17:07 2020
> New Revision: 361360
> URL: https://svnweb.freebsd.org/changeset/base/361360
>
> Log:
>   Socket AF_HYPERV should return failure when it is not running on HyperV
>
> [... snip ...]
> @@ -354,6 +354,9 @@ hvs_trans_attach(struct socket *so, int proto, struct
>  {
>         struct hvs_pcb *pcb = so2hvspcb(so);
>
> +       if (vm_guest != VM_GUEST_HV)
> +               return (ESOCKTNOSUPPORT);
> +
>         HVSOCK_DBG(HVSOCK_DBG_VERBOSE,
>             "%s: HyperV Socket hvs_trans_attach called\n", __func__);
>

This one may be OK, but none of these other methods should be able to
be invoked if the attach failed. See this comment at around
^/sys/kern/uipc_socket.c#50:

50 * pru_detach() disassociates protocol layer state from an attached socket,
51 * and will be called exactly once for sockets in which pru_attach() has
52 * been successfully called.  If pru_attach() returned an error,
53 * pru_detach() will not be called.  Socket layer private.

That said, I wonder if VNET_DOMAIN_SET provides the correct semantics
here at all. You can't fail domain_init, but I don't think there's any
value in this domain actually getting registered on non-HyperV guests,
a fact which presumably won't change at runtime.

I have some concerns about how hvsock works out with vnets, but I'll
drop those on the initial commit thread.

Thanks,

Kyle Evans