On Tue, Dec 20, 2016 at 03:28:38PM +0100, Martin Pieuchot wrote:
> I don't have a solution for the moment and I want to be sure we know all
Me neither. That's why I did not send a diff, but only the stack trace.
> recursions before trying to write a fix. So here's a diff that mark the
> recursions with a XXXSMP like in the NFS case.
I think this approach is fine.
> ok?
You missed the returns in pflowioctl() between rw_exit_write() and
rw_enter_write(). I would suggest that you follow the splnet() and
put a rw_enter_write() after each splx().
bluhm
>
> Index: net/if_pflow.c
> ===================================================================
> RCS file: /cvs/src/sys/net/if_pflow.c,v
> retrieving revision 1.62
> diff -u -p -r1.62 if_pflow.c
> --- net/if_pflow.c 4 Oct 2016 13:54:32 -0000 1.62
> +++ net/if_pflow.c 20 Dec 2016 14:23:32 -0000
> @@ -267,7 +267,10 @@ pflow_clone_destroy(struct ifnet *ifp)
> pflow_flush(sc);
> m_freem(sc->send_nam);
> if (sc->so != NULL) {
> + /* XXXSMP breaks atomicity */
> + rw_exit_write(&netlock);
> error = soclose(sc->so);
> + rw_enter_write(&netlock);
> sc->so = NULL;
> }
> if (sc->sc_flowdst != NULL)
> @@ -375,6 +378,8 @@ pflowioctl(struct ifnet *ifp, u_long cmd
> }
> }
>
> + /* XXXSMP breaks atomicity */
> + rw_exit_write(&netlock);
> s = splnet();
> pflow_flush(sc);
>
> @@ -530,6 +535,7 @@ pflowioctl(struct ifnet *ifp, u_long cmd
> } else
> ifp->if_flags &= ~IFF_RUNNING;
>
> + rw_enter_write(&netlock);
> break;
>
> default:
