On 9/30/19 7:23 PM, David Ahern wrote:
> On 9/30/19 8:01 PM, Eric Dumazet wrote:
>>
>>
>> Do we need to keep the test on IF_READY done later in this function ?
>>
>> If IF_READY can disappear only under RTNL, we might clean this.
>>
>> (unless addrconf_dad_work() releases rtnl and reacquires it)
>
> Unless I am missing something none of the functions called by dad_work
> release the rtnl, but your comment did have me second guessing the locking.
>
> The interesting cases for changing the idev flag are addrconf_notify
> (NETDEV_UP and NETDEV_CHANGE) and addrconf_ifdown (reset the flag). The
> former does not have the idev lock - only rtnl. The latter has both.
> Checking the flag is inconsistent with respect to locks.
>
> As for your suggestion, the 'dead' flag is set only under rtnl in
> addrconf_ifdown and it means the device is getting removed (or IPv6 is
> disabled). Based on that I think the existing:
>
> if (idev->dead || !(idev->if_flags & IF_READY))
> goto out;
>
> can be moved to right after the rtnl_lock in addrconf_dad_work in place
> of the above change, so the end result is:
>
>
> diff --git a/net/ipv6/addrconf.c b/net/ipv6/addrconf.c
> index 6a576ff92c39..dd3be06d5a06 100644
> --- a/net/ipv6/addrconf.c
> +++ b/net/ipv6/addrconf.c
> @@ -4032,6 +4032,12 @@ static void addrconf_dad_work(struct work_struct *w)
>
> rtnl_lock();
>
> + /* check if device was taken down before this delayed work
> + * function could be canceled
> + */
> + if (idev->dead || !(idev->if_flags & IF_READY))
> + goto out;
> +
> spin_lock_bh(&ifp->lock);
> if (ifp->state == INET6_IFADDR_STATE_PREDAD) {
> action = DAD_BEGIN;
> @@ -4077,11 +4083,6 @@ static void addrconf_dad_work(struct work_struct *w)
> goto out;
>
> write_lock_bh(&idev->lock);
> - if (idev->dead || !(idev->if_flags & IF_READY)) {
> - write_unlock_bh(&idev->lock);
> - goto out;
> - }
> -
> spin_lock(&ifp->lock);
> if (ifp->state == INET6_IFADDR_STATE_DEAD) {
> spin_unlock(&ifp->lock);
>
>
> agree?
>
SGTM, thanks !
