On Fri, 22 Jan 2021 09:30:27 +0100 Ivan Vecera wrote:
> On Thu, 21 Jan 2021 18:34:52 -0800
> Jakub Kicinski <k...@kernel.org> wrote:
>
> > On Thu, 21 Jan 2021 11:29:37 +0100 Ivan Vecera wrote:
> > > On Wed, 20 Jan 2021 15:18:20 -0800
> > > Cong Wang <xiyou.wangc...@gmail.com> wrote:
> > > > On Wed, Jan 20, 2021 at 4:56 AM Ivan Vecera <ivec...@redhat.com> wrote:
> > > >
> > > > > Team driver protects port list traversal by its team->lock mutex
> > > > > in functions like team_change_mtu(), team_set_rx_mode(),
> >
> > The set_rx_mode part can't be true, set_rx_mode can't sleep and
> > team->lock is a mutex.
> >
> > > > > To fix the problem __team_compute_features() needs to be postponed
> > > > > for these cases.
> > > >
> > > > Is there any user-visible effect after deferring this feature change?
> > > >
> > >
> > > An user should not notice this change.
> >
> > I think Cong is right, can you expand a little on your assertion?
> > User should be able to assume that the moment syscall returns the
> > features had settled.
> >
> > What does team->mutex actually protect in team_compute_features()?
> > All callers seem to hold RTNL at a quick glance. This is a bit of
> > a long shot but isn't it just tryin to protect the iteration over
> > ports which could be under RCU?
>
> In fact the mutex could be removed at all because all port-list
> writers are running under rtnl_lock, some readers like team_change_mtu()
> or team_device_event() [notifier] as well and hot path readers are
> protected by RCU.
> I have discussed this with Jiri but he don't want to introduce any dependency
> on RTNL to team as it was designed as RTNL-independent from beginning.
>
> Anyway your idea to run team_compute_features under RCU could be fine
> as subsequent __team_compute_features() cannot sleep...
>
> Do you mean something like this?
>
> diff --git a/drivers/net/team/team.c b/drivers/net/team/team.c
> index c19dac21c468..dd7917cab2b1 100644
> --- a/drivers/net/team/team.c
> +++ b/drivers/net/team/team.c
> @@ -992,7 +992,8 @@ static void __team_compute_features(struct team *team)
> unsigned int dst_release_flag = IFF_XMIT_DST_RELEASE |
> IFF_XMIT_DST_RELEASE_PERM;
>
> - list_for_each_entry(port, &team->port_list, list) {
> + rcu_read_lock();
> + list_for_each_entry_rcu(port, &team->port_list, list) {
> vlan_features = netdev_increment_features(vlan_features,
> port->dev->vlan_features,
> TEAM_VLAN_FEATURES);
> @@ -1006,6 +1007,7 @@ static void __team_compute_features(struct team *team)
> if (port->dev->hard_header_len > max_hard_header_len)
> max_hard_header_len = port->dev->hard_header_len;
> }
> + rcu_read_unlock();
>
> team->dev->vlan_features = vlan_features;
> team->dev->hw_enc_features = enc_features | NETIF_F_GSO_ENCAP_ALL |
> @@ -1020,9 +1022,7 @@ static void __team_compute_features(struct team *team)
>
> static void team_compute_features(struct team *team)
> {
> - mutex_lock(&team->lock);
> __team_compute_features(team);
> - mutex_unlock(&team->lock);
> netdev_change_features(team->dev);
> }
Yup, like this, but if Jiri doesn't like it then I guess we need to
come up with something else?
How about doing the work on unlock? Have some bit set when we had to
defer and then run __team_compute_features() before releasing the lock
for real?
