Re: [RFC net-next] net: dsa: rtl8366rb: support bridge offloading

[Linus Walleij]

On Tue, Mar 2, 2021 at 5:11 PM Vladimir Oltean <olteanv@xxxxxxxxx> wrote:
> On Tue, Mar 02, 2021 at 05:05:00PM +0100, Linus Walleij wrote:
> > On Tue, Mar 2, 2021 at 4:58 AM DENG Qingfang <dqfext@xxxxxxxxx> wrote:
> > > On Mon, Mar 1, 2021 at 9:48 PM Linus Walleij <linus.walleij@xxxxxxxxxx> wrote:
> > > > With my minor changes:
> > > > Tested-by: Linus Walleij <linus.walleij@xxxxxxxxxx>
> > >
> > > How about using a mutex lock in port_bridge_{join,leave} ?
> > > In my opinion all functions that access multiple registers should be
> > > synchronized.
> >
> > That's one way, in some cases the framework (DSA) serialize
> > the accesses so I don't know if that already happens on a
> > higher level? Since it is accessed over a slow bus we should go
> > for mutex in that case indeed.
>
> DSA does not serialize this. The .port_bridge_join and
> .port_bridge_leave calls are initiated from the NETDEV_CHANGEUPPER net
> device event, which is called under rtnl_mutex (see call_netdevice_notifiers).
> This is pretty fundamental and I don't think it will ever change.
>
> However, if you still want to add an extra layer of locking (with code
> paths that for some reason are not under the rtnl_mutex), then go ahead,
> I suppose. It will be challenging to make sure they do something that
> isn't snake oil, though.

Nah, just didn't know if was already in place.

I suggest Qingfang go with a driver-local mutex (it may already be needed in
more places).

Yours,
Linus Walleij