Re: [PATCH net v2] bonding: fix feature flag setting at init time

From: Jarod Wilson <jarod@redhat.com>
To: Jay Vosburgh <jay.vosburgh@canonical.com>
Cc: LKML <linux-kernel@vger.kernel.org>,
	Ivan Vecera <ivecera@redhat.com>,
	Veaceslav Falico <vfalico@gmail.com>,
	Andy Gospodarek <andy@greyhouse.net>,
	"David S. Miller" <davem@davemloft.net>,
	Jakub Kicinski <kuba@kernel.org>, Thomas Davis <tadavis@lbl.gov>,
	Netdev <netdev@vger.kernel.org>
Subject: Re: [PATCH net v2] bonding: fix feature flag setting at init time
Date: Wed, 2 Dec 2020 15:54:16 -0500	[thread overview]
Message-ID: <CAKfmpScW_Ar91f1dd5QAeVmQa-G9c0qnJph+cBOJ_CoEGBoFpg@mail.gmail.com> (raw)
In-Reply-To: <21153.1606940246@famine>

On Wed, Dec 2, 2020 at 3:17 PM Jay Vosburgh <jay.vosburgh@canonical.com> wrote:
>
> Jarod Wilson <jarod@redhat.com> wrote:
>
> >On Wed, Dec 2, 2020 at 12:55 PM Jay Vosburgh <jay.vosburgh@canonical.com> wrote:
> >>
> >> Jarod Wilson <jarod@redhat.com> wrote:
> >>
> >> >Don't try to adjust XFRM support flags if the bond device isn't yet
> >> >registered. Bad things can currently happen when netdev_change_features()
> >> >is called without having wanted_features fully filled in yet. Basically,
> >> >this code was racing against register_netdevice() filling in
> >> >wanted_features, and when it got there first, the empty wanted_features
> >> >led to features also getting emptied out, which was definitely not the
> >> >intended behavior, so prevent that from happening.
> >>
> >>         Is this an actual race?  Reading Ivan's prior message, it sounds
> >> like it's an ordering problem (in that bond_newlink calls
> >> register_netdevice after bond_changelink).
> >
> >Sorry, yeah, this is not actually a race condition, just an ordering
> >issue, bond_check_params() gets called at init time, which leads to
> >bond_option_mode_set() being called, and does so prior to
> >bond_create() running, which is where we actually call
> >register_netdevice().
>
>         So this only happens if there's a "mode" module parameter?  That
> doesn't sound like the call path that Ivan described (coming in via
> bond_newlink).

Ah. I think there's actually two different pathways that can trigger
this. The first is for bonds created at module load time, which I was
describing, the second is for a new bond created via bond_newlink()
after the bonding module is already loaded, as described by Ivan. Both
have the problem of bond_option_mode_set() running prior to
register_netdevice(). Of course, that would suggest every bond
currently comes up with unintentionally neutered flags, which I
neglected to catch in earlier testing and development.

-- 
Jarod Wilson
jarod@redhat.com