Received-SPF: pass (google.com: domain of linux-kernel+bounces-9405-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Message-ID: <aaa613cf32a940eb26e6713650555a1efe181ea9.camel@thedillows.org>
Subject: Re: [PATCH next] bonding: Extending LACP MUX State Machine to
 include a Collecting State.
From: David Dillow <dave@thedillows.org>
To: jay.vosburgh@canonical.com
Cc: aahila@google.com, andy@greyhouse.net, daniel@iogearbox.net,
        davem@davemloft.net, edumazet@google.com, herbert@gondor.apana.org.au,
        kuba@kernel.org, linux-kernel@vger.kernel.org, maheshb@google.com,
        martin.lau@kernel.org, netdev@vger.kernel.org, pabeni@redhat.com
Date: Thu, 21 Dec 2023 21:38:01 -0800
In-Reply-To: <21529.1703184528@famine>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
User-Agent: Evolution 3.50.0-1 
Precedence: bulk
MIME-Version: 1.0

> I haven't read the patch in detail yet, but my overall question
> is: why do we need this?  This adds significant complexity to the
> state machine logic.  What real problem is this solving, i.e., what
> examples do you have of systems where a port is "in a state where
> it can receive incoming packets while not still distributing"?

Any time we add a new link to an aggregator, or the bond selects a new
aggegrator based on the selection policy, there is currently a race
where we start distributing traffic before our partner (usually a
switch) is ready to start collecting it, leading to dropped packets if
we're running traffic over the bond. We reliably hit this window,
making what should be a non-issue into a customer-visible packet-loss
event. Implementing the full state machine closes the window and makes
these maintenance events lossless.