Date:	Fri, 21 Jun 2013 14:03:20 +0200
From:	Michal Kubecek <mkubecek@...e.cz>
To:	Veaceslav Falico <vfalico@...hat.com>
Cc:	netdev@...r.kernel.org, fubar@...ibm.com, andy@...yhouse.net,
	davem@...emloft.net, linux@...2.net, nicolas.2p.debian@...e.fr,
	rick.jones2@...com, nikolay@...hat.com
Subject: Re: [PATCH v2 net-next 6/6] bonding: add an option to fail when any
 of arp_ip_target is inaccessible

On Fri, Jun 21, 2013 at 01:00:31PM +0200, Veaceslav Falico wrote:
> On Fri, Jun 21, 2013 at 12:23:18PM +0200, Michal Kubecek wrote:
> >On Thu, Jun 20, 2013 at 06:35:05PM +0200, Veaceslav Falico wrote:
> >>@@ -1712,6 +1721,8 @@ int bond_enslave(struct net_device *bond_dev, struct net_device *slave_dev)
> >>
> >> 	new_slave->last_arp_rx = jiffies -
> >> 		(msecs_to_jiffies(bond->params.arp_interval) + 1);
> >>+	for (i = 0; i < BOND_MAX_ARP_TARGETS; i++)
> >>+		new_slave->target_last_arp_rx[i] = jiffies;
> >>
> >> 	if (bond->params.miimon && !bond->params.use_carrier) {
> >> 		link_reporting = bond_check_dev_link(bond, slave_dev, 1);
> >
> >For cards with slow initial negotiation, this can cause a down -> up ->
> >down -> up flap on enslaving. This is why initial walue of last_arp_rx
> >was modified in commit f31c7937. Is there a reason not to initialize
> >target_last_arp_rx[i] to the same value?
> 
> Yep, I've seen this commit, however I didn't really understand it.
> 
> My logic is:
> 
> 1) on enslaving, we suppose that the new slave is up and give it a chance
> to prove it.
> 	1.1) if there is no active slave, lets try the new one, anyway
> 	     we're down.
> 	1.2) if there is one - nothing changes
> 
> 2) if, as you've said, it's still initializing - then it basically will just
> be marked as down until it finishes the initialization, and after that will
> go up. So, it goes up -> down (while initializing) -> up (when arps are
> received).
> 
> So, by using jiffies, we can start using the slave immediately, without
> waiting to receive the confirmation - if we don't have an active one,
> obviously. If we have one - nothing changes.
> 
> Did I miss something?

Experiments I've done show that most cards fall into one of two groups:

1. device is ready after dev_open() and netif_carrier_ok() reflects it
2. device is not ready for some time after dev_open()

For some cards from group 2, especially modern gigabit cards, this delay
can be surprisingly long, e.g. for some igb based cards it can take more
than two seconds until the card is ready and working. The original logic
(always start in up state) then caused ARP monitor to detect a failure
which was recorded and shown in statistics. I was not a functional
problem but it confused some customers and their monitoring tools.

Therefore commit f31c7937 changed logic to start a new slave in down
state if bond uses ARP monitoring and netif_carrier_ok() returns false.
This allows slaves from group 1 to start as up and stay that way and
slaves from group 2 to start as down and do only one down -> up
transition once the card is really ready; to be more precise: with a bit
of delay but exactly at the same time the slave would be finally up
without the patch.

This also required setting last_arp_rx not to "now" but to "more that
arp_interval ago", otherwise with arp_interval short enough (with
respect to the initialization delay), ARP monitor would falsely detect
up state on first opportunity, switch the slave to up, then after
arp_interval back to down once more and later finally to up. And unless
I overlooked something, if you set target_last_arp_rx[i] to jiffies,
this is exactly what happens with the "all" setting.

                                                        Michal Kubecek

--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists