Message-ID: <20110713052848.GB1799@minipsycho>
Date:	Wed, 13 Jul 2011 07:28:48 +0200
From:	Jiri Pirko <jpirko@...hat.com>
To:	David Lamparter <equinox@...c24.net>
Cc:	netdev@...r.kernel.org, davem@...emloft.net,
	shemminger@...ux-foundation.org, kaber@...sh.net, fubar@...ibm.com,
	eric.dumazet@...il.com, nicolas.2p.debian@...il.com,
	andy@...yhouse.net, greearb@...delatech.com, mirqus@...il.com
Subject: Re: [patch net-next-2.6] net: allow multiple rx_handler registration

Tue, Jul 12, 2011 at 06:03:30PM CEST, equinox@...c24.net wrote:
>On Tue, Jul 12, 2011 at 05:01:22PM +0200, Jiri Pirko wrote:
>> Tue, Jul 12, 2011 at 04:29:38PM CEST, equinox@...c24.net wrote:
>> >On Tue, Jul 12, 2011 at 03:20:08PM +0200, Jiri Pirko wrote:
>> >> Not possible. See netdev_set_master(). Anyway, before rx_handler was
>> >> introduced, this was possible and no one cared.
>> >
>> >I don't see how this is related. I'm talking about the other end of your
>> >bond. Like for example the 802.3ad capable switch you're bonding to.
>> 
>> Well it is related in way that you cannot have one device in br an bond
>> in same time....
>
>Grah, I was looking at our production kernel tree, which doesn't have
>the netdev_set_master calls from the bridging code. Sorry, my fault.
>
>> >> >b) a device having macvlans and being a bond slave
>> >> > -> Fully incompatible. Same as above, packets to the macvlan will end
>> >> >    up on other bond member devices.
>
>But case b) is still up & alive, macvlan doesn't use netdev_set_master.
>
>> >> This patch doen't introduce anything new which wasn't possible before
>> >> rx_handler times. Anyway removing bond from using rx_handler as you
>> >> suggested pushes us back.
>> >
>> >I would actually consider this a regression, if the clashing rx_handler
>> >is the only thing that gets bonding an 'exclusive' hold of the device.
>> 
>> No regression. Regression it would be if something wouldn't work on same
>> setup. But this is not the case!
>
>Your patch allows a setup (bond+macvlan) that is not only a violation of
>the specification's letters, but will also wreak rather big havoc and
>may cause parts of itself to become non-functioning.
>
>What happens when the user does this?:
> eth0 -> bond0
>    -> macvlan0 -> bond1
>
>My complaint is primary centering on the inclusion of bonding code into
>this. There might be bonding modes where this is acceptable, but in
>802.3ad mode this royally breaks things.

Well as I pointed out, this is not a regression. User should not
configure this. And as I said, I plan to cook some follow up patches to
make this configs not possible in future. But anyway, user should be
responsible for his config and if it's wrong he should not expect it to
work. I can imagine a large set of screwed up configs which are not
forbidden. Forbidding all wrong configs is not the right way I think.

>
>> >> And to your idea about multi-bridge support, br co needs to be
>> >> adjusted as well. And in relation with PRIO, my idea (inspired from RFC
>> >> of this patch comments) is to allow users to change priorities
>> >> dynamically from userspace. Also then it could be a range of prios for
>> >> bridge for example.
>> >
>> >Hoping I can convey my point,
>> >
>> >
>> >-David
>> >
>> >
>> >P.S.: Could you please provide some sample usage cases for this feature?
>> 
>> Converting vlan to rx_handler needs this at least.
>
>Hm, yes. I guess this patch is needed to pave the way. I uphold my fears
>about including bonding (read: 802.3ad) in this though. Maybe I should
>cook up some code to give 802.3ad an exclusive grip on the slaves?

Sure you can. But I was thinking about some more generic way.

>
>
>-David
--
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