Re: network regression: cannot rename netdev twice

From: Kay Sievers <kay.sievers@vrfy.org>
To: Jiri Slaby <jslaby@suse.cz>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>,
	Greg KH <greg@kroah.com>, LKML <linux-kernel@vger.kernel.org>,
	ML netdev <netdev@vger.kernel.org>
Subject: Re: network regression: cannot rename netdev twice
Date: Tue, 31 Jan 2012 12:13:41 +0100	[thread overview]
Message-ID: <CAPXgP12pj6esZVmN-PYTVgJ+_MaAfmpa9hKoKKQZ+Y=Y-Kb4eA@mail.gmail.com> (raw)
In-Reply-To: <4F27C9D4.3040105@suse.cz>

On Tue, Jan 31, 2012 at 12:00, Jiri Slaby <jslaby@suse.cz> wrote:
> On 01/31/2012 11:52 AM, Kay Sievers wrote:
>> On Tue, Jan 31, 2012 at 11:41, Jiri Slaby <jslaby@suse.cz> wrote:
>>> On 01/30/2012 11:52 PM, Kay Sievers wrote:
>>>> 2012/1/30 Jiri Slaby <jslaby@suse.cz>:
>>>>> I cannot boot properly with this commit:
>>>>> commit 524b6c5b39b931311dfe5a2f5abae2f5c9731676
>>>>> Author: Eric W. Biederman <ebiederm@xmission.com>
>>>>> Date:   Sun Dec 18 20:09:31 2011 -0800
>>>>>
>>>>>    sysfs: Kill nlink counting.
>>>>>
>>>>> 1) network systemd rule doesn't start network
>>>>
>>>> What does that mean? What's a network systemd rule?
>>>
>>> Oh, perhaps you call it a service file, not rule file?
>>>
>>> Anyway this is a different bug. Revert of the patch above does not help.
>>
>> Ok, fine. I checked too, and systemd does not play any silly games
>> with link counts.
>>
>>> The bug lays in the network layer. udev is unable to perform persistent
>>> eth naming:
>>> # ip link set eth0 name eth1    -- this one is OK
>>> # ip link set eth1 name eth0
>>> RTNETLINK answers: No such file or directory
>>
>> Please make sure nothing tries to swap netif names in userspace. We
>> have given up that approach, because it is far too fragile to
>> temporary rename devices to be able to swap the names, and race
>> against the loading of new kernel network drivers at the same time.
>>
>> This might be a new kernel problem here, but in general that approach
>> is just broken, we have have given up fiddling around here. Udev does
>> not do that anymore, and also the code that currently *can* be used to
>> do this, will be removed from udev in the future.
>>
>> Network devices can only be renamed to a namespace that isn't ethX,
>> and which does not race against kernel names.
>
> I have two eth interfaces. The one on the motherboard is named eth0, an
> added PCI card is eth1. But kernel enumerates them in the opposite order.
>
> So udev does this sequence:
> eth1 -> rename3
> eth0 -> eth1
> rename3 -> eth0
>
> How it can do it differently? (This is openSUSE factory.)

A future udev will not help you doing that. We have given up
supporting this approach. Renaming is done during booting, at the same
time we load new kernel drivers, and all breaks in non-interesting
ways. Apart from all the other unsolvable problems with this model.

Pretending we are able to rename netif names in the same namespace the
kernel is allocating new names is just plain wrong. There are races
you can't control. The entire approach creates far more problems than
it solves. We just have to admit it was wrong to do that.
Custom/to-rename netif names can just not be ethX.

>> Does is work, if you rename the devices to something else than ethX?
>
> Negative:
> # ip link set eth0 name krtek
> # ip link set krtek name jezek
> RTNETLINK answers: No such file or directory

This is a command sequence you type manually?

You are sure that userspace is not working in the background,
triggered by uevents, and comes into your way here?

Kay