Q: netdev: generate kobject uevent on network events.

Q: netdev: generate kobject uevent on network events.

[Johannes Stezenbach]

Hi Stephen,

about a year ago you posted a patch which allows to manage
network hotplug via udev (or busybox mdev), which is useful
for low-end embedded platforms that don't want to run ifplugd.

http://article.gmane.org/gmane.linux.hotplug.devel/13386

The patch still works with 2.6.32.  Can it get merged or
were there any issues with it?  It seems the discussion
just stopped and the patch was forgotten...


Johannes

Re: Q: netdev: generate kobject uevent on network events.

[Kay Sievers]

On Wed, Dec 9, 2009 at 18:02, Johannes Stezenbach <js@sig21.net> wrote:
> about a year ago you posted a patch which allows to manage
> network hotplug via udev (or busybox mdev), which is useful
> for low-end embedded platforms that don't want to run ifplugd.
>
> http://article.gmane.org/gmane.linux.hotplug.devel/13386
>
> The patch still works with 2.6.32.  Can it get merged or
> were there any issues with it?  It seems the discussion
> just stopped and the patch was forgotten...

Can't, in some setups, these events happen at a rather high frequency?
I heard of people running many hundreds of ppp interfaces on a single
box acting as a DSL concentrator. They state to already have trouble
handling the amount of uevents generated on such boxes just for the
"add/remove" events of all the interfaces if something goes wrong with
the network. If we add more for state transitions, such events would
probably need to be rate-limited. In general, uevents/udev are not
really suitable for high-frequency events, and if such behavior can be
expected, we might better stick with the current netlink interface.

Thanks,
Kay

Re: Q: netdev: generate kobject uevent on network events.

[Benjamin LaHaise]

On Wed, Dec 09, 2009 at 07:46:26PM +0100, Kay Sievers wrote:
> Can't, in some setups, these events happen at a rather high frequency?
> I heard of people running many hundreds of ppp interfaces on a single
> box acting as a DSL concentrator.

/me waves

Yes, DSL concentration is looking at tens of thousands to a hundred thousand 
interfaces on current hardware with 10G links.

> They state to already have trouble
> handling the amount of uevents generated on such boxes just for the
> "add/remove" events of all the interfaces if something goes wrong with
> the network. If we add more for state transitions, such events would
> probably need to be rate-limited. In general, uevents/udev are not
> really suitable for high-frequency events, and if such behavior can be
> expected, we might better stick with the current netlink interface.

Duplicating events in different delivery mechanisms seems pointless.  A 
simple netlink listener can be done in less than a hundred lines, so I don't 
see the need for the duplication.

		-ben

Re: Q: netdev: generate kobject uevent on network events.

[Dan Williams]

On Wed, 2009-12-09 at 19:46 +0100, Kay Sievers wrote:
> On Wed, Dec 9, 2009 at 18:02, Johannes Stezenbach <js@sig21.net> wrote:
> > about a year ago you posted a patch which allows to manage
> > network hotplug via udev (or busybox mdev), which is useful
> > for low-end embedded platforms that don't want to run ifplugd.
> >
> > http://article.gmane.org/gmane.linux.hotplug.devel/13386
> >
> > The patch still works with 2.6.32.  Can it get merged or
> > were there any issues with it?  It seems the discussion
> > just stopped and the patch was forgotten...
> 
> Can't, in some setups, these events happen at a rather high frequency?
> I heard of people running many hundreds of ppp interfaces on a single
> box acting as a DSL concentrator. They state to already have trouble
> handling the amount of uevents generated on such boxes just for the
> "add/remove" events of all the interfaces if something goes wrong with
> the network. If we add more for state transitions, such events would
> probably need to be rate-limited. In general, uevents/udev are not
> really suitable for high-frequency events, and if such behavior can be
> expected, we might better stick with the current netlink interface.

And then lets direct energy towards making the netlink interface simpler
for app writers.  Either by fixing up libnl to be dead-simple to use, or
some other mechanism.  Let's fix the *actual* problem (netlink is hard
to use from shell scripts) instead of creating more interfaces that work
around it.  Or maybe shellscripts are simply the wrong tool for this
job?

Dan

Re: Q: netdev: generate kobject uevent on network events.

[Johannes Stezenbach]

On Wed, Dec 09, 2009 at 11:21:26AM -0800, Dan Williams wrote:
> On Wed, 2009-12-09 at 19:46 +0100, Kay Sievers wrote:
> > 
> > Can't, in some setups, these events happen at a rather high frequency?
> > I heard of people running many hundreds of ppp interfaces on a single
> > box acting as a DSL concentrator. They state to already have trouble
> > handling the amount of uevents generated on such boxes just for the
> > "add/remove" events of all the interfaces if something goes wrong with
> > the network. If we add more for state transitions, such events would
> > probably need to be rate-limited. In general, uevents/udev are not
> > really suitable for high-frequency events, and if such behavior can be
> > expected, we might better stick with the current netlink interface.
> 
> And then lets direct energy towards making the netlink interface simpler
> for app writers.  Either by fixing up libnl to be dead-simple to use, or
> some other mechanism.  Let's fix the *actual* problem (netlink is hard
> to use from shell scripts) instead of creating more interfaces that work
> around it.  Or maybe shellscripts are simply the wrong tool for this
> job?

Hm, yeah, I liked the simplicity of the uevent patch, but I see
it doesn't work for everyone.

BTW, I just saw that a current busybox version (1.15+) has
an iflugd applet, which uses a mix of (open coded) netlink
and polling via various socket ioctls.
http://git.busybox.net/busybox/tree/networking/ifplugd.c


Thanks,
Johannes