Re: [PATCH] dracut-module-setup.sh: skip installing driver for loopback device

[Coiby Xu]

On Thu, Dec 15, 2022 at 05:32:02PM +0100, Martin Pitt wrote:
> Hello Coiby,
>
> Coiby Xu [2022-12-15 18:46 +0800]:
> > Hi Martin,
> > Thanks for raising the above concerns! To find a simpler solution, I
> > checked linux/drivers/net/loopback.c and find there is only one loopback
> > interface and the name always "lo". So the following code would be
> > sufficient,
> >
> > [ $_netif != lo ] || continue
>
> Sure, that looks nice and clean, especially if you *don't* want to apply this
> to other virtual devices.
>
> > Btw, ip and grep are used quite often in kexec-tools. I'm curious in
> > what way do you feel they are heavy? In fact, the cost of ip and grep
> > is negligible compared with the cost of calling dracut to build
> > initramfs.
>
> It's ok, I just generally try to avoid grepping text meant for human
> consumption (this might even be translated!), and avoiding calling two programs
> and a shell when a shell builtin and proper API is available. 
Thanks for the explanation! Yeah, I also prefer to avoid parsing text
using grep/sed if possible. In fact, one motivation behind the recent
network changes is also to avoid parsing NetworkManager connection
profiles.

> > > Hence I would recommend this instead:
> > >
> > >    [ -e "/sys/class/net/$_netif/device/driver" ] || continue
> > >
> > > and rename the commit to "skip installing driver for virtual devices"
> or so.
> >
> > >
> > > I tested this with ssh/localhost as well. It has the added benefit of not
> > > relying on "ip show" format, is more robust, more generic, and not
> even a new
> > > concept: that script already looks at /sys/class/net anyway.
> >
> > Unfortunately what you propose won't work.
> >
> > 1. kexec-tools wants to only install needed drivers to save memory and
> >    it's this dracut module's job to install needed drivers which may
> >    include the drivers of virtual devices e.g. bond, team or vlan device
> >    as well i.e. no one else will install these drivers.
>
> Right, I just meant that it might crash the same way for other virtual devices
> like veth or wireguard. I didn't test this. 
Thanks for the clarification! I just checked veth has the API for
ethtool to get driver info but this is not case for wireguard. Note even
if ethtool can get driver info for wireguard, kexec-tools doesn't
support wireguard as it depends on dracut to install the necessary
binary's like the ones in wireguard-tools.

This discussion makes me realize ethtool may not be robust as I have
expected and it also inspires me to do an exhaustive test on how many
drivers supports ethtool by iterating all network drivers. So thank you!

> > 2. ethtool talks to the driver via the netlink interface [1]. It doesn't
> >    use /sys/class/net/$_netif/device/driver to get the driver info.
>
> Ack, if that's what the dracut function uses, that sounds fine.
>
> So in summary, your `[ $_netif != lo ]` looks preferable to me (simpler, more
> obvious, and much more robust), and it seems you ruled out my generalized
> patch, so it seems we have a winner? 
Glad to know that:)

> Cheers!
>
> Martin
>  
-- 
Best regards,
Coiby