Re: [PATCH V2] Use firmware provided index to register a network interface

[Greg KH <greg@kroah.com>]

On Thu, Oct 07, 2010 at 11:31:13AM -0500, Matt Domsch wrote:
> On Thu, Oct 07, 2010 at 08:11:34AM -0700, Greg KH wrote:
> > On Thu, Oct 07, 2010 at 07:23:35AM -0700, Narendra_K@Dell.com wrote:
> > > Hello,
> > > 
> > > V1 -> V2:
> > > 
> > > This patch addresses the scenario of buggy firmware/BIOS tables. The
> > > patch introduces a command line parameter 'no_netfwindex', passing which
> > > firmware provided index will not be used to derive 'eth' names. By
> > > default, firmware index will be used and the parameter can be used to
> > > work around buggy firmware/BIOS tables.
> > > 
> > > Please find the patch below.
> > > 
> > > From: Narendra K <narendra_k@dell.com>
> > > Subject: [PATCH] Use firmware provided index to register a network device
> > > 
> > > This patch uses the firmware provided index to derive the ethN name.
> > > If the firmware provides an index for the corresponding pdev, the N
> > > is derived from the index.
> > 
> > No, this has the very big chance to reorder existing network names and
> > should all be done in userspace with the firmware information exported
> > in sysfs, if the user so desires.
> 
> Existing names that use 70-persistent-net.rules won't change.

But users that don't use it would cause their names to change, and you
can't break that, no matter how naturally fragile those types of systems
are.  Sorry.

> The kernel patch has the advantage of not requiring users to change
> their scripts, by reserving the first N values for onboard devices as
> BIOS describes them.

So you feel that updating a kernel is easier than getting a user to
update their userspace scripts?  While I might also feel it is that way,
in reality, it isn't, sorry.

> Userspace udev rules require us to change user behavior, and likely
> their scripts, to use a new namespace such as ethlom1, in order to get
> deterministic naming.

Then use that, don't put this in the kernel please.

> > > It took some time to find out the details asked above. Right,
> > > windows does not use SMBIOS type 41 record to derive names. But as
> > > a datapoint, windows also has the same problem.
> 
> > If windows does not use this field, then you can guarantee it will
> > show up incorrectly in BIOSes and never be tested by manufacturers
> > before they ship their machines.
> 
> There are 2 methods of exporting the information that have gotten
> confusing here.
> 
> 1) SMBIOS type 41 method.  Windows does not use this today, and I
>    can't speak to their future plans.  Narendra's kernel patch does,
>    as has biosdevname, the udev helper we first wrote for this
>    purpose, for several years.

Then stick with that udev helper please :)

> 2) soon-to-be-released PCIe Firmware Spec, exporting label and index
>    as an ACPI _DSM().  It is anticipated that Windows will use this
>    information in some fashion at some point, per our conversations
>    with Microsoft as part of the PCI SIG proposal process.  We've held
>    off submitting a kernel patch until the spec is public.

Let's worry about that _when_ it is public, and when there is a Windows
version that supports it please.  Until then, there will not be support
for this in BIOSes in any usable way.

> Cases:
> 1) BIOS doesn't provide this information (the common case today for
>    all but Dell 10G and newer servers), nothing is reserved, and
>    therefore the existing 70-persistent-net.rules take effect to name
>    devices.  No change in behavior for existing systems.  New installs
>    will have NICs named non-deterministically, and recorded in
>    70-persistent-net.rules.  Users desiring consistent naming must
>    adjust 70-persistent-net.rules after install, and to avoid
>    eth0_rename collisions, must rename into another namespace.

That's fine and is how things work today, right?

> 2) BIOS provides this information correctly (Dell 10G and newer servers, or
>    anyone else implementing the spec).  The first N values for onboard
>    devices are reserved and assigned by the kernel to onboard
>    devices.

And you just broke people's machines that don't use udev for network
naming.  Sorry, that's not going to work.

>    70-persistent-net.rules may still try to rename the
>    ports, and may fail with eth0_rename collisions.

So you just broke their machines as well, when it was working, again,
not acceptable.

> 3) BIOS provides this information incorrectly (none known today).

Only because we have no way of testing for this :)

> What I can't find here is a solution where the user gets determistic
> naming, without being forced to change the namespace and adjust their
> scripts.  Can anyone?

No, and they shoudn't.

You should only have "deterministic" naming if you want it, and you set
it up to do so, from userspace.  The kernel is not responsible for this,
sorry.

If you want to do it in userspace, you have all the tools, and the data
exported from the kernel to do so.

> I'm not opposed to "do it all in userspace" - really, I'm not.  But
> the 'where' is only part of the problem.  No userspace solution has
> yet been proposed that doesn't require a namespace change, and I'm
> still hoping to avoid that.

Just use a new script for people who want this, they will have to know
they want it anyway, right?

thanks,

greg k-h