[systemd-devel] Match section of .link file not working

[Simon Peeters]

2014-08-25 19:46 GMT+02:00 Kay Sievers <kay at vrfy.org>:
> On Mon, Aug 25, 2014 at 7:25 PM, Lennart Poettering
> <lennart at poettering.net> wrote:
>> On Sun, 24.08.14 06:49, Simon Peeters (peeters.simon at gmail.com) wrote:
>>
>>> >> This should be the value of ID_PATH property, not sysfs path. Check with
>>> >>
>>> >> udevadm info --query
>>> >> property /sys/devices/ocp.3/47400000.usb/musb-hdrc.1.auto/usb1/1-1/1-1:1.0/net/wlv3
>>> >
>>> > oops, I confused ID_PATH with DEVPATH, now i feel like an idiot.
>>
>> ID_PATH is basically the "reduced + stable" part of the full sysfs path.
>>
>>> Ok, so there is only one problem now:
>>> ID_PATH is the same for all my vif interfaces, and i need to select
>>> only one of them.
>
> ID_PATH usually describes the "parent hardware", and the same hardware
> can have multiple independent devices. Unique names for several
> devices need to be composed by appending something to the "parent
> hardware".
>
> I don't have a good idea what a working model for your case would be,
> where multiple devices share the same parent. This is not a common
> model.
>
>> This is some embedded device with your hw connected to some platform
>> bus?
>>
>> Currently the ID_PATH logic doesn't understand platform busses (since it
>> is not clear how "stable" they are). This has come up before, and could
>> probably be fixed. Maybe we should assume that platform busses are
>> always enuemrated in a stable way, dunno.
>>
>> Kay probably has something useful to say here. Kay?
>
> Special buses need explicit support in the code if ID_PATH is not
> generated, but that does not look like the problem here.
>
> Seems the USB interface exports 3 different network devices. Do they
> have any unique property? Like dev_id, dev_port in /sys? Or are the
> types of the netifs unique?

Well, as posted earlier , these are virtual interfaces, kind of the
vlans for wifi (allows to connect ot multiple networks on a same phy).
(sory for not precise enough in the original post)
So they do not realy have something unique and stable appart from the name.

I think there are 3 posible ways to go forward:
 - fix iw (and maybe the kernel to) to support a mac address parameter
at interface creation time (so the mac is set even before udev sees
it)
 - Add vif netdev support to networkd similar to how vlans are
handled, also alowing to set the mac address.
 - Support [Match] Name= in .link files, even if it is with a large
warning, or even as an unsupported "ymmv" option (because names are
normaly not stable)

1. could be hard if the kernel does not support this, which i havn't
checked yet.
2. would be ideal and could also work in most cases if the kernel does
not support this in an atomic operation (since networkd can not race
with itself)
3. this would be the easy, almost lazy, way, but would also work with
wathever wierd exotic virtual device the kernel developrs come up
with.

I noted this boils down to almost the same discution there was a while
ago under the title: "[systemd-devel] systemd-networkd: setting the
MAC address for a .netdev unit?"[0] with the only difference that
.netdev support for vif's not yet exists. (so I do this in a separate
unit)

> If anything like that is available, that could be appended to the
> parent ID_PATH string and identify the device uniquely, if you get
> what I mean.

nah, I don't think so, I can check but the only thing i can imagine is
the name passed during creation, or maybe some sequence id, but I am
not sure wether that would be usable.

> Kay

[0] http://lists.freedesktop.org/archives/systemd-devel/2014-April/018712.html