Re: [patch iproute2 1/6] iproute2: ipa: show switch id

[ebiederm@xmission.com (Eric W. Biederman)]

Jiri Pirko <jiri@resnulli.us> writes:

> Thu, Dec 04, 2014 at 09:06:14PM CET, ebiederm@xmission.com wrote:
>>ebiederm@xmission.com (Eric W. Biederman) writes:
>>
>>> Jiri Pirko <jiri@resnulli.us> writes:
>>>
>>>>>So this id needs to be globally unique?
>>>>
>>>> No. It is enough to be unique within a single system. It serves for no
>>>> more than to find out 2 ids are same or not, no other info value.
>>>>
>>>> So when the drivers uses sane ids (like mac for example, or in case of
>>>> rocker an id which is passed by qemu command line), the chances of
>>>> collision are very very close to none (never say never).
>>
>>Thinking about what you said a little more.
>>
>>Two different sources of persistent numbers picking numbers by
>>completely different algorithms can give you no assurance that you don't
>>produce conflicts.
>>
>>The switch id as desisgned can not work.
>>
>>There are expected to be between 2**36 to 2**40 devices in this world.
>>Your first switch id is a 64it number.  At the very best by the birthday
>>pardox predicts there will be a conflict ever 2**32 devices or between
>>2**4 and 2**8 devices in the world with conflicts.  If the ids are not
>>randomly distributed (which they won't be) things could easily be much
>>much worse.
>>
>>That is just good enough the code could get out there and run for years
>>before you have the nightmare of having to fix all of userspace.   That
>>is a nightmare no one needs.
>>
>>So please remove this broken code, and this broken concept from the
>>kernel and go back to the drawing board.
>
> In that case the phys port id is broken in the same way. Let's rather
> think about how to avoid conflicts for both. Given the fact the
> conflicts should be avoided only on a single baremetal, that should be
> doable (for (bad) example using driver name mixed with driver created
> id).

No.  phys_port_id is not broken in the same way, and phys_port_id does
not have the same set of properties.

phys_port_id's in practice all have an IEEE prefix that identifies the
manufacturer and a manufacture assigned serial number.  Aka a mac
address or a EUID-64.  What the mlx4 ethernet driver is doing retunring
a 64bit EUID-64 I don't know.  If there are problems in the worst
case issues with phys_port_id are fixable by simple driver tweaks,
because fundamentally we are working with globally uniuqe identifiers.
Well globally unique baring manufacturing bugs in eeproms.



I agree with you that the switch id concept can be saved.  But I think
we should fix switch id before we export it to userspace so we don't
have to break userspace later.

My intuition says we want something like ifindex, but I am not at all
certain how switch id is planned to be used.  Given that it is single
box I don't expect you are sending it out over the wire.

*shrug*

Why does switch id need to be persistent?  Why can't switch id be
property like ifindex?

What are the actual requirements.

Eric