Peter Memishian wrote:
>  > "Is there a reason I'd want to do a temporary delete instead of    (say) 
>  > forcing the link down?  (And can I force the link down?)"
>  > 
>  > I have several questions regarding this:
>  > 
>  > - What's the real user case that needs a "link down"?
> 
> I'm not sure; what's the reason a real user would want to temporarily
> delete a link?
> 
The only case I can think of is that the user wants to use the associated 
device for other purpose. One example is the case I gave (to plumb the link 
directly instead of put it in an aggregation).

>  > - After think it more, I don't think "link down" can solve all the 
> scenario 
>  > that "temporary delete" can solve. Assume that one wants to (temporarily) 
>  > plumb an link which is a port of an established aggregation. (Note that 
> one 
>  > cannot plumb an link being aggregated). Today, one could temporarily 
> delete 
>  > this aggregation, and plumb that link. But "down" the aggregation won't 
> work.
> 
> Or one could just temporarily remove that link from the aggregation and
> plumb it, right?  

Only if this link is not the only port in the aggregation. But in most 
cases, yes. And I don't know why create an aggregation with a single port in 
the first place.

> But yes, it wouldn't make sense to force a given link
> down in this case.
> 
>  > wes-1
>  > 
>  > "In the absence of any way to map from a physical device path to a slot 
>  > label, etc., shouldn't there at least be some way to associate an 
>  > administratively-chosen descriptive field with each device?"
>  > 
>  > Will we add this support? and how? In the form of a linkprop?
> 
> To me, this still feels like unrelated work. 

I agree. I think the comments originated from that user cannot tell the 
/devices path from the link name, but as there is no directly relationship 
between the /devices path and the physical slot number, I don't understad 
why /devices path is important.

> But yes, we could add a
> "comment" link property that's a free-form string.  It just feels strange.
> 

Personally, I don't like to support it though.

>  > 1. libdevinfo impact
>  > 
>  > A general answer is that we will provide a dladm_walk() function to walk 
> all 
>  > the link names, and check all the places in ON to make sure every place 
>  > calls libdevinfo needs to be changed to call dladm_walk().
>  > 
>  > But at the same time, I remembered that I also discussed the possibility 
>  > that we might change the current libdevinfo implementation, to make the 
>  > current di_walk_minor(DDI_NT_NET) to return nothing, (and to provide 
> another 
>  > form of libdevinfo function to walk the list of network devices), so that 
>  > the application doing libdevinfo walking today will break badly, and will 
> be 
>  > changed to do the right thing in the earliest time.
>  > 
>  > But since we won't change the default link name, I don't know whether this 
>  > is still necessary.
> 
> I think it's still worthwhile, though we might want to do it after the
> initial putback.  For instance, we could do:
> 
>       * Initial putback: compatible names by default; no change to
>         di_walk_minor().
>       * Follow-up putback: change di_walk_minor(); examine and address
>         impact.
>       * Follow-up putback #2: vanity names by default; examine and
>         address impact.
> 
If so, do we need to mention this phases plan in the PSARC document?

>  > 2. access policy for network device (per-driver or per-link or default 
>  > access policy only).
>  > 
>  > I am thinking of keeping the current implementation as is (keep the the 
>  > per-driver policy and not introduce the per-link one).
> 
> Seems like per-driver policy becomes effectively "random" after vanity
> naming, so I'm not sure I like this approach.
> 
>  > If this is not a good idea, I will send a mail to the network-discuss 
> alias 
>  > to see whether there is anyone knows who is making use of the per-driver 
> policy.
> 
> I think we should get that data.
> 
I've sent the question.

Thanks
- Cathy

Reply via email to