Broadcom bge and 802.1Q vlan tags

[Sam Leffler]

Ruslan Ermilov wrote:
> On Wed, Oct 13, 2004 at 08:51:02AM -0700, Sam Leffler wrote:
> 
>>Ruslan Ermilov wrote:
>>
>>>Hi Sam,
>>>
>>>On Tue, Oct 12, 2004 at 01:20:07PM -0700, Sam Leffler wrote:
>>>
>>>
>>>>>>This pessimizes normal traffic.
>>>>>
>>>>>m_tag_locate() doesn't look like a very expensive function.  And
>>>>>with the "normal traffic", I don't expect to be more than one tag,
>>>>>no?  Also, if if_nvlans > 0, this is already "pessimized".
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>>We should look for a solution in the 
>>>>>>driver(s) to avoid sending packets up with tags when no vlans are 
>>>>>>configured.
>>>>>>
>>>>>
>>>>>I'd be opposed to such a change in behavior.  The VLAN consumer can
>>>>>be not only vlan(4), it can equally be the ng_vlan(4) node, etc.
>>>>
>>>>I'm not sure what you are opposed to or why.  The issue I have is that 
>>>>m_tag_locate can be expensive if many packets have tags.  The check for 
>>>>the existence of vlans configured on the interface short-circuits this 
>>>>work.  That vlan-tagged packets may be generated when no vlans are 
>>>>configured seems wrong to me and breaks the assumption used to write the 
>>>>code.  Changing the driver to drop the frame if ifp->if_nvlans is zero 
>>>>seems straightforward and could probably be hidden in the existing macro.
>>>>
>>>
>>>Please take a moment and re-read what I've already said: vlan(4) is not
>>>the only consumer of VLAN frames: ng_vlan(4) is another such one, and I
>>>have a proprietary Netgraph node here that demultiplexes VLANs.  If you
>>>start dropping VLAN frames in drivers when if_nvlans == 0, this will be
>>>a problem for me.  Is that clear now?
>>>
>>>
>>>Cheers,
>>
>>I've read what you've written but you also haven't explained why you 
>>can't signal the presence of these other entities in some way.
>>
> 
> Because these other entities don't have an access to "ifp", and can
> even exist on remote host (please see below).
> 
> 
>>The 
>>current mechanism to signal the presence of "interested parties" for 
>>vlan-tagged frames is ifp->if_nvlans. You are saying you have new 
>>(proprietary) code that is interested in vlans but will not use the 
>>existing mechanism. My reaction is fix your code, don't pessimize the 
>>code everyone else uses without netgraph.
>>
> 
> But ng_vlan(4) is part of the standard FreeBSD distribution, and you
> don't have access to "ifp" inside ng_vlan(4), because it's connected
> to the interface indirectly, through the ng_ether(4) node.  Even worse,
> ng_vlan(4) may not even be connected to a local interface, for example,
> you can capture and tunnel all Ethernet traffic to another host, and
> do the VLAN processing there, FWIW.
> 
> So while ifp->if_nvlans seems to be a good signalling mechamism for
> vlan(4), it's not suitable for ng_vlan(4) and other Netgraph code
> that works with VLAN.  This code works now, and I'm afraid it will
> break if we change drivers to drop VLAN frames if if_nvlans == 0,
> and I fail to see how I can make it work again after that.
> 
> In other words, I want that ng_ether(4) continues to see VLAN frames
> even if no vlan(4) interfaces are configured, like it does now: the
> ng_ether processing is done in ether_input() before ether_demux()
> that checks for ifp->if_nvlans.
> 
> OTOH, you may be right that one option would be to make ng_ether(4)
> increment ifp->if_nvlans, but I'm a little worried about the effect
> of doing this on the VLAN_OUTPUT_TAG macro (it looks safe, but I'm
> not sure).

I'm open to changing the mechanism by which we signal the presence of 
vlans and/or "interested parties" (or the presence of a h/w tag). 
However I believe if_nvlans is managed entirely in if_vlan.c and bumping 
it when ng_vlan is present would be ok (if that's feasible).

OTOH this issue exists only because running the tag list for every 
packet can potentially be expensive.  I recall cjsp (?) wanting to 
optimize this better so another tact is to look (again) at how to speed 
this up for the most common/important cases.  Vlans may be important 
enough to assign an mbuf flag bit though I'd prefer to do that as a last 
resort (and I'm sure other folks would popup and want a bit too :)).

	Sam