Re: [PATCH 2/2] [e1000 VLAN] Disable vlan hw accel when promiscuousmode

[Kok, Auke]

Willy Tarreau wrote:
> On Mon, Nov 12, 2007 at 03:19:23PM -0800, David Miller wrote:
>> From: Willy Tarreau <w@xxxxxx>
>> Date: Tue, 13 Nov 2007 00:15:16 +0100
>>
>>> I can say that sometimes you'd like to be aware that one of your
>>> VLANs is wrong and you'd simply like to sniff the wire to guess the
>>> correct tag. And on production, you simply cannot remove other
>>> VLANs, otherwise you disrupt the service.
>> If you were plugged into the wrong physical switch, how might
>> you debug that problem in production? :-)
> 
> tcpdump on an unfiltered RJ45 tells you a lot of things. Some switches
> for instance send keepalive packets (ether proto 9000) with their own
> MAC as source+dest, others send CDP packets, etc... I've very rarely
> stayed plugged to the wrong switch for too long before discovering it.
> But that requires the ability to sniff.
> 
>> Really, it's the same issue, just virtualized, as in the name
>> for the feature, VLAN.
> 
> No, it's not the same, because as soon as you pass through a switch
> which is not configured for VLANs, it does not make any distinction,
> and the same DMAC is considered on a single port whatever the VLAN.
> This is not true with multiple switches. Also, sometimes, being able
> to see that your port gets flooded with traffic for a VLAN on which
> you're not configured helps you understand why you cannot fill the
> port.
> 
> I'd like that we can use the hardware correctly without having to
> buy TAPs. It reminds me the discussions about TOE. You claim
> yourself that TOE is bad in part because you have little if any
> control on the bugs on the TCP stack on the chip. This is the same
> here. If I know that the chip does not send me what it receives
> when configured in promiscuous mode, I can I swear it never received
> the missing packet ? There's always a doubt. Maybe sometimes the
> filter is buggy, maybe it only passes tags when the remaining bits
> are all zero, etc... Promiscuous mode generally means that you're
> the closest possible to the wire. We already do not receive pause
> frames and sometimes that's annoying. But having no control over
> what we want to see is frustrating.
> 
> At least, being able to disable the feature at module load time
> would be acceptable. Many people who often need to sniff on decent
> machines would always keep it disabled.

I really do not want to add another module parameter to the driver for this. It
seems bogus as well: if we can agree on one sane choice it's much better, and if
we don't then we should use ethtool to provide a uniform way of implementing this
for all drivers sanely.

Auke
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/