Re: [PATCH] bonding: reset queue mapping prior to transmission to physical device

[David Miller <davem@davemloft.net>]

From: Ben Hutchings <bhutchings@solarflare.com>
Date: Thu, 02 Jun 2011 21:51:46 +0100

> On Thu, 2011-06-02 at 13:46 -0700, David Miller wrote:
>> From: Nicolas de Pesloüan <nicolas.2p.debian@gmail.com>
>> Date: Thu, 02 Jun 2011 22:13:57 +0200
>> 
>> > To be more precise, due to the way bonding use queue mapping for slave
>> > selection, it is desirable to clear the mapping before sending to the
>> > slave, because the meaning of the mapping for the slave interface
>> > might be really different from the meaning for the bonding
>> > interface. Arguably, this is the mapping usage in bonding which is
>> > "different" from other usages, but...
>> 
>> This just confirms my reasoning behind why I wanted to discourage
>> drivers from providing explicit ->ndo_select_queue() methods unless
>> absolutely necessary.
>> 
>> Information now gets lost in cases like this bonding issue.
>> 
>> Bonding should definitely, as I suggested, remember the original
>> rxhash value and restore it when sending to the slave.
> 
> Surely RX queue (queue_mapping), not RX hash (rxhash, which is unchanged
> anyway AFAIK).

Right.