LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: lblc + fwmark + persist = bug?

To: Damien CLERMONTE <damien.clermonte@xxxxxxx>
Subject: Re: lblc + fwmark + persist = bug?
Cc: lvs-users@xxxxxxxxxxxxxxxxxxxxxx
From: Julian Anastasov <ja@xxxxxx>
Date: Fri, 5 Jul 2002 15:56:21 +0300 (EEST)
        Hello,

On Tue, 4 Jun 2002, Damien CLERMONTE wrote:

> [Please CC me, I'm not on the list]

        Done

> I think I found an annoying bug when using lblc, fwmark and persist.
> When in fwmark mode, ipvs core don't keep destination address, only fwmark.
> The problem is: lblc(r) loadbalances on destination.
>
> In my config, the netmask is 0.0.0.0 (I want all clients to use the same real 
> server

        Persistence Netmask 0.0.0.0 means "treat all clients in this
network as one client". The result: only one RS is used from the
world.

> for a vip), and I expect the persistance to keep a dynamic mapping between 
> vips and
> real ips. Unfortunately, as the persistance code only keeps 0.0.0.fwmark as 
> destination,
> all the traffic goes to the same server and never switch to another...

        The relations are:

FWMARK-based VS:
        Client_Net -> FWMARK -> RS

non-FWMARK-based VS:
        Client_Net -> VIP -> RS

        Note that with netmask 0.0.0.0 Client_Net is 0/0

        Also note that the client persistence checks occur before
scheduling, i.e. before LBLC.

> I know that removing this special case would prevent multiple fwmark on same 
> vip to work,
> but now a unique fwmark on multiple vips is broken!
> I think the solution would be to keep both fwmark and dest and to compare 
> both, so both
> cases would work...
>
> Am I wrong?
> Any thoughts?

        Try with default netmask. Any difference?

> Regards,
>
> Damien

Regards

--
Julian Anastasov <ja@xxxxxx>



<Prev in Thread] Current Thread [Next in Thread>