Re: [rfc 0/3] IPVS: checksum updates

[Simon Horman <horms@verge.net.au>]

On Mon, Sep 08, 2008 at 02:14:12PM +0200, Julius Volz wrote:
> On Mon, Sep 8, 2008 at 2:04 PM, Simon Horman <horms@verge.net.au> wrote:
> > On Mon, Sep 08, 2008 at 09:57:35PM +1000, Simon Horman wrote:
> >> On Mon, Sep 08, 2008 at 01:42:59PM +0200, Julius Volz wrote:
> >> > On Mon, Sep 08, 2008 at 08:41:22PM +1000, Simon Horman wrote:
> >> > > On Mon, Sep 08, 2008 at 12:03:04PM +0200, Julius Volz wrote:
> >> > > > On Mon, Sep 8, 2008 at 4:04 AM, Simon Horman wrote:
> >> > > > > Hi,
> >> > > > >
> >> > > > > The impetus for this series of patches is Julian Anastasov noting
> >> > > > > that "load balance IPv4 connections from a local process" checks
> >> > > > > for 0 TCP checksums. Herbert Xu confirmed that this is not legal,
> >> > > > > even on loopback traffic, but that rather partial checksums are
> >> > > > > possible.
> >> > > > >
> >> > > > > The first patch in this series is a proposed solution to handle
> >> > > > > partial checksums for both TCP and UDP.
> >> > > > >
> >> > > > > The other two patches clean things up a bit.
> >> > > > >
> >> > > > > I have not tested this code beyond compilation yet.
> >> > > >
> >> > > > After some first tests, remote connections are still working, but not
> >> > > > local ones from the director. The TCP handshake works and the
> >> > > > connection is established, but all following packets arriving at the
> >> > > > real server have an incorrect TCP checksum.
> >> > > >
> >> > > > Btw., this happens both with and without this last series of patches,
> >> > > > so I can't get the local client feature working at all. Looking at it
> >> > > > further...
> >> > >
> >> > > Ok, is this for both IPv4 & IPv6? Does it still occur with just the first
> >> > > patch in this series applied?
> >> >
> >> > It's for both, although I only tested IPv4 at first. Here is a complete
> >> > test matrix of what works when:
> >> >
> >> > CR = connection refused
> >> > T = connection timeout
> >> > C = connection established, but not working afterwards
> >> > OK = working
> >> >
> >> >                     remote client | local client
> >> > COMMIT                      v4      v6    | v4      v6
> >> > ======================================|=================
> >> > CSUM 3/3            OK      T     | C       T
> >> > CSUM 2/3            OK      T     | C       T
> >> > CSUM 1/3            OK      T     | OK      T
> >> > W/O CSUM            OK      T     | C       T
> >> > ...                               |
> >> > f2428ed5            OK      T     | CR      CR
> >> > 4856c84c            OK      CR    | CR      CR
> >> > f94fd041 (my last one)      OK      OK    | CR      CR
> >> >
> >> > So the last time that IPv6 was working _at all_ was at my last commit of
> >> > the big v6 series...
> >>
> >> Ok, I'm really sorry about that :-(
> >>
> >> Do you want me to revert f2428ed5 & 4856c84c until this has been tracked down?
> >>
> >
> > Hi,
> >
> > Does 4856c84c + the following change (which you pointed out over the
> > weekend) work for remote IPv6 ?
> >
> > diff --git a/net/ipv4/ipvs/ip_vs_core.c b/net/ipv4/ipvs/ip_vs_core.c
> > index 26e3d99..c413444 100644
> > --- a/net/ipv4/ipvs/ip_vs_core.c
> > +++ b/net/ipv4/ipvs/ip_vs_core.c
> > @@ -1282,7 +1282,7 @@ ip_vs_in(unsigned int hooknum, struct sk_buff *skb,
> >         *      Don't handle local packets on IPv6 for now
> >         */
> >        if (unlikely(skb->pkt_type != PACKET_HOST ||
> > -                    (af == AF_INET6 || (skb->dev->flags & IFF_LOOPBACK ||
> > +                    (af == AF_INET6 && (skb->dev->flags & IFF_LOOPBACK ||
> >                                         skb->sk)))) {
> >                IP_VS_DBG_BUF(12, "packet type=%d proto=%d daddr=%s ignored\n",
> >                              skb->pkt_type,
> >
> 
> No, that just changes the behavior from "connection refused" to timeout...
> 
> I'm actually looking at that case now (4856c84c1358b, but with the fix
> above). It seems that the NAT isn't working (DR works, by the way!).
> At least the first packet arriving at the real server still has the
> client's IP as the source (in the v6 case)...

Ok, I'm looking at NAT with 4856c84c1358b + that fix too too,
but on v4 :-)

> Let's wait with reverting the local client patches until tomorrow...
> maybe I can find the problem until then.

Ok, I was just concerned that this might hold up merging
your code into Dave's tree for too long, thats all.