Re: [Lxc-users] Bad checksums and lost packets with macvlan on dummy

[Eric Dumazet <eric.dumazet@gmail.com>]

Le dimanche 27 février 2011 à 21:35 +0100, Daniel Lezcano a écrit :
> On 02/27/2011 08:50 PM, Eric Dumazet wrote:
> > Le dimanche 27 février 2011 à 16:14 +0100, Daniel Lezcano a écrit :
> >> On 02/23/2011 06:13 PM, Andrian Nord wrote:
> >>> On Mon, Feb 21, 2011 at 05:07:31PM +0100, Daniel Lezcano wrote:
> >>>> I Cc'ed the netdev mailing list and Patrick in case my analysis is wrong
> >>>> or incomplete.
> >>> I'm confirming, that this happens only when macvlan's are onto dummy net
> >>> device. In case of some physical interface under macvlan there is no lost
> >>> packages and no broken checksums.
> >> I did some tests with a 2.6.35 kernel version and it seems the checksum
> >> errors do not appear.
> >> I noticed there are some changes in the dummy setup function:
> >>
> >>     dev->features   |= NETIF_F_SG | NETIF_F_FRAGLIST | NETIF_F_TSO;
> >>     dev->features   |= NETIF_F_NO_CSUM | NETIF_F_HIGHDMA | NETIF_F_LLTX;
> >>
> >>
> >> May be that was introduced by commit:
> >>
> >> commit 6d81f41c58c69ddde497e9e640ba5805aa26e78c
> >> Author: Eric Dumazet<eric.dumazet@gmail.com>
> >> Date:   Mon Sep 27 20:50:33 2010 +0000
> >>
> >>       dummy: percpu stats and lockless xmit
> >>
> >>       Converts dummy network device driver to :
> >>
> >>       - percpu stats
> >>
> >>       - 64bit stats
> >>
> >>       - lockless xmit (NETIF_F_LLTX)
> >>
> >>       - performance features added (NETIF_F_SG | NETIF_F_FRAGLIST |
> >>       NETIF_F_TSO | NETIF_F_NO_CSUM | NETIF_F_HIGHDMA)
> >>
> >>       Signed-off-by: Eric Dumazet<eric.dumazet@gmail.com>
> >>       Signed-off-by: David S. Miller<davem@davemloft.net>
> >>
> >>
> >> Eric,
> >>
> >> Andrian is observing, with a couple of macvlan (in bridge mode) on top
> >> of a dummy interface, a lot of checksums error and packets drop.
> >> Each macvlan is in a different network namespace and the dummy interface
> >> is in the init_net.
> >>
> >> Any ideas ?
> > Not sure I understand... I thought dummy was dropping all frames
> > anyway ?
> >
> > static netdev_tx_t dummy_xmit(struct sk_buff *skb, struct net_device *dev)
> > {
> >          struct pcpu_dstats *dstats = this_cpu_ptr(dev->dstats);
> >
> >          u64_stats_update_begin(&dstats->syncp);
> >          dstats->tx_packets++;
> >          dstats->tx_bytes += skb->len;
> >          u64_stats_update_end(&dstats->syncp);
> >
> >          dev_kfree_skb(skb);
> >          return NETDEV_TX_OK;
> > }
> >
> >
> > Maybe you could describe the setup ?
> 
> Yes, it is very simple.
> 
> There are two network namespaces.
> 
> macvlan1 is in network namespace 1
> macvlan2 is in network namespace 2
> 
> Both are in "bridge" mode, so they can communicate together.
> The lower device is dummy0 in the init network namespace.
> 
> IMO the problem is coming from the macvlan driver:
> 
> dev->features           = lowerdev->features & MACVLAN_FEATURES
> 
> As dummy0 has the offloading capabilities set on, the macvlan driver 
> inherit these features.
> 
> In the normal case, dummy0 is supposed to drop the packets. But with 
> macvlan these packets are broadcasted to the other macvlan ports, so no 
> checksum is computed when the packets are transmitted between macvlan1 
> and macvlan2.

So where frames get bad checksums ?

In this "bridge" mode, I suspect the broadcast is done _before_ sending
frame to dummy, so maybe macvlan should not inherit from lowerdev in
this particular case ?