Message-ID: <CAFJtzm3pwAXKOxYLi+-EgCXYxA90UCGvRvn=qW=HD9AKzoheSQ@mail.gmail.com>
Date:   Tue, 31 Mar 2020 15:48:33 +0200
From:   Charles DAYMAND <charles.daymand@...irst.fr>
To:     netdev <netdev@...r.kernel.org>
Subject: Re: [PATCH net] r8169: fix multicast tx issue with macvlan interface

Hello,
We tested to enable tx checksumming manually (via ethtool) on a kernel
4.19.0-5-amd64 which is the oldest kernel compatible with our software
and we observed exactly the same issue.
For information when connecting a laptop directly to the interface we
can't see any multicast packet when tx checksumming is enabled on
tcpdump.
Our network is composed of a cisco switch and we can still see the
multicast counters correctly increasing even when we have the issue.

I also confirm that when not using macvlan but the real interface
there is no issue on the multicast packets, they are correctly sent
and received.
I have a stupid question, if the IP checksum was bad on the multicast
packet, would the receiver NIC drop the packet or would it be seen by
tcpdump by the receiver ?

Le ven. 27 mars 2020 à 20:43, Eric Dumazet <edumazet@...gle.com> a écrit :
>
> On Fri, Mar 27, 2020 at 12:17 PM Heiner Kallweit <hkallweit1@...il.com> wrote:
> >
> > On 27.03.2020 19:52, Eric Dumazet wrote:
> > > On Fri, Mar 27, 2020 at 10:41 AM Heiner Kallweit <hkallweit1@...il.com> wrote:
> > >>
> > >> On 27.03.2020 10:39, Heiner Kallweit wrote:
> > >>> On 27.03.2020 10:08, Charles Daymand wrote:
> > >>>> During kernel upgrade testing on our hardware, we found that macvlan
> > >>>> interface were no longer able to send valid multicast packet.
> > >>>>
> > >>>> tcpdump run on our hardware was correctly showing our multicast
> > >>>> packet but when connecting a laptop to our hardware we didn't see any
> > >>>> packets.
> > >>>>
> > >>>> Bisecting turned up commit 93681cd7d94f
> > >>>> "r8169: enable HW csum and TSO" activates the feature NETIF_F_IP_CSUM
> > >>>> which is responsible for the drop of packet in case of macvlan
> > >>>> interface. Note that revision RTL_GIGA_MAC_VER_34 was already a specific
> > >>>> case since TSO was keep disabled.
> > >>>>
> > >>>> Deactivating NETIF_F_IP_CSUM using ethtool is correcting our multicast
> > >>>> issue, but we believe that this hardware issue is important enough to
> > >>>> keep tx checksum off by default on this revision.
> > >>>>
> > >>>> The change is deactivating the default value of NETIF_F_IP_CSUM for this
> > >>>> specific revision.
> > >>>>
> > >>>
> > >>> The referenced commit may not be the root cause but just reveal another
> > >>> issue that has been existing before. Root cause may be in the net core
> > >>> or somewhere else. Did you check with other RTL8168 versions to verify
> > >>> that it's indeed a HW issue with this specific chip version?
> > >>>
> > >>> What you could do: Enable tx checksumming manually (via ethtool) on
> > >>> older kernel versions and check whether they are fine or not.
> > >>> If an older version is fine, then you can start a new bisect with tx
> > >>> checksumming enabled.
> > >>>
> > >>> And did you capture and analyze traffic to verify that actually the
> > >>> checksum is incorrect (and packets discarded therefore on receiving end)?
> > >>>
> > >>>
> > >>>> Fixes: 93681cd7d94f ("r8169: enable HW csum and TSO")
> > >>>> Signed-off-by: Charles Daymand <charles.daymand@...irst.fr>
> > >>>> ---
> > >>>>  net/drivers/net/ethernet/realtek/r8169_main.c | 3 +++
> > >>>>  1 file changed, 3 insertions(+)
> > >>>>
> > >>>> diff --git a/net/drivers/net/ethernet/realtek/r8169_main.c b/net/drivers/net/ethernet/realtek/r8169_main.c
> > >>>> index a9bdafd15a35..3b69135fc500 100644
> > >>>> --- a/net/drivers/net/ethernet/realtek/r8169_main.c
> > >>>> +++ b/net/drivers/net/ethernet/realtek/r8169_main.c
> > >>>> @@ -5591,6 +5591,9 @@ static int rtl_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
> > >>>>              dev->vlan_features &= ~(NETIF_F_ALL_TSO | NETIF_F_SG);
> > >>>>              dev->hw_features &= ~(NETIF_F_ALL_TSO | NETIF_F_SG);
> > >>>>              dev->features &= ~(NETIF_F_ALL_TSO | NETIF_F_SG);
> > >>>> +            if (tp->mac_version == RTL_GIGA_MAC_VER_34) {
> > >>>> +                    dev->features &= ~NETIF_F_IP_CSUM;
> > >>>> +            }
> > >>>>      }
> > >>>>
> > >>>>      dev->hw_features |= NETIF_F_RXALL;
> > >>>>
> > >>>
> > >>
> > >> After looking a little bit at the macvlen code I think there might be an
> > >> issue in it, but I'm not sure, therefore let me add Eric (as macvlen doesn't
> > >> seem to have a dedicated maintainer).
> > >>
> > >> r8169 implements a ndo_features_check callback that disables tx checksumming
> > >> for the chip version in question and small packets (due to a HW issue).
> > >> macvlen uses passthru_features_check() as ndo_features_check callback, this
> > >> seems to indicate to me that the ndo_features_check callback of lowerdev is
> > >> ignored. This could explain the issue you see.
> > >>
> > >
> > > macvlan_queue_xmit() calls dev_queue_xmit_accel() after switching skb->dev,
> > > so the second __dev_queue_xmit() should eventually call the real_dev
> > > ndo_features_check()
> > >
> > Thanks, Eric. There's a second path in macvlan_queue_xmit() calling
> > dev_forward_skb(vlan->lowerdev, skb). Does what you said apply also there?
>
> This path wont send packets to the physical NIC, packets are injected
> back via dev_forward_skb()
>
> >
> > Still I find it strange that a tx hw checksumming issue should affect multicasts
> > only. Also the chip version in question is quite common and I would expect
> > others to have hit the same issue.
> > Maybe best would be to re-test on the affected system w/o involving macvlen.
> >
> > >
> > >
> > >> Would be interesting to see whether it fixes your issue if you let the
> > >> macvlen ndo_features_check call lowerdev's ndo_features_check. Can you try this?
> > >>
> > >> By the way:
> > >> Also the ndo_fix_features callback of lowerdev seems to be ignored.
> >



-- 

Charles Daymand

Développeur infrastructure

26 rue de Berri 75008 Paris

Assistance dédiée responsable de site - 01 70 70 46 70
Assistance utilisateur - 01 70 70 46 26



-- 

Charles Daymand

Développeur infrastructure

26 rue de Berri 75008 Paris

Assistance dédiée responsable de site - 01 70 70 46 70
Assistance utilisateur - 01 70 70 46 26

Powered by blists - more mailing lists