CentOS virt - Apr 2015 - Seeing dropped packets / tcp retrans on latest 4.4.1-10el6

So I might have been misinterpreting things here and might be way off base. I
think you can ignore this thread and I'll follow up if I get anything
concrete down the road =) The retranmissions I'm seeing and reproducing are
probably within normal allowances and can't reproduce the issue that
originally lead me down this path.

- Nathan
> -----Original Message-----
> From: centos-virt-bounces at centos.org [mailto:centos-virt-
> bounces at centos.org] On Behalf Of Nathan March
> Sent: Wednesday, April 15, 2015 1:13 PM
> To: 'Discussion about the virtualization on CentOS'
> Subject: Re: [CentOS-virt] Seeing dropped packets / tcp retrans on latest
> 4.4.1-10el6
> 
> Hi All,
> 
> Some more data on this, I've reproduced this on another host that's
a
> completely stock centos/xen deployment with a centos 6.6 domU.
> 
> Since I?m seeing the retransmissions on the VIF, I don't think it's
related to
> the network stack but just in case.. Each host is connected via LACP with
vlan
> tagging to a pair of stacked cisco 3750's. Host networking config is
here:
> 
> http://dpaste.com/1Q6NY3Y
> 
> The vm is on br99 here.
> 
> This is easily reproducable by just generating a 250mb random file and
doing
> an scp, while watching with tshark:
> 
> tshark -R "tcp.analysis.retransmission"
> 
> There's no visible impact to the connection the vast majority of the
time,
> which is why I think this has gone unnoticed.
> 
> Just to confirm this wasn't related to hardware / nics, I've
reproduced this on:
> 
>  - Dell PowerEdge M620 with broadcom nics
>  - Dell C6220 with intel nics
>  - Supermicro X8DTT with intel nics
> 
> Any ideas? =)
> 
> - Nathan
> 
> _______________________________________________
> CentOS-virt mailing list
> CentOS-virt at centos.org
> http://lists.centos.org/mailman/listinfo/centos-virt

Hi All,

I've tracked this down... We do rate limiting of our vms with a mix of
ebtables/tc.

Running these commands (replace vif1.0 with the correct vif for your VM) will
reproduce this:

ebtables -A FORWARD -i vif1.0 -j mark --set-mark 990 --mark-target CONTINUE

tc qdisc add dev bond0 root handle 1: htb default 2 
tc class add dev bond0 parent 1: classid 1:0 htb rate 10000mbit 

tc class add dev bond0 parent 1: classid 1:990 htb rate 10000mbit
tc filter add dev bond0 protocol ip parent 1:0 prio 990 handle 990 fw flowid
1:990

Note that the speed limits being applied here are 10gb and I'm testing this
on a 1gb network, so TC shouldn't really be doing anything here except
letting the packets through. These same commands worked fine on gentoo xen 4.1 /
kernel 3.2.57, compared to this now not working on centos xen 4.4.1 / kernel
3.10.68.

Easiest way to reproduce is simply generate a large file, scp it to a remote
host and on the remote host run:
tshark -Y "tcp.analysis.duplicate_ack_num"

If you run the ssh in a loop + tshark in another window, you can see the Dup
ACK's begin immediately after adding the last filter rule:

25790294 1752.756733 xxx.xxx.xxx.13 -> xxx.xxx.xxx.205 TCP 78 [TCP Dup ACK
25790286#4] ssh > 51515 [ACK] Seq=15994 Ack=50769840 Win=1544704 Len=0
TSval=738150929 TSecr=4294944346 SLE=50785768 SRE=50790596
25790296 1752.756742 xxx.xxx.xxx.13 -> xxx.xxx.xxx.205 TCP 78 [TCP Dup ACK
25790286#5] ssh > 51515 [ACK] Seq=15994 Ack=50769840 Win=1544704 Len=0
TSval=738150929 TSecr=4294944346 SLE=50785768 SRE=50792044

- Nathan

[cc'ing xen-users to see if anyone there has any familiarity with ebtables]

On Fri, Apr 17, 2015 at 8:20 PM, Nathan March <nathan at gt.net>
wrote:
> Hi All,
>
> I've tracked this down... We do rate limiting of our vms with a mix of
ebtables/tc.
>
> Running these commands (replace vif1.0 with the correct vif for your VM)
will reproduce this:
>
> ebtables -A FORWARD -i vif1.0 -j mark --set-mark 990 --mark-target CONTINUE
>
> tc qdisc add dev bond0 root handle 1: htb default 2
> tc class add dev bond0 parent 1: classid 1:0 htb rate 10000mbit
>
> tc class add dev bond0 parent 1: classid 1:990 htb rate 10000mbit
> tc filter add dev bond0 protocol ip parent 1:0 prio 990 handle 990 fw
flowid 1:990
>
> Note that the speed limits being applied here are 10gb and I'm testing
this on a 1gb network, so TC shouldn't really be doing anything here except
letting the packets through. These same commands worked fine on gentoo xen 4.1 /
kernel 3.2.57, compared to this now not working on centos xen 4.4.1 / kernel
3.10.68.
So just to be clear, we have 3 variables to consider here?

* gentoo -> CentOS
* Xen 4.1 -> Xen 4.4
* kernel 3.2.57 -> 3.10.68

Unfortunately I'm not very familiar with ebtables, so I don't have a
clear idea what sort of thing might cause duplicate ACKs.

Are you able to narrow down any of those?

You can find CentOS packages for Xen 4.2 and kernel 3.4 here:

http://vault.centos.org/6.4/xen4/x86_64/Packages/

If you could build a more recent kernel and see if it's been fixed,
that might be helpful as well.

 -George
>
> Easiest way to reproduce is simply generate a large file, scp it to a
remote host and on the remote host run:
> tshark -Y "tcp.analysis.duplicate_ack_num"
>
> If you run the ssh in a loop + tshark in another window, you can see the
Dup ACK's begin immediately after adding the last filter rule:
>
> 25790294 1752.756733 xxx.xxx.xxx.13 -> xxx.xxx.xxx.205 TCP 78 [TCP Dup
ACK 25790286#4] ssh > 51515 [ACK] Seq=15994 Ack=50769840 Win=1544704 Len=0
TSval=738150929 TSecr=4294944346 SLE=50785768 SRE=50790596
> 25790296 1752.756742 xxx.xxx.xxx.13 -> xxx.xxx.xxx.205 TCP 78 [TCP Dup
ACK 25790286#5] ssh > 51515 [ACK] Seq=15994 Ack=50769840 Win=1544704 Len=0
TSval=738150929 TSecr=4294944346 SLE=50785768 SRE=50792044
>
> - Nathan
>
> _______________________________________________
> CentOS-virt mailing list
> CentOS-virt at centos.org
> http://lists.centos.org/mailman/listinfo/centos-virt