Skip to contentSkip to conversation

Load balancing using connmark

LARTC Francis Brosnan Blazquez 5 participants 12 messages May 9, 2007 – May 10, 2007

lore
Share & tools
lore
#0 lore raw

Hi,

I've been implementing a load balancing solution using CONNMARK, based on solution described by Luciano Ruete at [1]. Gracias por el post y por apuntar en la direcci�n correcta Luciano!

Once implemented, I've found that due to some reason packets aren't properly marked (or improperly remarked) and sent out using the wrong interface.

My topo setup is:

[82.123.136.74]: eth1 : mark:0x1 --\
                                    +--[FW BOX] -- eth0: 192.168.0.53

[217.146.74.82]: eth2 : mark:0x2 --/

Using conntrack tool, shows that after a while, it starts to appear packets marked with 0x2 or 0x1 not comming from the proper source IP.

Quoted 1 line
Quoted 1 line

conntrack -L | grep mark=2 | grep '82.123.136.74'; conntrack -L |

grep mark=1 | grep '217.146.74.82'

tcp 6 425543 ESTABLISHED src2.168.0.178 dst�.216.53.249
sport52 dportD3 packets�8 bytes�471 src�.216.53.249 dst�.123.136.74 sportD3 dport52 packets�5 bytes�909 [ASSURED] mark=2 use=1
tcp 6 428681 ESTABLISHED src2.168.0.177 dst�.139.122.12
sport#61 dportD3 packets2 bytes)381 src�.139.122.12 dst�.123.136.74 sportD3 dport#61 packets9 bytes120 [ASSURED] mark=2 use=1

This is quite odd since solution proposed at [1] looks good. I'll cite it here for clarity (suppose I already have all ip rule stuff installed):

iptables -t mangle -A POSTROUTING -m mark --mark ! 0 -j ACCEPT iptables -t mangle -A POSTROUTING -o eth1 -j MARK --set-mark 0x1 iptables -t mangle -A POSTROUTING -o eth2 -j MARK --set-mark 0x2 iptables -t mangle -A POSTROUTING -j CONNMARK --save-mark iptables -t mangle -A PREROUTING -j CONNMARK --restore-mark

After giving a try during several days, I've found that another firewall solution, shorewall [2], implements built-in load balacing for free by using the following set of instructions:

iptables -t mangle -A PREROUTING -m connmark ! --mark 0/0xFF -j CONNMARK --restore-mark --mask 0xFF

iptables -t mangle -A OUTPUT -m connmark ! --mark 0/0xFF -j CONNMARK --restore-mark --mask 0xFF

iptables -t mangle -N routemark
iptables -t mangle -A PREROUTING -i eth1 -m mark --mark 0/0xFF -j routemark

iptables -t mangle -A routemark -i eth1 -j MARK --set-mark 1 iptables -t mangle -A PREROUTING -i eth2 -m mark --mark 0/0xFF -j routemark

iptables -t mangle -A routemark -i eth2 -j MARK --set-mark 2 iptables -t mangle -A routemark -m mark ! --mark 0/0xFF -j CONNMARK --save-mark --mask 0xFF

After a bit of testing with the second solution, it seems to behave better, doing all marking job at the PREROUTING and OUTPUT.

Did anybody find that some packages doesn't get properly routed according to the mark with the first solution? What you do think about the second solution?

Cheers!

[1] http://mailman.ds9a.nl/pipermail/lartc/2006q2/018964.html
[2] http://www.shorewall.net

Signature · 7 lines
Francis Brosnan Blazquez <[email protected]>
Advanced Software Production Line, S.L.

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc
#1 lore raw
Quoted 10 lines · Francis Brosnan Blazquez

Hi,

I've been implementing a load balancing solution using CONNMARK, based on solution described by Luciano Ruete at [1]. Gracias por el post y por apuntar en la dirección correcta Luciano!

Once implemented, I've found that due to some reason packets aren't properly marked (or improperly remarked) and sent out using the wrong interface.

Quoted 6 lines

<snip>

iptables -t mangle -A POSTROUTING -m mark --mark ! 0 -j ACCEPT iptables -t mangle -A POSTROUTING -o eth1 -j MARK --set-mark 0x1 iptables -t mangle -A POSTROUTING -o eth2 -j MARK --set-mark 0x2 iptables -t mangle -A POSTROUTING -j CONNMARK --save-mark

This is wrong. POSTROUTING is exactly what is is _POST_ routing. By the time you do your marks and stuff the kernel has _already_ assigned a packet to an interface, and you can not alter this anymore.

Quoted 2 lines

After a bit of testing with the second solution, it seems to behave better, doing all marking job at the PREROUTING and OUTPUT.

This is flawed too. OUTPUT suffers from the very same problem as POSTROUTING - by the time the packets hit the NF stack the process has already bound itself to an interface, which you can not change anymore.

Peter

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#2 lore raw
In reply to Francis Brosnan Blazquez
Quoted 4 lines · Francis Brosnan Blazquez

Hi,

I've been implementing a load balancing solution using CONNMARK, based on solution described by Luciano Ruete at [1]. Gracias por el post y

por

Quoted 6 lines

apuntar en la dirección correcta Luciano!

Once implemented, I've found that due to some reason packets aren't properly marked (or improperly remarked) and sent out using the wrong interface.

Quoted 6 lines

<snip>

iptables -t mangle -A POSTROUTING -m mark --mark ! 0 -j ACCEPT iptables -t mangle -A POSTROUTING -o eth1 -j MARK --set-mark 0x1 iptables -t mangle -A POSTROUTING -o eth2 -j MARK --set-mark 0x2 iptables -t mangle -A POSTROUTING -j CONNMARK --save-mark

This is wrong. POSTROUTING is exactly what is is _POST_ routing. By the time you do your marks and stuff the kernel has _already_ assigned a packet to an interface, and you can not alter this anymore.

Quoted 2 lines

After a bit of testing with the second solution, it seems to behave better, doing all marking job at the PREROUTING and OUTPUT.

This is flawed too. OUTPUT suffers from the very same problem as POSTROUTING - by the time the packets hit the NF stack the process has already bound itself to an interface, which you can not change anymore.

Peter

Disagree with Peter. The marking in postrouting table is CONNMARK. This is for marking the connection, which has already had a route decided for it, so that all packets of the connection passes through this interface. This marking is done for packets with NEW state, see the check for mark==0 in the prev. line. The restore mark in PREROUTING will restore the connmark and route the subsequent packets.
This approach will work, but you need some sort of stateful-ness in netfilter.

The second point in Brosnan Blazquez’s mail about shorewall: They seem to be doing Policy Routing, not real load balancing.

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#3 lore raw
In reply to Francis Brosnan Blazquez

On closer look, I am wrong about shorewall. It seems to be a different approach to load balancing. They connmark the incoming packets from WAN, rather than outgoing packets. I think it should work well, but I wonder why this approach is not popular. There must be some drawback to it. I can’t think of one,though.

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Salim S I
Sent: Thursday, May 10, 2007 2:15 PM
To: [email protected]
Subject: Re: [LARTC] Load balancing using connmark

Quoted 4 lines · Francis Brosnan Blazquez

Hi,

I've been implementing a load balancing solution using CONNMARK, based on solution described by Luciano Ruete at [1]. Gracias por el post y

por

Quoted 6 lines

apuntar en la dirección correcta Luciano!

Once implemented, I've found that due to some reason packets aren't properly marked (or improperly remarked) and sent out using the wrong interface.

Quoted 6 lines

<snip>

iptables -t mangle -A POSTROUTING -m mark --mark ! 0 -j ACCEPT iptables -t mangle -A POSTROUTING -o eth1 -j MARK --set-mark 0x1 iptables -t mangle -A POSTROUTING -o eth2 -j MARK --set-mark 0x2 iptables -t mangle -A POSTROUTING -j CONNMARK --save-mark

This is wrong. POSTROUTING is exactly what is is _POST_ routing. By the time you do your marks and stuff the kernel has _already_ assigned a packet to an interface, and you can not alter this anymore.

Quoted 2 lines

After a bit of testing with the second solution, it seems to behave better, doing all marking job at the PREROUTING and OUTPUT.

This is flawed too. OUTPUT suffers from the very same problem as POSTROUTING - by the time the packets hit the NF stack the process has already bound itself to an interface, which you can not change anymore.

Peter

Disagree with Peter. The marking in postrouting table is CONNMARK. This is for marking the connection, which has already had a route decided for it, so that all packets of the connection passes through this interface. This marking is done for packets with NEW state, see the check for mark==0 in the prev. line. The restore mark in PREROUTING will restore the connmark and route the subsequent packets.
This approach will work, but you need some sort of stateful-ness in netfilter.

The second point in Brosnan Blazquez’s mail about shorewall: They seem to be doing Policy Routing, not real load balancing.

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#4 lore raw
In reply to Francis Brosnan Blazquez

El jue, 10-05-2007 a las 16:01 +0800, Salim S I escribió:
Hi Salim,

Thanks for your reply,

Quoted 5 lines

On closer look, I am wrong about shorewall. It seems to be a different approach to load balancing. They connmark the incoming packets from WAN, rather than outgoing packets. I think it should work well, but I wonder why this approach is not popular. There must be some drawback to it. I can’t think of one,though.

I think the main advantage of shorewall solution is that it applies connmark to incoming packets from the wan as you point, leaving load balancing to outgoing connections to the main table.

In any case, with this second solution I don't see wrong routed packages on wan interfaces using tcpdump, whereas with the first solution I do. More testing is required.

Regarding to your previous reply, can you elaborate more on "...This approach will work, but you need some sort of stateful-ness in netfilter..."

Cheers!

Signature · 7 lines
Francis Brosnan Blazquez <[email protected]>
Advanced Software Production Line, S.L.

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc
#5 lore raw
In reply to Francis Brosnan Blazquez

-----Original Message-----
From: Salim S I [mailto:[email protected]]
Sent: Thursday, May 10, 2007 5:22 PM
To: 'Francis Brosnan Blazquez'
Subject: RE: [LARTC] Load balancing using connmark

"I think the main advantage of shorewall solution is that it applies connmark to incoming packets from the wan as you point, leaving load balancing to outgoing connections to the main table"

Actually, the main table/multipath route only routes the first packet of a connection. The subsequent routing for that connection is done based on connmark, for outgoing packets too. Otherwise replies to packets coming from WAN1 may go through WAN2. The difference in the two solutions is only in where packets are marked and which packets are marked. Routing is the same.

For a detailed discussion on the first approach, you can refer to this thread.

http://mailman.ds9a.nl/pipermail/lartc/2006q2/018964.html

-----Original Message-----
From: Francis Brosnan Blazquez [mailto:[email protected]]
Sent: Thursday, May 10, 2007 5:07 PM
To: Salim S I
Cc: [email protected]
Subject: RE: [LARTC] Load balancing using connmark

El jue, 10-05-2007 a las 16:01 +0800, Salim S I escribió:
Hi Salim,

Thanks for your reply,

Quoted 5 lines

On closer look, I am wrong about shorewall. It seems to be a different approach to load balancing. They connmark the incoming packets from WAN, rather than outgoing packets. I think it should work well, but I wonder why this approach is not popular. There must be some drawback to it. I can’t think of one,though.

I think the main advantage of shorewall solution is that it applies connmark to incoming packets from the wan as you point, leaving load balancing to outgoing connections to the main table.

In any case, with this second solution I don't see wrong routed packages on wan interfaces using tcpdump, whereas with the first solution I do. More testing is required.

Regarding to your previous reply, can you elaborate more on "...This approach will work, but you need some sort of stateful-ness in netfilter..."

Cheers!

Signature · 8 lines
Francis Brosnan Blazquez <[email protected]>
Advanced Software Production Line, S.L.


_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc
#6 lore raw
In reply to Francis Brosnan Blazquez

hi people

Quoted 5 lines · Francis Brosnan Blazquez

I've been implementing a load balancing solution using CONNMARK, based

After giving a try during several days, I've found that another firewall solution, shorewall [2], implements built-in load balacing for free by using the following set of instructions:

did somebody try the shorewall solution with centos 4?

with centos 4 and the first solution i always had the problem, that it routes correctly only for passing through connections (forwarded). connections starting from the machine or hoing to the machine (input/output chain) had exactly the same behaviour as you stated before.

i noticed with centos 4 that packets do not pass the prerouting magle chain if going to the local host (passing the input filter chain thereafter). therefore certainly the mark will not be restored and there will be no influence on the routing decision.
someone noticed similar behaviour?

peter

Signature · 9 lines
:: e n d i a n
:: open source - open minds

:: peter warasin
:: http://www.endian.com   :: [email protected]
_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc
#7 lore raw
In reply to Francis Brosnan Blazquez
Quoted 69 lines · Salim S I
Quoted 24 lines · Francis Brosnan Blazquez

Hi,

I've been implementing a load balancing solution using CONNMARK, based

on solution described by Luciano Ruete at [1]. Gracias por el post y por

apuntar en la dirección correcta Luciano!

Once implemented, I've found that due to some reason packets aren't

properly marked (or improperly remarked) and sent out using the wrong

interface.

<snip>

Quoted 7 lines

iptables -t mangle -A POSTROUTING -m mark --mark ! 0 -j ACCEPT

iptables -t mangle -A POSTROUTING -o eth1 -j MARK --set-mark 0x1

iptables -t mangle -A POSTROUTING -o eth2 -j MARK --set-mark 0x2

iptables -t mangle -A POSTROUTING -j CONNMARK --save-mark

This is wrong. POSTROUTING is exactly what is is _POST_ routing. By the

time you do your marks and stuff the kernel has _already_ assigned a

packet to an interface, and you can not alter this anymore.

Quoted 3 lines

After a bit of testing with the second solution, it seems to behave

better, doing all marking job at the PREROUTING and OUTPUT.

This is flawed too. OUTPUT suffers from the very same problem as

POSTROUTING - by the time the packets hit the NF stack the process has

already bound itself to an interface, which you can not change anymore.

Peter

Disagree with Peter. The marking in postrouting table is CONNMARK. This is for marking the connection, which has already had a route decided for it, so that all packets of the connection passes through this interface. This marking is done for packets with NEW state, see the check for mark=0 in the prev. line. The restore mark in PREROUTING will restore the connmark and route the subsequent packets.

Quoted 3 lines

This approach will work, but you need some sort of stateful-ness in netfilter.

Connmark is exactly the statefullness you are talking about. The problem is that the marks by themselves do not mean anything. You mark packets and expect iproute to classify the packet in the correct routing table etc. CONNMARK is invisible to iproute - this is why you have only --save-mark and --restore-mark, and the rest of the rules deal with real MARKs.

Further you (and the OP) seem to be confused by a mix of routing tasks. In the case of _forwarded_ traffic, you need to make sure that all packets within a connection leave to WAN over the same interface, and are SNATed to the same ip, so that they will come bak the same interface. The SNATting is trivial (as it can be done in POSTROUTING only), but you need to set all marks before the routing takes place (which is anywhere _but_ POSTROUTING). You might mark the connection with the proper CONNMARK. and subsequent packets might get routed correctly, but the _first_ packet (the one that you use to set the mark) is already assigned to an interface, and there is nothing you can do about it.

In the case of _local_ traffic - it becomes even trickier. The problem is that when sockets are created they already have a source IP (the kernel determines that by looking at the default routing table, your marks do not exist yet). So since you can not alter the socket binding, the only way to make it leave on a different interface is by treating it as a forwarded connection and performing NAT on it. It is arguable if NATting locally originating connections is a good idea, but it can be done in OUTPUT just like it is done for forwarder connections in PREROUTING.

I hope this clarifies things a bit, feel free to point out any inconsistencies you may find.

Peter

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#8 lore raw
In reply to Francis Brosnan Blazquez
Quoted 5 lines · Peter Rabbitson

...
In the case of _local_ traffic - it becomes even trickier. The problem is that when sockets are created they already have a source IP (the kernel determines that by looking at the default routing table, your marks do not exist yet).

This is misleading - it will happen only when the application does not request a specific ip/interface to bind to. Only then the kernel default table is consulted, and the best interface is determined based on the destination that is supplied on socket creation.
_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#9 lore raw
In reply to Francis Brosnan Blazquez

Let me explain why the marking is done in POSTROUTING.

The first packet of any connection get routed by the multipath routing entry. This happens AFTER PREROUTING, as you know. And this is what we want, letting the kernel decide based on the weights. (some people do think that we shouldn't let multipath decide routing, but thatz a different story).

So where can this packet be marked? Obviously in POSTROUTING (so that local pkts also can be caught). We mark it and save it.(connmark).The mark is decoded by the chosen interface. (eg:-o WAN1 --set mark 1,-o WAN2 --set-mark 2)

In PREROUTING, there is a restore-mark. You see

iptables -t mangle -A PREROUTING -j CONNMARK --restore-mark.

If this packet belong to a connection that has already sent a packet,this will restore the mark set in POSTROUTING. Then it will be routed by the corresponding routing table.(wan1 table lookup mark1 and wan2 table lookup mark2)
If it is a new pkt, it will be routed by multipath routing statement,since no mark exists.

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Peter Rabbitson
Sent: Thursday, May 10, 2007 6:51 PM
Cc: [email protected]
Subject: Re: [LARTC] Load balancing using connmark

Quoted 7 lines · Salim S I
Quoted 5 lines · Francis Brosnan Blazquez

Hi,

I've been implementing a load balancing solution using CONNMARK,

based

Quoted 2 lines
Quoted 2 lines

on solution described by Luciano Ruete at [1]. Gracias por el post y

por

Quoted 21 lines
Quoted 20 lines

apuntar en la dirección correcta Luciano!

Once implemented, I've found that due to some reason packets aren't

properly marked (or improperly remarked) and sent out using the wrong

interface.

<snip>

iptables -t mangle -A POSTROUTING -m mark --mark ! 0 -j ACCEPT

iptables -t mangle -A POSTROUTING -o eth1 -j MARK --set-mark 0x1

Quoted 7 lines
Quoted 3 lines

iptables -t mangle -A POSTROUTING -o eth2 -j MARK --set-mark 0x2

iptables -t mangle -A POSTROUTING -j CONNMARK --save-mark

This is wrong. POSTROUTING is exactly what is is _POST_ routing. By

the

Quoted 13 lines

time you do your marks and stuff the kernel has _already_ assigned a

packet to an interface, and you can not alter this anymore.

Quoted 3 lines

After a bit of testing with the second solution, it seems to behave

better, doing all marking job at the PREROUTING and OUTPUT.

Quoted 5 lines

This is flawed too. OUTPUT suffers from the very same problem as

POSTROUTING - by the time the packets hit the NF stack the process has

already bound itself to an interface, which you can not change

anymore.

Quoted 8 lines

Peter

Disagree with Peter. The marking in postrouting table is CONNMARK.

This

Quoted 1 line

is for marking the connection, which has already had a route decided

for

Quoted 1 line

it, so that all packets of the connection passes through this

interface.

Quoted 7 lines

This marking is done for packets with NEW state, see the check for mark=0 in the prev. line. The restore mark in PREROUTING will restore the connmark and route the subsequent packets.

This approach will work, but you need some sort of stateful-ness in netfilter.

Connmark is exactly the statefullness you are talking about. The problem is that the marks by themselves do not mean anything. You mark packets and expect iproute to classify the packet in the correct routing table etc. CONNMARK is invisible to iproute - this is why you have only --save-mark and --restore-mark, and the rest of the rules deal with real MARKs.

Further you (and the OP) seem to be confused by a mix of routing tasks. In the case of _forwarded_ traffic, you need to make sure that all packets within a connection leave to WAN over the same interface, and are SNATed to the same ip, so that they will come bak the same interface. The SNATting is trivial (as it can be done in POSTROUTING only), but you need to set all marks before the routing takes place (which is anywhere _but_ POSTROUTING). You might mark the connection with the proper CONNMARK. and subsequent packets might get routed correctly, but the _first_ packet (the one that you use to set the mark) is already assigned to an interface, and there is nothing you can do about it.

In the case of _local_ traffic - it becomes even trickier. The problem is that when sockets are created they already have a source IP (the kernel determines that by looking at the default routing table, your marks do not exist yet). So since you can not alter the socket binding, the only way to make it leave on a different interface is by treating it as a forwarded connection and performing NAT on it. It is arguable if NATting locally originating connections is a good idea, but it can be done in OUTPUT just like it is done for forwarder connections in PREROUTING.

I hope this clarifies things a bit, feel free to point out any inconsistencies you may find.

Peter

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

#11 lore raw
In reply to Francis Brosnan Blazquez
Quoted 5 lines · Salim S I

Let me explain why the marking is done in POSTROUTING.

want, letting the kernel decide based on the weights. (some people do think that we shouldn't let multipath decide routing, but thatz a different story).

I apologize, as I am one of these people, and subsequently assumed the OP wanted this. In this light I agree with Salim. On an unrelated note the OP should be aware that letting multipath do the balancing is impractical (i.e. does not work) in real life scenarios, but this indeed is a topic for a separate thread.
_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc