neighbor table overflow

Post by Marco C. Coelho
# Added to stop "neighbor table overflow" messages in the kernel
net.ipv4.neigh.default.gc_thresh1=512
net.ipv4.neigh.default.gc_thresh2=2048
net.ipv4.neigh.default.gc_thresh3=4096
# Added to increase IP contrack number (was getting to max)
net.ipv4.ip_conntrack_max=99999
to sysctl.conf to increase the size, but this only seems to delay the
problem.
Any thoughts?

<skip />

try arpd?

--
Peter V. Saveliev

Grant Taylor

2007-10-22 22:35:19 UTC

Post by Peter V. Saveliev
try arpd?

You took the statement right out from under my finger tips.

Grant. . . .

Alexandru Dragoi

2007-10-23 11:56:48 UTC

Post by Marco C. Coelho
I've got a linux router pushing 600-1000 pppoe connections through
it. I'm getting a screen error "Neighbor Table Overflow" after this
box has been up for between 1 week and 1 month. When this is
# Added to stop "neighbor table overflow" messages in the kernel
net.ipv4.neigh.default.gc_thresh1=512
net.ipv4.neigh.default.gc_thresh2=2048
net.ipv4.neigh.default.gc_thresh3=4096
# Added to increase IP contrack number (was getting to max)
net.ipv4.ip_conntrack_max=99999
to sysctl.conf to increase the size, but this only seems to delay the
problem.
Any thoughts?
Marco
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

What about checking your routing table? you may have link routes for
massive subnets (like 85.0.0.0/8 or 140.20.0.0/16). Some programs prefer
to use "standard" netmask of classes A and B.

# ip rou |grep link

Grant Taylor

2007-10-23 20:32:19 UTC

Post by Alexandru Dragoi
What about checking your routing table? you may have link routes for
massive subnets (like 85.0.0.0/8 or 140.20.0.0/16). Some programs
prefer to use "standard" netmask of classes A and B.

I'm betting that the OP has other things going on seeing has how s/he
mentioned PPPoE, which to my knowledge is a layer 2 protocol, and thus
not subject to typical routing scenarios. In essence the OP could have
thousands of PPPoE connections terminating on one system with the ARP
cache having to deal with where to send traffic to which MAC address.
There is not a lot of room for routing in such a scenario.

Grant. . . .

Jeff Welling

2007-10-23 20:43:06 UTC

Post by Alexandru Dragoi
What about checking your routing table? you may have link routes
for massive subnets (like 85.0.0.0/8 or 140.20.0.0/16). Some
programs prefer to use "standard" netmask of classes A and B.

I'm betting that the OP has other things going on seeing has how s/
he mentioned PPPoE, which to my knowledge is a layer 2 protocol,
and thus not subject to typical routing scenarios. In essence the
OP could have thousands of PPPoE connections terminating on one
system with the ARP cache having to deal with where to send traffic
to which MAC address. There is not a lot of room for routing in
such a scenario.

I agree with Peter's suggestion, arpd. I ran into the neighbor table
overflow problem recently, at the hands of our ISP. I was in the
process of recompiling the kernel and mucking with arpd (I couldn't
get it to run/start properly) when the problem disappeared as quickly
as it showed up. Lucky for me, this was some kind of ISP problem, I
was able to determine that much through `tcpdump -i X -n arpd`.

My 'two cents' is that you try arpd, I did a bit of looking when I
came across that problem and it seemed to be the last ditch effort
when changing the gc threshold had no effect. Wasn't able to confirm
that it worked for sure though.

Cheers.

Marco C. Coelho

2007-10-23 21:10:32 UTC

This box is doing a lot. It terminates 1000 PPPoE connections, provides
traffic shaping using TC/HTB, authenticates all users via Radius. It
also runs OSPF routing for the internal network. Looking at a simple
route output I see all the PPP connections coming through the box, and
due to the OSPF I also see the rest of my network announcements. The
only strange things are:

1. The last man working on this box had mistakenly edited the hosts
file and added the machine name and complete domain name to the local
host 127.0.0.1 name. It should only be pointed to the eth0 interface.
I have changed this.

2. The route output is making an announcement

64.0.0.0 argontech.net 255.0.0.0 UG 20
0 0 eth0

My public IP space is a /20 within that space, not the whole Class A. I
have not found which box is announcing this within my network yet.

I'm betting that the OP has other things going on seeing has how s/he
mentioned PPPoE, which to my knowledge is a layer 2 protocol, and
thus not subject to typical routing scenarios. In essence the OP
could have thousands of PPPoE connections terminating on one system
with the ARP cache having to deal with where to send traffic to which
MAC address. There is not a lot of room for routing in such a scenario.

I agree with Peter's suggestion, arpd. I ran into the neighbor table
overflow problem recently, at the hands of our ISP. I was in the
process of recompiling the kernel and mucking with arpd (I couldn't
get it to run/start properly) when the problem disappeared as quickly
as it showed up. Lucky for me, this was some kind of ISP problem, I
was able to determine that much through `tcpdump -i X -n arpd`.
My 'two cents' is that you try arpd, I did a bit of looking when I
came across that problem and it seemed to be the last ditch effort
when changing the gc threshold had no effect. Wasn't able to confirm
that it worked for sure though.
Cheers.
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

Grant Taylor

2007-10-23 21:04:43 UTC

That's just a few things to do on one box. How well is it handling it
if I can ask (aside from the problem that you are working on)?

Post by Marco C. Coelho
1. The last man working on this box had mistakenly edited the hosts
file and added the machine name and complete domain name to the local
host 127.0.0.1 name. It should only be pointed to the eth0 interface.
I have changed this.

Dough!

Post by Marco C. Coelho
2. The route output is making an announcement
64.0.0.0 argontech.net 255.0.0.0 UG 20 0
0 eth0
My public IP space is a /20 within that space, not the whole Class A. I
have not found which box is announcing this within my network yet.

I would think that you could extract that information from OSPF, or at
least the system that is advertising and work backwards until you find
the ultimate culprit.

Grant. . . .

Marco C. Coelho

2007-10-23 21:27:47 UTC

Is there a way to probe the kernel to find out how big the neighbor
table is on a regular basis? Without making a smoking hole of course.

BTW, Traffic shaping is also controlled via Radius.

That's just a few things to do on one box. How well is it handling it
if I can ask (aside from the problem that you are working on)?

It's actually pretty happy on a single processor, single core AMD 3000
with 1.5 G of RAM (it was not happy with 512K!!!). I've actually got a
new Dual Core, Dual Processor box loaded and ready to place in
production, but would like to fix this problem first. Unfortunately it
takes between a week and a month for the problem to surface.

model name : AMD Athlon(tm) 64 Processor 3000+
stepping : 0
cpu MHz : 2000.000
cache size : 512 KB

total used free shared buffers cached
Mem: 1554796 1044324 510472 0 221180 430860
-/+ buffers/cache: 392284 1162512
Swap: 4096496 148 4096348

ping times through this box:

64 bytes from f1.www.vip.mud.yahoo.com (209.191.93.52): icmp_seq=1
ttl=58 time=7.74 ms
64 bytes from f1.www.vip.mud.yahoo.com (209.191.93.52): icmp_seq=2
ttl=58 time=8.25 ms
64 bytes from f1.www.vip.mud.yahoo.com (209.191.93.52): icmp_seq=3
ttl=58 time=8.36 ms
64 bytes from f1.www.vip.mud.yahoo.com (209.191.93.52): icmp_seq=4
ttl=58 time=11.9 ms
64 bytes from f1.www.vip.mud.yahoo.com (209.191.93.52): icmp_seq=5
ttl=58 time=8.39 ms

Dough!

Post by Marco C. Coelho
2. The route output is making an announcement
64.0.0.0 argontech.net 255.0.0.0 UG 20
0 0 eth0
My public IP space is a /20 within that space, not the whole Class
A. I have not found which box is announcing this within my network yet.

I would think that you could extract that information from OSPF, or at
least the system that is advertising and work backwards until you find
the ultimate culprit.

I'm working on it, but time is slim today (but not me)!

Post by Grant Taylor
Grant. . . .
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

Grant Taylor

2007-10-23 21:23:41 UTC

Post by Marco C. Coelho
Is there a way to probe the kernel to find out how big the neighbor
table is on a regular basis? Without making a smoking hole of course.

Other than querying the ARP cache, I'm not aware of any thing. I'm sure
there is a way with in the kernel to see how many entries are in the ARP
cache, but I am the wrong person to ask.

Post by Marco C. Coelho
BTW, Traffic shaping is also controlled via Radius.

*nod*

Post by Marco C. Coelho
It's actually pretty happy on a single processor, single core AMD 3000
with 1.5 G of RAM (it was not happy with 512K!!!). I've actually got a
new Dual Core, Dual Processor box loaded and ready to place in
production, but would like to fix this problem first. Unfortunately it
takes between a week and a month for the problem to surface.

Good. It is nice to see Linux doing some things that Cisco and others
tried to dominate for so long.

Post by Marco C. Coelho
I'm working on it, but time is slim today (but not me)!

I wonder if you can turn up debugging on your OSPF daemon to see who /
what is being advertised.

Grant. . . .

Alexandru Dragoi

2007-10-24 10:19:49 UTC

Post by Marco C. Coelho
This box is doing a lot. It terminates 1000 PPPoE connections,
provides traffic shaping using TC/HTB, authenticates all users via
Radius. It also runs OSPF routing for the internal network. Looking
at a simple route output I see all the PPP connections coming through
the box, and due to the OSPF I also see the rest of my network
1. The last man working on this box had mistakenly edited the hosts
file and added the machine name and complete domain name to the local
host 127.0.0.1 name. It should only be pointed to the eth0
interface. I have changed this.
2. The route output is making an announcement
64.0.0.0 argontech.net 255.0.0.0 UG 20
0 0 eth0

This doesn't look dangerous for your problem, I was only talking about
directly connected networks:

# ip route |grep link

Post by Marco C. Coelho
My public IP space is a /20 within that space, not the whole Class A.
I have not found which box is announcing this within my network yet.

Post by Alexandru Dragoi
What about checking your routing table? you may have link routes
for massive subnets (like 85.0.0.0/8 or 140.20.0.0/16). Some
programs prefer to use "standard" netmask of classes A and B.

I'm betting that the OP has other things going on seeing has how
s/he mentioned PPPoE, which to my knowledge is a layer 2 protocol,
and thus not subject to typical routing scenarios. In essence the
OP could have thousands of PPPoE connections terminating on one
system with the ARP cache having to deal with where to send traffic
to which MAC address. There is not a lot of room for routing in such
a scenario.

I agree with Peter's suggestion, arpd. I ran into the neighbor table
overflow problem recently, at the hands of our ISP. I was in the
process of recompiling the kernel and mucking with arpd (I couldn't
get it to run/start properly) when the problem disappeared as quickly
as it showed up. Lucky for me, this was some kind of ISP problem, I
was able to determine that much through `tcpdump -i X -n arpd`.
My 'two cents' is that you try arpd, I did a bit of looking when I
came across that problem and it seemed to be the last ditch effort
when changing the gc threshold had no effect. Wasn't able to confirm
that it worked for sure though.
Cheers.
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

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

Marco C. Coelho

2007-10-24 15:19:19 UTC

the ip route with a grep for link returns:

snip** too long
64.202.227.198 dev ppp436 proto kernel scope link src 10.20.1.1
64.202.227.196 dev ppp421 proto kernel scope link src 10.20.1.1
64.202.227.197 dev ppp211 proto kernel scope link src 10.20.0.1
64.202.227.194 dev ppp13 proto kernel scope link src 10.20.1.1
64.202.227.192 dev ppp404 proto kernel scope link src 10.20.1.1
64.202.227.254 dev ppp194 proto kernel scope link src 10.20.1.1
64.202.227.253 dev ppp130 proto kernel scope link src 10.20.1.1
64.202.227.252 dev ppp243 proto kernel scope link src 10.20.1.1
64.202.227.249 dev ppp195 proto kernel scope link src 10.20.1.1
64.202.227.248 dev ppp254 proto kernel scope link src 10.20.1.1
64.202.227.247 dev ppp235 proto kernel scope link src 10.20.1.1
64.202.227.242 dev ppp78 proto kernel scope link src 10.20.1.1
64.202.227.240 dev ppp328 proto kernel scope link src 10.20.1.1
64.202.227.237 dev ppp44 proto kernel scope link src 10.20.1.1
64.202.227.236 dev ppp122 proto kernel scope link src 10.20.1.1
64.202.227.234 dev ppp316 proto kernel scope link src 10.20.1.1
64.202.227.232 dev ppp132 proto kernel scope link src 10.20.1.1
64.202.227.231 dev ppp104 proto kernel scope link src 10.20.0.1
64.202.227.226 dev ppp179 proto kernel scope link src 10.20.0.1
64.202.224.0/24 dev eth0 proto kernel scope link src 64.202.224.8
192.168.1.0/24 dev eth3 proto kernel scope link src 192.168.1.8
169.254.0.0/16 dev eth3 scope link

All the pppoe terminations (pppd) are shown, as well as the last three
subnets. I'll have to see where the 169.254.0.0/16 is coming from?

mc

This doesn't look dangerous for your problem, I was only talking about
# ip route |grep link

Post by Marco C. Coelho
My public IP space is a /20 within that space, not the whole Class A.
I have not found which box is announcing this within my network yet.

Post by Alexandru Dragoi
What about checking your routing table? you may have link routes
for massive subnets (like 85.0.0.0/8 or 140.20.0.0/16). Some
programs prefer to use "standard" netmask of classes A and B.

I'm betting that the OP has other things going on seeing has how
s/he mentioned PPPoE, which to my knowledge is a layer 2 protocol,
and thus not subject to typical routing scenarios. In essence the
OP could have thousands of PPPoE connections terminating on one
system with the ARP cache having to deal with where to send traffic
to which MAC address. There is not a lot of room for routing in such
a scenario.

I agree with Peter's suggestion, arpd. I ran into the neighbor table
overflow problem recently, at the hands of our ISP. I was in the
process of recompiling the kernel and mucking with arpd (I couldn't
get it to run/start properly) when the problem disappeared as quickly
as it showed up. Lucky for me, this was some kind of ISP problem, I
was able to determine that much through `tcpdump -i X -n arpd`.
My 'two cents' is that you try arpd, I did a bit of looking when I
came across that problem and it seemed to be the last ditch effort
when changing the gc threshold had no effect. Wasn't able to confirm
that it worked for sure though.
Cheers.
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

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

Alexandru Dragoi

2007-10-24 16:06:12 UTC

Post by Marco C. Coelho
snip** too long
64.202.227.198 dev ppp436 proto kernel scope link src 10.20.1.1
64.202.227.196 dev ppp421 proto kernel scope link src 10.20.1.1
64.202.227.197 dev ppp211 proto kernel scope link src 10.20.0.1
64.202.227.194 dev ppp13 proto kernel scope link src 10.20.1.1
64.202.227.192 dev ppp404 proto kernel scope link src 10.20.1.1
64.202.227.254 dev ppp194 proto kernel scope link src 10.20.1.1
64.202.227.253 dev ppp130 proto kernel scope link src 10.20.1.1
64.202.227.252 dev ppp243 proto kernel scope link src 10.20.1.1
64.202.227.249 dev ppp195 proto kernel scope link src 10.20.1.1
64.202.227.248 dev ppp254 proto kernel scope link src 10.20.1.1
64.202.227.247 dev ppp235 proto kernel scope link src 10.20.1.1
64.202.227.242 dev ppp78 proto kernel scope link src 10.20.1.1
64.202.227.240 dev ppp328 proto kernel scope link src 10.20.1.1
64.202.227.237 dev ppp44 proto kernel scope link src 10.20.1.1
64.202.227.236 dev ppp122 proto kernel scope link src 10.20.1.1
64.202.227.234 dev ppp316 proto kernel scope link src 10.20.1.1
64.202.227.232 dev ppp132 proto kernel scope link src 10.20.1.1
64.202.227.231 dev ppp104 proto kernel scope link src 10.20.0.1
64.202.227.226 dev ppp179 proto kernel scope link src 10.20.0.1
64.202.224.0/24 dev eth0 proto kernel scope link src 64.202.224.8
192.168.1.0/24 dev eth3 proto kernel scope link src 192.168.1.8
169.254.0.0/16 dev eth3 scope link

The one above must be deleted, many redhat-like distros attach
169.254.0.0/16.

Post by Marco C. Coelho
All the pppoe terminations (pppd) are shown, as well as the last three
subnets. I'll have to see where the 169.254.0.0/16 is coming from?
mc

This doesn't look dangerous for your problem, I was only talking about
# ip route |grep link

Post by Marco C. Coelho
My public IP space is a /20 within that space, not the whole Class A.
I have not found which box is announcing this within my network yet.

Post by Alexandru Dragoi
What about checking your routing table? you may have link routes
for massive subnets (like 85.0.0.0/8 or 140.20.0.0/16). Some
programs prefer to use "standard" netmask of classes A and B.

I'm betting that the OP has other things going on seeing has how
s/he mentioned PPPoE, which to my knowledge is a layer 2 protocol,
and thus not subject to typical routing scenarios. In essence the
OP could have thousands of PPPoE connections terminating on one
system with the ARP cache having to deal with where to send traffic
to which MAC address. There is not a lot of room for routing in such
a scenario.

I agree with Peter's suggestion, arpd. I ran into the neighbor table
overflow problem recently, at the hands of our ISP. I was in the
process of recompiling the kernel and mucking with arpd (I couldn't
get it to run/start properly) when the problem disappeared as quickly
as it showed up. Lucky for me, this was some kind of ISP problem, I
was able to determine that much through `tcpdump -i X -n arpd`.
My 'two cents' is that you try arpd, I did a bit of looking when I
came across that problem and it seemed to be the last ditch effort
when changing the gc threshold had no effect. Wasn't able to confirm
that it worked for sure though.
Cheers.
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

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

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

Marco C. Coelho

2007-10-25 15:08:16 UTC

Looking into it further an ip route shows:

64.0.0.0/8 via 64.202.224.1 dev eth0 proto zebra metric 20 equalize

So the 64.0.0.0 announce is coming into this box through OSPF (zebra)

The 169.254.0.0/16 is being automajically added through the sysconfig
network scripts. I'm looking into why.

In either case I still don't see why these entries would make the
neighbor table overflow. Could it have been the previous fix to the
hosts file?

mc

The one above must be deleted, many redhat-like distros attach
169.254.0.0/16.

Post by Marco C. Coelho
All the pppoe terminations (pppd) are shown, as well as the last three
subnets. I'll have to see where the 169.254.0.0/16 is coming from?
mc

This doesn't look dangerous for your problem, I was only talking about
# ip route |grep link

Post by Marco C. Coelho
My public IP space is a /20 within that space, not the whole Class A.
I have not found which box is announcing this within my network yet.

Post by Alexandru Dragoi
What about checking your routing table? you may have link routes
for massive subnets (like 85.0.0.0/8 or 140.20.0.0/16). Some
programs prefer to use "standard" netmask of classes A and B.

I'm betting that the OP has other things going on seeing has how
s/he mentioned PPPoE, which to my knowledge is a layer 2 protocol,
and thus not subject to typical routing scenarios. In essence the
OP could have thousands of PPPoE connections terminating on one
system with the ARP cache having to deal with where to send traffic
to which MAC address. There is not a lot of room for routing in such
a scenario.

I agree with Peter's suggestion, arpd. I ran into the neighbor table
overflow problem recently, at the hands of our ISP. I was in the
process of recompiling the kernel and mucking with arpd (I couldn't
get it to run/start properly) when the problem disappeared as quickly
as it showed up. Lucky for me, this was some kind of ISP problem, I
was able to determine that much through `tcpdump -i X -n arpd`.
My 'two cents' is that you try arpd, I did a bit of looking when I
came across that problem and it seemed to be the last ditch effort
when changing the gc threshold had no effect. Wasn't able to confirm
that it worked for sure though.
Cheers.
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

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

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

Alexandru Dragoi

2007-10-25 16:30:52 UTC

Post by Marco C. Coelho
64.0.0.0/8 via 64.202.224.1 dev eth0 proto zebra metric 20 equalize

This /8 doesn't affect the neighbor table. There must be a problem on
the site that announce it

Post by Marco C. Coelho
So the 64.0.0.0 announce is coming into this box through OSPF (zebra)
The 169.254.0.0/16 is being automajically added through the sysconfig
network scripts. I'm looking into why.
In either case I still don't see why these entries would make the
neighbor table overflow. Could it have been the previous fix to the
hosts file?

Well, when somebody try to make traffic with somebody from
169.254.0.0/16 throught you, your server will ask for arp on eth3, and
most probably will record an <incomplete> entry in arp table. Virii and
others can make this worse. Another quick fix is to drop arps from/to
169.254.0.0/16 with arptables.

Post by Marco C. Coelho
mc

The one above must be deleted, many redhat-like distros attach
169.254.0.0/16.

Post by Marco C. Coelho
All the pppoe terminations (pppd) are shown, as well as the last three
subnets. I'll have to see where the 169.254.0.0/16 is coming from?
mc

This doesn't look dangerous for your problem, I was only talking about
# ip route |grep link

Post by Marco C. Coelho
My public IP space is a /20 within that space, not the whole Class A.
I have not found which box is announcing this within my network yet.

Post by Alexandru Dragoi
What about checking your routing table? you may have link routes
for massive subnets (like 85.0.0.0/8 or 140.20.0.0/16). Some
programs prefer to use "standard" netmask of classes A and B.

I'm betting that the OP has other things going on seeing has how
s/he mentioned PPPoE, which to my knowledge is a layer 2 protocol,
and thus not subject to typical routing scenarios. In essence the
OP could have thousands of PPPoE connections terminating on one
system with the ARP cache having to deal with where to send traffic
to which MAC address. There is not a lot of room for routing in such
a scenario.

I agree with Peter's suggestion, arpd. I ran into the neighbor table
overflow problem recently, at the hands of our ISP. I was in the
process of recompiling the kernel and mucking with arpd (I couldn't
get it to run/start properly) when the problem disappeared as quickly
as it showed up. Lucky for me, this was some kind of ISP problem, I
was able to determine that much through `tcpdump -i X -n arpd`.
My 'two cents' is that you try arpd, I did a bit of looking when I
came across that problem and it seemed to be the last ditch effort
when changing the gc threshold had no effect. Wasn't able to confirm
that it worked for sure though.
Cheers.
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

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

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

Marco C. Coelho

2007-11-19 22:36:31 UTC

Still beating the same bush!

I've done all the possible suggestions so far. I still was getting a
neighbor table overflow.
Looking at the MAN 7 ARP pages, I see:

gc_thresh1
The minimum number of entries to keep in the ARP cache.
The garbage collector will not run if there are
fewer than this number of entries in the cache. Defaults
to 128.

gc_thresh2
The soft maximum number of entries to keep in the ARP
cache. The garbage collector will allow the num-
ber of entries to exceed this for 5 seconds before
collection will be performed. Defaults to 512.

gc_thresh3
The hard maximum number of entries to keep in the ARP
cache. The garbage collector will always run if
there are more than this number of entries in the cache.
Defaults to 1024.

Since this box never gets less than 500 pppoe connections, this Sat I
changed
WAS NOW
gc_thresh1 512 1024
gc_thresh2 2048 2048
gc_thresh3 4096 4096

what's strange is when I do an 'arp -an' I only get three entries back.
(ips changed to protect the guilty). Shouldn't this show the arp entries

? (x.202.x.3) at 00:03:47:2D:8B:F9 [ether] on eth0
? (x.202.x.1) at 00:03:E3:88:EC:C2 [ether] on eth0
? (x.202.x.2) at 00:18:8B:76:EC:D8 [ether] on eth0
? (x.202.x.9) at 00:90:27:43:C2:CF [ether] on eth0

ip route | grep link provides:

snip (lots of pppoe connects)
x.202.x.237 dev ppp53 proto kernel scope link src 10.20.1.1
x.202.x.235 dev ppp339 proto kernel scope link src 10.20.1.1
x.202.x.232 dev ppp185 proto kernel scope link src 10.20.1.1
x.202.x.231 dev ppp313 proto kernel scope link src 10.20.1.1
x.202.x.230 dev ppp67 proto kernel scope link src 10.20.1.1
x.202.x.226 dev ppp74 proto kernel scope link src 10.20.1.1
x.202.x.224 dev ppp150 proto kernel scope link src 10.20.1.1
x.202.x.0/24 dev eth0 proto kernel scope link src x.202.224.8
192.168.1.0/24 dev eth3 proto kernel scope link src 192.168.1.8

I don't think we are doing anything too special with this box that we
would see a kernel issue no one else is seeing. Can arp poisoning cause
this?

a dmesg after a clean reboot only gives:

Shorewall:all2all:REJECT:IN=ppp413 OUT= MAC= SRC=x.202.x.165
DST=10.20.1.1 LEN=60 TOS=0x00 PREC=0x00 TTL=254 ID=39752 PROTO=ICMP
TYPE=8 CODE=0 ID=25040 SEQ=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=72 TOS=0x00 PREC=0x00 TTL=126 ID=48363 PROTO=UDP
SPT=427 DPT=427 LEN=52
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48492 DF PROTO=TCP
SPT=36005 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48493 DF PROTO=TCP
SPT=36005 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48517 DF PROTO=TCP
SPT=36005 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48518 DF PROTO=TCP
SPT=33969 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=72 TOS=0x00 PREC=0x00 TTL=126 ID=48519 PROTO=UDP
SPT=427 DPT=427 LEN=52
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48522 DF PROTO=TCP
SPT=33969 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48526 DF PROTO=TCP
SPT=33969 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48614 DF PROTO=TCP
SPT=35790 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48630 DF PROTO=TCP
SPT=35790 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48x6 DF PROTO=TCP
SPT=35790 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48x8 DF PROTO=TCP
SPT=34718 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48663 DF PROTO=TCP
SPT=34718 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48679 DF PROTO=TCP
SPT=34718 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.y.x.110
DST=192.168.1.7 LEN=72 TOS=0x00 PREC=0x00 TTL=126 ID=48724 PROTO=UDP
SPT=427 DPT=427 LEN=52

Kernel Version 2.6.18-8.1.6

Looking for any suggestions.

Marco

x.0.0.0/8 via x.y.224.1 dev eth0 proto zebra metric 20 equalize
So the x.0.0.0 announce is coming into this box through OSPF (zebra)
The 169.254.0.0/16 is being automajically added through the sysconfig
network scripts. I'm looking into why.

Add "NOZEROCONF=yes" in /etc/sysconfig/network and the 169.254.0.0/16
network won't be created anymore.

In either case I still don't see why these entries would make the neighbor
table overflow. Could it have been the previous fix to the hosts file?
mc
snip** too long
x.y.x.198 dev ppp436 proto kernel scope link src 10.20.1.1
x.y.x.196 dev ppp421 proto kernel scope link src 10.20.1.1
x.y.x.197 dev ppp211 proto kernel scope link src 10.20.0.1
x.y.x.194 dev ppp13 proto kernel scope link src 10.20.1.1
x.y.x.192 dev ppp404 proto kernel scope link src 10.20.1.1
x.y.x.254 dev ppp194 proto kernel scope link src 10.20.1.1
x.y.x.253 dev ppp130 proto kernel scope link src 10.20.1.1
x.y.x.252 dev ppp243 proto kernel scope link src 10.20.1.1
x.y.x.249 dev ppp195 proto kernel scope link src 10.20.1.1
x.y.x.248 dev ppp254 proto kernel scope link src 10.20.1.1
x.y.x.247 dev ppp235 proto kernel scope link src 10.20.1.1
x.y.x.242 dev ppp78 proto kernel scope link src 10.20.1.1
x.y.x.240 dev ppp328 proto kernel scope link src 10.20.1.1
x.y.x.237 dev ppp44 proto kernel scope link src 10.20.1.1
x.y.x.236 dev ppp122 proto kernel scope link src 10.20.1.1
x.y.x.234 dev ppp316 proto kernel scope link src 10.20.1.1
x.y.x.232 dev ppp132 proto kernel scope link src 10.20.1.1
x.y.x.231 dev ppp104 proto kernel scope link src 10.20.0.1
x.y.x.226 dev ppp179 proto kernel scope link src 10.20.0.1
x.y.224.0/24 dev eth0 proto kernel scope link src x.y.224.8
192.168.1.0/24 dev eth3 proto kernel scope link src 192.168.1.8
169.254.0.0/16 dev eth3 scope link
The one above must be deleted, many redhat-like distros attach
169.254.0.0/16.
All the pppoe terminations (pppd) are shown, as well as the last three
subnets. I'll have to see where the 169.254.0.0/16 is coming from?
mc
This box is doing a lot. It terminates 1000 PPPoE connections,
provides traffic shaping using TC/HTB, authenticates all users via
Radius. It also runs OSPF routing for the internal network. Looking
at a simple route output I see all the PPP connections coming through
the box, and due to the OSPF I also see the rest of my network
1. The last man working on this box had mistakenly edited the hosts
file and added the machine name and complete domain name to the local
host 127.0.0.1 name. It should only be pointed to the eth0
interface. I have changed this.
2. The route output is making an announcement
x.0.0.0 argontech.net 255.0.0.0 UG 20
0 0 eth0
This doesn't look dangerous for your problem, I was only talking about
# ip route |grep link
My public IP space is a /20 within that space, not the whole Class A.
I have not found which box is announcing this within my network yet.
What about checking your routing table? you may have link routes
for massive subnets (like 85.0.0.0/8 or 140.20.0.0/16). Some
programs prefer to use "standard" netmask of classes A and B.
I'm betting that the OP has other things going on seeing has how
s/he mentioned PPPoE, which to my knowledge is a layer 2 protocol,
and thus not subject to typical routing scenarios. In essence the
OP could have thousands of PPPoE connections terminating on one
system with the ARP cache having to deal with where to send traffic
to which MAC address. There is not a lot of room for routing in such
a scenario.
I agree with Peter's suggestion, arpd. I ran into the neighbor table
overflow problem recently, at the hands of our ISP. I was in the
process of recompiling the kernel and mucking with arpd (I couldn't
get it to run/start properly) when the problem disappeared as quickly
as it showed up. Lucky for me, this was some kind of ISP problem, I
was able to determine that much through `tcpdump -i X -n arpd`.
My 'two cents' is that you try arpd, I did a bit of looking when I
came across that problem and it seemed to be the last ditch effort
when changing the gc threshold had no effect. Wasn't able to confirm
that it worked for sure though.
Cheers.
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc
------------------------------------------------------------------------
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc
_______________________________________________
LARTC mailing list
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

Marco C. Coelho

2007-12-07 17:17:09 UTC

Ok, I hope this helps someone else out there when they google neighbor
table overflow solution linux kernel:

This is just an update to state that since gc_thresh1 was increased to a
number greater than the number of simultaneous connected PPPoE clients
on this box, it has not given me the neighbor table problem.

So set gc_thresh1 greater than the number of local connections you get with:

ip route | grep link | wc -l

So in /etc/sysctl.conf add something like:

# Added to stop "neighbor table overflow" messages in the kernel
net.ipv4.neigh.default.gc_thresh1=1024
net.ipv4.neigh.default.gc_thresh2=2048
net.ipv4.neigh.default.gc_thresh3=4096
# Added to increase IP contrack number (was getting to max)
net.ipv4.ip_conntrack_max=99999

Have a Merry Christmas!

Marco Coelho
Argon Technologies Inc.
www.argontech.net

Post by Marco C. Coelho
Still beating the same bush!
I've done all the possible suggestions so far. I still was getting a
neighbor table overflow.
gc_thresh1
The minimum number of entries to keep in the ARP cache.
The garbage collector will not run if there are
fewer than this number of entries in the cache.
Defaults to 128.
gc_thresh2
The soft maximum number of entries to keep in the ARP
cache. The garbage collector will allow the num-
ber of entries to exceed this for 5 seconds before
collection will be performed. Defaults to 512.
gc_thresh3
The hard maximum number of entries to keep in the ARP
cache. The garbage collector will always run if
there are more than this number of entries in the
cache. Defaults to 1024.
Since this box never gets less than 500 pppoe connections, this Sat I
changed
WAS NOW
gc_thresh1 512 1024
gc_thresh2 2048 2048
gc_thresh3 4096 4096
what's strange is when I do an 'arp -an' I only get three entries
back. (ips changed to protect the guilty). Shouldn't this show the
arp entries
? (x.202.x.3) at 00:03:47:2D:8B:F9 [ether] on eth0
? (x.202.x.1) at 00:03:E3:88:EC:C2 [ether] on eth0
? (x.202.x.2) at 00:18:8B:76:EC:D8 [ether] on eth0
? (x.202.x.9) at 00:90:27:43:C2:CF [ether] on eth0
snip (lots of pppoe connects)
x.202.x.237 dev ppp53 proto kernel scope link src 10.20.1.1
x.202.x.235 dev ppp339 proto kernel scope link src 10.20.1.1
x.202.x.232 dev ppp185 proto kernel scope link src 10.20.1.1
x.202.x.231 dev ppp313 proto kernel scope link src 10.20.1.1
x.202.x.230 dev ppp67 proto kernel scope link src 10.20.1.1
x.202.x.226 dev ppp74 proto kernel scope link src 10.20.1.1
x.202.x.224 dev ppp150 proto kernel scope link src 10.20.1.1
x.202.x.0/24 dev eth0 proto kernel scope link src x.202.224.8
192.168.1.0/24 dev eth3 proto kernel scope link src 192.168.1.8
I don't think we are doing anything too special with this box that we
would see a kernel issue no one else is seeing. Can arp poisoning
cause this?
Shorewall:all2all:REJECT:IN=ppp413 OUT= MAC= SRC=x.202.x.165
DST=10.20.1.1 LEN=60 TOS=0x00 PREC=0x00 TTL=254 ID=39752 PROTO=ICMP
TYPE=8 CODE=0 ID=25040 SEQ=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=72 TOS=0x00 PREC=0x00 TTL=126 ID=48363 PROTO=UDP
SPT=427 DPT=427 LEN=52
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48492 DF
PROTO=TCP SPT=36005 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48493 DF
PROTO=TCP SPT=36005 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48517 DF
PROTO=TCP SPT=36005 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48518 DF
PROTO=TCP SPT=33969 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=72 TOS=0x00 PREC=0x00 TTL=126 ID=48519 PROTO=UDP
SPT=427 DPT=427 LEN=52
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48522 DF
PROTO=TCP SPT=33969 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48526 DF
PROTO=TCP SPT=33969 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48614 DF
PROTO=TCP SPT=35790 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48630 DF
PROTO=TCP SPT=35790 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48x6 DF PROTO=TCP
SPT=35790 DPT=9220 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48x8 DF PROTO=TCP
SPT=34718 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48663 DF
PROTO=TCP SPT=34718 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.202.x.110
DST=192.168.1.7 LEN=48 TOS=0x00 PREC=0x00 TTL=126 ID=48679 DF
PROTO=TCP SPT=34718 DPT=16398 WINDOW=16384 RES=0x00 SYN URGP=0
Shorewall:all2all:REJECT:IN=ppp160 OUT=eth3 SRC=x.y.x.110
DST=192.168.1.7 LEN=72 TOS=0x00 PREC=0x00 TTL=126 ID=48724 PROTO=UDP
SPT=427 DPT=427 LEN=52
Kernel Version 2.6.18-8.1.6
Looking for any suggestions.
Marco

Add "NOZEROCONF=yes" in /etc/sysconfig/network and the 169.254.0.0/16
network won't be created anymore.

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

darko

2007-11-19 23:15:04 UTC