Invalid truesize messages since 2.6.25

Invalid truesize messages since 2.6.25

[Beschorner Daniel]

Since 2.6.25 we rarely see these messages:

SKB BUG: Invalid truesize (408) len=198, sizeof(sk_buff)=216

Yesterday I saw it the first time after a runtime of one week with
2.6.26.
The box is web server / IPSEC gateway.
Obviously nothing seems to be affected by the message but <= 2.6.24 no
messages appeared.

Daniel

Re: Invalid truesize messages since 2.6.25

[Ben Hutchings]

Beschorner Daniel wrote:
> Since 2.6.25 we rarely see these messages:
> 
> SKB BUG: Invalid truesize (408) len=198, sizeof(sk_buff)=216
> 
> Yesterday I saw it the first time after a runtime of one week with
> 2.6.26.
> The box is web server / IPSEC gateway.
> Obviously nothing seems to be affected by the message but <= 2.6.24 no
> messages appeared.

Invalid truesize can be caused by bugs in many different places.  So it
might be helpful to say which driver(s) and any unusual network protocol
or filter you are using.

Ben.

Re: Invalid truesize messages since 2.6.25

[Beschorner Daniel]

>> Since 2.6.25 we rarely see these messages:
>> 
>> SKB BUG: Invalid truesize (408) len=198, sizeof(sk_buff)=216
>> 
>> Yesterday I saw it the first time after a runtime of one week with
>> 2.6.26.
>> The box is web server / IPSEC gateway.
>> Obviously nothing seems to be affected by the message but <= 2.6.24
no
>> messages appeared.

> Invalid truesize can be caused by bugs in many different places.  So
it
> might be helpful to say which driver(s) and any unusual network
protocol
> or filter you are using.

Network driver: tg3
IPSEC: esp tunnels, ipcomp, 3des/aes
I use this config so far on other servers too without this message.

The unusal part special to this server may be:
Bandwidth management via htb & sfq, CONFIG_IP_MULTIPLE_TABLES for source
IP routing

Daniel

Re: Invalid truesize messages since 2.6.25

[Markus Werner]

2008/7/22 Beschorner Daniel <Daniel.Beschorner@facton.com>:
>>> Since 2.6.25 we rarely see these messages:
>>>
>>> SKB BUG: Invalid truesize (408) len=198, sizeof(sk_buff)=216

Same problem here.

>>> Yesterday I saw it the first time after a runtime of one week with
>>> 2.6.26.
>>> The box is web server / IPSEC gateway.
>>> Obviously nothing seems to be affected by the message but <= 2.6.24
> no
>>> messages appeared.
>
>> Invalid truesize can be caused by bugs in many different places.  So
> it
>> might be helpful to say which driver(s) and any unusual network
> protocol
>> or filter you are using.
>
> Network driver: tg3
> IPSEC: esp tunnels, ipcomp, 3des/aes
> I use this config so far on other servers too without this message.

Kernel: 2.6.26
Tested drivers: pcnet32, e1000, vmxnet

> The unusal part special to this server may be:
> Bandwidth management via htb & sfq, CONFIG_IP_MULTIPLE_TABLES for source
> IP routing

Both are used here too, but no IPSec.

This box is a router to the internet and an openvpn gateway. The
problem is not reliably reproducible, but it seems to happen when the
box is routing udp traffic, for example heavy openvpn usage or video
streaming via udp.


Markus

Re: Invalid truesize messages since 2.6.25

[Marco Berizzi]

Markus Werner wrote:

> 2008/7/22 Beschorner Daniel <Daniel.Beschorner@facton.com>:
> >>> Since 2.6.25 we rarely see these messages:
> >>>
> >>> SKB BUG: Invalid truesize (408) len=198, sizeof(sk_buff)=216
>
> Same problem here.

I haven't seen the problem after 7 days uptime with
2.6.26
With 2.6.25.x was very rare. My boxes are ipsec gateways
driven by openswan, squid and netfilter.
Network driver is 3c59x were the problem happened.

I was getting the warning both on system with and
without sfq/htb

On boxes without ipsec/netfilter (sendmail server)
I did never get this message.

> > The unusal part special to this server may be:
> > Bandwidth management via htb & sfq, CONFIG_IP_MULTIPLE_TABLES for
source
> > IP routing

Same for me. I do policy routing based on fwmark.
Just to be clear: I get this message with ipsec/
netfilter/policy routing and with/without sfq/htb

Invalid truesize messages since 2.6.25

[Beschorner Daniel]

I did some more investigations and can reproduce it now in a simple way:

I take my tg3 device into promiscuous mode and capture ESP packets:
Just "tcpdump -i eth1 esp" and ping 4 times through the IPSEC tunnel to
generate ESP traffic, after that dmesg shows

device eth1 entered promiscuous mode
SKB BUG: Invalid truesize (336) len=134, sizeof(sk_buff)=208
SKB BUG: Invalid truesize (336) len=134, sizeof(sk_buff)=208
SKB BUG: Invalid truesize (336) len=134, sizeof(sk_buff)=208
SKB BUG: Invalid truesize (336) len=134, sizeof(sk_buff)=208
device eth1 left promiscuous mode

Daniel

----------
Since 2.6.25 we rarely see these messages:

SKB BUG: Invalid truesize (408) len=198, sizeof(sk_buff)=216

Yesterday I saw it the first time after a runtime of one week with
2.6.26.
The box is web server / IPSEC gateway.
Obviously nothing seems to be affected by the message but <= 2.6.24 no
messages appeared.

Daniel

Re: Invalid truesize messages since 2.6.25

[Ben Hutchings]

On Fri, 2008-09-26 at 19:53 +0200, Beschorner Daniel wrote:
> I did some more investigations and can reproduce it now in a simple way:
> 
> I take my tg3 device into promiscuous mode and capture ESP packets:
> Just "tcpdump -i eth1 esp" and ping 4 times through the IPSEC tunnel to
> generate ESP traffic, after that dmesg shows
> 
> device eth1 entered promiscuous mode
> SKB BUG: Invalid truesize (336) len=134, sizeof(sk_buff)=208
> SKB BUG: Invalid truesize (336) len=134, sizeof(sk_buff)=208
> SKB BUG: Invalid truesize (336) len=134, sizeof(sk_buff)=208
> SKB BUG: Invalid truesize (336) len=134, sizeof(sk_buff)=208
> device eth1 left promiscuous mode
> 
> Daniel
> 
> ----------
> Since 2.6.25 we rarely see these messages:
> 
> SKB BUG: Invalid truesize (408) len=198, sizeof(sk_buff)=216
> 
> Yesterday I saw it the first time after a runtime of one week with
> 2.6.26.
> The box is web server / IPSEC gateway.
> Obviously nothing seems to be affected by the message but <= 2.6.24 no
> messages appeared.

Commit 3ab224be6d69de912ee21302745ea45a99274dbc, applied between 2.6.24
and 2.6.25, appears to ensure that truesize is checked for all skbs
where previously it was only done for some.  The bug probably already
existed in the ipsec code, which never updates truesize.

Ben.

-- 
Ben Hutchings, Senior Software Engineer, Solarflare Communications
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.