| Author |
Message |
|
|
Post subject: Weird Connectivity issues with 2.6.17 Kernel versions
 Posted: Jul 06, 2006 - 07:47 PM
|
|
Joined: Jun 21, 2006
Posts: 4
|
|
I already posted it in the German network folder, but it might be worth to put this here too.
In Kernel version 2.6.17 is a redesign for allocating rmem and wmem parameters of the IP stack. This result in probable connection issues.
Depending on other boxes on the way like Firewalls, Routers or even the Server connection might fail or data retrieval will not succeed.
A possible workaround is to set the window-scaling option to "disable" via the /etc/sysctl.conf
Just enter
net.ipv4.tcp_window_scaling=0
Another way is to adjust the "max" Value for both the
net.ipv4.tcp_rmem and
net.ipv4.tcp_wmem
The original discussion was found here
http://kerneltrap.org/node/6723
Keep on the good work
Regards,
Jochen |
|
|
| |
|
|
|
 |
|
|
|
Post subject: RE: Weird Connectivity issues with 2.6.17 Kernel versions
Posted: Jul 06, 2006 - 07:54 PM
|
|
Joined: Oct 20, 2005
Posts: 278
|
|
It may be because I use the 2.6.17.3-slh-up-1 version, but everything seems fine with me. www.everymac.com looks fine to me. (I could be missing something)
-Crust |
|
|
| |
|
|
|
|
|
|
|
Post subject: RE: Weird Connectivity issues with 2.6.17 Kernel versions
Posted: Jul 06, 2006 - 08:00 PM
|
|
Team Member
Joined: May 03, 2005
Posts: 1544
Location: out there somewhere
|
|
|
|
|
|
|
|
|
Post subject: RE: Weird Connectivity issues with 2.6.17 Kernel versions
Posted: Jul 06, 2006 - 08:15 PM
|
|
Joined: Aug 16, 2004
Posts: 1905
|
|
| It will only fail if one of the routers higher up in the chain is broken and doesn't support scaling properly, this is something to fix - not in the linux kernel which conforms to accepted standards, but in affected systems. |
|
|
| |
|
|
|
|
|
|
|
Post subject: Re: RE: Weird Connectivity issues with 2.6.17 Kernel version
Posted: Jul 06, 2006 - 08:19 PM
|
|
Joined: Jun 21, 2006
Posts: 4
|
|
|
Crust wrote:
It may be because I use the 2.6.17.3-slh-up-1 version, but everything seems fine with me. www.everymac.com looks fine to me. (I could be missing something)
-Crust
I guess that they switched the server to window-scaling disabled.
We made the experience by accessing a 2.6.17er box via a vpn. It worked perfect with 2.6.14 and 2.6.16. Switching to 2.6.17 we were unable to get the ssh connection.
Analyzing the tcpdump, we found that the window-size was lowered extremely when the Init-Key-Exchange took place. We did not receive the Key-Exchange from the ssh-server.
Playing around we saw some differences in the tcp-parms.
The rest was googling until I found the mentioned article.
We have a similar network structure than the one that failed and did not had any issues.
So, it is a weird issue that gives no clear indication why something doesn't work. For this reason I thought I should post it here to give a hint if something fails.
Jochen |
|
|
| |
|
|
|
|
|
|
|
Post subject: Re: RE: Weird Connectivity issues with 2.6.17 Kernel version
Posted: Jul 06, 2006 - 08:35 PM
|
|
Team Member
Joined: May 03, 2005
Posts: 1544
Location: out there somewhere
|
|
|
slh wrote:
It will only fail if one of the routers higher up in the chain is broken and doesn't support scaling properly, this is something to fix - not in the linux kernel which conforms to accepted standards, but in affected systems.
So is my router broke ?  |
_________________
h2's d-u script
h2's rdiff-backup script
|
| |
|
|
|
|
|
|
|
Post subject: Re: RE: Weird Connectivity issues with 2.6.17 Kernel version
Posted: Jul 06, 2006 - 11:17 PM
|
|
Joined: Aug 16, 2004
Posts: 1905
|
|
|
piper wrote:
So is my router broke ? ;)
Maybe your router, maybe on of your ISP's routers, maybe a router just in front of the webserver you're about to visit, maybe some router you're coincidently passing - time to investigate. |
|
|
| |
|
|
|
|
|
|
|
Post subject: RE: Re: RE: Weird Connectivity issues with 2.6.17 Kernel ver
Posted: Jul 07, 2006 - 12:21 AM
|
|
Joined: Jan 09, 2006
Posts: 1720
|
|
No piper not by default. Your router NIC's has a fixed RX/TX-Buffer size between 256 and 32K for each Dir and Port.
What they discuss is the buffer in the PC corresponding to NEW Cpu-handled NIC's. This means the cpu does the convert parallel to serial and the linking to the PHY for Voltage converting. Older NIC's may not have this kind of problem. Even so older Chipsets which controlls the memory. |
|
|
| |
|
|
|
|
|
|
|
Post subject: RE: Re: RE: Weird Connectivity issues with 2.6.17 Kernel ver
Posted: Jul 07, 2006 - 01:57 AM
|
|
Team Member
Joined: May 03, 2005
Posts: 1544
Location: out there somewhere
|
|
|
|
|
|
|
|
|
Post subject:
Posted: Jul 07, 2006 - 07:39 AM
|
|
Joined: Oct 20, 2004
Posts: 64
Location: Kaiserslautern
|
|
same to me, unreachable
then with
Code:
sysctl -w net.ipv4.tcp_window_scaling=0
its working.
so i will put
Code:
net.ipv4.tcp_window_scaling=0
in the sysctl.conf
Alex |
|
|
| |
|
|
|
|
|
|
|
Post subject:
Posted: Jul 07, 2006 - 08:51 AM
|
|
Joined: Dec 05, 2005
Posts: 414
Location: Auckland, New Zealand
|
|
I can't get to http://www.everymac.com/ from either:
2.6.17-kanotix-1
or:
2.6.17.3-slh-up-1
Putting:
net.ipv4.tcp_window_scaling=0
in:
/etc/sysctl.conf
works after a reboot
... or this works without having to reboot (as I discovered on my other pc):
/sbin/sysctl -p /etc/sysctl.conf |
_________________
Linux is evolution, not intelligent design - Linus Torvalds
|
| |
|
|
|
|
|
|
|
Post subject:
Posted: Jul 07, 2006 - 05:28 PM
|
|
Team Member
Joined: May 03, 2005
Posts: 1544
Location: out there somewhere
|
|
|
|
|
|
|
|
|
Post subject:
Posted: Jul 07, 2006 - 07:46 PM
|
|
Joined: Jun 21, 2006
Posts: 4
|
|
Just as a quick reminder:
Setting window-scaling to 0 means that there might be some performance degradations. Especially with High-Speed Links this can delay the response.
I am currently looking for a setting of
net.ipv4.tcp_rmem
and
net.ipv4.tcp_wmem
that will work without any trouble.
There is a discussion that the old values (until Kernel 2.6.16) might result in some issues - at least they worked...
Also, the amount of Memory is related to this issue, because both vaules for rmem and wmem are calculated from there.
So setting window_scaling=0 is just a "quick-and-dirty" Workaround
It really seems that there are specific boxes involved in this issue (in the middle).
Cheers,
Jochen |
|
|
| |
|
|
|
|
|
|
|
Post subject:
Posted: Jul 07, 2006 - 11:49 PM
|
|
Joined: Jan 09, 2006
Posts: 1720
|
|
@jgrotepass, this missleading widow_scaling can have more than one reason.
R1: memory mapping between NIC's-Ram and DMA. let it lead to bufferoverfflow when Buffersize is to large for dma.
R2: Paket to large to fit in Buffer either in NIC's-Ram or w(w)mem. Let it lead to a NAK-cycle and a resend-paket.
R3: window_size and position has changed within the addresss-space where the buffers should found
R4:New Soft-NIC is used and w/rmen window is within the cache. Cachemiss while the next paket should written or read by the CPU.
R5: more exceptions can occure at the same time
Best way to get rid use window_scaling=0 |
|
|
| |
|
|
|
|
|
|
