DragonFly bugs List (threaded) for 2004-08
Re: Internet problem after recent rewrite of mbuf
On 10:15, Mon 09 Aug 04, Matthew Dillon wrote:
> I don't think it could be the 29 July commit, but it could
> be one of the later ones.
I backed out changes to 3 August when the new network code
went in, and that wasn't the problem.
> If possible I would like you to try the latest kernel again, but
> this time see if the net.inet.tcp.avoid_pure_win_update sysctl
> makes a difference. It defaults to 1. See if setting it to 0
> solves your problem.
I tried new kernel from yesterday, and same thing happens.
Jeffrey also suggested setting
net.inet.tcp.avoid_out_win_update to 0, but that didn't help.
I haven't tried it today, but there weren't any changes to
the network code.
> If not I will give you a patch to back-out the other tcp related
> windowing change we made and we will see if that has an effect.
> There is a known issue with exponential window expansion potentially
> not working correctly with the current patches
> Another thing that would be very useful would be tcpdump output showing
> the session. e.g. in one window do this:
> tcpdump -n -i <network_ifc> -l port 80 > outputfile
OK, I did this 3 times:
1. Todays generic kernel(cvsuped few hours ago), trying to
load amazon.com, in which nothing happened: the website was
completely unloaded. Elinks sent packets, but amazon.com
Results are here:
2. Todays generic kernel, trying to load www.amazon.com (note
the 'www' part). The connection timed out after around 10
minutes, and only few KiB were transfered.
I didn't notice that there is a difference when loading
amazon.com and www.amazon.com until today, when nothing was
transferred when loading amazon.com. I can't remember did I
run amazon.com or www.amazon.com yesterday, but I'm pretty
sure it was www.amazon.com.
Results are here:
3. 29 July's generic kernel, in which every think works fine.
Results are here:
Why this problem doesn't happen to all websites and all network
> I tried 'simulating' the issue by using IPFW to create a dummynet
> pipe limited to 1 KByte/s on my test box, and then used elinks to
> connect to amazon.com, but it seemed to work fairly well. It loaded
> amazon.com in around ~20 seconds.
My combinations is unique, who has a working old 34KB serial
modem that they use? Not many, its just I couldn't get any
of the winmodems working, and some guy who I bought an old
box from cheaply had a few of these modems laying around....
I should be getting wireless broadband soon.
> :Hi Matt and others,
> :After recent kernel update(4 August from 20 July) I noticed
> :that some internet websites take really long time to load
> :(connection speed drops to around 50B/s(but is not active)
> :with all browsers that I tried, links, elinks, w3m, firefox).
> :This doesn't happen on all websites (there is a not so
> :noticeable slow down on google, but sites such as amazon.com
> :become totally unloadable. I was able to load 8.1KiB of
> :amazon.com in 9 minutes with elinks!). Connecting to my
> :ISP's mail and news servers doesn't seem to be affected.
> :With the help of Jeffrey Hsu, the source of the problem
> :seems to be a commit on 29 July to
> :sys/kern/uicp_mbuf.c and others (I'm not fully sure is it
> :that commit, or a commit on 31 July).
> :After backing out my src/sys to 29 July, and recompiling
> :the kernel the problem seems to disappear (amazon.com loads
> :in a few seconds with elinks).
> :I use an old serial modem (Pace Linnet 34 Plus) with a few
> :scripts from slackware (IIRC version 9) to connect to my ISP.
> :Sarunas Vancevicius