Bug 105853

Summary:	window_scaling broken when syn packet sent twice
Product:	[openSUSE] SUSE LINUX 10.0	Reporter:	Bjoern Jacke <bjacke>
Component:	Kernel	Assignee:	Olaf Kirch <okir>
Status:	RESOLVED INVALID	QA Contact:	E-mail List <qa-bugs>
Severity:	Normal
Priority:	P5 - None
Version:	Beta 1
Target Milestone:	---
Hardware:	Other
OS:	All
Whiteboard:
Found By:	Other	Services Priority:
Business Priority:		Blocker:	---
Marketing QA Status:	---	IT Deployment:	---
Attachments:	dump of 3 tries of conntection setup with window_scaling = 4 set

Description Bjoern Jacke 2005-08-19 14:06:38 UTC

for example

echo 4 > /proc/sys/net/ipv4/tcp_window_scaling
(the bug happens with all window_scaling > 0)

then setting up a connection will first send a SYN package, which contains
window_scaling factor 2 (shoud this not be 4, might be another bug).
In the attached sniff the window size is 5840 byte multiplied by 2^2.

The other host is unreachabel, that's why we send another SYN package, this
time, our window size is 23360 bytes still multiplied by 2^2 (!).

This way all requested window sizes are double factored by the window_scaling
value when the first SYN package gets lost.

Comment 1 Bjoern Jacke 2005-08-19 14:08:32 UTC

Created attachment 46698 [details]
dump of 3 tries of conntection setup with window_scaling = 4 set

Comment 2 Olaf Kirch 2005-08-19 14:31:48 UTC

The tcp_window_scaling sysctl is a yes/no affair, so it doesn't matter 
if you write 1, 4 or 4000 :) 
 
I also don't understand the other part of the report. The tcpdump shows 
three identical packets: 
 
win 5840 <mss 1460,sackOK,timestamp 33005845 0,nop,wscale 2> 
 
Which looks right to me. The initial congestion window defaults to 4, 
an with a MSS of 1460, that's a window of 5840. And as per the RFC 1323, 
the window in a SYN packet is never scaled (because we don't know yet 
whether the other end understands window scaling at all). 
 
Can you explain exactly what is wrong here?

Comment 3 Andreas Kleen 2005-08-19 14:39:50 UTC

The window scaling is determined by the socket buffer, not by the sysctl.

The SYNs also look correct to me - they are all identical.

Suggest INVALID.

Comment 4 Bjoern Jacke 2005-08-19 14:43:55 UTC

urg, I looked with ethereal at it and ethereal showes the first package window
size unscaled and the second and later SYN packages with window size scaled.
You'r right the packages are identical, ethereal just fooled me, sorry.

Comment 5 Olaf Kirch 2005-08-19 14:47:12 UTC

Aw, thanks. Lucky I'm still a tcpdump guy :) 
 
You may want to send a notice to the ethereal folks though telling them 
about the problem in their code.