Hi Dave,

On Sun, Sep 21, 2008 at 12:22:15AM -0700, David Miller wrote:

From: Tom Quetchenbach <redacted>
Date: Tue, 26 Aug 2008 11:11:52 -0700

quoted

This patch (applies to latest net-2.6) reduces the advertised value to
that requested by the user as long as setsockopt() is called before
connect() or accept(). This seems like the behavior that one would expect
as well as that which is documented.

 ...

quoted

Signed-off-by: Tom Quetchenbach <redacted>

I've applied this patch to net-next-2.6, thanks Tom.

Yesterday I worked on a similar patch to address the same issue which
I was experiencing on 2.6.27.25. Before posting my tiny patch, I
preferred to check the list and found this one which addresses the
same issue and which is already merged in mainline !

Would you accept to queue it for -stable please ? The commit ID was
f5fff5dc8a7a3f395b0525c02ba92c95d42b7390. I did not figure how to find
when it was merged, because its version pre-dates 2.6.27 (it matches
your net-2.6 queue I think). git-describe --contains tells me it was
merged in 2.6.28-rc1.

BTW, I found the same behaviour in 2.4 too, and from my quick analysis,
the mss is correctly computed till the moment the SYN-ACK is built,
where the one from the route cache is used instead. I'll check deeper
when I have time, but most likely a similar fix is needed there too.

Thanks!
Willy