Re: Small TCP-buglet in 2.1.80..

Andi Kleen (ak@muc.de)
21 Jan 1998 06:13:30 +0100


Linus Torvalds <torvalds@transmeta.com> writes:

> This should fix a TCP problem in 2.1.80 that makes at least one
> version of netscape able to get a oops message out of it otherwise.
>
> freitag, you may want to fix this up?

I already sent a fix for this to Alan yesterday, but it seems it was too late
or got lost.

Another question: 4.4.BSD always signal writable in select when an error occurs
on a socket. Should Linux do this too?

Here is the fix again:

--- net/ipv4/tcp.c-o Tue Jan 20 00:28:04 1998
+++ net/ipv4/tcp.c Tue Jan 20 00:29:31 1998
@@ -614,7 +614,7 @@
mask |= POLLIN | POLLRDNORM;

#if 1 /* This needs benchmarking and real world tests */
- space = sk->dst_cache->pmtu + 128;
+ space = (sk->dst_cache ? sk->dst_cache->pmtu : sk->mss) + 128;
if (space < 2048) /* XXX */
space = 2048;
#else /* 2.0 way */

-Andi

P.S.: I would appreciate any benchmark data that test these two options,
especially on bigger squids. If anyone is willing to do this please contact
me.