netdev
[Top] [All Lists]

Re: [Fwd: Re: possible bug x86 2.4.2 SMP in IP receive stack]

To: feldy@xxxxxxxx (Bob Felderman)
Subject: Re: [Fwd: Re: possible bug x86 2.4.2 SMP in IP receive stack]
From: kuznet@xxxxxxxxxxxxx
Date: Thu, 8 Mar 2001 22:49:01 +0300 (MSK)
Cc: feldy@xxxxxxxx, ak@xxxxxx, andrewm@xxxxxxxxxx, davem@xxxxxxxxxx, netdev@xxxxxxxxxxx, pp@xxxxxxxxxxxxxx
In-reply-to: <Pine.SUN.4.21.0103081133350.5124-100000@orion.myri.com> from "Bob Felderman" at Mar 8, 1 11:38:08 am
Sender: owner-netdev@xxxxxxxxxxx
Hello!

> [root@rcc intel_linux]# cat /proc/net/softnet_stat
> 00fe20ec 0000ae4e 00000ece 000000b6 00000000 00000000 00000000 00000000
> 0012e228
> 00eff495 0000b941 000010d0 000000cc 00000000 00000000 00000000 00000000
> 0006f8bd

Yes, statistics is sad. Each ~300th frame is lost, serious latency
problems.

Is this without zerocopy? I still hope, it looks better with it.

> I'd like them to go away, but UDP losees due to socket overflows
> are quite common on most operating systems.

Well, provided we do not attach gigabit interface to 20MHz i386
and do not experience floods of tiny frames, all the rest is problem
of user. If he selects enough high rcvbuf and do not sleep instead
of working, he may expect that kernel will not not lose.

Alexey

<Prev in Thread] Current Thread [Next in Thread>