Re: 2.4.23 crash on Intel SDS2

From: Marcelo Tosatti
Date: Mon Nov 17 2003 - 06:48:51 EST




On Sun, 16 Nov 2003, Shane Wegner wrote:

> > > It's a database machine running MySQL and Postgres. The
> > > MySQL server runs about 4 queries/sec and PostGres only as
> > > needed. It also does some minor mail service, say 2
> > > messages per minute and runs apache at about 10 requests
> > > per minute.
> > >
> > > > > There are no significant driver changes in -pre4 that could affect you.
> > > > >
> > > > > Can you please try with mem=900M? I suspect something in the VM changes
> > > > > might be causing this.
> > >
> > > Just tried with mem=900m and subsequently mem=850m so as no
> > > himem pages were available with no effect. Machine still
> > > crashed.
>
> Hi,
>
> Well, I tried backing out the vm changes from pre4, no
> luck so started disabling things. So far, it seems my
> firewall script is at fault. I looked through the
> pre3-pre4 diff and the only change to the nat code is a
> one-liner.
>
> # The following is the BitKeeper ChangeSet Log
> # --------------------------------------------
> # 03/09/04 laforge@xxxxxxxxxxxxx 1.1063.41.4
> # [NETFILTER]: NAT range calculation fix.
> #
> # This patch fixes a logic bug in NAT range calculations, which also
> # causes a large slowdown when ICMP floods go through NAT.
> #
> # Author: Karlis Piesenieks
> # --------------------------------------------
> diff -Nru a/net/ipv4/netfilter/ip_nat_core.c b/net/ipv4/netfilter/ip_nat_core.c
> --- a/net/ipv4/netfilter/ip_nat_core.c Sun Nov 16 13:41:25 2003
> +++ b/net/ipv4/netfilter/ip_nat_core.c Sun Nov 16 13:41:25 2003
> @@ -157,8 +157,8 @@
> continue;
> }
>
> - if ((mr->range[i].flags & IP_NAT_RANGE_PROTO_SPECIFIED)
> - && proto->in_range(&newtuple, IP_NAT_MANIP_SRC,
> + if (!(mr->range[i].flags & IP_NAT_RANGE_PROTO_SPECIFIED)
> + || proto->in_range(&newtuple, IP_NAT_MANIP_SRC,
> &mr->range[i].min, &mr->range[i].max))
> return 1;
> }
>
> Reversing that change has thus far fixed things over here
> but time will tell. Is there any possible way that that
> particular change is somehow not smp safe?

That change is broken, its known to break other setups.

It has been reverted in the BK tree.


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/