svn commit: r193941 - head/sys/netinet

[John Baldwin]

On Thursday 11 June 2009 12:04:32 am Bruce Evans wrote:
> On Wed, 10 Jun 2009, John Baldwin wrote:
> 
> > On Wednesday 10 June 2009 2:27:15 pm John Baldwin wrote:
> >> Author: jhb
> >> Date: Wed Jun 10 18:27:15 2009
> >> New Revision: 193941
> >> URL: http://svn.freebsd.org/changeset/base/193941
> >>
> >> Log:
> >>   Change a few members of tcpcb that store cached copies of ticks to be ints
> >>   instead of unsigned longs.  This fixes a few overflow edge cases on 64-bit
> >>   platforms.  Specifically, if an idle connection receives a packet shortly
> >>   before 2^31 clock ticks of uptime (about 25 days with hz=1000) and the keep
> >>   alive timer fires after 2^31 clock ticks, the keep alive timer will think
> >>   that the connection has been idle for a very long time and will immediately
> >>   drop the connection instead of sending a keep alive probe.
> >>
> >>   Reviewed by:	silby, gnn, lstewart
> >>   MFC after:	1 week
> >
> > Note that the MFC patch for 7 is very different.  I can't change the members
> > to int in 7 since tcpcb is part of the ABI (for netstat, etc.).  So, for 7 I
> > added explicit casts in the math operations using t_rcvtime and ticks.
> 
> That has a better chance of working too.

I tested both patches by cranking the keep alive timers to 5 seconds and
manually adjusting ticks to INT_MAX - 15000.  Then I letting a TCP connection
sit idle while ticks wrapped around to -ve.  Without either patch the
connection always died with ETIMEDOUT.  With each patch (including what was
committed above), the connection survived and the keepalive timer sent a
probe instead of dropping the connection.  Thus, both patches "work".

-- 
John Baldwin