Jukka Marin <jmarin@pyy.jmp.fi> writes:

> > I have a large number of connections which drop off the face of the planet
> > (www servers suck that way) and I noticed slowly increasing memory usage
> > in the networking code until I applied a variant of pr 2335...
> 
> Please tell us more.

The only changes I made was to put it into a -current kernel.  I made
no functional changes, really.

Here is the text from the pr.  I noticed the problem in Description, so
I applied the patch and it went away until I went and increased the
number of connections my machine gets.  :)

>Description:
Not all the 4.4BSD-Lite2 TCP patches, in particular the persist timeout 
patches seem to have made it into NetBSD-Current.  This means TCP sockets
can get into CLOSE_WAIT or FIN_WAIT_1 with data queued on them and neither
persist nor keepalive timers running, tieing up mbus forever and eventually
filling up the mbuf map

>How-To-Repeat:
Run a machine with many TCP connections from dial-up clients which
may disappear from the net before closing down the TCP sockets (eg a web 
server or web cache).  Note that netstat -m shows network memory in use
monotonically increaing until mb_map fills up after some period.

>Fix:
This patch makes sure that TCP send window reaches zero that the persist
timer is running, and that if enough persist probes fail the socket will be
dropped (Stevens, "TCP/IP Illustrated, Volume 3", pp196--200).