samba - Oct 2004 - which kernel should be used to solve sendfile problem on linux?

I have seen some folks on the list with what seems to be a similar issue
as I have.  I work in a large FX shop and this has been a problem for us
since the 3.0.7 upgrade.  If anyone has a known to work config,  could you
share your set up info please?

Jerry what kernel, and other system bits versions do you test on?

The system has a 2.4.21 kernel with xfs filesystems. It is a redhat 9
machine on IBM hardware with a large FC attached array connected via a
qla2300.

The problem we are having from a client perspective is: when loading large
files over the network, part way into the xfer it will stop and usually
lock up the application.  Sometimes an error will display similar to

"Cannot copy sz20q007s004_lf: the specified network name is no longer
available."

they also complain about general slowness,  the same file which loads in 5
seconds on the station beside them takes 10 minutes,  yet xfer tests with
other protocals run at line speed.  This is an intermitting issue too, 
sometimes a week will pass without issues,  then it happens 20 times in
one week to the client.

On the server I will see this in the log file:

[2004/10/30 17:37:33, 1] smbd/service.c:make_connection_snum(648)
  wk26 (10.0.1.26) connect to service nova1 initially as user dmay
(uid=618, gid=100) (pid 29575)

[2004/10/30 17:38:39, 0] smbd/reply.c:send_file_readX(2134)
  send_file_readX: sendfile failed for file
shoebox/sz20/playblast/layout/sz20q007s004_lf.avi (Connection reset by
peer). Terminating

[2004/10/30 17:38:39, 1] smbd/service.c:close_cnum(837)
  wk26 (10.0.1.26) closed connection to service nova1

i've no problem upgrading system bits,  but i'd like to make it a one
time
thing as its in a major production role and requires a planed down time.

thanks
-g

> since the 3.0.7 upgrade.  If anyone has a known to work config,  could you
> share your set up info please?
> The system has a 2.4.21 kernel with xfs filesystems. It is a redhat 9
> machine on IBM hardware with a large FC attached array connected via a
> qla2300.
Did you test that the sendfile implementation is really the cause for 
the failure. That means did you try
use sendfile = no
in global?