Poor performance, samba bug, DOS client
Frank Giessler
This was a long way for me but I will tell it in detail so others might benefit. I have very
little experience in C-programming and the internals of Solaris. Maybe someone with more
experience can check my reasoning and my patch.
I had set up a samba server (2.0.4) on a SUN running Solaris 2.6. It used to be running
great with just one client (DOS PC running MS-DOS client 3.0).
After we had some problems with the SUN's HD I changed it and upgraded to Solaris 8 and
Samba 2.2.1a. After some some time I noticed that the writing speed from the client to the
server had gone down to a crawl while reading seemed to be quite normal. I did some
searching and tried several suggestions from others with similar problems (such as 'write
raw', 'read raw', 'socket options', etc.). Nothing helped. I got so frustrated that I
plugged the old HD back in and checked again. Wonderful performance! I upgraded to 2.2.1a on
the old HD to see if the Solaris version matters. Obviously, it did. Performance was still
high.
I put the new HD in and used tcpdump to check what was going on. Although I have no
experience with such dumps I compared two with good and bad connection, respectively. They
didn't look much different. In fact, it seemed that the number of packets were the same, but
on Solaris 8 they came with a very slow speed.
Then I tried to connect the PC client to some other servers. WinNT 4.0 gave good
performance, but Win2k was as slow as Samba. I checked at support.microsoft.com and got the
following:
http://support.microsoft.com/support/kb/articles/Q244/8/26.ASP
I checked the receive window size (netstat |grep .139) under Solaris 2.6 and 8, and indeed
SUN changed it from about 8k to about 24k. Setting
'socket options = SO_RCVBUF=8192'
in smb.conf didn't change anything, whereas
'ndd -set /dev/tcp tcp_recv_hiwat 8192'
worked!
The man page of setsockopt() on Solaris 8 finally revealed that the function chould be
called before any listen() or connect() call. After applying the following patch everything
worked fine:
diff -wc smbd/server.c smbd/server.c.new ***
smbd/server.c Fri Jul 6 04:02:03 2001 ---
smbd/server.c.new Thu Aug 9 11:06:12 2001
***************
*** 159,164 ****
--- 159,166 ----
if(s == -1)
return False;
/* ready to listen */ +
set_socket_options(s,"SO_KEEPALIVE");
+ set_socket_options(s,user_socket_options);
if (listen(s, 5) == -1) {
DEBUG(0,("listen: %s\n",strerror(errno)));
close(s);
***************
*** 178,183 ****
--- 180,187 ----
return(False);
/* ready to listen */ +
set_socket_options(s,"SO_KEEPALIVE");
+ set_socket_options(s,user_socket_options);
if (listen(s, 5) == -1) {
DEBUG(0,("open_sockets: listen: %s\n",
strerror(errno)));
***************
*** 264,272 ****
close_low_fds();
am_parent = 0;
-
set_socket_options(smbd_server_fd(),"SO_KEEPALIVE");
- set_socket_options(smbd_server_fd(),user_socket_options); -
/* Reset global variables in util.c so
that client substitutions will be
done correctly in the process. */
--- 268,273 ----
The problems I see are:
1) I don't know what happens when smbd is called from inetd.
2) I don't know in what sequence setsockopt() and listen() should be called in OSs other
than Solaris 8.
So, please, can someone have a look at it?
Thanks
Frank.
__
Frank Giessler
Klinikum der Universitaet Jena Tel.: +49-3641-9 35259
Biomagnetisches Zentrum Fax : +49-3641-9 35355
Andrew Bartlett
>
> Hi all,
>
> This was a long way for me but I will tell it in detail so others might benefit. I have very
> little experience in C-programming and the internals of Solaris. Maybe someone with more
> experience can check my reasoning and my patch.
Tridge thought it looked sane, and I've applied it.
--snip--
> The man page of setsockopt() on Solaris 8 finally revealed that the function chould be
> called before any listen() or connect() call. After applying the following patch everything
> worked fine:
>
--snip--
> The problems I see are:
>
> 1) I don't know what happens when smbd is called from inetd.
> 2) I don't know in what sequence setsockopt() and listen() should be called in OSs other
> than Solaris 8.
>
> So, please, can someone have a look at it?
Applied, thanks!
As for inetd, because I didn't drop the previous setting of socket
options it should 'just work', except on solaris as indicated. (calling
setsockopt() twice shouldn't do any harm I'm told).
Could you please check out the HEAD branch and confirm I fixed it
correctly?
Thanks,
Andrew Bartlett
--
Andrew Bartlett
abar...@pcug.org.au
abar...@samba.org