With a few days leave available, it seemed a good time to upgrade the FreeBSD server, crimson, from 4.11 to 5.3. I'd held off migrating to the newer version due to my experiences with the earliest release of 5.0. By coincidence, I also had just acquired a new laptop, an IBM Thinkpad T42, so I also installed 5.3 on this device.
The web/mail server is hosted on a very old (and slow) hardware platform. Its memory is now 64MB, with a hard drive of 8GB. I could just upgrade to 5.3 using cvsup, but I decided to buy a new, bigger (80GB), disk for the 5.3 install. This was partly so I could use the new UFS2 filesystem that comes with 5.3 and partly because it offered less risk if something went horribly wrong in the 5.3 install.
I downloaded a snapshot i386 iso from the FreeBSD snapshot location and burnt it to a CD. Since 5.4 was so close to arriving I took the chance of starting from 5.3-STABLE. To perform the actual install, I temporarily put the new hard disk in silver (normally a Windows 2000 machine). Its 2GHz processor should also help the install run a little faster.
The 5.3-STABLE version came up fine. I named it crimson5 as it would have to coexist with the 4.11 version of crimson for a while. I created the following partition scheme (in one slice):
Filesystem 1K-blocks Used Avail Capacity Mounted on /dev/ad0s1a 126702 52384 64182 45% / devfs 1 1 0 100% /dev /dev/ad0s1g 30462636 219134 27806492 1% /home /dev/ad0s1h 40401964 4 37169804 0% /opt /dev/ad0s1d 253678 12 233372 0% /tmp /dev/ad0s1f 4058062 1444964 2288454 39% /usr /dev/ad0s1e 247726 4392 223516 2% /var
The snapshot mini-ISO contains one package - perl-5.8.6, so once the
main install is complete, everything else has to be installed over
the Internet. I used /stand/sysinstall to perform the
initial package install, but I had trouble providing a release name
(specified via the Options menu under Custom) that
corresponded to a package set on the FreeBSD servers; I figured I
could use 5-STABLE, but that didn't work. The only value that
persuaded /stand/sysinstall/ to continue was 5.3-RELEASE.
The original crimson had X, which I didn't really need, so the only packages I downloaded were:
apache cvsup-without-gui emacs gmake perl portaudit portupgrade procmail py24-bsddb python qpopper rsync ruby ruby18-bdb1 sendmail
Needless to say, these packages pulled in a lot of others as
dependencies. Emacs ended up downloading elements of X, which kind
of negated my simplification. However, I re-built emacs using
WITHOUT_X11 so I could safely delete:
Xaw3d jpeg tiff libungif png imake xorg-libraries freetype2 fontconfig pkgconfig (?)
I didn't really want perl, but it seems to be needed by apache (I have my doubts).
Now that cvsup and portupgrade were available, I
could bring the ports tree up-to-date and upgrade any out of date
applications. I also created a custom kernel for crimson, which was
basically GENERIC without SCSI, most network adapters and USB.
However, I didn't install the new kernel; I wanted to see if the
GENERIC kernel would boot up in the target hardware first.
Last thing before swapping the hard drive to the target hardware was
to rename crimson5 as crimson. This involves changing
/etc/rc.conf, /etc/hosts and
/etc/dhclient.conf (which sends the host name to the DHCP
server).
To ensure that my many visitors (ha!) had an uninterrupted experience, before shutting down crimson, I booted the backup server (chrome), and set the NAPT of the Alcatel Speedtouch 510 router to direct HTTP and SMTP traffic to it. Crimson was shutdown, the hard drive replaced and rebooted. Everything worked just fine, once I had told the Compaq firmware to save the change of disk size. Even the custom kernel worked, once I installed it and rebooted.
I did encounter a slight problem with sendmail. I had
incorrectly created the /etc/mail/local-host-names file
(which defines the domains to accept mail for) with the wrong name
(local_host_names). When I tested the mail server with an
incoming mail message from outside the domain (via my ISP mail
account), the message bounced madly between crimson and the ISP,
until:
554 5.4.6 Too many hops 26 (25 max): from <me@isp.com> via localhost, to <testing@hydrus.org.uk>
Once the file name was fixed, the problem vanished. Then testing the web server revealed another, more serious, problem. Crimson froze completely when attempting to deliver the home page; the hydrus logo half displayed and then all activity stopped. The console was unresponsive, as were any open ssh sessions. Strangely, the ssh sessions did not die (at least not until I power cycled the machine). My immediate suspicion fell upon the lance (PCNet/PCI) network adapter, as I had previously seen one message like:
lnc0: Start of packet found before end of previous in transmit ring -- Resetting lnc0: Missed packet -- no receive buffer
However, that message had been issued a long time before the apache
testing. Also, I'd seen Missed packet -- no receive buffer
messages in 4.x, which had not seemed to cause problems. Having
been lead up the garden path many times based on hunches, I decided
to test all the elements I could.
Firstly, could I reproduce the problem? Yep, running a big machine-to-machine copy using rsync and hitting the apache server would make crimson freeze in a very short time. Rsync on it's own seemed to be fine. OK, was the problem the custom kernel? I rebooted back into GENERIC and tested again; another freeze. Was this some 5.3 related problem, independent of crimson? I put the hard disk back into silver; couldn't make it freeze, with or without loading ACPI (crimson didn't support ACPI).
OK, now to try a new network adapter. I installed the spare Netgear
FA331 adapter I had (which was nice). In FreeBSD,
this uses the sis driver. I re-installed the hard drive in
crimson, re-booted and tested. No freeze. Hmm, for once the
initial hunch was correct. Since this test used the GENERIC kernel,
I rebuilt and installed a custom CRIMSON kernel, specifying the
sis driver, rather than the lnc driver. Testing
with the custom kernel also seemed fine. I decided to let the
system bed down for a week or so before doing anything to the backup
server. Finally, the new server was up and running.
[N.B. 3rd March, 2008] I can no longer reproduce the problem with
the PCI/PCNet card, once I'd replaced the lnc driver by
if_le. The lnc driver is deprecated. However,
many other things have changed; now on FreeBSD 6.3 and the network
card is connected to a different switch - still my money's on the
driver.
It's apparent that the new version of gcc (3.4.2) is a lot slower at
compilation (especially when using the -MM -MG options); a
kernel build on the crimson platform used to take about one hour for
4.11; for 5.3 it takes about three hours. I therefore set the
variable NO_MODULES=true in the /etc/make.conf
file, since crimson uses no modules at all. This brings the
buildkernel time down to just less than one hour.
The install was pretty easy, once I'd found a site which saved all the hard work.
In preparation, I shrank the Windows XP partition to 25GB, leaving about 8GB for FreeBSD. I used Knoppix to shrink the NTFS partition, which works, but then some arithmetic is required (and some prayer) since the original partition must be deleted and then recreated with a new, smaller size using Linux fdisk. The IBM recovery partition takes a massive 4GB - I've since requested the Windows recovery CDs from IBM so that I can use this space. The partition/slice layout as shown by FreeBSD fdisk is below:
******* Working on device /dev/ad0 *******
parameters extracted from in-core disklabel are:
cylinders=77520 heads=16 sectors/track=63 (1008 blks/cyl)
Figures below won't work with BIOS for partitions not in cyl 1
parameters to be used for BIOS calculations are:
cylinders=77520 heads=16 sectors/track=63 (1008 blks/cyl)
Media sector size is 512
Warning: BIOS sector numbering starts with sector 1
Information from DOS bootblock is:
The data for partition 1 is:
sysid 7 (0x07),(OS/2 HPFS, NTFS, QNX-2 (16 bit) or Advanced UNIX)
start 63, size 52919937 (25839 Meg), flag 80 (active)
beg: cyl 0/ head 1/ sector 1;
end: cyl 1023/ head 29/ sector 63
The data for partition 2 is:
sysid 165 (0xa5),(FreeBSD/NetBSD/386BSD)
start 52920000, size 15597225 (7615 Meg), flag 0
beg: cyl 1023/ head 255/ sector 63;
end: cyl 1023/ head 254/ sector 63
The data for partition 3 is:
sysid 18 (0x12),(Compaq diagnostics)
start 68523840, size 9616320 (4695 Meg), flag 0
beg: cyl 1023/ head 255/ sector 63;
end: cyl 1023/ head 254/ sector 63
The data for partition 4 is:
<UNUSED>
According the miraclenet web site, if the IBM recovery partition is not on the second partition, the recovery won't work. Naturally, I didn't read that bit until it was too late. I sure hope the recovery CDs arrive before I have any problems with the Windows installation.
The FreeBSD 5.3 snapshot ISO booted and installed without any
problems. Well, there was one small one: the boot hung for a little
while during the probe for disks, and produced the error message:
"ata1-slave: FAILURE - ATAPI_IDENTIFY timed out". It
doesn't seem to cause any other problems, which is just as well, as
there is no obvious fix according to Google. Anyway, here's the
partitions defined within the FreeBSD slice:
Filesystem 1K-blocks Used Avail Capacity Mounted on /dev/ad0s2a 126702 36290 80276 31% / devfs 1 1 0 100% /dev /dev/ad0s2g 1847458 285462 1414200 17% /home /dev/ad0s2d 247726 16 227892 0% /tmp /dev/ad0s2f 4058062 1759546 1973872 47% /usr /dev/ad0s2e 247726 14514 213394 6% /var
I set the swap partition up as 1GB, probably overkill.
I had elected not to install the FreeBSD boot manager, since I
intended to use the Windows boot manager instead. When I tried to
reboot the laptop back into Windows, my heart nearly stopped. The
message "No operating system found" appeared. Had I
accidently trashed the NTFS partition? I booted from the Knoppix CD
to have a look at the partition table; the new FreeBSD
partition had been left marked as bootable. I set the NTFS
partition as bootable, and restarted. This time Windows came up.
To use the Windows boot manager, copy the file /boot/boot1
from the FreeBSD install CD to the C:\ directory under Windows. I
called the file FreeBSD.bin. [Note: /boot/boot1
can only be used when Windows and FreeBSD reside on the same disk].
Then, add C:\FreeBSD.bin to the end of
c:\boot.ini, like so:
[boot loader] timeout=30 default=multi(0)disk(0)rdisk(0)partition(1)\WINDOWS [operating systems] multi(0)disk(0)rdisk(0)partition(1)\WINDOWS="Microsoft Windows XP Professional" /fastdetect c:\FreeBSD.bin="FreeBSD"
Since the wireless card needed some additional software (see below), I used the inbuilt ethernet port to download the initial packages. The in-built port is an Intel(R) PRO/1000 Network Connection device, which uses the em driver,
The Thinkpad T42 is Centrino-based and therefore has an in-built wireless adapter, the Intel(R) PRO/Wireless 2200BG MiniPCI. Support for this network adapter is not available in the standard 5.3 release, but the driver, firmware and instructions for setup can be found on Damien Bergamini's site (no longer).
Building the driver is straightforward:
# tar xzvf iwi-freebsd-VERSION.tgz # cd iwi-freebsd-VERSION/ # make && make install
If everything goes well, it will have created the following files:
/boot/kernel/if_iwi.ko /usr/sbin/iwicontrol /usr/share/man/man4/iwi.4.gz /usr/share/man/man8/iwicontrol.8.gz
The card requires firmware, available separately. Damien supplies it as a FreeBSD package:
# pkg_add http://damien.bergamini.free.fr/ipwfw/FreeBSD/ipw-firmware-1.1.tbz
While the instructions ensured I had a working wireless card, it
wasn't obvious how the settings could be made to happen at boot
time. It turns out to be fairly easy. I added the following lines to
/boot/loader.conf:
if_iwi_load="YES" snd_ich_load="YES"
This ensure that the module for 2200BG wireless adapter (iwi) is loaded at boot time. The other line loads the driver for the on-board sound card.
To configure the iwi0 adapter, I created the file
/etc/start_if.iwi0 with the following lines:
iwicontrol -i iwi0 -d /usr/local/libdata/if_iwi -m bss ifconfig iwi0 ssid default wepkey 0xdeadbeef11 wepmode on
The first command ensures the firmware is loaded into the wireless
adapter, and the second sets the wep parameters. I see the error
message "iwi0: unknown authentication state 1" at boot time
but, once again, there seems to be no ill effects.
The on-board graphics adapter is an ATI Radeon Mobility M7 LW. I
specified radeon as the driver in the xorg.conf file
(which I had just copied from my other Thinkpad FreeBSD machine).
Once X was running, I noticed the mouse pointer was sluggish when
moving the mouse using a Microsoft USB mouse (an IntelliMouse
Explorer 4.0 USB/PS2 compatible). The touchpad moved the mouse
ok. I swapped the mouse for the version I used on the Thinkpad T30
(an IntelliMouse Optical 1.1A), which worked perfectly. Ah well, I
didn't really like that newer mouse anyway.
While testing which modules I needed for the sound card, I noticed
that loading modules after FreeBSD was booted could cause either the
wireless port or the USB mouse to stop working. However, loading at boot
time via /boot/loader.conf didn't cause these problems.
Subsequently, I installed FreeBSD 6.0 (at the RC-1 stage) on the IBM laptop. This was mostly because 6.0 includes the iwi drivers, which meant that I could forget about that extra step of building/installing the iwi drivers when performing any further kernel changes.
The upgrade was mostly painless. The iwi driver worked ok, once I
had included the wlan_wep module in the build process, as
this is now separate. I no longer needed the start_if.iwi0
script, as its work could be managed via rc.conf, with the
additional lines:
iwi_enable="YES" ifconfig_iwi0="DHCP ssid default nwkey 0xdeadbeef99"
The other change affected /etc/make/conf, where a number of
the NOXXX make options have been renamed to NO_XXX. Note that the
mergemaster setup takes quite a long time, as there are many changes
to the /etc/ files. I was glad to see that the "ata1-slave:
FAILURE - ATAPI_IDENTIFY timed out" error at boottime had
vanished.
Here's the new /etc/make.conf for 6.0:
# make.conf for OPAL # # Modification History: # Date Who # 20050311 mpw # Added options to prevent unused functionality being built # Added BATCH=true # 20051014 mpw # 6.0 changes: added iwi wlan_wep to modules # NOXXX -> NO_XXX # standard CFLAGS CFLAGS= -O -pipe # no profiled libraries NO_PROFILE=true # The following modules, only, are required MODULES_OVERRIDE= sound acpi linux linprocfs ntfs iwi wlan_wep # omit these elements from buildkernel/buildworld NO_BIND=true NO_IPFILTER=true NO_PF=true NO_AUTHPF=true NO_ATM=true NO_LPR=true NO_VINUM=true NO_BLUETOOTH=true NO_I4B=true NO_OBJC=true # can't use AAAA queries against the Alcatel DNS NO_INET6=true # Build ports in unattended mode BATCH=true # added by use.perl 2005-03-09 12:06:02 PERL_VER=5.8.6 PERL_VERSION=5.8.6
