Jim's Depository

After my dismal failure at Pohmelfs, I shopped for alternative distributed files systems. I can’t tell if CODA is abandoned or not, but I’ll give it a try.

1. Build into kernel 2.6.33, no problems.
2. Debian has let their user mode utilities rot away.
3. CMU has prebuilt .debs, including for Squeeze. That looks good.
4. Add the CMU apt sources, looks like this now… # cat /etc/apt/sources.list.d/coda.list deb http://www.coda.cs.cmu.edu/debian lenny/
5. aptitude update ; aptitude install coda-server, scary text about unsigned packages follows, but it is OK.
6. Configure server with vice-setup
7. Similar for the client machines, except aptitude install coda-client… ends with 00:35:02 Venus starting… 00:35:02 /coda now mounted
8. Test client with CMU’s server: # cfs lv /coda/testserver.coda.cs.cmu.edu Status of volume 7f000000 (2130706432) named “/” Volume type is ReadWrite Looks good. I can read the sample documents.
9. Test client with my server: Much hanging and server not founding. Most likely related to my “hostname” being an internal address unreachable by the client. There is a spot in the /etc/coda/*.conf file and in one in the /vice directory, but it still attempts to use the unreachable address and no hints in vice-setup on how to force an address.
10. I temporarily change my hostname to the external interface address. No luck. The client does a pile of DNS lookups for SRV records for my host with _codasrv._udp prepended, then does one for a regular A record. Gets my address, but never tries to contact.
11. Reboot client. Success! Now we can connect. Something evil must have been cached.
12. Many failures to authenticate with clog until I realize that the example I’m working from only has an account of admin because that is what they gave vice-setup. Use the account I gave with the password “changeme” and I’m in.
13. Find cpasswd, change that password.
14. Try to make myself a user account. au -h MYHOST nu, coda is a pretty hostile system. It echoes my new password in the clear even though it knows how not to, then fails with AuthNameToId() --> AUTH_FAILED
15. Find pdbtool. Use it to make my account from the server side, but I can’t see how to set the password. Go back and try au from a client, and now I can set my password.
16. Notice I have no group. Use pdbtool to make a group with my name, program explodes with an assert failure. That’s going to be a problem.
17. Unpack a linux kernel on the client. It flew until arch/arm, not it is crawling along at about 1 file per second. No load on server, no load on client, no load on network, just glacial progress.
18. Abandon Code!

Coda is badly documented, user hostile, and not robust.

I do not trust it.

I’ve enjoyed reading zbr’s blog postings about POHMELFS and the Elliptics network for years. With POHMELFS in the staging directories of Linux 2.6.33 and me needing to share some files across several machines, it is time to try it out.

1. Build myself some fresh new 2.6.33 kernels with POHMELFS enabled.
2. Look for a HOWTO of sorts, fail.
3. Check the server out with git: git clone http://www.ioremap.net/git/pohmelfs-server.git
4. Follow the included README
5. Fail, install autoconf.
6. Fail, finally discover that --with-kdir-path=<mylinuxsourc>/drivers/staging/pohmelfs/ is the right answer.
7. Fail, install libssl-dev.
8. make
9. make install
10. Copy /usr/local/bin/cfg over to the client machines.
11. Make a new filesystem to export mounted on /jim.
12. Start the server: fserver -r /jim
13. Go to a client machine:  #cfg -A add -a 1.1.1.1 -p 1025 -i 1 # mount -t pohmel -o idx=1 none /jim Killed
14. That isn’t good. dmesg reports there is a null pointer reference in the kernel.
15. Find two patches that will be in 2.6.33.1 related to “bdi” and not having a real backing device. Add them and make a new kernel.
16. Fail, badly. Please remember to change your version number in the kernel make file, lest you also forget the init ramdisk and make an unbootable system.
17. Success! With these patches I can now mount my pohmelfs partition! Unpacking a linux source tree is pleasently quick.
18. Failure! Chowning the source tree fills dmesg with things like…  pohmelfs_insert_hash: exist: parent: 856, ino: 879, hash: 9f4b42c5, len: 14, data: ‘README.syncppp’, new: ino: 37643, hash: 9f4b42c5, len: 14, data: ‘README.syncppp’. Chown silently and slowly fails.
19. Unmount and remount. Now chown hangs after a small amount of progress. No messages in dmesg this time. No changes make their way to the server.
20. I’m beginning to think Pohmelfs is not ready for me.

I will go shopping for other distributed file systems.

Let’s just say you have taken your old PPC G4 mac Mini with the dead hard disk and fitted it with an 8GB compact flash card in a 2.5” IDE adapter. Installing Debian Lenny on there is easy, but you end up with an ext3 journaled filesystem that may be concentrating wear on the journal area.

NILFS2 is a spiffy new filesystem that probably doesn’t concentrate wear and is attractive when writing is slower than reading. Sounds great, but yaboot doesn’t know how to find the kernel and initrd.img on it, so you need to make an ext2 /boot partion and then tell yaboot to look there.

The secret is in the naming of the files in yaboot.conf. Here is mine:

boot=/dev/hda2
device=hd:
ofboot=hd:2
partition=4
root=/dev/hda4
timeout=50
install=/usr/lib/yaboot/yaboot
magicboot=/usr/lib/yaboot/ofboot
enablecdboot

image=hd:3,/vmlinux
    partition 3
    label=Linux
    read-only
    initrd=hd:3,/initrd.img

image=hd:3,/vmlinux.old
    partition 3
    label=old
    read-only
    initrd=hd:3,/initrd.img.old

As you can see, hda3 is my /boot, hda4 is my root.

Astute readers will also notice that you have to either save a partition or migrate from another drive to pull this off since you need to build yourself a shiny new 2.6.31 kernel with NILFS2 compiled in.

Note: My G4 Mini has developed an odd hanging problem when using X programs with the root on NILFS2. I’ll go back to ext2, 2.6.26 for now then try ext2 2.6.31 and see where the trouble lies.

I just finished finding a four day bug. It sounds simple in retrospect, but oh the wrong paths I have trod and wild codes I have explored.

If, like me, you run Debian on tiny routers, like WRTSL54GS or other Broadcom units from Linksys, you probably use an OpenWRT kernel and a Debian root. Somewhere between 2.6.25 and 2.6.28 the OpenWRT people chopped the FPU emulator out of their Linux tree. Sadly for Debian people this causes any program you try to run, like init, to hang in an infinite loop.

The solution is now enshrined over at https://dev.openwrt.org/ticket/5774 where it talks about the invisible kernel config option: MIPS_FPU_EMU.

Addendum: Oh yeah, you are going to want to go into drivers/usb/host/ehci-ssb.c and pull the static off of ssb_ehci_attach and ssb_ehci_detach if you want USB 2.0 built into the kernel.

If you are going to run virtual machines in LXC containers it is nice to let each one have a console on your physical box. You can do this thusly…

Note: we are going under the knickers of lxc. It may not appreciate having its config files changed, but I didn’t find an exposed command to accomplish this.

1. Take some virtual consoles away from the main container. Edit your /etc/inittab and comment out the four lines for tty3 to tty6. We will save these for the containers.
2. kill -HUP 1
3. Now we will give tty3 to my container named rattus. You will replace rattus with your machine name.
4. Go to /var/lib/lxc/rattus.
5. Edit cgroup. The devices like 4:N are the virtual console terminals. We want to have a line 4:3 and no other 4:N where N is an integer.
6. Do the same for config.
7. Cut rattus down to two virtual consoles, the kind that can be used with lxc-console. This will be tty1 and tty2 leaving tty3 to be the real one. Edit /var/lib/lxc/rattus/ttys to be a 2 instead of a 4. Beware: There is a NUL after the digit. It may be important. Emacs will preserve it.
8. Edit the ROOT/etc/inittab file and comment out the getty line for tty4 since we don’t get to use it and it won’t have a virtual console on it. While you are in here, take out the console line as well. I think it is responsible for the spurious login prompts while you are working in the other consoles.

Linux Containers are a mechanism by which you can run multiple virtual instances of linux inside a single machine.

• + more efficient than a virtual solution like Xen, VirtualBox, or KVM.
• - all instances share a single Linux kernel.
• - I think if you expect to have actively hostile users then you should not use containers. Just a gut feeling here. They are new-ish.

Following are my notes for using Linux containers with Debian Lenny. Read along on that IBM article I just linked. It is good stuff.

1. Following a catastrophic disk failure, I begin with a clean install on a new hard drive. (Bah on you Hitachi.)
2. Install lenny. Minimal install.
3. Add openssh and emacs for my sanity. Add build-essential for kernel building.
4. Try the 2.6.30 kernel from backports. Crash on ifconfig in an inner container. Try my own 2.6.30, same problem. Try 2.6.31-rc7. All is good. So checkout 2.6.31-rc7 (or probably a real 2.6.31 when it comes out) and build that kernel. I started with the .config file from the Debian 2.6.30 backports and answered the extra questions from there. My resulting .config file is attached to the end of this posting. I no longer use make-kpkg, it rotted away too far. I just install my kernel with “make install ; make modules_install” in the kernel build directory. You then have to “update-initramfs -k 2.6.31-rc7 -c” to get the init ram disk in place. Finally “update-grub”.
5. Fail to reboot. Ha ha! The linux kernel now names my hda drives as sda and fstab fails miserably. Use “ls -l /dev/disk/by-uuid” to find my UUIDs and use them in fstab.
6. The swap partition will still fail, but I can remake it with “mkswap -L SWAP /dev/hdaXXX” and then use LABEL=SWAP in fstab after the next boot.
7. Seek out the user space tools. Decide to try to 0.6.3-1 package from unstable. wget http://ftp.us.debian.org/debian/pool/main/l/lxc/lxc_0.6.3-1_i386.deb
8. Install manually with dpkg: dpkg -i lxc_0.6.3-1_i386.deb (note: As squeeze diverges, later versions of this may have linkage problems that won’t let them install on lenny)
9. Read the /usr/share/doc/lxc/README.Debian file. See some neat stuff.
10. Run lxc-checkconfig to verify the kernel has what it takes. Notice the cgroup memory controller is disabled, but don’t care.
11. Decision time: The cgroup filesystem has to get mounted somewhere. It doesn’t matter much where. I chose to put mine in /cgroup to avoid any mount order dependencies. Some people will have ontological problems with this. Imkdir /cgroup and add this to my /etc/fstab file: cgroup  /cgroup cgroup  defaults 0 0

12. I think I have all the pieces now. Pop over to LXC: Linux container tools and slide down to “Configure networking” and lets get going. I am going to have to change my ethernet to a bridge. Here is the before and after of my /etc/network/interfaces file

    # The primary network interface
    allow-hotplug eth1
    iface eth1 inet static
          address 172.18.36.4
          netmask 255.255.255.0
          gateway 172.18.36.1
    

    and the after

    # The primary network interface
    auto br0
    iface br0 inet static
          bridge_ports eth1
          bridge_maxwait 0
          address 172.18.36.4
          netmask 255.255.255.0
          gateway 172.18.36.1auto eth1
    iface eth1 inet manual
    

    Use ifdown -a and ifup -a to make the change, from the console. I failed here because I didn’t have bridge-utils installed. Make sure you do that first. If you are using dhcp just reboot and save yourself the heartache when your old dhcp trashes your address.

13. Try using lxc-debian to make a container. You can find this in /usr/share/doc/lxc/examples. Pull it out and uncompress it, then run it. You will need to install debootstrap first. It appears to be a trainwreck. I had to patch it a little to make it run. The patch is attached below.

14. Note: lxc-debian doesn’t allow you to specify your package source, it is hard coded to ftp.debian.org. If you have a local cache, or a faster source you will want to edit a line in there. debootstrap handles its own cache, so it only sucks once.

15. When you get the “Configuring Locales” dialog, take the time to scroll down to en_US.UTF-8 and turn it on, then select it on the next page to prevent a cloud of error lines. I suppose other ones work, but not generating and selecting makes noise.

16. The last line will tell you how to start your container, mine says: 

    You can run your container with the 'lxc-start -n test1'
    

17. You now have a container with a bunch of bad config files and no editor. Fortunately you can edit the afflicted files from the main linux.

18. Go look at your ROOT/etc/network/interfaces. Mine had a bad netmask.

19. Check your ROOT/etc/apt/sources.list, it is probably missing security and volatile.

20. Go ahead and start your container now and start working form the inside.

21. I am going to use aptitude, so “apt-get update ; apt-get install aptitude”

22. Let’s get the “normal” packages installed. “aptitude update” and then aptitude install "?priority(standard)" and aptitude install "?priority(important)"

23. There are still issues starting up as a daemon and some collisions with the system console, but I’ll address those in a later post.

This should be enough to get started.

You should read the article Setting up a server for PXE network booting 

It is a bit dated, but swapping all the “etch” references for “lenny” works. I don’t get the “boot” menu when I booted my test system, so when the prompt says “boot:” just type in something like “lenny_i386_install”.

I was wondering if I needed to keep as many physical development machines or if I could just use virtual machines on my laptop and main desktop machine. I think I can certainly lose the P4 3GHz. It offers no real advantage over the Atom 330 1.6GHz.

I don’t think I’ll rely on the VirtualBox machine for Linux kernel work, but perhaps for mortal scale projects it will serve.

(pardon my confusing terminology: Compile time vs. Concurrent Jobs would have been a better graph title. i.e. make -j3 )

The blue line is different. That is a different version of gcc so you can’t directly compare the hardware with the others, but for developer experience you can make the comparison. 

I guess the bottom line is that I still lust for a Core i7 machine but will hold off for the lower power versions coming in the fall.

I used to test web pages for IE compatibility by walking around the office, finding a person with IE and asking them to pull up the page. I can’t do that anymore, and it doesn’t seem likely to work well if I go to the local coffee house and try it on a stranger, so I have found a new solution.

Microsoft makes VPC (virtual PC) system images available for minimal OS + IE installs for IE6, IE7, and IE8, some on both Vista and XP. Download details: IE App Compat VHD (These are timebombed and die after a few months, and look in the ReadMe.txt file for passwords.)

These can be run in VirtualBox on Windows, Mac OS X, Linux and Solaris hosts. The initial boot is a little ugly as Windows gropes for drivers, but once you get the VirtualBox guest additions installed it is smooth.

On my 12+ month old iMac booting an XP machine with IE takes about 22 seconds (10 seconds in BIOS, 10 booting XP, 2 logging in), restarting one from saved state takes about 6 seconds. Once running they are snappy. They don’t feel any different from native applications.

You can find excellent instructions for running these under VirtualBox at zytzagoo’s den. These are written for linux users. I used linux for the unrar step but used my Mac for the VirtualBox machine.

If you want to test with more than one version of IE you will discover that Microsoft used the same UUID for all of the disk images (challenging the meaning of unique) and that VirtualBox is offended by that. You can read the also good instructions at Shape Shed and look at the Fixing Microsoft’s Duplicate Identifiers section for details on converting the image to a raw image and back to a VDI to get a new UUID.

My EEE PC is nearly a 900A, it was a bit crippled to be a super lower than low end Best Buy machine.

First I will create a nilfs2 rooted system on an SD card to verify the steps, then I will migrate that to the internal flash. I am running 2.6.30 with no init ramdisk. Most of the important bits are compiled in, some fringe stuff is a module.

1. Read about aligning filesystems to flash erase boundaries. I have no idea what my device’s erase block size is, but surely it is a power of two and less than 128k, so I will use 128k. I can conceive of a test I could write to establish this, but will not do it.  Go read Aligning filesystems to an SSD’s erase block size | Thoughts by Ted to learn what I know.
2. For some silly reason, my internal drive is sdb and my SD reader is sda. Keep that in mind, don’t trash your system if yours is different.
3. Per Ted’s thoughts (see above) I did an fdisk -H 224 -S 56 /dev/sda and then created three partitions. A /boot at the front, this will be ext3 for grub’s sake, a single track partition 4 typed as 0xef for the BIOS quick boot, and the rest in the middle as my nilfs2 root. (I am willing to believe that this is a bad idea and that the GC of nilfs2 might be happier if I put my log files and such in a different place from my static files, but oh well.)
4. mkfs -t nilfs2 /dev/sda2 is amusingly fast.
5. Mount the new root on /mnt and the new boot on /mnt/boot
6. debootstrap –arch i386 lenny /mnt YOUR-PACKAGE-SOURCE (find that from your /etc/apt/sources file) (I have to do this twice, the first time it complains that Packages.gz is corrupt.)
7. chroot /mnt  this moves you to the new system. I “touch” a file at the root to identify it so I can tell where I am later… touch /SDCARD
8. exit to leave the chroot.
9. cp /boot/*2.6.30 /mnt/boot/  – you’ll want your kernel
10. cp -arv /lib/modules/2.6.30 /mnt/lib/modules/2.6.30 – and it will want its modules.
11. cp /etc/apt/sources.list /mnt/etc/apt/sources.list
12. chroot /mnt
13. aptitude update
14. aptitude install locales  (we need to fix that storm of locale errors pending)
15. dpkg-reconfigure locales (I choose nothing.)
16. aptitude full-upgrade
17. aptitude install openssh-server
18. aptitude install nilfs2-tools (sad, 100MB of crap to install this)
19. aptitude install grub
20. mkdir /boot/grub
21. leave the chroot and get some config files
22. cp /etc/network/interfaces /mnt/etc/network/
23. cp /etc/hostname /mnt/etc/
24. cp /etc/hosts /mnt/etc/
25. reboot, catch the boot at grub and change to the new root
26. grub-install /dev/sda  (DANGER! This make break going back to the internal, since I used the internal grub to get here. Maybe manually make the /dev/sda* nodes so grub will work before the reboot?)
27. Here goes… time to repeat in reverse to install on the internal flash.
28. All is good. I am confused why my menu.lst file needs root to be (hd0,0) when I am on sdb, but I swapped the SD card for a blank one and I can still boot, so I’m not getting grub bits from there. Probably some BIOS stunt.
29. Get the nilfs2-tools from squeeze, the one in lenny is too old to work.
30. The kernel command line option root=UUID=xxxx or root-LABEL=xxxx don’t work because they are not kernel options, they only look like it and are handled in some common initrd programs.

After several days of use I can add:

1. It feels better than the old ext3 install. That is a terrible metric, I’ll have to do something more rigorous when I get home.
2. As installed it doesn’t reclaim disk space, it just fills up. If I manually run the nilfs cleanerd then it does reclaim space. For some reason the one Debian starts doesn’t work.

After 25000km of travel with the a nilfs2 root I have to say I like it a lot. Now to figure out why the Debian start of the cleanerd doesn’t reclaim space.