Zugschlusbeobachtungen

Yes, You Should Not discard cached ssh host keys without looking. An unexpected change of an ssh host key is always a reason to step back from the keyboard and think. However, there are situations when you know that a systems' ssh host key has changed, for example when the system reachable under this host name has been redeployed, which happens increasingly often proportionally to the devopsness of your environment, or for example in test environments.

Later versions of ssh offer you the ssh-keygen -R command line to paste from the error message, so that you can abort the connection attempt, paste the command and reconnect again. This will still ask for confirmation of the new host key though.

Almost every sysadmin has an alias or wrapper to make handling of this situation easier. Solutions range from using "StrictHostKeyChecking no" and/or "UserKnownHostsFile /dev/null", turning off this layer of securit altogether either globally or usually too broadly, to more-or-less sophisticated solutions that involve turning off know-host file hashing, parsing client output and/or grep-sed-awk magic. grml even comes with an insecssh script that is rather neat and that I used until I developed my own.

Starting with OpenSSH 7.2, a new "restrict" option for authorized_keys lines has become available. It sets all available restrictions that the current OpenSSH version can do (like no-agent-forwarding, no-x11-forwarding etc). One can individually turn on those features again by corresponding new options.

This saves one from sorrows when a new capability of OpenSSH is introduced through an update which is enabled by default, since one has to remember that restricted authorized_keys lines are in unse and then to manually add the restrictions.

On the downside, Debian jessie and CentOS 7 don't have a recent enough OpenSSH. So we'll have to continue worrying about new features being inadvertendly enabled for a while.

P.S.: Yes, I haven't blogged about Linux and Debian in English in a while.

I have just uploaded PowerDNS 3.3-1~exp1 to Debian experimental. This new upstream version has introduced its own include directive, so Debian was able to drop its patch. Hence, our conffiles had to grow a .conf extension, which most of them didn't have in previous version.

If anybody wants to test updates from PowerDNS 3.1 to the new 3.3-1~exp1 in Debian unstable, please go ahead and report bugs in the Debian BTS. The package is known to not offer seamless DNSSEC, I'll work on that before I upload to unstable.

While we're at it: I would appreciate help with the PostgreSQL backend. Myself, I use mainly MySQL and am not too proficient in PostgreSQL. I'll accept both patches and more formal co-maintenance.

Migrating a Debian installation between architectures has always been difficult. The recommended way to "crossgrade" an i386 Debian to amd64 Debian is to reinstall the system, move over data and configuration. For the more brave, in-place crossgrades usually involved chroots, rescue CDs, a lot of ar p | tar xf - data.tar.gz and luck.

I have never been brave when it comes to system administration, have done a lot of architecture migrations with reinstallation, and have always taken the opportunity to clear out the contamination that accumulates itself when a system is running for a long time. I would even recommend doing this to most people even now. However, I have a few very ugly systems in place that are still on i386 because I didn't dare going the reinstallation path.

Doing in-place crossgrades has become a lot easier since wheezy's release, since once now can have both i386 and amd64 libraries installed in parallel, which allows to replace foo:i386 with foo:amd64 without influencing the other parts of the system. The process is still full of pitfalls:

• In wheezy, many library packages are multiarch capable. This means that you can have those library packages installed for more than one architecture. This is a technical must for this way of crossgrade, so never use that for an older-than-wheezy system. It won't work, it needs at least Debian wheezy. Unfortunately, not all libraries in wheezy are multiarch capable. This makes the process harder and a lot less predictable, since a crosscrade including such packages is going to spew incomprehensible and misleading apt error messages. In my experience, for example the libaprutil-1-dbd-* packages and libonig2 are of this kind.
• apt removes a package before it reinstalls its new counterpart. This results in apt calling dpkg to remove dpkg, and then calling dpkg again to install dpkg. Guess which operation fails and the state of the system after this failure. Same applies to coreutils, which leaves the system without rm, which in turn dpkg of either architecture doesn't like. Using apt-get --download-only install to resolve dependencies and downloading the debs, followed by a traditional dpkg --install solves this issue since multiarch dpkg will replace a package with another one without deinstalling the first one first.
• At least for the process, you need a kernel that can run both 32bit and 64bit binaries for the i386 architecture. AFAIR, setting CONFIG_64BIT, CONFIG_X86_64 and CONFIG_IA32_EMULATION in the kernel configuration takes care of this.
• During the process, apt will temporary go into a badly broken state where it will refuse most operations. Be aware that you might need to manually download packages from the Internet. Be sure to have wget, curl, or a browser (maybe a text based one like elinks) available. dget is not going to help you here since it will only downloda packages for the native arch.
• During the process, apt wants to remove the better part of your system. It is important to not let it do this, as it wants to deinstall essential packages as well.
• Watch what your system does. During some steps, it might remove packages you might need. Keep track of the packages that were removed during the process and re-install them manually after finishing the crossgrade. Be sure not to purge packages that you might still need.
• It looks like the process is not always exactly reproducible. During the first tries, I found myself without an initrd at all, with an initrd that lacked the ext[234].ko kernel modules, without working e2fstools and in a number of other undesireable states of the system.

I have only tried this yet with a freshly installed minimal wheezy server system. Trying the process with "real life" systems has shown to be full of more surprises. I will document other pitfalls I have fallen into here at a later time. My minimal wheezy system was running in a KVM VM with its virtual disk as a LVM LV in the host system. I took a snapshot before beginning and used lvconvert --merge numerous time to return my LV to the original state. Be aware that lvconvert --merge removes the snapshot after merging it, so you'll need to re-create the snapshot before trying again.

The process is absolutely not for the faint of the heart, and intimate knowlegde of Debian mechanisms is required at many points in the process. Please seriously consider a reinstall+migrate approach instead of using this process, and be sure to practice it on a working copy of your system before touching the live system. And always have a backup.

During the process, I discussed things with Paul Tagliamonte, who has done this before, but on a live system and with a slightly more invasive approach. He has blogged about this. Thank you very much, your hints were very helpful.

I have published PowerDNS version 3.1-1.0 on https://ivanova.notwork.de/~mh/debian/pdns/

This is a preliminary package and a release candidate to be 3.1-2 in Debian. If you're interested in PowerDNS on Debian, please test this package.

I plan to upload next week. This package will vanish from the web server once the package is visible in Debian.

Eight days ago, I uploaded atop 1.26-1 to DELAYED/8, listing me as new maintainer. This means that the package has in the mean time appeared in unstable, and I hope that it'll swiftly migrate to testing.

Dear Lazyweb, for a long time I have been using iproute2's label feature to assign arbitrary labels to IP addresses configured on Interfaces:

40: int152@dotqa: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP
link/ether 00:25:b3:01:e5:6c brd ff:ff:ff:ff:ff:ff
inet 10.1.152.254/24 brd 10.1.152.255 scope global int152:98fe8

Recently, this has shown to at least confuse both isc-dhcp-relay (#617258) and dhcp-helper (#617264).

As I have never seen interfaces labels used outside my firewalls (which happen to use ifupdown-scripts-zg2 (Debian PTS)) and ifupdown's rather twisted handling of multiple IP addresses per interface (using Alias Interfaces), I'd like to know whether my usage is a legitimate one and whether there are other uses for interface labels.

At the moment, I'm tempted to remove label support from ifupdown-scripts-zg2 in the next release, or to make it optional. Please comment if you have an opinion.

It is well known that apt has an issue when it comes to resolving circular dependencies. Therefore, Debian bug reporters have set out to eradicate circular dependencies from the archive. This does, however, add significant bloat to the actual packages, and I am questioning why this is really necessary.

In the last few days, I found the time to spend some with KVM and libvirt. Unfortunately, there is a subject that I haven't yet found a satisfying solution: Naming of block devices in guest instances.

This is surely a common issue, but solutions are rare. Neither an article on Usenet (in German) nor the German version of this blog article has found solutions for the main question. I should have written this in English in the first place and am thus translating from German to english, hoping that there will be some answers and suggestions.

KVM is quite inflexible when it coms to configure block devices. It is possible to define on the host, which files or whole devices from the host should be visible in the guest. The documentation suggests that devices should be brought into the guest with the virtio model, which needs suppport in the guest kernel. Importing a device as emulated ATA or SCSI device brings a performance penalty.

The devices brought into the guest via virtio appear in the guest's dev as /dev/vd<x> and do also have their corresponding entries in /dev/disk/by-uuid and /dev/disk/by-path. The vd<x> node is simply numbered in consecutive order as hd<x> and sd<x>. /dev/disk/by-uuid is the correct UUID of the file system found on the device, at least if it's a block device partitioned inside the guest and formatted with ext3 (I didn't try anything else yet). The terminology of the /dev/disk/by-path node is not yet understood, and I am somewhat reluctant to assume the PCI paths of emulated hardware as stable.

Four days before my wedding, I spent some time researching booting a PC from a large hard disk, where large means "larger than two Terabytes". These days, single disks are approaching this size, so we are near the state where this issue pops up for your run-of-the-mill computer rather than the data store RAID. Today, the per-gigabyte price is however still significantly cheaper if you go for a 1 T or an 1.5 T disk.

The old blog article shows that I spent considerable time in finding out today's limitations below the 2 T limit by using conventional partitioning schemes to boot a 2 T disk. Since I don't have this much storage available at the moment, I had to use virtualization and to take advantage of nearly empty virtual disks taking up much less space than their raw capacity suggests. This works fine as long as you don't start actually using the disk.

Back then, the only combination that worked for a raw disk larger than 2 T (only using the first 2 T) was Virtualbox and grub 0 (now grub-legacy). I regret to admit that the results of my experiments from June are not any more reproducible (most probably due to changes in Virtualbox since then) and that I was not able to boot any disk larger than 2 T any more, even if the partitions were well below the 2 T limit. I chose to ignore these results and to finally start the GPT research.

Steinar H. Gunderson, sesse, has written an interesting article about TCP performance. I didn't find your blog's comment function, so I am commenting with a trackback. (note: which didn't work either, "The auto-discovered trackback URI does not match our target URI")

I frequently use mobile internet, using various of the German GSM/UMTS network operators, out of a moving train. As you have written, this frequently causes packet loss which is not only not caused by congestion, but sends the congestion avoidance algorithms on a false path.

For example, when the train passes through the 3575 m long Distelrasentunnel between Frankfurt and Fulda, my network link is broken for like two minutes. Passing through other parts of Germany sometimes gives me a ping response of hundreds of thousands of microseconds by virtue of the rather huge send buffer the UMTS equipment has.

In these circumstances, ssh sessions frequently take tens of minutes to notice that the network is back before the session is useable again. Frequently, it doesn't come up again before an hour has passed. And I have not found a way to work myself around this. Can you explain what's happening here, and do you have any ideas to solve the issue?

Dear Lazyweb, sorry to bother you again, but I have tried to get this question answered on IRC, on Usenet and on the Samba Mailing List, and was not able to get an answer (not even a remotely clueless one) there. Can you help?

I currently have an "interesting" task to accomplish: An IT environment with about 90 % Windows and 10 % Linux machines would like to unify backup. Currently, the Windows world backs itself up to tape using Backup Exec; the Linux world has Amanda backing up to a big disk RAID.

This RAID is acting up and is scheduled to disappear. The current plan is to back up the Linux world with Amanda to a Samba share which is then backed up to tape by the Backup Exec installation running in the Windows world.

The Linux systems are in a diffent network, and the firewall people would like to keep the ports being open between the two networks to the bare minimum. I don't want to see NETBIOS Broadcasts inside the Linux world, I don't want to see this server in any network neighborhood, and the system acting as the Samba server for the backup should have as few open ports as possible. Of course, the share should be read only and to be as secure as possible.

Dear Lazyweb, can anybody with some advanced socat-foo tell me the command line needed to have socat create a socket in the local file system and to listen on it, so that I can have Virtualbox connect a virtual serial console to it?

The material available on socat on the web is sparse, and virtualbox-related docs usually contain "tick the create pipe option", which is not helpful here since I would like to see the first output the virtual machine prints to its serial port. It would be vastly more useful to have the socket already created with socat listening so that I can immediately see what is being printed to the socket.

Traditionally, the Linux kernel is software that I compile myself from pristine upstream sources for various reasons. I have three major kernel flavours that get built (server, desktop and notebook), and I am pretty current in running a bleeding edge kernel. This is not really necessary any more nowadays, but it's a tradition that works pretty well.

My kernels get built on sid and are packaged up with kernel-package, and equivs builds a dependency helper package which pulls in the kernel's dependencies such as initramfs-tools and takes care of cross-version updates like going from 2.6.29 to 2.6.30. Up to now, I was always able to run a kernel built this way on all my systems which can range from oldstable to unstable.

I have been using current KDE since most of my Linux time (having converted over from WindowMaker to KDE 2 back in 2002). But currently, I am seriously pondering to ditch KDE since KDE upstream seems to be wildly decided to kill KDE.

I have accidentally upgraded my desktop box to KDE4 because I missed putting KDE on hold before doing a major sid update after a couple of months. KDE4's first regression immediately showed itself - the right display doesn't get any attention from KDE. It just shows up in a grey checkerboard background, it doesn't have a panel, it doesn't have a menu, right click doesn't work. It looks like the only thing one can do with it is dragging windows onto it.

With help of #debian-kde, I quickly found out about this bug in Upstream Bugzilla, which is referred from #529487 and which was marked as Duplicate of this bug in upstream bugzilla, which is one and a half years old and was marked as "severity wishlist".

Despite the splendid job that the Debian KDE team has done to sort out the KDE4 mess, it looks like KDE upstream has managed to break Dual Head Setups for one and a half years and doesn't seem to be too interested in providing KDE4 in a way that it can be compared with past versions. This is very sad and will have me shopping for a new desktop environment soon, I am afraid.

Maybe it was not a so good idea to take away KDE 3 so soon and it might have been better to keep KDE 3 in Debian. Maybe it's time to re-introduce KDE 3 as co-installable packages? I would be willing to participate in this effort as a team member.

Which other Desktop Environments and/or Window Managers should I be shopping for? I'd like to have:

• Dual-Head support (preferably with the possibility to switch desktops only on one display, but that's something that even KDE 3 cannot do yet)
• Shortcuts like "gg:search words" or "wp:search words" to immediately open google, wikipedia, the BTS or the PTS
• Overlapping windows that are not automatically resized
• A terminal like konsole which allows me to have different session in tabs and to send my input to all tabs
• A clipboard handler that will automatically pop up a window asking me whether I want to open the URL that I just marked in a browser
• Integration with the Debian menu system
I will try adding to this list over the next days when I notice a feature that I have accustomed to so badly that I don't even notice any more when I'm using it.