Search Results: "sto"

17 July 2025

Arnaud Rebillout: Acquire-By-Hash for APT packages repositories, and the lack of it in Kali Linux

This is a lenghty blog post. It features a long introduction that explains how apt update acquires various files from a package repository, what is Acquire-By-Hash, and how it all works for Kali Linux: a Debian-based distro that doesn't support Acquire-By-Hash, and which is distributed via a network of mirrors and a redirector. In a second part, I explore some "Hash Sum Mismatch" errors that we can hit with Kali Linux, errors that would not happen if only Acquire-By-Hash was supported. If anything, this blog post supports the case for adding Acquire-By-Hash support in reprepro, as requested at https://bugs.debian.org/820660. All of this could have just remained some personal notes for myself, but I got carried away and turned it into a blog post, dunno why... Hopefully others will find it interesting, but you really need to like troubleshooting stories, packed with details, and poorly written at that. You've been warned! Introducing Acquire-By-Hash Acquire-By-Hash is a feature of APT package repositories, that might or might not be supported by your favorite Debian-based distribution. A repository that supports it says so, in the Release file, by setting the field Acquire-By-Hash: yes. It's easy to check. Debian and Ubuntu both support it:
$ wget -qO- http://deb.debian.org/debian/dists/sid/Release   grep -i ^Acquire-By-Hash:
Acquire-By-Hash: yes
$ wget -qO- http://archive.ubuntu.com/ubuntu/dists/devel/Release   grep -i ^Acquire-By-Hash:
Acquire-By-Hash: yes
What about other Debian derivatives?
$ wget -qO- http://http.kali.org/kali/dists/kali-rolling/Release   grep -i ^Acquire-By-Hash:   echo not supported
not supported
$ wget -qO- https://archive.raspberrypi.com/debian/dists/trixie/Release   grep -i ^Acquire-By-Hash:   echo not supported
not supported
$ wget -qO- http://packages.linuxmint.com/dists/faye/Release   grep -i ^Acquire-By-Hash:   echo not supported
not supported
$ wget -qO- https://apt.pop-os.org/release/dists/noble/Release   grep -i ^Acquire-By-Hash:   echo not supported
not supported
Huhu, Acquire-By-Hash is not ubiquitous. But wait, what is Acquire-By-Hash to start with? To answer that, we have to take a step back and cover some basics first. The HTTP requests performed by 'apt update' What happens when one runs apt update? APT first requests the Release file from the repository(ies) configured in the APT sources. This file is a starting point, it contains a list of other files (sometimes called "Index files") that are available in the repository, along with their hashes. After fetching the Release file, APT proceeds to request those Index files. To give you an idea, there are many kinds of Index files, among which: There's an excellent Wiki page that details the structure of a Debian package repository, it's there: https://wiki.debian.org/DebianRepository/Format. Note that APT doesn't necessarily download ALL of those Index files. For simplicity, we'll limit ourselves to the minimal scenario, where apt update downloads only the Packages files. Let's try to make it more visual: here's a representation of a apt update transaction, assuming that all the components of the repository are enabled:
apt update -> Release -> Packages (main/amd64)
                      -> Packages (contrib/amd64)
                      -> Packages (non-free/amd64)
                      -> Packages (non-free-firmware/amd64)
Meaning that, in a first step, APT downloads the Release file, reads its content, and then in a second step it downloads the Index files in parallel. You can actually see that happen with a command such as apt -q -o Debug::Acquire::http=true update 2>&1 grep ^GET. For Kali Linux you'll see something pretty similar to what I described above. Try it!
$ podman run --rm kali-rolling apt -q -o Debug::Acquire::http=true update 2>&1   grep ^GET
GET /kali/dists/kali-rolling/InRelease HTTP/1.1    # <- returns a redirect, that is why the file is requested twice
GET /kali/dists/kali-rolling/InRelease HTTP/1.1
GET /kali/dists/kali-rolling/non-free/binary-amd64/Packages.gz HTTP/1.1
GET /kali/dists/kali-rolling/main/binary-amd64/Packages.gz HTTP/1.1
GET /kali/dists/kali-rolling/non-free-firmware/binary-amd64/Packages.gz HTTP/1.1
GET /kali/dists/kali-rolling/contrib/binary-amd64/Packages.gz HTTP/1.1
However, and it's now becoming interesting, for Debian or Ubuntu you won't see the same kind of URLs:
$ podman run --rm debian:sid apt -q -o Debug::Acquire::http=true update 2>&1   grep ^GET
GET /debian/dists/sid/InRelease HTTP/1.1
GET /debian/dists/sid/main/binary-amd64/by-hash/SHA256/22709f0ce67e5e0a33a6e6e64d96a83805903a3376e042c83d64886bb555a9c3 HTTP/1.1
APT doesn't download a file named Packages, instead it fetches a file named after a hash. Why? This is due to the field Acquire-By-Hash: yes that is present in the Debian's Release file. What does Acquire-By-Hash mean for 'apt update' The idea with Acquire-By-Hash is that the Index files are named after their hash on the repository, so if the MD5 sum of main/binary-amd64/Packages is 77b2c1539f816832e2d762adb20a2bb1, then the file will be stored at main/binary-amd64/by-hash/MD5Sum/77b2c1539f816832e2d762adb20a2bb1. The path main/binary-amd64/Packages still exists (it's the "Canonical Location" of this particular Index file), but APT won't use it, instead it downloads the file located in the by-hash/ directory. Why does it matter? This has to do with repository updates, and allowing the package repository to be updated atomically, without interruption of service, and without risk of failure client-side. It's important to understand that the Release file and the Index files are part of a whole, a set of files that go altogether, given that Index files are validated by their hash (as listed in the Release file) after download by APT. If those files are simply named "Release" and "Packages", it means they are not immutable: when the repository is updated, all of those files are updated "in place". And it causes problems. A typical failure mode for the client, during a repository update, is that: 1) APT requests the Release file, then 2) the repository is updated and finally 3) APT requests the Packages files, but their checksum don't match, causing apt update to fail. There are variations of this error, but you get the idea: updating a set of files "in place" is problematic. The Acquire-By-Hash mechanism was introduced exactly to solve this problem: now the Index files have a unique, immutable name. When the repository is updated, at first new Index files are added in the by-hash/ directory, and only after the Release file is updated. Old Index files in by-hash/ are retained for a while, so there's a grace period during which both the old and the new Release files are valid and working: the Index files that they refer to are available in the repo. As a result: no interruption of service, no failure client-side during repository updates. This is explained in more details at https://www.chiark.greenend.org.uk/~cjwatson/blog/no-more-hash-sum-mismatch-errors.html, which is the blog post from Colin Watson that came out at the time Acquire-By-Hash was introduced in... 2016. This is still an excellent read in 2025. So you might be wondering why I'm rambling about a problem that's been solved 10 years ago, but then as I've shown in the introduction, the problem is not solved for everyone. Support for Acquire-By-Hash server side is not for granted, and unfortunately it never landed in reprepro, as one can see at https://bugs.debian.org/820660. reprepro is a popular tool for creating APT package repositories. In particular, at Kali Linux we use reprepro, and that's why there's no Acquire-By-Hash: yes in the Kali Release file. As one can guess, it leads to subtle issues during those moments when the repository is updated. However... we're not ready to talk about that yet! There's still another topic that we need to cover: this window of time during which a repository is being updated, and during which apt update might fail. The window for Hash Sum Mismatches, and the APT trick that saves the day Pay attention! In this section, we're now talking about packages repositories that do NOT support Acquire-By-Hash, such as the Kali Linux repository. As I've said above, it's only when the repository is being updated that there is a "Hash Sum Mismatch Window", ie. a moment when apt update might fail for some unlucky clients, due to invalid Index files. Surely, it's a very very short window of time, right? I mean, it can't take that long to update files on a server, especially when you know that a repository is usually updated via rsync, and rsync goes to great length to update files the most atomically as it can (with the option --delay=updates). So if apt update fails for me, I've been very unlucky, but I can just retry in a few seconds and it should be fixed, isn't it? The answer is: it's not that simple. So far I pictured the "package repository" as a single server, for simplicity. But it's not always what it is. For Kali Linux, by default users have http.kali.org configured in their APT sources, and it is a redirector, ie. a web server that redirects requests to mirrors that are nearby the client. Some context that matters for what comes next: the Kali repository is synced with ~70 mirrors all around the world, 4 times a day. What happens if your apt update requests are redirected to 2 mirrors close-by, and one was just synced, while the other is still syncing (or even worse, failed to sync entirely)? You'll get a mix of old and new Index files. Hash Sum Mismatch! As you can see, with this setup the "Hash Sum Mismatch Window" becomes much longer than a few seconds: as long as nearby mirrors are syncing the window is opened. You could have a fast and a slow mirror next to you, and they can be out of sync with each other for several minutes every time the repository is updated, for example. For Kali Linux in particular, there's a "detail" in our network of mirrors that, as a side-effect, almost guarantees that this window lasts several minutes at least. This is because the pool of mirrors includes kali.download which is in fact the Cloudflare CDN, and from the redirector point of view, it's seen as a "super mirror" that is present in every country. So when APT fires a bunch of requests against http.kali.org, it's likely that some of them will be redirected to the Kali CDN, and others will be redirected to a mirror nearby you. So far so good, but there's another point of detail to be aware of: the Kali CDN is synced first, before the other mirrors. Another thing: usually the mirrors that are the farthest from the Tier-0 mirror are the longest to sync. Packing all of that together: if you live somewhere in Asia, it's not uncommon for your "Hash Sum Mismatch Window" to be as long as 30 minutes, between the moment the Kali CDN is synced, and the moment your nearby mirrors catch up and are finally in sync as well. Having said all of that, and assuming you're still reading (anyone here?), you might be wondering... Does that mean that apt update is broken 4 times a day, for around 30 minutes, for every Kali user out there? How can they bear with that? Answer is: no, of course not, it's not broken like that. It works despite all of that, and this is thanks to yet another detail that we didn't go into yet. This detail lies in APT itself. APT is in fact "redirector aware", in a sense. When it fetches a Release file, and if ever the request is redirected, it then fires the subsequent requests against the server where it was initially redirected. So you are guaranteed that the Release file and the Index files are retrieved from the same mirror! Which brings back our "Hash Sum Mismatch Window" to the window for a single server, ie. something like a few seconds at worst, hopefully. And that's what makes it work for Kali, literally. Without this trick, everything would fall apart. For reference, this feature was implemented in APT back in... 2016! A busy year it seems! Here's the link to the commit: use the same redirection mirror for all index files. To finish, a dump from the console. You can see this behaviour play out easily, again with APT debugging turned on. Below we can see that only the first request hits the Kali redirector:
$ podman run --rm kali-rolling apt -q -o Debug::Acquire::http=true update 2>&1   grep -e ^Answer -e ^HTTP
Answer for: http://http.kali.org/kali/dists/kali-rolling/InRelease
HTTP/1.1 302 Found
Answer for: http://mirror.freedif.org/kali/dists/kali-rolling/InRelease
HTTP/1.1 200 OK
Answer for: http://mirror.freedif.org/kali/dists/kali-rolling/non-free-firmware/binary-amd64/Packages.gz
HTTP/1.1 200 OK
Answer for: http://mirror.freedif.org/kali/dists/kali-rolling/contrib/binary-amd64/Packages.gz
HTTP/1.1 200 OK
Answer for: http://mirror.freedif.org/kali/dists/kali-rolling/main/binary-amd64/Packages.gz
HTTP/1.1 200 OK
Answer for: http://mirror.freedif.org/kali/dists/kali-rolling/non-free/binary-amd64/Packages.gz
HTTP/1.1 200 OK
Interlude Believe it or not, we're done with the introduction! At this point, we have a good understanding of what apt update does (in terms of HTTP requests), we know that Release files and Index files are part of a whole, and we know that a repository can be updated atomically thanks to the Acquire-By-Hash feature, so that users don't experience interruption of service or failures of any sort, even with a rolling repository that is updated several times a day, like Debian sid. We've also learnt that, despite the fact that Acquire-By-Hash landed almost 10 years ago, some distributions like Kali Linux are still doing without it... and yet it works! But the reason why it works is more complicated to grasp, especially when you add a network of mirrors and a redirector to the picture. Moreover, it doesn't work as flawlessly as with the Acquire-By-Hash feature: we still expect some short (seconds at worst) "Hash Sum Mismatch Windows" for those unlucky users that run apt update at the wrong moment. This was a long intro, but that really sets the stage for what comes next: the edge cases. Some situations in which we can hit some Hash Sum Mismatch errors with Kali. Error cases that I've collected and investigated over the time... If anything, it supports the case that Acquire-By-Hash is really something that should be implemented in reprepro. More on that in the conclusion, but for now, let's look at those edge cases. Edge Case 1: the caching proxy If you put a caching proxy (such as approx, my APT caching proxy of choice) between yourself and the actual packages repository, then obviously it's the caching proxy that performs the HTTP requests, and therefore APT will never know about the redirections returned by the server, if any. So the APT trick of downloading all the Index files from the same server in case of redirect doesn't work anymore. It was rather easy to confirm that by building a Kali package during a mirror sync, and watch if fail at the "Update chroot" step:
$ sudo rm /var/cache/approx/kali/dists/ -fr
$ gbp buildpackage --git-builder=sbuild
+------------------------------------------------------------------------------+
  Update chroot                                Wed, 11 Jun 2025 10:33:32 +0000  
+------------------------------------------------------------------------------+
Get:1 http://http.kali.org/kali kali-dev InRelease [41.4 kB]
Get:2 http://http.kali.org/kali kali-dev/contrib Sources [81.6 kB]
Get:3 http://http.kali.org/kali kali-dev/main Sources [17.3 MB]
Get:4 http://http.kali.org/kali kali-dev/non-free Sources [122 kB]
Get:5 http://http.kali.org/kali kali-dev/non-free-firmware Sources [8297 B]
Get:6 http://http.kali.org/kali kali-dev/non-free amd64 Packages [197 kB]
Get:7 http://http.kali.org/kali kali-dev/non-free-firmware amd64 Packages [10.6 kB]
Get:8 http://http.kali.org/kali kali-dev/contrib amd64 Packages [120 kB]
Get:9 http://http.kali.org/kali kali-dev/main amd64 Packages [21.0 MB]
Err:9 http://http.kali.org/kali kali-dev/main amd64 Packages
  File has unexpected size (20984689 != 20984861). Mirror sync in progress? [IP: ::1 9999]
  Hashes of expected file:
   - Filesize:20984861 [weak]
   - SHA256:6cbbee5838849ffb24a800bdcd1477e2f4adf5838a844f3838b8b66b7493879e
   - SHA1:a5c7e557a506013bd0cf938ab575fc084ed57dba [weak]
   - MD5Sum:1433ce57419414ffb348fca14ca1b00f [weak]
  Release file created at: Wed, 11 Jun 2025 07:15:10 +0000
Fetched 17.9 MB in 9s (1893 kB/s)
Reading package lists...
E: Failed to fetch http://http.kali.org/kali/dists/kali-dev/main/binary-amd64/Packages.gz  File has unexpected size (20984689 != 20984861). Mirror sync in progress? [IP: ::1 9999]
   Hashes of expected file:
    - Filesize:20984861 [weak]
    - SHA256:6cbbee5838849ffb24a800bdcd1477e2f4adf5838a844f3838b8b66b7493879e
    - SHA1:a5c7e557a506013bd0cf938ab575fc084ed57dba [weak]
    - MD5Sum:1433ce57419414ffb348fca14ca1b00f [weak]
   Release file created at: Wed, 11 Jun 2025 07:15:10 +0000
E: Some index files failed to download. They have been ignored, or old ones used instead.
E: apt-get update failed
The obvious workaround is to NOT use the redirector in the approx configuration. Either use a mirror close by, or the Kali CDN:
$ grep kali /etc/approx/approx.conf 
#kali http://http.kali.org/kali <- do not use the redirector!
kali  http://kali.download/kali
Edge Case 2: debootstrap struggles What if one tries to debootstrap Kali while mirrors are being synced? It can give you some ugly logs, but it might not be fatal:
$ sudo debootstrap kali-dev kali-dev http://http.kali.org/kali
[...]
I: Target architecture can be executed
I: Retrieving InRelease 
I: Checking Release signature
I: Valid Release signature (key id 827C8569F2518CC677FECA1AED65462EC8D5E4C5)
I: Retrieving Packages 
I: Validating Packages 
W: Retrying failed download of http://http.kali.org/kali/dists/kali-dev/main/binary-amd64/Packages.gz
I: Retrieving Packages 
I: Validating Packages 
W: Retrying failed download of http://http.kali.org/kali/dists/kali-dev/main/binary-amd64/Packages.gz
I: Retrieving Packages 
I: Validating Packages 
W: Retrying failed download of http://http.kali.org/kali/dists/kali-dev/main/binary-amd64/Packages.gz
I: Retrieving Packages 
I: Validating Packages 
W: Retrying failed download of http://http.kali.org/kali/dists/kali-dev/main/binary-amd64/Packages.gz
I: Retrieving Packages 
I: Validating Packages 
I: Resolving dependencies of required packages...
I: Resolving dependencies of base packages...
I: Checking component main on http://http.kali.org/kali...
I: Retrieving adduser 3.152
[...]
To understand this one, we have to go and look at the debootstrap source code. How does debootstrap fetch the Release file and the Index files? It uses wget, and it retries up to 10 times in case of failure. It's not as sophisticated as APT: it doesn't detect when the Release file is served via a redirect. As a consequence, what happens above can be explained as such:
  1. debootstrap requests the Release file, gets redirected to a mirror, and retrieves it from there
  2. then it requests the Packages file, gets redirected to another mirror that is not in sync with the first one, and retrieves it from there
  3. validation fails, since the checksum is not as expected
  4. try again and again
Since debootstrap retries up to 10 times, at some point it's lucky enough to get redirected to the same mirror as the one from where it got its Release file from, and this time it gets the right Packages file, with the expected checksum. So ultimately it succeeds. Edge Case 3: post-debootstrap failure I like this one, because it gets us to yet another detail that we didn't talk about yet. So, what happens after we successfully debootstraped Kali? We have only the main component enabled, and only the Index file for this component have been retrieved. It looks like that:
$ sudo debootstrap kali-dev kali-dev http://http.kali.org/kali
[...]
I: Base system installed successfully.
$ cat kali-dev/etc/apt/sources.list
deb http://http.kali.org/kali kali-dev main
$ ls -l kali-dev/var/lib/apt/lists/
total 80468
-rw-r--r-- 1 root root    41445 Jun 19 07:02 http.kali.org_kali_dists_kali-dev_InRelease
-rw-r--r-- 1 root root 82299122 Jun 19 07:01 http.kali.org_kali_dists_kali-dev_main_binary-amd64_Packages
-rw-r--r-- 1 root root    40562 Jun 19 11:54 http.kali.org_kali_dists_kali-dev_Release
-rw-r--r-- 1 root root      833 Jun 19 11:54 http.kali.org_kali_dists_kali-dev_Release.gpg
drwxr-xr-x 2 root root     4096 Jun 19 11:54 partial
So far so good. Next step would be to complete the sources.list with other components, then run apt update: APT will download the missing Index files. But if you're unlucky, that might fail:
$ sudo sed -i 's/main$/main contrib non-free non-free-firmware/' kali-dev/etc/apt/sources.list
$ cat kali-dev/etc/apt/sources.list
deb http://http.kali.org/kali kali-dev main contrib non-free non-free-firmware
$ sudo chroot kali-dev apt update
Hit:1 http://http.kali.org/kali kali-dev InRelease
Get:2 http://kali.download/kali kali-dev/contrib amd64 Packages [121 kB]
Get:4 http://mirror.sg.gs/kali kali-dev/non-free-firmware amd64 Packages [10.6 kB]
Get:3 http://mirror.freedif.org/kali kali-dev/non-free amd64 Packages [198 kB]
Err:3 http://mirror.freedif.org/kali kali-dev/non-free amd64 Packages
  File has unexpected size (10442 != 10584). Mirror sync in progress? [IP: 66.96.199.63 80]
  Hashes of expected file:
   - Filesize:10584 [weak]
   - SHA256:71a83d895f3488d8ebf63ccd3216923a7196f06f088461f8770cee3645376abb
   - SHA1:c4ff126b151f5150d6a8464bc6ed3c768627a197 [weak]
   - MD5Sum:a49f46a85febb275346c51ba0aa8c110 [weak]
  Release file created at: Fri, 23 May 2025 06:48:41 +0000
Fetched 336 kB in 4s (77.5 kB/s)  
Reading package lists... Done
E: Failed to fetch http://mirror.freedif.org/kali/dists/kali-dev/non-free/binary-amd64/Packages.gz  File has unexpected size (10442 != 10584). Mirror sync in progress? [IP: 66.96.199.63 80]
   Hashes of expected file:
    - Filesize:10584 [weak]
    - SHA256:71a83d895f3488d8ebf63ccd3216923a7196f06f088461f8770cee3645376abb
    - SHA1:c4ff126b151f5150d6a8464bc6ed3c768627a197 [weak]
    - MD5Sum:a49f46a85febb275346c51ba0aa8c110 [weak]
   Release file created at: Fri, 23 May 2025 06:48:41 +0000
E: Some index files failed to download. They have been ignored, or old ones used instead.
What happened here? Again, we need APT debugging options to have a hint:
$ sudo chroot kali-dev apt -q -o Debug::Acquire::http=true update 2>&1   grep -e ^Answer -e ^HTTP
Answer for: http://http.kali.org/kali/dists/kali-dev/InRelease
HTTP/1.1 304 Not Modified
Answer for: http://http.kali.org/kali/dists/kali-dev/contrib/binary-amd64/Packages.gz
HTTP/1.1 302 Found
Answer for: http://http.kali.org/kali/dists/kali-dev/non-free/binary-amd64/Packages.gz
HTTP/1.1 302 Found
Answer for: http://http.kali.org/kali/dists/kali-dev/non-free-firmware/binary-amd64/Packages.gz
HTTP/1.1 302 Found
Answer for: http://kali.download/kali/dists/kali-dev/contrib/binary-amd64/Packages.gz
HTTP/1.1 200 OK
Answer for: http://mirror.sg.gs/kali/dists/kali-dev/non-free-firmware/binary-amd64/Packages.gz
HTTP/1.1 200 OK
Answer for: http://mirror.freedif.org/kali/dists/kali-dev/non-free/binary-amd64/Packages.gz
HTTP/1.1 200 OK
As we can see above, for the Release file we get a 304 (aka. "Not Modified") from the redirector. Why is that? This is due to If-Modified-Since also known as RFC-7232. APT supports this feature when it retrieves the Release file, it basically says to the server "Give me the Release file, but only if it's newer than what I already have". If the file on the server is not newer than that, it answers with a 304, which basically says to the client "You have the latest version already". So APT doesn't get a new Release file, it uses the Release file that is already present locally in /var/lib/apt/lists/, and then it proceeeds to download the missing Index files. And as we can see above: it then hits the redirector for each requests, and might be redirected to different mirrors for each Index file. So the important bit here is: the APT "trick" of downloading all the Index files from the same mirror only works if the Release file is served via a redirect. If it's not, like in this case, then APT hits the redirector for each files it needs to download, and it's subject to the "Hash Sum Mismatch" error again. In practice, for the casual user running apt update every now and then, it's not an issue. If they have the latest Release file, no extra requests are done, because they also have the latest Index files, from a previous apt update transaction. So APT doesn't re-download those Index files. The only reason why they'd have the latest Release file, and would miss some Index files, would be that they added new components to their APT sources, like we just did above. Not so common, and then they'd need to run apt update at a unlucky moment. I don't think many users are affected in practice. Note that this issue is rather new for Kali Linux. The redirector running on http.kali.org is mirrorbits, and support for If-Modified-Since just landed in the latest release, version 0.6. This feature was added by no one else than me, a great example of the expression "shooting oneself in the foot". An obvious workaround here is to empty /var/lib/apt/lists/ in the chroot after debootstrap completed. Or we could disable support for If-Modified-Since entirely for Kali's instance of mirrorbits. Summary and Conclusion The Hash Sum Mismatch failures above are caused by a combination of things: At the same time: All in all, it seems that all those issues would go away if only Acquire-By-Hash was supported in the Kali packages repository. Now is not a bad moment to try to land this feature in reprepro. After development halted in 2019, there's now a new upstream, and patches are being merged again. But it won't be easy: reprepro is a C codebase of around 50k lines of code, and it will take time and effort for the newcomer to get acquainted with the codebase, to the point of being able to implement a significant feature like this one. As an alternative, aptly is another popular tool to manage APT package repositories. And it seems to support Acquire-By-Hash already. Another alternative: I was told that debusine has (experimental) support for package repositories, and that Acquire-By-Hash is supported as well. Options are on the table, and I hope that Kali will eventually get support for Acquire-By-Hash, one way or another. To finish, due credits: this blog post exists thanks to my employer OffSec. Thanks for reading!

16 July 2025

Sven Hoexter: Windows 10 to Ubuntu Migration

I know that Canonical / Ubuntu people are sometimes not well received due to promotion of Canonical tooling (some might remember upstart and Mir, or more recently snap and netplan). Thus for some positive vibes consider that I could hand out the Ubuntu Desktop image on a USB flash drive to a family member, and the family member could just replace Windows 10 without any assistance. It just worked. This was made possible by the will to keep a slightly dated ThinkPad in use, which it's not supported by Windows 11. I've to admit that I never looked at Ubuntu Desktop before, but the user experience is on par with everything else I know. Thanks to all the folks at Canonical who made that possible! Luckily the times when you had to fiddle with modelines for XFree86, and sleepless nights about configuring lpd to get printing up and running are long gone. I believe now that Microsoft is doing Microsoft things with rolling Windows updates which force users to replace completely fine working hardware is the time to encourage more people to move to open operating systems, and Ubuntu Desktop seems to be a very suitable choice. Things to Improve Albeit I think the out of the box experience is great, there are a few niche topics where things could improve. Default Access to Apt / Ubuntu Universe Well snaps are promoted as the primary application source, but having some graphical interface like synaptic available by default to just install from Ubuntu Universe would be helpful. In this case we wanted to install keepass2 to access the users keepass file kept from the Windows setup. Having to tell someone "open the terminal and type sudo apt install" is something that requires support. Snaps and Isolation Overrides I'm fine with snaps having least privileges, but it would be nice if one could add overrides easily. Here the family member was playing with an Arduino Uno and there is one sample in the workbook that utilizes a Java application called Processing. It's available as a snap, but that one doesn't have access to the required serial port device file. I tried hard to make it work - full details in the snapcraft forum - but failed, and opted to use the --devmode to install it without isolation enforcement. As far as I understood snap that results in no more automatic updates for the application. If someone from the Ubuntu crowd with more snap debug experience has additional hints on how to narrow down which change is required, I would love to improve that and create a PR for the processing developers. Either reply in the forum or reach out via mail sven at stormbind dot net.

Sven Hoexter: Google Cloud Oddities Summer 2025 Edition

Latest oddities I ran into with Google Cloud products before I start to forget about them again. e2 Compute Instances vs CloudNAT Years ago I already had a surprising encounter with the Google Cloud e2 instances. Back then we observed CPU steal time from 20-60%, which made the instances unusable for anything remotely latency sensitive. Now someone started to run a workload which creates many outbound connections to the same destination IP:Port. To connect to the internet we utilize the Google Cloud product "CloudNAT" which implements a NAT solution somewhere in the network layer. Starting the workload let after a few seconds to all sort of connection issues, and of course logs from CloudNAT that it dropped connections. The simplest reproducer I could find was while true; do curl http://sven.stormbind.net; done which already let to connection drops on CloudNAT. We starred a bit at output of gcloud compute routers get-nat-mapping-info our-default-natgw, but allocating additional ports still worked fine in general. Further investigation let to two differences between a project which was fine and those that failed:
  1. c2d or n2d machine types instead of e2 and
  2. usage of gVNIC.
Moving away from the e2 instances instantly fixed our issue. Only some connection drops could be observed on CloudNAT if we set the min_ports_per_vm value too low and it could not allocate new ports in time. Thus we did some additional optimizations: See also upstream documentation regarding timeouts. To complete the setup alignment we also enabled gVNIC on all GKE pools. Noteworthy detail a colleague figured out: If you use terraform to provision GKE pools make sure to use at least google provider v6.33.0 to avoid a re-creation of your node pool. GKE LoadBalancer Force allPorts: true on Forwarding Rule Technically it's possible to configure a forwarding rule to listen on some or all ports. That gets more complicated if you do not configure the forwarding rule via terraform or gcloud cli, but use a GKE resource kind: Service with spec.type: LoadBalancer. The logic documented by Google Cloud is that the forwarding rule will have per port configuration if it's five or less, and above that it will open for all ports. Sadly that does not work e.g. in cases where you've an internal load balancer and a serviceAttachment attached to the forwarding rule. In my experience reconfiguring was also unreliable in cases without a serviceAttachment and required a manual deletion of the service load balancer to have the operator reconcile it and create it correctly. Given that we wanted to have all ports open to allow us to dynamically add more ports on a specific load balancer, but there is no annotation for that, I worked around with this beauty:
      ports:
        - name: dummy-0
          port: 2342
          protocol: TCP
          targetPort: 2342
        - name: dummy-1
          port: 2343
          protocol: TCP
          targetPort: 2343
        - name: dummy-2
          port: 2344
          protocol: TCP
          targetPort: 2344
        - name: dummy-3
          port: 2345
          protocol: TCP
          targetPort: 2345
        - name: service-1
          port: 4242
          protocol: TCP
          targetPort: 4242
        - name: service-2
          port: 4343
          protocol: TCP
          targetPort: 4343
If something in that area did not work out there's basically two things to check:
  1. Is the port open on the forwarding rule / is the forwarding rule configured with allPorts: true?
  2. Got the VPC firewall rule created by the service operator in GKE updated to open all required ports?
Rate Limiting with Cloud Armor on Global TCP Proxy Load Balancer According to the Google Cloud support rate limiting on a TCP proxy is a preview feature. That seems to be the excuse why it's all very inconsistent right now, but it works. Terraform Sample Snippet
  rule  
    description = "L3-L4 Rate Limit"
    action      = "rate_based_ban"
    priority    = "2342"
    match  
      versioned_expr = "SRC_IPS_V1"
      config  
        src_ip_ranges = ["*"]
       
     
    rate_limit_options  
      enforce_on_key = "IP"
      # exceed_action only supports deny() with a response code
      exceed_action = "deny(429)"
      rate_limit_threshold  
        count        = 320
        interval_sec = 60
       
      ban_duration_sec = 240
      ban_threshold  
        count        = 320
        interval_sec = 60
       
      conform_action = "allow"
     
   
Config-Connector Sample Snippet
  - action: rate_based_ban
    description: L3-L4 Rate Limit
    match:
      config:
        srcIpRanges:
          - "*"
      versionedExpr: SRC_IPS_V1
    preview: false
    priority: 2342
    rateLimitOptions:
      banDurationSec: 240
      banThreshold:
        count: 320
        intervalSec: 60
      conformAction: allow
      enforceOnKey: IP
      exceedAction: deny
      rateLimitThreshold:
         count: 320
         intervalSec: 60

15 July 2025

Alberto Garc a: Converting QEMU qcow2 images directly to stdout

Introduction Some months ago, my colleague Madeeha Javed and I wrote a tool to convert QEMU disk images into qcow2, writing the result directly to stdout. This tool is called qcow2-to-stdout.py and can be used for example to create a new image and pipe it through gzip and/or send it directly over the network without having to write it to disk first. This program is included in the QEMU repository: https://github.com/qemu/qemu/blob/master/scripts/qcow2-to-stdout.py If you simply want to use it then all you need to do is have a look at these examples: $ qcow2-to-stdout.py source.raw > dest.qcow2
$ qcow2-to-stdout.py -f dmg source.dmg gzip > dest.qcow2.gz If you re interested in the technical details, read on.

A closer look under the hood QEMU uses disk images to store the contents of the VM s hard drive. Images are often in qcow2, QEMU s native format, although a variety of other formats and protocols are also supported. I have written in detail about the qcow2 format in the past (for example, here and here), but the general idea is very easy to understand: the virtual drive is divided into clusters of a certain size (64 KB by default), and only the clusters containing non-zero data need to be physically present in the qcow2 image. So what we have is essentially a collection of data clusters and a set of tables that map guest clusters (what the VM sees) to host clusters (what the qcow2 file actually stores).
A qcow2 file is a collection of data clusters plus some metadata to map them to what the guest VM sees.
qemu-img is a powerful and versatile tool that can be used to create, modify and convert disk images. It has many different options, but one question that sometimes arises is whether it can use stdin or stdout instead of regular files when converting images. The short answer is that this is not possible in general. qemu-img convert works by checking the (virtual) size of the source image, creating a destination image of that same size and finally copying all the data from start to finish. Reading a qcow2 image from stdin doesn t work because data and metadata blocks can come in any arbitrary order, so it s perfectly possible that the information that we need in order to start writing the destination image is at the end of the input data . Writing a qcow2 image to stdout doesn t work either because we need to know in advance the complete list of clusters from the source image that contain non-zero data (this is essential because it affects the destination file s metadata). However, if we do have that information then writing a new image directly to stdout is technically possible. The bad news is that qemu-img won t help us here: it uses the same I/O code as the rest of QEMU. This generic approach makes total sense because it s simple, versatile and is valid for any kind of source and destination image that QEMU supports. However, it needs random access to both images. If we want to write a qcow2 file directly to stdout we need new code written specifically for this purpose, and since it cannot reuse the logic present in the QEMU code this was written as a separate tool (a Python script). The process itself goes like this:
  • Read the source image from start to finish in order to determine which clusters contain non-zero data. These are the only clusters that need to be present in the new image.
  • Write to stdout all the metadata structures of the new image. This is now possible because after the previous step we know how much data we have and where it is located.
  • Read the source image again and copy the clusters with non-zero data to stdout.
Images created with this program always have the same layout: header, refcount tables and blocks, L1 and L2 tables, and finally all data clusters. One problem here is that, while QEMU can read many different image formats, qcow2-to-stdout.py is an independent tool that does not share any of the code and therefore can only read raw files. The solution here is to use qemu-storage-daemon. This program is part of QEMU and it can use FUSE to export any file that QEMU can read as a raw file. The usage of qemu-storage-daemon is handled automatically and the user only needs to specify the format of the source file: $ qcow2-to-stdout.py -f dmg source.dmg > dest.qcow2 qcow2-to-stdout.py can only create basic qcow2 files and does not support features like compression or encryption. However, a few parameters can be adjusted, like the cluster size (-c), the width of the reference count entries (-r) and whether the new image is created with the input as an external data file (-d and -R). And this is all, I hope that you find this tool useful and this post informative. Enjoy!

Acknowledgments This work has been developed by Igalia and sponsored by Outscale, a Dassault Syst mes brand.
Logos of Igalia and Outscale

This problem would not happen if the input data was in raw format but in this case we would not know the size in advance.

Valhalla's Things: Federated instant messaging, 100% debianized

Posted on July 15, 2025
Tags: madeof:bits, topic:xmpp, topic:debian
This is an approximation of what I told at my talk Federated instant messaging, 100% debianized at DebConf 25, for people who prefer reading text. There will also be a video recording, as soon as it s ready :) at the link above. Communicating is a basic human need, and today some kind of computer-mediated communication is a requirement for most people, especially those in this room. With everything that is happening, it s now more important than ever that these means of communication aren t controlled by entities that can t be trusted, whether because they can stop providing the service at any given time or worse because they are going to abuse it in order to extract more profit. If only there was a well established chat system based on some standard developed in an open way, with all of the features one expects from a chat system but federated so that one can choose between many different and independent providers, or even self-hosting. But wait, it does exist! I m not talking about IRC, I m talking about XMPP! While it has been around since the last millennium, it has not remained still, with hundred of XMPP Extension Protocols, or XEPs that have been developed to add all of the features that nobody in 1999 imagined we could need in Instant Messaging today, and more, such as IoT devices or even social networks. There is a myth that this makes XMPP a mess of incompatible software, but there is an XEP for that: XEP-0479: XMPP Compliance Suites 2023, which is a list of XEPs that needs to be supported by Instant Messaging servers and clients, including mobile ones, and all of the recommended ones will mostly just work. These include conversations.im on android, dino on linux, which also works pretty nicely on linux phones, gajim for a more fully featured option that includes the kitchen sink, profanity for text interface fanatics like me, and I ve heard that monal works decently enough on the iThings. One thing that sets XMPP apart from other federated protocols, is that it has already gone through the phase where everybody was on one very big server, which then cut out federation, and we ve learned from the experience. These days there are still a few places that cater to newcomers, like https://account.conversations.im/, https://snikket.org/ (which also includes tools to make it easier to host your own instance) and https://quicksy.im/, but most people are actually on servers of a manageable size. My strong recommendation is for community hosting: not just self-hosting for yourself, but finding a community you feel part of and trust, and share a server with them, whether managed by volunteers from the community itself, or by a paid provider. If you are a Debian Developer, you already have one: you can go to https://db.debian.org/ , select Change rtc password to set your own password, wait an hour or so and you re good to go, as described at the bottom of https://wiki.debian.org/Teams/DebianSocial. A few years ago it had remained a bit behind, but these days it s managed by an active team, and if you re missing some features, or just want to know what s happening with it, you can join their BoF on Friday afternoon (and also thank them for their work). But for most people in this room, I d also recommend finding a friend or two who can help as a backup, and run a server for your own families or community: as a certified lazy person who doesn t like doing sysadmin jobs, I can guarantee it s perfectly feasible, about in the same range of difficulty as running your own web server for a static site. The two most popular servers for this, prosody and ejabberd, are well maintained in Debian, and these days there isn t a lot more to do than installing them, telling them your hostname, setting up a few DNS entries, and then you mostly need to keep the machine updated and very little else. After that, it s just applying system security updates, upgrading everything every couple years (some configuration updates may be needed, but nothing major) and maybe helping some non-technical users, if you are hosting your non-technical friends (the kind who would need support on any other platform).
Question time (including IRC questions) included which server would be recommended for very few users (I use prosody and I m very happy with it, but I believe ejabberd works also just fine), then somebody reminded me that I had forgotten to mention https://www.chatons.org/ , which lists free, ethical and decentralized services, including xmpp ones. I was also asked a comparison with matrix, which does cover a very similar target as XMPP, but I am quite biased against it, and I d prefer to talk well of my favourite platform than badly of its competitor.

12 July 2025

Bits from Debian: Debconf25 welcomes its sponsors

DebConf25 logo DebConf25, the 26th edition of the Debian conference is taking place in Brest Campus of IMT Atlantique Bretagne-Pays de la Loire, France. We appreciate the organizers for their hard work, and hope this event will be highly beneficial for those who attend in person as well as online. This event would not be possible without the help from our generous sponsors. We would like to warmly welcome the sponsors of DebConf 25, and introduce them to you. We have five Platinum sponsors. Our Gold sponsors are: Our Silver sponsors are: Bronze sponsors: And finally, our Supporter level sponsors: A special thanks to the IMT Atlantique Bretagne-Pays de la Loire, our Venue Partner and our Network Partner ResEl! Thanks to all our sponsors for their support! Their contributions enable a diverse global community of Debian developers and maintainers to collaborate, support one another, and share knowledge at DebConf25.

Reproducible Builds: Reproducible Builds in June 2025

Welcome to the 6th report from the Reproducible Builds project in 2025. Our monthly reports outline what we ve been up to over the past month, and highlight items of news from elsewhere in the increasingly-important area of software supply-chain security. If you are interested in contributing to the Reproducible Builds project, please see the Contribute page on our website. In this report:
  1. Reproducible Builds at FOSSY 2025
  2. Distribution work
  3. diffoscope
  4. OSS Rebuild updates
  5. Website updates
  6. Upstream patches
  7. Reproducibility testing framework

Reproducible Builds at FOSSY 2025 On Saturday 2nd August, Vagrant Cascadian and Chris Lamb will be presenting at this year s FOSSY 2025. Their talk, titled Never Mind the Checkboxes, Here s Reproducible Builds!, is being introduced as follows:
There are numerous policy compliance and regulatory processes being developed that target software development but do they solve actual problems? Does it improve the quality of software? Do Software Bill of Materials (SBOMs) actually give you the information necessary to verify how a given software artifact was built? What is the goal of all these compliance checklists anyways or more importantly, what should the goals be? If a software object is signed, who should be trusted to sign it, and can they be trusted forever?
The talk will introduce the audience to Reproducible Builds as a set of best practices which allow users and developers to verify that software artifacts were built from the source code, but also allows auditing for license compliance, providing security benefits, and removes the need to trust arbitrary software vendors. Hosted by the Software Freedom Conservancy and taking place in Portland, Oregon, USA, FOSSY aims to be a community-focused event: Whether you are a long time contributing member of a free software project, a recent graduate of a coding bootcamp or university, or just have an interest in the possibilities that free and open source software bring, FOSSY will have something for you . More information on the event is available on the FOSSY 2025 website, including the full programme schedule. Vagrant and Chris will also be staffing a table this year, where they will be available to answer any questions about Reproducible Builds and discuss collaborations with other projects.

Distribution work In Debian this month:
  • Holger Levsen has discovered that it is now possible to bootstrap a minimal Debian trixie using 100% reproducible packages. This result can itself be reproduced, using the debian-repro-status tool and mmdebstrap s support for hooks:
      $ mmdebstrap --variant=apt --include=debian-repro-status \
           --chrooted-customize-hook=debian-repro-status \
           trixie /dev/null 2>&1   grep "Your system has"
       INFO  debian-repro-status > Your system has 100.00% been reproduced.
    
  • On our mailing list this month, Helmut Grohne wrote an extensive message raising an issue related to Uploads with conflicting buildinfo filenames:
    Having several .buildinfo files for the same architecture is something that we plausibly want to have eventually. Imagine running two sets of buildds and assembling a single upload containing buildinfo files from both buildds in the same upload. In a similar vein, as a developer I may want to supply several .buildinfo files with my source upload (e.g. for multiple architectures). Doing any of this is incompatible with current incoming processing and with reprepro.
  • 5 reviews of Debian packages were added, 4 were updated and 8 were removed this month adding to our ever-growing knowledge about identified issues.

In GNU Guix, Timothee Mathieu reported that a long-standing issue with reproducibility of shell containers across different host operating systems has been solved. In their message, Timothee mentions:
I discovered that pytorch (and maybe other dependencies) has a reproducibility problem of order 1e-5 when on AVX512 compared to AVX2. I first tried to solve the problem by disabling AVX512 at the level of pytorch, but it did not work. The dev of pytorch said that it may be because some components dispatch computation to MKL-DNN, I tried to disable AVX512 on MKL, and still the results were not reproducible, I also tried to deactivate in openmpi without success. I finally concluded that there was a problem with AVX512 somewhere in the dependencies graph but I gave up identifying where, as this seems very complicated.

The IzzyOnDroid Android APK repository made more progress in June. Not only have they just passed 48% reproducibility coverage, Ben started making their reproducible builds more visible, by offering rbtlog shields, a kind of badge that has been quickly picked up by many developers who are proud to present their applications reproducibility status.
Lastly, in openSUSE news, Bernhard M. Wiedemann posted another monthly update for their work there.

diffoscope diffoscope is our in-depth and content-aware diff utility that can locate and diagnose reproducibility issues. This month, Chris Lamb made the following changes, including preparing and uploading versions 298, 299 and 300 to Debian:
  • Add python3-defusedxml to the Build-Depends in order to include it in the Docker image. [ ]
  • Handle the RPM format s HEADERSIGNATURES and HEADERIMMUTABLE as a special-case to avoid unnecessarily large diffs. Thanks to Daniel Duan for the report and suggestion. [ ][ ]
  • Update copyright years. [ ]
In addition, @puer-robustus fixed a regression introduced in an earlier commit which resulted in some differences being lost. [ ][ ] Lastly, Vagrant Cascadian updated diffoscope in GNU Guix to version 299 [ ][ ] and 300 [ ][ ].

OSS Rebuild updates OSS Rebuild has added a new network analyzer that provides transparent HTTP(S) interception during builds, capturing all network traffic to monitor external dependencies and identify suspicious behavior, even in unmodified maintainer-controlled build processes. The text-based user interface now features automated failure clustering that can group similar rebuild failures and provides natural language failure summaries, making it easier to identify and understand patterns across large numbers of build failures. OSS Rebuild has also improved the local development experience with a unified interface for build execution strategies, allowing for more extensible environment setup for build execution. The team also designed a new website and logo.

Website updates Once again, there were a number of improvements made to our website this month including:
  • Arnaud Brousseau added Stage , a new Linux distribution, to our Tools page.
  • Chris Lamb improved the docker instructions on the diffoscope website. [ ]


Upstream patches The Reproducible Builds project detects, dissects and attempts to fix as many currently-unreproducible packages as possible. We endeavour to send all of our patches upstream where appropriate. This month, we wrote a large number of such patches, including:

Reproducibility testing framework The Reproducible Builds project operates a comprehensive testing framework running primarily at tests.reproducible-builds.org in order to check packages and other artifacts for reproducibility. In June, however, a number of changes were made by Holger Levsen, including:
  • reproduce.debian.net-related:
    • Installed and deployed rebuilderd version 0.24 from Debian unstable in order to make use of the new compression feature added by Jarl Gullberg for the database. This resulted in massive decrease of the SQLite databases:
      • 79G 2.8G (all)
      • 84G 3.2G (amd64)
      • 75G 2.9G (arm64)
      • 45G 2.1G (armel)
      • 48G 2.2G (armhf)
      • 73G 2.8G (i386)
      • 72G 2.7G (ppc64el)
      • 45G 2.1G (riscv64)
      for a combined saving from 521G 20.8G. This naturally reduces the requirements to run an independent rebuilderd instance and will permit us to add more Debian suites as well.
    • During migration to the latest version of rebuilderd, make sure several services are not started. [ ]
    • Actually run rebuilderd from /usr/bin. [ ]
    • Raise temperatures for NVME devices on some riscv64 nodes that should be ignored. [ ][ ]
    • Use a 64KB kernel page size on the ppc64el architecture (see #1106757). [ ]
    • Improve ordering of some failed to reproduce statistics. [ ]
    • Detect a number of potential causes of build failures within the statistics. [ ][ ]
    • Add support for manually scheduling for the any architecture. [ ]
  • Misc:
    • Update the Codethink nodes as there are now many kernels installed. [ ][ ]
    • Install linux-sysctl-defaults on Debian trixie systems as we need ping functionality. [ ]
    • Limit the fs.nr_open kernel turnable. [ ]
    • Stop submitting results to deprecated buildinfo.debian.net service. [ ][ ]
In addition, Jochen Sprickerhof greatly improved the statistics and the logging functionality, including adopting to the new database format of rebuilderd version 0.24.0 [ ] and temporarily increasing maximum log size in order to debug a nettlesome build [ ]. Jochen also dropped the CPUSchedulingPolicy=idle systemd flag on the workers. [ ]

Finally, if you are interested in contributing to the Reproducible Builds project, please visit our Contribute page on our website. However, you can get in touch with us via:

Freexian Collaborators: Monthly report about Debian Long Term Support, June 2025 (by Roberto C. S nchez)

Like each month, have a look at the work funded by Freexian s Debian LTS offering.

Debian LTS contributors In June, 20 contributors have been paid to work on Debian LTS, their reports are available:
  • Abhijith PA did 14.0h (out of 14.0h assigned).
  • Adrian Bunk did 23.5h (out of 23.5h assigned).
  • Andreas Henriksson did 3.0h (out of 3.0h assigned and 17.0h from previous period), thus carrying over 17.0h to the next month.
  • Andrej Shadura did 2.0h (out of 3.0h assigned and 7.0h from previous period), thus carrying over 8.0h to the next month.
  • Bastien Roucari s did 20.0h (out of 20.0h assigned).
  • Ben Hutchings did 8.0h (out of 7.5h assigned and 16.0h from previous period), thus carrying over 15.5h to the next month.
  • Carlos Henrique Lima Melara did 12.0h (out of 12.0h assigned).
  • Chris Lamb did 18.0h (out of 18.0h assigned).
  • Daniel Leidert did 22.0h (out of 22.5h assigned and 1.0h from previous period), thus carrying over 1.5h to the next month.
  • Emilio Pozuelo Monfort did 23.5h (out of 16.75h assigned and 6.75h from previous period).
  • Guilhem Moulin did 14.0h (out of 11.5h assigned and 3.5h from previous period), thus carrying over 1.0h to the next month.
  • Jochen Sprickerhof did 21.0h (out of 0.5h assigned and 22.75h from previous period), thus carrying over 2.25h to the next month.
  • Lucas Kanashiro did 20.0h (out of 20.0h assigned).
  • Markus Koschany did 23.25h (out of 17.0h assigned and 6.25h from previous period).
  • Roberto C. S nchez did 21.25h (out of 20.75h assigned and 3.25h from previous period), thus carrying over 2.75h to the next month.
  • Santiago Ruano Rinc n did 12.75h (out of 15.0h assigned), thus carrying over 2.25h to the next month.
  • Sean Whitton did 1.0h (out of 4.25h assigned and 1.75h from previous period), thus carrying over 5.0h to the next month.
  • Sylvain Beucler did 23.5h (out of 23.5h assigned).
  • Thorsten Alteholz did 15.0h (out of 15.0h assigned).
  • Tobias Frost did 2.5h (out of 12.0h assigned), thus carrying over 9.5h to the next month.

Evolution of the situation In June, we released 35 DLAs.
  • Notable security updates:
    • mariadb-10.5, prepared by Otto Kek l inen, fixes vulnerabilities which could result in denial of service, information disclosure, or unauthorized data modification
    • python-django, prepared by Chris Lamb, fixes vulnerabilities which would result in log injection or denial of service
    • webkit2gtk, prepared by Emilio Pozuelo Monfort, fixes many vulnerabilities which could results in a wide range of issues
    • xorg-server, prepared by Emilio Pozuelo Monfort, fixes multiple vulnerabilities which may result in privilege escalation
    • sudo, prepared by Thorsten Alteholz, fixes a vulnerability which could result in privilege escalation
  • Notable non-security updates:
    • debian-security-support, prepared by Santiago Ruano Rinc n, updates status of packages which receive limited security support or which have reached the end of security support
    • dns-root-data, prepared by Sylvain Beucler, updates the DNSSEC trust anchors
This month s contributions from outside the regular team include the mariadb-10.5 update mentioned above, prepared by Otto Kek l inen (the package maintainer); an update to libfile-find-rule-perl, prepared by Salvatore Bonaccorso (a member of the Debian Security Team); an update to activemq, prepared by Emmanuel Arias (a maintainer of the package). Additionally, LTS Team members contributed stable updates of the following packages:
  • curl, prepared by Carlos Henrique Lima Melara
  • python-tornado, prepared by Daniel Leidert
  • python-flask-cors, prepared by Daniel Leidert
  • common-vfs, prepared by Daniel Leidert
  • cjson, prepared by Adrian Bunk
  • icu, prepared by Adrian Bunk
  • node-tar-fs, prepared by Adrian Bunk
  • rar, prepared by Adrian Bunk
Something of particular noteworthiness is that LTS contributor Carlos Henrique Lima Melara discovered a regression in the upstream fix for CVE-2023-2753 in curl. The corrective action which he took included providing a patch to upstream, uploading a stable update of curl, and further updating the version of curl in LTS. DebConf, the annual Debian Conference, is coming up in July and, as is customary each year, the week preceding the conference will feature an event called DebCamp. The DebCamp week provides an opportunity for teams and other interested groups/individuals to meet together in person in the same venue as the conference itself, with the purpose of doing focused work, often called sprints . LTS coordinator Roberto C. S nchez has announced that the LTS Team is planning to hold a sprint primarily focused on the Debian security tracker and the associated tooling used by the LTS Team and the Debian Security Team.

Thanks to our sponsors Sponsors that joined recently are in bold.

10 July 2025

David Bremner: Hibernate on the pocket reform 4/n

Context

Log from (failed) platform test After some fun I got the serial console working and re-ran the platform test. After a bit of reading the serial console, I realized that rmmod dwc3 was causing more problems than it solved, in particularly reliable hard lockup on one of the CPUs. My revised test script is
set -x
echo platform >  /sys/power/pm_test
echo reboot > /sys/power/disk
sleep 2
rmmod mt76x2u
sleep 2
echo disk >  /sys/power/state
sleep 2
modprobe mt76x2u
The current problem seems to be pcie not resuming properly.
[   65.306842] usbcore: deregistering interface driver mt76x2u
[   65.343606] wlx000a5205eb2d: deauthenticating from 20:05:b7:00:2d:89 by local choice (Reason: 3=DEAUTH_LEAVING)
[   67.995239] PM: hibernation: hibernation entry
[   68.048103] Filesystems sync: 0.022 seconds
[   68.049005] Freezing user space processes
[   68.051075] Freezing user space processes completed (elapsed 0.001 seconds)
[   68.051760] OOM killer disabled.
[   68.052597] PM: hibernation: Basic memory bitmaps created
[   68.053108] PM: hibernation: Preallocating image memory
[   69.719040] PM: hibernation: Allocated 366708 pages for snapshot
[   69.719650] PM: hibernation: Allocated 1466832 kbytes in 1.66 seconds (883.63 MB/s)
[   69.720370] Freezing remaining freezable tasks
[   69.723558] Freezing remaining freezable tasks completed (elapsed 0.002 seconds)
[   69.728002] rk_gmac-dwmac fe1b0000.ethernet end0: Link is Down
[   69.992324] rockchip-dw-pcie a40c00000.pcie: Failed to receive PME_TO_Ack
[   69.993405] PM: hibernation: debug: Waiting for 5 seconds.
[   76.059484] rockchip-dw-pcie a40c00000.pcie: Phy link never came up
[   76.060043] rockchip-dw-pcie a40c00000.pcie: fail to resume
[   76.060546] rockchip-dw-pcie a40c00000.pcie: PM: dpm_run_callback(): genpd_restore_noirq returns -110
[   76.061363] rockchip-dw-pcie a40c00000.pcie: PM: failed to restore noirq: error -110
previous episode next episode

Russell Coker: Bad Product Comparisons and EVs

When companies design products a major concern seems to be what the reviewers will have to say about it. For any product of significant value the users are unable to perform any reasonable test before buying, for a casual user some problems may only be apparent after weeks of use so professional reviews are important to many people. The market apparently doesn t want reviews of the form here s a list of products that are quite similar and all do the job well, you can buy any of them, it s no big deal which would be the most technically accurate way of doing it. So the reviewers compare the products on the criteria that are easiest to measure, this lead to phones being compared by how light and thin they are. I think it s often the case that users would be better served by thicker heavier phones that have larger batteries but instead they are being sold phones that have good battery life in a fresh installation but which don t last a day with a full load of apps installed. The latest issue with bad reviews driving poor product design is electric cars. For a while the advocates of old fashioned cars have touted the range of petrol cars which has become an issue for comparing EVs. I have been driving cars for 35 years and so far I have never driven anywhere that s out of range of the current electric charging network, even with the range of the LEAF (which is smaller than many other EVs). If I ever felt the need to drive across the Nullarbor Plain then I could rent a car to do that and the costs of such car rental would be small compared to the money I m saving by driving an EV and also small when compared to the premium I would have to pay for an EV with a larger range. Some of the recent articles I ve seen about EVs have covered vehicles with a battery range over 700Km which is greater than the legal distance a commercial driver can drive without a break. I ve also seen articles about plans to have a small petrol or Diesel motor in an EV to recharge the battery without directly driving the wheels. A 9KW Diesel motor could provide enough electricity on average to keep the charge maintained in a LEAF battery and according to the specs of Diesel generators would take about 55Kg of fuel to provide the charge a LEAF needs to drive 1000Km. The idea of a mostly electric hybrid car that can do 1000Km on one tank of fuel is interesting as a thought experiment but doesn t seem to have much actual use. Apparently a Chinese company is planning to release a car that can do 1400Km one one tank of fuel using such technology which is impressive but not particularly useful. The next issue of unreasonable competition is in charge speed. Charging a car at 2KW from a regular power socket is a real limit to what you can do with a car. It s a limit that hasn t bothered me so far because the most driving I typically do in a week is less than one full charge, so at most I have to charge overnight twice in a week. But if I was going to drive to another city without hiring a car that has better range I d need a fast charger. Most current models of the Nissan LEAF support charging speeds up to 50KW which means fully charging the battery in under an hour (or slightly over an hour for the long range version). If I was to drive from Melbourne to Canberra in my LEAF I d have to charge twice which would be an annoyance at those speeds. There are a variety of EVs that can charge at 100KW and some as high as 350KW. 350KW is enough to fully charge the largest EV batteries in half an hour which seems to be as much as anyone would need. But there are apparently plans for 1MW car chargers which would theoretically be able to charge a Hummer (the EV with the largest battery) in 12 minutes. One obvious part of the solution to EV charging times is to not drive a Hummer! Another thing to note is that batteries can t be charged at a high rate for all charge levels, this is why advertising for fast chargers makes claims like 80% charge in half an hour which definitely doesn t mean 100% charge in 37.5 minutes ! There are significant engineering issues with high power applications. A 1MW cable is not just a bigger version of a regular power cable, there are additional safety issues, user training is required and cooling of the connector is probably required. That s a lot to just get a better number in the table at the end of a review. There is research in progress on the Megawatt Charging System which is designed to charge heavy vehicles (presumably trucks and buses) at up to 3.75MW. Charging a truck at that rate is reasonable as the process of obtaining and maintaining a heavy vehicle license requires a significant amount of effort and some extra training in 3.75MW charging probably doesn t make much difference. A final issue with fast charging is the capacity of the grid. A few years ago I attended a lecture by an electrical engineer who works for the Victorian railway system which was very interesting. The Vic rail power setup involved about 100MW of grid connectivity with special contracts with the grid operators due to the fact that 1MW trains suddenly starting and stopping causes engineering problems that aren t trivial to solve. They were also working on battery packs and super capacitors to deal with regenerative braking and to avoid brownouts in long sections of track. For a medium size petrol station 14 bays for fuelling cars is common. If 6 such petrol stations were replaced with fast charging stations that can charge cars at 1MW each that would draw the same power as the train network for the entire state! There is a need for significant engineering work to allow most cars to be electric no matter how it s done, but we don t need to make that worse just for benchmarks.

Tianon Gravi: Yubi Whati? (YubiKeys, ECDSA, and X.509)

Off-and-on over the last several weeks, I've been spending time trying to learn/understand YubiKeys better, especially from the perspective of ECDSA and signing. I had a good mental model for how "slots" work (canonically referenced by their hexadecimal names such as 9C), but found that it had a gap related to "objects"; while closing that, I was annoyed that the main reference table for this gap lives primarily in either a PDF or inside several implementations, so I figured I should create the reference I want to see in the world, but that it would also be useful to write down some of my understanding for my own (and maybe others') future reference. So, to that end, I'm going to start with a bit ( ) of background information, with the heavy caveat that this only applies to "PIV" ("FIPS 201") usage of YubiKeys, and that I only actually care about ECDSA, although I've been reassured that it's the same for at least RSA (anything outside this is firmly Here Be Not Tianon; "gl hf dd"). (Incidentally, learning all this helped me actually appreciate the simplicity of cloud-based KMS solutions, which was an unexpected side effect. ) At a really high level, ECDSA is like many other (asymmetric) cryptographic solutions you've got a public key and a private key, the private key can be used to "sign" data (tiny amounts of data, in fact, like P-256 can only reasonably sign 256 bits of data, which is where cryptographic hashes like SHA256 come in as secure analogues for larger data in small bit sizes), and the public key can then be used to verify that the data was indeed signed by the private key, and only someone with the private key could've done so. There's some complex math and RNGs involved, but none of that's actually relevant to this post, so find that information elsewhere. Unfortunately, this is where things go off the rails: PIV is X.509 ("x509") heavy, and there's no X.509 in the na ve view of my use case. In a YubiKey (or any other PIV-signing-supporting smart card? do they actually have competitors in this specific niche? ), a given "slot" can hold one single private key. There are ~24 slots which can hold a private key and be used for signing, although "Slot 9c" is officially designated as the "Digital Signature" slot and is encouraged for signing purposes. One of the biggest gotchas is that with pure-PIV (and older YubiKey firmware ) the public key for a given slot is only available at the time the key is generated, and the whole point of the device in the first place is that the private key is never, ever available from it (all cryptographic operations happen inside the device), so if you don't save that public key when you first ask the device to generate a private key in a particular slot, the public key is lost forever (asterisk).
$ # generate a new ECDSA P-256 key in "slot 9c" ("Digital Signature")
$ # WARNING: THIS WILL GLEEFULLY WIPE SLOT 9C WITHOUT PROMPTING
$ yubico-piv-tool --slot 9c --algorithm ECCP256 --action generate
-----BEGIN PUBLIC KEY-----
MFkwEwYHKoZIzj0CAQYIKoZIzj0DAQcDQgAEtGoWRGyjjUlJFXpu8BL6Rnx8jjKR
5+Mzl2Vepgor+k7N9q7ppOtSMWefjFVR0SEPmXqXINNsCi6LpLtNEigIRg==
-----END PUBLIC KEY-----
Successfully generated a new private key.
$ # this is the only time/place we (officially) get this public key
With that background, now let's get to the second aspect of "slots" and how X.509 fits. For every aforementioned slot, there is a corresponding "object" (read: place to store arbitrary data) which is corresponding only by convention. For all these "key" slots the (again, by convention) corresponding "object" is explicitly supposed to be an X.509 certificate (see also the PDF reference linked above). It turns out this is a useful and topical place to store that public key we need to keep handy! It's also an interesting place to shove additional details about what the key in a given slot is being used for, if that's your thing. Converting the raw public key into a (likely self-signed) X.509 certificate is an exercise for the reader, but if you want to follow the conventions, you need some way to convert a given "slot" to the corresponding "object", and that is the lookup table I wish existed in more forms. So, without further ado, here is the anti-climax:
Slot Object Description
0x9A 0x5FC105 X.509 Certificate for PIV Authentication
0x9E 0x5FC101 X.509 Certificate for Card Authentication
0x9C 0x5FC10A X.509 Certificate for Digital Signature
0x9D 0x5FC10B X.509 Certificate for Key Management
0x82 0x5FC10D Retired X.509 Certificate for Key Management 1
0x83 0x5FC10E Retired X.509 Certificate for Key Management 2
0x84 0x5FC10F Retired X.509 Certificate for Key Management 3
0x85 0x5FC110 Retired X.509 Certificate for Key Management 4
0x86 0x5FC111 Retired X.509 Certificate for Key Management 5
0x87 0x5FC112 Retired X.509 Certificate for Key Management 6
0x88 0x5FC113 Retired X.509 Certificate for Key Management 7
0x89 0x5FC114 Retired X.509 Certificate for Key Management 8
0x8A 0x5FC115 Retired X.509 Certificate for Key Management 9
0x8B 0x5FC116 Retired X.509 Certificate for Key Management 10
0x8C 0x5FC117 Retired X.509 Certificate for Key Management 11
0x8D 0x5FC118 Retired X.509 Certificate for Key Management 12
0x8E 0x5FC119 Retired X.509 Certificate for Key Management 13
0x8F 0x5FC11A Retired X.509 Certificate for Key Management 14
0x90 0x5FC11B Retired X.509 Certificate for Key Management 15
0x91 0x5FC11C Retired X.509 Certificate for Key Management 16
0x92 0x5FC11D Retired X.509 Certificate for Key Management 17
0x93 0x5FC11E Retired X.509 Certificate for Key Management 18
0x94 0x5FC11F Retired X.509 Certificate for Key Management 19
0x95 0x5FC120 Retired X.509 Certificate for Key Management 20
See also "piv-objects.json" for a machine-readable copy of this data. (Major thanks to paultag and jon gzip johnson for helping me learn and generally putting up with me, but especially dealing with my live-stream-of-thoughts while I stumble through the dark. )

8 July 2025

Sahil Dhiman: Five Years of Writing - Now What?

Okay, here s the deal, I pushed my first post on Reimagined Doodle - Alias Command, five years ago on July 8th, 2020. Don t think I ever mentioned that post started out as a Github Gist which I later transferred here seeking a more long-term home on an independent platform. Writing about writings, motivations, and the blog itself has been a recurring theme here over the years. 1 2 3 4 5 6 7 8 9 I m unsure how I sustained expressing myself and writing here for this long. Now and then, I go months without any thought of writing, and then all of a sudden I start in bursts with sequential posts one after another. There isn t a pattern per se in topics other than whatever burning question I have at the moment. So here s to a milestone and then some.

7 July 2025

Birger Schacht: Debian on Framework 12

For some time now I was looking for a device to replace my Thinkpad. Its a 14" device, but thats to big for my taste. I am a big fan of small notebooks, so when frame.work announced their 12" laptop, I took the chance and ordered one right away. I was in one of the very early batches and got my package a couple of days ago. When ordering, I chose the DIY edition, but in the end there was not that much of DIY to do: I had to plug in the storage and the memory, put the keyboard in and tighten some screws. There are very detailed instructions with a lot of photos that tell you which part to put where, which is nice. Image of the Framework 12 laptop, assembled but powered off My first impressions of the device are good - it is heavier than I anticipated, but very vell made. It is very easy to assemble and disassemble and it feels like it can take a hit. When I started it the first time it took some minutes to boot because of the new memory module, but then it told me right away that it could not detect an operating system. As usual when I want to install a new system, I created a GRML live usb system and tried to boot from this USB device. But the Framwork BIOS did not want to let me boot GRML, telling me it is blocked by the current security policy. So I started to look in the BIOS where I could find the SecureBoot configuration, but there was no such setting anywhere. I then resorted to a Debian Live image, which was allowed to boot. Image of the screen of the Framework 12 laptop, saying it could not detect an operating system I only learned later, that the SecureBoot setting is in a separate section that is not part of the main BIOS configuration dialog. There is an Administer Secure Boot icon which you can choose when starting the device, but apparently only before you try to load an image that is not allowed. I always use my personal minimal install script to install my Debian systems, so it did not make that much of a difference to use Debian Live instead of GRML. I only had to apt install debootstrap before running the script. I updated the install script to default to trixie and to also install shim-signed and after successful installation booted into Debian 13 on the Framwork 12. Everthing seems to work fine so far. WIFI works. For sway to start I had to install firmware-intel-graphics. The touchscreen works without me having to configure anything (though I don t have frame.work stylus, as they are not yet available), also changing the brightness of the screen worked right away. The keyboard feels very nice, likewise the touchpad, which I configured to allow tap-to-click using the tap enabled option of sway-input. Image of the a Framework 12 laptop, showing the default Sway background image One small downside of the keyboard is that it does not have a backlight, which was a surprise. But given that this is a frame.work laptop, there are chances that a future generation of the keyboard will have backlight support. The screen of the laptop can be turned all the way around to the back of the laptops body, so it can be used as a tablet. In this mode the keyboard gets disabled to prevent accidently pushing keys when using the device in tablet mode. For online meetings I still prefer using headphones with cables over bluetooth once, so I m glad that the laptop has a headphone jack on the side. Above the screen there are a camera and a microphone, which both have separate physical switches to disable them. I ordered a couple of expansion cards, in the current setup I use two USB-C, one HDMI and one USB-A. I also ordered a 1TB expansion card and only used this to transfer my /home, but I soon realized that the card got rather hot, so I probably won t use it as a permanent expansion. I can not yet say a lot about how long the battery lasts, but I will bring the laptop to DebConf 25, I guess there I ll find out. There I might also have a chance to test if the screen is bright enough to be usable outdoors ;)

5 July 2025

Bits from Debian: Bits from the DPL

Dear Debian community, This is bits from the DPL for June. The Challenge of Mentoring Newcomers In June there was an extended discussion about the ongoing challenges around mentoring newcomers in Debian. As many of you know, this is a topic I ve cared about deeply--long before becoming DPL. In my view, the issue isn t just a matter of lacking tools or needing to try harder to attract contributors. Anyone who followed the discussion will likely agree that it s more complex than that. I sometimes wonder whether Debian s success contributes to the problem. From the outside, things may appear to just work , which can lead to the impression: Debian is doing fine without me--they clearly have everything under control. But that overlooks how much volunteer effort it takes to keep the project running smoothly. We should make it clearer that help is always needed--not only in packaging, but also in writing technical documentation, designing web pages, reaching out to upstreams about license issues, finding sponsors, or organising events. (Speaking from experience, I would have appreciated help in patiently explaining Free Software benefits to upstream authors.) Sometimes we think too narrowly about what newcomers can do, and also about which tasks could be offloaded from overcommitted contributors. In fact, one of the most valuable things a newcomer can contribute is better documentation. Those of us who ve been around for years may be too used to how things work--or make assumptions about what others already know. A person who just joined the project is often in the best position to document what s confusing, what s missing, and what they wish they had known sooner. In that sense, the recent "random new contributor s experience" posts might be a useful starting point for further reflection. I think we can learn a lot from positive user stories, like this recent experience of a newcomer adopting the courier package. I'm absolutely convinced that those who just found their way into Debian have valuable perspectives--and that we stand to learn the most from listening to them. We should also take seriously what Russ Allbery noted in the discussion: "This says bad things about the project's sustainability and I think everyone knows that." Volunteers move on--that s normal and expected. But it makes it all the more important that we put effort into keeping Debian's contributor base at least stable, if not growing. Project-wide LLM budget for helping people Lucas Nussbaum has volunteered to handle the paperwork and submit a request on Debian s behalf to LLM providers, aiming to secure project-wide access for Debian Developers. If successful, every DD will be free to use this access--or not--according to their own preferences. Kind regards Andreas.

Taavi V n nen: Tracking my train travel by parsing tickets in emails

Rumour has it that I might be a bit of a train nerd. At least I want to collect various nerdy data about my travels. Historically that data has lived in manual form in several places,1 but over the past year and a half I've been working on a toy project to collect most of that information into a custom tool. That toy project2 uses various sources to get information about trains to fill up its database: for example, in Finland Fintraffic, the organization responsible for railway traffic management, publishes very comprehensive open data about almost everything that's moving on the Finnish railway network. Unfortunately, I cannot be on all of the trains.3 Thus I need to tell the system details about my journeys. The obvious solution is to make a form that lets me save that data. Which I did, but I got very quickly bored of filling out that form, and as regular readers of this blog know, there is no reason to settle for a simple but boring solution when the alternative is to make something that is ridiculously overengineered.

Parsing data out of my train tickets Finnish long-distance trains generally require train-specific seat reservations, which means VR (the train company) knows which trains I am on. We just need to find a way to extract that information in some machine-readable format. So my plan for the ridiculously overengineered solution was to parse the booking emails to get the details I need. Now, VR ticket emails include the data I want in a couple of different formats: they're included as text in the HTML email body, they're in the embedded calendar invite, as text in the included PDF ticket, and encoded in the Aztec Code in the included PDF ticket. I chose to parse the last option with the hopes of building something that could be ported to parse other operators' tickets with relative ease.
Example Aztec code
Example Aztec code
After a bit of digging (thank you to the KDE Itinerary people for documenting this!) I stumbled upon an European Union Agency for Railways PDF titled ELECTRONIC SEAT/BERTH RESERVATION AND ELECTRONIC PRODUCTION OF TRANSPORT DOCUMENTS - TRANSPORT DOCUMENTS (RCT2 STANDARD) which, in its Appendix C.1, describes how the information is encoded in the code.4 (As a side note, various sources call these codes SSB version 1 codes, although that term isn't used in this specification. So maybe there are more specifications about the format that I haven't discovered yet!) I then wrote a parser in Go for the binary data embedded in these codes. So far it works, although I wouldn't be surprised if there are some edge cases that it doesn't handle. In particular, the spec specifies a custom lookup table for converting between text and binary data, and that only has support for characters 0-9 and A-Z. But Finnish railway station codes can also use and .. maybe I need to buy a ticket to a station with one of those.

Extracting barcodes out of emails A parser just for the binary format isn't enough here if the intended source input is the emails that VR sends upon making a booking. Time to write a single-purpose email server! In short, the logic in the server, again written in Go and with the help of go-smtp and go-message, is:
  • Accept any mail with a reasonable body size
  • Process through all body parts
  • For all PDF parts, extract all images
  • For all images, run them through ZXing
  • For all decoded barcodes, try to parse them with my new ticket parsing library I mentioned earlier
  • If any tickets are found, send the data from them and any metadata to the main backend, which will save them to a database
The custom mail server exposes an LMTP interface over TCP for my internet-facing mail servers to forward to. I chose LMTP for this because it seemed like a better fit in theory than normal (E)SMTP. I've since discovered that curl doesn't support LMTP which makes development much harder, and in practice there's no benefit of LMTP here as all mails are being sent to the backend in a single request regardless of the number of recipients, so maybe I'll migrate it to regular SMTP at some point.

Side quest time The last missing part is automatically forwarding the ticket mails to the new service. I've routed a dedicated subdomain to the new service, and the backend is configured to allocate addresses like i2v44g2pygkcth64stjgyuqz@somedomain.example for each user. That's great if we wanted to manually forward mails to the service, but we can go one step above that. I created a dedicated email alias in my mail server config that routes both to my regular mailbox and the service address. That way I can update my VR account to use the alias and have mails automatically processed while still receiving backup copies of the tickets (and any other important mail that VR might send me). Unfortunately that last part turns out something that's easier said than done. Logging in on the website, I'm greeted by this text stating I need to contact customer service by phone to change the address associated with my account.5 After a bit of digging, I noticed that the mobile app suggests filling out a feedback form in order to change the address. So I filled that, and after a day or two I got a "confirm you want to change your email" mail. Success!

  1. Including (but not limited to): a page of this website, the notes app on my phone, and an uMap map.
  2. Which I'm not directly naming here because I still think it needs a lot more work before being presentable, but if you're really interested it's not that hard to find out.
  3. Someone should invent human cloning so that we can fix this.
  4. People who know much more about railway ticketing than I do were surprised when I told them this format is still in use somewhere. So, uh, sorry if you were expecting a nice universal worldwide standard!
  5. In case you have not guessed yet, I do not like making phone calls.

4 July 2025

Sahil Dhiman: Secondary Authoritative Name Server Options for Self-Hosted Domains

In the past few months, I have moved authoritative name servers (NS) of two of my domains (sahilister.net and sahil.rocks) in house using PowerDNS. Subdomains of sahilister.net see roughly 320,000 hits/day across my IN and DE mirror nodes, so adding secondary name servers with good availability (in addition to my own) servers was one of my first priorities. I explored the following options for my secondary NS, which also didn t cost me anything:

1984 Hosting

Hurriance Electric

Afraid.org

Puck

NS-Global

Asking friends Two of my friends and fellow mirror hosts have their own authoritative name server setup, Shrirang (ie albony) and Luke. Shirang gave me another POP in IN and through Luke (who does have an insane amount of in-house NS, see dig ns jing.rocks +short), I added a JP POP. If we know each other, I would be glad to host a secondary NS for you in (IN and/or DE locations).

Some notes
  • Adding a third-party secondary is putting trust that the third party would serve your zone right.
  • Hurricane Electric and 1984 hosting provide multiple NS. One can use some or all of them. Ideally, you can get away with just using your own with full set from any of these two. Play around with adding and removing secondaries, which gives you the best results. . Using everyone is anyhow overkill, unless you have specific reasons for it.
  • Moving NS in-house isn t that hard. Though, be prepared to get it wrong a few times (and some more). I have already faced partial outages because:
    • Recursive resolvers (RR) in the wild behave in a weird way and cache the wrong NS response for longer time than in TTL.
    • NS expiry took more than time. 2 out of 3 of my Netim s NS (my domain registrar) had stopped serving my domain, while RRs in the wild hadn t picked up my new in-house NS. I couldn t really do anything about it, though.
    • Dot is pretty important at the end.
    • With HE.net, I forgot to delegate my domain on their panel and just added in my NS set, thinking I ve already done so (which I did but for another domain), leading to a lame server situation.
  • In terms of serving traffic, there s no distinction between primary and secondary NS. RR don t really care who they re asking the query to. So one can have hidden primary too.
  • I initially thought of adding periodic RIPE Atlas measurements from the global set but thought against it as I already host a termux mirror, which brings in thousands of queries from around the world leading to a diverse set of RRs querying my domain already.
  • In most cases, query resolution time would increase with out of zone NS servers (which most likely would be in external secondary). 1 query vs. 2 queries. Pay close attention to ADDITIONAL SECTION Shrirang s case followed by mine:
$ dig ns albony.in
; <<>> DiG 9.18.36 <<>> ns albony.in
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 60525
;; flags: qr rd ra; QUERY: 1, ANSWER: 4, AUTHORITY: 0, ADDITIONAL: 9
;; OPT PSEUDOSECTION:
; EDNS: version: 0, flags:; udp: 65494
;; QUESTION SECTION:
;albony.in.			IN	NS
;; ANSWER SECTION:
albony.in.		1049	IN	NS	ns3.albony.in.
albony.in.		1049	IN	NS	ns4.albony.in.
albony.in.		1049	IN	NS	ns2.albony.in.
albony.in.		1049	IN	NS	ns1.albony.in.
;; ADDITIONAL SECTION:
ns3.albony.in.		1049	IN	AAAA	2a14:3f87:f002:7::a
ns1.albony.in.		1049	IN	A	82.180.145.196
ns2.albony.in.		1049	IN	AAAA	2403:44c0:1:4::2
ns4.albony.in.		1049	IN	A	45.64.190.62
ns2.albony.in.		1049	IN	A	103.77.111.150
ns1.albony.in.		1049	IN	AAAA	2400:d321:2191:8363::1
ns3.albony.in.		1049	IN	A	45.90.187.14
ns4.albony.in.		1049	IN	AAAA	2402:c4c0:1:10::2
;; Query time: 29 msec
;; SERVER: 127.0.0.53#53(127.0.0.53) (UDP)
;; WHEN: Fri Jul 04 07:57:01 IST 2025
;; MSG SIZE  rcvd: 286
vs mine
$ dig ns sahil.rocks
; <<>> DiG 9.18.36 <<>> ns sahil.rocks
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 64497
;; flags: qr rd ra; QUERY: 1, ANSWER: 11, AUTHORITY: 0, ADDITIONAL: 1
;; OPT PSEUDOSECTION:
; EDNS: version: 0, flags:; udp: 65494
;; QUESTION SECTION:
;sahil.rocks.			IN	NS
;; ANSWER SECTION:
sahil.rocks.		6385	IN	NS	ns5.he.net.
sahil.rocks.		6385	IN	NS	puck.nether.net.
sahil.rocks.		6385	IN	NS	colin.sahilister.net.
sahil.rocks.		6385	IN	NS	marvin.sahilister.net.
sahil.rocks.		6385	IN	NS	ns2.afraid.org.
sahil.rocks.		6385	IN	NS	ns4.he.net.
sahil.rocks.		6385	IN	NS	ns2.albony.in.
sahil.rocks.		6385	IN	NS	ns3.jing.rocks.
sahil.rocks.		6385	IN	NS	ns0.1984.is.
sahil.rocks.		6385	IN	NS	ns1.1984.is.
sahil.rocks.		6385	IN	NS	ns-global.kjsl.com.
;; Query time: 24 msec
;; SERVER: 127.0.0.53#53(127.0.0.53) (UDP)
;; WHEN: Fri Jul 04 07:57:20 IST 2025
;; MSG SIZE  rcvd: 313
  • Theoretically speaking, a small increase/decrease in resolution would occur based on the chosen TLD and the popularity of the TLD in query originators area (already cached vs. fresh recursion).
  • One can get away with having only 3 NS (or be like Google and have 4 anycast NS or like Amazon and have 8 or like Verisign and make it 13 :P).
  • Nowhere it s written, your NS needs not to be called dns* or ns1, ns2 etc. Get creative with naming NS; be deceptive with the naming :D.
  • A good understanding of RR behavior can help engineer a good authoritative NS system.

Further reading

3 July 2025

Russell Coker: The Fuss About AI

There are many negative articles about AI (which is not about actual Artificial Intelligence also known as AGI ). Which I think are mostly overblown and often ridiculous. Resource Usage Complaints about resource usage are common, training Llama 3.1 could apparently produce as much pollution as 10,000 round trips by car between Los Angeles and New York City . That s not great but when you compare to the actual number of people doing such drives in the US and the number of people taking commercial flights on that route it doesn t seem like such a big deal. Apparently commercial passenger jets cause CO2 emissions per passenger about equal to a car with 2 people. Why is it relevant whether pollution comes from running servers, driving cars, or steel mills? Why not just tax polluters for the damage they do and let the market sort it out? People in the US make a big deal about not being communist, so why not have a capitalist solution, make it more expensive to do undesirable things and let the market sort it out? ML systems are a less bad use of compute resources than Bitcoin, at least ML systems give some useful results while Bitcoin has nothing good going for it. The Dot-Com Comparison People often complain about the apparent impossibility of AI companies doing what investors think they will do. But this isn t anything new, that all happened before with the dot com boom . I m not the first person to make this comparison, The Daily WTF (a high quality site about IT mistakes) has an interesting article making this comparison [1]. But my conclusions are quite different. The result of that was a lot of Internet companies going bankrupt, the investors in those companies losing money, and other companies then bought up their assets and made profitable companies. The cheap Internet we now have was built on the hardware from bankrupt companies which was sold for far less than the manufacture price. That allowed it to scale up from modem speeds to ADSL without the users paying enough to cover the purchase of the infrastructure. In the early 2000s I worked for two major Dutch ISPs that went bankrupt (not my fault) and one of them continued operations in the identical manner after having the stock price go to zero (I didn t get to witness what happened with the other one). As far as I m aware random Dutch citizens and residents didn t suffer from this and employees just got jobs elsewhere. There are good things being done with ML systems and when companies like OpenAI go bankrupt other companies will buy the hardware and do good things. NVidia isn t ever going to have the future sales that would justify a market capitalisation of almost 4 Trillion US dollars. This market cap can support paying for new research and purchasing rights to patented technology in a similar way to the high stock price of Google supported buying YouTube, DoubleClick, and Motorola Mobility which are the keys to Google s profits now. The Real Upsides of ML Until recently I worked for a company that used ML systems to analyse drivers for signs of fatigue, distraction, or other inappropriate things (smoking which is illegal in China, using a mobile phone, etc). That work was directly aimed at saving human lives with a significant secondary aim of saving wear on vehicles (in the mining industry drowsy drivers damage truck tires and that s a huge business expense). There are many applications of ML in medical research such as recognising cancer cells in tissue samples. There are many less important uses for ML systems, such as recognising different types of pastries to correctly bill bakery customers technology that was apparently repurposed for recognising cancer cells. The ability to recognise objects in photos is useful. It can be used for people who want to learn about random objects they see and could be used for helping young children learn about their environment. It also has some potential for assistance for visually impaired people, it wouldn t be good for safety critical systems (don t cross a road because a ML system says there are no cars coming) but could be useful for identifying objects (is this a lemon or a lime). The Humane AI pin had some real potential to do good things but there wasn t a suitable business model [2], I think that someone will develop similar technology in a useful way eventually. Even without trying to do what the Humane AI Pin attempted, there are many ways for ML based systems to assist phone and PC use. ML systems allow analysing large quantities of data and giving information that may be correct. When used by a human who knows how to recognise good answers this can be an efficient way of solving problems. I personally have solved many computer problems with the help of LLM systems while skipping over many results that were obviously wrong to me. I believe that any expert in any field that is covered in the LLM input data could find some benefits from getting suggestions from an LLM. It won t necessarily allow them to solve problems that they couldn t solve without it but it can provide them with a set of obviously wrong answers mixed in with some useful tips about where to look for the right answers. Jobs and Politics Noema Magazine has an insightful article about how AI can allow different models of work which can enlarge the middle class [3]. I don t think it s reasonable to expect ML systems to make as much impact on society as the industrial revolution, and the agricultural revolutions which took society from more than 90% farm workers to less than 5%. That doesn t mean everything will be fine but it is something that can seem OK after the changes have happened. I m not saying apart from the death and destruction everything will be good , the death and destruction are optional. Improvements in manufacturing and farming didn t have to involve poverty and death for many people, improvements to agriculture didn t have to involve overcrowding and death from disease. This was an issue of political decisions that were made. The Real Problems of ML Political decisions that are being made now have the aim of making the rich even richer and leaving more people in poverty and in many cases dying due to being unable to afford healthcare. The ML systems that aim to facilitate such things haven t been as successful as evil people have hoped but it will happen and we need appropriate legislation if we aren t going to have revolutions. There are documented cases of suicide being inspired by Chat GPT systems [4]. There have been people inspired towards murder by ChatGPT systems but AFAIK no-one has actually succeeded in such a crime yet. There are serious issues that need to be addressed with the technology and with legal constraints about how people may use it. It s interesting to consider the possible uses of ChatGPT systems for providing suggestions to a psychologist, maybe ChatGPT systems could be used to alleviate mental health problems. The cases of LLM systems being used for cheating on assignments etc isn t a real issue. People have been cheating on assignments since organised education was invented. There is a real problem of ML systems based on biased input data that issue decisions that are the average of the bigotry of the people who provided input. That isn t going to be worse than the current situation of bigoted humans making decisions based on hate and preconceptions but it will be more insidious. It is possible to search for that so for example a bank could test it s mortgage approval ML system by changing one factor at a time (name, gender, age, address, etc) and see if it changes the answer. If it turns out that the ML system is biased on names then the input data could have names removed. If it turns out to be biased about address then there could be weights put in to oppose that. For a long time there has been excessive trust in computers. Computers aren t magic they just do maths really fast and implement choices based on the work of programmers who have all the failings of other humans. Excessive trust in a rule based system is less risky than excessive trust in a ML system where no-one really knows why it makes the decisions it makes. Self driving cars kill people, this is the truth that Tesla stock holders don t want people to know. Companies that try to automate everything with AI are going to be in for some nasty surprises. Getting computers to do everything that humans do in any job is going to be a large portion of an actual intelligent computer which if it is achieved will raise an entirely different set of problems. I ve previously blogged about ML Security [5]. I don t think this will be any worse than all the other computer security problems in the long term, although it will be more insidious. How Will It Go? Companies spending billions of dollars without firm plans for how to make money are going to go bankrupt no matter what business they are in. Companies like Google and Microsoft can waste some billions of dollars on AI Chat systems and still keep going as successful businesses. Companies like OpenAI that do nothing other than such chat systems won t go well. But their assets can be used by new companies when sold at less than 10% the purchase price. Companies like NVidia that have high stock prices based on the supposed ongoing growth in use of their hardware will have their stock prices crash. But the new technology they develop will be used by other people for other purposes. If hospitals can get cheap diagnostic ML systems because of unreasonable investment into AI then that could be a win for humanity. Companies that bet their entire business on AI even when it s not necessarily their core business (as Tesla has done with self driving) will have their stock price crash dramatically at a minimum and have the possibility of bankruptcy. Having Tesla go bankrupt is definitely better than having people try to use them as self driving cars.

1 July 2025

Guido G nther: Free Software Activities June 2025

Another short status update of what happened on my side last month. Phosh 0.48.0 is out with nice improvements, phosh.mobi e.V. is alive, helped a bit to get cellbroadcastd out, osk bugfixes and some more: See below for details on the above and more: phosh phoc phosh-mobile-settings stevia (formerly phosh-osk-stub) phosh-osk-data phosh-vala-plugins pfs xdg-desktop-portal-phosh xdg-desktop-portal phosh-debs meta-phosh feedbackd feedbackd-device-themes gmobile GNOME clocks Debian Mobian ModemManager libmbim libqmi mobile-broadband-provider-info Cellbroadcastd iio-sensor-proxy twenty-twenty-hugo gotosocial Reviews This is not code by me but reviews on other peoples code. The list is (as usual) slightly incomplete. Thanks for the contributions! Help Development If you want to support my work see donations. Comments? Join the Fediverse thread

30 June 2025

Colin Watson: Free software activity in June 2025

My Debian contributions this month were all sponsored by Freexian. This was a very light month; I did a few things that were easy or that seemed urgent for the upcoming trixie release, but otherwise most of my energy went into Debusine. I ll be giving a talk about that at DebConf in a couple of weeks; this is the first DebConf I ll have managed to make it to in over a decade, so I m pretty excited. You can also support my work directly via Liberapay or GitHub Sponsors. PuTTY After reading a bunch of recent discourse about X11 and Wayland, I decided to try switching my laptop (a Framework 13 AMD running Debian trixie with GNOME) over to Wayland. I don t remember why it was running X; I think I must have either inherited some configuration from my previous laptop (in which case it could have been due to anything up to ten years ago or so), or else I had some initial problem while setting up my new laptop and failed to make a note of it. Anyway, the switch was hardly noticeable, which was great. One problem I did notice is that my preferred terminal emulator, pterm, crashed after the upgrade. I run a slightly-modified version from git to make some small terminal emulation changes that I really must either get upstream or work out how to live without one of these days, so it took me a while to notice that it only crashed when running from the packaged version, because the crash was in code that only runs when pterm has a set-id bit. I reported this upstream, they quickly fixed it, and I backported it to the Debian package. groff Upstream bug #67169 reported URLs being dropped from PDF output in some cases. I investigated the history both upstream and in Debian, identified the correct upstream patch to backport, and uploaded a fix. libfido2 I upgraded libfido2 to 1.16.0 in experimental. Python team I upgraded pydantic-extra-types to a new upstream version, and fixed some resulting fallout in pendulum. I updated python-typing-extensions in bookworm-backports, to help fix python3-tango: python3-pytango from bookworm-backports does not work (10.0.2-1~bpo12+1). I upgraded twisted to a new upstream version in experimental. I fixed or helped to fix a few release-critical bugs:

Russell Coker: Links June 2025

Jonathan McDowell wrote part 2 of his blog series about setting up a voice assistant on Debian, I look forward to reading further posts [1]. I m working on some related things for Debian that will hopefully work with this. I m testing out OpenSnitch on Trixie inspired by this blog post, it s an interesting package [2]. Valerie wrote an informative article about creating mesh networks using LORA for emergency use [3]. Interesting article about Signal and Windows Recall. That gives us some things to consider regarding ML features on Linux systems [4]. Insightful article about AI and the end of prestige [5]. We should all learn about LLMs. Jonathan Dowland wrote an informative blog post about how to manage namespaces on Linux [6]. The Consumer Rights wiki is a great resource for raising awareness of corporations exploiting their customers for computer related goods and services [7]. Interesting article about Schizophrenia and the cliff-edge function of evolution [8].

Next.

Previous.