Search Results: "reg"

3 December 2021

Evgeni Golov: Dependency confusion in the Ansible Galaxy CLI

I hope you enjoyed my last post about Ansible Galaxy Namespaces. In there I noted that I originally looked for something completely different and the namespace takeover was rather accidental. Well, originally I was looking at how the different Ansible content hosting services and their client (ansible-galaxy) behave in regard to clashes in naming of the hosted content. "Ansible content hosting services"?! There are currently three main ways for users to obtain Ansible content:
  • Ansible Galaxy - the original, community oriented, free hosting platform
  • Automation Hub - the place for Red Hat certified and supported content, available only with a Red Hat subscription, hosted by Red Hat
  • Ansible Automation Platform - the on-premise version of Automation Hub, syncs content from there and allows customers to upload own content
Now the question I was curious about was: how would the tooling behave if different sources would offer identically named content? This was inspired by Alex Birsan: Dependency Confusion: How I Hacked Into Apple, Microsoft and Dozens of Other Companies and zofrex: Bundler is Still Vulnerable to Dependency Confusion Attacks (CVE - 2020 - 36327), who showed that the tooling for Python, Node.js and Ruby can be tricked into fetching content from "the wrong source", thus allowing an attacker to inject malicious code into a deployment. For the rest of this article, it's not important that there are different implementations of the hosting services, only that users can configure and use multiple sources at the same time. The problem is that, if the user configures their server_list to contain multiple Galaxy-compatible servers, like Ansible Galaxy and Automation Hub, and then asks to install a collection, the Ansible Galaxy CLI will ask every server in the list, until one returns a successful result. The exact order seems to differ between versions, but this doesn't really matter for the issue at hand. Imagine someone wants to install the redhat.satellite collection from Automation Hub (using ansible-galaxy collection install redhat.satellite). Now if their configuration defines Galaxy as the first, and Automation Hub as the second server, Galaxy is always asked whether it has redhat.satellite and only if the answer is negative, Automation Hub is asked. Today there is no redhat namespace on Galaxy, but there is a redhat user on GitHub, so The canonical answer to this issue is to use a requirements.yml file and setting the source parameter. This parameter allows you to express "regardless which sources are configured, please fetch this collection from here". That's is nice, but I think this not being the default syntax (contrary to what e.g. Bundler does) is a bad approach. Users might overlook the security implications, as the shorter syntax without the source just "magically" works. However, I think this is not even the main problem here. The documentation says: Once a collection is found, any of its requirements are only searched within the same Galaxy instance as the parent collection. The install process will not search for a collection requirement in a different Galaxy instance. But as it turns out, the source behavior was changed and now only applies to the exact collection it is set for, not for any dependencies this collection might have. For the sake of the example, imagine two collections: evgeni.test1 and evgeni.test2, where test2 declares a dependency on test1 in its galaxy.yml. Actually, no need to imagine, both collections are available in version 1.0.0 from galaxy.ansible.com and test1 version 2.0.0 is available from galaxy-dev.ansible.com. Now, given our recent reading of the docs, we craft the following requirements.yml: 
collections:
- name: evgeni.test2
  version: '*'
  source: https://galaxy.ansible.com
In a perfect world, following the documentation, this would mean that both collections are fetched from galaxy.ansible.com, right? However, this is not what ansible-galaxy does. It will fetch evgeni.test2 from the specified source, determine it has a dependency on evgeni.test1 and fetch that from the "first" available source from the configuration. Take for example the following ansible.cfg: 
[galaxy]
server_list = test_galaxy, release_galaxy, test_galaxy
[galaxy_server.release_galaxy]
url=https://galaxy.ansible.com/
[galaxy_server.test_galaxy]
url=https://galaxy-dev.ansible.com/
And try to install collections, using the above requirements.yml: 
% ansible-galaxy collection install -r requirements.yml -vvv                 
ansible-galaxy 2.9.27
  config file = /home/evgeni/Devel/ansible-wtf/collections/ansible.cfg
  configured module search path = ['/home/evgeni/.ansible/plugins/modules', '/usr/share/ansible/plugins/modules']
  ansible python module location = /usr/lib/python3.10/site-packages/ansible
  executable location = /usr/bin/ansible-galaxy
  python version = 3.10.0 (default, Oct  4 2021, 00:00:00) [GCC 11.2.1 20210728 (Red Hat 11.2.1-1)]
Using /home/evgeni/Devel/ansible-wtf/collections/ansible.cfg as config file
Reading requirement file at '/home/evgeni/Devel/ansible-wtf/collections/requirements.yml'
Found installed collection theforeman.foreman:3.0.0 at '/home/evgeni/.ansible/collections/ansible_collections/theforeman/foreman'
Process install dependency map
Processing requirement collection 'evgeni.test2'
Collection 'evgeni.test2' obtained from server explicit_requirement_evgeni.test2 https://galaxy.ansible.com/api/
Opened /home/evgeni/.ansible/galaxy_token
Processing requirement collection 'evgeni.test1' - as dependency of evgeni.test2
Collection 'evgeni.test1' obtained from server test_galaxy https://galaxy-dev.ansible.com/api
Starting collection install process
Installing 'evgeni.test2:1.0.0' to '/home/evgeni/.ansible/collections/ansible_collections/evgeni/test2'
Downloading https://galaxy.ansible.com/download/evgeni-test2-1.0.0.tar.gz to /home/evgeni/.ansible/tmp/ansible-local-133/tmp9uqyjgki
Installing 'evgeni.test1:2.0.0' to '/home/evgeni/.ansible/collections/ansible_collections/evgeni/test1'
Downloading https://galaxy-dev.ansible.com/download/evgeni-test1-2.0.0.tar.gz to /home/evgeni/.ansible/tmp/ansible-local-133/tmp9uqyjgki
As you can see, evgeni.test1 is fetched from galaxy-dev.ansible.com, instead of galaxy.ansible.com. Now, if those servers instead would be Galaxy and Automation Hub, and somebody managed to snag the redhat namespace on Galaxy, I would be now getting the wrong stuff Another problematic setup would be with Galaxy and on-prem Ansible Automation Platform, as you can have any namespace on the later and these most certainly can clash with namespaces on public Galaxy. I have reported this behavior to Ansible Security on 2021-08-26, giving a 90 days disclosure deadline, which expired on 2021-11-24. So far, the response was that this is working as designed, to allow cross-source dependencies (e.g. a private collection referring to one on Galaxy) and there is an issue to update the docs to match the code. If users want to explicitly pin sources, they are supposed to name all dependencies and their sources in requirements.yml. Alternatively they obviously can configure only one source in the configuration and always mirror all dependencies. I am not happy with this and I think this is terrible UX, explicitly inviting people to make mistakes.

2 December 2021

Dirk Eddelbuettel: drat 0.2.2 on CRAN: Package Maintenance

drat user A fresh and new minor release of drat arrived on CRAN overnight. This is another small update relative to the 0.2.0 release in April followed by a 0.2.1 update in July. This release follows the changes made in digest yesterday. We removed the YAML file (and badge) for the disgraced former continuous integration service we shall not name (yet that we all used to use). And we converted the vignette from using the minidown package to the (fairly new) simplermarkdown package which is so much more appropriate for our use of the minimal water.css style. drat stands for drat R Archive Template, and helps with easy-to-create and easy-to-use repositories for R packages. Since its inception in early 2015 it has found reasonably widespread adoption among R users because repositories with marked releases is the better way to distribute code. See below for a few custom reference examples. Because for once it really is as your mother told you: Friends don t let friends install random git commit snapshots. Properly rolled-up releases it is. Just how CRAN shows us: a model that has demonstrated for two-plus decades how to do this. And you can too: drat is easy to use, documented by six vignettes and just works. Detailed information about drat is at its documentation site. The NEWS file summarises the release as follows:

Changes in drat version 0.2.2 (2021-12-01)
  • Travis artifacts and badges have been pruned
  • Vignettes now use simplermarkdown

Courtesy of my CRANberries, there is a comparison to the previous release. More detailed information is on the drat page as well as at the documentation site. If you like this or other open-source work I do, you can now sponsor me at GitHub.

This post by Dirk Eddelbuettel originated on his Thinking inside the box blog. Please report excessive re-aggregation in third-party for-profit settings.

1 December 2021

Dirk Eddelbuettel: digest 0.6.29 on CRAN: Package Maintenance

Release 0.6.29 of the digest package arrived at CRAN earlier today, and will be uploaded Debian shortly. digest creates hash digests of arbitrary R objects (using the md5, sha-1, sha-256, sha-512, crc32, xxhash32, xxhash64, murmur32, spookyhash, and blake3 algorithms) permitting easy comparison of R language objects. It is a mature and widely-used as many tasks may involve caching of objects for which it provides convenient general-purpose hash key generation. This release only contains two smaller internal changes. We removed the YAML file (and badge) for the disgraced former continuous integration service we shall not name (yet that we all used to use). And we converted the vignette from using the minidown package to the (fairly new) simplermarkdown package which is so much more appropriate for our use of the minimal water.css style. My CRANberries provides the usual summary of changes to the previous version. For questions or comments use the issue tracker off the GitHub repo. If you like this or other open-source work I do, you can now sponsor me at GitHub.

This post by Dirk Eddelbuettel originated on his Thinking inside the box blog. Please report excessive re-aggregation in third-party for-profit settings.

Utkarsh Gupta: FOSS Activites in December 2021

Here s my (twenty-sixth) monthly but brief update about the activities I ve done in the F/L/OSS world.

Debian
This was my 35th month of actively contributing to Debian. I became a DM in late March 2019 and a DD on Christmas 19! \o/ Just churning through the backlog again this month. Ugh. Anyway, I did the following stuff in Debian:

Uploads and bug fixes:
  • rails (2:6.1.4.1+dfsg-3) - No-change rebuild for unstable.

Other $things:
  • Mentoring for newcomers.
  • Moderation of -project mailing list.

Ubuntu
This was my 10th month of actively contributing to Ubuntu. Now that I ve joined Canonical to work on Ubuntu full-time, there s a bunch of things I do! \o/ I mostly worked on different things, I guess. I was too lazy to maintain a list of things I worked on so there s no concrete list atm. Maybe I ll get back to this section later or will start to list stuff from next year onward, as I was doing before. :D

Debian (E)LTS
Debian Long Term Support (LTS) is a project to extend the lifetime of all Debian stable releases to (at least) 5 years. Debian LTS is not handled by the Debian security team, but by a separate group of volunteers and companies interested in making it a success. And Debian Extended LTS (ELTS) is its sister project, extending support to the Jessie release (+2 years after LTS support). This was my twenty-sixth month as a Debian LTS and seventeenth month as a Debian ELTS paid contributor.
I was assigned 30.00 hours for LTS and 45.00 hours for ELTS and worked on the following things:

LTS CVE Fixes and Announcements:
  • Issued DLA 2813-1, fixing CVE-2021-33829 and CVE-2021-37695, for ckeditor.
    For Debian 9 stretch, these problems have been fixed in version 4.5.7+dfsg-2+deb9u1.
  • Issued DLA 2817-1, fixing CVE-2021-23214 and CVE-2021-23222, for postgresql-9.6.
    For Debian 9 stretch, these problems have been fixed in version 9.6.24-0+deb9u1.
  • Issued DLA 2836-1, fixing CVE-2021-43527, for nss.
    For Debian 9 stretch, these problems have been fixed in version 2:3.26.2-1.1+deb9u3.
  • Started working on src:samba for CVE-2020-25717 to CVE-2020-25722 and CVE-2021-23192 for jessie and stretch, both.
    The version difference b/w the suites are a bit too much for the patch(es) to be easily backported. I ve talked to Anton to work something out. \o/
  • Found the problem w/ libjdom1-java. Will have to roll the regression upload.
    I ve prepared the patch but needs some testing to be finally rolled out. Same for jessie.
  • Started working on libgit2.

ELTS CVE Fixes and Announcements:

Other (E)LTS Work:
  • Front-desk duty from 29-11 to 05-12 for both LTS and ELTS.
  • Triaged udisk2, wordpress, samba, gmp, nss, ntfs-3g, and openssh.
  • Auto EOL ed dwarfutils, radare2, mongodb, linux for jessie.
  • As FD, did a deep dive into the no-pu-update issue. Will write to list shortly.
  • Attended monthly Debian LTS meeting.
  • Answered questions (& discussions) on IRC (#debian-lts and #debian-elts).
  • General and other discussions on LTS private and public mailing list.

Debian LTS Survey I ve spent 3 hours on the LTS survey on the following bits:
  • Talking to Laura to revive the old a/c on survey.d.net.
  • Setting up stuff there.
  • Discussing the survey questions and other bits w/ Jeremiah.
  • Partly reviewing the questions of the survey.
  • Doing a walkthru of the LimeSurvey instance we have to make sure there are no changes .
Until next time.
:wq for today.

30 November 2021

Russell Coker: Links November 2021

The Guardian has an amusing article by Sophie Elmhirst about Libertarians buying a cruise ship to make a seasteading project off the coast of Panama [1]. It turns out that you need permits etc to do this and maintaining a ship is expensive. Also you wouldn t want to mine cryptocurrency in a ship cabin as most cabins are small and don t have enough airconditioning to remain pleasant if you dump 1kW or more into the air. NPR has an interesting article about the reaction of the NRA to the Columbine shootings [2]. Seems that some NRA person isn t a total asshole and is sharing their private information, maybe they are dying and are worried about going to hell. David Brin wrote an insightful blog post about the singleton hypothesis where he covers some of the evidence of autocratic societies failing [3]. I think he makes a convincing point about a single centralised government for human society not being viable. But something like the EU on a world wide scale could work well. Ken Shirriff wrote an interesting blog post about reverse engineering the Yamaha DX7 synthesiser [4]. The New York Times has an interesting article about a Baboon troop that became less aggressive after the alpha males all died at once from tuberculosis [5]. They established a new more peaceful culture that has outlived the beta males who avoided tuberculosis. The Guardian has an interesting article about how sequencing the genomes of the entire population can save healthcare costs while improving the health of the population [6]. This is somthing wealthy countries should offer for free to the world population. At a bit under $1000 per test that s only about $7 trillion to test everyone, and of course the price should drop significantly if there were billions of tests being done. The Strategy Bridge has an interesting article about SciFi books that have useful portrayals of military strategy [7]. The co-author is Major General Mick Ryan of the Australian Army which is noteworthy as Major General is the second highest rank in use by the Australian Army at this time. Vice has an interesting article about the co-evolution of penises and vaginas and how a lot of that evolution is based on avoiding impregnation from rape [8]. Cory Doctorow wrote an insightful Medium article about the way that governments could force interoperability through purchasing power [9]. Cory Doctorow wrote an insightful article for Locus Magazine about imagining life after capitalism and how capitalism might be replaced [10]. We need a Star Trek future! Arstechnica has an informative article about new developmenet in the rowhammer category of security attacks on DRAM [11]. It seems that DDR4 with ECC is the best current mitigation technique and that DDR3 with ECC is harder to attack than non-ECC RAM. So the thing to do is use ECC on all workstations and avoid doing security critical things on laptops because they can t juse ECC RAM.
Related posts:
  1. Links April 2021 Dr Justin Lehmiller s blog post comparing his official (academic style)...
  2. Links October 2021 Bloomburg has an insightful article about Juniper, the NSA, and...
  3. Links September 2021 Matthew Garrett wrote an interesting and insightful blog post about...

Russell Coker: Your Device Has Been Improved

I ve just started a Samsung tablet downloading a 770MB update, the description says:
Technically I have no doubt that both those claims are true and accurate. But according to common understanding of the English language I think they are both misleading. By stability improved they mean fixed some bugs that made it unstable and no technical person would imagine that after a certain number of such updates the number of bugs will ever reach zero and the tablet will be perfectly reliable. In fact if you should consider yourself lucky if they fix more bugs than they add. It s not THAT uncommon for phones and tablets to be bricked (rendered unusable by software) by an update. In the past I got a Huawei Mate9 as a warranty replacement for a Nexus 6P because an update caused so many Nexus 6P phones to fail that they couldn t be replaced with an identical phone [1]. By security improved they usually mean fixed some security flaws that were recently discovered to make it almost as secure as it was designed to be . Note that I deliberately say almost as secure because it s sometimes impossible to fix a security flaw without making significant changes to interfaces which requires more work than desired for an old product and also gives a higher probability of things going wrong. So it s sometimes better to aim for almost as secure or alternatively just as secure but with some features disabled. Device manufacturers (and most companies in the Android space make the same claims while having the exact same bugs to deal with, Samsung is no different from the others in this regards) are not making devices more secure or more reliable than when they were initially released. They are aiming to make them almost as secure and reliable as when they were released. They don t have much incentive to try too hard in this regard, Samsung won t suffer if I decide my old tablet isn t reliable enough and buy a new one, which will almost certainly be from Samsung because they make nice tablets. As a thought experiment, consider if car repairers did the same thing. Getting us to service your car will improve fuel efficiency , great how much more efficient will it be than when I purchased it? As another thought experiment, consider if car companies stopped providing parts for car repair a few years after releasing a new model. This is effectively what phone and tablet manufacturers have been doing all along, software updates for stability and security are to devices what changing oil etc is for cars.
Related posts:
  1. Long-term Device Use It seems to me that Android phones have recently passed...
  2. Android Device Service Life In recent years Android devices have been the most expensive...
  3. Xen and eth device renaming Recently I rebooted one of my Debian Xen servers and...

29 November 2021

Evgeni Golov: Getting access to somebody else's Ansible Galaxy namespace

TL;DR: adding features after the fact is hard, normalizing names is hard, it's patched, carry on. I promise, the longer version is more interesting and fun to read! Recently, I was poking around Ansible Galaxy and almost accidentally got access to someone else's namespace. I was actually looking for something completely different, but accidental finds are the best ones! If you're asking yourself: "what the heck is he talking about?!", let's slow down for a moment:
  • Ansible is a great automation engine built around the concept of modules that do things (mostly written in Python) and playbooks (mostly written in YAML) that tell which things to do
  • Ansible Galaxy is a place where people can share their playbooks and modules for others to reuse
  • Galaxy Namespaces are a way to allow users to distinguish who published what and reduce name clashes to a minimum
That means that if I ever want to share how to automate installing vim, I can publish evgeni.vim on Galaxy and other people can download that and use it. And if my evil twin wants their vim recipe published, it will end up being called evilme.vim. Thus while both recipes are called vim they can coexist, can be downloaded to the same machine, and used independently. How do you get a namespace? It's automatically created for you when you login for the first time. After that you can manage it, you can upload content, allow others to upload content and other things. You can also request additional namespaces, this is useful if you want one for an Organization or similar entities, which don't have a login for Galaxy. Apropos login, Galaxy uses GitHub for authentication, so you don't have to store yet another password, just smash that octocat! Did anyone actually click on those links above? If you did (you didn't, right?), you might have noticed another section in that document: Namespace Limitations. That says:
Namespace names in Galaxy are limited to lowercase word characters (i.e., a-z, 0-9) and _ , must have a minimum length of 2 characters, and cannot start with an _ . No other characters are allowed, including . , - , and space. The first time you log into Galaxy, the server will create a Namespace for you, if one does not already exist, by converting your username to lowercase, and replacing any - characters with _ .
For my login evgeni this is pretty boring, as the generated namespace is also evgeni. But for the GitHub user Evil-Pwnwil-666 it will become evil_pwnwil_666. This can be a bit confusing. Another confusing thing is that Galaxy supports two types of content: roles and collections, but namespaces are only for collections! So it is Evil-Pwnwil-666.vim if it's a role, but evil_pwnwil_666.vim if it's a collection. I think part of this split is because collections were added much later and have a much more well thought design of both the artifact itself and its delivery mechanisms. This is by the way very important for us! Due to the fact that collections (and namespaces!) were added later, there must be code that ensures that users who were created before also get a namespace. Galaxy does this (and I would have done it the same way) by hooking into the login process, and after the user is logged in it checks if a Namespace exists and if not it creates one and sets proper permissions. And this is also exactly where the issue was! The old code looked like this: 
    # Create lowercase namespace if case insensitive search does not find match
    qs = models.Namespace.objects.filter(
        name__iexact=sanitized_username).order_by('name')
    if qs.exists():
        namespace = qs[0]
    else:
        namespace = models.Namespace.objects.create(**ns_defaults)
    namespace.owners.add(user)
See how namespace.owners.add is always called? Even if the namespace already existed? Yepp! But how can we exploit that? Any user either already has a namespace (and owns it) or doesn't have one that could be owned. And given users are tied to GitHub accounts, there is no way to confuse Galaxy here. Now, remember how I said one could request additional namespaces, for organizations and stuff? Those will have owners, but the namespace name might not correspond to an existing user! So all we need is to find an existing Galaxy namespace that is not a "default" namespace (aka a specially requested one) and get a GitHub account that (after the funny name conversion) matches the namespace name. Thankfully Galaxy has an API, so I could dump all existing namespaces and their owners. Next I filtered that list to have only namespaces where the owner list doesn't contain a username that would (after conversion) match the namespace name. I found a few. And for one of them (let's call it the_target), the corresponding GitHub username (the-target) was available! Jackpot! I've registered a new GitHub account with that name, logged in to Galaxy and had access to the previously found namespace. This felt like sufficient proof that my attack worked and I mailed my findings to the Ansible Security team. The issue was fixed in d4f84d3400f887a26a9032687a06dd263029bde3 by moving the namespace.owners.add call to the "new namespace" branch. And this concludes the story of how I accidentally got access to someone else's Galaxy namespace (which was revoked after the report, no worries).

24 November 2021

Dirk Eddelbuettel: nanotime 0.3.4 on CRAN: Maintenance Update

Another (minor) nanotime release, now at version 0.3.4, arrived at CRAN overnight. It exports some nanoperiod functionality via a C++ header, and Leonardo and I will use this in an upcoming package that we hope to talk about a little more in a few days. It also adds a few as.character.*() methods that had not been included before. nanotime relies on the RcppCCTZ package for (efficient) high(er) resolution time parsing and formatting up to nanosecond resolution, and the bit64 package for the actual integer64 arithmetic. Initially implemented using the S3 system, it has benefitted greatly from a rigorous refactoring by Leonardo who not only rejigged nanotime internals in S4 but also added new S4 types for periods, intervals and durations. The NEWS snippet adds more details.

Changes in version 0.3.4 (2021-11-24)
  • Added a few more as.character conversion function (Dirk)
  • Expose nanoperiod functionality via header file for use by other packages (Leonardo in #95 fixing #94).

Thanks to CRANberries there is also a diff to the previous version. More details and examples are at the nanotime page; code, issue tickets etc at the GitHub repository. If you like this or other open-source work I do, you can now sponsor me at GitHub.

This post by Dirk Eddelbuettel originated on his Thinking inside the box blog. Please report excessive re-aggregation in third-party for-profit settings.

21 November 2021

Antoine Beaupr : mbsync vs OfflineIMAP

After recovering from my latest email crash (previously, previously), I had to figure out which tool I should be using. I had many options but I figured I would start with a popular one (mbsync). But I also evaluated OfflineIMAP which was resurrected from the Python 2 apocalypse, and because I had used it before, for a long time. Read on for the details.

Benchmark setup All programs were tested against a Dovecot 1:2.3.13+dfsg1-2 server, running Debian bullseye. The client is a Purism 13v4 laptop with a Samsung SSD 970 EVO 1TB NVMe drive. The server is a custom build with a AMD Ryzen 5 2600 CPU, and a RAID-1 array made of two NVMe drives (Intel SSDPEKNW010T8 and WDC WDS100T2B0C). The mail spool I am testing against has almost 400k messages and takes 13GB of disk space: 
$ notmuch count --exclude=false
372758
$ du -sh --exclude xapian Maildir
13G Maildir
The baseline we are comparing against is SMD (syncmaildir) which performs the sync in about 7-8 seconds locally (3.5 seconds for each push/pull command) and about 10-12 seconds remotely. Anything close to that or better is good enough. I do not have recent numbers for a SMD full sync baseline, but the setup documentation mentions 20 minutes for a full sync. That was a few years ago, and the spool has obviously grown since then, so that is not a reliable baseline. A baseline for a full sync might be also set with rsync, which copies files at nearly 40MB/s, or 317Mb/s! 
anarcat@angela:tmp(main)$ time rsync -a --info=progress2 --exclude xapian  shell.anarc.at:Maildir/ Maildir/
 12,647,814,731 100%   37.85MB/s    0:05:18 (xfr#394981, to-chk=0/395815)    
72.38user 106.10system 5:19.59elapsed 55%CPU (0avgtext+0avgdata 15988maxresident)k
8816inputs+26305112outputs (0major+50953minor)pagefaults 0swaps
That is 5 minutes to transfer the entire spool. Incremental syncs are obviously pretty fast too: 
anarcat@angela:tmp(main)$ time rsync -a --info=progress2 --exclude xapian  shell.anarc.at:Maildir/ Maildir/
              0   0%    0.00kB/s    0:00:00 (xfr#0, to-chk=0/395815)    
1.42user 0.81system 0:03.31elapsed 67%CPU (0avgtext+0avgdata 14100maxresident)k
120inputs+0outputs (3major+12709minor)pagefaults 0swaps
As an extra curiosity, here's the performance with tar, pretty similar with rsync, minus incremental which I cannot be bothered to figure out right now: 
anarcat@angela:tmp(main)$ time ssh shell.anarc.at tar --exclude xapian -cf - Maildir/   pv -s 13G   tar xf - 
56.68user 58.86system 5:17.08elapsed 36%CPU (0avgtext+0avgdata 8764maxresident)k
0inputs+0outputs (0major+7266minor)pagefaults 0swaps
12,1GiO 0:05:17 [39,0MiB/s] [===================================================================> ] 92%
Interesting that rsync manages to almost beat a plain tar on file transfer, I'm actually surprised by how well it performs here, considering there are many little files to transfer. (But then again, this maybe is exactly where rsync shines: while tar needs to glue all those little files together, rsync can just directly talk to the other side and tell it to do live changes. Something to look at in another article maybe?) Since both ends are NVMe drives, those should easily saturate a gigabit link. And in fact, a backup of the server mail spool achieves much faster transfer rate on disks: 
anarcat@marcos:~$ tar fc - Maildir   pv -s 13G > Maildir.tar
15,0GiO 0:01:57 [ 131MiB/s] [===================================] 115%
That's 131Mibyyte per second, vastly faster than the gigabit link. The client has similar performance: 
anarcat@angela:~(main)$ tar fc - Maildir   pv -s 17G > Maildir.tar
16,2GiO 0:02:22 [ 116MiB/s] [==================================] 95%
So those disks should be able to saturate a gigabit link, and they are not the bottleneck on fast links. Which begs the question of what is blocking performance of a similar transfer over the gigabit link, but that's another question altogether, because no sync program ever reaches the above performance anyways. Finally, note that when I migrated to SMD, I wrote a small performance comparison that could be interesting here. It show SMD to be faster than OfflineIMAP, but not as much as we see here. In fact, it looks like OfflineIMAP slowed down significantly since then (May 2018), but this could be due to my larger mail spool as well.

mbsync The isync (AKA mbsync) project is written in C and supports syncing Maildir and IMAP folders, with possibly multiple replicas. I haven't tested this but I suspect it might be possible to sync between two IMAP servers as well. It supports partial mirorrs, message flags, full folder support, and "trash" functionality.

Complex configuration file I started with this .mbsyncrc configuration file: 
SyncState *
Sync New ReNew Flags
IMAPAccount anarcat
Host imap.anarc.at
User anarcat
PassCmd "pass imap.anarc.at"
SSLType IMAPS
CertificateFile /etc/ssl/certs/ca-certificates.crt
IMAPStore anarcat-remote
Account anarcat
MaildirStore anarcat-local
# Maildir/top/sub/sub
#SubFolders Verbatim
# Maildir/.top.sub.sub
SubFolders Maildir++
# Maildir/top/.sub/.sub
# SubFolders legacy
# The trailing "/" is important
#Path ~/Maildir-mbsync/
Inbox ~/Maildir-mbsync/
Channel anarcat
# AKA Far, convert when all clients are 1.4+
Master :anarcat-remote:
# AKA Near
Slave :anarcat-local:
# Exclude everything under the internal [Gmail] folder, except the interesting folders
#Patterns * ![Gmail]* "[Gmail]/Sent Mail" "[Gmail]/Starred" "[Gmail]/All Mail"
# Or include everything
Patterns *
# Automatically create missing mailboxes, both locally and on the server
#Create Both
Create slave
# Sync the movement of messages between folders and deletions, add after making sure the sync works
#Expunge Both
Long gone are the days where I would spend a long time reading a manual page to figure out the meaning of every option. If that's your thing, you might like this one. But I'm more of a "EXAMPLES section" kind of person now, and I somehow couldn't find a sample file on the website. I started from the Arch wiki one but it's actually not great because it's made for Gmail (which is not a usual Dovecot server). So a sample config file in the manpage would be a great addition. Thankfully, the Debian packages ships one in /usr/share/doc/isync/examples/mbsyncrc.sample but I only found that after I wrote my configuration. It was still useful and I recommend people take a look if they want to understand the syntax. Also, that syntax is a little overly complicated. For example, Far needs colons, like: 
Far :anarcat-remote:
Why? That seems just too complicated. I also found that sections are not clearly identified: IMAPAccount and Channel mark section beginnings, for example, which is not at all obvious until you learn about mbsync's internals. There are also weird ordering issues: the SyncState option needs to be before IMAPAccount, presumably because it's global. Using a more standard format like .INI or TOML could improve that situation.

Stellar performance A transfer of the entire mail spool takes 56 minutes and 6 seconds, which is impressive. It's not quite "line rate": the resulting mail spool was 12GB (which is a problem, see below), which turns out to be about 29Mbit/s and therefore not maxing the gigabit link, and an order of magnitude slower than rsync. The incremental runs are roughly 2 seconds, which is even more impressive, as that's actually faster than rsync: 
===> multitime results
1: mbsync -a
            Mean        Std.Dev.    Min         Median      Max
real        2.015       0.052       1.930       2.029       2.105       
user        0.660       0.040       0.592       0.661       0.722       
sys         0.338       0.033       0.268       0.341       0.387
Those tests were performed with isync 1.3.0-2.2 on Debian bullseye. Tests with a newer isync release originally failed because of a corrupted message that triggered bug 999804 (see below). Running 1.4.3 under valgrind works around the bug, but adds a 50% performance cost, the full sync running in 1h35m. Once the upstream patch is applied, performance with 1.4.3 is fairly similar, considering that the new sync included the register folder with 4000 messages: 
120.74user 213.19system 59:47.69elapsed 9%CPU (0avgtext+0avgdata 105420maxresident)k
29128inputs+28284376outputs (0major+45711minor)pagefaults 0swaps
That is ~13GB in ~60 minutes, which gives us 28.3Mbps. Incrementals are also pretty similar to 1.3.x, again considering the double-connect cost: 
===> multitime results
1: mbsync -a
            Mean        Std.Dev.    Min         Median      Max
real        2.500       0.087       2.340       2.491       2.629       
user        0.718       0.037       0.679       0.711       0.793       
sys         0.322       0.024       0.284       0.320       0.365
Those tests were all done on a Gigabit link, but what happens on a slower link? My server uplink is slow: 25 Mbps down, 6 Mbps up. There mbsync is worse than the SMD baseline: 
===> multitime results
1: mbsync -a
Mean        Std.Dev.    Min         Median      Max
real        31.531      0.724       30.764      31.271      33.100      
user        1.858       0.125       1.721       1.818       2.131       
sys         0.610       0.063       0.506       0.600       0.695
That's 30 seconds for a sync, which is an order of magnitude slower than SMD.

Great user interface Compared to OfflineIMAP and (ahem) SMD, the mbsync UI is kind of neat: 
anarcat@angela:~(main)$ mbsync -a
Notice: Master/Slave are deprecated; use Far/Near instead.
C: 1/2  B: 204/205  F: +0/0 *0/0 #0/0  N: +1/200 *0/0 #0/0
(Note that nice switch away from slavery-related terms too.) The display is minimal, and yet informative. It's not obvious what does mean at first glance, but the manpage is useful at least for clarifying that:
This represents the cumulative progress over channels, boxes, and messages affected on the far and near side, respectively. The message counts represent added messages, messages with updated flags, and trashed messages, respectively. No attempt is made to calculate the totals in advance, so they grow over time as more information is gathered. (Emphasis mine).
In other words:
  • C 2/2: channels done/total (2 done out of 2)
  • B 204/205: mailboxes done/total (204 out of 205)
  • F: changes on the far side
  • N: +10/200 *0/0 #0/0: changes on the "near" side:
    • +10/200: 10 out of 200 messages downloaded
    • *0/0: no flag changed
    • #0/0: no message deleted
You get used to it, in a good way. It does not, unfortunately, show up when you run it in systemd, which is a bit annoying as I like to see a summary mail traffic in the logs.

Interoperability issue In my notmuch setup, I have bound key S to "mark spam", which basically assigns the tag spam to the message and removes a bunch of others. Then I have a notmuch-purge script which moves that message to the spam folder, for training purposes. It basically does this: 
notmuch search --output=files --format=text0 "$search_spam" \
      xargs -r -0 mv -t "$HOME/Maildir/$ PREFIX junk/cur/"
This method, which worked fine in SMD (and also OfflineIMAP) created this error on sync: 
Maildir error: duplicate UID 37578.
And indeed, there are now two messages with that UID in the mailbox: 
anarcat@angela:~(main)$ find Maildir/.junk/ -name '*U=37578*'
Maildir/.junk/cur/1637427889.134334_2.angela,U=37578:2,S
Maildir/.junk/cur/1637348602.2492889_221804.angela,U=37578:2,S
This is actually a known limitation or, as mbsync(1) calls it, a "RECOMMENDATION":
When using the more efficient default UID mapping scheme, it is important that the MUA renames files when moving them between Maildir fold ers. Mutt always does that, while mu4e needs to be configured to do it: 
(setq mu4e-change-filenames-when-moving t)
So it seems I would need to fix my script. It's unclear how the paths should be renamed, which is unfortunate, because I would need to change my script to adapt to mbsync, but I can't tell how just from reading the above. (A manual fix is actually to rename the file to remove the U= field: mbsync will generate a new one and then sync correctly.) Fortunately, someone else already fixed that issue: afew, a notmuch tagging script (much puns, such hurt), has a move mode that can rename files correctly, specifically designed to deal with mbsync. I had already been told about afew, but it's one more reason to standardize my notmuch hooks on that project, it looks like. Update: I have tried to use afew and found it has significant performance issues. It also has a completely different paradigm to what I am used to: it assumes all incoming mail has a new and lays its own tags on top of that (inbox, sent, etc). It can only move files from one folder at a time (see this bug) which breaks my spam training workflow. In general, I sync my tags into folders (e.g. ham, spam, sent) and message flags (e.g. inbox is F, unread is "not S", etc), and afew is not well suited for this (although there are hacks that try to fix this). I have worked hard to make my tagging scripts idempotent, and it's something afew doesn't currently have. Still, it would be better to have that code in Python than bash, so maybe I should consider my options here.

Stability issues The newer release in Debian bookworm (currently at 1.4.3) has stability issues on full sync. I filed bug 999804 in Debian about this, which lead to a thread on the upstream mailing list. I have found at least three distinct crashes that could be double-free bugs "which might be exploitable in the worst case", not a reassuring prospect. The thing is: mbsync is really fast, but the downside of that is that it's written in C, and with that comes a whole set of security issues. The Debian security tracker has only three CVEs on isync, but the above issues show there could be many more. Reading the source code certainly did not make me very comfortable with trusting it with untrusted data. I considered sandboxing it with systemd (below) but having systemd run as a --user process makes that difficult. I also considered using an apparmor profile but that is not trivial because we need to allow SSH and only some parts of it... Thankfully, upstream has been diligent at addressing the issues I have found. They provided a patch within a few days which did fix the sync issues. Update: upstream actually took the issue very seriously. They not only got CVE-2021-44143 assigned for my bug report, they also audited the code and found several more issues collectively identified as CVE-2021-3657, which actually also affect 1.3 (ie. Debian 11/bullseye/stable). Somehow my corpus doesn't trigger that issue, but it was still considered serious enough to warrant a CVE. So one the one hand: excellent response from upstream; but on the other hand: how many more of those could there be in there?

Automation with systemd The Arch wiki has instructions on how to setup mbsync as a systemd service. It suggests using the --verbose (-V) flag which is a little intense here, as it outputs 1444 lines of messages. I have used the following .service file: 
[Unit]
Description=Mailbox synchronization service
ConditionHost=!marcos
Wants=network-online.target
After=network-online.target
Before=notmuch-new.service
[Service]
Type=oneshot
ExecStart=/usr/bin/mbsync -a
Nice=10
IOSchedulingClass=idle
NoNewPrivileges=true
[Install]
WantedBy=default.target
And the following .timer: 
[Unit]
Description=Mailbox synchronization timer
ConditionHost=!marcos
[Timer]
OnBootSec=2m
OnUnitActiveSec=5m
Unit=mbsync.service
[Install]
WantedBy=timers.target
Note that we trigger notmuch through systemd, with the Before and also by adding mbsync.service to the notmuch-new.service file: 
[Unit]
Description=notmuch new
After=mbsync.service
[Service]
Type=oneshot
Nice=10
ExecStart=/usr/bin/notmuch new
[Install]
WantedBy=mbsync.service
An improvement over polling repeatedly with a .timer would be to wake up only on IMAP notify, but neither imapnotify nor goimapnotify seem to be packaged in Debian. It would also not cover for the "sent folder" use case, where we need to wake up on local changes.

Password-less setup The sample file suggests this should work: 
IMAPStore remote
Tunnel "ssh -q host.remote.com /usr/sbin/imapd"
Add BatchMode, restrict to IdentitiesOnly, provide a password-less key just for this, add compression (-C), find the Dovecot imap binary, and you get this: 
IMAPAccount anarcat-tunnel
Tunnel "ssh -o BatchMode=yes -o IdentitiesOnly=yes -i ~/.ssh/id_ed25519_mbsync -o HostKeyAlias=shell.anarc.at -C anarcat@imap.anarc.at /usr/lib/dovecot/imap"
And it actually seems to work: 
$ mbsync -a
Notice: Master/Slave are deprecated; use Far/Near instead.
C: 0/2  B: 0/1  F: +0/0 *0/0 #0/0  N: +0/0 *0/0 #0/0imap(anarcat): Error: net_connect_unix(/run/dovecot/stats-writer) failed: Permission denied
C: 2/2  B: 205/205  F: +0/0 *0/0 #0/0  N: +1/1 *3/3 #0/0imap(anarcat)<1611280><90uUOuyElmEQlhgAFjQyWQ>: Info: Logged out in=10808 out=15396642 deleted=0 expunged=0 trashed=0 hdr_count=0 hdr_bytes=0 body_count=1 body_bytes=8087
It's a bit noisy, however. dovecot/imap doesn't have a "usage" to speak of, but even the source code doesn't hint at a way to disable that Error message, so that's unfortunate. That socket is owned by root:dovecot so presumably Dovecot runs the imap process as $user:dovecot, which we can't do here. Oh well? Interestingly, the SSH setup is not faster than IMAP. With IMAP: 
===> multitime results
1: mbsync -a
            Mean        Std.Dev.    Min         Median      Max
real        2.367       0.065       2.220       2.376       2.458       
user        0.793       0.047       0.731       0.776       0.871       
sys         0.426       0.040       0.364       0.434       0.476
With SSH: 
===> multitime results
1: mbsync -a
            Mean        Std.Dev.    Min         Median      Max
real        2.515       0.088       2.274       2.532       2.594       
user        0.753       0.043       0.645       0.766       0.804       
sys         0.328       0.045       0.212       0.340       0.393
Basically: 200ms slower. Tolerable.

Migrating from SMD The above was how I migrated to mbsync on my first workstation. The work on the second one was more streamlined, especially since the corruption on mailboxes was fixed:
  1. install isync, with the patch: 
    dpkg -i isync_1.4.3-1.1~_amd64.deb
  2. copy all files over from previous workstation to avoid a full resync (optional): 
    rsync -a --info=progress2 angela:Maildir/ Maildir-mbsync/
  3. rename all files to match new hostname (optional): 
    find Maildir-mbsync/ -type f -name '*.angela,*' -print0    rename -0 's/\.angela,/\.curie,/'
  4. trash the notmuch database (optional): 
    rm -rf Maildir-mbsync/.notmuch/xapian/
  5. disable all smd and notmuch services: 
    systemctl --user --now disable smd-pull.service smd-pull.timer smd-push.service smd-push.timer notmuch-new.service notmuch-new.timer
  6. do one last sync with smd: 
    smd-pull --show-tags ; smd-push --show-tags ; notmuch new ; notmuch-sync-flagged -v
  7. backup notmuch on the client and server: 
    notmuch dump   pv > notmuch.dump
  8. backup the maildir on the client and server: 
    cp -al Maildir Maildir-bak
  9. create the SSH key: 
    ssh-keygen -t ed25519 -f .ssh/id_ed25519_mbsync
cat .ssh/id_ed25519_mbsync.pub
  10. add to .ssh/authorized_keys on the server, like this: command="/usr/lib/dovecot/imap",restrict ssh-ed25519 AAAAC...
  11. move old files aside, if present: 
    mv Maildir Maildir-smd
  12. move new files in place (CRITICAL SECTION BEGINS!): 
    mv Maildir-mbsync Maildir
  13. run a test sync, only pulling changes: mbsync --create-near --remove-none --expunge-none --noop anarcat-register
  14. if that works well, try with all mailboxes: mbsync --create-near --remove-none --expunge-none --noop -a
  15. if that works well, try again with a full sync: mbsync register mbsync -a
  16. reindex and restore the notmuch database, this should take ~25 minutes: 
    notmuch new
pv notmuch.dump   notmuch restore
  17. enable the systemd services and retire the smd-* services: systemctl --user enable mbsync.timer notmuch-new.service systemctl --user start mbsync.timer rm ~/.config/systemd/user/smd* systemctl daemon-reload
During the migration, notmuch helpfully told me the full list of those lost messages: 
[...]
Warning: cannot apply tags to missing message: CAN6gO7_QgCaiDFvpG3AXHi6fW12qaN286+2a7ERQ2CQtzjSEPw@mail.gmail.com
Warning: cannot apply tags to missing message: CAPTU9Wmp0yAmaxO+qo8CegzRQZhCP853TWQ_Ne-YF94MDUZ+Dw@mail.gmail.com
Warning: cannot apply tags to missing message: F5086003-2917-4659-B7D2-66C62FCD4128@gmail.com
[...]
Warning: cannot apply tags to missing message: mailman.2.1316793601.53477.sage-members@mailman.sage.org
Warning: cannot apply tags to missing message: mailman.7.1317646801.26891.outages-discussion@outages.org
Warning: cannot apply tags to missing message: notmuch-sha1-000458df6e48d4857187a000d643ac971deeef47
Warning: cannot apply tags to missing message: notmuch-sha1-0079d8e0c3340e6f88c66f4c49fca758ea71d06d
Warning: cannot apply tags to missing message: notmuch-sha1-0194baa4cfb6d39bc9e4d8c049adaccaa777467d
Warning: cannot apply tags to missing message: notmuch-sha1-02aede494fc3f9e9f060cfd7c044d6d724ad287c
Warning: cannot apply tags to missing message: notmuch-sha1-06606c625d3b3445420e737afd9a245ae66e5562
Warning: cannot apply tags to missing message: notmuch-sha1-0747b020f7551415b9bf5059c58e0a637ba53b13
[...]
As detailed in the crash report, all of those were actually innocuous and could be ignored. Also note that we completely trash the notmuch database because it's actually faster to reindex from scratch than let notmuch slowly figure out that all mails are new and all the old mails are gone. The fresh indexing took: 
nov 19 15:08:54 angela notmuch[2521117]: Processed 384679 total files in 23m 41s (270 files/sec.).
nov 19 15:08:54 angela notmuch[2521117]: Added 372610 new messages to the database.
While a reindexing on top of an existing database was going twice as slow, at about 120 files/sec.

Current config file Putting it all together, I ended up with the following configuration file: 
SyncState *
Sync All
# IMAP side, AKA "Far"
IMAPAccount anarcat-imap
Host imap.anarc.at
User anarcat
PassCmd "pass imap.anarc.at"
SSLType IMAPS
CertificateFile /etc/ssl/certs/ca-certificates.crt
IMAPAccount anarcat-tunnel
Tunnel "ssh -o BatchMode=yes -o IdentitiesOnly=yes -i ~/.ssh/id_ed25519_mbsync -o HostKeyAlias=shell.anarc.at -C anarcat@imap.anarc.at /usr/lib/dovecot/imap"
IMAPStore anarcat-remote
Account anarcat-tunnel
# Maildir side, AKA "Near"
MaildirStore anarcat-local
# Maildir/top/sub/sub
#SubFolders Verbatim
# Maildir/.top.sub.sub
SubFolders Maildir++
# Maildir/top/.sub/.sub
# SubFolders legacy
# The trailing "/" is important
#Path ~/Maildir-mbsync/
Inbox ~/Maildir/
# what binds Maildir and IMAP
Channel anarcat
Far :anarcat-remote:
Near :anarcat-local:
# Exclude everything under the internal [Gmail] folder, except the interesting folders
#Patterns * ![Gmail]* "[Gmail]/Sent Mail" "[Gmail]/Starred" "[Gmail]/All Mail"
# Or include everything
#Patterns *
Patterns * !register  !.register
# Automatically create missing mailboxes, both locally and on the server
Create Both
#Create Near
# Sync the movement of messages between folders and deletions, add after making sure the sync works
Expunge Both
# Propagate mailbox deletion
Remove both
IMAPAccount anarcat-register-imap
Host imap.anarc.at
User register
PassCmd "pass imap.anarc.at-register"
SSLType IMAPS
CertificateFile /etc/ssl/certs/ca-certificates.crt
IMAPAccount anarcat-register-tunnel
Tunnel "ssh -o BatchMode=yes -o IdentitiesOnly=yes -i ~/.ssh/id_ed25519_mbsync -o HostKeyAlias=shell.anarc.at -C register@imap.anarc.at /usr/lib/dovecot/imap"
IMAPStore anarcat-register-remote
Account anarcat-register-tunnel
MaildirStore anarcat-register-local
SubFolders Maildir++
Inbox ~/Maildir/.register/
Channel anarcat-register
Far :anarcat-register-remote:
Near :anarcat-register-local:
Create Both
Expunge Both
Remove both
Note that it may be out of sync with my live (and private) configuration file, as I do not publish my "dotfiles" repository publicly for security reasons.

OfflineIMAP I've used OfflineIMAP for a long time before switching to SMD. I don't exactly remember why or when I started using it, but I do remember it became painfully slow as I started using notmuch, and would sometimes crash mysteriously. It's been a while, so my memory is hazy on that. It also kind of died in a fire when Python 2 stop being maintained. The main author moved on to a different project, imapfw which could serve as a framework to build IMAP clients, but never seemed to implement all of the OfflineIMAP features and certainly not configuration file compatibility. Thankfully, a new team of volunteers ported OfflineIMAP to Python 3 and we can now test that new version to see if it is an improvement over mbsync.

Crash on full sync The first thing that happened on a full sync is this crash: 
Copy message from RemoteAnarcat:junk:
 ERROR: Copying message 30624 [acc: Anarcat]
  decoding with 'X-EUC-TW' codec failed (AttributeError: 'memoryview' object has no attribute 'decode')
Thread 'Copy message from RemoteAnarcat:junk' terminated with exception:
Traceback (most recent call last):
  File "/usr/share/offlineimap3/offlineimap/imaputil.py", line 406, in utf7m_decode
    for c in binary.decode():
AttributeError: 'memoryview' object has no attribute 'decode'
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
  File "/usr/share/offlineimap3/offlineimap/threadutil.py", line 146, in run
    Thread.run(self)
  File "/usr/lib/python3.9/threading.py", line 892, in run
    self._target(*self._args, **self._kwargs)
  File "/usr/share/offlineimap3/offlineimap/folder/Base.py", line 802, in copymessageto
    message = self.getmessage(uid)
  File "/usr/share/offlineimap3/offlineimap/folder/IMAP.py", line 342, in getmessage
    data = self._fetch_from_imap(str(uid), self.retrycount)
  File "/usr/share/offlineimap3/offlineimap/folder/IMAP.py", line 908, in _fetch_from_imap
    ndata1 = self.parser['8bit-RFC'].parsebytes(data[0][1])
  File "/usr/lib/python3.9/email/parser.py", line 123, in parsebytes
    return self.parser.parsestr(text, headersonly)
  File "/usr/lib/python3.9/email/parser.py", line 67, in parsestr
    return self.parse(StringIO(text), headersonly=headersonly)
  File "/usr/lib/python3.9/email/parser.py", line 56, in parse
    feedparser.feed(data)
  File "/usr/lib/python3.9/email/feedparser.py", line 176, in feed
    self._call_parse()
  File "/usr/lib/python3.9/email/feedparser.py", line 180, in _call_parse
    self._parse()
  File "/usr/lib/python3.9/email/feedparser.py", line 385, in _parsegen
    for retval in self._parsegen():
  File "/usr/lib/python3.9/email/feedparser.py", line 298, in _parsegen
    for retval in self._parsegen():
  File "/usr/lib/python3.9/email/feedparser.py", line 385, in _parsegen
    for retval in self._parsegen():
  File "/usr/lib/python3.9/email/feedparser.py", line 256, in _parsegen
    if self._cur.get_content_type() == 'message/delivery-status':
  File "/usr/lib/python3.9/email/message.py", line 578, in get_content_type
    value = self.get('content-type', missing)
  File "/usr/lib/python3.9/email/message.py", line 471, in get
    return self.policy.header_fetch_parse(k, v)
  File "/usr/lib/python3.9/email/policy.py", line 163, in header_fetch_parse
    return self.header_factory(name, value)
  File "/usr/lib/python3.9/email/headerregistry.py", line 601, in __call__
    return self[name](name, value)
  File "/usr/lib/python3.9/email/headerregistry.py", line 196, in __new__
    cls.parse(value, kwds)
  File "/usr/lib/python3.9/email/headerregistry.py", line 445, in parse
    kwds['parse_tree'] = parse_tree = cls.value_parser(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 2675, in parse_content_type_header
    ctype.append(parse_mime_parameters(value[1:]))
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 2569, in parse_mime_parameters
    token, value = get_parameter(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 2492, in get_parameter
    token, value = get_value(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 2403, in get_value
    token, value = get_quoted_string(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 1294, in get_quoted_string
    token, value = get_bare_quoted_string(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 1223, in get_bare_quoted_string
    token, value = get_encoded_word(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 1064, in get_encoded_word
    text, charset, lang, defects = _ew.decode('=?' + tok + '?=')
  File "/usr/lib/python3.9/email/_encoded_words.py", line 181, in decode
    string = bstring.decode(charset)
AttributeError: decoding with 'X-EUC-TW' codec failed (AttributeError: 'memoryview' object has no attribute 'decode')
Last 1 debug messages logged for Copy message from RemoteAnarcat:junk prior to exception:
thread: Register new thread 'Copy message from RemoteAnarcat:junk' (account 'Anarcat')
ERROR: Exceptions occurred during the run!
ERROR: Copying message 30624 [acc: Anarcat]
  decoding with 'X-EUC-TW' codec failed (AttributeError: 'memoryview' object has no attribute 'decode')
Traceback:
  File "/usr/share/offlineimap3/offlineimap/folder/Base.py", line 802, in copymessageto
    message = self.getmessage(uid)
  File "/usr/share/offlineimap3/offlineimap/folder/IMAP.py", line 342, in getmessage
    data = self._fetch_from_imap(str(uid), self.retrycount)
  File "/usr/share/offlineimap3/offlineimap/folder/IMAP.py", line 908, in _fetch_from_imap
    ndata1 = self.parser['8bit-RFC'].parsebytes(data[0][1])
  File "/usr/lib/python3.9/email/parser.py", line 123, in parsebytes
    return self.parser.parsestr(text, headersonly)
  File "/usr/lib/python3.9/email/parser.py", line 67, in parsestr
    return self.parse(StringIO(text), headersonly=headersonly)
  File "/usr/lib/python3.9/email/parser.py", line 56, in parse
    feedparser.feed(data)
  File "/usr/lib/python3.9/email/feedparser.py", line 176, in feed
    self._call_parse()
  File "/usr/lib/python3.9/email/feedparser.py", line 180, in _call_parse
    self._parse()
  File "/usr/lib/python3.9/email/feedparser.py", line 385, in _parsegen
    for retval in self._parsegen():
  File "/usr/lib/python3.9/email/feedparser.py", line 298, in _parsegen
    for retval in self._parsegen():
  File "/usr/lib/python3.9/email/feedparser.py", line 385, in _parsegen
    for retval in self._parsegen():
  File "/usr/lib/python3.9/email/feedparser.py", line 256, in _parsegen
    if self._cur.get_content_type() == 'message/delivery-status':
  File "/usr/lib/python3.9/email/message.py", line 578, in get_content_type
    value = self.get('content-type', missing)
  File "/usr/lib/python3.9/email/message.py", line 471, in get
    return self.policy.header_fetch_parse(k, v)
  File "/usr/lib/python3.9/email/policy.py", line 163, in header_fetch_parse
    return self.header_factory(name, value)
  File "/usr/lib/python3.9/email/headerregistry.py", line 601, in __call__
    return self[name](name, value)
  File "/usr/lib/python3.9/email/headerregistry.py", line 196, in __new__
    cls.parse(value, kwds)
  File "/usr/lib/python3.9/email/headerregistry.py", line 445, in parse
    kwds['parse_tree'] = parse_tree = cls.value_parser(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 2675, in parse_content_type_header
    ctype.append(parse_mime_parameters(value[1:]))
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 2569, in parse_mime_parameters
    token, value = get_parameter(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 2492, in get_parameter
    token, value = get_value(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 2403, in get_value
    token, value = get_quoted_string(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 1294, in get_quoted_string
    token, value = get_bare_quoted_string(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 1223, in get_bare_quoted_string
    token, value = get_encoded_word(value)
  File "/usr/lib/python3.9/email/_header_value_parser.py", line 1064, in get_encoded_word
    text, charset, lang, defects = _ew.decode('=?' + tok + '?=')
  File "/usr/lib/python3.9/email/_encoded_words.py", line 181, in decode
    string = bstring.decode(charset)
Folder junk [acc: Anarcat]:
 Copy message UID 30626 (29008/49310) RemoteAnarcat:junk -> LocalAnarcat:junk
Command exited with non-zero status 100
5252.91user 535.86system 3:21:00elapsed 47%CPU (0avgtext+0avgdata 846304maxresident)k
96344inputs+26563792outputs (1189major+2155815minor)pagefaults 0swaps
That only transferred about 8GB of mail, which gives us a transfer rate of 5.3Mbit/s, more than 5 times slower than mbsync. This bug is possibly limited to the bullseye version of offlineimap3 (the lovely 0.0~git20210225.1e7ef9e+dfsg-4), while the current sid version (the equally gorgeous 0.0~git20211018.e64c254+dfsg-1) seems unaffected.

Tolerable performance The new release still crashes, except it does so at the very end, which is an improvement, since the mails do get transferred: 
 *** Finished account 'Anarcat' in 511:12
ERROR: Exceptions occurred during the run!
ERROR: Exception parsing message with ID (<20190619152034.BFB8810E07A@marcos.anarc.at>) from imaplib (response type: bytes).
 AttributeError: decoding with 'X-EUC-TW' codec failed (AttributeError: 'memoryview' object has no attribute 'decode')
Traceback:
  File "/usr/share/offlineimap3/offlineimap/folder/Base.py", line 810, in copymessageto
    message = self.getmessage(uid)
  File "/usr/share/offlineimap3/offlineimap/folder/IMAP.py", line 343, in getmessage
    data = self._fetch_from_imap(str(uid), self.retrycount)
  File "/usr/share/offlineimap3/offlineimap/folder/IMAP.py", line 910, in _fetch_from_imap
    raise OfflineImapError(
ERROR: Exception parsing message with ID (<40A270DB.9090609@alternatives.ca>) from imaplib (response type: bytes).
 AttributeError: decoding with 'x-mac-roman' codec failed (AttributeError: 'memoryview' object has no attribute 'decode')
Traceback:
  File "/usr/share/offlineimap3/offlineimap/folder/Base.py", line 810, in copymessageto
    message = self.getmessage(uid)
  File "/usr/share/offlineimap3/offlineimap/folder/IMAP.py", line 343, in getmessage
    data = self._fetch_from_imap(str(uid), self.retrycount)
  File "/usr/share/offlineimap3/offlineimap/folder/IMAP.py", line 910, in _fetch_from_imap
    raise OfflineImapError(
ERROR: IMAP server 'RemoteAnarcat' does not have a message with UID '32686'
Traceback:
  File "/usr/share/offlineimap3/offlineimap/folder/Base.py", line 810, in copymessageto
    message = self.getmessage(uid)
  File "/usr/share/offlineimap3/offlineimap/folder/IMAP.py", line 343, in getmessage
    data = self._fetch_from_imap(str(uid), self.retrycount)
  File "/usr/share/offlineimap3/offlineimap/folder/IMAP.py", line 889, in _fetch_from_imap
    raise OfflineImapError(reason, severity)
Command exited with non-zero status 1
8273.52user 983.80system 8:31:12elapsed 30%CPU (0avgtext+0avgdata 841936maxresident)k
56376inputs+43247608outputs (811major+4972914minor)pagefaults 0swaps
"offlineimap  -o " took 8 hours 31 mins 15 secs
This is 8h31m for transferring 12G, which is around 3.1Mbit/s. That is nine times slower than mbsync, almost an order of magnitude! Now that we have a full sync, we can test incremental synchronization. That is also much slower: 
===> multitime results
1: sh -c "offlineimap -o   true"
            Mean        Std.Dev.    Min         Median      Max
real        24.639      0.513       23.946      24.526      25.708      
user        23.912      0.473       23.404      23.795      24.947      
sys         1.743       0.105       1.607       1.729       2.002
That is also an order of magnitude slower than mbsync, and significantly slower than what you'd expect from a sync process. ~30 seconds is long enough to make me impatient and distracted; 3 seconds, less so: I can wait and see the results almost immediately.

Integrity check That said: this is still on a gigabit link. It's technically possible that OfflineIMAP performs better than mbsync over a slow link, but I Haven't tested that theory. The OfflineIMAP mail spool is missing quite a few messages as well: 
anarcat@angela:~(main)$ find Maildir-offlineimap -type f -type f -a \! -name '.*'   wc -l 
381463
anarcat@angela:~(main)$ find Maildir -type f -type f -a \! -name '.*'   wc -l 
385247
... although that's probably all either new messages or the register folder, so OfflineIMAP might actually be in a better position there. But digging in more, it seems like the actual per-folder diff is fairly similar to mbsync: a few messages missing here and there. Considering OfflineIMAP's instability and poor performance, I have not looked any deeper in those discrepancies.

Other projects to evaluate Those are all the options I have considered, in alphabetical order
  • doveadm-sync: requires dovecot on both ends, can tunnel over SSH, may have performance issues in incremental sync, written in C
  • fdm: fetchmail replacement, IMAP/POP3/stdin/Maildir/mbox,NNTP support, SOCKS support (for Tor), complex rules for delivering to specific mailboxes, adding headers, piping to commands, etc. discarded because no (real) support for keeping mail on the server, and written in C
  • getmail: fetchmail replacement, IMAP/POP3 support, supports incremental runs, classification rules, Python
  • interimap: syncs two IMAP servers, apparently faster than doveadm and offlineimap, but requires running an IMAP server locally, Perl
  • isync/mbsync: TLS client certs and SSH tunnels, fast, incremental, IMAP/POP/Maildir support, multiple mailbox, trash and recursion support, and generally has good words from multiple Debian and notmuch people (Arch tutorial), written in C, review above
  • mail-sync: notify support, happens over any piped transport (e.g. ssh), diff/patch system, requires binary on both ends, mentions UUCP in the manpage, mentions rsmtp which is a nice name for rsendmail. not evaluated because it seems awfully complex to setup, Haskell
  • nncp: treat the local spool as another mail server, not really compatible with my "multiple clients" setup, Golang
  • offlineimap3: requires IMAP, used the py2 version in the past, might just still work, first sync painful (IIRC), ways to tunnel over SSH, review above, Python
Most projects were not evaluated due to lack of time.

Conclusion I'm now using mbsync to sync my mail. I'm a little disappointed by the synchronisation times over the slow link, but I guess that's on par for the course if we use IMAP. We are bound by the network speed much more than with custom protocols. I'm also worried about the C implementation and the crashes I have witnessed, but I am encouraged by the fast upstream response. Time will tell if I will stick with that setup. I'm certainly curious about the promises of interimap and mail-sync, but I have ran out of time on this project.

Antoine Beaupr : The last syncmaildir crash

My syncmaildir (SMD) setup failed me one too many times (previously, previously). In an attempt to migrate to an alternative mail synchronization tool, I looked into using my IMAP server again, and found out my mail spool was in a pretty bad shape. I'm comparing mbsync and offlineimap in the next post but this post talks about how I recovered the mail spool so that tools like those could correctly synchronise the mail spool again.

The latest crash On Monday, SMD just started failing with this error: 
nov 15 16:12:19 angela systemd[2305]: Starting pull emails with syncmaildir...
nov 15 16:12:22 angela systemd[2305]: smd-pull.service: Succeeded.
nov 15 16:12:22 angela systemd[2305]: Finished pull emails with syncmaildir.
nov 15 16:14:08 angela systemd[2305]: Starting pull emails with syncmaildir...
nov 15 16:14:11 angela systemd[2305]: smd-pull.service: Main process exited, code=exited, status=1/FAILURE
nov 15 16:14:11 angela systemd[2305]: smd-pull.service: Failed with result 'exit-code'.
nov 15 16:14:11 angela systemd[2305]: Failed to start pull emails with syncmaildir.
nov 15 16:16:14 angela systemd[2305]: Starting pull emails with syncmaildir...
nov 15 16:16:17 angela smd-pull[27178]: smd-client: ERROR: Network error.
nov 15 16:16:17 angela smd-pull[27178]: smd-client: ERROR: Unable to get any data from the other endpoint.
nov 15 16:16:17 angela smd-pull[27178]: smd-client: ERROR: This problem may be transient, please retry.
nov 15 16:16:17 angela smd-pull[27178]: smd-client: ERROR: Hint: did you correctly setup the SERVERNAME variable
nov 15 16:16:17 angela smd-pull[27178]: smd-client: ERROR: on your client? Did you add an entry for it in your ssh
nov 15 16:16:17 angela smd-pull[27178]: smd-client: ERROR: configuration file?
nov 15 16:16:17 angela smd-pull[27178]: smd-client: ERROR: Network error
nov 15 16:16:17 angela smd-pull[27188]: register: smd-client@localhost: TAGS: error::context(handshake) probable-cause(network) human-intervention(avoidable) suggested-actions(retry)
nov 15 16:16:17 angela systemd[2305]: smd-pull.service: Main process exited, code=exited, status=1/FAILURE
nov 15 16:16:17 angela systemd[2305]: smd-pull.service: Failed with result 'exit-code'.
nov 15 16:16:17 angela systemd[2305]: Failed to start pull emails with syncmaildir.
What is frustrating is that there's actually no network error here. Running the command by hand I did see a different message, but now I have lost it in my backlog. It had something to do with a filename being too long, and I gave up debugging after a while. This happened suddenly too, which added to the confusion. In a fit of rage I started this blog post and experimenting with alternatives, which led me down a lot of rabbit holes. Reviewing my previous mail crash documentation, it seems most solutions involve talking to an IMAP server, so I figured I would just do that. Wanting to try something new, i gave isync (AKA mbsync) a try. Oh dear, I did not expect how much trouble just talking to my IMAP server would be, which wasn't not isync's fault, for what that's worth. It was the primary tool I used to debug things, and served me well in that regard.

Mailbox corruption The first thing I found out is that certain messages in the IMAP spool were corrupted. mbsync would stop on a FETCH command and Dovecot would give me those errors on the server side.

"wrong W value" 
nov 16 15:31:27 marcos dovecot[3621800]: imap(anarcat)<3630489><wAmSzO3QZtfAqAB1>: Error: Mailbox junk: Maildir filename has wrong W value, renamed the file from /home/anarcat/Maildir/.junk/cur/1454623938.M101164P22216.marcos,S=2495,W=2578:2,S to /home/anarcat/Maildir/.junk/cur/1454623938.M101164P22216.marcos,S=2495:2,S
nov 16 15:31:27 marcos dovecot[3621800]: imap(anarcat)<3630489><wAmSzO3QZtfAqAB1>: Error: Mailbox junk: Deleting corrupted cache record uid=1582: UID 1582: Broken virtual size in mailbox junk: read(/home/anarcat/Maildir/.junk/cur/1454623938.M101164P22216.marcos,S=2495,W=2578:2,S): FETCH BODY[] got too little data: 2540 vs 2578
At least this first error was automatically healed by Dovecot (by renaming the file without the W= flag). The problem is that the FETCH command fails and mbsync exits noisily. So you need to constantly restart mbsync with a silly command like: 
while ! mbsync -a; do sleep 1; done

"cached message size larger than expected" 
nov 16 13:53:08 marcos dovecot[3520770]: imap(anarcat)<3594402><M5JHb+zQ3NLAqAB1>: Error: Mailbox Sent: UID=19288: read(/home/anarcat/Maildir/.Sent/cur/1224790447.M898726P9811V000000000000FE06I00794FB1_0.marvin,S=2588:2,S) failed: Cached message size larger than expected (2588 > 2482, box=Sent, UID=19288) (read reason=mail stream)
nov 16 13:53:08 marcos dovecot[3520770]: imap(anarcat)<3594402><M5JHb+zQ3NLAqAB1>: Error: Mailbox Sent: Deleting corrupted cache record uid=19288: UID 19288: Broken physical size in mailbox Sent: read(/home/anarcat/Maildir/.Sent/cur/1224790447.M898726P9811V000000000000FE06I00794FB1_0.marvin,S=2588:2,S) failed: Cached message size larger than expected (2588 > 2482, box=Sent, UID=19288)
nov 16 13:53:08 marcos dovecot[3520770]: imap(anarcat)<3594402><M5JHb+zQ3NLAqAB1>: Error: Mailbox Sent: UID=19288: read(/home/anarcat/Maildir/.Sent/cur/1224790447.M898726P9811V000000000000FE06I00794FB1_0.marvin,S=2588:2,S) failed: Cached message size larger than expected (2588 > 2482, box=Sent, UID=19288) (read reason=)
nov 16 13:53:08 marcos dovecot[3520770]: imap-login: Panic: epoll_ctl(del, 7) failed: Bad file descriptor
This second problem is much harder to fix, because dovecot does not recover automatically. This is Dovecot complaining that the cached size (the S= field, but also present in Dovecot's metadata files) doesn't match the file size. I wonder if at least some of those messages were corrupted in the OfflineIMAP to syncmaildir migration because part of that procedure is to run the strip_header script to remove content from the emails. That could easily have broken things since the files do not also get renamed.

Workaround So I read a lot of the Dovecot documentation on the maildir format, and wrote an extensive fix script for those two errors. The script worked and mbsync was able to sync the entire mail spool. And no, rebuilding the index files didn't work. Also tried doveadm force-resync -u anarcat which didn't do anything. In the end I also had to do this, because the wrong cache values were also stored elsewhere. 
service dovecot stop ; find -name 'dovecot*' -delete; service dovecot start
This would have totally broken any existing clients, but thankfully I'm starting from scratch (except maybe webmail, but I'm hoping it will self-heal as well, assuming it only has a cache and not a full replica of the mail spool).

Incoherence between Maildir and IMAP Unfortunately, the first mbsync was incomplete as it was missing about 15,000 mails: 
anarcat@angela:~(main)$ find Maildir -type f -type f -a \! -name '.*'   wc -l 
384836
anarcat@angela:~(main)$ find Maildir-mbsync/ -type f -a \! -name '.*'   wc -l 
369221
As it turns out, mbsync was not at fault here either: this was yet more mail spool corruption. It's actually 26 folders (out of 205) with inconsistent sizes, which can be found with: 
for folder in * .[^.]* ; do 
  printf "%s\t%d\n" $folder $(find "$folder" -type f -a \! -name '.*'   wc -l );
done
The special \! -name '.*' bit is to ignore the mbsync metadata, which creates .uidvalidity and .mbsyncstate in every folder. That ignores about 200 files but since they are spread around all folders, which was making it impossible to review where the problem was. Here is what the diff looks like: 
--- Maildir-list    2021-11-17 20:42:36.504246752 -0500
+++ Maildir-mbsync-list 2021-11-17 20:18:07.731806601 -0500
@@ -6,16 +6,15 @@
[...]
 .Archives  1
 .Archives.2010 3553
-.Archives.2011 3583
-.Archives.2012 12593
+.Archives.2011 3582
+.Archives.2012 620
 .Archives.2013 8576
 .Archives.2014 11057
-.Archives.2015 8173
+.Archives.2015 8165
 .Archives.2016 54
 .band  34
 .bitbuck   1
@@ -38,13 +37,12 @@
 .couchsurfers  2
-cur    11285
+cur    11280
 .current   130
 .cv    2
 .debbug    262
-.debian    37544
-drafts 1
-.Drafts    4
+.debian    37533
+.Drafts    2
 .drone 241
 .drupal    188
 .drupal-devel  303
[...]

Misfiled messages It's a bit all over the place, but we can already notice some huge differences between mailboxes, for example in the Archives folders. As it turns out, at least 12,000 of those missing mails were actually misfiled: instead of being in the Maildir/.Archives.2012/cur/ folder, they were directly in Maildir/.Archives.2012/. This is something that doesn't matter for SMD (and possibly for notmuch? it does matter, notmuch suddenly found 12,000 new mails) but that definitely matters to Dovecot and therefore mbsync... After moving those files around, we still have 4,000 message missing: 
anarcat@angela:~(main)$ find Maildir-mbsync/  -type f -a \! -name '.*'   wc -l 
381196
anarcat@angela:~(main)$ find Maildir/  -type f -a \! -name '.*'   wc -l 
385053
The problem is that those 4,000 missing mails are harder to track. Take, for example, .Archives.2011, which has a single message missing, out of 3,582. And the files are not identical: the checksums don't match after going through the IMAP transport, so we can't use a tool like hashdeep to compare the trees and find why any single file is missing.

"register" folder One big chunk of the 4,000, however, is a special folder called register in my spool, which I am syncing separately (see Securing registration email for details on that setup). That actually covers 3,700 of those messages, so I actually have a more modest 300 messages to figure out, after (easily!) configuring mbsync to sync that folder separately: 
 @@ -30,9 +33,29 @@ Slave :anarcat-local:
  # Exclude everything under the internal [Gmail] folder, except the interesting folders
  #Patterns * ![Gmail]* "[Gmail]/Sent Mail" "[Gmail]/Starred" "[Gmail]/All Mail"
  # Or include everything
 -Patterns *
 +#Patterns *
 +Patterns * !register  !.register
  # Automatically create missing mailboxes, both locally and on the server
  #Create Both
  Create slave
  # Sync the movement of messages between folders and deletions, add after making sure the sync works
  #Expunge Both
 +
 +IMAPAccount anarcat-register
 +Host imap.anarc.at
 +User register
 +PassCmd "pass imap.anarc.at-register"
 +SSLType IMAPS
 +CertificateFile /etc/ssl/certs/ca-certificates.crt
 +
 +IMAPStore anarcat-register-remote
 +Account anarcat-register
 +
 +MaildirStore anarcat-register-local
 +SubFolders Maildir++
 +Inbox ~/Maildir-mbsync/.register/
 +
 +Channel anarcat-register
 +Master :anarcat-register-remote:
 +Slave :anarcat-register-local:
 +Create slave

"tmp" folders and empty messages After syncing the "register" messages, I end up with the measly little 160 emails out of sync: 
anarcat@angela:~(main)$ find Maildir-mbsync/  -type f -a \! -name '.*'   wc -l 
384900
anarcat@angela:~(main)$ find Maildir/  -type f -a \! -name '.*'   wc -l 
385059
Argh. After more digging, I have found 131 mails in the tmp/ directories of the client's mail spool. Mysterious! On the server side, it's even more files, and not the same ones. Possible that those were mails that were left there during a failed delivery of some sort, during a power failure or some sort of crash? Who knows. It could be another race condition in SMD if it runs while mail is being delivered in tmp/... The first thing to do with those is to cleanup a bunch of empty files (21 on angela): 
find .[^.]*/tmp -type f -empty -delete
As it turns out, they are all duplicates, in the sense that notmuch can easily find a copy of files with the same message ID in its database. In other words, this hairy command returns nothing 
find .[^.]*/tmp -type f   while read path; do
  msgid=$(grep -m 1  -i ^message-id "$path"   sed 's/Message-ID: //i;s/[<>]//g');
  if notmuch count --exclude=false  "id:$msgid"   grep -q 0; then
    echo "$path <$msgid> not in notmuch" ;
  fi;
done
... which is good. Or, to put it another way, this is safe: 
find .[^.]*/tmp -type f -delete
Poof! 314 mails cleaned on the server side. Interestingly, SMD doesn't pick up on those changes at all and still sees files in tmp/ directories on the client side, so we need to operate the same twisted logic there.

notmuch to the rescue again After cleaning that on the client, we get: 
anarcat@angela:~(main)$ find Maildir/  -type f -a \! -name '.*'   wc -l 
384928
anarcat@angela:~(main)$ find Maildir-mbsync/  -type f -a \! -name '.*'   wc -l 
384901
Ha! 27 mails difference. Those are the really sticky, unclear ones. I was hoping a full sync might clear that up, but after deleting the entire directory and starting from scratch, I end up with: 
anarcat@angela:~(main)$ find Maildir -type f -type f -a \! -name '.*'   wc -l 
385034
anarcat@angela:~(main)$ find Maildir-mbsync -type f -type f -a \! -name '.*'   wc -l 
384993
That is: even more messages missing (now 37). Sigh. Thankfully, this is something notmuch can help with: it can index all files by Message-ID (which I learned is case-insensitive, yay) and tell us which messages don't make it through. Considering the corruption I found in the mail spool, I wouldn't be the least surprised those messages are just skipped by the IMAP server. Unfortunately, there's nothing on the Dovecot server logs that would explain the discrepancy. Here again, notmuch comes to the rescue. We can list all message IDs to figure out that discrepancy: 
notmuch search --exclude=false --output=messages '*'   pv -s 18M   sort > Maildir-msgids
notmuch --config=.notmuch-config-mbsync search --exclude=false --output=messages '*'   pv -s 18M   sort > Maildir-mbsync-msgids
And then we can see how many messages notmuch thinks are missing: 
$ wc -l *msgids
372723 Maildir-mbsync-msgids
372752 Maildir-msgids
That's 29 messages. Oddly, it doesn't exactly match the find output: 
anarcat@angela:~(main)$ find Maildir-mbsync -type f -type f -a \! -name '.*'   wc -l 
385204
anarcat@angela:~(main)$ find Maildir -type f -type f -a \! -name '.*'   wc -l 
385241
That is 10 more messages. Ugh. But actually, I know what those are: more misfiled messages (in a .folder/draft/ directory, bizarrely, so the totals actually match. In the notmuch output, there's a lot of stuff like this: 
id:notmuch-sha1-fb880d673e24f5dae71b6b4d825d4a0d5d01cde4
Those are messages without a valid Message-ID. Notmuch (presumably) constructs one based on the file's checksum. Because the files differ between the IMAP server and the local mail spool (which is unfortunate, but possibly inevitable), those do not match. There are exactly the same number of those on both sides, so I'll go ahead and assume those are all accounted for. What remains is: 
anarcat@angela:~(main)$ diff -u Maildir-mbsync-msgids Maildir-msgids    grep '^\-[^-]'   grep -v sha1   wc -l 
2
anarcat@angela:~(main)$ diff -u Maildir-mbsync-msgids Maildir-msgids    grep '^\+[^+]'   grep -v sha1   wc -l 
21
anarcat@angela:~(main)$
ie. 21 missing from mbsync, and, surprisingly, 2 missing from the original mail spool. Further inspection also showed they were all messages with some sort of "corruption": no body and only headers. I am not sure that is a legal email format in the first place. Since they were mostly spam or administrative emails ("You have been unsubscribed from mailing list..."), it seems fairly harmless to ignore those.

Conclusion As we'll see in the next article, SMD has stellar performance. But that comes at a huge cost: it accesses the mail storage directly. This can (and has) created significant problems on the mail server. It's unclear exactly why those things happen, but Dovecot expects a particular storage format on its file, and it seems unwise to bypass that. In the future, I'll try to remember to avoid that, especially since mechanisms like SMD require special server access (SSH) which, in the long term, I am not sure I want to maintain or expect. In other words, just talking with an IMAP server opens up a lot more possibilities of hosting than setting up a custom synchronisation protocol over SSH. It's also safer and more reliable, as we have seen. Thankfully, I've been able to recover from all the errors I could find, but it could have gone differently and it would have been possible for SMD to permanently corrupt significant part of my mail archives. In the end, however, the last drop was just another weird bug which, ironically, SMD mysteriously recovered from on its own while I was writing this documentation and migrating away from it. In any case, I recommend SMD users start looking for alternatives. The project has been archived upstream, and the Debian package has been orphaned. I have seen significant mail box corruption, including entire mail spool destruction, mostly due to incorrect locking code. I have filed a release-critical bug in Debian to make sure it doesn't ship with Debian bookworm. Alternatives like mbsync provide fast and reliable transport, including over SSH. See the next article for further discussion of the alternatives.

19 November 2021

Dirk Eddelbuettel: RcppArmadillo 0.10.7.3.0 on CRAN: Bugfix, New Features

armadillo image Armadillo is a powerful and expressive C++ template library for linear algebra aiming towards a good balance between speed and ease of use with a syntax deliberately close to a Matlab. RcppArmadillo integrates this library with the R environment and language and is widely used by (currently) 928 other packages on CRAN. I somehow missed to blog and tweet about the recent release based on the Armadillo 10.7.3 upstream release. Conrad is in long-term support mode , and 10.7.* is meant to provide fixes and stability relative to the most recent release which we did on September 30. We did actually find a regression when checking reverse-dependencies requiring an upstream move to 10.7.3. At the same time, we folded pull request #352 in. It addresses an old bug of ours where Armadillo fields types were not converted correctly in all dimensions. So no we do have 0.10.7.3.0 on CRAN, as well as 0.10.7.3.1 with the (opt-in) field) fixes on the drat repo) repo. As the change perturbs a few existing packages, it is opt-in for now. We will likely aim at a proper deprecation of the old behaviour and give packages time to adjust. Stay tuned. With that, big thanks to Jonathan Berrisch for filing issue #351 and basically addressing it in pull request #352. Very nice work, and I basically just wrapped a few more tests around it. The full set of changes (since the last CRAN release 0.10.7.0.) follows.

Changes in RcppArmadillo version 0.10.7.3.1 (2021-11-18)
  • Correct dimensions setting in import/export of arma::field types, protected by #define (Jonathan Berrisch in #352 fixing #351)
  • Add unit tests for fields both with and without new #define (Dirk)

Changes in RcppArmadillo version 0.10.7.3.0 (2021-11-04)
  • Upgraded to Armadillo release 10.7.3
    • fix regression in alias handling by fliplr(), flipud(), reverse()

Changes in RcppArmadillo version 0.10.7.2.0 (2021-11-02)
  • Upgraded to Armadillo release 10.7.2
    • more robust handling of diagonal matrices by pinv()

Changes in RcppArmadillo version 0.10.7.1.0 (2021-10-08)
  • Upgraded to Armadillo release 10.7.1
    • fix regression in interactions between dense matrix subviews and sparse matrices

Courtesy of my CRANberries, there is a diffstat report relative to previous release. More detailed information is on the RcppArmadillo page. Questions, comments etc should go to the rcpp-devel mailing list off the R-Forge page. If you like this or other open-source work I do, you can sponsor me at GitHub.

This post by Dirk Eddelbuettel originated on his Thinking inside the box blog. Please report excessive re-aggregation in third-party for-profit settings.

15 November 2021

Vincent Bernat: Git as a source of truth for network automation

The first step when automating a network is to build the source of truth. A source of truth is a repository of data that provides the intended state: the list of devices, the IP addresses, the network protocols settings, the time servers, etc. A popular choice is NetBox. Its documentation highlights its usage as a source of truth:
NetBox intends to represent the desired state of a network versus its operational state. As such, automated import of live network state is strongly discouraged. All data created in NetBox should first be vetted by a human to ensure its integrity. NetBox can then be used to populate monitoring and provisioning systems with a high degree of confidence.
When introducing Jerikan, a common feedback we got was: you should use NetBox for this. Indeed, Jerikan s source of truth is a bunch of YAML files versioned with Git.

Why Git? If we look at how things are done with servers and services, in a datacenter or in the cloud, we are likely to find users of Terraform, a tool turning declarative configuration files into infrastructure. Declarative configuration management tools like Salt, Puppet,1 or Ansible take care of server configuration. NixOS is an alternative: it combines package management and configuration management with a functional language to build virtual machines and containers. When using a Kubernetes cluster, people use Kustomize or Helm, two other declarative configuration management tools. Tapped together, these tools implement the infrastructure as code paradigm.
Infrastructure as code is an approach to infrastructure automation based on practices from software development. It emphasizes consistent, repeatable routines for provisioning and changing systems and their configuration. You make changes to code, then use automation to test and apply those changes to your systems. Kief Morris, Infrastructure as Code, O Reilly.
A version control system is a central tool for infrastructure as code. The usual candidate is Git with a source code management system like GitLab or GitHub. You get:
Traceability and visibility
Git keeps a log of all changes: what, who, why, and when. With a bit of discipline, each change is explained and self-contained. It becomes part of the infrastructure documentation. When the support team complains about a degraded experience for some customers over the last two months or so, you quickly discover this may be related to a change to an incoming policy in New York.
Rolling back
If a change is defective, it can be reverted quickly, safely, and without much effort, even if other changes happened in the meantime. The policy change at the origin of the problem spanned over three routers. Reverting this specific change and deploying the configuration let you solve the situation until you find a better fix.
Branching, reviewing, merging
When working on a new feature or refactoring some part of the infrastructure, a team member creates a branch and works on their change without interfering with the work of other members. Once the branch is ready, a pull request is created and the change is ready to be reviewed by the other team members before merging. You discover the issue was related to diverting traffic through an IX where one ISP was connected without enough capacity. You propose and discuss a fix that includes a change of the schema and the templates used to declare policies to be able to handle this case.
Continuous integration
For each change, automated tests are triggered. They can detect problems and give more details on the effect of a change. Branches can be deployed to a test infrastructure where regression tests are executed. The results can be synthesized as a comment in the pull request to help the review. You check your proposed change does not modify the other existing policies.

Why not NetBox? NetBox does not share these features. It is a database with a REST and a GraphQL API. Traceability is limited: changes are not grouped into a transaction and they are not documented. You cannot fork the database. Usually, there is one staging database to test modifications before applying them to the production database. It does not scale well and reviews are difficult. Applying the same change to the production database can be hazardous. Rolling back a change is non-trivial.

Update (2021-11) Nautobot, a fork of NetBox, will soon address this point by using Dolt, an SQL database engine allowing you to clone, branch, and merge, like a Git repository. Dolt is compatible with MySQL clients. See Nautobots, Roll Back! for a preview of this feature.

Moreover, NetBox is not usually the single source of truth. It contains your hardware inventory, the IP addresses, and some topology information. However, this is not the place you put authorized SSH keys, syslog servers, or the BGP configuration. If you also use Ansible, this information ends in its inventory. The source of truth is therefore fragmented between several tools with different workflows. Since NetBox 2.7, you can append additional data with configuration contexts. This mitigates this point. The data is arranged hierarchically but the hierarchy cannot be customized.2 Nautobot can manage configuration contexts in a Git repository, while still allowing to use of the API to fetch them. You get some additional perks, thanks to Git, but the remaining data is still in a database with a different lifecycle. Lastly, the schema used by NetBox may not fit your needs and you cannot tweak it. For example, you may have a rule to compute the IPv6 address from the IPv4 address for dual-stack interfaces. Such a relationship cannot be easily expressed and enforced in NetBox. When changing the IPv4 address, you may forget the IPv6 address. The source of truth should only contain the IPv4 address but you also want the IPv6 address in NetBox because this is your IPAM and you need it to update your DNS entries.

Why not Git? There are some limitations when putting your source of truth in Git:
  1. If you want to expose a web interface to allow an external team to request a change, it is more difficult to do it with Git than with a database. Out-of-the-box, NetBox provides a nice web interface and a permission system. You can also write your own web interface and interact with NetBox through its API.
  2. YAML files are more difficult to query in different ways. For example, looking for a free IP address is complex if they are scattered in multiple places.
In my opinion, in most cases, you are better off putting the source of truth in Git instead of NetBox. You get a lot of perks by doing that and you can still use NetBox as a read-only view, usable by other tools. We do that with an Ansible module. In the remaining cases, Git could still fit the bill. Read-only access control can be done through submodules. Pull requests can restrict write access: a bot can check the changes only modify allowed files before auto-merging. This still requires some Git knowledge, but many teams are now comfortable using Git, thanks to its ubiquity.
  1. Wikimedia manages its infrastructure with Puppet. They publish everything on GitHub. Creative Commons uses Salt. They also publish everything on GitHub. Thanks to them for doing that! I wish I could provide more real-life examples.
  2. Being able to customize the hierarchy is key to avoiding repetition in the data. For example, if switches are paired together, some data should be attached to them as a group and not duplicated on each of them. Tags can be used to partially work around this issue but you lose the hierarchical aspect.

12 November 2021

Jonathan Dowland: Frictionless external backups with systemd

Here's a description of how my monthly external backups are managed at a technical level. I didn't realise I hadn't written this all down anywhere yet. What
blinkenlights! blinkenlights!
I plug in one of two (prepared) external hard drives into my headless NAS. The NAS contains my primary data backup. A job automatically decrypts the encrypted filesystem on the drive, mounts it and synchronises the copy of my backup data on the drive from that on the NAS. Whilst this is going on, the blinkstick LED on the NAS switches to a colour to signal "in progress". When it's done, the light changes to green to signal "done" and I can remove it. If something goes wrong, it turns red and I get mail. Why I want a third-strand, off-site backup of my and my family's data in case of a disaster in our house. For it to be useful it has to be regular, so I needed to remove as much of the friction of performing the backup as possible. I use two drives alternately so that I don't have all my eggs in one basket in the window when I bring one of them home and perform the backup. How As much as possible I lean on systemd and its ability to trigger actions based on events.
  1. External drive is plugged in. systemd instantiates a corresponding device unit, named dev-disk-by\x2duuid-aaaaaaaa\x2daaaa\x2daaaa\x2daaaa\x2daaaaaaaaaaaa.device, where aaaa is the UUID of a partition on the device
  2. The backup job is a systemd service which has a WantedBy relationship on the device unit, so when the device appears, systemd starts the backup service.
  3. The backup service has Requires and After relationships on systemd-cryptsetup@extbackup.service, a service created by systemd's cryptsetup generator on start-up (but slightly customised, see below). The encrypted device is therefore unlocked.
  4. The backup service defines multiple start and stop commands with ExecStart and ExecStop. These are used to:
    1. set the blinkstick to the working colour (blue-ish)
    2. mount the now-decrypted filesystem
    3. get a lock on the backup repository (so nothing else writes to it) and synchronise the files
    4. unmount the filesystem
    5. set the blinkstick to the success colour (green)
  5. Finally, the systemd-cryptsetup@extbackup.service unit realises it is not required any more. It has been customised with StopWhenUnneeded=true1, so the encrypted filesystem is closed, ready for the drive to be removed.
  6. I notice the LED colour is green, remove the drive, and take it to its off-site home.
If anything goes wrong, all my custom systemd units have, as a matter of course, 
OnFailure=status-email-user@%n.service blinkstick-fail.service
Preparing a new backup disk This is mostly just a standard dm-crypt/cryptsetup/LUKS encrypted device, on top of a standard partition on the underlying disk, with a normal filesystem sitting on top: Basically, the most common way to encrypt a drive in Linux. See places like the cryptsetup docs for how to set something like that up. The key things here are The backup service Here's the backup service unit definition in its entirety: 
 [Unit]
 OnFailure=status-email-user@%n.service blinkstick-fail.service
 Requires=systemd-cryptsetup@extbackup.service backup.mount
 After=systemd-cryptsetup@extbackup.service backup.mount
 [Service]
 Type=oneshot
 ExecStart=/usr/local/bin/blinkstick --index 1 --limit 10 --set-color 33c280
 ExecStart=/bin/mount /extbackup
 ExecStart=/home/jon/bin/phobos-backup-monthly
 ExecStop=/bin/umount /extbackup
 ExecStop=/usr/local/bin/blinkstick --index 1 --limit 10 --set-color green
 [Install]
 WantedBy=dev-disk-by\x2duuid-aaaaaaaa\x2daaaa\x2daaaa\x2daaaa\x2daaaaaaaaaaaa.device
The dashes in the UUID in WantedBy= need to be encoded as \x2d and then the slashes from the path bit as dashes. Using dashes to encode slashes is possibly the single most frustrating systemd design decision. Issues Sadly (as detailed in Blinkenlights, part 2) there are 2 some frustrating limitations with trying to handle the mount (and unmount) of the filesystem in systemd-land, so instead, it's done using the traditional mount, umount and fstab. If you can point out any improvements to this approach, please let me know!
  1. I customized mine a while ago by copying the generated service file to a static file, but nowadays I think you could do systemd edit systemd-cryptsetup@extbackup.service to add the StopWhenUnneeded to an override file and not need the rest.
  2. or at least were. It's been a while since I revisited this part.

10 November 2021

Neil Williams: LetsEncrypt with Apache, Gunicorn and Debian Bullseye

This took me far too long to identify and debug, so I'm going to write it up here for my own reference and to possibly help others.
Background Upgrading an old codebase from Python2 on Buster to Python3 ready for Bullseye and from Django1 to Django2 (prepared for Django3). Everything is fine at this stage - the Django test server is happy with HTTP and it gives enough support to do the actual code changes to get to Python3. All well and good so far. The main purpose of this particular code was to support payments, so a chunk of the testing cannot be done without HTTPS, which is where things got awkward. This particular service needs HTTPS using LetsEncrypt and Apache2. To support Django, I typically use Gunicorn. All of this works with HTTP. Moving to HTTPS was easy to test using the default-ssl virtual host that comes with Apache2 in Debian. It's a static page and it worked well with https. The problems all start when trying to use this known-working HTTPS config with the other Apache virtual host to add support for the gunicorn proxy.
Apache reverse proxy AH00898 Error during SSL Handshake with remote server
Investigating Now that I know why this happened, it's easier to see what was happening. At the time, I was swamped in a plethora of options and permutations between the Django HTTPS options and the Apache VirtualHost SSL and proxy commands. Going through all of those took up huge amounts of time, all in the wrong area. In previous configurations using packages in Buster, gunicorn could simply run on http://localhost:8000 and Apache would proxy that as https. In versions in Bullseye, this no longer works and it is that handover from https in Apache to http in the proxy is where it is failing. Apache is using HTTPS because the LetsEncrypt certificates, created using dehydrated, are specified in the VirtualHost configuration. To fix the handshake error, the proxy server needs to know about the certificates created by dehydrated as well.
Gunicorn needs the certificates The clue is in the gunicorn help: 
--keyfile FILE        SSL key file [None]
--certfile FILE       SSL certificate file [None]
The final part of the puzzle is that the certificates created by dehydrated are in a private location: 
drwx------ 2 root root /var/lib/dehydrated/certs/
To test gunicorn, this will mean using sudo but that's just a step towards running gunicorn as a systemd service (when access to the certs will not be a problem). Starting gunicorn using these options shows the proxy now being available at https://localhost:8000 which is a subtle but very important change. 
Environment=LOGLEVEL=DEBUG WORKERS=4 LOGFILE=/var/log/gunicorn/site.log
ExecStart=/usr/bin/gunicorn3 site.wsgi --log-level $LOGLEVEL --log-file $LOGFILE --workers $WORKERS \
--certfile /var/lib/dehydrated/certs/site/cert.pem \
--keyfile /var/lib/dehydrated/certs/site/privkey.pem
The specified locations are symbolic links created by dehydrated to cope with regular updates of the certificates using cron.

Craig Small: Changing Grafana Legends

I m not sure if I just can search Google properly, or this really is just not written down much, but I have had problems with Grafana Legends (I would call them the series labels). The issue is that Grafana queries Prometheus for a time series and you want to display multiple lines, but the time-series labels you get are just not quite right. A simple example is you might be using the black-box exporter to monitor an external TCP port and you would just like the port number separate to display. The default output would look like this: 
probe_duration_seconds instance="example.net:5222",job="blackbox",module="xmpp_banner"  = 0.01
probe_duration_seconds instance="example.net:5269",job="blackbox",module="xmpp_banner"  = 0.01
I can graph the number of seconds that it takes to probe the 5222 and 5269 TCP ports, but my graph legend is going to have the hostname, making it cluttered. I just want the legend to be the port numbers on Grafana. The answer is to use a Prometheus function called label_replace that takes an existing label, applies a regular expression, then puts the result into another label. That s right, regular expressions, and if you get them wrong then the label just doesn t appear.
Perl REGEX Problems courtesy of XKCD
The label_replace documentation is a bit terse, and in my opinion, the order of parameters is messed up, but after a few goes I had what I needed: 
label_replace(probe_duration_seconds module="xmpp_banner" , "port", "$1", "instance", ".*:(.*)")
probe_duration_seconds instance="example.net:5222",job="blackbox",module="xmpp_banner",port="5222" 	0.001
probe_duration_seconds instance="example.net:5269",job="blackbox",module="xmpp_banner",port="5269" 	0.002
The response now has a new label (or field if you like) called port. So what is this function to our data coming from probe_duration_seconds? The function format is: label_replace(value, dst_label, replacement, src_label, regex) So the function does the following:
  1. Evaluate value, which is generally some sort of query such as probe_duration_seconds
  2. Find the required source label src_label, in this example is instance, in this case the values are example.net:5222 and example.net:5269
  3. Apply regular expression regex, for us its .*:(.*) That says skip everying before : then capture/store everything past : . The brackets mean copy what is after the colon and put it in match #1
  4. Make a new label specified in dst_label, for us this is port
  5. Whatever is in replacement goes into dst_label. For this example it is $1 which means match #1 in our regular expression in the label called port.
In short, the function captures everything after the colon in the instance label and puts that into a new label called port. It does this for each value that is returned into the first parameter. This means I can use the port in my Grafana graph Legend and it will show 5222 or 5269 respectively. I have made the Legend TCP port to give the below result, but I could have used port in Grafana Legend and made the result TCP $1 in the label_replace function to get the same result.
Grafana console showing the use of the label_replace function

9 November 2021

Joachim Breitner: How to audit an Internet Computer canister

I was recently called upon by Origyn to audit the source code of some of their Internet Computer canisters ( canisters are services or smart contracts on the Internet Computer), which were written in the Motoko programming language. Both the application model of the Internet Computer as well as Motoko bring with them their own particular pitfalls and possible sources for bugs. So given that I was involved in the creation of both, they reached out to me. In the course of that audit work I collected a list of things to watch out for, and general advice around them. Origyn generously allowed me to share that list here, in the hope that it will be helpful to the wider community.

Inter-canister calls The Internet Computer system provides inter-canister communication that follows the actor model: Inter-canister calls are implemented via two asynchronous messages, one to initiate the call, and one to return the response. Canisters process messages atomically (and roll back upon certain error conditions), but not complete calls. This makes programming with inter-canister calls error-prone. Possible common sources for bugs, vulnerabilities or simply unexpected behavior are:
  • Reading global state before issuing an inter-canister call, and assuming it to still hold when the call comes back.
  • Changing global state before issuing an inter-canister call, changing it again in the response handler, but assuming nothing else changes the state in between (reentrancy).
  • Changing global state before issuing an inter-canister call, and not handling failures correctly, e.g. when the code handling the callback rolls backs.
If you find such pattern in your code, you should analyze if a malicious party can trigger them, and assess the severity that effect These issues apply to all canisters, and are not Motoko-specific.

Rollbacks Even in the absence of inter-canister calls the behavior of rollbacks can be surprising. In particular, rejecting (i.e. throw) does not rollback state changes done before, while trapping (e.g. Debug.trap, assert , out of cycle conditions) does. Therefore, one should check all public update call entry points for unwanted state changes or unwanted rollbacks. In particular, look for methods (or rather, messages, i.e. the code between commit points) where a state change is followed by a throw. This issues apply to all canisters, and are not Motoko-specific, although other CDKs may not turn exceptions into rejects (which don t roll back).

Talking to malicious canisters Talking to untrustworthy canisters can be risky, for the following (likely incomplete) reasons:
  • The other canister can withhold a response. Although the bidirectional messaging paradigm of the Internet Computer was designed to guarantee a response eventually, the other party can busy-loop for as long as they are willing to pay for before responding. Worse, there are ways to deadlock a canister.
  • The other canister can respond with invalidly encoded Candid. This will cause a Motoko-implemented canister to trap in the reply handler, with no easy way to recover. Other CDKs may give you better ways to handle invalid Candid, but even then you will have to worry about Candid cycle bombs that will cause your reply handler to trap.
Many canisters do not even do inter-canister calls, or only call other trustwothy canisters. For the others, the impact of this needs to be carefully assessed.

Canister upgrade: overview For most services it is crucial that canisters can be upgraded reliably. This can be broken down into the following aspects:
  1. Can the canister be upgraded at all?
  2. Will the canister upgrade retain all data?
  3. Can the canister be upgraded promptly?
  4. Is three a recovery plan for when upgrading is not possible?

Canister upgradeability A canister that traps, for whatever reason, in its canister_preupgrade system method is no longer upgradeable. This is a major risk. The canister_preupgrade method of a Motoko canister consists of the developer-written code in any system func preupgrade() block, followed by the system-generated code that serializes the content of any stable var into a binary format, and then copies that to stable memory. Since the Motoko-internal serialization code will first serialize into a scratch space in the main heap, and then copy that to stable memory, canisters with more than 2GB of live data will likely be unupgradeable. But this is unlikely the first limit: The system imposes an instruction limit on upgrading a canister (spanning both canister_preupgrade and canister_postupgrade). This limit is a subnet configuration value, and sepearate (and likely higher) than the normal per-message limit, and not easily determined. If the canister s live data becomes too large to be serialized within this limit, the canister becomes non-upgradeable. This risk cannot be eliminated completely, as long as Motoko and Stable Variables are used. It can be mitigated by appropriate load testing: Install a canister, fill it up with live data, and exercise the upgrade. If this succeeds with a live data set exceeding the expected amount of data by a margin, this risk is probably acceptable. Bonus points for adding functionality that will prevent the canister s live data to increase above a certain size. If this testing is to be done on a local replica, extra care needs to be taken to make sure the local replica actually performs instruction counting and has the same resource limits as the production subnet. An alternative mitigation is to avoid canister_pre_upgrade as much as possible. This means no use of stable var (or restricted to small, fixed-size configuration data). All other data could be
  • mirrored off the canister (possibly off chain), and manually re-hydrated after an upgrade.
  • stored in stable memory manually, during each update call, using the ExperimentalStableMemory API. While this matches what high-assurance Rust canisters (e.g. the Internet Identity) do, This requires manual binary encoding of the data, and is marked experimental, so I cannot recommend this at the moment.
  • not put into a Motoko canister until Motoko has a scalable solution for stable variable (for example keeping them in stable memory permanently, with smart caching in main memory, and thus obliterating the need for pre-upgrade code.)

Data retention on upgrades Obviously, all live data ought to be retained during upgrades. Motoko automatically ensures this for stable var data. But often canisters want to work with their data in a different format (e.g. in objects that are not shared and thus cannot be put in stable vars, such as HashMap or Buffer objects), and thus may follow following idiom: 
stable var fooStable =  ;
var foo = fooFromStable(fooStable);
system func preupgrade()   fooStable := fooToStable(foo);  )
system func postupgrade()   fooStable := (empty);  )
In this case, it is important to check that
  • All non-stable global vars, or global lets with mutable values, have a stable companion.
  • The assignments to foo and fooStable are not forgotten.
  • The fooToStable and fooFromStable form bijections.
An example would be HashMaps stored as arrays via Iter.toArray( .entries()) and HashMap.fromIter( .vals()). It is worth pointiong out that a code view will only look at a single version of the code, but cannot check whether code changes will preserve data on upgrade. This can easily go wrong if the names and types of stable variables are changed in incompatible way. The upgrade may fail loudly in this cases, but in bad cases, the upgrade may even succeed, losing data along the way. This risk needs to be mitigated by thorough testing, and possibly backups (see below).

Prompt upgrades Motoko and Rust canisters cannot be safely upgraded when they are still waiting for responses to inter-canister calls (the callback would eventually reach the new instance, and because of infelicities of the IC s System API, could possibly call arbitrary internal functions). Therefore, the canister needs to be stopped before upgrading, and started again. If the inter-canister calls take a long time, this mean that upgrading may take a long time, which may be undesirable. Again, this risk is reduced if all calls are made to trustworthy canisters, and elevated when possibly untrustworthy canisters are called, directly or indirectly.

Backup and recovery Because of the above risk around upgrades it is advisable to have a disaster recovery strategy. This could involve off-chain backups of all relevant data, so that it is possible to reinstall (not upgrade) the canister and re-upload all data. Note that reinstall has the same issue as upgrade described above in prompt upgrades : It ought to be stopped first to be safe. Note that the instruction limit for messages, as well as the message size limit, limit the amount of data returned. If the canister needs to hold more data than that, the backup query method might have to return chunks or deltas, with all the extra complexity that entails, e.g. state changes between downloading chunks. If large data load testing is performed (as Irecommend anyways to test upgradeability), one can test whether the backup query method works within the resource limits.

Time is not strictly monotonic The timestamps for current time that the Internet Computer provides to its canisters is guaranteed to be monotonic, but not strictly monotonic. It can return the same values, even in the same messages, as long as they are processed in the same block. It should therefore not be used to detect happens-before relations. Instead of using and comparing time stamps to check whether Y has been performed after X happened last, introduce an explicit var y_done : Bool state, which is set to False by X and then to True by Y. When things become more complex, it will be easier to model that state via an enumeration with speaking tag names, and update this state machine along the way. Another solution to this problem is to introduce a var v : Nat counter that you bump in every update method, and after each await. Now v is your canister s state counter, and can be used like a timestamp in many ways. While we are talking about time: The system time (typically) changes across an await. So if you do let now = Time.now() and then await, the value in now may no longer be what you want.

Wrapping arithmetic The Nat64 data type, and the other fixed-width numeric types provide opt-in wrapping arithmetic (e.g. +%, fromIntWrap). Unless explicitly required by the current application, this should be avoided, as usually a too large or negatie value is a serious, unrecoverable logic error, and trapping is the best one can do.

Cycle balance drain attacks Because of the IC s canister pays model, all canisters are prone to DoS attacks by draining their cycle balance, and this risk needs to be taken into account. The most elementary mitigation strategy is to monitor the cycle balance of canisters and keep it far from the (configurable) freezing threshold. On the raw IC-level, further mitigation strategies are possible:
  • If all update calls are authenticated, perform this authentication as quickly as possible, possibly before decoding the caller s argument. This way, a cycle drain attack by an unauthenticated attacker is less effective (but still possible).
  • Additionally, implementing the canister_inspect_message system method allows the above checks to be performed before the message even is accepted by the Internet Computer. But it does not defend against inter-canister messages and is therefore not a complete solution.
  • If an attack from an authenticated user (e.g. a stakeholder) is to be expected, the above methods are not effective, and an effective defense might require relatively involved additional program logic (e.g. per-caller statistics) to detect such an attack, and react (e.g. rate-limiting).
  • Such defenses are pointless if there is only a single method where they do not apply (e.g. an unauthenticated user registration method). If the application is inherently attackable this way, it is not worth the bother to raise defenses for other methods.
Related: A justification why the Internet Identity does not use canister_inspect_message) A motoko-implemented canister currently cannot perform most of these defenses: Argument decoding happens unconditionally before any user code that may reject a message based on the caller, and canister_inspect_message is not supported. Furthermore, Candid decoding is not very cycle defensive, and one should assume that it is possible to construct Candid messages that require many instructions to decode, even for simple argument type signatures. The conclusion for the audited canisters is to rely on monitoring to keep the cycle blance up, even during an attack, if the expense can be born, and maybe pray for IC-level DoS protections to kick in.

Large data attacks Another DoS attack vector exists if public methods allow untrustworthy users to send data of unlimited size that is persisted in the canister memory. Because of the translation of async-await code into multiple message handlers, this applies not only to data that is obviously stored in global state, but also local data that is live across an await point. The effectiveness of such attacks is limited by the Internet Computer s message size limit, which is in the order of a few megabytes, but many of those also add up. The problem becomes much worse if a method has an argument type that allows a Candid space bomb: It is possible to encode very large vectors with all values null in Candid, so if any method has an argument of type [Null] or [?t], a small message will expand to a large value in the Motoko heap. Other types to watch out:
  • Nat and Int: This is an unbounded natural number, and thus can be arbitrarily large. The Motoko representation will however not be much larger than the Candid encoding (so this does not qualify as a space bomb).
It is still advisable to check if the number is reasonable in size before storing it or doing an await. For example, when it denotes an index in an array, throw early if it exceeds the size of the array; if it denotes a token amount to transfer, check it against the available balance, if it denotes time, check it against reasonable bounds.
  • Principal: A Principal is effectively a Blob. The Interface specification says that principals are at most 29 bytes in length, but the Motoko Candid decoder does not check that currently (fixed in the next version of Motoko). Until then, a Principal passed as an argument can be large (the principal in msg.caller is system-provided and thus safe). If you cannot wait for the fix to reach you, manually check the size of the princial (via Principal.toBlob) before doing the await.

Shadowing of msg or caller Don t use the same name for the message context of the enclosing actor and the methods of the canister: It is dangerous to write shared(msg) actor, because now msg is in scope across all public methods. As long as these also use public shared(msg) func , and thus shadow the outer msg, it is correct, but it if one accidentially omits or mis-types the msg, no compiler error would occur, but suddenly msg.caller would now be the original controller, likely defeating an important authorization step. Instead, write shared(init_msg) actor or shared( caller = controller ) actor to avoid using msg.

Conclusion If you write a serious canister, whether in Motoko or not, it is worth to go through the code and watch out for these patterns. Or better, have someone else review your code, as it may be hard to spot issues in your own code. Unfortunately, such a list is never complete, and there are surely more ways to screw up your code in addition to all the non-IC-specific ways in which code can be wrong. Still, things get done out there, so best of luck!

8 November 2021

Dima Kogan: mrcal 2.0: triangulation and stereo

mrcal is my big toolkit for geometric computer vision: making models (camera calibration) and using models (mapping, ranging, etc). Since the release of mrcal 1.0 back in February I've been busy using the tools in the field, fixing things and improving things. Today I'm happy to finally be able to announce the release of mrcal 2.0. A big part of this release is maintenance and cleanup that resulted from me heavily using the tools over the course of this past year, and improving whatever was bugging me. The most notable result of that effort, is that splined models are no longer "experimental". They work well and they're awesome. Go try them. And there're a number of new features, most notably nice dense stereo support and nice sparse triangulation support (with uncertainty propagation!) These are awesome. Go try them. As before, the tour of mrcal provides a good overview of the capabilities of the toolkit, and is a good place to start reading the documentation. Reading these docs would be very illuminating for anybody that calibrates cameras, even for those that have no intent to actually use the mrcal tools. Let me know if you try it out! The most list of most notable improvements, from the release notes:

Enrico Zini: An educational debugging session

This morning we realised that a test case failed on Fedora 34 only (the link is in Italian) and we set to debugging. The initial analysis This is the initial reproducer: 
$ PROJ_DEBUG=3 python setup.py test
test_recipe (tests.test_litota3.TestLITOTA3NordArkimetIFS) ... pj_open_lib(proj.db): call fopen(/lib64/../share/proj/proj.db) - succeeded
proj_create: Open of /lib64/../share/proj/proj.db failed
pj_open_lib(proj.db): call fopen(/lib64/../share/proj/proj.db) - succeeded
proj_create: no database context specified
Cannot instantiate source_crs
EXCEPTION in py_coast(): ProjP: cannot create crs to crs from [EPSG:4326] to [+proj=merc +lon_0=0 +k=1 +x_0=0 +y_0=0 +ellps=WGS84 +datum=WGS84 +over +units=m +no_defs]
ERROR
Note that opening /lib64/../share/proj/proj.db sometimes succeeds, sometimes fails. It's some kind of Schr dinger path, which works or not depending on how you observe it: 
# ls -lad /lib64
lrwxrwxrwx 1 1000 1000 9 Jan 26  2021 /lib64 -> usr/lib64
$ ls -la /lib64/../share/proj/proj.db
-rw-r--r-- 1 root root 8925184 Jan 28  2021 /lib64/../share/proj/proj.db
$ cd /lib64/../share/proj/
$ cd /lib64
$ cd ..
$ cd share
-bash: cd: share: No such file or directory
And indeed, stat(2) finds it, and sqlite doesn't (the file is a sqlite database): 
$ stat /lib64/../share/proj/proj.db
  File: /lib64/../share/proj/proj.db
  Size: 8925184     Blocks: 17432      IO Block: 4096   regular file
Device: 33h/51d Inode: 56907       Links: 1
Access: (0644/-rw-r--r--)  Uid: (    0/    root)   Gid: (    0/    root)
Access: 2021-11-08 14:09:12.334350779 +0100
Modify: 2021-01-28 05:38:11.000000000 +0100
Change: 2021-11-08 13:42:51.758874327 +0100
 Birth: 2021-11-08 13:42:51.710874051 +0100
$ sqlite3 /lib64/../share/proj/proj.db
Error: unable to open database "/lib64/../share/proj/proj.db": unable to open database file
A minimal reproducer Later on we started stripping layers of code towards a minimal reproducer: here it is. It works or doesn't work depending on whether proj is linked explicitly, or via MagPlus: 
$ cat tc.cc
#include <magics/ProjP.h>
int main()  
    magics::ProjP p("EPSG:4326", "+proj=merc +lon_0=0 +k=1 +x_0=0 +y_0=0 +ellps=WGS84 +datum=WGS84 +over +units=m +no_defs");
    return 0;
 
$ g++ -o tc  tc.cc -I/usr/include/magics  -lMagPlus
$ ./tc
proj_create: Open of /lib64/../share/proj/proj.db failed
proj_create: no database context specified
terminate called after throwing an instance of 'magics::MagicsException'
  what():  ProjP: cannot create crs to crs from [EPSG:4326] to [+proj=merc +lon_0=0 +k=1 +x_0=0 +y_0=0 +ellps=WGS84 +datum=WGS84 +over +units=m +no_defs]
Aborted (core dumped)
$ g++ -o tc  tc.cc -I/usr/include/magics -lproj  -lMagPlus
$ ./tc
What is going on here? A difference between the two is the path used to link to libproj.so: 
$ ldd ./tc   grep proj
    libproj.so.19 => /lib64/libproj.so.19 (0x00007fd4919fb000)
$ g++ -o tc  tc.cc -I/usr/include/magics   -lMagPlus
$ ldd ./tc   grep proj
    libproj.so.19 => /lib64/../lib64/libproj.so.19 (0x00007f6d1051b000)
Common sense screams that this should not matter, but we chased an intuition and found that one of the ways proj looks for its database is relative to its shared library. Indeed, gdb in hand, that dladdr call returns /lib64/../lib64/libproj.so.19. From /lib64/../lib64/libproj.so.19, proj strips two paths from the end, presumably to pass from something like /something/usr/lib/libproj.so to /something/usr. So, dladdr returns /lib64/../lib64/libproj.so.19, which becomes /lib64/../, which becomes /lib64/../share/proj/proj.db, which exists on the file system and is used as a path to the database. But depending how you look at it, that path might or might not be valid: it passes the stat(2) check that stops the lookup for candidate paths, but sqlite is unable to open it. Why does the other path work? By linking libproj.so in the other way, dladdr returns /lib64/libproj.so.19, which becomes /share/proj/proj.db, which doesn't exist, which triggers a fallback to a PROJ_LIB constant defined at compile time, which is a path that works no matter how you look at it. Why that weird path with libMagPlus? To complete the picture, we found that libMagPlus.so is packaged with a rpath set, which is known to cause trouble 
# readelf -d /usr/lib64/libMagPlus.so grep rpath
 0x000000000000000f (RPATH)              Library rpath: [$ORIGIN/../lib64]
The workaround We found that one can set PROJ_LIB in the environment to override the normal proj database lookup. Building on that, we came up with a simple way to override it on Fedora 34 only: 
    if distro is not None and distro.linux_distribution()[:2] == ("Fedora", "34") and "PROJ_LIB" not in os.environ:
         self.env_overrides["PROJ_LIB"] = "/usr/share/proj/"
This has been a most edifying and educational debugging session, with only the necessary modicum of curses and swearwords. Working in a team of excellent people really helps.

6 November 2021

Reproducible Builds: Reproducible Builds in October 2021

Welcome to the October 2021 report from the Reproducible Builds project!
This month Samanta Navarro posted to the oss-security security mailing on a novel category of exploit in the .tar archive format, where a single .tar file contains different contents depending on the tar utility being used. Naturally, this has consequences for reproducible builds as Samanta goes onto reply:

Arch Linux uses libarchive (bsdtar) in its build environment. The default tar program installed is GNU tar. It is possible to create a source distribution which leads to different files seen by the build environment than compared to a careful reviewer and other Linux distributions.
Samanta notes that addressing the tar utilities themselves will not be a sufficient fix:
I have submitted bug reports and patches to some projects but eventually I had to conclude that the problem itself cannot be fixed by these implementations alone. The best choice for these tools would be to only allow archives which are fully compatible to standards but this in turn would render a lot of archives broken.
Reproducible builds, with its twin ideas of reaching consensus on the build outputs as well as precisely recording and describing the build environment, would help address this problem at a higher level.
Codethink announced that they had achieved ISO-26262 ASIL D Tool Certification, a way of determining specific safety standards for software. Codethink used open source tooling to achieve this, but they also leverage:
Reproducibility, repeatability and traceability of builds, drawing heavily on best-practices championed by the Reproducible Builds project.

Elsewhere on the internet, according to a comment on Hacker News, Microsoft are now comparing NPM Javascript packages with their original source repositories:
I got a PR in my repository a few days ago leading back to a team trying to make it easier for packages to be reproducible from source.

Lastly, Martin Monperrus started an interesting thread on our mailing list about Github, specifically that their autogenerated release tarballs are not deterministic . The thread generated a significant number of replies that are worth reading.

Events and presentations

Community news On our mailing list this month:
There were quite a few changes to the Reproducible Builds website and documentation this month as well, including Feng Chai updating some links on our publications page [ ] and marco updated our project metadata around the Bitcoin Core building guide [ ].
Lastly, we ran another productive meeting on IRC during October. A full set of notes from the meeting is available to view.

Distribution work Qubes was heavily featured in the latest edition of Linux Weekly News, and a significant section was dedicated to discussing reproducibility. For example, it was mentioned that the Qubes project has been working on incorporating reproducible builds into its continuous integration (CI) infrastructure . But the LWN article goes on to describe that:
The current goal is to be able to build the Qubes OS Debian templates solely from packages that can be built reproducibly. Templates in Qubes OS are VM images that can be used to start an application qube quickly based on the template. The qube will have read-only access to the root filesystem of the template, so that the same root filesystem can be shared with multiple application qubes. There are official templates for several variants of both Fedora and Debian, as well as community maintained templates for several other distributions.
You can view the whole article on LWN, and Fr d ric also published a lengthy summary about their work on reproducible builds in Qubes as well for those wishing to learn more.
In Debian this month, 133 reviews of Debian packages were added, 81 were updated and 24 were removed this month, adding to Debian s ever-growing knowledge about identified issues. A number of issues were categorised and added by Chris Lamb and Vagrant Cascadian too [ ][ ][ ]. In addition, work on alternative snapshot service has made progress by Fr d ric Pierret and Holger Levsen this month, including moving from the existing host (snapshot.notset.fr) to snapshot.reproducible-builds.org (more info) thanks to OSUOSL for the machine and hosting and Debian for the disks.
Finally, Bernhard M. Wiedemann posted his monthly reproducible builds status report.

diffoscope diffoscope is our in-depth and content-aware diff utility. Not only can it locate and diagnose reproducibility issues, it can provide human-readable diffs from many kinds of binary formats. This month, Chris Lamb made the following changes, including preparing and uploading versions 186, 187, 188 and 189 to Debian
  • New features:
    • Add support for Python Sphinx inventory files (usually named objects.inv on-disk). [ ]
    • Add support for comparing .pyc files. Thanks to Sergei Trofimovich for the inspiration. [ ]
    • Try some alternative suffixes (e.g. .py) to support distributions that strip or retain them. [ ][ ]
  • Bug fixes:
    • Fix Python decompilation tests under Python 3.10+ [ ] and for Python 3.7 [ ].
    • Don t raise a traceback if we cannot unmarshal Python bytecode. This is in order to support Python 3.7 failing to load .pyc files generated with newer versions of Python. [ ]
    • Skip Python bytecode testing where we do not have an expected diff. [ ]
  • Codebase improvements:
    • Use our file_version_is_lt utility instead of accepting both versions of uImage expected diff. [ ]
    • Split out a custom call to assert_diff for a .startswith equivalent. [ ]
    • Use skipif instead of manual conditionals in some tests. [ ]
In addition, Jelle van der Waa added external tool references for Arch Linux for ocamlobjinfo, openssl and ffmpeg [ ][ ][ ] and added Arch Linux as a Continuous Integration (CI) test target. [ ] and Vagrant Cascadian updated the testsuite to skip Python bytecode comparisons when file(1) is older than 5.39. [ ] as well as added external tool references for the Guix distribution for dumppdf and ppudump. [ ][ ]. Vagrant Cascadian also updated the diffoscope package in GNU Guix [ ][ ]. Lastly, Guangyuan Yang updated the FreeBSD package name on the website [ ], Mattia Rizzolo made a change to override a new Lintian warning due to the new test files [ ], Roland Clobus added support to detect and log if the GNU_BUILD_ID field in an ELF binary been modified [ ], Sandro J ckel updated a number of helpful links on the website [ ] and Sergei Trofimovich made the uImage test output support file() version 5.41 [ ].

reprotest reprotest is the Reproducible Build s project end-user tool to build same source code twice in widely differing environments, checking the binaries produced by the builds for any differences. This month, reprotest version 0.7.18 was uploaded to Debian unstable by Holger Levsen, which also included a change by Holger to clarify that Python 3.9 is used nowadays [ ], but it also included two changes by Vasyl Gello to implement realistic CPU architecture shuffling [ ] and to log the selected variations when the verbosity is configured at a sufficiently high level [ ]. Finally, Vagrant Cascadian updated reprotest to version 0.7.18 in GNU Guix.

Upstream patches The Reproducible Builds project detects, dissects and attempts to fix unreproducible packages. We try to send all of our patches upstream where appropriate. We authored a large number of such patches this month, including:

Testing framework The Reproducible Builds project runs a testing framework at tests.reproducible-builds.org, to check packages and other artifacts for reproducibility. This month, the following changes were made:
  • Holger Levsen:
    • Debian-related changes:
      • Incorporate a fix from bremner into builtin-pho related to binary-NMUs. [ ]
      • Keep bullseye environments around longer, in an attempt to fix a Jenkins issue. [ ]
      • Improve the documentation of buildinfos.debian.net. [ ]
      • Improve documentation for the builtin-pho setup. [ ][ ]
    • OpenWrt-related changes:
      • Also use -j1 for better debugging. [ ]
      • Document that that Python 3.x is now used. [ ]
      • Enable further debugging for the toolchain build. [ ]
    • New snapshot.reproducible-builds.org service:
      • Actually add new node. [ ][ ]
      • Install xfsprogs on snapshot.reproducible-builds.org. [ ]
      • Create account for fpierret on new node. [ ]
      • Run node_health_check job on new node too. [ ]
  • Mattia Rizzolo:
    • Debian-related changes:
      • Handle schroot errors when invoking diffoscope instead of masking them. [ ][ ]
      • Declare and define some variables separately to avoid masking the subshell return code. [ ]
      • Fix variable name. [ ]
      • Improve log reporting. [ ]
      • Execute apt-get update with the -q argument to get more decent logs. [ ]
      • Set the Debian HTTP mirror and proxy for snapshot.reproducible-builds.org. [ ]
      • Install the libarchive-tools package (instead of bsdtar) when updating Jenkins nodes. [ ]
    • Be stricter about errors when starting the node agent [ ] and don t overwrite NODE_NAME so that we can expect Jenkins to properly set for us [ ].
    • Explicitly warn if the NODE_NAME is not a fully-qualified domain name (FQDN). [ ]
    • Document whether a node runs in the future. [ ]
    • Disable postgresql_autodoc as it not available in bullseye. [ ]
    • Don t be so eager when deleting schroot internals, call to schroot -e to terminate the schroots instead. [ ]
    • Only consider schroot underlays for deletion that are over a month old. [ ][ ]
    • Only try to unmount /proc if it s actually mounted. [ ]
    • Move the db_backup task to its own Jenkins job. [ ]
Lastly, Vasyl Gello added usage information to the reproducible_build.sh script [ ].

Contributing 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:

Vincent Bernat: How to rsync files between two remotes?

scp -3 can copy files between two remote hosts through localhost. This comes in handy when the two servers cannot communicate directly or if they are unable to authenticate one to the other.1 Unfortunately, rsync does not support such a feature. Here is a trick to emulate the behavior of scp -3 with SSH tunnels. When syncing with a remote host, rsync invokes ssh to spawn a remote rsync --server process. It interacts with it through its standard input and output. The idea is to recreate the same setup using SSH tunnels and socat, a versatile tool to establish bidirectional data transfers. The first step is to connect to the source server and ask rsync the command-line to spawn the remote rsync --server process. The -e flag overrides the command to use to get a remote shell: instead of ssh, we use echo. 
$ ssh web04
$ rsync -e 'sh -c ">&2 echo $@" echo' -aLv /data/. web05:/data/.
web05 rsync --server -vlogDtpre.iLsfxCIvu . /data/.
rsync: connection unexpectedly closed (0 bytes received so far) [sender]
rsync error: error in rsync protocol data stream (code 12) at io.c(228) [sender=3.2.3]
The second step is to connect to the destination server with local port forwarding. When connecting to the local port 5000, the TCP connection is forwarded through SSH to the remote port 5000 and handled by socat. When receiving the connection, socat spawns the rsync --server command we got at the previous step and connects its standard input and output to the incoming TCP socket. 
$ ssh -L 127.0.0.1:5000:127.0.0.1:5000 web05
$ socat tcp-listen:5000,reuseaddr exec:"rsync --server -vlogDtpre.iLsfxCIvu . /data/."
The last step is to connect to the source with remote port forwarding. socat is used in place of a regular SSH connection and connects its standard input and output to a TCP socket connected to the remote port 5000. Thanks to the remote port forwarding, SSH forwards the data to the local port 5000. From there, it is relayed back to the destination, as described in the previous step. 
$ ssh -R 127.0.0.1:5000:127.0.0.1:5000 web04
$ rsync -e 'sh -c "socat stdio tcp-connect:127.0.0.1:5000"' -aLv /data/. remote:/data/.
sending incremental file list
haproxy.debian.net/
haproxy.debian.net/dists/buster-backports-1.8/Contents-amd64.bz2
haproxy.debian.net/dists/buster-backports-1.8/Contents-i386.bz2
[ ]
media.luffy.cx/videos/2021-frnog34-jerikan/progressive.mp4
sent 921,719,453 bytes  received 26,939 bytes  7,229,383.47 bytes/sec
total size is 7,526,872,300  speedup is 8.17
This little diagram may help understand how everything fits together:
Diagram showing how all processes are connected together: rsync, socat and ssh
How each process is connected together. Arrows labeled stdio are implemented as two pipes connecting the process to the left to the standard input and output of the process to the right. Don't be fooled by the apparent symmetry!
The rsync manual page prohibits the use of --server. Use this hack at your own risk!
The options --server and --sender are used internally by rsync, and should never be typed by a user under normal circumstances. Some awareness of these options may be needed in certain scenarios, such as when setting up a login that can only run an rsync command. For instance, the support directory of the rsync distribution has an example script named rrsync (for restricted rsync) that can be used with a restricted ssh login.

Addendum I was hoping to get something similar with a one-liner. But this does not work! 
$ socat \
>  exec:"ssh web04 rsync --server --sender -vlLogDtpre.iLsfxCIvu . /data/." \
>  exec:"ssh web05 rsync --server -vlogDtpre.iLsfxCIvu /data/. /data/." \
over-long vstring received (511 > 255)
over-long vstring received (511 > 255)
rsync error: requested action not supported (code 4) at compat.c(387) [sender=3.2.3]
rsync error: requested action not supported (code 4) at compat.c(387) [Receiver=3.2.3]
socat[878291] E waitpid(): child 878292 exited with status 4
socat[878291] E waitpid(): child 878293 exited with status 4
  1. And SSH agent forwarding is dangerous. Don t use it if you can.

Next.

Previous.