This seems like the kind of technology that could make the problem described in https://www.gnu.org/philosophy/can-you-trust.en.html a lot worse. Do you have any plans for making sure it doesn't get used for that?
I've been a FOSS guy my entire adult life, I wouldn't put my name to something that would enable the kinds of issues you describe.
It would be a lot more reassuring if we knew what the business model actually was, or indeed anything else at all about this. I remain somewhat confused as to the purpose of this announcement when no actual information seems to be forthcoming. The negative reactions seen here were quite predictable, given the sensitive topic and the little information we do have.
Better security is good in theory, as long as the user maintains control and the security is on the user end. The last thing we need is required ID linked attestation for accessing websites or something similar.
Suppose you wanted to identify potential agitators by scanning all communication for indications in a fascist state one could require this technology in all trusted environments and require such an environment to bank, connect to an ISP, or use Netflix.
One could even imagine a completely benign usage which only identified actual wrong doing alongside another which profiled based almost entirely on anti regime sentiment or reasonable discontent.
The good users would argue that the only problem with the technology is its misuse but without the underlying technology such misuse is impossible.
One can imagine two entirely different parallel universes one in which a few great powers went the wrong way in part enabled by trusted computing and the pervasive surveillance enabled by the capability of AI to do the massive and boring task of analyzing a massive glut of ordinary behaviour and communication + tech and law to ensure said surveillance is carried out.
Even those not misusing the tech may find themselves worse off in such a world.
Why again should we trust this technology just because you are a good person?
Until you get acquired, receive a golden parachute and use it when realizing that the new direction does not align with your views anymore.
But, granted, if all you do is FOSS then you will anyway have a hard time keeping evil actors from using your tech for evil things. Might as well get some money out of it, if they actually dump money on you.
Argument should be technical.
It’s less of an issue in the case of a normal job than in an open source project where often the commitment of particular founding individuals to the long-term future of the project is a big part of people’s decision to use or not use that tech in their solutions. Here, given that “Trusted computing” can potentially lock you out of devices you have bought, it’s important for people to be able to judge the risk of getting “legal ransomware”d if the trusted computing base ends up depending on a proprietary component that they can’t back out of.
That said, there is absolutely zero chance that I use this (systemd is already enough Poettering software for me in this lifetime) so I’m not personally affected either way.
Yes. You correctly stated the important point.
Yes. Aleksa made no technical argument.
A rational and intelligent engineer cannot possibly believe that he'll be able to control what a technology is used for after he creates it, unless his salary depends on him not understanding it.
A lot of the concerns in this thread center on TPMs, but TPMs are really more akin to very limited HSMs that are actually under the user's control (I gave a longer explanation in a sibling comment but TPMs fundamentally trust the data given to them when doing PCR extensions -- the way that consumer hardware is fundamentally built and the way TPMs are deployed is not useful for physical "attacks" by the device owner).
Yes, you can imagine DRM schemes that make use of them but you can also imagine equally bad DRM schemes that do not use them. DRM schemes have been deployed for decades (including "lovely" examples like the Sony rootkit from the 2000s[1], and all of the stuff going on even today with South Korean banks[2]). I think using TPMs (and other security measures) for something useful to users is a good thing -- the same goes for cryptography (which is also used for DRM but I posit most people wouldn't argue that we should eschew all cryptography because of the existence of DRM).
[1]: https://en.wikipedia.org/wiki/Sony_BMG_copy_protection_rootk... [2]: https://palant.info/2023/01/02/south-koreas-online-security-...
it will be railroaded through in the same way that systemD was railroaded onto us.
This is basically true today with Secure Boot on modern hardware (at least in the default configuration -- Microsoft's soft-power policies for device manufacturers actually requires that you can change this on modern machines). This is bad, but it is bad because platform vendors decide which default keys are trusted for secure boot by default and there is no clean automated mechanism to enroll your own keys programmatically (at least, without depending on the Microsoft key -- shim does let you do this programmatically with the MOK).
The set of default keys ended up being only Microsoft (some argue this is because of direct pressure from Microsoft, but this would've happened for almost all hardware regardless and is a far more complicated story), but in order to permit people to run other operating systems on modern machines Microsoft signed up to being a CA for every EFI binary in the universe. Red Hat then controls which distro keys are trusted by the shim binary Microsoft signs[1].
This system ended up centralised because the platform vendor (not the device owner) fundamentally controls the default trusted key set and is what caused the whole nightmare of the Microsoft Secure Boot keys and rh-boot signing of shim. Getting into the business of being a CA for every binary in the world is a very bad idea, even if you are purely selfish and don't care about user freedoms (and it even makes Secure Boot less useful of a protection mechanism because it means that machines where users only want to trust Microsoft also necessarily trust Linux and every other EFI binary they sign -- there is no user-controlled segmentation of trust, which is the classic CA/PKI problem). I don't personally know how the Secure Boot / UEFI people at Microsoft feel about this, but I wouldn't be surprised if they also dislike the situation we are all in today.
Basically none of these issues actually apply to TPMs, which are more akin to limited HSMs where the keys and policies are all fundamentally user-controlled in a programmatic way. It also doesn't apply to what we are building either, but we need to finish building it before I can prove that to you.
This is the intent of the Poettering and Brauner.
Attestation of what to whom for which purpose? Which freedom does it allow users to control their keys, how does it square with remote attestation and the wishes of enterprise users?
As an aside, it is a bit amusing to me that an initial announcement about a new company working on Linux systems caused the vast majority of people to discuss the impact on personal computers (and games!) rather than servers. I guess we finally have arrived at the fabled "Year of the Linux Desktop" in 2026, though this isn't quite how I expected to find out.
> Attestation of what to whom for which purpose? Which freedom does it allow users to control their keys, how does it square with remote attestation and the wishes of enterprise users?
We do have answers for these questions, and a lot of the necessary components exist already (lots of FOSS people have been working on problems in this space for a while). The problem is that there is still the missing ~20% (not an actual estimate) we are building now, and the whole story doesn't make sense without it. I don't like it when people announce vapourware, so I'm really just trying to not contribute to that problem by describing a system that is not yet fully built, though I do understand that it comes off as being evasive. It will be much easier to discuss all of this once we start releasing things, and I think that very theoretical technical discussions can often be quite unproductive.
In general, I will say that there a lot of unfortunate misunderstandings about TPMs that lead people to assume their only use is as a mechanism for restricting users. This is really not the case, TPMs by themselves are actually more akin to very limited HSMs with a handful of features that can (cooperatively with firmware and operating systems) be used to attest to some aspects of the system state. They are also fundamentally under the users' control, completely unlike the PKI scheme used by Secure Boot and similar systems. In fact, TPMs are really not a useful mechanism for protecting against someone with physical access to the machine -- they have to trust that the hashes they are given to extend into PCRs are legitimate and on most systems the data is even provided over an insecure data line. This is why the security of locked down systems like Xbox One[1] don't really depend on them directly and don't use them at all in the way that they are used on consumer hardware. They are only really useful at protecting against third-party software-based attacks, which is something users actually want!
All of the comments about DRM obviously come from very legitimate concerns about user freedoms, but my views on this are a little too long to fit in a HN comment -- in short, I think that technological measures cannot fix a social problem and the history of DRM schemes shows that the absence of technological measures cannot prevent a social problem from forming either. It's also not as if TPMs haven't been around for decades at this point.
The absence of technological measures used to implement societal problems totally does help though. Just look at social media.
I fear the outlaw evil maid or other hypothetical attackers (good old scare-based sales tactics) much less than already powerful entities (enterprises, states) lawfully encroaching on my devices using your technology. So, I don't care about "misunderstandings" of the TPM or whatever other wall of text you are spewing to divert attention.
Anyway, "full control over your keys" isn't the issue, it's the way that normalization of this kind of attestation will enable corporations and governments to infringe on traditional freedoms and privacy. People in an autocratic state "have full control over" their identity papers, too.
PE or EIT?
I can understand corporate use case - the person with access to the machine is not its owner, and corporation may want to ensure their property works the way they expect it to be. Not something I care about, personally.
But when it’s a person using their own property, I don’t quite get the practical value of attestation. It’s not a security mechanism anymore (protecting a person from themselves is an odd goal), and it has significant abuse potential. That happened to mobile, and the outcome was that users were “protected” from themselves, that is - in less politically correct words - denied effective control over their personal property, as larger entities exercised their power and gated access to what became de-facto commonplace commodities by forcing to surrender any rights. Paired with awareness gap the effects were disastrous, and not just for personal compute.
So, what’s the point and what’s the value?
https://doc.qubes-os.org/en/latest/user/security-in-qubes/an... For laptops, it helps make tampering obvious. (a different attestation scheme with smaller scope however)
This might not be useful to you personally, however.
Of course you can already do the above with secure boot coupled with a CPU that implements an fTPM. So I can't speak to the value of this project specifically, only build and boot integrity in general. For example I have no idea what they mean by the bullet "runtime integrity".
And that’s why I said “not a security mechanism”. Attestation is for protecting against actors with local hardware access. I have FDE and door locks for that already.
Given secure boot and a TPM you can remotely attest, using your own keys, that the system booted up to a known good state. What exactly that means though depends entirely on what you configured the image to contain.
> it won’t protect from malicious updates to configuration files
It will if you include the verified correct state of the relevant config file in a merkel tree.
> It won’t let me run arbitrary binaries (putting a nail to any local development), or if it will - it would be a temporary security theater (as attackers would reuse the same processes to sign their malware).
Shouldn't it permit running arbitrary binaries that you have signed? That places the root of trust with the build environment.
Now if you attempt to compile binaries and then sign them on the production system yeah that would open you up to attack (if we assume a process has been compromised at runtime). But wasn't that already the case? Ideally the production system should never be used to sign anything. (Some combination of SGX, TPM, and SEV might be an exception to that but I don't know enough to say.)
> Attestation is for protecting against actors with local hardware access. I have FDE and door locks for that already.
If you remotely boot a box sitting in a rack on the other side of the world how can you be sure it hasn't been compromised? However you go about confirming it, isn't that what attestation is?
Production servers are a whole different story - it's usually not my hardware to begin with. But given how things are mostly immutable those days (shipped as images rather than installed the old-fashioned sysadmin way), I'm not really sure what to think of it...
Granted these technologies can also be abused. But that involves running third party binaries that require SGX or other DRM measures before they will unlock or decrypt content or etc. Or querying a security element to learn who signed the image that was originally booted. Devices that support those things are already widespread. I don't think that's what this project is supposed to be. (Although I could always be wrong. There's almost no detail provided.)
This is for example dm-verity (e.g. `/usr/` is an erofs partiton with matching dm-verity). Lennart always talks about either having files be RW (backed by encryption) or RX (backed by kernel signature verification).
That is impossible.
"secure" devices get silently tampered with everyday.
You can never guarantee that.
History is pretty consistent here:
WhatsApp: privacy-first, founders with principles, both left once monetization and policy pressure kicked in.
Google: 'Don’t be evil' didn’t disappear by accident — it became incompatible with scale, revenue, and government relationships.
Facebook/Meta: years of apologies and "we'll do better," yet incentives never changed.
Mobile OS attestation (iOS / Android): sold as security, later became enforcement and gatekeeping.
Ruby on Rails ecosystem: strong opinions, benevolent control, then repeated governance, security, and dependency chaos once it became critical infrastructure. Good intentions didn't prevent fragility, lock-in, or downstream breakage.
Common failure modes:
Enterprise customers demand guarantees - policy creeps in.
Governments demand compliance - exceptions appear.
Liability enters the picture - defaults shift to "safe for the company."
Revenue depends on trust decisions - neutrality erodes.
Core maintainers lose leverage - architecture hardens around control.
Even if keys are user-controlled today, the key question is architectural: Can this system resist those pressures long-term, or does it merely promise to?
Most systems that can become centralized eventually do, not because engineers change, but because incentives do. That’s why skepticism here isn't personal — it's based on pattern recognition.
I genuinely hope this breaks the cycle. History just suggests it's much harder than it looks.
If user control of keys becomes the linchpin for retaining full control over one's own computer, doesn't it become easy for a lobby or government to exert control by banning user-controlled keys? Today, such interest groups would need to ban Linux altogether to achieve such a result.
The road to hell is paved with good intentions.
FOR NOW. Policies and laws always change. Corporations and governments somehow always find ways to work against their people, in ways which are not immediately obvious to the masses. Once they have a taste of this there's no going back.
Please have a hard and honest think on whether you should actually build this thing. Because once you do, the genie is out and there's no going back.
This WILL be used to infringe on individual freedoms.
The only question is WHEN? And your answer to that appears to be 'Not for the time being'.
- your bank won't let you log in from an "insecure" device.
- you won't be able to play videos on an "insecure" device.
- you won't be able to play video games on an "insecure" device.
And so on, and so forth.
Of course its all nonsense make believe, the "trust root" is literally a Microsoft signed stub. For this dummy implementation you can't modify your own kernel anymore.
The attestation portion of those systems is happening on locked down devices, and if you gain ownership of the devices they no longer attest themselves.
This is the curse of the duopoly of iOS and Android.
BankID in Sweden will only run with one of these devices, they used to offer a card system but getting one seems to be impossible these days. So you're really stuck with a mobile device as your primary means of identification for banking and such.
There's a reason that general purpose computers are locked to 720p on Netflix and Disney+; yet AppleTV's are not.
Now, if you want to use your phone as a debit/credit card substitute that is different (Google Pay cares, and I don't use it thus).
Anyway, why should banking apps care? It is not like they care when I use the bank from Firefox on my Linux laptop.
"Any bank"? Although the bank I use locks NFC payments behind such checks (which is not a big loss since a physical debit card offers the same functionality), anything else still works otherwise. Most of the things are available through the website (which fits well on mobile too), and mobile BLIK payments can be done from the Android app which works inside Waydroid with microG.
There's no reason other banks can't work the same way and it's outraging when they don't. Look around for a better bank.
The way you defeat things like that is through political maneuvering and guile rather than submission to their artificial narrative. Publish your own papers and documentation that recommends apps not support any device with that feature or require it to be off because it allows malware to use the feature to evade malware scans, etc. Or point out that it prevents devices with known vulnerabilities from being updated to third party firmware with the patch because the OEM stopped issuing patches but the more secure third party firmware can't sign an attestation, i.e. the device that can do the attestation is vulnerable and the device that can't is patched.
The way you break the duopoly is by getting open platforms that refuse to support it to have enough market share that they can't ignore it. And you have to solve that problem before they would bother supporting your system even if you did implement the treachery. Meanwhile implementing it makes your network effect smaller because then it only applies to the devices and configurations authorized to support it instead of every device that would permissionlessly and independently support ordinary open protocols with published specifications and no gatekeepers.
Another point is (often )the apps that banks makes are 3rd party developed by outsourcing (even if within the same developed country). If someone uses some MiTM or logcat to see some traffic and publishes it then banks get bad publicity. So to prevent this the banks, devs tell anything that is not normal (i.e) non-stock ROM is bad.
FOSS is also something many app-based software devs don't like on their products. While people in cloud, infra like it the app devs like these tools while developing or building a company but not when making end resulting apps.
I think ever compute professional needs to spend at least a year trying to secure a random companies windows network to appreciate how impossible this actually is without hardware based roots of trust like TPMs and HSMs
It's not. Mobile applications just don't have unrestricted access to everything in your user directory, attestation have nothing to do with it.
Even if you stopped supporting desktops, then they would just reverse engineer the mobile app instead of the web app and extract the attestation keys from any unpatched model of phone and still run their code on a server, and then it would show up as "mobile fraud" because they're pretending to be a phone instead of a desktop, when in reality it was always a server rather than a phone or a desktop.
And even if attestation actually worked (which it doesn't), that still wouldn't prevent fraud, because it only tries to prove that the person requesting the transfer is using a commercial device. If the user's device is compromised then it doesn't matter if it can pass attestation because the attacker is only running the fake, credential stealing "bank app" on the user's device, not the real bank app. Then they can run the official bank app on an official device and use the stolen credentials to transfer the money. The attestation buys you nothing.
Also, as there is ample discussion elsewhere, having attestation does NOT eliminate the ability for your account to become compromised.
As restated.
"If the user's device isn't compromised then everything is fine regardless of whether or not it can pass attestation. If the user's device is compromised, the device doesn't need to pass attestation to run a fake bank app and steal the user's credentials. Once the attacker has the user's credentials they can use them to transfer money regardless of whether or not they have to use a different device that can pass attestation.
It doesn't really provide any security."
IT DOES however completely rewrite the paradigm of general purpose computing in very asymmetrical ways.
I think there's also an argument in favor of a person having the right to access their money (and I'd argue that accessing your bank's website/app is accessing your money) however they want, and that access to their money is more of an important right than the bank's right to control how that access happens.
I think we can all agree to some "within reason" clauses on both sides (eg not allowing HTTP only access seems reasonable), and I guess a lot of this debate is "is requiring attestation within reason?"
To me, any asymmetry between the rights of the consumer and the rights of the bank should be in the favor of the consumer.
It doesn't really provide any security.
On top of that, there are tons of devices that can pass attestation that have known vulnerabilities, so the attacker could just use one of those (or extract the keys from it) if they had any reason to. But in the mobile banking threat model they don't actually need to.
It's not that you can't solve the problem, it's that you can't solve the problem using that mechanism. Attestation is useless for this.
The thing that would actually work for this is to have an open standard supported by PCs and phones to read the chip in payment/ATM cards, because then you could do "card-present" transactions remotely. You touch your card to the phone/PC and enter your PIN to authorize a new merchant. That actually solves the problem because then instead of the bank trusting every commercially available phone on the market, they only trust the specific card that they mailed to the cardholder, and you can only authorize a new merchant with physical possession of the card because it contains a private key. But that doesn't require attestation because then you don't need the keys to be in the phone since they're in the card, and it doesn't require a third party to sign anything because the bank puts the private key into the card before sending it to the cardholder without any need for Google or Apple to certify anything.
Yes you can use a chip that the bank trusts (that's your card), however the bank wants to trust that the hardware you use to read that chip is not compromised and does not try to do things on the behalf of the user that the user didn't authorize. A non trusted device can operate in a different way than the user demands of it, and the user might never know.
That's the use case that hardware attestation can prevent. Or so the theory says...
Also IIRC, linux foundation etc are not interested in doing such standardisations.
However the problem is that A LOT of things only work with the mobile app.
Finances, just pay everything by cheque or physical pennies. Fight back. Starve the tyrants to death where you can, force the tyrants to incur additional costs and inefficiencies where you can't.
the anti-user attestation will at least be full of security holes, and likely won't work at all
It took us nearly a decade and a half to unfuck the pulseaudio situation and finally arrive at a simple solution (pipewire).
SystemD has a lot more people refining it down but a clean (under the hood) implementation probably won't be witnessed in my lifetime.
for systemd, I don't think I have a single linux system that boots/reboots reliably 100% of the time these days
The people who had no issues with Pulseaudio; used a mainstream distribution. Those distributions did the heavy lifting of making sure stuff fit together in a cohesive way.
SystemD is very opinionated, so you'd assume it wouldn't have the same results, but it does.. if you use a popular distro then they've done a lot of the hard work that makes systemd function smooth.
I was today years old when I realised this is true for both bits of poetter-ware. Weird.
pulseaudio I had to fight every single day, with my "exotic" setup of one set of speakers and a headset
with pipewire, I've never had to even touch it
systemd: yesterday I had a network service on one machine not start up because the IP it was trying to bind to wasn't available yet
the dependencies for the .service file didn't/can't express the networking semantics correctly
this isn't some hacked up .service file I made, it's that from an extremely popular package from a very popular distro
(yeah I know, use a socket activated service......... more tight coupling to the garbage software)
the day before that I had a service fail to start because the wall clock was shifted by systemd-timesyncd during startup, and then the startup timeout fired because the clock advanced more than the timeout
then the week before that I had a load of stuff start before the time was synced, because chrony has some weird interactions with time-sync.target
it's literally a new random problem every other boot because of this non-deterministic startup, which was never a problem with traditional init or /etc/rc
for what? to save maybe a second of boot time
if the distro maintainers don't understand the systemd dependency model after a decade then it's unfit for purpose
well, systemd's got them beat there!
… and the warrant canary we publish every Monday morning.
Ubuntu boxes: usually ok as long as you stay away from anything python related in the core system.
This gave me a good chuckle. Systemd literally was created to solve the awful race conditions and non-determinism in other init systems. And it has done a tremendous job at it. Hence the litany of options to ensure correct order and execution: https://www.freedesktop.org/software/systemd/man/latest/syst...
And outside of esoteric setups I haven't ever encountered the problems you mentioned with service files.
like "at least one real IP address is available" or "time has been synced"
and it's not esoteric, even ListenAddress with sshd doesn't even work reliably
the ONLY piece of systemd I've not had problems with is systemd-boot, and then it turned out they didn't write that
"network-online.target is a target that actively waits until the network is “up”, where the definition of “up” is defined by the network management software. Usually it indicates a configured, routable IP address of some kind. Its primary purpose is to actively delay activation of services until the network has been set up."
For time sync checks, I assume one of the targets available will effectively mean a time sync has happened. Or you can do something with ExecStartPre. You could run a shell command that checks for the most recent time sync or forces one.
It is possible for a specification to be so abstract that it's useless.
this service (untouched by me) had:
After=local-fs.target network-online.target remote-fs.target time-sync.target
but it was still started without an IP address, and then failed to bind
just like this sort of problem: https://github.com/systemd/systemd/issues/4880#issuecomment-...
the entire thing is unreliable and doesn't act like you'd expect
> Or you can do something with ExecStartPre. You could run a shell command that checks for the most recent time sync or forces one.
at that point I might as well go back to init=/etc/rc
You could also try targeting NetworkManager or networkd's "wait-online" services. Or if that doesn't work, something is telling systemd that you have an IP when you don't. NetworkManager has "ipv4.may-fail" and "ipv6.may-fail" that might be errenously true.
> at that point I might as well go back to init=/etc/rc
The difference is that systemd is much better at ensuring correctness. If you write the invoked shell command properly, it'll communicate failure or success correct and systemd will then communicate that state to the unit. It's still a lot more robust than before.
the problem is systemd
> The difference is that systemd is much better at ensuring correctness.
yeah, whatever mate
There is so much granularity and flexibility in what you can do it seems rather unlikely you cannot make it happen correctly. And if it is truly a bug... open an issue? They're rather responsive to it. And it isn't like the legacy init systems were bug free from inception (well, lord knows they were still chock full of bugs even when they were replaced).
Edit: sitting here with a grin .. HN downvoting the advice of checking logs, debugging and opening an issue. I wish the companies y'all work at good luck.. they'll need it.
I'm a pragmatist: I just want it to work
my solution to MULTIPLE different services failing to IP bind is to turn on the non-local ip binding sysctl, bypassing systemd's brokenness entirely
> There is so much granularity and flexibility in what you can do it seems rather unlikely you cannot make it happen correctly.
I've written an init before (in C), I know how the netlink interface to set an IP address and add routing table entries works
I understand the difference between monotonic and wall clocks
I understand the difference between Wants and Require
I know what's going on at every, single, level
and I can't stand how unreliable systemd makes nearly every single one of my, bluntly, completely vanilla systems
> And if it is truly a bug... open an issue?
did you read the link I pasted earlier?
I'm not wasting my time with that level of idiocy (from LP himself)
So basically it just doesn't work sometimes for no particular reason.
> The difference is that systemd is much better at ensuring correctness
Uh, well, you just said that it isn't, because some targets like network-online.target don't work from user unit files.
I'm not a systemd hater or anything, but I continue to read stuff from Poettering which to me is deeply disturbing given the programs he works on.
Saying it's not a bug that service is launched despite a stated required prerequisite dependency failed... WTF?
Sure, I agree with him that most computers should probably boot despite NTP being unable to sync. But proposing that the solution to that is breaking Requires is just wild to me.
The question in that issue is around the semantics of time-sync.target. Targets are synchronization points for the system and don't (afaik) generally make promises about the units that are ordered before them (in this case chrony-wait.service.
Does that answer your specific objection of "proposing that the solution to that is breaking Requires is just wild to me"? Basically, what is proposed in that issue is not breaking Requires=. The proposition is that the user add their own, specific Requires= as a drop-in configuration since that's not a generally-applicable default.
Targets[1]: Target units do not offer any additional functionality on top of the generic functionality provided by units. They merely group units, allowing a single target name to be used in Wants= and Requires= settings to establish a dependency on a set of units defined by the target, and in Before= and After= settings to establish ordering.
boot-complete.target[2]: Order units that shall only run when the boot process is considered successful after the target unit and pull in the target from it, also with Requires=.
Note use of "only run" with a reference to Requires=.
time-sync.target[3]: This target provides stricter clock accuracy guarantees than time-set.target (see above), but likely requires network communication and thus introduces unpredictable delays. Services that require clock accuracy and where network communication delays are acceptable should use this target.
Especially note the last sentence there.
The documentation clearly indicates that targets can fail, and that services that needs the target to be successful, should use Requires= to specify that.
If the above is not true, the Requires= and Targets documentation should be rewritten to explicitly say that targets might fulfill Requires= regardless of state. Also, the documentation for time-sync.target should explicitly remove the last sentence and instead state there is no functional difference between Requires=time-sync.target and Wants=time-sync.target, it is best-effort only.
[1]: https://www.freedesktop.org/software/systemd/man/latest/syst...
[2]: https://www.freedesktop.org/software/systemd/man/latest/syst...
[3]: https://www.freedesktop.org/software/systemd/man/latest/syst...
If he really don't want targets to deliver failed/success guarantees, then they've massively miscommunicated in their documentation. That in my book is a huge deal.
In either case the issue should in no circumstance be casually dismissed as not-a-bug without further action.
I'm sure the project would accept a documentation patch to amend this discrepancy. At the end of the day (despite what some people on the internet might like to allege), systemd is a free software project that, despite having (more or less) a BIFL, is ultimately a relatively bazaar-like project.
Though since these targets and unit properties are very core to systemd-the-service manager, I do think that this is a bigger documentation oversight than most.
To me this is a huge red flag for a senior contributer to a core systems component, signalling some fundamental lack of understanding or imagination.
I very much disagree with not fixing time-sync.target, but if he had instead written a well-reasoned explanation for why time-sync.target should not propagate failed states and flagging it as a documentation bug, then that's something I'd respect and would be fine with. Or, even better IMHO, he'd fix time-sync.target and state that users who wants to boot regardless should use Wants instead.
runit's approach is to just keep trying to start the shell script every 2 seconds until it works. One of those worse–is–better ideas, it's really dumb, and effective. You can check for arbitrary conditions and error–exit, and it will keep trying. If you need the time synced you can just make your script fail if the time is not synced.
traditional inittab is older than that and there's not any reason to use it when you could be using runit, really.
Doubtful the motivation was /etc/rc being too slow
daemontools, runit, s6 solve that problem
IIRC before PulseAudio we had to mess around with ALSA directly (memory hazy, it was a while ago). It could be a bit of a pain.
ALSA was kind of OK after mixing was enabled by default and if you didn't need to switch outputs of a running application between anything but internal speakers and headphones (which worked basically in hardware). With any additional devices that you could add and remove, ALSA became a more serious limitation, depending. You could usually choose your audio devices (including microphones) at least at the beginning of a video conference / playing a movie etc, but it was janky (unreliable, list of 20 devices for one multi-channel sound card) and needed explicit support from all applications. Not sure if it ever worked with Bluetooth.
It does, with the help of BlueALSA[0].
I get ALSA followed the Unix philosophy of doing one thing but I want my audio mixer to play multiple sounds at once.
I didn't actually know anything about Linux at the time and started with Gentoo because I saw a meme saying "install Gentoo" and people told me not to start with that distro. So it's possible I messed up the default config by accident.
Either way PulseAudio worked after I emerged it.
I am not seeing these kind of systemd issues with Fedora / RHEL.
It just works
I believe that you are genuinely being sincere here, thinking this is good advice.
But this is an absolutely terrible philosophy. This statement is ignorant as well as inconsiderate. (again, I do believbe you don't intend to be inconsiderate consciously, that is just the result.)
It's ignorant of history and inconsiderate of everyone else but yourself.
Go back a few years and this same logic says "The trick is, just use Windows and do whatever it wants and don't fight."
So why in the world are you even using Linux at all in the first place with that attitude? For dishonest reasons (when unpacked to show the double standard).
Since you are using Linux instead of Windows, then you actually are fine with fighting the tide. You want the particular bits of control you want, and as long as you are lucky enough to get whatever you happen to care about without fighting too much, then you have no sympathy for anyone else who cares aboiut anything else.
You don't see yourself as fighting any tides because you are benefitting from being able to use a mainstream distro without customizing it. But the only reason you get to enjoy any such thing at all in the first place is because a lot of other people before you fought the tide to bring some mainstream distros into existence, and actually use them for ordinary activities enough despite all the difficulties, to force at least some companies and government agencies to acknowledge them. So now you can say things like "just use a mainstream distro as it comes and don't try to do what you actually want".
This is basically exactly what I saw people saying in Windows subreddits. There's one post that particularly sticks out in my memory[0] that basically had everybody telling the OP to just not make any of the changes that they wanted to make. The advice seemed to revolve around adapting to the OS rather than adapting the OS to you, and it made me sad at the time.
[0] https://www.reddit.com/r/Windows10/comments/hehrqe/what_are_...
Incorrect. I used mainstream distro, still had issues, that just solved itself moving to pipewire. Issues like it literally crashing or emitting spur of max volume noise once every few months for no discernable reason.
Pulseaudio also completely denies existence of people trying to do music on Linux, there is no real way to make latency on it be good.
> SystemD is very opinionated, so you'd assume it wouldn't have the same results, but it does.. if you use a popular distro then they've done a lot of the hard work that makes systemd function smooth.
Over the years of using the "opinion" of SystemD seems to be "if it is not problem on Lennart's laptop, it's not a real problem and it can be closed or ignored completely".
For example systemd have no real method to tell it "turn off all apps after 5 minutes regardless of what silly package maintainers think". Now what happens if you have a server on UPS that have say 5 minutes of battery and one of the apps have some problem and doesn't want to close?
In SysV, it gets killed, and system gets remounted read only. You have app crash recovery but at least your filesystem is clean In systemd ? No option to do that. You can set default timeout but it can be override in each service so you'd have to audit every single package and tune it to achieve that. That was one bug that was closed.
Same problem also surfaced if you have say app with a bug that prevented it from closing from sigterm and you wanted to reboot a machine. Completely stuck
But wait, there is another method, systemd have an override, you can press (IIRC) ctrl+alt+delete 7 times within 2 seconds to force it to restart ( which already confuses some people that expect it to just restart machine clean(ish) regardless https://github.com/systemd/systemd/issues/11285 ).
...which is also impossible if your only method of access is software KVM where you need to navigate to menu to send ctrl+alt+del. So I made ticket with proposal to just make it configurable timeout for the CAD ( https://github.com/systemd/systemd/issues/29616 ), the ticket wasn't even read completely because Mr. Poettering said "this is not actionable, give a proposal", so I pasted the thing he decided to ignore in original ticket, and got ignored. Not even "pull requests welcome" (which I'd be fine with, I just wanted confirmation that the feature like that won't be rejected if I start writing it).
There is also issue of journald disk format being utter piece of garbage ("go thru entire journal just to get app's last few lines bad", hundreds of disk reads on simple systemctl status <appname> bad) that is consistently ignored thru many tickets from different people.
Or the issue that resolvconf replacement in systemd will just roll a dice on DNS ordering, but hey, Mr. Lennart doesn't use openvpn so it's not real issue ( https://github.com/systemd/systemd/issues/27543 )
I'm not writing it to shit on systemd and praise what was before, as a piece of software it's very useful for my job as sysadmin (we literally took tens of thousands lines of fixed init scripts out because all of the features could be achieved in unit files) and I mean "saved tons of time and few demons running" in some cases, but Mr. Poettering is showing same ignorant "I know better" attitude he got scolded at by kernel maintainers.
I don't care much about PA at this point tbh and don't know much about the inner workings; it always worked just fine for me. But from what I read from people more "in the know" at the time, I'd heard that a lot of the (very real) user-facing problems with PA were ultimately caused by driver and other low-level problems. Those were hacky, had poor assumptions, etc. PA ultimately exposed those failures, and largely got better over time because those problems got fixed upstream of PA.
My takeaway from what I read was basically that PA had to stumble and walk so that pipewire could run.
> For example systemd have no real method to tell it "turn off all apps after 5 minutes regardless of what silly package maintainers think". Now what happens if you have a server on UPS that have say 5 minutes of battery and one of the apps have some problem and doesn't want to close?
Add a TimeoutStopSec= to /etc/systemd/system/service.d/my-killing-dropin.conf more or less, I think? These are documented in the systemd.service and systemd.unit manpages respectively.
> Same problem also surfaced if you have say app with a bug that prevented it from closing from sigterm and you wanted to reboot a machine. Completely stuck
See the --force option on the halt, poweroff, and reboot subcommands of systemctl. The kill subcommand if you want to target that specific service.
> so I pasted the thing he decided to ignore in original ticket, and got ignored. Not even "pull requests welcome" (which I'd be fine with, I just wanted confirmation that the feature like that won't be rejected if I start writing it).
I'm certainly sympathetic to this pain point. I'd take Lennart at his word that he's not opposed. Generally speaking, from following the systemd project somewhat, it's a very busy project and it's hard for all issues to get serviced. But they're very open to PRs, generally speaking.
> Or the issue that resolvconf replacement in systemd will just roll a dice on DNS ordering, but hey, Mr. Lennart doesn't use openvpn so it's not real issue ( https://github.com/systemd/systemd/issues/27543 )
Quickly taking a peek here (and speaking as a relatively superficial user of resolved myself), isn't the proposed solution to define interface ordering?
> it will ask on all links in parallel if there's no better routing info available. In your case there is none (i.e. no ~. listed among your network interfaces), hence it will be asked on all interfaces at the same time.
What set systemd apart is the collection of tightly integrated utilities such as a dns resolver, sntp client, core dump handler, rpc-like api linking to complex libraries in the hot path and so on and so forth that has been a constant stream of security exploits for over a decade now.
This is a case where the critics were proven to be right. Complexity increases the cognitive burden.
I think he will succeed and we will be worse off, collectively.
For example, the part of systemd that fills DNS will put them in random order (like actual random, not "code happened to dump it in map order)
The previous, while very much NOT perfect, system, put the DNSes in order of one in latest interface, which had useful side-feature that if your VPN had different set of DNSes, it got added in front
The systemd one just randomizes it ( https://github.com/systemd/systemd/issues/27543 ) which means that using standard openvpn wrapper script for it will need to be reran sometimes few times to "roll" the right address, I pretty much have to run
systemctl restart systemd-resolved ; sleep 1 ; cat /etc/resolv.conf
The OTHER problem is pervasive NIH in codebase.
Like, they decided to use binary log format. Okay, I can see advantages, it can be indexed or sharded for faster access to app's files...
oh wait it isn't, if you want to get last few lines of a service the worst case is "mmap every single journal file for hundreds of MBs of reads"
It can be optimized so some long but constant fields like bootid are not repeated...
oh wait it doesn't do that either, is massively verbose. I guess I can understand it, at least that would make it less crash-proof...
oh wait no, after crash it just spams logs that previous log file is corrupted and it won't be used.
So we have a log format that only systemd tools can read, takes few times as much space per line as text or even JSON version would, and it still craps out on unclean shutdown
They could've just integrated SQLite. Hell I literally made a lil prototype that took journalctl logs and wrote it to indexed SQLite file and it was not only faster but smaller (as there is no need to write bootid with each line, and log lines can be sharded or indexed so lookup is faster). But nah, Mr. Poettering always wanted to make a binary log format so he did.
don't get me wrong, i use pipewire all day every day, and wrote one of the APIs (JACK) that it implements (pretty well, too!).
but pipewire is an order of magnitude more complex than pulseaudio.
My 0.02 bits.
Audio server for linux: Great idea! Pulseaudio: Genuinely a terrible implementation of it, Pipewire is a drop in replacement that actually works.
Launchd but for Linux: Great idea! SystemD: generally works now at least, but packed with insane defaults and every time this is brought up with the devs they say its the distro packagers jobs to wipe SystemD's ass and clean up the mess before users see it.
Security bug in SystemD when the user has a digit in their username: Lennart closes the bug and says that SystemD is perfect, the distros erred by permitting such usernames. Insane ego-driven response.
I am not going to struggle with systemd if I have to build containers specifically for it. If I have to rearrange everything I am doing I would just learn to do it on a minimal Kubernetes install instead.
I don't think the setting is exposed to regular service units (it might be able to in the future, I don't know) and I don't think podman has any integration with it.
What kinda service do you have where you need a full range of UIDs?
Let's say Home Assistant. It would be nice to have a have some system user "homeassistant" with no home directory that owns the process and owns its /var/whereever/config.conf . It would be nice to have the isolation on host in addition to the isolation via container. But I don't want to be rebuilding any containers to get that, unless I am misunderstanding something on nsresourced.
I'd be really pleased with that setup. MQTT could be its own system user. And HA could depend on MQTT so I have nice startup behavior. Etc.
IDK how to have system users like this run a container without the subuid range. Even when I create the users with ranges in the file, there seems to be problems with informing systemd (as a non-root user) that the running process is different from the one it started.
> But I don't want to be rebuilding any containers to get that, unless I am misunderstanding something on nsresourced.
Setting up the user namespace would be part of the container manager and not the containers themselves, so they shouldn't need any rebuilding or special handling (possibly the files might need to be shifted into the "foreign ID" range[1, 2], but I might be lying with this and this isn't necessary for this usecase) but the container manager needs to be specifically make use of nsresourced.
I really think currently the best option is to go with either systemd as your "container manager" (e.g. just regular system files with sandboxing or nspawn images or maybe systemd-portabled[3]) or podman as your container manager. As much as I too would love to mix them, I don't think it's the best idea (at least in the current state) and just go with what is more suited for the task (in your case it sounds like podman would be the most suited option).
> there seems to be problems with informing systemd (as a non-root user) that the running process is different from the one it started.
Yea, I don't think systemd likes double forking. The best option would be to keep the process that spawned your actual process alive until the child exists and just bubble up the exit code. There is the `PIDFile=` option with `Type=forking`, but I haven't used it, nor looked much into it.
[0]: https://docs.podman.io/en/v5.7.1/markdown/podman-systemd.uni...
[1]: https://www.freedesktop.org/software/systemd/man/latest/syst...
[2]: https://systemd.io/UIDS-GIDS/#special-systemd-uid-ranges
"In fact, the way I see things the Linux API has been taking the role of the POSIX API and Linux is the focal point of all Free Software development. Due to that I can only recommend developers to try to hack with only Linux in mind and experience the freedom and the opportunities this offers you. So, get yourself a copy of The Linux Programming Interface, ignore everything it says about POSIX compatibility and hack away your amazing Linux software. It's quite relieving!" -- https://archive.fosdem.org/2011/interview/lennart-poettering...