|
Top-level
 @leftpaddotpy @ariadne @dysfun There is really a clash of two *philosophies* here: One that believes software systems should be tightly coupled with every piece of software interacting with the events and data of one another, mediated through object oriented abstractions. And one that believes software should be decoupled, staying out of each other's way and exchanging data and reacting to events only through standard data formats and system interfaces. 37 comments
@dalias @leftpaddotpy @ariadne @dysfun Can you cite user hostile behavior of systemd? @raito @leftpaddotpy @ariadne @dysfun The post you're replying to is not even particularly about systemd. It's about systems where software is tightly coupled. One obvious place it ends up being user hostile is in the presence of low-trust or untrusted applications which get visibility into what other applications you have running and what you're doing in them (Android), partly via the IPC/RPC system. @dalias @leftpaddotpy @ariadne @dysfun why tight coupling cause what you're describing?  @dalias @leftpaddotpy @dysfun i don’t think this really applies to systemd. you can just use the parts you want, without the others. you can even replace the parts you don’t want with other implementations. all of that is possible. it just happens to be a bunch of projects worked on by the same people in a monorepo with a unified release schedule, but the projects are actually fairly independent of each other. @ariadne @leftpaddotpy @dysfun Is it actually possible to run systemd the service manager without systemd the udev controller that decides how it wants to rename all your devices? @ariadne To replace udev with a different device manager you need libudev-zero in most cases. Even then udisk2 and other don't work since they rely on udev internals (source libudev-zero). @sertonix @dalias @leftpaddotpy @dysfun the fact that libudev-zero exists kinda proves my point. yes, if you are reimplementing an API, then you have to reimplement the API. i know, shocking. @ariadne @sertonix @leftpaddotpy @dysfun I was only aware that you could use libudev-zero as a udev replacement on a non-systemd system, not that you could opt not to use the systemd-integrated udev behaviors on a systemd-based system. I'm still not convinced the latter is practical unless someone demonstrates it. @ariadne And even if you need an api libseatd has shown how to do it right: allow multiple backends. If libudev would work ok when udev isn't running it could be considered independent.  @ariadne @sertonix @dalias @leftpaddotpy @dysfun As someone that is one the server side, the total lack of service manager that seems to have understood the problems we deal with is... fascinating. Systemd is definitely the easiest to work with and the one with the most hooks we need. It does have sooooo many issues, both in UX and in implementation (and code quality) but at least it... works and allows us to do the job @ariadne @sertonix @dalias @leftpaddotpy @dysfun Would I love something better and have ideas for it? Sure. Do any of us have the time and expertise and money to build something better? Nope. Same reason we are stuck with autotools and make.... @Di4na @ariadne meson, cmake, and others exist and make my life as a packager and developer much less bad. well, cmake is about on par for packaging suffering as make. but still, they're way nicer to write and use. on the other hand, bazel exists and has caused hours of pain every time I've touched it. but so is the case for googleware of any kind. @ariadne Actively developing better designed software and push the use of it is. And most importantly explain how to use it! The Alpine Wiki tends to include "use this systemd component and it just works(TM)" instead of "you have a choice. One might not work but you can help to improve it here".  @sertonix never heard about libudev-zero before.
Thank you for mentioning it.I had to patch out udev dependency from kmscon few weeks ago. A stub/minimal implementation would've been ideal there, but I didn't knew about it. @dalias uhhhhh > software should be decoupled, staying out of each other's way and exchanging data and reacting to events only through standard data formats and system interfaces. "standard data formats". do you mean "text", for which you have a "broken handrolled C parser full of CVEs" because it's not actually structured *or* standard? or is there some magical format I have forgot about? dbus is a standard data format. the systemd-standardized interfaces are a standard, structured format. @leftpaddotpy Um, you're reading some weird ideas you had in your head into what I said...? @dalias every time I see people accusing systemd of not using the standard APIs, the so-called standard APIs in question did not exist before, and the functionality was, prior to systemd, achieved by the client software mutating things in such a way that assumes they were the only possible client (e.g. /etc/resolv.conf with VPN software, vs under systemd-resolved) I simply do not understand your argument. @leftpaddotpy The post you replied to was not even about systemd. It was about dbus, CORBA, DCOM, OLE & DDE, etc. @leftpaddotpy @dalias oh, systemd-resolved handling of /etc/resolv.conf is great example how to do it wrong way. Yes, there was maybe resolvconf only, hackish command. But because no other service claimed to be best #DNS cache everyone should use. I am quite confident resolved is broken. It supports only itself, unlike other solutions. Okay, it has decent dbus API. But also unfixed issues. Systemd has excellent integration skills, but fails to export it for use by 3rd party services. @pemensik @leftpaddotpy Nobody actually wants resolv.conf to get rewritten based on what network they connect to, with a garbage ns serving fake results with ads injected. They want it statically pointed at 127.0.0.1, 8.8.8.8, or something they trust. That something *should* be on 127.0.0.1, with full dnssec validation & efficient upstream querying, which systemd kinda tried to do, but botched... @dalias @leftpaddotpy that depends. If systemd is preventing some features like dnssec or single label queries, then I want working resolvers instead. Statically configured servers won't know site-specific domains. Sure, you want to override DNS servers network has provided *sometime*. But you would want privacy VPN on such networks anyway, if you need to use that network at all. I get broken experience even on trusted networks with resolved. @pemensik @leftpaddotpy VPN UX on Linux is just messed up. We now have the tools to do it right (user+net namespaces with their own localhost & own ns running on it with no access to outside-vpn network), but are doing it still with fragile hacks messing up global system network config. 🤦 @dalias @leftpaddotpy I am not sure, why would you want to lock a process into separate namespace with a VPN. In most cases I *want* site-enabling VPN to modify *global* system configuration. I just want to choose which parts this VPN can provide, for example limiting it to only selected domains. It would be nice to have simpler ways to bind a service to just one network interface, but I don't need it often. @pemensik @leftpaddotpy I was speaking more with sn assumption of a privacy providing VPN where you don't want any traffic to leak. But a site-enabling VPN also makes sense to scope only to processes meant to have access. A compromised nobody account shouldn't have access to VPN your user account has access to. @dalias @leftpaddotpy I think general assumption is once local machine is compromised, security falls apart. I would say untrusted nobody user (and it's services) should have filtered access to the system network. Especially if you run a code without trusted source from it. Sort of podman solves it. Okay, it would be nice to restrict domains and IP ranges such process can access. I don't know anything simple offering that. @pemensik @leftpaddotpy If that's true, sounds like really, really bad news for most Android users... @dalias @leftpaddotpy who said VPN server on Android cannot filter out ads or unwanted traffic? I think iptables should be able to filter based on userid or pid. I admit it would be great for flatpaks also. Name resolution usually does not have user or pid information for each request, it would have to use Unix socket instead of IP+port domain. So it would be difficult to apply ACL on those. But having alternative DNS cache on non-default address would be simple. @dalias @leftpaddotpy Android apps have good sandbox from other users. Each app is separate user there. Unless given rights to media access, they are well separated. Except network and DNS, correct. I would not expect Google to make ads blocking trivial task for obvious reasons. @dalias @leftpaddotpy what I thing we miss is a good implementation of provisioning domains RFC draft, well integrated with the system. And proper support for site-specific domains (discovery), not ~example.com hacks in dns-search. @dalias @pemensik i think you may have missed the real point of systemd-resolved here: you no longer have different vpn software tampering resolv.conf and breaking your DNS queries when you change networks (e.g.). resolved proxies even bad software like gpg that ignores NSS into hitting the VPN DNS server for only VPN stuff and the network one for others. there is literally no standard way to do this besides everyone having the same DNS proxy, which has APIs, on their computer. @leftpaddotpy The dbus protocol framing/transport is a "standardized data format". The addresses you interact with over it are *explicitly named for application software* that you're communicating with. And forget about whether it's standard or not, it's not even a data format. It's a command channel. @leftpaddotpy What I mean by "standard data formats" vs tightly coupled applications goes back to the early Windows idea of "inserting a paintbrush object" into your Word document rather than "inserting an image file in BMP format". Or the early Mac idea of resource forks identifying an "owning application" for a file rather than the data format of the file. |
Gregory's Server
@leftpaddotpy @ariadne @dysfun The former is largely seen as "polished" and "professional looking" and "user-friendly" by techies and industry UX folks.
But it also enables a lot of really user-hostile behavior.