Does anybody know why dbus exists? I’ve been wracking my brain trying to come up with a usecase for dbus that isn’t already covered by Unix sockets.
You want to remotely control a daemon? Use sockets. You want the daemon to respond to the client? Sockets. Want to exchange information in json? plaintext? binary data? Sockets can do it. Want to restrict access to a socket? Go ahead, change the socket’s permissions. Want to prevent unauthorized programs from pretending to be someone they’re not? Change the permissions of the directory containing the socket. Want network transparency? That’s why we have abstract sockets.
Plenty of well-established software uses sockets. Music player daemon uses sockets. BSPWM uses sockets. Tmux uses sockets. Pipewire uses sockets. Dhcpcd uses sockets. Heck, dbus itself relies on sockets!
For developers, using sockets is easy. I once wrote a program that interfaced with BSPWM, and it was a breeze. Dbus, on the other hand, not so much. I tried writing a Python script that would contact Network Manager and check the WiFi signal strength. Right off the bat I’m using some obscure undocumented package for interfacing with dbus. What is an introspection? What is a proxy object? What is an interface? Why do I need 60 lines of (Python!) code for a seemingly trivial operation?
So why do some developers decide to use dbus when they could just use unix sockets and save a lot of hassle for themselves and others?
With pipes/sockets, each program has to coordinate the establishment of the connection with the other program. This is especially problematic if you want to have modular daemons, e.g. to support drop-in replacements with alternative implementations, or if you have multiple programs that you need to communicate with (each with a potentially different protocol).
To solve this problem, you want to standardize the connection establishment and message delivery, which is what dbus does.
With dbus, you just write your message to the bus. Dbus will handle delivering the message to the right program. It can even start the receiving daemon if it is not yet running.
It’s a bit similar to the role of an intermediate representation in compilers.
This reminds me of QT’s signal/slot system. I.e. instead of calling functions directly, you just emit a signal and then any number of functions may have the receiving slot enabled.
Lot’s of similar systems in other frameworks too I’m sure.
It occurs to me that sendmsg() is already kind of a standard, and the problem of drop in replacements could be solved by just making the replacement bind to the same file path and emulate the same protocol, and the problem of automatically starting the daemon could be handled by a systemd socket (or even inetd if you wanna go old school). The only advantage that I can see dbus really having over Unix sockets is allowing multiple programs to respond to the same message, which is a definite advantage but AFAIK not many things take advantage of that.
modular daemons
A message bus won’t magically remove the need for developers to sit down together and agree on how some API would work. And not having a message bus also doesn’t magically prevent you from allowing for alternative implementations. Pipewire is an alternative implementation of pulseaudio, and neither of those rely on dbus (pulse can optionally use dbus, but not for its core features). When using dbus, developers have to agree on which path the service owns and which methods it exposes. When using unix sockets, they have to agree where the socket lives and what data format it uses. It’s all the same.
It can even start the receiving daemon if it is not yet running.
We have a tool for that, it’s called an init system. Init systems offer a large degree of control over daemons (centralized logging? making sure things are started in the correct order? letting the user disable and enable different daemons?). Dbus’ autostart mechanism is a poor substitute. Want to run daemons per-user instead of as root? Many init systems let you do that too (I know systemd and runit do).
“Bro just use sockets lol” completely misses the point. When you decide you want message based IPC, you need to then design and implement:
- Message formatting
- Service addressing
- Data marshalling
- Subscriptions and publishing
- Method calling, marshalling of arguments and responses
- Broadcast and 1:1 messaging
And before you know it you’ve reimplemented dbus, but your solution is undocumented, full of bugs, has no library, no introspection, no debugging tools, can only be used from one language, and in general is most likely pure and complete garbage.
Well said. There are so many details to making code work that can so often be avoided by using the right tooling. OP said it was harder to get started, which implies they did not handle the details and have code not actually robust to all kinds of edge cases. Maybe they don’t need it but they probably do.
Message formatting
still have to do that with dbus
Service addressing
They’re Unix sockets, dude, they’re file paths in /run
Data marshalling
Still have to do that with dbus, also that’s the same thing as message formatting
Pubsub
Again, sockets. One application binds and many can connect (how often do you really need more than one application to respond to a method call? That’s a valid reason to use dbus in lieu of sockets, but how often do you need it?)
Method calling, marshalling of arguments and responses
They’re called “unix doors”, and that’s the third time you’ve said marshalling. As for that, language agnostic data marshalling is kind of a solved problem. I’m personally a fan of msgpack but JSON works too if you want to maximize compatibility. Or XML if you really want to piss off the people who interact with your API.
Broadcast and 1:1 messaging
Sockets and doors can only do 1:1, and that’s true enough, but it occurs to me that 99% of use cases don’t need that and thus don’t need dbus. dbus can still be used for those cases, but less load on dbus daemon = less load on system. Also you said that already with pubsub.
As for that blob at the bottom, again, who said anything about there not being a language agnostic library? It’d be a lot of work to make one, sure, but that doesn’t mean it’s impossible. Besides, most of the work has been done for you in the form of language agnostic marshalling libraries which as you said are like 50% of the problem. The rest is just syscalls and minor protocol standardization (how to reference FDs passed through the door in the msgpack data etc.)
And what I’ve just described isn’t a reimplementation of dbus without any of the good parts, it’s a reimplementation of dbus on top of the kernel instead of on top of a daemon that doesn’t need to be there.
You still have to do this with dbus
It can even start the receiving daemon if it is not yet running.
We have a tool for that, it’s called an init system.
The init system is for trusted system services that can talk directly to hardware. Unless you are working on a single-user system with no security concerns of any kind, you might consider using init to launch persistent user land or GUI processes.
DBus is for establishing a standard publish/subscribe communication protocol between user applications, and in particular, GUI applications. And because it is standard, app developers using different GUI frameworks (Gtk, Qt, WxWidgets, FLTK, SDL2) can all publish/subscribe to each other using a common protocol.
It would be certainly be possible to establish a standard place in the /tmp directory and a standard naming scheme for sockets and temporary files so that applications can obtain a view of other running applications and request to receive message from other applications using ordinary filesystem APIs, but DBus does this without needing the /tmp directory. A few simple C APIs replace the need for naming and creating your temporary files and sockets.
…systemd very much does use the init system to launch userland and GUI processes. That’s how GNOME works.
Dbus is for interprocess communication. The fact that its primary use case is communication between desktop applications is hardly relevant to its design. I don’t see how GUI frameworks are at all relevant, or how it would be possible to create an interprocess communication mechanism that only worked with one GUI framework without some heroic levels of abstraction violation (which I would not put past Qt, but that’s another story).
I don’t see why having an entire dbus daemon running in the background is better than having a cluttered /tmp or /run directory.
Let’s say you want to write a GUI for connecting to networks.
In the backend, you have NetworkManager, systemd-networkd, ConnMann, netctl, dhcpcd, …
Dbus could be a good way to expose a common API surface for clients.
/thread
But it’s not. Why is it better to have a different degree distribution?
…and why is that…better?
Programs still have to be written to accommodate the specific protocol that the program on the other end speaks, and dbus paths could translate pretty directly to subdirectories of /run. All adding dbus in the middle does is add a daemon where there doesn’t need to be one and force the programs to talk to each other through that rather than directly to each other
No, you see, additional layers of abstraction are always a good thing. Haven’t you been paying attention to the latest software memes? /s
Easy… decoupling. You hit the pause button on your keyboard, it does not need to “know” (in code or compile time or at runtime) what your music player is, and it can still pause it. Similarly, you can write a new media player, and not have to convince 1000 different projects to support or implement your custom api. https://en.m.wikipedia.org/wiki/Enterprise_service_bus
There’s nothing about dbus that makes decoupling easier, you can do it just as well with sockets. Pipewire and pulse both speak the same protocol, and they both rely on sockets, not dbus. The vast majority of the apps on my system don’t know or care that they’re speaking with pipewire instead of pulse. Read my comment here https://lemmy.world/comment/6284859
Did you read more than the first two words of my post? https://wiki.archlinux.org/title/MPRIS
Also, that same logic would “prove” that p2p has no value b/c you can do just as well with fixed servers and http.
My serious answer, not an argument: Use d-feet to inspect what’s available on the system and session buses. That’ll show the benefit of introspection and a common serialization mechanism.
About the security comments: Some access control mechanisms aren’t just allow/deny, and many need more than socket permissions. Those benefit from DBus policies, and PolicyKit integration helps for more complex needs. You can always DIY it, that’s Linux/FOSS life, but these are great tools to have in your toolbox. I’ll avoid credential passing via sockets whenever I can and have something else do it.
Great point about policies! Setting permissions on sockets only gets you so far… I guess if you really wanted to, you could create an individual socket for every method of every resource, and have granular permissions that way. But that would be quite messy
Sockets are effectively point-to-point communication. Dbus is a bus. Your question is similar to “what is the point of I2, or an ATA bus when directly wiring ICs gets the job done”. Both have different strengths and weaknesses.
Some heavy schooling happening in this thread. Glad to see it. Learning is a good thing. 🙌
I’m learning a lot, so I’m not a fan of the people flaming and downvoting OP for having genuine confusion. I want us to incentivize more posts like this.
Same here. I’ve always wondered what dbus actually was and I’m glad OP asked
Want to exchange information in json? plaintext? binary data? Sockets can do it.
This is exactly why you need something like dbus. If you just have a socket, you know nothing about how the data is structured, what the communication protocol is, etc. dbus defines all this.
In either case you still need to read the documentation of whatever daemon you’re trying to interface with to understand how it actually works. Dbus just adds the extra overhead of also needing to understand how dbus itself works. Meanwhile sockets can be explained in sixteen words: “It’s like a TCP server, but listening on a path instead of an ip and port”.
This makes no sense because dbus uses unix domain sockets.
It sounds like you don’t actually understand what dbus is.
It’s much easier to understand how dbus works once than to understand how every daemon you connect to works every time you interface with a new daemon.
but you still have to understand the methods available in every daemon
Yeah that’s the case with programming… well anything. This at least gives you a way to automatically receive all of that data from any app without excessive prior knowledge. With a small amount of info you can filter for specific events and create all kinds of robust functionality. That’s the power of a set protocol - it is to make things widely compatible with one another by only depending on the dbus protocol and app name. Otherwise you may need to depend on some shared objects which makes deployment and maintenance a total clusterfuck.
https://en.wikipedia.org/wiki/Coupling_(computer_programming)
You got downvoted for speaking the truth. You can’t talk to a dbus app without understanding how it communicates. You can’t talk to a sockets app without understanding how it communicates.
Look into history of object brokering in object oriented environments. I was around when KDE went from CORBA to DCOP to DBUS, but not involved in the decisions. Basically: object sharing between processes with security, type translation, and a few other things. In the Microsoft world, this was called “component object model” if my memory is correct.
DBUS is pretty nice for complex interactions.
Based on the various other descriptions of the DBUS features, I kept thinking “this sounds like a message passing model with a bit of CORBA hiding in there”. It’s got a bit of SLP and AMQP/MQTT to it, just on a local machine instead of a distributed network. It’s solving a lot of problems with service discovery, message passing structure, and separating transmission layer details from service API design. Raw sockets/pipes can always be used to pass data (it’s how DBUS does it!), but there’s additional problems of where to send the data and how to ensure data formatting that sockets/pipes do not have the capability of solving by design since they’re simple and foundational to how interprocess communication works in the kernel.
Dbus attempts to provide a standardized system for inter process communication (IPC) without the complexity of managing individual connections between each and every process that needs to share certain types of data.
But the implemenation has it’s fair share of issues and the attempted reimplementation even more so, because it wants to leverage everything important to Systemd instead.
Is the dbus spec too big/broad?
I don’t know but I’m interested in the answer.
New comments have appeared
Sometimes you want loose coupling.
Sockets are just streams of bytes - no defined structure to them at all. Dbus is about defining a common interface that everything can talk. That means when writing a program you don’t need to learn how every program you want to talk to talks over its own socket - just can just use a dbus library and query what is available on the system.
At least that is the idea - IMO its implementation has a lot to be desired but a central event bus IMO is a good idea. Just needs to be easy to integrate with which I think is what dbus fails at.
A great example is music player software - rather than every music player software creating its own socket and each having its own API to basically all do the same operations. So anything that want to just play/pause some music would need to understand all the differences between all the various different music applications. Instead with a central event bus system each music app could integrate with that and each application that wants to talk to a music app would just need to talk to the event bus and not need to understand every single music app out there.
But isn’t this the reason why client libraries to talk to programs behind sockets exist? Kinda like an SDK library behind an HTTP protocol API?
The Linux kernel is already a central event bus of sockets, btw
Wouldn’t this also be possible with plain sockets tho? To continue with your example of music players, the current standard is MPRIS, which uses dbus. But in an alternate universe, the people behind MPRIS could just have decided that music players shall create sockets at
/run/user/1000/mpris/[player name]that all speak the same standardized protocol. If a player wanted to add functionality beyond MPRIS, it could accept nonstandard requests on its socket, or create a new socket altogether for extended control.I just don’t see how this would require any more coordination between developers than the current solution. And I don’t see how dbus can save you from having to “understanding every single […] app out there”. If anything, it adds the overhead of learning how dbus itself works, on top of how a specific app’s dbus interface works.
Wouldn’t this also be possible with plain sockets tho?
You keep using this phrase. Given time and money anything is possible. Technically we don’t need to use http - every server could implement their own standard using raw sockets. You then could download a simple client library for every site!
With a well defined dbus interface your application can talk to any number of applications that implement that interface. Even those you didn’t know about it at time of development. It provides a structure for ipc other than “go fetch libblah” and also “libblarg” and “libfloob” and read all of their docs and implement each one separately.
Many toolkits make dbus usage simple. It’s also introspectable so very easy to explore or generate bindings for dynamically.
It’s pretty nice to use IME.
this is the real answer - programming language bindings
Anything is possible with sockets… and that is a meaningless statement. It is like saying you can build anything with bricks. Technically true, but missing all the important details of how.
In an alternative universe we could have done so many things differently to solve the same problems. But we don’t live there and in our universe dbus was the attempt to solve that problem among others. And yes you can create a standardization for music players easily enough - but what about notifications, and everything else? DBus tries to be a generic interface anything can talk over at a logical level - rather that just being the basic way two process can physically send bytes between each other.
Yes, of course, the sockets are the answer to everything (and BTW, d-bus uses sockets as well, e.g.
/run/dbus/system_bus_socketon my current system), but the problem is no standard for the communication over these sockets (or where is the socket located). For example, X11 developed one system of communicating over their socket, but it was used just by few X11 programs, and everybody else had their other system of communication. And even if an app found some socket, there was absolutely no standard how exactly should programs communicate over it. How to send more than just plain ASCII strings? Each program had to write their own serialization/deserialization code, their own format for marshalling binary data, etc. Now there is just one standard for those protocols, and even libraries with the standard (and well tested) code for it.There’s still no standard on how to use dbus so I don’t get your point
dbus can also start a program. For example when one notification was generated and no notification daemon is running, then dbus launch one to handle the request.
Doesn’t systemd have the ability to do this as well with unix sockets?
I posted this in another comment, but to me it just sounds like this autostart mechanism in dbus is just a poor re-implementation of an init system
Its so that your system can hold passengers
*can hold dpassengers
Get into de bus?
