Plan 9 is a uniquely complete operating system
posixcafe.org
posixcafe.org
Plan 9’s filesystem is a very simple network filesystem protocol to share files between systems. They are specifically using 9P2000.L.
They considered using Samba and SMB instead but can’t rely on Samba being installed and usable in the Linux guest OS and didn’t want to ship it because Samba is GPL licensed.
They picked Plan 9 because it’s much simpler to implement. Also Microsoft already had Plan 9 server code for some other Linux container project they’d done.
The \\wsl$\ path is handled in the Windows system by the MUP, an existing hook for network-like filesystems. They added a new one for Plan 9.
The $ is in the name so that it can’t be confused with a computer whose hostname is wsl.
The Plan 9 server in Linux communicates with the Windows Plan 9 client via a Unix socket. (Windows supports Unix sockets; who knew?)
Windows can access your Linux files even if no Linux is instance is running. There’s a new Windows service called LXSManagerUser that mediates user identity and permissions.Only since Windows 10 build 17063 (December 2017 pre-release) [0] [1], which was released as Windows 10 April 2018 Update. So for the first 25+ years of Windows' existence, it didn't.
And although it does implement the basic functionality, it is missing features found on mainstream Unix-like platforms, e.g. file descriptor passing (SCM_RIGHTS)
[0] https://devblogs.microsoft.com/commandline/af_unix-comes-to-...
Same with other hypervisors, virtualbox etc do the same. If you have docker installed on macOS it also uses 9p to share data with the host.
But IMO 9p is a terrible choice for this, particularly because it doesn’t support hard links. It breaks a lot of software like sccache etc which rely on hardlinks to work.
The reply from the plan9 devs on why this is the case hits staggering levels of arrogance:
> If you look at what a hard link is, you'll realize why they are not in Plan 9.
https://groups.google.com/g/comp.os.plan9/c/24mMVoy6wXA/m/JW...
I'd even suggest it reflects humility not arrogance.
It’s phrased as if to say “if you gave it a moments thought, you’d see I’m right”, which is the epitome of arrogance to me.
I get this impression from Rob Pike as well… I’m sure it comes from decades of being tired of arguing with people, but he comes off as utterly dismissive, as if to say “you either agree with me or you’re an idiot”. It doesn’t help that he continually throws shade on Linux (which is approximately infinity times more successful than Plan9) as you can see from other comments in the same thread. I don’t come away with a good impression of him or any of the other plan9/9front devs. Their whole attitude seems to be “everyone in the OS/systems world is dumber than us, and even too dumb to see why they’re dumber than us. We have a perfect system beyond any reproach and you’re an idiot if you disagree.”
>On Plan 9, the rule tends to be: if feature(X) can't be implemented in a way that works for everything, don't do it.
TBH I don't think hard links would be missed in Linux either, symlinks are good enough.
They really aren’t though. Anyone reading from a symlink needs access to the path it points to. If I have a network share with hard links pointing to common storage that’s outside the share’s root, clients are none the wiser and see them as normal files. With symlinks they would see links to paths they can’t see. It’s not really the same thing.
He is right however, you are also right: it's because Windows and Docker publish 9p server in a stupid way. It shouldn't be just the guest fs, you should be able to make any file server you like so hard links would be useless (as they are in plan9) because you would decide what filesystem organization layout you need.
You could make do with FUSE inside the virtualized OSes I guess.
One of the “stupid” ideas I have in my back-burner is to rewrite rio so that it works like Mac OS 7 (the platinum look with window shading), which in my mind was always a very sane and efficient way to manage windows — but time is not on my side…
I have one of my usual lists of resources for it on https://taoofmac.com/space/os/plan9 - comment here if it’s missing anything you particularly like.
This is IMO the biggest drawback. Why wouldn't any user want the software to be feature rich? In fact, looking at Plan 9, I often feel that the provided software is just a MVP.
Composable software is something I’m highly interested in. There were efforts in the 1990s to make desktop software more composable, such as COM from Windows and OpenDoc from Apple, but the desktop world is still dominated by large applications such as those that constitute Microsoft Office and the Adobe Creative Suite. It would have been a wonderful opportunity for the Linux desktop world to embrace components, but, alas, the community embraced OpenOffice, GIMP, and other large applications.
COM is everywhere on Windows, specially since Vista, as the WinDev regained the control they thought Longhorn was going to take away from them.
One of Powershell strengths is the easy access to COM, just like .NET, frameworks.
Linux could do the same with D-Bus, but alas so is the distributions wars, and hate on anything like proprietary OSes, that it only has a minor role on systemd, GNOME and KDE.
Imagine writing a huge complex program that is dependent of communication between smaller existing programs. Either you use the default programs that where shipped with all the different versions of OS'es with different distro's (never going to work, too many different versions of programs and their communication interfaces) or you ship certain fixed versions of all of the small programs that form your bigger program.
In case of the latter: why not just use libraries? It's basically the same thing with an easier API?
Maybe I'm missing something...
I’d guess the trick is that you’re not thinking small enough. GNU coreutils, etc. are not minimal by any means, but compared with even late-90s graphical desktop software they are still fairly compact, and you’re rarely using each and every tool at once. And the smaller the tool and the problem it targets, the more likely it is that the problem is mostly a solved one, so interface churn is less necessary.
I’m not sure every problem area is amenable to this—GUIs and things best expressed as GUIs seem particularly stubborn. (It would be sad if OLE/ActiveX was the best possible solution.) But some are, and few enough people are trying to expand the simplicity frontier for their real-world tasks in recent years that I don’t believe the state of the art of the 1980s is the farthest it can reach, either.
Except at the end of the day, the interface can be as simple as a stream of bytes over a socket. If I want a h264 decoder in my application I could just pipe a stream to a specific program made to decode it. That could be written in Python, in Rust, in C, in Go, etc. whereas dynamic libraries don't give you that freedom as you have to abide by the ABI defined by the host application.
Also there's the old viewpoint of "let's put everything in XML. Everything can read and write XML", and that seems similar to the argument of "everything can be an object that is exposed to other programs". The problem in both cases, objects have requirements for function call names, function arguments, nuanced conventions, as XML has specific tree structures. Interfacing programs must be wise to these conventions and sometimes nuances so it doesn't "just work". Windows is like - programs CAN share objects and interface BUT it needs developer work for program A to work with program B. Unix is like "programs X and Y, whatever they are, work together with no additional work". The magic of unix is that if there are 1000 programs, now there are 1,000,000 combinations that "just work" vs. 1,000,000 developer hours
Unix on the other hand recognizes that it's a lot easier to just say "everything is a text stream, and all programs are responsible for inputting and outputting only fairly clean text streams without extraneous output" . This is extended to "everything is a file", a file also typically made to be read into a clean text stream so it can feed into that whole ecosystem.
Once you've written some very basic boostrap tools, the "second generation" of stuff that adds convenience and flexibility are a lot simpler.
A trivial example: 20 seconds after you wrote "directory listing", someone will say "I want a directory listing, but sorted by date, and it would be awesome if it didn't immediately scroll past the end of my screen."
With Unix Philosophy tools, you might already have have a "sort" and "paginate" command, so it's just piping stuff together. They can do it themselves, or it will take 20 seconds to explain.
Without it, you're going to have to add additional options to "directory listing" (or parallel commands) to handle the sorting and pagination features. The tools get bigger and buggier for the same functionality.
Early Unix machines weren't much bigger than mid-80s PCs-- 512K of memory or less-- but offered a very rich command line experience compared with DOS machines of similar sizes.
Programs like database or CAD packages probably go monolithic because they're more "state dependent" than your usual command-line tools. "sort" and "more" can take their inputs from stdin and feed them out to stdout, and when they're done, forget everything with no damage.
That wouldn't work well for other packages. You could probably make a database or CAD system that worked as composable units, like `echo db.sql | db-query "select username from accounts where credit < 0" | xargs delete-account` or `echo image.dxf | add-circle -x 200 -y 400 -r 60 > image.dxf` But you'd spend a lot of time reloading and reparsing the same files. A persistent monolith that keeps the data file open and in whatever internal representation is most efficient is going to perform better.
Some use cases also have limited composability, because the user can only plan a few moves ahead. Tools that encourage interactive/experimental usage, like drafting software, might involve the user stopping every step or two to make sure they're staying on plan, and queuing up a series of commands could wreak havoc. Some of these packages ended up simulating composable tools through internal macro/scripting languages which still avoided the penalty of having to rely on the OS to orchestrate every single action.
Sorting the output of textual tools like ls requires parsing which can be non-trivial. It's easier to do it by using a modern structural shell such as nushell.
(Offhand, I've never touched Plan 9 but...) Hypothetically /proc/SOMEPID/db/DATABASE/SCHEMA/TABLE/various views which provide expressions in some order. Or /proc/SOMEPID/containerofthings/ and the directory listing is serviced by the application, as an enumeration of keys (filenames) to values (datasets). For a database the API would behave similarly to how ORMs operate since filesystems are inherently similar to objects.
It kind of feels a bit of cargo cult, praising it all the time.
It's pleasant to use a minimalist, viable product.
9front is not the only OS I use, but it is one of my daily drivers.
Additionally, since even minor versions introduce breaking changes, getting something from e.g. Python 1.6 to run on 3.12 is an exercise in trial and error, or unexpected surprises at some moment at runtime.
Would be interesting to hear what levels you think those are because as someone who has been learning and writing C++ and Python for over 20 years now, I'd say there's no level of Python which comes even close to the complexity of C++ at a comparative level.
> getting something from e.g. Python 1.6 to run on 3.12 is an exercise in trial and error, or unexpected surprises at some moment at runtime.
Fair enough. But then again: that's irrelevant for any new project written since about the start of the last decade. And to continue the C++ comparison: whereas Python 1 -> 2 -> 3 imo solved some real issues, the consecutive C++ standards never did this resulting in something which is, yes, backwards comptible (roughly - try getting 30-40 year old code getting to compile with /std:c++latest and /Wall - or look at all those tiny behavior changes between the last couple of standard iterations) but also seriously plagued by that as it holds back innovation. Modern C++ minus a lot of the old UB-prone stuff would definitely be better and less complex than what we have now.
Then just like C++, add the set of key implementations, CPython, Cython, PyPy, key libraries everyone uses (parallel to Boost).
Since you have such a wide Python experience, naturally you already printed out all the PDFs of Python documentation, and read them cover to cover.
I did such thing back on Python 2.0 days, and it only grew bigger.
How many pages was it again?
If you give the tiniest of Fs for your user, do not write your thing in python.
Perhaps vast differences in ergonomics and language-culture-fit but that's orthogonal/unrelated?
The supposed simplicity of Python over Perl has nothing to do with Perl being more feature-rich and that compromising its simplicity.
It seems like Plan9 would get closer to "Python simplicity" by adding features and extending interfaces. Which would be at conflict with the "minimalist MVP simplicity".
You present it as mutually exclusive. I believe it's the same word used for two fundamentally different aspects of software - less of a nit.
Complexity is in conflict with usable composition. Because of simple, small interfaces, the surface area needed for composition to cover all cases in a plan 9 environment is much smaller.
This allows for things like remote login to be implemented as a small number of 'mount' calls, git to serve repos in a way that can be scripted without writing gobs of porcelain, or sshnet to trivially replace the entire network stack in userspace (namespace by namespace), so that software doesn't need to implement features like socks proxy support.
The simplicity and uniformity of the interfaces and tools is an enabler.
But, you're right. I should have picked scheme or lua as an analogy. Scheme is a particularly good one: the simplicity of its syntax enables easier macro manipulation. How many special cases would you need to have if you were implementing a lisp style macro system for C++?
Smaller interfaces and implementations, lead to smaller sets of special cases and contexts to keep in mind, easier interposition and emulation, and faster debugging.
> You present it as mutually exclusive.
I honestly have no idea what you mean. What is the "it" that I am presenting, and what do you think I presented it as excluding?
> I honestly have no idea what you mean. What is the "it" that I am presenting, and what do you think I presented it as excluding?
Simplicity vs feature-rich. Your last follow-up mostly resolves that, I think.
BTW, I think you just sold me on diving into Plan9 some time soonish (:
I should probably make time look again in case I missed something or it's improved in the last decade.
Users want feature rich systems. Individual programs are best feature-complete, but focused on a single task and capable of cooperating with others when something out of scope is desired.
From my personal viewpoint: It's not easy to hack on large monoliths, even for senior software engineers. But if every logical piece of the monolith tries to be as small as it meaningfully could, the barriers are drastically lower.
Lastly, it's really not for general users. It's for (academic?) computer people who are dexterous and willing to try new things.
To talk about the article itself, the only reason plan 9 can achieve such a design is because it's developed and used by the same small group of people. If linux is a bazaar and BSDs are cathedrals, then 9front is a monastery's citadel. Another thing that isn't mentioned is that both linux and BSD (and pretty much anything based on posix) has a lot of third party software that would be hard to maintain along with the rest of the system, if the monks even include it to begin with. And that software could include something like jq which a lot of software depends on and would love to just assume it's there.
And really, what more does someone get from something like this over, say, having a more or less formal standard on what a true plan9 system includes and waving it in someone's face when they choose to ignore it? This is pretty much what modern unices do and it works out great in cases when it's actually important. Most people don't care what commit your system is built from as long as it works as their programs expect it to.
> ... has a lot of third party software that would be hard to maintain along with the rest of the system
This is the point that the article is trying to challenge. I think 9front proves that it's doable.
> Most people don't care what commit your system is built from as long as it works as their programs expect it to.
The former helps the later a lot. Everything is tested with each other and for a lot of functionality there is only one option.
A few years ago I thought about what it would take to implement a more conventional desktop GUI on top of Plan 9, but I’ve oscillated back and forth between wanting Plan 9 with a Mac-like desktop versus wanting a modern Lisp/Smalltalk machine (with object-oriented underpinnings instead of Plan 9’s “everything is a file” interfaces) with a Mac-like desktop.
Alef => Limbo => Go.
Being able to backport Go into Plan 9, doesn't make sense in this context, that isn't how historical evolution works.
Also even Inferno has the necessary C infrastructure to port Go, if someone hasn't done it already.
Though I do know that it puts some strong emphasis on mouse usage, something that for someone that grew to use the keyboard a lot like me (ironically, as a graphic designer) seems to be really awkward, to say the less. Its strengths seem to be its overlaying concepts and that it intended to be "the next gen Unix" - alas it won't take over for a myriad of reasons, and some would argue Unix-es have already borrowed some of its concepts for themselves.
(That said, there may have been an X server brought in at some point, but don't quote me on that. That's the least of anyone's problems in undertaking a CDE port though.)
Of course lots of it might be skill issue.
https://web.archive.org/web/20240728004832/https://posixcafe...
> Put in another way, if someone wanted the ability to touch every line of code (in the upstream sense), they would have to be a member of some non trivial amount of communities.
On a typical distro you can just download sources and start hacking, you don't need to be member of any community.
While something like Debian might not be monorepo in the strictest sense, on a conceptual level it is very close. They still have all the sources under their control and are not dependent on anything outside. They are at full liberty to accept or reject any patches regardless of where they come from, from "upstream" or "downstream".
This idea that distros are actually independent full-featured operating systems is an idea that I think is getting forgotten way too often. Distros are (or rather can be) much more than mere repackaging of upstream software.
The illustrate how I think Plan 9 is different in this regard. A patch for 9front could include a new feature for our compilers and then also show how useful it is by using it within other parts of our code. In plan 9 you can interact fully with every component.
I would say introducing a backdoor (xz) without downstream knowing is probably the biggest "privilege" you can have on a system or distribution no?
[1] Systems Software Research is Irrelevant (2000) [PDF]:
I always have the impression the discussion stops on a gas station the middle of the road, instead of on the destination.
Otherwise we could also go back to Windows NT with POSIX subsystem, fork it to a version where the UNIX experience was first class like on NeXTSTEP and macOS, and then praise Windows NT UNIX qualities, that weren't there in first place. (Naturally ignoring the access to source code issue).
From the man page: bullshit - assemble a stream of bullshit from words in a file
It will produce output like (ran it 3 times)
persistence firewall markup realtime-java callback-scale generator
virtual polling polling SQL out-scaling blockchain
converged converged singleton property self-signing-based element polling just-in-time control