To be charitable, Linus initially called Linux "a hobby, won't be big and professional like GNU" and "NOT protable [sic](uses 386 task switching etc), probably never will support anything other than AT-harddisks".
I don't think this is an accurate description of why linux was created. But I do think it's time to design something that better fits the needs of the modern world.
Our posts were solely why it was created but rather the design choices made as part of building it from the beginning to the present day.about the Linux kernel itself and not software used on top of it. We weren't talking about
Yes, smaller and verifiable kernels would help a lot against buggy and malicious code. Let alone stability and robustness.
Then, I for one, would not change C for anything else. The problem is not the language (which is just a tool), the problem is the coding (and the coders).
The UNIX command line userland philosophy was not to architect a rigid user facing system in advance but enable ad hoc, as needed edge case problem solving.
Linus solved the problem of not having UNIX kernel sources by designing the Linux kernel from his experience as a UNIX user solving problems as they appeared not through advanced speculation and architecture. The connection seems obvious to me.
And this doesn't say anything that wasn't obvious this whole time. It's just in the spirit of the don't architect based on speculation but reality, LLMs didn't exist and didn't need to be accounted for. No corporation, nation state or script kiddie had enough compute back then.
This is all like config DSLs; it all been coming up again and again the entire 30ish I been paying attention to Linux/Unix.
Never contributed code back then though. I was more interested in EE, building electronics back then but was curious enough about Linux to build kernels and eventually Slackware. Really not all that interested in code today even; it's just how it works
Unix ans Linux are 20+ years apart with all the changes that happened to hardware, software and even to C: they could hardly have the same general architecture.
Of course, I am talking about the kernel, not the entire OS.
Linux stood up on the shoulders of a giant like Minix, Unix was basically designed and coded from scratch.
We weren't talking about motivations or goals for making an open source kernel. Why Linus wanted to make an open source kernel is not the same topic as why he has been so strongly against microkernels for years.
We were talking about the chosen architecture and design choices. Those have been explicitly made to prioritize performance and features over robustness and security. More robust and secure ways of doing things have been explicitly rejected and often ridiculed.
> Linux was created as a rejection of robust software architecture and tools. It's a monolithic kernel with no internal isolation written as memory unsafe code. A typo in an obscure driver for unused functionality can easily create a fully privileged code execution vulnerability.
Which is, of course, why GrapheneOS is based on minix, right?
Edit: Really, this is just GOS being GOS; only ever Security (by their own very specific, poorly defined definition thereof), with nothing else allowed to be a consideration. There is a reason the industry went with monolithic kernels when other options existed. Even today in 2026, we've merely worked out how to reduce the performance hit of microkernels, and in the 90s Torvalds was very justified in preferring to ship something that would actually run decently on extant hardware.
Edit2: To be fair. I like minix. I have actually tried to make HURD into a usable system. I want these things to work. But there are harsh tradeoffs in play that are still painful even after decades of work, and this kind of post does not exactly convince people that GOS is making level-headed evaluations of technical choices.
Arguably the only reason GOS is relevant is because they do exactly what you criticize them for.
Lots of options exist for people who do what you describe as the "right" way. People who aren't in the security-at-all-costs group have plenty to choose from.
Are you saying that software that promotes security at the price of usability should not be permitted to exist, nor that people should, whatever their use case, desire it?
> Are you saying that software that promotes security at the price of usability should not be permitted to exist, nor that people should, whatever their use case, desire it?
I'm saying maybe they could desire and work on their thing without having to constantly yell about how everyone and everything else is awful and sucks.
You're falsely accusing us of what you're doing. You lie about what we've said and done. Your posts contribute no value to this discussion. It's clear you can't participate in threads about GrapheneOS in a constructive way and should stop.
We are literally having this conversation because you opened with
> Linux was created as a rejection of robust software architecture and tools. It's a monolithic kernel with no internal isolation written as memory unsafe code. A typo in an obscure driver for unused functionality can easily create a fully privileged code execution vulnerability.
Anytime anyone mentions other ROMs like LineageOS you immediately jump in to proclaim them insecure. The shoe fits. If you don't like being called out for slinging mud at other projects including those that your project is built on, feel free to stop doing it.
We were participating in a discussion about a discussion with high importance to protecting people's privacy and security. You're participated in threads about GrapheneOS in bad faith misrepresenting and directly lying about what we've said and done. You're continuing vile personal attacks on our team including continuing to try to portray as being insane and delusional as you've done many times before. You aren't contributing anything constructive to discussions about it but rather are trying to use it as a platform to harm the project and our team.
(I'm going to ignore the wild personal attacks in this comment.)
> We were participating in a discussion about a discussion with high importance to protecting people's privacy and security.
Okay? How do you get from there to
> Linux was created as a rejection of robust software architecture and tools.
?
Like, the rest of this is completely reasonable and (materially) correct:
> It's a monolithic kernel with no internal isolation written as memory unsafe code. A typo in an obscure driver for unused functionality can easily create a fully privileged code execution vulnerability.
If you'd written "Linux was created when hardware was weaker so it made trade-offs against security" I wouldn't mind. If you said "Linux was created to favor performance over security" I'd quibble a little but that'd be a reasonable argument. If you'd just said "Linux prefers performance over security and we prefer the opposite so we're investigating a migration to seL4" I'd agree and be super interested. Why would you jump straight to "a rejection of robust software architecture and tools"? What productive discussion is going to come from that?
> Why would you jump straight to "a rejection of robust software architecture and tools"?
Linux has explicitly rejected isolation within the kernels for years and heavily attacked it as an approach for any purpose. It's a core part of the project's historical philosophy and messaging from the beginning until the recent past. Linux is changing now and Linus is quite open to new ways of doing things but that's not what's being discussed.
How is a project which has essentially campaigned against microkernels on many occasions throughout the years not doing exactly that? Linux could already have multiple forms of internal isolation for components. It has immense complexity throughout the codebase for very niche purposes but not basic measures for robustness and security.
Security features were expected to be self-contained without requiring major overhauls to the code. The work on security was repeatedly ridiculously and downplayed as not important in an official capacity. Things have changed a lot in the past few years and especially this year where it has been so clearly demonstrated the status quo is unacceptable.
> What productive discussion is going to come from that?
The context was someone calling the Linux kernel and other projects AI slop for using AI models in any way in a thread about access to frontier models for cybersecurity work where we posted about why we need those. Our response was about the poor state of core infrastructure projects we ship including the Linux kernel and the immense scale of how much needs to be fixed. The topic was why these tools are needed to improve the status quo. We brought up how the Linux kernel is heavily shifting their approach and considering major improvements to the architecture and tools due to it too. If you actually read what we wrote, it's quite optimistic about major changes due to the status quo being widely shown to be so awful.
> (I'm going to ignore the wild personal attacks in this comment.)
You're continuing to heavily misrepresent what we've said and done as you've been doing for a long time across these things. Pointing that out is not a personal attack. Your post above has a poorly veiled form of personal attacks you've made before much more blatantly. It's clear that when you start trying to claim our team isn't in their right mind or isn't rational that you're participating in ongoing bullying and harassment you've supported and engaged in for a long time. It's absolutely vile and we aren't going to respond to your posts as if you aren't doing it and haven't been doing it for a long time.
I think most users of GOS don't have a stick by which to measure the security claims of GOS. And I'm guessing most are there for privacy, not security.
You're lying about what we've said and done. You're lying about what we've built and are building. Your posts in discussions about GrapheneOS are irrational attacks on the project. You post in threads about it maliciously to attack the project and team in any way you can. You've consistently engaged in vile personal attacks towards our team and supported harassment. You're referencing those things here.
Same old narrative about the entire world attacking you. No surprise people feel forced to steer away from your project by choosing less toxic ones.
In the same time you silence people quoting your public takes on your discussion board and whining about other platforms (like Hacker News) to express their opinion freely. Recent example thereof: https://grapheneos.social/@GrapheneOS/117566606444951363
The posts were justified. I was one of the people who was involved in the email chain mentioned in the posts there. We reported someone who was steering people to actual harassment content. We provided proof, and the moderators did nothing. I've seen personal attacks go without being moderated many times.
It's also strange how there's often this double standard. People are allowed to attack the project, but we're not allowed to call out platforms that permit it? And is there something about those posts that you found incorrect?
Perhaps processor architectures like the https://en.wikipedia.org/wiki/Intel_iAPX_432 which do away with such coarse grained and therefore dangerous abstractions by using capability-based addressing should have succeeded earlier too. This is how modern programming languages should be structured too, instead of giving no guarantees to programmers (and therefore their users) about what (parts of) programs are able to access like they do now.
Do you have a way to do it on the hardware people actually have that doesn't destroy performance every time the CPU context switches? You can always move all your filesystems to FUSE if you like the performance...
> Yes, it works alright for Windows, macOS, iOS, iPadOS, Android, Fuchsia.
Other than Fuchsia (I assume), all of those are at most hybrid systems with the performance-sensitive drivers still in ring 0.
> And all those cars, rockets, factories powered by QNX, INTEGRITY,....
Yes, you can get away with a lot if you're willing to sacrifice performance for safety.
> Ah, and those Switch games, nothing really that relevant.
That actually is an interesting point. Do you know how that works? I'd guess either they're doing something fun in hardware or they've actually got something going on in software worth considering. (Unless the thing they're doing is impractical to port to more general-purpose systems)
> Other than Fuchsia (I assume), all of those are at most hybrid systems with the performance-sensitive drivers still in ring 0.
Not really.
Apple iDevices don't allow third parties to ship drivers to ring 0, all third party drivers are in userspace.
While macOS still allows for kexts, the official roadmap is that ring 0 is only for Apple, everyone else has to move into userspace drivers.
Android only allows ring 0 drivers for "legacy" Linux kernel drivers, all other drivers must use Android IPC since Android 8.
Windows 11 runs on top of Hyper-V, drivers run virtualised (Device Guard) and do several key kernel components (Secure Kernel), everything together is known as Virtualization-Based Security.
These features are also available on Windows 10, however there they need to be explicitly enabled.
Additionally, since Windows 8 several drivers have been pushed into UMDF, and since the whole fallout in 2024, that there is a new framework on the way for security folks to take their stuff out of the kernel.
Across all three vendors, the amount of application sandboxing keeps going beyond a plain Flatpak or Snap can do, and there is a lot more going on with XPC, Android IPC, COM (out-proc) than D-Bus gets used for.
Horizon works, because Nintendo is not doing Minix, like everyone does when they talk about microkernels, rather taking the lessons out of operating systems like QNX on how to pass capabilities around.
None of the core Linux kernel or drivers are written in Rust. Drivers can theoretically be written in Rust but there's immense resistance to landing the bindings needed to do it. In practice, a few core Android drivers were ported to Rust and in very rare cases new drivers for a previously unsupported GPU, etc. are using it. The positive impact from such a tiny portion of the kernel being written in Rust is negligible. A lot of the cost of supporting it has been paid without reaping the benefits.
Haskell was a hyped language with new features and a promise of safety. It was popular for a few years and then died out. It's best ideas were stolen by other languages and it stopped making sense to use in production.
Rust feels very similar, and while things could always be different (Rust makes sense for vibe-coders) we won't know if it's here to stay until the future arrives.
I love this discussion, because it demonstrates what the market/developers actually care about.
Anyone who studies kernel development long enough will eventually say micro kernels are much safer, we have better languages now, or there are better ways to access OS resources, and they’re sort of right. But the reason Linux succeeds is because it focuses on being the fastest, fully featured kernel, and that’s what people actually wants. There are dozens of alternatives that nobody uses because at the end of the day, nobody cares if your language is “safer” and your drivers don’t have kernel access if it’s 10% slower when reading IO. They’ll just stick with Linux
I strongly doubt this is (still) true. Most devices are networked now. What is 10% less performance, in exchange for proper security? I would switch in a heartbeat. The only problem is that existing microkernel projects are not exactly usable yet. That their research, development and adoption are not more strongly funded and supported seems strange to me. As a programmer, OF COURSE i'd prefer to work with systems that gave me fine grained guarantees about resource (file system, network, CPU, memory) access, -isolation and -consumption. When driving, one doesn't go as fast as one can either, if it would risk other's and one's own lives - that would be irresponsible.
Microkernels are incredibly widely used. They're used on many of the components used in a modern computer including radios for cellular, Wi-Fi/Bluetooth, GNSS, NFC, etc. Many components are now written in a more secure way than mainstream general purpose operating systems. macOS and iOS have also gradually been moving more towards microkernel design by moving code out of the kernel. They started from a foundation where that was part of the design but not heavily used.
The main reason Linux succeeds is network effects. Once everyone has been coding to Linux APIs for upwards of 35 years, the ability to move to something else is incredibly limited. Even if a new OS attracted 100% mind share for all new development, it'd take many, many years for it to approach even the beginnings of ubiquity that Linux had in the early 2000s.
Replacing Linux (and likely having to change some parts of POSIX as well) would be an even taller order than rewriting C code in some other language. At least for userland rewrites, you have parity and an API (POSIX, other libraries in userland) to code to.
It might still be a good idea to do both of those, mind you. But let's not fool ourselves: the main reasons the incumbents are the incumbents is that it's hard, likely-to-fail, and incredibly slow work to replace them.
Replacing the Linux kernel ABI isn't necessary to have a microkernel written in a memory safe language. It can provide the same ABI. gVisor is not what we mean but shows it's doable.
See iDevices (what matters is the Objective-C/Swift userland) and Android (what matters is the Java/Kotin/C17/C++17 userland), it is doable, however it requires lots of money and willingness to push it through no matter what.
A great pragmatical axiom of security is that if it doesn't come cheap in terms of speed and convenience, it's likely going to fail in the end. People will always pick practicality over security in the real world: if it takes too much time to lock and unlock the front door, the average person will just leave it ajar.
Linux was a monolithic kernel because in 1991 microkernel were in earnest 25 years too early. Computers didn't really have the extra performance to spend for things like IPC and stuff. lots of other kernel designs that started back then like NT and Nextstep (was the NextStep kernel called XNU already? I guess not) all tried the microkernel route and then ended up doing stuff like networking at kernel level anyway because it just didn't work well enough.
arguably nowadays we have extra headroom to spare so stuff like containers, sandboxing, userspace drivers are becoming more and more commonplace.
Yes, I never realised that Mach had become a microkernel only after version 3, before it was a hybrid. NeXT and then Apple simply continued the older lineage which mixed BSD with userspace bits
> What's simplistic about it? The industry could have heavily adopted microkernels and memory safe languages 30 years ago.
Ignoring the pesky little problem that these microkernels and languages did not exist 30 years ago. It was Linux with their garbage drivers that enabled people to use it in their computers. It is very easy to imagine something rather than creating it in reality.
Linux only mattered to me in 1995, because doing the UNIX work assignments from the university DG/UX and Solaris timesharing servers wasn't possible on the joke of Windows NT POSIX subsystem.
Had Microsoft been serious about their POSIX support, I would not have bothered one second.
It wasn't FOSS nor hardware support, rather not having to travel to campus laboratory and fight for a terminal, losing a couple of hours.
Ignoring the pesky little problem that these microkernels and languages did not exist 30 years ago.
L4, which had much faster IPC than Mach, started in 1996. QNX4 was released in 1990, which had POSIX support (and was a microkernel before that) and was used in many critical embedded systems. There were a lot of exciting microkernel projects in the mid to late 90s, but they never gained much traction because of the network effects that Linux already had (drivers, etc.).
Memory-safe languages go way back before the 90ies. ML is from the 70ies and was quite performant early on. Rust can in many ways be traced back to ML (despite not being a functional language). Not only was the original Rust bootstrapping compiler written in OCaml, is also uses many concepts from ML (algebraic data types, type inference, etc.).
Doh, a dude's opinion about something done completely for free as a hobbyist 30 years ago, that had world success, actually killing the robust corporate "only a lab can afford" alternative. Give yourself some peace
Unix itself was created as a rejection of robust software architecture and tools.
"I remarked to Dennis that easily half the code I was writing in Multics was error recovery code. He said, "We left all that stuff out of Unix. If there's an error, we have this routine called panic, and when it is called, the machine crashes, and you holler down the hall, 'Hey, reboot it.'"
-- Tom Van Vleck
Not that this is a bad thing, but unix was created as a hacker's toy, a small elegant system to play games on, This is why it is so fun to use. I mean, it is not really small or elegant anymore. but that is how it started, not as a robust correct business system. That is also why it got so popular. "A small hackable lowcost operating system with full source" collages: yes please, and at a critical time in computing history, a very influential generation of hackers learned on it. Something now is better than perfect that does not exist, or that old cliche, worse is better.
Lets remind ourselves that initially, DoD assigned an higher security score to Multics than UNIX, with PL/I proper strings and arrays, with bounds checking available (configurable), as one of the key reasons.
> Speaking as someone who has delved into the intricacies of PL/I, I am sure that only Real Men could have written such a machine-hogging, cycle-grabbing, all-encompassing monster. Allocate an array and free the middle third? Sure! Why not? Multiply a character string times a bit string and assign the result to a float decimal? Go ahead! Free a controlled variable procedure parameter and reallocate it before passing it back? Overlay three different types of variable on the same memory location? Anything you say! Write a recursive macro? Well, no, but Real Men use rescan. How could a language so obviously designed and written by Real Men not be intended for Real Man use?
Interestingly, at least for me, the author of the original "worse is better" series of essays (they're still online) has flip flopped many times on his instance.
I think it speaks to the core of the "software crisis" literature which is its own genre. Big Ball of Mud, No Silver Bullet, The Three Cultures of Computer Science and so on.
Correctness and expediency are at odds and we have to make choices along these trade-offs constantly. The costs each imply accumulate. I don't think there can ever be a simple answer to the question of how to design and run complex systems. Every facile argument about this or that tool chain devolves into flame wars precisely because these arguments are about "the right way" (or the wrong way) to design and run complex systems and there simply isn't any (right or wrong).
"Systemantics" is a perpetual gem that should be High School required reading.
The original methodology is arguably the riskiest aspect. By not including comprehensive automated testing and verification infrastructure to guarantee correctness, it's basically fallible humans doing meatcloud beta testing "works for me".
Another consideration is design: everything having access to everything as a monolithic kernel running in a privileged execution environment (ring 0) is an inherently unsafe and insecure design, however it offers apparent simplicity compared hybrid microkernels that have to coordinate messy transactions across multiple areas of responsibility and organize the exposure/control of particular data and operations that would otherwise be exposed in a monolith.
My 2 centidollars: a basic kernel should be shrunk to include as little as possible and formally prove that (under all combinations of configuration options that change it). Then, almost all functionality should be loaded as a module with an accompanying allowlist metadata of granted functions/data structures. On top of that, modules should be thoroughly testable and also provably correct.
65 comments
[ 2.7 ms ] story [ 24.3 ms ] threadWe need another Bell Labs.
Yeah I think they're conflating UNIX-like general philosophy and Linux.
The UNIX method has been shown to have many upsides.
Though implementations due to slow old computers without fancy GPUs are user hostile
It's all computer state so IMO applying the same concept and principle is more valuable a mindset than something more specific "Linux 4eva!"
I don't think they are, actually; they appear to be arguing microkernel vs monolithic kernel, and we have unix-likes based on microkernels.
Then, I for one, would not change C for anything else. The problem is not the language (which is just a tool), the problem is the coding (and the coders).
The UNIX command line userland philosophy was not to architect a rigid user facing system in advance but enable ad hoc, as needed edge case problem solving.
Linus solved the problem of not having UNIX kernel sources by designing the Linux kernel from his experience as a UNIX user solving problems as they appeared not through advanced speculation and architecture. The connection seems obvious to me.
And this doesn't say anything that wasn't obvious this whole time. It's just in the spirit of the don't architect based on speculation but reality, LLMs didn't exist and didn't need to be accounted for. No corporation, nation state or script kiddie had enough compute back then.
This is all like config DSLs; it all been coming up again and again the entire 30ish I been paying attention to Linux/Unix.
Never contributed code back then though. I was more interested in EE, building electronics back then but was curious enough about Linux to build kernels and eventually Slackware. Really not all that interested in code today even; it's just how it works
Of course, I am talking about the kernel, not the entire OS.
Linux stood up on the shoulders of a giant like Minix, Unix was basically designed and coded from scratch.
The fact that we now have multi gigabyte RAM devices, does not mean that the kernel can go multi megabyte.
I'm also think that CPU designers should talk more to kernel developers: Linus is right on this topic.
We weren't talking about motivations or goals for making an open source kernel. Why Linus wanted to make an open source kernel is not the same topic as why he has been so strongly against microkernels for years.
We were talking about the chosen architecture and design choices. Those have been explicitly made to prioritize performance and features over robustness and security. More robust and secure ways of doing things have been explicitly rejected and often ridiculed.
Which is, of course, why GrapheneOS is based on minix, right?
Edit: Really, this is just GOS being GOS; only ever Security (by their own very specific, poorly defined definition thereof), with nothing else allowed to be a consideration. There is a reason the industry went with monolithic kernels when other options existed. Even today in 2026, we've merely worked out how to reduce the performance hit of microkernels, and in the 90s Torvalds was very justified in preferring to ship something that would actually run decently on extant hardware.
Edit2: To be fair. I like minix. I have actually tried to make HURD into a usable system. I want these things to work. But there are harsh tradeoffs in play that are still painful even after decades of work, and this kind of post does not exactly convince people that GOS is making level-headed evaluations of technical choices.
Lots of options exist for people who do what you describe as the "right" way. People who aren't in the security-at-all-costs group have plenty to choose from.
Are you saying that software that promotes security at the price of usability should not be permitted to exist, nor that people should, whatever their use case, desire it?
I'm saying maybe they could desire and work on their thing without having to constantly yell about how everyone and everything else is awful and sucks.
> Linux was created as a rejection of robust software architecture and tools. It's a monolithic kernel with no internal isolation written as memory unsafe code. A typo in an obscure driver for unused functionality can easily create a fully privileged code execution vulnerability.
Anytime anyone mentions other ROMs like LineageOS you immediately jump in to proclaim them insecure. The shoe fits. If you don't like being called out for slinging mud at other projects including those that your project is built on, feel free to stop doing it.
> We were participating in a discussion about a discussion with high importance to protecting people's privacy and security.
Okay? How do you get from there to
> Linux was created as a rejection of robust software architecture and tools.
?
Like, the rest of this is completely reasonable and (materially) correct:
> It's a monolithic kernel with no internal isolation written as memory unsafe code. A typo in an obscure driver for unused functionality can easily create a fully privileged code execution vulnerability.
If you'd written "Linux was created when hardware was weaker so it made trade-offs against security" I wouldn't mind. If you said "Linux was created to favor performance over security" I'd quibble a little but that'd be a reasonable argument. If you'd just said "Linux prefers performance over security and we prefer the opposite so we're investigating a migration to seL4" I'd agree and be super interested. Why would you jump straight to "a rejection of robust software architecture and tools"? What productive discussion is going to come from that?
Linux has explicitly rejected isolation within the kernels for years and heavily attacked it as an approach for any purpose. It's a core part of the project's historical philosophy and messaging from the beginning until the recent past. Linux is changing now and Linus is quite open to new ways of doing things but that's not what's being discussed.
How is a project which has essentially campaigned against microkernels on many occasions throughout the years not doing exactly that? Linux could already have multiple forms of internal isolation for components. It has immense complexity throughout the codebase for very niche purposes but not basic measures for robustness and security.
Security features were expected to be self-contained without requiring major overhauls to the code. The work on security was repeatedly ridiculously and downplayed as not important in an official capacity. Things have changed a lot in the past few years and especially this year where it has been so clearly demonstrated the status quo is unacceptable.
> What productive discussion is going to come from that?
The context was someone calling the Linux kernel and other projects AI slop for using AI models in any way in a thread about access to frontier models for cybersecurity work where we posted about why we need those. Our response was about the poor state of core infrastructure projects we ship including the Linux kernel and the immense scale of how much needs to be fixed. The topic was why these tools are needed to improve the status quo. We brought up how the Linux kernel is heavily shifting their approach and considering major improvements to the architecture and tools due to it too. If you actually read what we wrote, it's quite optimistic about major changes due to the status quo being widely shown to be so awful.
> (I'm going to ignore the wild personal attacks in this comment.)
You're continuing to heavily misrepresent what we've said and done as you've been doing for a long time across these things. Pointing that out is not a personal attack. Your post above has a poorly veiled form of personal attacks you've made before much more blatantly. It's clear that when you start trying to claim our team isn't in their right mind or isn't rational that you're participating in ongoing bullying and harassment you've supported and engaged in for a long time. It's absolutely vile and we aren't going to respond to your posts as if you aren't doing it and haven't been doing it for a long time.
In the same time you silence people quoting your public takes on your discussion board and whining about other platforms (like Hacker News) to express their opinion freely. Recent example thereof: https://grapheneos.social/@GrapheneOS/117566606444951363
It's also strange how there's often this double standard. People are allowed to attack the project, but we're not allowed to call out platforms that permit it? And is there something about those posts that you found incorrect?
Perhaps processor architectures like the https://en.wikipedia.org/wiki/Intel_iAPX_432 which do away with such coarse grained and therefore dangerous abstractions by using capability-based addressing should have succeeded earlier too. This is how modern programming languages should be structured too, instead of giving no guarantees to programmers (and therefore their users) about what (parts of) programs are able to access like they do now.
Do you have a way to do it on the hardware people actually have that doesn't destroy performance every time the CPU context switches? You can always move all your filesystems to FUSE if you like the performance...
And all those cars, rockets, factories powered by QNX, INTEGRITY,....
Ah, and those Switch games, nothing really that relevant.
Other than Fuchsia (I assume), all of those are at most hybrid systems with the performance-sensitive drivers still in ring 0.
> And all those cars, rockets, factories powered by QNX, INTEGRITY,....
Yes, you can get away with a lot if you're willing to sacrifice performance for safety.
> Ah, and those Switch games, nothing really that relevant.
That actually is an interesting point. Do you know how that works? I'd guess either they're doing something fun in hardware or they've actually got something going on in software worth considering. (Unless the thing they're doing is impractical to port to more general-purpose systems)
Not really.
Apple iDevices don't allow third parties to ship drivers to ring 0, all third party drivers are in userspace.
While macOS still allows for kexts, the official roadmap is that ring 0 is only for Apple, everyone else has to move into userspace drivers.
Android only allows ring 0 drivers for "legacy" Linux kernel drivers, all other drivers must use Android IPC since Android 8.
Windows 11 runs on top of Hyper-V, drivers run virtualised (Device Guard) and do several key kernel components (Secure Kernel), everything together is known as Virtualization-Based Security.
These features are also available on Windows 10, however there they need to be explicitly enabled.
Additionally, since Windows 8 several drivers have been pushed into UMDF, and since the whole fallout in 2024, that there is a new framework on the way for security folks to take their stuff out of the kernel.
Across all three vendors, the amount of application sandboxing keeps going beyond a plain Flatpak or Snap can do, and there is a lot more going on with XPC, Android IPC, COM (out-proc) than D-Bus gets used for.
Horizon works, because Nintendo is not doing Minix, like everyone does when they talk about microkernels, rather taking the lessons out of operating systems like QNX on how to pass capabilities around.
https://news.ycombinator.com/item?id=17534593
What do you mean by that?
Rust feels very similar, and while things could always be different (Rust makes sense for vibe-coders) we won't know if it's here to stay until the future arrives.
Anyone who studies kernel development long enough will eventually say micro kernels are much safer, we have better languages now, or there are better ways to access OS resources, and they’re sort of right. But the reason Linux succeeds is because it focuses on being the fastest, fully featured kernel, and that’s what people actually wants. There are dozens of alternatives that nobody uses because at the end of the day, nobody cares if your language is “safer” and your drivers don’t have kernel access if it’s 10% slower when reading IO. They’ll just stick with Linux
Replacing Linux (and likely having to change some parts of POSIX as well) would be an even taller order than rewriting C code in some other language. At least for userland rewrites, you have parity and an API (POSIX, other libraries in userland) to code to.
It might still be a good idea to do both of those, mind you. But let's not fool ourselves: the main reasons the incumbents are the incumbents is that it's hard, likely-to-fail, and incredibly slow work to replace them.
Linux was a monolithic kernel because in 1991 microkernel were in earnest 25 years too early. Computers didn't really have the extra performance to spend for things like IPC and stuff. lots of other kernel designs that started back then like NT and Nextstep (was the NextStep kernel called XNU already? I guess not) all tried the microkernel route and then ended up doing stuff like networking at kernel level anyway because it just didn't work well enough.
arguably nowadays we have extra headroom to spare so stuff like containers, sandboxing, userspace drivers are becoming more and more commonplace.
If it had been given half a chance, this would be true. But not even trying is not the same as failing by a long shot.
https://en.wikipedia.org/wiki/Mach_kernel
Ignoring the pesky little problem that these microkernels and languages did not exist 30 years ago. It was Linux with their garbage drivers that enabled people to use it in their computers. It is very easy to imagine something rather than creating it in reality.
Had Microsoft been serious about their POSIX support, I would not have bothered one second.
It wasn't FOSS nor hardware support, rather not having to travel to campus laboratory and fight for a terminal, losing a couple of hours.
L4, which had much faster IPC than Mach, started in 1996. QNX4 was released in 1990, which had POSIX support (and was a microkernel before that) and was used in many critical embedded systems. There were a lot of exciting microkernel projects in the mid to late 90s, but they never gained much traction because of the network effects that Linux already had (drivers, etc.).
Memory-safe languages go way back before the 90ies. ML is from the 70ies and was quite performant early on. Rust can in many ways be traced back to ML (despite not being a functional language). Not only was the original Rust bootstrapping compiler written in OCaml, is also uses many concepts from ML (algebraic data types, type inference, etc.).
https://news.ycombinator.com/newsguidelines.html
"I remarked to Dennis that easily half the code I was writing in Multics was error recovery code. He said, "We left all that stuff out of Unix. If there's an error, we have this routine called panic, and when it is called, the machine crashes, and you holler down the hall, 'Hey, reboot it.'"
-- Tom Van Vleck
Not that this is a bad thing, but unix was created as a hacker's toy, a small elegant system to play games on, This is why it is so fun to use. I mean, it is not really small or elegant anymore. but that is how it started, not as a robust correct business system. That is also why it got so popular. "A small hackable lowcost operating system with full source" collages: yes please, and at a critical time in computing history, a very influential generation of hackers learned on it. Something now is better than perfect that does not exist, or that old cliche, worse is better.
> Speaking as someone who has delved into the intricacies of PL/I, I am sure that only Real Men could have written such a machine-hogging, cycle-grabbing, all-encompassing monster. Allocate an array and free the middle third? Sure! Why not? Multiply a character string times a bit string and assign the result to a float decimal? Go ahead! Free a controlled variable procedure parameter and reallocate it before passing it back? Overlay three different types of variable on the same memory location? Anything you say! Write a recursive macro? Well, no, but Real Men use rescan. How could a language so obviously designed and written by Real Men not be intended for Real Man use?
I think it speaks to the core of the "software crisis" literature which is its own genre. Big Ball of Mud, No Silver Bullet, The Three Cultures of Computer Science and so on.
Correctness and expediency are at odds and we have to make choices along these trade-offs constantly. The costs each imply accumulate. I don't think there can ever be a simple answer to the question of how to design and run complex systems. Every facile argument about this or that tool chain devolves into flame wars precisely because these arguments are about "the right way" (or the wrong way) to design and run complex systems and there simply isn't any (right or wrong).
"Systemantics" is a perpetual gem that should be High School required reading.
Another consideration is design: everything having access to everything as a monolithic kernel running in a privileged execution environment (ring 0) is an inherently unsafe and insecure design, however it offers apparent simplicity compared hybrid microkernels that have to coordinate messy transactions across multiple areas of responsibility and organize the exposure/control of particular data and operations that would otherwise be exposed in a monolith.
My 2 centidollars: a basic kernel should be shrunk to include as little as possible and formally prove that (under all combinations of configuration options that change it). Then, almost all functionality should be loaded as a module with an accompanying allowlist metadata of granted functions/data structures. On top of that, modules should be thoroughly testable and also provably correct.