Tilck: A tiny Linux-compatible kernel | Hacker News Reader

Tilck: A tiny Linux-compatible kernel

github.com

286 pointschubot58 comments

Nice! Looks like an interesting mid-point between xv6 (modern reimplementation of Sixth Edition Unix from 1975 e.g. Lion book) and a full Linux kernel.

Great to see it runs on the LicheeRV Nano, a $9 RISC-V board with 1.0 GHz 64 bit CPU (C906) with MMU, FPU, 128 bit vector unit (RVV draft 0.7.1) and 256 MB of in-package DDR3. That's comparable to a midlife Pentium III or PowerPC G4.

It should be a very easy port to the Milk-V Duo 256M (same SG2002 SoC) or Duo S (SG2000 with 512 MB RAM for $9.90) or original Duo (closely related CV1800B with 64 MB RAM for $5).

No network or block device support at present, and no multi-core either, by the looks.

A G4 with Altivec was almost on par on a PIV with SSE2 for multimedia.

Multimedia is important for some uses, not for others.

And what speed G4, against what speed Pentium 4?

G4 started from 450 MHz and went to 1.6 GHz, with several different µarch along the way.

Conversely, the Pentium 4 ranged from 1.3 GHz to 3.8 GHz.

G4 and P3 at least covered a nearly identical MHz range, and with similar MHz in the market at the same times.

Pentium 4 had an exceptionally long pipeline to be able to hit higher frequencies, trading-off high IPC numbers in the process, most of the time. On top of that, Pentium 4 also “looped” the same instruction while waiting data resulting in lower efficiency and higher heat output.

So, I assume, they are “step by step” comparable. i.e. slowest P4 equal to slowest G3 and they are again equal at next stepping and so on.

Core Duo's are almost duct-taped PIIIs.

Isn't Core Duo is an evolution of Pentium-M which is an evolution of Pentium 3, omitting Pentium 4 in the process like an unwanted child?

Feels like they rolled back a couple of commits, forked to a new branch and started over...

Which given hardware description languagesm that might indeed have been in the case. :)

The a version of the Pentium 4 was slower than the Pentium IIi per clock (launch model) and used goofy rambus memory

Their prediction about rambus hasn't aged well, but the technical advantages vs. SDRAM of the time are still interesting:

https://www.hardwarecentral.com/sdram-vs-rdram-facts-and-fan...

Surprisingly enough the N64 used RDRAM (and supported a memory upgrade, lowering the cost of the console at launch but enabling an upgrade path for games that needed more memory.)

It actually was between PIII and IV on capabilities. Altivec allows to play h264 videos on some settings. A Pentium III, well, SSE and FFMPEG can do magic, but not as well as Altivec. SSE2 it's better, much better. Basically the minimum for x86 and 'modern' browsers.

On the architecture itself, PowerPC did far more per cycle.

I recall folks claiming the G5 outperformed Pentium D, but I had both and not only was the Pentium snappier, "Roxio Toast" could burn a DVD in half the amount of time on it.

Maybe it was a software optimization issue and Toast just didn't use Altivec, but I was not surprised at all when Apple moved to Intel.

it was mostly marketing "oooh altivec can do sooo many FP calculations that it has to be licensed as a deadly weapon". Same kinda marketing for MMX "if a programmer was an a MMX guru then basically you have a supercomputer bro"

That's all great but it was all awaiting a killer app to make it all worth it and reasonable, versus random programs rarely supporting it for a little boost.

And the name of that killer app, using these DSP type features? Local AI.

In case of Altivec, having it or not was THE difference betweeen watching a video in MPlayer or a slideshow.

"No network or block device support at present, and no multi-core either, by the looks."

The first versions of Linux I used did not have multi-core support and I can imagine a Linux without networking, but no block devices?

Does that mean just character devices? How does the FAT driver work then?

It supports a read-only boot drive, but other than that just tempfs -- just like running a LiveCD version of Ubuntu etc.

Implementing those things is on the TODO list.

The earliest versions of Linux were pretty much just a terminal emulator for the PC console. Gotta start somewhere!

It looks like TILCK uses RAM images to provide FAT support, which is mainly used for initrd.

Every couple of months a new "Operating System" kernel which does not provide hardware abstraction and only runs in a VM pops up.

Tilck is not like that!

Tilck is a real operating system. It runs on real hardware.

The space for the former is overly crowded but I think Tilck really fills a gap that has been mostly unoccupied before.

Very nice. Its great to see how fast it boots, and it can run doom (framebuffer): https://www.youtube.com/watch?v=Ce1pMlZO_mI (also nice to see the dev takes the time to reply to an aspiring CS student on what it takes to grow in this field - comments in youtube)

Nice and short video. Demonstrates Vim, which seems like a sizable piece of software to get compiled with a subset of all Linux syscalls.

I can see it's marked "educational", but would it be also suitable for smallish embedded devices ? Assuming the bootloader part is swapped out, etc.

A kernel is mostly "hardware drivers" with an ABI.

That's why performant "standard" hardware programing interfaces are so much important. Those interfaces should be the simple and stable in time as much as possible.

Many hardware realms have such _performant_ interface: nvme, usb, etc.

Basically it means DMA, doorbells/IRQ "ring buffers", command ring buffers("queues").

Because for those alternative kernel initiatives, this would allow them to become 'real-life' much faster.

And a NOTE related to RISC-V hardware: keep an eye on arm piggybacking RISC-V chips(they are the bad guys with their blocking/$$$ intellectual property), but the target goal does include AV1 video dec/enc blocks instead of mpeg, and DisplayPort instead of hdmi, because those are hardly less worse than arm.

And some hardware is going even further, the "next step": user level hardware queues (look at AMD GPUs which started to implement those). I know there is 3D pipeline programming behind this hardware interface, but I believ that if they started to clean the base hardware interface, they will cleanup their 3D pipeline programming too.

The README was surprisingly lengthy and interesting. Worth a read for osdev people!

A couple of other kernels compatible with the Linux userspace ABI: Fiwix Asterinas

For us, the "elderly", who grew up with Minix in the educational space, how useful is Tilck in this regard?

Tilck is a small and deterministic monokernel that currently implements around 100 Linux syscalls in a Linux-compatible way. It’s a good educational tool, but is planned to be a Linux-compatible RTOS kernel. It currently supports statically linked binaries built against musl and boots and runs in about 3 megs of RAM.

One of the goals is to be small enough and simple enough of a codebase to be useful for educational purposes. Another is to eventually be useful in production in embedded systems.

Previous discussions:

3 years ago, 75 comments. https://news.ycombinator.com/item?id=34295165 (no riscv64 then)

5 years ago, 7 comments. https://news.ycombinator.com/item?id=28040210

It is interesting, although the lack of multiuser support is unfortunate. I really hope that the author reconsiders his stance on multiuser support, even if it is merely to allow things like chmod/chgrp to set file ownership to other users and groups. That would open the door to hosting a NFS server.

Its file system compatibility story is a bigger hurdle there.

To be honest, you’re much better off with a battle hardened platform for your use case. Tilck is meant to be educational, not secure in either the InfoSec nor data robustness sense.

Upon seeing this, thoughts keep occurring to me such as:

  * “It could be fun to be the author of a block layer, VFS and an AHCI driver for SATA. NCQ support is a must if I do.”
  * “It could be fun to port ZFS: ztest and zdb, plus the userland version of the driver into which they hook need pthreads, but the important userland utilities do not. Having to statically link everything would be a pain. I would probably have to reimplement the entire SPL for this to work and that would probably be at least 10,000 lines of code.”
  * “If I port ZFS, a NFS server needs to follow so ZFS has an application. This will need support for setting/getting user/group ownership and mode bits. If I rewrite the VFS, I could maybe sneak that feature into it for a NFS server to use while preserving the documented syscall behavior that requires everything be root. If I port the NFS server from illumos, it could share the SPL code with ZFS. NFSv4 permissions will be needed to make it fully happy. Beyond that, I will need a network stack.”
  * “It could be fun to port a network stack. Maybe LwIP could be used.”
  * “It could be fun to write an e1000 driver.”

I have already found the documentation I need if I actually were to implement AHCI and e1000 drivers:

https://www.intel.com/content/dam/www/public/us/en/documents...

https://www.intel.com/content/dam/doc/manual/pci-pci-x-famil...

If I were to do all of this, I would likely try my best to make it a production platform. The purpose would be fun.

Anyway, I wonder if these thoughts will continue if I sleep on them.

Go for it. All that sounds like fun.

If you're serious and the maintainer is willing to accept these contributions, then go for it.

It's a lot of work but definitely rewarding stuff. And you'll learn a lot along the way too.

I am still considering it.

Won't be big and professional like GNU

Different era then.

Your options were basically just:

- GNU: which didn't have a kernel

- BSD: which had massive question marks over ownership

- Minix: which wasn't free back then

- or a commercial Unix

These days you have a dozen BSDs, Darwin, several OpenSolaris forks, hundreds of Linux distros as well as non-commercial licenses for many commercial Unixes. Minix is free. Free reimplementations of BeOS too. And that's before you count the plethora of free non-POSIX systems like Plan 9, ReactOS, and so on and so forth.

And, ironically, there's now less of a need for competition in this space because web engines have replaced OS kernels for a lot of common use cases and WASM is fast becoming the new ABI.

To be clear, I'm not saying this project has no merit. Nor am I saying there is zero chance this project might evolve into something big. I'm just saying that quoting Linus like you have is extremely simplistic.

Your comment goes on and on about different variants but fails to see the germ of a good idea. Sorry to say.

Oh I see the good idea. I also read their readme, which is how i know it's not yet ready to work as a file server ;)

>Oh I see the good idea. I also read their readme, which is how i know it's not yet ready to work as a file server ;)

this is called "judging a book by its cover"

> this is called "judging a book by its cover"

So you think they're lying about not having implemented block devices nor a network stack...? Or are you suggesting that NFS doesn't need network nor file system support?

Please calm down.

Could probably just record user:group in the filesystem and ignore it at runtime except when serving clients? Ex. a Linux file server running as root can still check permissions and ownership without having to change its own user. (Although as an interesting thought experiment, forking per session and switching to the client user would give you file permissions for "free" by way of the kernel imposing them on you...)

As long as the VFS permits those calls to go to the filesystem driver, that is doable. A casual inspction of the documentation suggests that doing this will return an error from the syscall. I am not sure what a kernel driver calling the VFS can do in terms of getting/setting these yet.

This is really cool. It reminds me of the running linux on a 3.5" floppy as a NAT firewall.

Doesn't look like you can run WINE on it, which would make for an interesting Windows replacement.

Why would you just not use Linux to run WINE. It's not realistic to replace all the drivers that make Linux so universally usable as it is. If you specifically want it as a Windows replacement, the Linux distro that Valve is building seems like the best bet.

The readme explicitly states it’s not suitable for desktop usage. For staters, it doesn’t support all the features needed to run x11.

This looks like a very interesting project for OS devs but if you’re after a Windows replacement then there’s already a few options out there from ReactOS to Linux-distros that come with Wine configured as part of the base install.

Details on what is missing that a X11 server needs are here:

https://github.com/vvaltchev/tilck/discussions/81

To be honest, that reads less like details on what's missing to port x11, and more like a frustrated exchange between a project author who's clearly stating what the project cannot do, and an overly enthusiastic member of the community who's vastly underestimated the amount of work required to run x11.

Credit to the authors patience throughout that exchange, though. I've had similar exchanges with some of my open source projects too. It's great when people do want to try things but it can cost the author a lot of time and effort helping people on endeavors that you already know will fail and already know it isn't where you're time should be spent. And can be particularly frustrating when you know those participants are unlikely to contribute any code that would further the project's goals. But you still have to show them courtesy because the last thing you want to do is kill their enthusiasm for open source.

Once you're running something as heavy as a desktop application or game, then the full Linux kernel is not the heavy part anymore.

This is fantastic. The problem with Linux is the dev culture at the top. I've long wished for an alternative kernel project that might serve as a better dev model. This has the potential.

The kernel dev culture is excellent, it's pure meritocracy.

We all know that Rust in the kernel is a political operation to change this to woke.

I'm not even sure what that's supposed to mean. Is it a joke that I'm not getting?

> I've long wished for an alternative kernel project that might serve as a better dev model.

Do you have something against FreeBSD, NetBSD, OpenBSD, illumos, Redox, or (if you don't need a unix-like) Haiku?

Linux syscall compatibility.

Doesn't BSD have that?

Some of them do.

FreeBSD, NetBSD, and illumos have that.

Linux won BECAUSE of whatever dev culture at the top you mention, not DESPITE. There are plenty other more permissive Unix-like OS, and none are more relevant than Linux.

A lot of the top maintainers have probably been in that role for too long. 30-40% fresh blood (read: replacements) within a few months would probably do wonders.

Reply on news.ycombinator.com