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.
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.
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.
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.
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)
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.
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:
- 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.
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.
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 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.
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.
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.
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.
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.
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.
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.
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?
Implementing those things is on the TODO list.
It looks like TILCK uses RAM images to provide FAT support, which is mainly used for initrd.
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.
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.
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.
Feels like they rolled back a couple of commits, forked to a new branch and started over...
On the architecture itself, PowerPC did far more per cycle.
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.
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.
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
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.
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.
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.
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.
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.
https://github.com/vvaltchev/tilck/discussions/81
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.
Do you have something against FreeBSD, NetBSD, OpenBSD, illumos, Redox, or (if you don't need a unix-like) Haiku?
We all know that Rust in the kernel is a political operation to change this to woke.