Hacker News Reader

Comment by blippage | Hacker News Reader

Linux user here, not NetBSD. My exposure to NetBSD is infinitesimal. But ...

I saw a video on "rump" kernels awhile back. I also heard that Andrew Tanenbaum ported/is porting Minix to NetBSD. NetBSD has an "anykernel" approach which separates the kernel for the drivers. So you can impose your own flavour of kernel on a bunch of pre-written drivers.

This struck me as a potentially mind-blowing win for NetBSD. Why isn't every RTOS, embedded system OS, toy OS,experimental OS etc. based on NetBSD? I don't know the reason for this. Maybe it's part NIH, part open secret that nobody's heard about, and a large part to do with getting over the hump of integration.

View on HN

Yeah, rump kernel tech seems like it should be taking the world by storm - in particular, it looks to me like NetBSD should be a "universal donor" of drivers to other systems. So for instance, as I understand it, with minimal shims, you should be able to use NetBSD hardware drivers on Linux and Illumos and HURD and Redox and [your hobby project here]. It'd be worse than "native" drivers, but especially for smaller projects it'd save a ton of work.

> I also heard that Andrew Tanenbaum ported/is porting Minix to NetBSD

Kernel, or userspace? Last I paid attention, MINIX 3 used pkgsrc for packages and I think base/coreutils in userspace, but I hadn't heard anything about kernel space. Though, as I mentioned above, it does seem like the obvious way to get drivers for a lot of hardware cheaply, especially if you're already a microkernel.

Not NetBSD, but iirc a lot of the network drivers in Haiku were taken from FreeBSD.

> with minimal shims, you should be able to use NetBSD hardware drivers on [...] HURD

The Hurd is already using rumpkernel for some servers, e.g. for audio and work is underway for disks.

Public release of RUMP predated Docker by at least two years, maybe it was more. I just remember when all the Docker hype began I was thinking "WTF, this is not even half as good." But that's what's different about NetBSD. No hype. To answer the question of why it isn't more widely used, IMO, besides absence of hype, it's lack of hardware support, and comfort among those who use it with notions like DIY and minimalism. Often you have to put in some effort to get some hardware feature to work.1 With Linux, chances are someone else has done that work for you.

1 First time I ever tried NetBSD, the audio did not work on the laptop I was using. I added a couple of lines in a kernel header file, recompiled and voila, I had audio. That simplicity made me want to keep using it.

> I added a couple of lines in a kernel header file, recompiled and voila, I had audio

How did you figure out what to add?

IIRC, dmesg. It was probably pure luck that it worked.

Close to my change to the BTTV drivers for a Conceptronic TV card.

Most of hardware use common devices and chipsets.

> NetBSD has an "anykernel" approach which separates the kernel for the drivers. So you can impose your own flavour of kernel on a bunch of pre-written drivers.

As a NetBSD developer, I don't recognize what you mean by this.

Maybe read this whitepaper:

https://www.netbsd.org/gallery/presentations/justin/2015_Asi...

“ The rump kernel has been used as a way to supply device drivers in other new operating systems, which do not yet have a full set of device drivers. For example, Genode[4] is a framework for building microkernel operating sys- tems using the L4 family of microkernels. Genode uses the rump kernel to provide file system support, so that it does not have to develop its own file systems.”

Here’s a 362 page dissertation on rump kernels which has a netbsd focus:

http://lib.tkk.fi/Diss/2012/isbn9789526049175/isbn9789526049...

There’s also the Wikipedia entry.

https://en.m.wikipedia.org/wiki/Rump_kernel

That isn't how device drivers are used in NetBSD though, it is a different way of compiling them for other targets.

From the white paper published on netbsd.org:

“ An implementation that does this ships with NetBSD, allowing the NetBSD drivers to be run in NetBSD userspace… The first major PCI driver developed was the In- tel Centrino 7260 driver developed for NetBSD and OpenBSD by Antti Kantee. The commit message said “This is probably the world’s first Canadian cross device driver: it was created for OpenBSD by writing and port- ing a NetBSD driver which was developed in a rump ker- nel in Linux userspace.”[13]

Further, just another example, from page 149 in the dissertation is an extended many page discussion of how the rump kernel was used to provide usb support in netbsd.

http://lib.tkk.fi/Diss/2012/isbn9789526049175/isbn9789526049...

“ We implemented a host controller called ugenhc. When the kernel’s device autocon- figuration subsystem calls the ugenhc driver to probe the device, the ugenhc driver tries to open /dev/ugen on the host. If the open is successful, the host kernel has attached a device to the respective ugen instance and ugenhc can return a successful match. Next, the ugenhc driver is attached in the rump kernel, along with a USB bus and a USB root hub. The root hub driver explores the bus to see which devices are connected to it, causing the probes to be delivered first to ugenhc and through /dev/ugen to the host kernel and finally to the actual hardware. Figure 3.24 con- tains a “dmesg” of a server with four ugenhc devices configured and one USB mass media attached.”

¯\_(ツ)_/¯

> ¯\_(ツ)_/¯

UTSL.

The author used rump to develop the drivers, it isn't used to run them in a production NetBSD kernel.

But running on other targets is exactly what people are excited about?

QNX RTOS uses netbsd userspace applications, actually. And a lot of the drivers in QNX are from NetBSD. It’s more properly a “microkernel” which means the drivers run as userspace applications with a defined interface to the kernel for memory maps to underlying hardware.

Are new versions of QNX available on X86?

yes. (newest one might just be x86_64, not 100% sure on that)

That's cool, I can emulate 32 if necessary :)

How might one obtain it?

I'm guessing, call blackberry and prepare to break your piggybank?

Annnnd please let us know how many limbs you had to sell. I'm actually interested in ballpark numbers for this, as I have very fond memories of QNX as the OS where my 100us timers did fire even under heavy IO, threading churn, etc. I miss that.

Go there:

http://www.qnx.com/download/group.html?programid=29178

Download:

QNX Software Center

http://www.qnx.com/download/feature.html?programid=51624

> newest one might just be x86_64

This would be weird. I believe 32-bit x86 still is widely used by the target audience.

Some subset, but even there it's waning, especially for new designs. QNX is seems to be not mentioning it explicitly anymore, and I don't remember seeing it in the options last time I looked, but it's very possible that it is still available if necessary.

> This struck me as a potentially mind-blowing win for NetBSD. Why isn't every RTOS, embedded system OS, toy OS,experimental OS etc. based on NetBSD?

I think it is used a lot as a starting point for porting other OS, but it doesn't necessarily persist.

Supposedly Apple started with NetBSD when porting OS X to x86, not that x86 is exactly esoteric, but NetBSD lends itself to being a starting point. From what I read nothing remains of NetBSD in OSX, the current BSD userland is from FreeBSD but is pretty old still.

Apple Base Station and related firmwares (e.g. Extreme) ran NetBSD.

BSD-based products are more common than believed, but because of the liberal licensing they seem to mostly fly under the radar, not unlike various proprietary options like VxWorks and TRON.

Just to be clear vxworks is not bsd based that I could tell from Wikipedia. I work with Schneider electric plcs and sometimes vxworks errors bubble up.

I think they meant to say that BSD-based stuff is common but unknown to most, and that the same goes for VxWorks and TRON, in that both open source BSD and proprietary solutions have licenses that make their inclusion in other software products not noticed by many people.

Same for Minix3... until someone dug it out of IntelME no one knew it was one of the most ubiquitous deployed OS.

That is super interesting. I will say the package manager (pkgsrc) seems to be very portable as well. It’s default on netbsd, smartos/illumos, and minix, but also available for macOS, Linux, and other operating systems. I use it routinely on smartos and it’s a breeze.

IMO: NetBSD might be superior but the lack of GPL means less work on it is published. The net effect is that everyone puts up with Linux because it actually runs on everything.

>Linux because it actually runs on everything.

Good joke. Linux barely runs on some legacy arches, having to use really old kernels and making a lot of current software useless.

NetBSD just /runs/ new releases on older hardware such as m68k machines with no crazy tweaks.

>but the lack of GPL means less work on it is published.

You realize a lot of the components on the graphics substack on Linux are non-GPL, right? Not just X, but the backends on the kernel.

Also Linux is full of propietary drivers and blobs. Check Linux-Libre's changelog if you don't believe me.

There was a time when NetBSD's portability was impressive. A handful of open source operating systems supported a handful of architectures, then there was NetBSD (and OpenBSD). Then Linux started to look beyond the x86 and started gathering steam on other architectures. NetBSD couldn't really compete for developers and it began stagnating. It probably didn't help that a lot of the "Unix" userspace also became more tightly coupled to Linux.

Don't get me wrong: I was enormously enthusiastic about NetBSD in the mid-1990's and I still prefer the idea of a unified core OS. The thing is, it's a lot easier to deal with one operating system than multiple. If you're developing under Linux, it's going to be easier to develop code for an Arm based system under Linux because it defines your expectations.

Linux probably supports more platforms these days. But it often requires vendor-specific kernel forks, specific distributions, etc. NetBSD is impressive because it can be (cross-)compiled for many systems using the stock distribution without much effort (build.sh script).

Disclaimer: I haven't used or looked at NetBSD since 2006-2007 or so, but I think this is still accurate.

This. Linux under ARM is a nightmare, NetBSD for ARM it's just a single distribution per subarch.

Linux is just a Kernel NetBSD is a System.

Linux doesn't even properly work on legacy non-x86 devices. NetBSD for m68k runs much better than GNU/Linux.

EDIT: Downvote me if you want, but the reality is that hard.

You have to use specific distros and hard patches to get things running well. For example under PPC32 there's no support for common distros.

Don't let me start on Mips, Sparc or Alpha, where NetBSD runs on it just fine.

Reply on news.ycombinator.com