Linux: microkernels are crap, it's clearly impossible to do anything useful with message passing.

Also Linux: if you want to do anything with network device configurations, here's the AF_NETLINK socket type and a bunch of messages you can pass to/from the kernel to get everything done.


I know there's nuance that I'm ignoring here, but it amuses me.

· · Web · 1 · 0 · 1

@jens Theres obviously a need to communicate between userspace and the kernel somehow, but now imagine the network subsystem doing the exact same thing when working with the memory management or filesystem subsystems

@espen @jens I can, and I have used systems that worked exactly in that mode before (e.g., QNX), and the systems were very responsive. This, on 386 and 486-class hardware.

Maybe the problem isn't microkernels, but the in-built assumptions kernel devs make about what kind of environment their code is running in.

There's no question microkernels will always take longer to satisfy a service request than a monolithic kernel; so, instead of building APIs that are composed of lots and lots of tiny functions, you create requests that are basically aggregate common functionality.

Prior art: X11 works this way (where xlib queues "lots of little functions" into work units that are big enough that a network request's overhead is well amortized), 9P works this way (where Plan 9 queues seek and read/write interfaces into a single 9P transaction to minimize overhead), the L4 programming interface frequently works this way (where just about every message exchange is always a send-then-receive or receive-then-send, and is one reason why it outperforms Mach), etc.

And, speaking of 9P, although Plan 9 is not technically a microkernel, it is not exactly a monolithic kernel either. It's somewhere in between these two extremes, and any typical Plan 9 environment will make extensive use of microkernel-like functionality, since the vast majority of its system services are actually provided through user-space daemons and applications. ACME, the GUI itself, Plumber, etc. are all user-space, not kernel.

A lot of R&D has gone into making microkernel environments responsive. If they can make hard real-time devices using them, frequently in medical devices that keep people alive, I feel the issue of latency when invoking other system services is a problem which is now tolerable, if not solved.

@vertigo @espen I had that opinion for a while, but frankly never had that much experience with these systems, other than as an end user.

@jens @espen Contrary to what my post might suggest, I'm not actually a microkernel apologist. My preferred architecture is exactly what Plan 9 uses; a hybrid design where you can get the best of both worlds, leaving it up to a system integrator to decide what resides in the kernel and what doesn't. Putting something in the kernel should be a performance optimization, not a hard requirement for basic operation.

Linux modules are a great compromise here, but I think their potential was never truly realized. I agree with Rob Pike that the basic philosophy behind Unix I/O primitives are under-utilized. Yeah, we now have /proc and /sys, but ... that's it?

I think it's obscene that Linux has over a thousand system calls when I've used, maybe, a total of 40 system calls in all my code since 1995, when I switched to using Linux as my desktop for the first time.

Old man shouting at clouds again, I suppose.

@vertigo @espen Well, I'm another old man like that.

And though I'm actually doing only userspace things for a bunch of reasons, a kind of planet scale 9P-ish thing is sort of a goal.

My focus is currently on network and security topics. @alcinnz is, in a roundabout way, closer to the file-like protocol. It'll have to be a combination of both to become useful.

TL;DR, a better file-like API would probably make a lot of application code simpler and safer.

@jens @espen I stand corrected; Linux has, depending upon configuration and processor architecture, between 300 and 450 system calls. I must have been recalling figures for a different kernel.

Still, that's only a factor of two to three, not an entire order of magnitude, difference. I'd argue that my point still stands.

@vertigo @jens @espen take the unikernel pill: everything should be in the kernel

@lunch @jens @espen +1 if the hypervisor offered the right services. I grew up on AmigaOS; if I can use contemporary terminology, arguably the first open-ended uni-microkernel environment. 😏 There are still some things about it that I miss dearly.

Sign in to participate in the conversation
Finkhäuser Social

A private instance for the Finkhäuser family.