Core work still going on 33 years later [LWN.net]

Core work still going on 33 years later

Posted Oct 20, 2024 0:55 UTC (Sun) by willy (subscriber, #9762)
In reply to: Core work still going on 33 years later by atnot
Parent article: The long road to lazy preemption

https://doc.cat-v.org/bell_labs/utah2000/utah2000.html

You're not necessarily wrong, but I certainly didn't start from "I want a kernel that manages memory in large chunks, which one shall I work on?" Other projects may have. I started from "Here are some problems I see in Linux. How could we solve them?"

I did consciously ask "Which filesystem shall I start with?" and chose XFS for a number of reasons (mostly working with the people and perceiving iomap as being the future of the VFS).

to post comments

Core work still going on 33 years later

Posted Oct 20, 2024 22:39 UTC (Sun) by Paf (subscriber, #91811) [Link] (24 responses)

About Rob Pike's essay, which I've seen before.

Frankly, this view is a little harsh, but Pike has always seemed to me bitter that he came along a little too late and too uninterested in the idea that a problem could be, well, fairly well solved by existing systems. I think Plan 9 wound up irrelevant not because industry is stupid or hidebound, but because it didn't represent enough of an improvement (where it was an improvement at all).

Some problems do, at a certain point, end up, for some sense of the word solved, mostly solved. And the kind of innovation that previously defined a system/product/whatever trails off and moves elsewhere.

At a certain point, we stopped changing the basic design of how cars work with their driver and the road. A car from much after 1950 (or, really, even the 1930s for many designs) has, to a first approximation, the same human interface and basic functionality as one from 2024, leaving aside any partial self driving features (which are brand new in any case). There has been a lot of innovation, but not in the controls or basic shape. This was a problem that was solved well enough that new solutions couldn't generate space. And yeah, they might be superior, but probably not all that superior. It's a lot more than just "cost of switching" and "everyone gave up". There was, in fact, little need to make larger changes. I know that's depressing if you're a researcher, and, yes, some promising innovations end up moribund... But they always did.

Core work still going on 33 years later

Posted Oct 20, 2024 23:11 UTC (Sun) by willy (subscriber, #9762) [Link] (20 responses)

Yeah, I'm not sure Pike is 100% right. Plan9 was a great piece of research. Had it been done earlier, it might have displaced Unix. But I do think that we're past the point where we can just throw away Unix compatibility. Linux shows the road forward; we can add new interfaces to make things more efficient, but we can't abandon Unix.

The car-human interface isn't as fixed as one might think. Is the gear shift in the centre console or on a stalk? Automatic or manual? Is there an ignition key or a "Start" button somewhere? How does one open the refuelling socket? Which @#$& side of the car is it on? Where are the windscreen wipers? How do you dip the headlights? None of these are big burdens if you own a car, but if you rent, you have to figure all these things out at some point (preferably before leaving the lot).

It's the same with showers. They mostly have controls for volume, temperature and which outlet(s) the water should come out of, but some hotels go out of their way to have super fancy ones that are utterly non-discoverable. I know how to work my shower at home, but if you stay at three different hotels in three nights, you're going to be confused and angry in the third shower.

Anyway, back to my point. Innovation has moved up the stack. We have something good enough, and we're all building on it. Linux has become the substrate on which we innovate. I don't think that's sad, I think that's progress.

Which @#$& side of the car is it [Fuel Socket] on.

Posted Oct 21, 2024 0:30 UTC (Mon) by gmatht (subscriber, #58961) [Link] (8 responses)

Well, there should be a little symbol on your dashboard, either <[FuelPump] or [FuelPump]>. The fuel should be on the side the arrow is pointing to, so you shouldn't have to stop your car to get out and check on the way to the fuel station. Not sure having a little arrow pointing to either Plan9 or Unix would help much in our case though :P

Which @#$& side of the car is it [Fuel Socket] on.

Posted Oct 21, 2024 1:14 UTC (Mon) by dskoll (subscriber, #1630) [Link]

LWN properly supports Unicode, so ◀⛽ or ⛽▶ 🙂

I didn't know about the fuel socket indicator until about 5 years ago.

Which @#$& side of the car is it [Fuel Socket] on.

Posted Oct 21, 2024 2:35 UTC (Mon) by sfeam (subscriber, #2841) [Link] (5 responses)

Yeah, great. Now how about the problem of finding where they put the little latch-release thingy that you have to pull before opening the cover to the fuel input. Under the seat? left or right? How far back? Or maybe it's under the dash. Or on the driver-side door. Or maybe there isn't one but the cover is stuck and needs to be pried. Or maybe it's a button on the key fob.

Which @#$& side of the car is it [Fuel Socket] on.

Posted Oct 21, 2024 13:43 UTC (Mon) by LtWorf (subscriber, #124958) [Link] (4 responses)

I've opened the engine compartment instead of the fuel lid a certain number of times.

Which @#$& side of the car is it [Fuel Socket] on.

Posted Oct 21, 2024 14:03 UTC (Mon) by Wol (subscriber, #4433) [Link] (3 responses)

???

Ime, the engine cover lever is always in the passenger footwell by the door. While the fuel cover switch in my car is in the driver's door. And while I have little experience of said switches, I've never known them on the passenger side ...

My bug bear is "flash headlamps" and "wash windscreen". The number of times I've flashed people by mistake ...

Cheers,
Wol

Which @#$& side of the car is it [Fuel Socket] on.

Posted Oct 21, 2024 20:24 UTC (Mon) by mathstuf (subscriber, #69389) [Link] (1 responses)

The engine cover lever has always been on the driver's side IME (US). AFAIK, it has always been a physical by-cable latch; the fuel door meanwhile is now a button and electronic in newer cars. No idea what a Tesla does; probably electronic though given that everything else is.

Which @#$& side of the car is it [Fuel Socket] on.

Posted Oct 22, 2024 8:13 UTC (Tue) by anselm (subscriber, #2796) [Link]

The engine-hood lever is usually on the left side of a car because most cars have the driver sitting on the left, and given how rarely that lever is used there is no point in moving it to the right for the others because all that will achieve is to make building the cars more complicated. So in the UK it is on the passenger side, and in places like the USA and Germany it is on the driver's side.

We can probably count ourselves lucky, though, that in right-hand-side-driver cars the pedals aren't in reverse order.

Which @#$& side of the car is it [Fuel Socket] on.

Posted Oct 22, 2024 8:55 UTC (Tue) by farnz (subscriber, #17727) [Link]

IME, engine cover latch is on the left of the car in the footwell, and some cars have the fuel cover release on the frame next to the driver's seat. In UK versions, that puts the fuel cover release on the right of the car, but in the French market version of the same car, the engine cover release and the fuel cover release are on the same side of the car - engine cover by your left foot, fuel cover under your left shoulder when the door is open.

Which @#$& side of the car is it [Fuel Socket] on.

Posted Oct 21, 2024 3:44 UTC (Mon) by interalia (subscriber, #26615) [Link]

Plenty of older cars don't have such an arrow in my experience, it's definitely common to have one but plenty don't have it. It was a good innovation, the kind that's obvious in hindsight.

Core work still going on 33 years later

Posted Oct 21, 2024 2:12 UTC (Mon) by Paf (subscriber, #91811) [Link]

"Anyway, back to my point. Innovation has moved up the stack. We have something good enough, and we're all building on it. Linux has become the substrate on which we innovate. I don't think that's sad, I think that's progress."

You know, perhaps this is the biggest point of all: It is good enough, and very challenging to change.

Why should hardware be involved in security?

Posted Oct 21, 2024 4:38 UTC (Mon) by ebiederm (subscriber, #35028) [Link] (1 responses)

The basic paradigm of premptive multitasking operating systems with protection provided by the hardware seem like a local maximum.

The most questionable decision in all of that seems to be relying on hardware to define the isolation of untrusted software.

Hardware is always buggy and so expensive to fix it might as well be unfixable. We are effectively 7 years into the era of Spectre and I am not aware of any high performance cpus that successfully isolate untrusted software.

So why do our operating system architectures by design rely on broken and unfixable hardware to get security right?

Which is to say when operating systems are failing at part of their core mission because of how they are designed I think there is room and need for innovation at that level.

Why should hardware be involved in security?

Posted Dec 13, 2024 13:21 UTC (Fri) by roblucid (guest, #48964) [Link]

The logic was if the hardware is not functioning correctly, how can you rely on software running on the hardware to do so?

A lot of security is about process isolation and correct virtual memory implementations, you simply cannot do something like logical->physical address translations efficiently in software, it needs to be initiated by L1 cache look up (hence the cache's tags to eliminate false positive hits) and available for L2/L3/DRAM fetches.

Then again software being mutable is what hostiles rely on, you need the OS & hardware support to harden a system against expoitation, a program that's reentrant, relocatable or dynamically linkable simply cannot know what logical addresses it uses. Even so without hardware support where would the immutable correct address tables be stored, so errors cannot be exploited to patch the program?

Core work still going on 33 years later

Posted Oct 21, 2024 17:00 UTC (Mon) by paulj (subscriber, #341) [Link] (7 responses)

It probably doesn't have a bearing on the computer side of the discussion, but the major differences in car UI over the decades has been the simplification of controls. In particular, there were some systems that had the human in the control loop, which were eventually automated (by mechanical means first, later electronic). Or there were controls to systems which required a number of co-ordinated steps, which are no longer necessary thanks to mechanical refinements.

E.g., if you got in a car from the 20s to early 30s, most people today would be unable to start it, for want of knowledge of 2 key engine controls (one of which was automated fairly early on - who here knows what the lever on the centre of the steering wheel [typically] did?; another was present until the 80s on many cars, a knob on the dash you had to pull in and out typically - I'll add a comment later with the answers ;) ). If they were able to start it, they might well damage the engine. They would also struggle to change gear without damaging the car.

UI has gotten simpler, and details hidden. A driver from the 20s would probably find it easier to get comfortable driving a modern car, than a modern driver getting into a car from 100 years ago - bit more to learn. The amazing speed of modern vehicles might be the 1 control thing the 20s driver might need to adapt to, but that wouldn't stop them driving at a slower speed.

Also, the maintenance of the car is now minimal compared to the earlier days.

Core work still going on 33 years later

Posted Oct 22, 2024 13:44 UTC (Tue) by paulj (subscriber, #341) [Link] (6 responses)

And the systems/steps I had in mind are:

- Ignition advance:

This used to be something that had to be manually adjusted as you drove, to suit the engine speed, warmth and mixture.

- Choke:

Manual adjustment of mixture (interacting with previous), particularly for engine start.

- double-declutching for gear changes:

Changing gear required pausing the gear change in neutral, letting the clutch engage again, and matching the engine speed to the drive-shaft (either
by letting the engine speed fall a little, if changing up; or blipping the throttle, if changing down - often while continuing to hold the brake pedal),
before disengaging the clutch again and completing the gear change.

(Apparently truck drivers in the USA still have to do this on many models of tractor units).

Core work still going on 33 years later

Posted Oct 22, 2024 16:17 UTC (Tue) by farnz (subscriber, #17727) [Link] (5 responses)

You missed a couple more important details; the modern 2 or 3 pedal layout was not yet the standard in the 1920s, and some cars even then had the throttle as a lever on the steering wheel, rather than a pedal. And the pedals might well be gear selection, with possibly a foot brake, possibly not. The clutch could be a pedal, but it might also be a hand-operated lever, and even if it's a pedal, it might need lifting with your foot instead of pressing.

It's not until the 1940s that the industry finally settles on the modern control scheme.

Core work still going on 33 years later

Posted Oct 22, 2024 16:21 UTC (Tue) by paulj (subscriber, #341) [Link] (4 responses)

That could well be. I don't know. It was my dad who was the old car nut - not me.

The 1927 Austin 7 he once had, which I've driven, already had the familiar 3 pedal layout. The clutch was more like a button though. Very hard to get used to. So that layout already existed in the 20s.

Core work still going on 33 years later

Posted Oct 22, 2024 16:26 UTC (Tue) by farnz (subscriber, #17727) [Link] (3 responses)

The modern layout existed, but (e.g.) Fords from the 1920s had a mix of layouts - indeed, a Model T and a Model A had different control schemes, and some of the things I mentioned that now seem odd were used by different 1920s Ford models (foot pedals for gear selection, lifting the clutch not pressing it).

If you're used to that sort of array of different possibilities, where you need to read the fine manual before trying to drive because there's so many options, learning how to drive a modern car isn't that hard; just work out how the modern control map to what you expect, and complain because the car does timing advance, choking etc for you. But (as evidenced by people who can drive an automatic transmission, but can't drive a manual transmission) going the other way is harder - you have to do more things that a modern car does for you.

And I've not driven anything without a modern control layout - I've only seen them in museums with my grandfather, who wanted to show me the cars he dreamt about being able to own when he was a child.

Core work still going on 33 years later

Posted Oct 22, 2024 16:42 UTC (Tue) by paulj (subscriber, #341) [Link] (2 responses)

I'm not familiar with old US cars. My dad has had a variety of old European cars. Mostly English, but he's had Citroen Traction-Avant too (which is one of the coolest looking cars ever IMO - and surprisingly modern in design and build). He had some kind of 1920s Rolls-Royce at one point, a Phantom I think - if only he'd kept that one, would be worth loads today (he loved swapping cars)! I don't remember the controls of that, I was little - though I don't remember it being weird per se. I could ask him. ;)

Core work still going on 33 years later

Posted Oct 22, 2024 16:50 UTC (Tue) by farnz (subscriber, #17727) [Link] (1 responses)

These were all models that my English grandfather had heard about as a child and really wanted at the time, but could never afford - he was a serious car nut.

But that does lead to a serious point; it isn't unusual for different countries to have kept different standards from the past, even though they "could" unify with the rest of the world. For example, on pedal cycles, some countries put the front brake on the left lever, while others put it on the right lever. Arguably, the only reason this didn't happen with the motor car is that we had a large crowd of ex-military drivers in the late 1940s who all knew the same standard no matter where in the world they were going back to, and so everyone settled on one standard.

Core work still going on 33 years later

Posted Oct 22, 2024 18:17 UTC (Tue) by joib (subscriber, #8541) [Link]

> it isn't unusual for different countries to have kept different standards from the past, even though they "could" unify with the rest of the world.

Case in point, the International System of Units (SI) is adopted by almost the entire world, except Myanmar, Liberia, and some other country whose name escapes me at the moment.

Core work still going on 33 years later

Posted Oct 21, 2024 7:46 UTC (Mon) by roc (subscriber, #30627) [Link] (2 responses)

A few things happened.

Linux grabbed the "free-software OS for commodity PCs" ecosystem niche. Perhaps it could have been MINIX with a more community-oriented owner and a better license, but it was Linux. Reasonably well-run open-source projects in important niches accrue powerful network effects.

Over the same time period, what people expect from the OS --- userspace APIs and hardware support --- grew massively, making it much harder to build a viable competitor.

And yes, for a long time the Linux kernel design was good enough ... good enough that the cost of replacing it (including the cost of migrating higher-level software to a new design) has never been justified.

But I think it would be wrong to conclude that the Linux kernel, or the general Unix-style kernel interface, is in any sense optimal. Linux has a lot of serious problems that are becoming more serious over time. The monolithic design has led us to a point where the kernel is too big to trust and developers are overwhelmed with CVEs. Relying on namespaces and seccomp for isolation makes sandboxing brittle and very complicated; I wish the system was much more capability-oriented. ptrace and signals are notoriously problematic.

Core work still going on 33 years later

Posted Oct 22, 2024 0:27 UTC (Tue) by Cyberax (✭ supporter ✭, #52523) [Link]

On the third hand, Linux _is_ evolving. We now have resource partitioning via cgroups, we're evolving a new process management API that lacks braindeadness of the classic POSIX process management API. And we have io_uring for the async stuff.

I can see a world 15 years from now, where all the classic synchronous Linux system calls are reimplemented as user-space compat shims on top of a minimalistic native io_uring kernel.

Linux kernel and interface

Posted Oct 22, 2024 2:40 UTC (Tue) by interalia (subscriber, #26615) [Link]

All good points re the network effects etc, though I'm not sure anyone was saying the Linux kernel or its interface is optimal. I think we're all imprisoned by the tyranny of existing software and the effort required to port it.