| HN Mirror

1) Many will cry about that performance hit (including me).

For over a decade our computers have gotten faster marginally, but our software has gotten slower at a greater rate.

You can barely navigate the web now with a new low end computer (that isn't a Chromebook). Most on this site won't care though because our machines cost $2,000+ and the web is Fine(tm); many folk aren't buying anything over $300 though.

2) These are memory bugs, so the introduction of Rust into the kernel could help us here potentially, no need for an architectural revolution.

> Many will cry about that performance hit (including me) because for over a decade our computers have gotten faster marginally, but our software has gotten slower and bloatier at an increasingly rapid pace.

We're talking about network stacks and network drivers, not web browsers. Migrating the network stack from the kernel to a user-land process is not going to measurably slow down web browsers, especially on modern systems with gigabytes of RAM, multiple cores, IOMMUs and whatnots.

> These are memory bugs, so the introduction of Rust into the kernel could help us here potentially, no need for an architectural revolution.

That would require rewriting the network stack and network drivers in Rust (driver code is much more likely to have bugs than the rest of the kernel) for this to be effective, otherwise you'll still have a lot of C code in the network path. I'd argue that this would be a bigger architectural revolution than porting the existing code and running it in user-land. MINIX3 went through such a change when drivers were removed from the kernel (can't find the publication about it right now) and they only required reasonably small changes when porting these to user-land, there were not rewritten from scratch.

But this is not just about memory safety, Rust code can still be vulnerable in many other ways (memory leaks, unsafe blocks, wrong assumptions, incorrect algorithm implementation, buggy/compromised toolchains...). Code running inside the trusted computing base of a system is a liability, enforcing privilege separation and principle of least authority reduces it.

> We're talking about network stacks and network drivers, not web browsers.

Ah yes, the magic web-browser that doesn't do any kind of networking at all.

> Migrating the network stack from the kernel to a user-land process is not going to measurably slow down web browsers, especially on modern systems with IOMMUs and whatnots.

I don't know how you can possibly assert that, it's contradicting computer sciences' current understanding of operating system design as it relates to kernelmode/usermode switching, unless you're doing weird shared-memory things in userspace... which is terrifying.

> That would require rewriting the network stack and network drivers in Rust

Not really, C and Rust can interop just fine, you can have network drivers that are rust but the actual networking stack itself can remain C, if you want.

> but this is not just about memory safety, Rust code can still be vulnerable in many other ways

The post is literally memory safety bugs.

> Ah yes, the magic web-browser that doesn't do any kind of networking at all.

The web browser isn't Netflix trying to serve hundreds of gigabits per second of encrypted video streams from a single server. Do you really need the ability to reliably saturate a 40 Gb/s Ethernet link to browse Hacker News comfortably? You'll hit various other bottlenecks long before performance for practical usages of web browsers will be significantly impacted by a user-land network stack.

As I've said, there are use-cases where extreme throughput and latency requirements warrant a design focusing on performance. Smartphones aren't one of them.

> I don't know how you can possibly assert that, it's contradicting computer sciences' current understanding of operating system design as it relates to kernelmode/usermode switching, unless you're doing weird shared-memory things in userspace... which is terrifying.

Again, not everyone is Netflix. I'd rather have a computer capped at 1 Gb/s speed with a user-land network stack than a computer capable of saturating a 40 Gb/s Ethernet link with a kernel network stack when I'm managing my bank accounts. Most end-users don't need ludicrously fast network speeds to browse funny cat GIFs on their web browsers.

Also, I've contributed code to multiple operating systems (MINIX3, SerenityOS). Running an user-land network stack isn't going to turn your 1 Gb/s Ethernet card into a 10 Mb/s Ethernet card.

> Not really, C and Rust can interop just fine, you can have network drivers that are rust but the actual networking stack itself can remain C, if you want.

As far as I can tell, the bug is in the network stack itself. A network driver written in Rust wouldn't immunize your Linux kernel here from this bug.

> The post is literally memory safety bugs.

The consequence is about computer security, of which memory safety bugs are but one cause among many.

AndrewGaspar 1344 days ago

> The web browser isn't Netflix trying to serve hundreds of gigabits per second of encrypted video streams from a single server.

Ironically, server workloads are the ones that are increasingly moving to networking stacks that run in user space, using frameworks like DPDK, with performance as a motivator: https://en.wikipedia.org/wiki/Data_Plane_Development_Kit

Of course, there are some caveats - from my understanding, typical DPDK use cases would turn over the entire NIC to a single application, meaning you aren't contending with sharing the network between multiple, potentially adversarial user mode processes. This is fine for a server, but not really appropriate for a PC or smartphone.

Yes, the way Netflix and Co. are using Userspace drivers is by passing entire devices to a single application.

There's no general purpose IPC happening there.

vlovich123 1343 days ago

I wonder if io_uring will be able to maintain competitive against DPDK-like approaches. Multi tenant solutions are more attractive and seem like they could be extremely competitive since they should be largely equivalent in the case that you have a single tenant.

adrian_b 1344 days ago

> unless you're doing weird shared-memory things in userspace

Shared-memory things in userspace, i.e. buffers shared between 2 distinct user processes are no weirder than buffers that are shared between a user process and a kernel-mode driver. In both cases the buffers cannot be accessed by third parties.

Moreover, the transfer of data between 2 processes through a shared buffer can be done without any context switch (which could be slow), if the 2 processes are executed on distinct cores. Therefore having the network device driver as a distinct process does not have to cause any reduction in performance, if the means for inter-process communication are chosen wisely.

For any device driver that is implemented as a user process, the kernel can enable direct access to any I/O ports and memory-mapped I/O areas that are needed by the device, so the device driver can work in user mode without requiring any context switches.

Such direct I/O access cannot be enabled for ordinary processes, because those are not trusted enough and also because the direct I/O access could be enabled only for a single process at a time.

A dedicated device driver process solves both the trust problem and the multiple access problem equally well as a kernel-mode driver.

gpderetta 1344 days ago

Things are more complicated. You can indeed have a very fast network driver in userspace (in fact for many use cases userspace networking is faster than the kernel). But where do you put the rest of the network stack?

detaro 1344 days ago

> Ah yes, the magic web-browser that doesn't do any kind of networking at all.

They clearly didn't claim that. Your webbrowser being slow nowadays is not because it needs to do some networking.

They are claiming a loss in performance is ok.

I am claiming that people keep making this claim and it no longer holds true because software is already losing too much performance for the value we get back.

That's my whole thesis.

jjnoakes 1344 days ago

Your claim assumes that a small loss in performance in networking will lead to a loss in performance of the overall web browser, which is only true if networking is the bottleneck while browsing. And it usually isn't.

throw10920 1344 days ago

> I'd argue that this would be a bigger architectural revolution than porting the existing code and running it in user-land. MINIX3 went through such a change

MINIX is a microkernel architecture - running drivers in userspace is one of its core features/selling points, and one that differentiates it from (modular) monolithic kernels such as Linux. So, this isn't a very solid line of reasoning.

It seems to me that the situation is the opposite - that moving drivers to userspace is an architectural change, which is more complex than porting an existing architecture to a new language.

> Rust code can still be vulnerable in many other ways

Sure, but not vulnerable in the way that the vulnerability under discussion is.

> memory leaks

Much harder in Rust than C, and also unlike in C, not going to result in security vulnerabilities.

> buggy/compromised toolchains

If you're going to assume that your toolchain is compromised, than anything is on the table, including the toolchain inserting a backdoor into the kernel and completely bypassing the proposed architectural change of moving drivers into user-space. And, needless to say, compiler bugs are rare in general, and compiler bugs that cause software vulnerabilities are nearly unheard of (and I've literally never seen one before).

Nobody thinking rationally is going to tell you that Rust is going to eliminate all your bugs or make your code secure. However, by far, the majority of security bugs in the Linux kernel are due to mistakes that the design of Rust either completely eliminates or massively reduces.

And security is intrinsically a tradeoff - the Linux kernel is not optimized for maximum security (which would be something formally-verified like seL4), but a compromise between security, performance, and development velocity. The claim is that Rust will provide significantly better security at basically the same performance and possibly modestly improved development velocity - the very least that one should do is rewrite the existing architecture in it (or, again, a language that meets or exceeds the specs of Rust) and then see what the bug rate is before deciding to take a guaranteed performance hit through an architectural change.

boricj 1343 days ago

> MINIX is a microkernel architecture - running drivers in userspace is one of its core features/selling points, and one that differentiates it from (modular) monolithic kernels such as Linux. So, this isn't a very solid line of reasoning.

The first two versions of MINIX ran drivers inside the context of the kernel. The migration of drivers to user-space and overall emphasis on reliability didn't happen before MINIX3 in 2006.

Given that Linux nowadays has FUSE and UIO, still calling it a strictly monolithic kernel is probably a bit of a misnomer at this point. The same goes for Windows, NetBSD and others by the way.

> It seems to me that the situation is the opposite - that moving drivers to userspace is an architectural change, which is more complex than porting an existing architecture to a new language.

Trying to do a straightforward port of a C code base to Rust will quickly grind to a halt due to Rust's borrow checker. On a highly-optimized code base such as the Linux network stack, untangling every last optimization trick and shortcut to make the Rust compiler happy would require a large-scale refactoring that'll end up looking nothing like the original code base, at which point you might as well rewrite it from scratch.

In comparison, migrating that C code base to user-land would be less of a disruptive change to the code base (as was the case when MINIX3 did so with its drivers). It's still the same old network stack, adapted to run on a different environment.

> Sure, but not vulnerable in the way that the vulnerability under discussion is.

Rust isn't a silver bullet against every class of bug. Code that runs inside the trusted computing base means that it's one security bug away from system compromise. Writing that code in Rust doesn't change that.

loudmax 1344 days ago

A faster network is going to have a marginal effect on software getting slower. There are a lot of factors at play: network speed, cpu power, local caches, etc. The speed of the network driver is one factor among many and I'd be surprised if were the most common bottleneck.

The operators of websites that derive most of their revenue from advertising are going to run their sites at whatever level users will tolerate. Where network drivers are faster they'll either cram in more ad tracking, or won't bother optimizing their existing trackers. If users stop accessing sites because they're too slow to load, operators will either cut down on ad tracking, or more likely, put some effort into optimizing the performance of their ad trackers.

Rust is itself something of an architectural revolution. I believe network drivers in userspace is already a thing, and eBPF may also have a role here. All of this is worth exploring. This is what progress in Linux looks like.

> A faster network is going to have a marginal effect on software getting slower.

I don't mean to be glib but: citation needed?

One of the slowest moving hardware improvements (compared to CPU/Memory speeds) is networking.

It's going to take some serious convincing to tell me that we should just fork off performance; when software is already getting slower and slower.

It's not fair to blame advertisers exclusively, we also have electron and the hundreds of JS frameworks, that's before we get down to the low level abstractions that hook into basic programs.

tjoff 1344 days ago

The network stack isn't written in javascript.

The slowness of the web will not get any worse by moving the network stack to userspace. Guaranteed.

You have the entire body of computer science research around microkernels against you here.

What has changed?

cogman10 1344 days ago

CPUs got faster, memory got larger, compilers got better.

Microkernels aren't a dead idea and they are even making their way into consumer electronics. (see Zircon).

In fact, even newer oses are backing into microkernels. Windows adopted similar microkernel concepts in Vista with the HAL. Android adopted a HAL with project treble in android P.

The steady march of modern OSes has been moving driver logic out of kernel space and into user space.

binkHN 1344 days ago

> You can barely navigate the web now with a new low end computer...

You can barely navigate the web now with almost any computer. I have a high-end laptop and opening a few tabs from various sites on the Internet will cause the CPU usage and fan speed to spike. Just for a few tabs! Obviously not all sites do this, but the web has become a framework of advertising monstrosity and I can barely navigate and consume much of today's web content without enabling Reader Mode.

photochemsyn 1344 days ago

Well, you can navigate the web with older low end computers if you're running the full suite of blockers (NoScript & UBlock on Firefox seems to work pretty well), but that means many sites don't work unless you selectively allow certain scripts to work by fiddling with NoScript permissions and selective blocking some elements with Ublock. HTML + CSS sites are no problem, however.

miclill 1344 days ago

Regarding the first point. If I understand correctly you say that there is an inevitable performance hit when not running in kernel mode. But is that really so?

I don't like that you were downvoted for curiosity.

Yes, there is a performance hit but how large it is depends a lot on what you're doing.

"Switching" is one of the most costly operations and in Kernel mode you do not need to do it unless interacting with something in user space.. which you would only do because something in Userspace requested it somehow.

For other things, such as virtual memory, Microsoft found that the protections needed for virtual memory could be anywhere between 10 and 20%; but since there's no concept of virtual memory in kernel space: it's hard to say concretely that your program would be "20% faster". It would be too much of a different program.

galangalalgol 1344 days ago

The whole point of dpdk is that it bypasses the kernel to obtain speed. It almost certainly requires a lot of privileges, but the kernel itself isn't the source of the hit.

gpderetta 1344 days ago

The cost is in the switch between driver->kernel->network-consumer as running TCP and routing packets to the correct process are done by the kernel.

If you run dpdk you get raw ethernet frames and run TCP yourself. This means that a program can receive any data sent to the machine. More sophisticate cards can do routing in hardware and present multiple "virtual" cards, but this is not yet commonly available in consumer hardware.

dpdk is not general purpose networking. It's passing a device to a process and forgetting about it; the process itself needs to decode what is sent on the wire and make sense of it.

It's basically the kernel giving up on trying to do anything with the hardware, thus it's not available to any other process except the one that takes the hardware.

To have general purpose networking in user space you will end up with some other IPC which does not rely on sockets (because sockets are kernel) or shared memory which is dangerous as hell.

throwaway894345 1344 days ago

How do the micro kernel folks get around this? Are they paying the syscall toll between kernel components?

yjftsjthsd-h 1344 days ago

> How do the micro kernel folks get around this?

AIUI, in general, they don't. Avoiding the overhead of context switches is the research question in microkernels, and while there is progress being made, last I'd heard there wasn't a solution that didn't carry caveats. Now to be fair, sometimes those caveats are fine - ex. if you only write software in rust maybe you can get away with actually passing around memory without copies - but often they undermine the ability to run arbitrary software on general purpose hardware.

(If I'm behind and this has been solved in the general case, I'm happy to be corrected, but this was my undestanding of things as of a few years ago and I haven't heard about a breakthrough)

Yes, exactly!

Redox-OS is based on this approach and is honestly extremely clean and easy to understand (when compared to MINIX which also adopts the same approach).

Wikipedia has a much better explanation than I can give right now: https://en.wikipedia.org/wiki/Microkernel#Performance

mike_hock 1344 days ago

1) For this very reason, it doesn't make sense to optimize low-level components from the OS down to the hardware anymore, and hasn't for over a decade, if an improved user experience is your goal.

The short-term effect may be a slight improvement, but it's a treadmill and the next wave of web crapware will more than nullify it.

If a microkernel architecture with worse performance got established on all mainstream devices, the experience would be worse in the short term but in the medium term, the crapware would have to adapt so that it again becomes just barely usable as it is today.

The problem is that the short-term gains create an incentive for users to buy/use the marginally faster hardware or kernel, which then forces everyone else to follow suit.

amelius 1344 days ago

> Many will cry about that performance hit (including me).

Ok, we could make it optional, i.e. additional security for those who wish it (on top of other things we could do, like the things you mentioned but which aren't a panacea too).

devwastaken 1344 days ago

This is driver software, not relevant. Rust isn't a magic spell to be thrown out there when traditional solutions work quite well. The old reasons for hard kernel/userland separation are far less reasonable now.

See eg https://talawah.io/blog/linux-kernel-vs-dpdk-http-performanc...

fulafel 1344 days ago

Interestingly the user space way is also used in the performance absolutist end of the spectrum, so userspace can talk to the HW without kernel involvement (Snabb, dpdk, etc).

legulere 1344 days ago

For performance you want to reduce context switches between processes including the kernel. With kernel bypass you do everything in one process in the user space, but you are also losing out on features like sharing hardware resources between processes. With general user space drivers you will gain additional context switches. There are also ways to reduce the cost of those context switches though, like IO-uring that Linux recently got.

hansel_der 1344 days ago

indeed, but afaik these more or less depend on dedicating a cpu-thread to the software-thread because a cache eviction would completely wreck the performance

fulafel 1342 days ago

I think with networking appliances that are built from eg DPDK building blocks the main motivation for pinning and isolcpus usage is largely that users want to have a fixed topology of CPU cores, PCIe devices, NUMA memory pools/hugepages. But this doesn't really mean it would be necessary if we started using userspace networking drivers for general purpouse linux workloads.

matheusmoreira 1344 days ago

> Can we please stop running network drivers and network stacks in kernel mode by default?

No. I want the kernel to have as much functionality as possible. I have some zero dependency freestanding software that I boot Linux directly into. I really don't want to have to maintain additional user space in the form of C libraries. If I need to manage wifi connections, I should be able to make some system calls and be done with it without having to link to anything else.

Anything related to hardware belongs in the kernel so that all software can access it via Linux's amazing language agnostic system interface. If there are security problems, then that process should be improved without screwing up the interface by replacing it with user space C libraries. We have enough of that in the graphics stack.

mmis1000 1344 days ago

In the last decade, we have bump link speed from 100m 1G to some board even have 2.5G rj45 port now. I am not sure if move data between kernel and userland at 2.5G/s is even a good idea.

And even worse, because kernel still have to distribute data to other userland program, you actually need another round trip so the impact need to multiply by two.

zekica 1344 days ago

Worse, IP package sizes didn't grow at all. So it would require 200 000 context switches per second if you don't want to add code bundling packets kernel-side and dispatching them a-la Nagle's algorithm. This is both error-prone and adds latency. Even worse, trampolines and other mitigations make it so switches clear CPU cache.

legulere 1344 days ago

The trick is to not copy the data, but to pass pages around.

fulafel 1342 days ago

It's the opposite: In the last decade networking connection speeds have plateaued. In the 90s we had 100M ethernet, in the 00's 1G ethernet, in the 2010's 10G ethernet, and it stuck there for a long time. We had much much less CPU cycles available per packet when 1G came, for example.

mmis1000 1341 days ago

I don't think I have seen 10G ethernet network card in consumer grade hardware even recently. 10G is there in core infrastructure for a long time. But only until recently you see 2.5Gbps in endpoint devices. (Probably 2.5 makes more sense to a RJ45 head cable)

And also I don't think 10G routers/switchs ever use pure software based solution to handle the traffics. They are all hardware based or mixed solutions.

It's amazing that memory bandwidth / cpu speed / core counts grows so much that makes this even possible. But it still isn't a good idea.

GrabbinD33ze69 1344 days ago

I only have a very surface level understanding of linux, how does code running in user mode incur a bit of overhead as opposed to Kernel?

the_duke 1344 days ago

You really need to back up these assertions with some evidence.

Do you have benchmarks that show the impact of switching to userspace on a typical, loaded desktop system with all kinds of workloads? Or are you just guessing?

I did not anticipate a Hacker News discussion about a remotely exploitable Linux kernel WiFi vulnerability requiring some network benchmarks on an unusual network stack architecture, but I'll oblige:

This publication (http://www.minix3.org/docs/jorrit-herder/asci06.pdf) claims that MINIX3 could saturate a 1 Gb/s Ethernet link with an user-space network stack, with separate processes for the stack and the driver, on a rusty 32-bit micro-kernel that can't do SMP. In 2006.

nordis8 1344 days ago

> It's 2022 and we've got more than enough compute power nowadays to run in user-land

Oh hell no! User-land driver can barely handle WiFi 4 speeds (72Mbps), with terrible CPU performance (1000's of context switches and interrupts per second).

For WiFi 5 (ac), WiFi 6 (ac), you need heavy WiFi firmware involvement, multiple DMA queues and kernel driver, and even with all that it requires special care to reach target performance. There is no chance in hell to reach that kind of performance in user-land.

Manu40 1344 days ago

Uhm... no?

1. Date of year has nothing to do with doing things correctly, ever.

2. Ironically, pulling it out of the kernel and running it in user-land will probably bring about more bugs and issues. I would much rather we just fix the problem where it is, and leave it at that, instead of potentially introducing new problems, like backdoors and exploits in the software provided. Not saying it WOULD happen, but the potential for it alone is just not worth the risk in my honest opinion. Let's just fix the right way, and be done with it.

yardstick 1344 days ago

> I get that there are some use cases where performance really matters to the point where kernel network stack and drivers make a difference (high-throughput and/or low-latency services running on servers, high-performance routers...), but that should not be the default for everyone.

I’m on board so long as there is a choice. Routers with crappy hardware need as much help as possible. Also tangentially this is why the current darling of VPN tech, WireGuard, is implemented in kernel not userspace.

chasil 1344 days ago

Good gracious.

"WEINBERG'S SECOND LAW: If builders built buildings the way programmers wrote programs, then the first woodpecker that came along would destroy civilization."

meltyness 1344 days ago

This is more of a disparaging analogy than a law.

https://www.oreilly.com/openbook/opensources/book/appa.html

alfiedotwtf 1344 days ago

Obligatory comment:

Note that the Tanenbaum-Torvalds debate was in 1992, over thirty years ago. The security of computer systems might have improved since then, but the fallout of security issues has massively increased. Managing your bank accounts wirelessly on the Internet from anywhere in the world with a thin, battery-powered device that fits inside your pocket was a pipe dream (and no one could've possibly imagined that an Internet-ready smart toaster could compromise it and steal your money on your accounts).

dogleash 1344 days ago

> It's 2022

So someone else should have done it for you by now?

Be the change you want to see in the world.

I'm sure you have an excuse for not doing it personally. Just as I'm sure the person who you've mentally assigned responsibility has at least as good of an excuse too.

> Be the change you want to see in the world.

I have made dozens of commits to MINIX3, including a brand-new ISO 9660 file system implementation (https://github.com/Stichting-MINIX-Research-Foundation/minix...).

I have made more than a hundred commits to SerenityOS (https://github.com/SerenityOS/serenity/commits?author=boricj).

Just because I deplore the general state of security in mainstream operating systems doesn't mean that I demand that someone else does something about it for free.

I'm not paid to fix security bugs in the Linux kernel, do you expect me to fix these myself for free just because you want to? No one is entitled to my own free time spent hacking on random stuff.

314 1344 days ago

What kind of latency would this introduce on packet processing? A context switch used to be measured in microseconds. I don't have a linux system here to run lmbench on, and I wonder what the level of latency would be on a modern system.

[2] https://eli.thegreenplace.net/2018/measuring-context-switchi...

fulafel 1342 days ago

Current everyday networking has tens-of-microseconds latency[1] and on 2018 era CPUs the context switch overhead seems to be 1-2.5 microseconds[2]. So in benchmark and specialized use cases, context switch overhead would be measurable but many orders of magntude away from relevant when talking about communication over the internet.

[1] https://blog.cloudflare.com/how-to-achieve-low-latency/ - 30-45 microseconds on plain 10G ethernet - faster ethernet probably wouldn't improve much on this

throw10920 1344 days ago

"Drivers are exploitable so we should run them in userspace" is a hack, and not a good one.

The problem is that drivers are exploitable in the first place, so the solution is that we should make them not exploitable (using Rust, or a better language than Rust that fixes some of its problems) and try to preserve our performance that is rapidly being stolen away by bloated userspace software, rather than just shrug our shoulders and say "oh well, I guess that drivers are just intrinsically insecure".

legulere 1344 days ago

It's not a hack. Virtual address spaces and process isolation were built exactly for misbehaving code.