Comment by vlovich123 | Hacker News Reader

Comment by vlovich123 | Hacker News Reader

Parent Full thread

You’re answering a slightly different question but to me that’s a Debian packaging problem to solve. It’s weird to me that QEMU devs take this problem seriously enough to be putting in all sorts of workarounds to support old versions of the toolchain in the tip of tree just to privilege Debian support.

This feels more like a CI thing for the QEMU project and I’m sure solvable by using rustup or a trusted deb repo that makes the latest tool chain available on older Debian platforms.

As for Debian itself, for toolchains it really should do a better job back porting more recent versions of toolchains (not just Rust) or at least making them available to be installed. The current policy Debian is holding is really difficult to work with and causes downstream projects to do all sorts of workarounds to make Debian builds work (not just for Rust by the way - this applies to C++ as well). And it’s not like this is something it’s unfamiliar with - you can install multiple JVM versions in parallel and choose a different default.

It's not about our own CI -- we could easily use rustup as part of setting up the CI environment, and I think we might actually be doing exactly that at the moment.

Lots of QEMU users use it through their downstream distros. We even recommend that if you're using QEMU in a way that you care about its security then you should use a distro QEMU, because the distros will provide you timely security fix updates. Sure, we could throw all that cooperation away and say "tough, you need to use up-to-the-minute rust, if that's a problem for distro packagers we don't care". But we want to be a good citizen in the traditional distro packaging world, as we have been up til now. Not every open source project will want or need to cater to that, but I think for us it matters.

That doesn't mean that we always do the thing that is simplest for distros (that would probably be "don't use Rust at all"); but it does mean that we take distro pain into account as a factor when we're weighing up tradeoffs about what we do.

To be clear. I’m not criticizing the position the QEMU project is in. I recognize you have to work with Debian here. I’m more frustrated that Debian has such a stranglehold on packaging decisions and it effectively refuses to experiment or innovate on that in any way.

Out of curiosity though, have you explored having your own deb repo instead? I would trust QEMU-delivered security fixes on mainline far more than the Debian maintainers to backport patches.

I think that trust would be somewhat misplaced -- QEMU has historically not made particularly timely security fixes either on mainline or on branches. To the extent that our stable-branch situation is better today than it was some years ago, that is entirely because the person who does the downstream Debian packaging stepped up to do a lot more backporting work and stable-branch maintenance and releases. (I'm very grateful for that effort -- I think it's good for the project to have those stable branch releases but I certainly don't have time myself to do that work.)

As an upstream project, we really don't want to be in the business of making, providing and supporting binary releases. We just don't have the volunteer effort available and willing to do that work. It's much easier for us to stick to making source releases, and delegate the job of providing binaries to our downstreams.

These two statements to me seem contradictory:

> QEMU has historically not made particularly timely security fixes either on mainline or on branches

> It's much easier for us to stick to making source releases, and delegate the job of providing binaries to our downstreams

Am I correct that this is essentially saying "we're going to do a snapshot of the software periodically but end users are responsible for applying patches that are maintained by other users as part of building"? Where do these security patches come from and how do non-Debian distros pick them up? Are Arch maintainers in constant contact with Debian maintainers for security issues to know to apply those patches & rebuild?

Security patches are usually developed by upstream devs and get applied to mainline fairly promptly[1], but you don't want to run head-of-git in production. If you run a distro QEMU then the distro maintainers backport security fixes to whatever QEMU they're currently shipping and produce new packages. None of this is particularly QEMU specific. There's a whole infrastructure of security mailing lists and disclosure policies for people to tell distros about security bugs and patches, so if you're a distro you're going to be in contact with that and can get a headsup before public disclosure.

[1] and also to stable branches, but not day-of-cve-announcement level of urgency.

Sure, but then why does the mainline branch need to worry about supporting the rust that’s bundled with the last stable Debian release? By definition that’s not going into a distro (or the distro is building mainline with rusts latest release anyway).

Is it a precautionary concern that backporting patches gets more complicated if the vuln is in Rust code?

But then again Rust code isn’t even compiled by default so I guess I’m not sure why you’re bothering to support for old versions of the toolchain in mainline, at least this early in the development process. Certainly not a two year old toolchain.

We already make an exception in that we don't support Debian bullseye (which is supported by the rest of QEMU until the April 2025 release), but not supporting Debian stable at all seemed too much.

That said we will probably switch to Debian rustc-web soon, and bump the lower limit to 1.75 or so.

I think I'm missing why you need to require using the toolchain bundled with the last stable Debian release vs having devs just rustup the latest version of the toolchain (or via a PPA [1] or however else they want to install it).

The current approach basically guarantees that you're always targeting a ~2-4 year old version of the toolchain and that feels like a particularly weird maintenance burden given how many workarounds you're putting in to do so.

[1] https://launchpad.net/~jonathonf/+archive/ubuntu/rustlang

Because it's not about devs, it's about distro packagers. They are an extremely important audience for QEMU. As pm215 said above, we don't want to make their lives unnecessarily harder. For example Debian has the backports repository, and CentOS Stream has QEMU and Rust updates done by different teams.

Anyhow starting April (August release) we will be able to target 1.75.0 while being consistent with QEMU's (C-targeted) distro support policies, which is not that bad. Maybe newer than that depending on what Ubuntu 22.04 does between now and August.

> I’m more frustrated that Debian has such a stranglehold on packaging decisions and it effectively refuses to experiment or innovate on that in any way.

What Debian has is not a "stranglehold" but an ideology, and Debian continues to matter to (some) upstream projects because lots of users identify with Debian's hyperconservative, noncommercial ideology.

Your complaint is basically, "it's too bad that the userbase not sharing my values is large enough to matter".

> What Debian has is not a "stranglehold" but an ideology, and Debian continues to matter to (some) upstream projects because lots of users identify with Debian's hyperconservative, noncommercial ideology.

> Your complaint is basically, "it's too bad that the userbase not sharing my values is large enough to matter".

Arch and rolling releases have about the same market share as Debian. Indeed, ironically, Debian's widespread adoption is seen primarily in the enterprise space where it's free as in beer nature and peer adoption is a signal it's a suitable free (as in beer) alternative to RedHat. Without Ubuntu's popularity a while back making Debian not so crazy an idea, I think "Debian" philosophy would not have anywhere near the adoption we see in commercial environments.

> I’m more frustrated that Debian has such a stranglehold on packaging decisions and it effectively refuses to experiment or innovate on that in any way.

Does Debian have a stranglehold? AFAIK every other distro does the same thing, and all of them for good reasons.

Ubuntu and Arch are about equal market share penetration for the desktop from what I researched with other and steam deck being the main dominant categories. So ignoring corp fleet deployments, I'd say Arch and NixOS have stolen quite a bit of market share from Debian-based systems in terms of end-user preference. But yes, Debian does still have a stranglehold because corp $ are behind Debian-style deployments.

> As for Debian itself, for toolchains it really should do a better job back porting more recent versions of toolchains (not just Rust) or at least making them available to be installed.

Most toolchains don't have as much churn as Rust.

Disagree. No stable release of Debian today supports c++23 which is coming up on two years old at this point (this corresponds to a major Rust edition, not the minor stuff they publish on an ongoing basis every month).

Java in Bookworm installs JDK 17 which is three years old at this point. Java itself is on 23 with 21 being an LTS release.

This means that upstream users intentionally maintain an old toolchain just to support Debian packaging or maintain their own deb repo with newer tools.

You’re confusing cause and effect. People aren’t migrating because Debian packaging lags so badly, not because there aren’t improvements projects would love to otherwise use.

> Most toolchains don't have as much churn as Rust.

What churn? A release every 6 months? Unlike many others, toolchains (i count nodejs and co here) rust only need one toolchain because latest rustc always able to compile older rust code.

Now compare rust releases to this: https://gcc.gnu.org/releases.html

Rust’s release cadence is 6 weeks not 6 months.

My bad, I thought it was 6 months. Okay, a bit too often, but no reason why debian folk can't update it more often than it is now.

> latest rustc always able to compile older rust code.

That is not true. Adding any public method to any impl can cause existing working code not to compile, and Rust adds new methods to the stdlib all the time.

I don't believe this is correct. I have not once seen a new method being added to stdlib that causes code to not compile. And ABI concerns are a non-issue since Rust is statically linked.

It is correct. For example, this code (admittedly a contrived example, just to illustrate the point) compiles in 1.82 but not in 1.83:

  trait OptionExt {
      fn get_or_insert_default();
  }
  
  impl<T> OptionExt for Option<T> {
      fn get_or_insert_default() {}
  }
  
  fn main() {
      Option::<()>::get_or_insert_default();
  }

The reason is that `get_or_insert_default` was added to the stdlib in 1.83, and takes a different number of arguments, so it clashes with the user-defined one here.

Interesting, I would think that having OptionExt in scope would make it clear which method to use.

Reply on news.ycombinator.com