Lux under a pile of leaves 🍂🦊🍂

Asahi Linya (朝日りにゃ〜)'s posts Post Back to profile

@Sobex @lina there is a repo, it's where the previous work resides in the rust branch from before the merge in 6.1. That was where we would have wanted a branch, but review rules are what they are and we would have needed to go through LKMLs with netdev people and netdev people will not review code that is not aiming to be merged (which i understand, but then why force us through LKMLs?)

And it was all just diplomacy and sticking to how things are done

Like 29 August at 18:10 | Wall-to-wall | Open on eldritch.cafe

15 comments

Asahi Linya (朝日りにゃ〜)

@SharpLimefox @Sobex The kernel dev model is very hostile to major changes like this, because the upstream devs refuse to review anything out of tree but then expect submissions to always come with users. That's why I had to submit my DRM abstractions with one huge commit at the end of the series adding my whole GPU driver, even though as expected that got zero reviews...

It causes a lot of friction any time we need to touch any core code because "show us how you're going to use it in the series" is a very effective stalling tactic when that means adding 50 patches of extra stuff just to get to the point where you can merge the intended user...

Expand text...

29 August at 19:14 | Open on vt.social

Lux under a pile of leaves 🍂🦊🍂

@lina @Sobex the current model is fine for C developers trying to add simple things or fix bugs and so on, but when large architectural changes are needed (aka you welcome a new language in the codebase) it's definitely a major point of friction and literal slowdown

Regarding reviews, IIRC, the first time a network driver was sent to netdev, the first reply arrived 48h later at least and it was an apology that nobody was reviewing (i could pull the exact lore links when i'm back at a computer): there was just nobody available to review Rust. Concurring with that (and I'm sorry i'm putting my academic hat on because this is exactly what I'm focused on atm), Hongyu Li et al0 found that "RFL is bottlenecked by code review but not by code development" (insight 3), and it's pretty easy to see anytime you get an actual driver sent to the LKMLs for reviews, which is somewhat rare these days

i remember seeing your patchsets dying and your presence in the LKMLs diminishing, and it was quite sad, because of all the work put in it :( i'm also very interested in making abstractions, but finding an end user without paying for obscure or brand new hardware is so hard in net, so i just can't contribute anything, and i'm not alone...

Hopefully things fare better in the future but it's going to need a lot of change of mind from a lot of people..

Expand text...

29 August at 19:30 | Open on eldritch.cafe

Asahi Linya (朝日りにゃ〜)

@SharpLimefox @Sobex Honestly the thing that has burned me out the most has been the scheduler saga and seeing how slow the core Device/platform driver stuff (+ the device ID pain) is moving (due to pushback from upstream mostly). I don't want to take on that work myself because I didn't write the original code and I don't fully understand the nuances of the device binding cycle/lifetimes and how that was originally designed for the Rust abstractions (+ my GPU driver literally can't support rebind due to firmware limitations so it's a terrible test/demo case of how to do all that properly). But I depend on that to get anything done...

Once platform devices are supported I'll go back to sending my stuff and I hope it'll be a lot less blocked after that. For better or worse the approach ends up being "don't depend on subsystems where upstream doesn't want to cooperate and will stall, if you can help it"... I think I have a pretty good idea of what those are so things should move forward. I hope...

Expand text...

29 August at 19:49 | Open on vt.social

Sobex

@lina @SharpLimefox Is it possible to avoid the subsystem that are a pain in the ass, when writing drivers ?

Or is this more a case of not upstreaming those drivers and upstreaming stuff that goes in Rust friendly subsystem first ?

29 August at 19:55 | Open on social.sciences.re

Asahi Linya (朝日りにゃ〜)

@Sobex @SharpLimefox It's possible in some cases, not others. You can't avoid the subsystems "above" you providing your services (DRM for me) and "below" you interacting with other devices, but you can avoid "helper" subsystems and parts of them.

The DRM scheduler is a helper. The other thing I'm going to RiiR is the GPU page table management, which is currently using the IOMMU page table helpers (both because the maintainer stopped replying to my emails... and because since the original submission my needs have diverged enough from the standard code that I think it stopped making sense to reuse it and try to fight to shoehorn those features in).

Expand text...

29 August at 20:02 | Open on vt.social

Lux under a pile of leaves 🍂🦊🍂

@Sobex @lina IMO regardless of subsystem you'll have to endure the structural issues plaguing RFL, but a subsystem full of people willing and motivated to help should alleviate those significantly

And there's also the problem that if your interest is in a specific area of kernel dev (drm, net, whatever), you've only got so much wiggle room to avoid maintainers unwilling to help more.

Someone i talked about this with mentioned forking as well, but that comes with the human cost of maintaining the upstream source (mainline), people having to decide whether to leave the main project, stay, or invest twice as much energy, and the risk that nobody will really use your fork (Asahi Linux being the huge exception for that point lmao)

@Sobex @lina IMO regardless of subsystem you'll have to endure the structural issues plaguing RFL, but a subsystem full of people willing and motivated to help should alleviate those significantly

And there's also the problem that if your interest is in a specific area of kernel dev (drm, net, whatever), you've only got so much wiggle room to avoid maintainers unwilling to help more.

Expand text...

29 August at 20:05 | Open on eldritch.cafe

Asahi Linya (朝日りにゃ〜)

@SharpLimefox @Sobex Hard forks aren't sustainable. Asahi Linux tracks upstream and with that comes a maintenance cost proportional to how much it diverges. The plan was always to upstream almost everything... the project doesn't have a long term future if we get stuck maintaining large amounts out of tree forever (a few patches is OK and we already have a couple "upstream is never going to take this and that's fine, we can live with that" cases... but it can't grow without bound)

29 August at 20:11 | Open on vt.social

Sobex

@lina @SharpLimefox Urgh, what falls in the upstream won’t take that ever case ?

29 August at 20:14 | Open on social.sciences.re

Asahi Linya (朝日りにゃ〜) replied to Sobex

@Sobex @SharpLimefox TSO support for fast x86 emulation and the M1 cpuidle driver, both are NAKed by all the ARM employees (for politics reasons) and they own the arm64 platform subsystem.

The actual patches are pretty small and not intrusive though so it's not a big maintenance burden.

In theory the cpuidle stuff is supposed to be replaced by some PSCI transport alternative that doesn't exist, but that idea has been floated for years now and not gone anywhere...

29 August at 20:22 | Open on vt.social

Sobex replied to Asahi Linya (朝日りにゃ〜)

@lina @SharpLimefox Hmm, what are the politics reason (rather than undocumented, and not required to make the hardware run) ?

29 August at 20:24 | Open on social.sciences.re

Asahi Linya (朝日りにゃ〜) replied to Sobex

@Sobex @SharpLimefox For TSO they fear "fragmentation" as in developers using TSO to work around arm64 porting issues instead of fixing the code, even though there is no evidence of anyone ever even thinking to do that.

For cpuidle, the maintainers want PSCI (a firmware interface) to be the only cpuidle driver to avoid driver proliferation. Unfortunately the PSCI spec was written with assumptions about the platform that make it impossible to implement on Apple Silicon despite it being a conformant arm64 implementation (they require optional features).

Expand text...

29 August at 20:27 | Open on vt.social

mort

@lina This makes me wonder, how does the future actually look like for your driver? Will the M1 GPU driver just always live out-of-tree because upstream is institutionally incapable of accepting GPU drivers written in Rust? Or are there signs of progress?

(Also, thank you for all your work, I'm a happy user of your driver and it has been nothing but great for me :3)

29 August at 19:35 | Open on fosstodon.org

Asahi Linya (朝日りにゃ〜)

@mort The DRM people are actually very nice in general except for that one scheduler guy. So I don't expect any major trouble merging the driver itself once the dependencies are in. ^^ (and I might seriously just rewrite the scheduler in Rust as driver-internal code just so I don't have to deal with that guy and that mess of buggy C code... it's not just the lifetime stuff, I keep finding outright memory safety bugs.)

The good news is Nova (the new Nvidia driver) is also written in Rust with the same abstractions, so now we have two GPU drivers to put some pressure on the people blocking things...

Expand text...

29 August at 19:43 | Open on vt.social

mort

@lina Oh, that's great news! Being the odd one out is always more challenging, I wasn't aware that Nova was also Rust.

29 August at 19:49 | Open on fosstodon.org

Imikoy

offtopic

@lina@vt.social

there is a rust driver for the gpu that I have

I can touch gpu rust code in the kernel and see it do things... *appends to the list of remember-later topics*

29 August at 19:50 | Open on meow.miabaka.moe