（评论）

（评论）
(comments)

原始链接: https://news.ycombinator.com/item?id=41548474

该用户之前在 Firefox 上享受流畅的流媒体视频，但在 Chrome 中很难启用 YouTube 播放的硬件加速。由于在 Linux 上配置视频播放很困难，他们最近从 Linux 切换到了 Windows。用户发现与之前使用 Linux 的体验相比，Windows 更能支持硬件和软件需求。然而，他们提到桌面系统上的 Linux 缺乏官方支持，并且经常由于驱动程序质量差而崩溃。这会导致频繁出现的问题，例如配置损坏、窗口管理器错误和其他问题。尽管社区努力改善这些问题，但他们发现许多用户仍然遇到问题。在 Linux 上，AMDGPU 驱动程序给用户带来了问题，包括持续的帧缓冲区问题。解决此问题后，他们考虑改用 NVIDIA 卡并使用 Nouveau 驱动程序以获得更好的可靠性。为了测试不同的内核，用户提出了一个测试环境，其中多个设备（例如服务器、笔记本电脑和智能设备）进行网络启动并执行各种压力测试，以确保所有驱动程序顺利运行。不幸的是，建立这样的测试环境需要大量的 C 编程技能，而用户并不具备这些技能。

6.10 (TEN, the previous one) has been a very problematic release for me, with one desktop running into four major bugs in total: three separate amdgpu bugs resulting in video corruption, hangs and crashes, and now that I'm on 6.10.10 and those seem to be fixed, the system intermittently refuses to come up from sleep mode.

Anyone else having similar experience? This is the first time something like that happened in a decade of using the latest stable kernel release (in my experience, it's actually been stable for all that time except for 6.10).

I am really surprised with RNDA3 support. I have never seen so many issue with iGPU (APU). It started with VP9 decoder issue (e.g. just playing videos on YouTube was enough to trigger it), but that got fixed after a very long time (required a new firmware). Multiple constant [different] crashes, but you can workaround most of them by adding amdgpu.sg_display=0 to your bootargs. It's already listed in Arch Linux wiki, Gentoo wiki, etc.

Again, I was surprised by the number of firmware and driver issues since RNDA1/2/3 have been around for years now.

6.10 broke my fedora gaming proton box and I was on holiday at the time and so upset I just nuked it and put windows on it to play games. Now it's only powered on on the weekends to play games and I've moved all my linux needs to a VM on my overspecc'd Mac.

I also had an AMDGPU system. 5600X, AMD 6800 GPU, Fedora 40. -> now win11 (which has so much cruft out of the box I am considering nuking it and going back haha)

Broke my nixos steam/proton setup also. Using game scope as the compositor. Switch to LTS (6.6 I think) solved it. Frustrating bug but they don’t call it nixos-unstable for nothing. 7900 GRE

Yeah ... had I had an immutable system or something where rolling back was easy I would have ... I dunno Fedora/Linux + AMD hardware had been so rock solid I took it for granted.

Wow thanks for that link. I've had my machine crashing a couple of times as well the last couple of weeks, was absolutely stable before. I hope it's this and not hardware failure.

Just to add a non-problematic experience report to the mix: I've been using 6.10 for months on two AMD machines with different hardware (one with a 7840U and one with a 5700XT) without any issues whatsoever.

The amdgpu driver has been the main source of issues for me on 6.10, but I had issues on older 6.x versions as well: for example, on a desktop with 2 monitors, I had to turn on the 2 monitors simultaneously or the UI would freeze.

The past few releases were more problematic for me yeah. It's super anecdotal since I never used Linux before the v5.xx kernels but in comparison to those yes it's a bit less stable.

Kernel version 6.8 broke suspend ($ systemctl suspend) for me. I run two machines with near identical setups. I upgrade my "preview" machine first before updating my primary machine to test for defects.

If I boot from kernel version 6.5, suspend works fine. Hold shift while your machine is booting and the grub menu will allow you to select a different kernel version.

I too have been having AMD GPU video artifacting lately, but so far that is the only regression I've noticed in 6.10.x. I am still on 6.10.8 so I'm not sure if 6.10.10 will contain a fix for me just yet.

Is that just AMD? On my thinkpad X270 playing videos in Firefox is just a mess. All sorts of problems while Chromium is just fine. It's also fine on a copy of my system that I run on a thinkcentre tiny.

I do believe Chromium still has to be patched to support HW decode via vaapi, while recent Firefoxen turn it on out of the box. So it's likely that Chromium is using software rendering, bypassing any driver-related bugs.

Chromium doesn't document this but they support hardware video decode on Linux. Works best on Intel iGPU. AMD needed some patches to get it working using Vulkan IIRC and Nvidia is Nvidia.

What did you have to do to get it to work for Intel iGPU? Passing all combinations of command line flags never made it work.

intel_gpu_top showed both Video and VideoEnhance markers signaling hardware decode as inactive, and all commenters on message boards that I came across essentially came to the same conclusion, that it doesn't work at least at the time I as trying a few months ago. I think it was that Chromium devs don't support it and block it from functioning in the code because of this or that upstream dependency not being stable.

Have no problem using just this flag --enable-features=VaapiVideoDecodeLinuxGL on a laptop with HD Graphics 500. And I even think that flag is passed by default now.

Key thing is to make sure you are playing back supported codecs when checking for hw decode. E.g YouTube randomly forces av1 on clients and some older hardware don't support hw decode for it.

Also some distros like Fedora and OpenSUSE ship with partial codec support in their installation media.

An easy way to make sure you have all the right drivers in place is to test this in one of the Chromium flatpaks.

Though I have to state hardware video playback is not yet officially supported and it has broken a lot over the past due to this fact and the quality of the different GPU drivers in question here. The Chromium devs don't explicitly block vaapi on Intel hardware.

I always had all the codecs and Vaapi on Firefox worked fine, but could specifically never get Youtube on Chromium to use GPU decoding, that was the main problem you're reminding me.

In this past year, I've since moved to Windows after like 7 years of constantly messing and tweaking Linux to kinda sorta almost for some stretch of time get all the things to work after spending hours looking for the right configuration. Look at how long this thread is, for example[0]. Or here most saying it still doesn't work[1] and the flags constantly change. Before Linux I was about 6 yrs on Mac. I hadn't used Windows for so long, it actually felt weird to have so much hardware and software support for anything I ever need to use.

I'll take your word for it that Chromium sometimes uses hardware decoding now. But as you say, and they should put that slogan on a T-shirt: "Linux on Desktop: not officially supported and it has broken a lot over the past due to lack of support and quality of driver". I wouldn't really characterize that as "have no problem using". It feels nice to have a system that just works without all that hassle and tweaking so I can just get actual work done, or watch Youtube efficiently without brittle configuration, no window manager bugs etc.

[0] https://bbs.archlinux.org/viewtopic.php?id=244031&p=40

[1] https://www.reddit.com/r/linux/comments/1ey4n2k/heres_how_to...

amdgpu on my 6.10 kernel has been crashing constantly too. It makes me want to go back to Intel. My workaround has been to let the ryzen 7700 iGPU run at its max clockspeed of 2200 Mhz …

amdgpu is shit for me, my friend. Funny story: my headless server with a small Navi 1 workstation card (repurposed) couldn't be SSH'd into last week. Went and plugged in a monitor, rebooted, and the framebuffer got stuck during stage 1 while fsck'ing my disk. OK, I think to myself, it's probably taking a while to fsck after N boots, happens once every few months. Doesn't change for 48 hours. Turns out my machine had just run out its DHCP lease, so it had a new IP and I didn't realize it, which is why I couldn't log in. So I log in, and..

What was wrong? What actually happened was that on bootup, the amdgpu driver would panic and fault during boot exactly when fsck was happening, and the framebuffer would be stuck forever. So it just looked like a filesystem issue but in reality my graphics output was merely fubar; the system itself was otherwise fine tnough.

This is reliable and reproducible for me; it always faults at almost the exact same spot at boot every time, for this kernel version at least. In reality amdgpu has been unreliable for me for 5+ releases at this point on a card less than like 7 years old.

Really considering moving over to a small cheap nvidia card and just running Nouveau instead. At least then I might have a reliable framebuffer.

Yeah I've experienced a fair number crashes with 6.10 (multiple point releases) on my 680M. Earlier kernels were fine in my recollection although I didn't bother downgrading, and Windows remains rock solid on the same system.

Doing well on a recent AMD, although I didn't upgrade until about 6.10.7 or so.

There were problems on earlier kernels, but they were mostly fixed by a firmware update. Kernel already had support by that time.

How does Linux handle testing? They don't seem to have a CI system as far as I know. Presumably there's no big lab with automated testing on real hardware. Does anyone know how releases are tested?

Take a machine, set it up to network boot, plug it in to a ZigBee-controlled plug. Every time you have a new version, a script bumps the netboot kernel and power cycles the machine, then runs tests. Those tests include "does it boot and respond to ssh", but also "can it run render jobs on the GPU", "does KDE start and if we connect over VNC did it render correctly", "do NFS performance tests show expected numbers", etc. Ideally, this box has a lot of hardware plugged in and those tests exercise all the drivers you can think of. For bonus points, scale horizontally; each kernel version can get booted on a Dell server and a Raspberry Pi and a Macbook and a Raptor POWER workstation and a Steam Deck (probably with battery removed) and [...], and all of these can netboot and run tests in parallel.

Honestly, I'd love to set up this kind of thing, but I've not seen a job for it that doesn't need C expertise:(

Yeah, same problems with 6.10 and amdgpu. Radeon Pro WX 3200 fwiw. I've been on 6.8.9 for several weeks now. Just today I booted 6.10.7 and it's been stable so far. I haven't tried to put the system to sleep yet, though.

This isn't the first time I've had problems with the stable kernel, though. A while back I had problems, also graphics related, with Intel i915 (my onboard graphics that I used before I got the AMD card). It took a while but it eventually got fixed. I haven't looked to see if there's a bug tracked for the AMD problem.

Man, time flies. I remember Slashdot's thread announcing it.

Linux versioning now is the worst kind of arbitrary! It's the web-browser "just iterate the number" method, but with the appearance of semantic versioning.

No way. KHTML is a cautionary tale of what happens when open source is coopted by corporate power. How can you call KHTML successful when it is dead, with its progeny in tight control of Google/Apple?

Because it formed the basis for essentially every web engine that is used today? I would call that successful.

(Before you say "but gecko!", take a look at the usage stats.)

Not only release notes, he's still actively contributing! From said release notes:

> Linus Torvalds (2):

> mm: avoid leaving partial pfn mappings around in error case

> Linux 6.11

I mean, it's GPL. If something were to happen to him, anyone could just start their own fork. There would likely be some competition between forks for a while before two or three emerge as the successors eventually.

Linux is basically a national security issue at this point with how much runs on it. Linus is an interesting dictator because he's not bound to any corporation. I think some kind of kernel throne wars would be terrible for the community and could result in Red Hat or some other corporate overlord owning the kernel (not literally because of trademarks but effectively through the release and mainline).

PREEMPT_RT would be a boon for audio/video work, hope it finally goes in. Not everyone who needs it knows they need it, let alone how to compile it, and having it in-tree would massively simplify the process for distros to [re-]package a proper RT kernel.

Yes, that's why it's a config option; but if the code is in the kernel a distro can package up a preempt-rt kernel by flipping the config flag rather than maintaining an out-of-tree patchset, which is a _much_ lower burden.

When will be the year of GNU/Linux for smartphones? Android is not it. I wish we could install distros in a smartphone as easy as it is in desktop

"As easy as in desktop" not going to happen anytime soon. Phone "bootloader firmwares" lack full description of phone hardware in stardartized form like PC ACPI does, expecting customized OS kernel to know the phone it's running on. Phone devices lack capability to emulate something ancient-but-standard, expecting customized OS to include drivers for all of them. That's enough to make unified OS images impossible.

Fighting against well known hardware on PCs is one thing, doing the same against mobile manufacturers because they refuse to release any documentation is a nightmare. Linux will have a hard time becoming a reality in the mobile world because of hardware manufacturers hostility, not for any of its faults. Pine64 had to design the PinePhone platform from scratch for this exact reason, still they encountered so many blocks that when it finally became available it was already too old.

Nobody understands it outside of a tiny hobbyist segment, because there is no "definitive" Linux. No two "Linux" look the same.

The year of the Linux desktop will be when there's an OS like Android or macOS that is running Linux but the only way you would tell is by running uname.

I thought the issue with this nowadays was that hardware support for suspend-to-ram has been increasingly removed from hardware.

At the same time, S3 sleep worked just fine on supported hardware 10 years ago. So what does suspend-to-ram have to do with Linux 6.11 in particular?

Always worked for me on Thinkpads (T41/T61/T420/T520). T420 reports a somewhat optimistic "122 hours remaining" when coming out of sleep with a fully charged battery though.

Is suspend-to-ram something that often doesn't work on laptops with Linux, really? I used it on dozens and saw problems maybe once or twice. Though I usually pick ones which are 3+ years old.

The problem is that platforms are removing S3 and S2idle support wasn't great for a while. Those problems should be mostly resolved in 2024 though.

Not at all. It is a poor pattern because they throw a kernel over the wall right before the freeze, then refuse to fix any of the bugs, because of the freeze.

The kernel should go out to general testing as soon as the cycle starts, not right before it ends.

> but if you list the important advancements of humankind (like LLMs, drug development, basic science breakthroughs etc.)

Amusingly, free and open source kernels (including bsds) are key to enabling all of that.

No, not at all. GPUs are the key to enabling all that. Kernels are just some glue running in there. Most of the NVIDIA goodness is actually in userspace. You could just as well say that electricity or HVAC is the key.

If crypto is the bar, Linux looks even worse. At least crypto was novel in the past 30 years. Linux works on new hardware and... that's it. There's nothing new about Linux. As far as OSes go there is seL4 other interesting research. Linux is a toilet. It flushes. It has utility. That's good, but I don't need to hear about every new release of something which just isn't changing in fundamental ways.

And you don't think plumbing is important? Are you secretly a fan of cholera?

And stability is important. Things which change all the time for the sake of change are extremely tiring, e.g. the JS/TS/JSX/whatever is new over there these days. Yet stodgy old Linux provides more utility than almost all other software.

Pretty bad take and calling it “a collection of drivers, some VMM, and some syscalls” is a bit reductive.

Linux kernel is built and maintained across a variety of people. To my knowledge, nobody is directly paid by Linux foundation yet these people come together to slowly improve the kernel everybody uses either directly or indirectly. It’s also a demonstration of the greatness that humans can achieve if we work together to solve a problem(s) outside of the typical “capitalistic” motivations (ie, money)

A nuclear power plant is just a bunch of pipes and a pile of uranium.

/s

We like it because it is free, it is not a technological amazing breakthrought, but as a collaborative project it's kind of successful in time, you have to agree on that :)

I wouldn't put the kernel on the same footing as a nuclear power plant. A power plant is designed, the Linux kernel is accreted, like a stalagmite. It's a testament to the power of open source that something useful can come out of that process. Amazing, in fact. But the newsworthiness of the kernel is overstated, in my view.

（评论） (comments)

（评论）
(comments)