Conversation

Thorsten Leemhuis (acct. 1/4)

QOTD from @torvalds:

"'[…] no [] developer should spend one single second worrying about out-of-tree modules.

It's simply not a concern - never has been, and never will be.

Now, if some out-of-tree module is on the cusp of being integrated, and is out-of-tree just because it's not quite ready yet, that would maybe be then a case of "hey, wait a second".

But no. We are not going to start any kind of feature test macros for external modules […]'"

https://lore.kernel.org/all/CAHk-%3Dwg1w6%2BXup%3DamYtYUCLO-SRYoy9R0z6BG-uGV%3Dy2f6yFWA@mail.gmail.com/

1
4
3

2/ quick reminder, while at it:

That "'[…] no [] developer should spend one single second worrying about out-of-tree modules. […]" from Linus includes the still relative new Open Source Linux kernel driver from Nvidia a lot of people are so happy about which becomes the default with the new 560 driver series.

IOW: these drivers will continue to break occasionally when switching kernel versions, just like Nvidia's old kernel driver with its proprietary core did.

5
0
0
@kernellogger out of tree ebpf sched_ext code will be guaranteed to work though ;)
1
0
4

@ljs

😄

I recently already prepared mentally for the first regression report related to sched_ext, even if that likely is still some time away. But that's better then being surprised by it. 🥴

1
0
1
@kernellogger IOW confused me, I'm not good with Internet acronyms so I mixed it up with _IOW first ;-) ioctl's fit to the driver context...
0
0
0

@kernellogger honestly in practice it's not that big of a problem because NVIDIA is good enough at supporting those themselves

2
0
0

@never_released usually yes, but that was the case in the past already, too. Nevertheless people occasionally ran into problems, for example when distros did not pick up the updated drivers and/or when people were running the mainline -rc kernels.

Also do not forget that some (many?) kernel developers will ignore even apparently unrelated bug reports if you have out of tree modules loaded.

0
0
0

3/ fun fact:

Just a few hours after above toots OpenZFS released a new version[1] – one that finally brings compatibility with 6.9.y, which became EOL about 9 days earlier…[2] 🤷

[1] https://github.com/openzfs/zfs/releases/tag/zfs-2.2.5

[2] https://fosstodon.org/@kernellogger/112869284061295755

1
0
2

@never_released @kernellogger it depends. Last year the kernel driver broke as usual, and while the desktop driver was fixed almost immediately, the packaged cuda driver took three month+ before having a new rpm release.

0
0
0

@kernellogger The ZFS code isn't upstreamed because of licensing. What's stopping Nvidia from doing this?

1
0
0

@bitpirate

upstreaming would be doomed to fail for one reason alone:

the driver does not provide a stable userland interface.

I suspect there are more reasons (are they properly using the common infra bits these days or are they doing their own thing?)

0
0
0

@kernellogger While the OpenZFS release story wrt kernel versions could certainly stand some improvement, 2.2.5 does contain some fixes for 6.10, and in my experience it'll work with the next kernel version more often than not despite it not being officially supported.

1
0
0

@Lalufu

I know; YMMV, but in the end it remains a situation I would not feel comfortable with when it comes to storing my data.

0
0
0