FSGSBASE Testing Is Encouraged Ahead Of Linux 5.9

[phoronix]

Phoronix: FSGSBASE Testing Is Encouraged Ahead Of Linux 5.9

A few days ago I mentioned that it looked like the FSGSBASE patches could finally land for Linux 5.9 and indeed this performance-sensitive x86_64 feature is on track for premiering the next kernel cycle. But additional testing is encouraged...

http://www.phoronix.com/scan.php?page=news_item&px=FSGSBASE-Testing-Pre-5.9

[Space Heater]

This is the mailing list post the article is referencing.

[Michael]

This is the mailing list post the article is referencing.

Whoops, oversight on my part thought I had added it in there. It's in there now, thanks.

[kcrudup]

I'd tested it thoroughly back around spin 9 or 10 when you had to apply it manually via mbox patches and it worked well for many weeks (only stopped using it when a recent Linus' tip update caused merge conflicts). I'll E-mail him the findings of my success.

[Beherit]

Is it true this also speeds up AMD CPUs?

[milkylainen]

Just the sheer engineering keeping the old turd afloat...

I remember when cpuinfo features was a oneliner. Now it's like a novel...
... And f00f/fdiv was all you needed to know before Intel screwed everything up...

[Space Heater]

Is it true this also speeds up AMD CPUs?

Yes, but the improvement isn't likely to be as noticeable, as context switches are currently less expensive on AMD systems.

You can look at Michael's earlier benchmarks.

[_Alex_]

In the geometric mean, AMD got more performance (+7.3% vs intel 6.1%): https://openbenchmarking.org/embed.p...ha=432aff7&p=2

[damentz]

FSGSBASE patches seem to have issues with host systems running VMs through VirtualBox on Ryzen processors. I had to remove FSGSBASE patches from Zen Kernel after a report that a 3950x user was getting stalls and freezes after the patches were added: https://github.com/zen-kernel/zen-kernel/issues/172

[Space Heater]

FSGSBASE patches seem to have issues with host systems running VMs through VirtualBox on Ryzen processors.

Given that the bug report was posted before the current revision of the FSGSBASE patches came out, I'd suggest checking again to see if the problem still persists (and on an otherwise mainline kernel).