Re: [tip: x86/fpu] x86/fpu/xstate: Define new functions for clearing fpregs and xstates

[Ivan Zahariev]

Hello,

On 29.11.2022 г. 20:16, Dave Hansen wrote:

> On 11/29/22 03:19, Ivan Zahariev wrote:
> > Can you please take a look into this bug which syzbot tracked to a
> > commit of yours (b860eb8dce5906b14e3a7f3c771e0b3d6ef61b94). Even since
> > we switched from kernel 4.14 to 5.15 we are experiencing often random
> > segmentation faults with the following error in "dmesg":
> Which kernel are you running, exactly?  There is a fix for the commit
> that you identified:
>
> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=484cea4f362e
> but it should have been in 5.15.

We are running 5.15.75 (LTS) but the problem started when we upgraded 
from 5.15.31 to 5.15.59 and is present ever since. I erroneously said 
that it's present into every 5.15.

I didn't do my homework well and blamed the commit by Yu-cheng Yu. But 
this commit never landed into 5.15, nor the fix commit that you 
referred. There are no functions fpu__clear_all(), 
copy_init_fpstate_to_fpregs(), copy_user_to_xstate() anywhere in the 
sources of 5.15.75 or 5.15.31, so the 5.15 kernel is running with a 
different FPU implementation.

Additionally, I tested the reproducer on older 5.15 kernels and on the 
super-stable 4.14.256. They all emit the same "dmesg" error, so the 
reproducer is not reliable to detect our problem.

I am sorry for wasting your time.

> Is there a chance you could test current mainline and see if the issue
> is still there?

That's our only option, it seems. Thank you.

Best regards.
--Ivan