linuxkpi: make FPU sections thread-safe and NOCTX

[val_packett.cool]

Summary

FPU_KERN_NOCTX should be enough to match the Linux semantics.

+ add explanation for the nesting counter.

Test Plan

Test with an AMD Navi GPU.

[kib]

sys/compat/linuxkpi/common/include/asm/fpu/api.h

36–37

This is still completely melts if more than one thread does kernel_fpu_begin()

36–37

Use ANSI C definitions.

36–37

Comment should follow style.

[• hselasky]

sys/compat/linuxkpi/common/include/asm/fpu/api.h

36–37

Feel free to extend the LinuxKPI's "current" with this variable.

[val_packett.cool]

sys/compat/linuxkpi/common/include/asm/fpu/api.h

36–37

Feel free to extend the LinuxKPI's "current" with this variable.

OH! Thanks, that makes sense now. Where was this comment when I posted the original patch ;)

36–37

Huh? The only non-ANSI thing I know of is "K&R style" local variables being declared before the opening { bracket. But these functions do not declare any locals at all…??

[val_packett.cool]

sys/compat/linuxkpi/common/include/asm/fpu/api.h

36–37

Oh.. you mean the whole "blank () args doesn't mean no args" thing I guess

[val_packett.cool]

No one is stepping up to test on affected hardware, so I guess this can just be committed already

[• hselasky]

Make sure you bump the __FreeBSD_version when you commit this.

--HPS

[val_packett.cool]

well, I can't exactly commit anything myself ;)

[• hselasky]

I'll try to help you. Just be patient :-)

[vishwin]

This broke at least i915kms from drm-kmod despite building successfully. I've reported there as well, since I'm not sure where any fixes need to go.

[• hselasky]

Should be fixed now.