PTI build regression with nvidia drivers

PTI build regression with nvidia drivers

[Kees Cook]

Hi,

This was pointed out in a few places, but not forwarded to lkml yet that I saw:

https://devtalk.nvidia.com/default/topic/1028222/linux/lts-kernel-patch-for-intel-cpu-vulnerability-breaks-nvidia-driver/post/5230546

Before and after PTI, cpu_tlbstate is a GPL export:

$ git show v4.14:arch/x86/mm/init.c | grep cpu_tlbstate
DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate) = {
EXPORT_SYMBOL_GPL(cpu_tlbstate);

But after PTI, inlining or something is dragging cpu_tlbstate into the
open, causing build failures.

Technically, to avoid regressions for that module, we'll need to drop
the GPL marking on that symbol, or find some other solution...

-Kees

-- 
Kees Cook
Pixel Security

Re: PTI build regression with nvidia drivers

[Thomas Gleixner]

On Thu, 4 Jan 2018, Kees Cook wrote:
> 
> This was pointed out in a few places, but not forwarded to lkml yet that I saw:
> 
> https://devtalk.nvidia.com/default/topic/1028222/linux/lts-kernel-patch-for-intel-cpu-vulnerability-breaks-nvidia-driver/post/5230546
> 
> Before and after PTI, cpu_tlbstate is a GPL export:
> 
> $ git show v4.14:arch/x86/mm/init.c | grep cpu_tlbstate
> DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate) = {
> EXPORT_SYMBOL_GPL(cpu_tlbstate);
> 
> But after PTI, inlining or something is dragging cpu_tlbstate into the
> open, causing build failures.
> 
> Technically, to avoid regressions for that module, we'll need to drop
> the GPL marking on that symbol, or find some other solution...

Correct. It's indirect via the tlbflush inlines. We'll fix it as usual. Sigh....

Thanks,

	tglx