Re: [PATCH 5/8] x86/mmu: Add mm-based PASID refcounting

[Fenghua Yu]

Hi, Thomas,

On Wed, Sep 29, 2021 at 09:51:15AM -0700, Luck, Tony wrote:
> > There is zero requirement to look at TIF_NEED_FPU_LOAD or
> > fpregs_state_valid() simply because the #GP comes straight from user
> > space which means the FPU registers contain the current tasks user space
> > state.
> 
> Just to double confirm ... there is no point in the #GP handler up to this point
> where pre-emption can occur?

Same question here. The fixup function is called after cond_local_irq_enable().
If an interrupt comes before fixup_pasid_exception(), the interrupt may
use FPU and call kernel_fpu_begin_mask()->set(TIF_NEED_FPU_LOAD)->
__cpu_invalidate_fpregs_state(). Then writing to the IA32_PASID MSR. When
exiting to user, the FPU states will be restored to the FPU regs including
the IA32_PASID MSR. So the MSR could be different from the value written in
fixup_pasid_execption(). Is it possible?

Or should fixup_pasid_exception() be called before cond_local_irq_enable()?

Thanks.

-Fenghua