Re: [PATCH 2/2] x86/snp: Convert shared memory back to private on kexec

[Tom Lendacky]

On 2/22/24 04:50, Kirill A. Shutemov wrote:
> On Wed, Feb 21, 2024 at 02:35:13PM -0600, Tom Lendacky wrote:
> > > @@ -906,6 +917,206 @@ void snp_accept_memory(phys_addr_t start, phys_addr_t end)
> > >    	set_pages_state(vaddr, npages, SNP_PAGE_STATE_PRIVATE);
> > >    }
> > > +static inline bool pte_decrypted(pte_t pte)
> > > +{
> > > +	return cc_mkdec(pte_val(pte)) == pte_val(pte);
> > > +}
> > > +
> >
> > This is duplicated in TDX code, arch/x86/coco/tdx/tdx.c, looks like
> > something that can go in a header file, maybe mem_encrypt.h.
>
> I think <asm/pgtable.h> is a better fit.
>
> > > +void snp_kexec_stop_conversion(bool crash)
> > > +{
> > > +	/* Stop new private<->shared conversions */
> > > +	conversion_allowed = false;
> > > +	crash_requested = crash;
> > > +
> > > +	/*
> > > +	 * Make sure conversion_allowed is cleared before checking
> > > +	 * conversions_in_progress.
> > > +	 */
> > > +	barrier();
> >
> > This should be smp_wmb().
>
> Why?

IIUC, this is because conversions_in_progress can be set on another thread 
and so this needs an smp barrier. In this case, smp_wmb() just ends up 
being barrier(), but to me it is clearer this way. Just my opinion, though.

Thanks,
Tom


>