Re: [Xen-devel] [RFC XEN PATCH v3 05/39] x86/mm: exclude PMEM regions from initial frametable

[Haozhong Zhang <haozhong.zhang@xxxxxxxxx>]

On 11/03/17 13:58 +0800, Chao Peng wrote:
> 
> > +#ifdef CONFIG_NVDIMM_PMEM
> > +static void __init init_frametable_pmem_chunk(unsigned long s,
> > unsigned long e)
> > +{
> > +    static unsigned long pmem_init_frametable_mfn;
> > +
> > +    ASSERT(!((s | e) & (PAGE_SIZE - 1)));
> > +
> > +    if ( !pmem_init_frametable_mfn )
> > +    {
> > +        pmem_init_frametable_mfn = alloc_boot_pages(1, 1);
> > +        if ( !pmem_init_frametable_mfn )
> > +            panic("Not enough memory for pmem initial frame table
> > page");
> > +        memset(mfn_to_virt(pmem_init_frametable_mfn), -1, PAGE_SIZE);
> > +    }
> 
> Can zero_page be used instead?

No. I intend to make the frametable entries for NVDIMM as invalid at
boot time, in order to avoid/detect accidental accesses to NVDIMM
pages before they are registered to Xen hypervisor later (by part 2
patches 14 - 25).

> 
> > +
> > +    while ( s < e )
> > +    {
> > +        /*
> > +         * The real frame table entries of a pmem region will be
> > +         * created when the pmem region is registered to hypervisor.
> > +         * Any write attempt to the initial entries of that pmem
> > +         * region implies potential hypervisor bugs. In order to make
> > +         * those bugs explicit, map those initial entries as read-
> > only.
> > +         */
> > +        map_pages_to_xen(s, pmem_init_frametable_mfn, 1,
> > PAGE_HYPERVISOR_RO);
> > +        s += PAGE_SIZE;
> 
> Don't know how much the impact of 4K mapping on boot time when pmem is
> very large. Perhaps we need get such data on hardware.
>

Well, it will be very slow because the size of NVDIMM is usually very
large (e.g. from hundreds of giga-bytes to several tera-bytes). I can
make it to use huge page if possible.

> Another question is do we really need to map it, e.g. can we just skip
> the range here?

Sadly, I cannot remind why I did this. Maybe I can just leave the
frametable of NVDIMM unmapped and accidental access to them would just
trigger page fault in hypervisor, which can makes bugs explicit as well.


Haozhong

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel