Re: [PATCH 5/5] x86_64/mm: map and unmap page tables in destroy_m2p_mapping

[Jan Beulich <jbeulich@xxxxxxxx>]

On 23.03.2020 10:41, Hongyan Xia wrote:
> @@ -297,26 +298,33 @@ static void destroy_m2p_mapping(struct mem_hotadd_info 
> *info)
>              continue;
>          }
>  
> -        l2_ro_mpt = l3e_to_l2e(l3_ro_mpt[l3_table_offset(va)]);
> +        l2_ro_mpt = map_l2t_from_l3e(l3_ro_mpt[l3_table_offset(va)]);
>          if (!(l2e_get_flags(l2_ro_mpt[l2_table_offset(va)]) & _PAGE_PRESENT))
>          {
>              i = ( i & ~((1UL << (L2_PAGETABLE_SHIFT - 3)) - 1)) +
>                      (1UL << (L2_PAGETABLE_SHIFT - 3)) ;
> +            UNMAP_DOMAIN_PAGE(l2_ro_mpt);
>              continue;
>          }
>  
>          pt_pfn = l2e_get_pfn(l2_ro_mpt[l2_table_offset(va)]);
>          if ( hotadd_mem_valid(pt_pfn, info) )
>          {
> +            l2_pgentry_t *l2t;
> +
>              destroy_xen_mappings(rwva, rwva + (1UL << L2_PAGETABLE_SHIFT));
>  
> -            l2_ro_mpt = l3e_to_l2e(l3_ro_mpt[l3_table_offset(va)]);
> -            l2e_write(&l2_ro_mpt[l2_table_offset(va)], l2e_empty());
> +            l2t = map_l2t_from_l3e(l3_ro_mpt[l3_table_offset(va)]);

Why a 2nd mapping of the same L3 entry that you've already mapped
into l2_ro_mpt?

> +            l2e_write(&l2t[l2_table_offset(va)], l2e_empty());
> +            UNMAP_DOMAIN_PAGE(l2t);

If this then weren't to go away, it should again be the lowercase
variant imo, as the variable's scope ends here.

Jan