Re: [PATCH 08/22] x86/pv: rewrite how building PV dom0 handles domheap mappings

[Jan Beulich <jbeulich@xxxxxxxx>]

On 16.12.2022 12:48, Julien Grall wrote:
> From: Hongyan Xia <hongyxia@xxxxxxxxxx>
> 
> Building a PV dom0 is allocating from the domheap but uses it like the
> xenheap. This is clearly wrong. Fix.

"Clearly wrong" would mean there's a bug here, at lest under certain
conditions. But there isn't: Even on huge systems, due to running on
idle page tables, all memory is mapped at present.

> @@ -711,22 +715,32 @@ int __init dom0_construct_pv(struct domain *d,
>          v->arch.pv.event_callback_cs    = FLAT_COMPAT_KERNEL_CS;
>      }
>  
> +#define UNMAP_MAP_AND_ADVANCE(mfn_var, virt_var, maddr) \
> +do {                                                    \
> +    UNMAP_DOMAIN_PAGE(virt_var);                        \

Not much point using the macro when ...

> +    mfn_var = maddr_to_mfn(maddr);                      \
> +    maddr += PAGE_SIZE;                                 \
> +    virt_var = map_domain_page(mfn_var);                \

... the variable gets reset again to non-NULL unconditionally right
away.

> +} while ( false )

This being a local macro and all use sites passing mpt_alloc as the
last argument, I think that parameter wants dropping, which would
improve readability.

> @@ -792,9 +808,9 @@ int __init dom0_construct_pv(struct domain *d,
>              if ( !l3e_get_intpte(*l3tab) )
>              {
>                  maddr_to_page(mpt_alloc)->u.inuse.type_info = 
> PGT_l2_page_table;
> -                l2tab = __va(mpt_alloc); mpt_alloc += PAGE_SIZE;
> -                clear_page(l2tab);
> -                *l3tab = l3e_from_paddr(__pa(l2tab), L3_PROT);
> +                UNMAP_MAP_AND_ADVANCE(l2start_mfn, l2start, mpt_alloc);
> +                clear_page(l2start);
> +                *l3tab = l3e_from_mfn(l2start_mfn, L3_PROT);

The l2start you map on the last iteration here can be re-used ...

> @@ -805,9 +821,17 @@ int __init dom0_construct_pv(struct domain *d,
>          unmap_domain_page(l2t);
>      }

... in the code the tail of which is visible here, eliminating a
redundant map/unmap pair.

> @@ -977,8 +1001,12 @@ int __init dom0_construct_pv(struct domain *d,
>       * !CONFIG_VIDEO case so the logic here can be simplified.
>       */
>      if ( pv_shim )
> +    {
> +        l4start = map_domain_page(l4start_mfn);
>          pv_shim_setup_dom(d, l4start, v_start, vxenstore_start, 
> vconsole_start,
>                            vphysmap_start, si);
> +        UNMAP_DOMAIN_PAGE(l4start);
> +    }

The, at the first glance, redundant re-mapping of the L4 table here could
do with explaining in the description. However, I further wonder in how
far in shim mode eliminating the direct map is actually useful. Which is
to say that I question the need for this change in the first place. Or
wait - isn't this (unlike the rest of this patch) actually a bug fix? At
this point we're on the domain's page tables, which may not cover the
page the L4 is allocated at (if a truly huge shim was configured). So I
guess the change is needed but wants breaking out, allowing to at least
consider whether to backport it.

Jan