Re: [PATCH 3/4] x86/PVH: improve Dom0 memory size calculation

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 30/08/2021 14:03, Jan Beulich wrote:
> Assuming that the accounting for IOMMU page tables will also take care
> of the P2M needs was wrong: dom0_paging_pages() can determine a far
> higher value, high enough for the system to run out of memory while
> setting up Dom0. Hence in the case of shared page tables the larger of
> the two values needs to be used (without shared page tables the sum of
> both continues to be applicable).

I'm afraid that I can't follow this at all.

What causes the system to run out of RAM while constructing dom0?

>
> While there also account for two further aspects in the PV case: With
> "iommu=dom0-passthrough" no IOMMU page tables would get allocated, so
> none need accounting for. And if shadow mode is to be enabled, setting
> aside a suitable amount for the P2M pool to get populated is also
> necessary (i.e. similar to the non-shared-page-tables case of PVH).
>
> Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
>
> --- a/xen/arch/x86/dom0_build.c
> +++ b/xen/arch/x86/dom0_build.c
> @@ -318,7 +318,7 @@ unsigned long __init dom0_compute_nr_pag
>      struct domain *d, struct elf_dom_parms *parms, unsigned long initrd_len)
>  {
>      nodeid_t node;
> -    unsigned long avail = 0, nr_pages, min_pages, max_pages;
> +    unsigned long avail = 0, nr_pages, min_pages, max_pages, iommu_pages = 0;
>      bool need_paging;
>  
>      /* The ordering of operands is to work around a clang5 issue. */
> @@ -337,18 +337,23 @@ unsigned long __init dom0_compute_nr_pag
>          avail -= d->max_vcpus - 1;
>  
>      /* Reserve memory for iommu_dom0_init() (rough estimate). */
> -    if ( is_iommu_enabled(d) )
> +    if ( is_iommu_enabled(d) && !iommu_hwdom_passthrough )
>      {
>          unsigned int s;
>  
>          for ( s = 9; s < BITS_PER_LONG; s += 9 )
> -            avail -= max_pdx >> s;
> +            iommu_pages += max_pdx >> s;
> +
> +        avail -= iommu_pages;
>      }
>  
> -    need_paging = is_hvm_domain(d) &&
> -        (!iommu_use_hap_pt(d) || !paging_mode_hap(d));
> +    need_paging = is_hvm_domain(d)
> +                  ? !iommu_use_hap_pt(d) || !paging_mode_hap(d)
> +                  : opt_dom0_shadow;

As per patch 4, this condition needs adjusting.

~Andrew

>      for ( ; ; need_paging = false )
>      {
> +        unsigned long paging_pages;
> +
>          nr_pages = get_memsize(&dom0_size, avail);
>          min_pages = get_memsize(&dom0_min_size, avail);
>          max_pages = get_memsize(&dom0_max_size, avail);
> @@ -377,11 +382,20 @@ unsigned long __init dom0_compute_nr_pag
>          nr_pages = min(nr_pages, max_pages);
>          nr_pages = min(nr_pages, avail);
>  
> -        if ( !need_paging )
> -            break;
> +        paging_pages = paging_mode_enabled(d) || need_paging
> +                       ? dom0_paging_pages(d, nr_pages) : 0;
>  
>          /* Reserve memory for shadow or HAP. */
> -        avail -= dom0_paging_pages(d, nr_pages);
> +        if ( !need_paging )
> +        {
> +            if ( paging_pages <= iommu_pages )
> +                break;
> +
> +            avail -= paging_pages - iommu_pages;
> +        }
> +        else
> +            avail -= paging_pages;
> +        iommu_pages = paging_pages;
>      }
>  
>      if ( is_pv_domain(d) &&
>