Re: [PATCH v4 06/16] x86/virt/tdx: Improve PAMT refcounts allocation for sparse memory

[Nikolay Borisov <nik.borisov@suse.com>]

On 21.11.25 г. 2:51 ч., Rick Edgecombe wrote:
> From: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
> 
> init_pamt_metadata() allocates PAMT refcounts for all physical memory up
> to max_pfn. It might be suboptimal if the physical memory layout is
> discontinuous and has large holes.
> 
> The refcount allocation vmalloc allocation. This is necessary to support a

nit: Something's odd with the first sentence, perhaps an "is a" before 
is missing before "vmalloc"?

> large allocation size. The virtually contiguous property also makes it
> easy to find a specific 2MB range’s refcount since it can simply be
> indexed.
> 
> Since vmalloc mappings support remapping during normal kernel runtime,
> switch to an approach that only populates refcount pages for the vmalloc
> mapping when there is actually memory for that range. This means any holes
> in the physical address space won’t use actual physical memory.
> 
> The validity of this memory optimization is based on a couple assumptions:
> 1. Physical holes in the ram layout are commonly large enough for it to be
>     worth it.
> 2. An alternative approach that looks the refcounts via some more layered
>     data structure wouldn’t overly complicate the lookups. Or at least
>     more than the complexity of managing the vmalloc mapping.
> 
> Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> [Add feedback, update log]
> Signed-off-by: Rick Edgecombe <rick.p.edgecombe@intel.com>

<snip>

> ---
>   arch/x86/virt/vmx/tdx/tdx.c | 136 +++++++++++++++++++++++++++++++++---
>   1 file changed, 125 insertions(+), 11 deletions(-)
> 
> diff --git a/arch/x86/virt/vmx/tdx/tdx.c b/arch/x86/virt/vmx/tdx/tdx.c
> index c28d4d11736c..edf9182ed86d 100644
> --- a/arch/x86/virt/vmx/tdx/tdx.c
> +++ b/arch/x86/virt/vmx/tdx/tdx.c
> @@ -194,30 +194,135 @@ int tdx_cpu_enable(void)
>   }
>   EXPORT_SYMBOL_GPL(tdx_cpu_enable);
>   
> -/*
> - * Allocate PAMT reference counters for all physical memory.
> - *
> - * It consumes 2MiB for every 1TiB of physical memory.
> - */
> -static int init_pamt_metadata(void)
> +/* Find PAMT refcount for a given physical address */
> +static atomic_t *tdx_find_pamt_refcount(unsigned long pfn)
>   {
> -	size_t size = DIV_ROUND_UP(max_pfn, PTRS_PER_PTE) * sizeof(*pamt_refcounts);
> +	/* Find which PMD a PFN is in. */
> +	unsigned long index = pfn >> (PMD_SHIFT - PAGE_SHIFT);
>   
> -	if (!tdx_supports_dynamic_pamt(&tdx_sysinfo))
> -		return 0;
> +	return &pamt_refcounts[index];
> +}
>   
> -	pamt_refcounts = __vmalloc(size, GFP_KERNEL | __GFP_ZERO);
> -	if (!pamt_refcounts)
> +/* Map a page into the PAMT refcount vmalloc region */
> +static int pamt_refcount_populate(pte_t *pte, unsigned long addr, void *data)
> +{
> +	struct page *page;
> +	pte_t entry;
> +
> +	page = alloc_page(GFP_KERNEL | __GFP_ZERO);
> +	if (!page)
>   		return -ENOMEM;
>   
> +	entry = mk_pte(page, PAGE_KERNEL);
> +
> +	spin_lock(&init_mm.page_table_lock);
> +	/*
> +	 * PAMT refcount populations can overlap due to rounding of the
> +	 * start/end pfn. Make sure the PAMT range is only populated once.
> +	 */
> +	if (pte_none(ptep_get(pte)))
> +		set_pte_at(&init_mm, addr, pte, entry);
> +	else
> +		__free_page(page);
> +	spin_unlock(&init_mm.page_table_lock);

nit: Wouldn't it be better to perform the pte_none() check before doing 
the allocation thus avoiding needless allocations? I.e do the 
alloc/mk_pte only after we are 100% sure we are going to use this entry.

> +
>   	return 0;
>   }

    <snip>