Re: [PATCH v7 13/20] x86/virt/tdx: Allocate and set up PAMTs for TDMRs

[Dave Hansen <dave.hansen@intel.com>]

On 11/20/22 16:26, Kai Huang wrote:
> The TDX module uses additional metadata to record things like which
> guest "owns" a given page of memory.  This metadata, referred as
> Physical Address Metadata Table (PAMT), essentially serves as the
> 'struct page' for the TDX module.  PAMTs are not reserved by hardware
> up front.  They must be allocated by the kernel and then given to the
> TDX module.

... during module initialization.

> TDX supports 3 page sizes: 4K, 2M, and 1G.  Each "TD Memory Region"
> (TDMR) has 3 PAMTs to track the 3 supported page sizes.  Each PAMT must
> be a physically contiguous area from a Convertible Memory Region (CMR).
> However, the PAMTs which track pages in one TDMR do not need to reside
> within that TDMR but can be anywhere in CMRs.  If one PAMT overlaps with
> any TDMR, the overlapping part must be reported as a reserved area in
> that particular TDMR.
> 
> Use alloc_contig_pages() since PAMT must be a physically contiguous area
> and it may be potentially large (~1/256th of the size of the given TDMR).
> The downside is alloc_contig_pages() may fail at runtime.  One (bad)
> mitigation is to launch a TD guest early during system boot to get those
> PAMTs allocated at early time, but the only way to fix is to add a boot
> option to allocate or reserve PAMTs during kernel boot.

FWIW, we all agree that this is a bad permanent way to leave things.
You can call me out here as proposing that this wart be left in place
while this series is merged and is a detail we can work on afterword
with new module params, boot options, Kconfig or whatever.

> TDX only supports a limited number of reserved areas per TDMR to cover
> both PAMTs and memory holes within the given TDMR.  If many PAMTs are
> allocated within a single TDMR, the reserved areas may not be sufficient
> to cover all of them.
> 
> Adopt the following policies when allocating PAMTs for a given TDMR:
> 
>   - Allocate three PAMTs of the TDMR in one contiguous chunk to minimize
>     the total number of reserved areas consumed for PAMTs.
>   - Try to first allocate PAMT from the local node of the TDMR for better
>     NUMA locality.
> 
> Also dump out how many pages are allocated for PAMTs when the TDX module
> is initialized successfully.

... this helps answer the eternal "where did all my memory go?" questions.

> diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> index b36129183035..b86a333b860f 100644
> --- a/arch/x86/Kconfig
> +++ b/arch/x86/Kconfig
> @@ -1960,6 +1960,7 @@ config INTEL_TDX_HOST
>  	depends on KVM_INTEL
>  	depends on X86_X2APIC
>  	select ARCH_KEEP_MEMBLOCK
> +	depends on CONTIG_ALLOC
>  	help
>  	  Intel Trust Domain Extensions (TDX) protects guest VMs from malicious
>  	  host and certain physical attacks.  This option enables necessary TDX
> diff --git a/arch/x86/virt/vmx/tdx/tdx.c b/arch/x86/virt/vmx/tdx/tdx.c
> index 57b448de59a0..9d76e70de46e 100644
> --- a/arch/x86/virt/vmx/tdx/tdx.c
> +++ b/arch/x86/virt/vmx/tdx/tdx.c
> @@ -586,6 +586,187 @@ static int create_tdmrs(struct tdmr_info *tdmr_array, int *tdmr_num)
>  	return 0;
>  }
>  
> +/*
> + * Calculate PAMT size given a TDMR and a page size.  The returned
> + * PAMT size is always aligned up to 4K page boundary.
> + */
> +static unsigned long tdmr_get_pamt_sz(struct tdmr_info *tdmr, int pgsz)
> +{
> +	unsigned long pamt_sz, nr_pamt_entries;
> +
> +	switch (pgsz) {
> +	case TDX_PS_4K:
> +		nr_pamt_entries = tdmr->size >> PAGE_SHIFT;
> +		break;
> +	case TDX_PS_2M:
> +		nr_pamt_entries = tdmr->size >> PMD_SHIFT;
> +		break;
> +	case TDX_PS_1G:
> +		nr_pamt_entries = tdmr->size >> PUD_SHIFT;
> +		break;
> +	default:
> +		WARN_ON_ONCE(1);
> +		return 0;
> +	}
> +
> +	pamt_sz = nr_pamt_entries * tdx_sysinfo.pamt_entry_size;
> +	/* TDX requires PAMT size must be 4K aligned */
> +	pamt_sz = ALIGN(pamt_sz, PAGE_SIZE);
> +
> +	return pamt_sz;
> +}
> +
> +/*
> + * Pick a NUMA node on which to allocate this TDMR's metadata.
> + *
> + * This is imprecise since TDMRs are 1G aligned and NUMA nodes might
> + * not be.  If the TDMR covers more than one node, just use the _first_
> + * one.  This can lead to small areas of off-node metadata for some
> + * memory.
> + */
> +static int tdmr_get_nid(struct tdmr_info *tdmr)
> +{
> +	struct tdx_memblock *tmb;
> +
> +	/* Find the first memory region covered by the TDMR */
> +	list_for_each_entry(tmb, &tdx_memlist, list) {
> +		if (tmb->end_pfn > (tdmr_start(tdmr) >> PAGE_SHIFT))
> +			return tmb->nid;
> +	}

Aha, the first use of tmb->nid!  I wondered why that was there.

> +
> +	/*
> +	 * Fall back to allocating the TDMR's metadata from node 0 when
> +	 * no TDX memory block can be found.  This should never happen
> +	 * since TDMRs originate from TDX memory blocks.
> +	 */
> +	WARN_ON_ONCE(1);

That's probably better a pr_warn() or something.  A backtrace and all
that jazz seems a bit overly dramatic for this.

> +	return 0;
> +}
The rest of this actually looks fine.  It's nearing ack'able state.