Re: [PATCH, RFC 18/62] x86/mm: Implement syncing per-KeyID direct mappings

[Peter Zijlstra]

On Mon, Jun 17, 2019 at 05:43:28PM +0300, Kirill A. Shutemov wrote:
> On Mon, Jun 17, 2019 at 11:27:55AM +0200, Peter Zijlstra wrote:

> > > > And yet I don't see anything in pageattr.c.
> > > 
> > > You're right. I've hooked up the sync in the wrong place.

> I think something like this should do (I'll fold it in after testing):

> @@ -643,7 +641,7 @@ static int sync_direct_mapping_keyid(unsigned long keyid)
>   *
>   * The function is nop until MKTME is enabled.
>   */
> -int sync_direct_mapping(void)
> +int sync_direct_mapping(unsigned long start, unsigned long end)
>  {
>  	int i, ret = 0;
>  
> @@ -651,7 +649,7 @@ int sync_direct_mapping(void)
>  		return 0;
>  
>  	for (i = 1; !ret && i <= mktme_nr_keyids; i++)
> -		ret = sync_direct_mapping_keyid(i);
> +		ret = sync_direct_mapping_keyid(i, start, end);
>  
>  	flush_tlb_all();
>  
> diff --git a/arch/x86/mm/pageattr.c b/arch/x86/mm/pageattr.c
> index 6a9a77a403c9..eafbe0d8c44f 100644
> --- a/arch/x86/mm/pageattr.c
> +++ b/arch/x86/mm/pageattr.c
> @@ -347,6 +347,28 @@ static void cpa_flush(struct cpa_data *data, int cache)
>  
>  	BUG_ON(irqs_disabled() && !early_boot_irqs_disabled);
>  
> +	if (mktme_enabled()) {
> +		unsigned long start, end;
> +
> +		start = *cpa->vaddr;
> +		end = *cpa->vaddr + cpa->numpages * PAGE_SIZE;
> +
> +		/* Sync all direct mapping for an array */
> +		if (cpa->flags & CPA_ARRAY) {
> +			start = PAGE_OFFSET;
> +			end = PAGE_OFFSET + direct_mapping_size;
> +		}

Understandable but sad, IIRC that's the most used interface (at least,
its the one the graphics people use).

> +
> +		/*
> +		 * Sync per-KeyID direct mappings with the canonical one
> +		 * (KeyID-0).
> +		 *
> +		 * sync_direct_mapping() does full TLB flush.
> +		 */
> +		sync_direct_mapping(start, end);
> +		return;

But it doesn't flush cache. So you can't return here.

> +	}
> +
>  	if (cache && !static_cpu_has(X86_FEATURE_CLFLUSH)) {
>  		cpa_flush_all(cache);
>  		return;
> -- 
>  Kirill A. Shutemov