Re: [PATCH v3] p2m: use GNTTABOP_unmap_and_duplicate if available

[David Vrabel <david.vrabel@citrix.com>]

On 04/11/13 15:38, Roger Pau Monne wrote:
> The new GNTTABOP_unmap_and_duplicate operation doesn't zero the
> mapping passed in new_addr, allowing us to perform batch unmaps in p2m
> code without requiring the use of a multicall.

Should the grant maps should be batched in a similar way?

> +int m2p_remove_override_batch(struct page **pages,
> +		struct gnttab_map_grant_ref *kmap_ops, int count)
> +{
> +	struct gnttab_unmap_and_duplicate *unmap_ops = NULL;
> +	int ret = 0, i;
> +
> +	for (i = 0; i < count; i++) {
> +		ret = remove_page_override(pages[i]);
> +		if (ret)
> +			goto out;
> +	}
> +
> +	if (kmap_ops != NULL) {
> +		struct page *scratch_page = get_balloon_scratch_page();
> +		unsigned long scratch_page_address = (unsigned long)
> +					__va(page_to_pfn(scratch_page) << PAGE_SHIFT);
> +		int invcount = 0;
> +
> +		unmap_ops = kcalloc(count, sizeof(unmap_ops[0]), GFP_KERNEL);
> +		if (!unmap_ops) {
> +			put_balloon_scratch_page();
> +			ret = -ENOMEM;
> +			goto out;
> +		}

A memory allocation failure here looks bad as the grants will not be
unmapped which will have an impact on the granter domain.  Either: a)
the unmap ops should be constructed in advance (excluding high mem pages
at that point); or b) this should use a statically allocated per-cpu
array of unmap ops with the batch split as necessary to fit into the
static array.

David