Re: [PATCH -mm -v4 05/21] mm, THP, swap: Support PMD swap mapping in free_swap_and_cache()/swap_free()

[Daniel Jordan <daniel.m.jordan@oracle.com>]

On Fri, Jun 22, 2018 at 11:51:35AM +0800, Huang, Ying wrote:
> +static unsigned char swap_free_cluster(struct swap_info_struct *si,
> +				       swp_entry_t entry)
...
> +	/* Cluster has been split, free each swap entries in cluster */
> +	if (!cluster_is_huge(ci)) {
> +		unlock_cluster(ci);
> +		for (i = 0; i < SWAPFILE_CLUSTER; i++, entry.val++) {
> +			if (!__swap_entry_free(si, entry, 1)) {
> +				free_entries++;
> +				free_swap_slot(entry);
> +			}
> +		}

Is is better on average to use __swap_entry_free_locked instead of
__swap_entry_free here?  I'm not sure myself, just asking.

As it's written, if the cluster's been split, we always take and drop the
cluster lock 512 times, but if we don't expect to call free_swap_slot that
often, then we could just drop and retake the cluster lock inside the innermost
'if' against the possibility that free_swap_slot eventually makes us take the
cluster lock again.

...
> +		return !(free_entries == SWAPFILE_CLUSTER);

                return free_entries != SWAPFILE_CLUSTER;