Re: [PATCH -mm -V3 03/21] mm, THP, swap: Support PMD swap mapping in swap_duplicate()

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

Hi,

The series up to and including this patch doesn't build.  For this patch we
need:

diff --git a/mm/swap_state.c b/mm/swap_state.c
index c6b3eab73fde..2f2d07627113 100644
--- a/mm/swap_state.c
+++ b/mm/swap_state.c
@@ -433,7 +433,7 @@ struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
                /*
                 * Swap entry may have been freed since our caller observed it.
                 */
-               err = swapcache_prepare(entry);
+               err = swapcache_prepare(entry, false);
                if (err == -EEXIST) {
                        radix_tree_preload_end();
                        /*


On Wed, May 23, 2018 at 04:26:07PM +0800, Huang, Ying wrote:
> @@ -3516,11 +3512,39 @@ static int __swap_duplicate(swp_entry_t entry, unsigned char usage)

Two comments about this part of __swap_duplicate as long as you're moving it to
another function:

   } else if (count || has_cache) {
   
   	if ((count & ~COUNT_CONTINUED) < SWAP_MAP_MAX)          /* #1   */
   		count += usage;
   	else if ((count & ~COUNT_CONTINUED) > SWAP_MAP_MAX)     /* #2   */
   		err = -EINVAL;

#1:  __swap_duplicate_locked might use

    VM_BUG_ON(usage != SWAP_HAS_CACHE && usage != 1);

to document the unstated assumption that usage is 1 (otherwise count could
overflow).

#2:  We've masked off SWAP_HAS_CACHE and COUNT_CONTINUED, and already checked
for SWAP_MAP_BAD, so I think condition #2 always fails and can just be removed.

> +#ifdef CONFIG_THP_SWAP
> +static int __swap_duplicate_cluster(swp_entry_t *entry, unsigned char usage)
...
> +	} else {
> +		for (i = 0; i < SWAPFILE_CLUSTER; i++) {
> +retry:
> +			err = __swap_duplicate_locked(si, offset + i, 1);

I guess usage is assumed to be 1 at this point (__swap_duplicate_locked makes
the same assumption).  Maybe make this explicit with

			err = __swap_duplicate_locked(si, offset + i, usage);

, use 'usage' in cluster_set_count and __swap_entry_free too, and then
earlier have a

       VM_BUG_ON(usage != SWAP_HAS_CACHE && usage != 1);

?

> +#else
> +static inline int __swap_duplicate_cluster(swp_entry_t *entry,

This doesn't need inline.


Not related to your changes, but while we're here, the comment with
SWAP_HAS_CONT in swap_count() could be deleted: I don't think there ever was a
SWAP_HAS_CONT.

The rest looks ok up to this point.