mm/vmscan: do not try to reclaim page pinned by a GUP (get_user_page)

Page pinned by a get user as we won't be able to free them. This avoids
doing all the heavy work: scanning page mapping reference status, trying
to unmap the page, write it to disk (if file back and dirty or swap),
trying to release page (if file back) ...

Signed-off-by: Jérôme Glisse <jglisse@redhat.com>

1 files changed, 20 insertions, 3 deletions

diff --git a/mm/vmscan.c b/mm/vmscan.c
index 7e7d25504651..a4e4a4725afc 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -4138,15 +4138,32 @@ int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
  * Reasons page might not be evictable:
  * (1) page's mapping marked unevictable
  * (2) page is part of an mlocked VMA
- *
+ * (3) page is pinned by some GUP (get_user_pages())
  */
 int page_evictable(struct page *page)
 {
-	int ret;
+	struct address_space *mapping = page_mapping(page);
+	int ret, extra;
+
+	/*
+	 * If page has a mapping then it has an extra refcount and can another
+	 * extra one if it has its private fields set.
+	 */
+	extra = mapping ? 1 + page_has_private(page) : 0;
+	/* If page is isolated from lru then it has an extra refcount. */
+	extra += PageLRU(page) ? 0 : 1;
+	/*
+	 * GUP (get_user_pages()) will increment the refcount on the page so at
+	 * this point if the refcount is bigger than mapcount it means that the
+	 * page is pinned by some GUP and it is pointless to try to reclaim any
+	 * such page. Ignore any extra reference that are not from a GUP.
+	 */
+	if ((page_count(page) - extra) > page_mapcount(page))
+		return 0;
 
 	/* Prevent address_space of inode and swap cache from being freed */
 	rcu_read_lock();
-	ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
+	ret = !mapping_unevictable(mapping) && !PageMlocked(page);
 	rcu_read_unlock();
 	return ret;
 }