shrink_dcache_sb speedup

This patch makes shrink_dcache_sb consistent with dentry pruning policy.

On the first pass we iterate over dentry unused list and prepare some
dentries for removal.

However, since the existing code moves evicted dentries to the beginning of
the LRU it can happen that fresh dentries from other superblocks will be
inserted *before* our dentries.

This can result in significant slowdown of shrink_dcache_sb().  Moreover,
for virtual filesystems like unionfs which can call dput() during dentries
kill existing code results in O(n^2) complexity.

We observed 2 minutes shrink_dcache_sb() with only 35000 dentries.

To avoid this effects we propose to isolate sb dentries at the end
of LRU list.

Signed-off-by: Denis V. Lunev <[email protected]>
Signed-off-by: Kirill Korotaev <[email protected]>
Signed-off-by: Andrey Mirkin <[email protected]>
Cc: Neil Brown <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>

fs/dcache.c

@@ -553,18 +553,18 @@ void shrink_dcache_sb(struct super_block * sb)
 	 * superblock to the most recent end of the unused list.
 	 */
 	spin_lock(&dcache_lock);
-	list_for_each_safe(tmp, next, &dentry_unused) {
+	list_for_each_prev_safe(tmp, next, &dentry_unused) {
 		dentry = list_entry(tmp, struct dentry, d_lru);
 		if (dentry->d_sb != sb)
 			continue;
-		list_move(tmp, &dentry_unused);
+		list_move_tail(tmp, &dentry_unused);
 	}
 
 	/*
 	 * Pass two ... free the dentries for this superblock.
 	 */
 repeat:
-	list_for_each_safe(tmp, next, &dentry_unused) {
+	list_for_each_prev_safe(tmp, next, &dentry_unused) {
 		dentry = list_entry(tmp, struct dentry, d_lru);
 		if (dentry->d_sb != sb)
 			continue;

include/linux/list.h

@@ -477,6 +477,18 @@ static inline void list_splice_init_rcu(struct list_head *list,
 	for (pos = (head)->next, n = pos->next; pos != (head); \
 		pos = n, n = pos->next)
 
+/**
+ * list_for_each_prev_safe - iterate over a list backwards safe against removal
+			of list entry
+ * @pos:	the &struct list_head to use as a loop cursor.
+ * @n:		another &struct list_head to use as temporary storage
+ * @head:	the head for your list.
+ */
+#define list_for_each_prev_safe(pos, n, head) \
+	for (pos = (head)->prev, n = pos->prev; \
+	     prefetch(pos->prev), pos != (head); \
+	     pos = n, n = pos->prev)
+
 /**
  * list_for_each_entry	-	iterate over list of given type
  * @pos:	the type * to use as a loop cursor.