VFS: Remove dentry->d_lock locking from shrink_dcache_for_umount_subtree()
authorDavid Howells <dhowells@redhat.com>
Tue, 7 Jun 2011 13:09:20 +0000 (14:09 +0100)
committerAl Viro <viro@zeniv.linux.org.uk>
Mon, 1 Aug 2011 06:27:57 +0000 (02:27 -0400)
Locks of the dcache_lock were replaced by locks of dentry->d_lock in commits
such as:

2304450783dfde7b0b94ae234edd0dbffa865073
2fd6b7f50797f2e993eea59e0a0b8c6399c811dc

as part of the RCU-based pathwalk changes, despite the fact that the caller
(shrink_dcache_for_umount()) notes in the banner comment the reasons that
d_lock is not necessary in these functions:

/*
 * destroy the dentries attached to a superblock on unmounting
 * - we don't need to use dentry->d_lock because:
 *   - the superblock is detached from all mountings and open files, so the
 *     dentry trees will not be rearranged by the VFS
 *   - s_umount is write-locked, so the memory pressure shrinker will ignore
 *     any dentries belonging to this superblock that it comes across
 *   - the filesystem itself is no longer permitted to rearrange the dentries
 *     in this superblock
 */

So remove these locks.  If the locks are actually necessary, then this banner
comment should be altered instead.

The hash table chains are protected by 1-bit locks in the hash table heads, so
those shouldn't be a problem.

Note that to make this work, __d_drop() has to be split so that the RCUwalk
barrier can be avoided.  This causes problems otherwise as it has an assertion
that dentry->d_lock is locked - but there is no need for that as no one else
can be trying to access this dentry, except to step over it (and that should
be handled by d_free(), I think).

Signed-off-by: David Howells <dhowells@redhat.com>
Cc: Nick Piggin <npiggin@kernel.dk>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
fs/dcache.c

index 75590572ff7ad4930fe91d9fe1f61b50f30a8f5f..9df8b861e18e127f9fb5ce58df651731ad3ffb4c 100644 (file)
@@ -301,6 +301,27 @@ static struct dentry *d_kill(struct dentry *dentry, struct dentry *parent)
        return parent;
 }
 
+/*
+ * Unhash a dentry without inserting an RCU walk barrier or checking that
+ * dentry->d_lock is locked.  The caller must take care of that, if
+ * appropriate.
+ */
+static void __d_shrink(struct dentry *dentry)
+{
+       if (!d_unhashed(dentry)) {
+               struct hlist_bl_head *b;
+               if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
+                       b = &dentry->d_sb->s_anon;
+               else
+                       b = d_hash(dentry->d_parent, dentry->d_name.hash);
+
+               hlist_bl_lock(b);
+               __hlist_bl_del(&dentry->d_hash);
+               dentry->d_hash.pprev = NULL;
+               hlist_bl_unlock(b);
+       }
+}
+
 /**
  * d_drop - drop a dentry
  * @dentry: dentry to drop
@@ -319,17 +340,7 @@ static struct dentry *d_kill(struct dentry *dentry, struct dentry *parent)
 void __d_drop(struct dentry *dentry)
 {
        if (!d_unhashed(dentry)) {
-               struct hlist_bl_head *b;
-               if (unlikely(dentry->d_flags & DCACHE_DISCONNECTED))
-                       b = &dentry->d_sb->s_anon;
-               else
-                       b = d_hash(dentry->d_parent, dentry->d_name.hash);
-
-               hlist_bl_lock(b);
-               __hlist_bl_del(&dentry->d_hash);
-               dentry->d_hash.pprev = NULL;
-               hlist_bl_unlock(b);
-
+               __d_shrink(dentry);
                dentry_rcuwalk_barrier(dentry);
        }
 }
@@ -832,10 +843,8 @@ static void shrink_dcache_for_umount_subtree(struct dentry *dentry)
        BUG_ON(!IS_ROOT(dentry));
 
        /* detach this root from the system */
-       spin_lock(&dentry->d_lock);
        dentry_lru_del(dentry);
-       __d_drop(dentry);
-       spin_unlock(&dentry->d_lock);
+       __d_shrink(dentry);
 
        for (;;) {
                /* descend to the first leaf in the current subtree */
@@ -844,16 +853,11 @@ static void shrink_dcache_for_umount_subtree(struct dentry *dentry)
 
                        /* this is a branch with children - detach all of them
                         * from the system in one go */
-                       spin_lock(&dentry->d_lock);
                        list_for_each_entry(loop, &dentry->d_subdirs,
                                            d_u.d_child) {
-                               spin_lock_nested(&loop->d_lock,
-                                               DENTRY_D_LOCK_NESTED);
                                dentry_lru_del(loop);
-                               __d_drop(loop);
-                               spin_unlock(&loop->d_lock);
+                               __d_shrink(loop);
                        }
-                       spin_unlock(&dentry->d_lock);
 
                        /* move to the first child */
                        dentry = list_entry(dentry->d_subdirs.next,
@@ -885,10 +889,8 @@ static void shrink_dcache_for_umount_subtree(struct dentry *dentry)
                                list_del(&dentry->d_u.d_child);
                        } else {
                                parent = dentry->d_parent;
-                               spin_lock(&parent->d_lock);
                                parent->d_count--;
                                list_del(&dentry->d_u.d_child);
-                               spin_unlock(&parent->d_lock);
                        }
 
                        inode = dentry->d_inode;
@@ -935,9 +937,7 @@ void shrink_dcache_for_umount(struct super_block *sb)
 
        dentry = sb->s_root;
        sb->s_root = NULL;
-       spin_lock(&dentry->d_lock);
        dentry->d_count--;
-       spin_unlock(&dentry->d_lock);
        shrink_dcache_for_umount_subtree(dentry);
 
        while (!hlist_bl_empty(&sb->s_anon)) {