2 files changed, 30 insertions, 9 deletions

diff --git a/Documentation/filesystems/directory-locking b/Documentation/filesystems/directory-locking
index ff7b611abf33..09bbf9a54f80 100644
--- a/Documentation/filesystems/directory-locking
+++ b/Documentation/filesystems/directory-locking
@@ -2,6 +2,10 @@
 kinds of locks - per-inode (->i_mutex) and per-filesystem
 (->s_vfs_rename_mutex).
 
+        When taking the i_mutex on multiple non-directory objects, we
+always acquire the locks in order by increasing address.  We'll call
+that "inode pointer" order in the following.
+
         For our purposes all operations fall in 5 classes:
 
 1) read access.  Locking rules: caller locks directory we are accessing.
@@ -12,8 +16,9 @@ kinds of locks - per-inode (->i_mutex) and per-filesystem
 locks victim and calls the method.
 
 4) rename() that is _not_ cross-directory.  Locking rules: caller locks
-the parent, finds source and target, if target already exists - locks it
-and then calls the method.
+the parent and finds source and target.  If target already exists, lock
+it.  If source is a non-directory, lock it.  If that means we need to
+lock both, lock them in inode pointer order.
 
 5) link creation.  Locking rules:
         * lock parent
@@ -30,7 +35,9 @@ rules:
                 fail with -ENOTEMPTY
         * if new parent is equal to or is a descendent of source
                 fail with -ELOOP
-        * if target exists - lock it.
+        * If target exists, lock it.  If source is a non-directory, lock
+          it.  In case that means we need to lock both source and target,
+          do so in inode pointer order.
         * call the method.
 
 
@@ -56,9 +63,11 @@ objects - A < B iff A is an ancestor of B.
     renames will be blocked on filesystem lock and we don't start changing
     the order until we had acquired all locks).
 
-(3) any operation holds at most one lock on non-directory object and
-    that lock is acquired after all other locks.  (Proof: see descriptions
-    of operations).
+(3) locks on non-directory objects are acquired only after locks on
+    directory objects, and are acquired in inode pointer order.
+    (Proof: all operations but renames take lock on at most one
+    non-directory object, except renames, which take locks on source and
+    target in inode pointer order in the case they are not directories.)
 
         Now consider the minimal deadlock.  Each process is blocked on
 attempt to acquire some lock and already holds at least one lock.  Let's
@@ -66,9 +75,13 @@ consider the set of contended locks.  First of all, filesystem lock is
 not contended, since any process blocked on it is not holding any locks.
 Thus all processes are blocked on ->i_mutex.
 
-        Non-directory objects are not contended due to (3).  Thus link
-creation can't be a part of deadlock - it can't be blocked on source
-and it means that it doesn't hold any locks.
+        By (3), any process holding a non-directory lock can only be
+waiting on another non-directory lock with a larger address.  Therefore
+the process holding the "largest" such lock can always make progress, and
+non-directory objects are not included in the set of contended locks.
+
+        Thus link creation can't be a part of deadlock - it can't be
+blocked on source and it means that it doesn't hold any locks.
 
         Any contended object is either held by cross-directory rename or
 has a child that is also contended.  Indeed, suppose that it is held by
diff --git a/Documentation/filesystems/porting b/Documentation/filesystems/porting
index f0890581f7f6..fe2b7ae6f962 100644
--- a/Documentation/filesystems/porting
+++ b/Documentation/filesystems/porting
@@ -455,3 +455,11 @@ in your dentry operations instead.
         vfs_follow_link has been removed.  Filesystems must use nd_set_link
         from ->follow_link for normal symlinks, or nd_jump_link for magic
         /proc/<pid> style links.
+--
+[mandatory]
+        iget5_locked()/ilookup5()/ilookup5_nowait() test() callback used to be
+        called with both ->i_lock and inode_hash_lock held; the former is *not*
+        taken anymore, so verify that your callbacks do not rely on it (none
+        of the in-tree instances did).  inode_hash_lock is still held,
+        of course, so they are still serialized wrt removal from inode hash,
+        as well as wrt set() callback of iget5_locked().