9p: fix ->rename_sem exclusion

9p wants to be able to build a path from given dentry to fs root and keep
it valid over a blocking operation.

->s_vfs_rename_mutex would be a natural candidate, but there are places
where we need that and where we have no way to tell if ->s_vfs_rename_mutex
is already held deeper in callchain.  Moreover, it's only held for
cross-directory renames; name changes within the same directory happen
without it.

Solution:
	* have d_move() done in ->rename() rather than in its caller
	* maintain a 9p-private rwsem (per-filesystem)
	* hold it exclusive over the relevant part of ->rename()
	* hold it shared over the places where we want the path.

That almost works.  FS_RENAME_DOES_D_MOVE is enough to put all d_move()
and d_exchange() calls under filesystem's control.  However, there's
also __d_unalias(), which isn't covered by any of that.

If ->lookup() hits a directory inode with preexisting dentry elsewhere
(due to e.g. rename done on server behind our back), d_splice_alias()
called by ->lookup() will move/rename that alias.

Add a couple of optional methods, so that __d_unalias() would do
	if alias->d_op->d_unalias_trylock != NULL
		if (!alias->d_op->d_unalias_trylock(alias))
			fail (resulting in -ESTALE from lookup)
	__d_move(...)
	if alias->d_op->d_unalias_unlock != NULL
		alias->d_unalias_unlock(alias)
where it currently does __d_move().  9p instances do down_write_trylock()
and up_write() of ->rename_mutex.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

2 files changed, 25 insertions, 0 deletions

diff --git a/Documentation/filesystems/locking.rst b/Documentation/filesystems/locking.rst
index 146e7d8aa736..d20a32b77b60 100644
--- a/Documentation/filesystems/locking.rst
+++ b/Documentation/filesystems/locking.rst
@@ -31,6 +31,8 @@ prototypes::
 	struct vfsmount *(*d_automount)(struct path *path);
 	int (*d_manage)(const struct path *, bool);
 	struct dentry *(*d_real)(struct dentry *, enum d_real_type type);
+	bool (*d_unalias_trylock)(const struct dentry *);
+	void (*d_unalias_unlock)(const struct dentry *);
 
 locking rules:
 
@@ -50,6 +52,8 @@ d_dname:	   no		no		no		no
 d_automount:	   no		no		yes		no
 d_manage:	   no		no		yes (ref-walk)	maybe
 d_real		   no		no		yes 		no
+d_unalias_trylock  yes		no		no 		no
+d_unalias_unlock   yes		no		no 		no
 ================== ===========	========	==============	========
 
 inode_operations
diff --git a/Documentation/filesystems/vfs.rst b/Documentation/filesystems/vfs.rst
index 7c352ebaae98..31eea688609a 100644
--- a/Documentation/filesystems/vfs.rst
+++ b/Documentation/filesystems/vfs.rst
@@ -1265,6 +1265,8 @@ defined:
 		struct vfsmount *(*d_automount)(struct path *);
 		int (*d_manage)(const struct path *, bool);
 		struct dentry *(*d_real)(struct dentry *, enum d_real_type type);
+		bool (*d_unalias_trylock)(const struct dentry *);
+		void (*d_unalias_unlock)(const struct dentry *);
 	};
 
 ``d_revalidate``
@@ -1428,6 +1430,25 @@ defined:
 
 	For non-regular files, the 'dentry' argument is returned.
 
+``d_unalias_trylock``
+	if present, will be called by d_splice_alias() before moving a
+	preexisting attached alias.  Returning false prevents __d_move(),
+	making d_splice_alias() fail with -ESTALE.
+
+	Rationale: setting FS_RENAME_DOES_D_MOVE will prevent d_move()
+	and d_exchange() calls from the outside of filesystem methods;
+	however, it does not guarantee that attached dentries won't
+	be renamed or moved by d_splice_alias() finding a preexisting
+	alias for a directory inode.  Normally we would not care;
+	however, something that wants to stabilize the entire path to
+	root over a blocking operation might need that.  See 9p for one
+	(and hopefully only) example.
+
+``d_unalias_unlock``
+	should be paired with ``d_unalias_trylock``; that one is called after
+	__d_move() call in __d_unalias().
+
+
 Each dentry has a pointer to its parent dentry, as well as a hash list
 of child dentries.  Child dentries are basically like files in a
 directory.