Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs

Pull vfs updates from Al Viro:
 "All kinds of stuff this time around; some more notable parts:

   - RCU'd vfsmounts handling
   - new primitives for coredump handling
   - files_lock is gone
   - Bruce's delegations handling series
   - exportfs fixes

  plus misc stuff all over the place"

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: (101 commits)
  ecryptfs: ->f_op is never NULL
  locks: break delegations on any attribute modification
  locks: break delegations on link
  locks: break delegations on rename
  locks: helper functions for delegation breaking
  locks: break delegations on unlink
  namei: minor vfs_unlink cleanup
  locks: implement delegations
  locks: introduce new FL_DELEG lock flag
  vfs: take i_mutex on renamed file
  vfs: rename I_MUTEX_QUOTA now that it's not used for quotas
  vfs: don't use PARENT/CHILD lock classes for non-directories
  vfs: pull ext4's double-i_mutex-locking into common code
  exportfs: fix quadratic behavior in filehandle lookup
  exportfs: better variable name
  exportfs: move most of reconnect_path to helper function
  exportfs: eliminate unused "noprogress" counter
  exportfs: stop retrying once we race with rename/remove
  exportfs: clear DISCONNECTED on all parents sooner
  exportfs: more detailed comment for path_reconnect
  ...

1 files changed, 149 insertions, 118 deletions

diff --git a/fs/exportfs/expfs.c b/fs/exportfs/expfs.c
index a235f0016889..48a359dd286e 100644
--- a/fs/exportfs/expfs.c
+++ b/fs/exportfs/expfs.c
@@ -69,145 +69,162 @@ find_acceptable_alias(struct dentry *result,
 	return NULL;
 }
 
-/*
- * Find root of a disconnected subtree and return a reference to it.
- */
-static struct dentry *
-find_disconnected_root(struct dentry *dentry)
+static bool dentry_connected(struct dentry *dentry)
 {
 	dget(dentry);
-	while (!IS_ROOT(dentry)) {
+	while (dentry->d_flags & DCACHE_DISCONNECTED) {
 		struct dentry *parent = dget_parent(dentry);
 
-		if (!(parent->d_flags & DCACHE_DISCONNECTED)) {
+		dput(dentry);
+		if (IS_ROOT(dentry)) {
 			dput(parent);
-			break;
+			return false;
 		}
+		dentry = parent;
+	}
+	dput(dentry);
+	return true;
+}
+
+static void clear_disconnected(struct dentry *dentry)
+{
+	dget(dentry);
+	while (dentry->d_flags & DCACHE_DISCONNECTED) {
+		struct dentry *parent = dget_parent(dentry);
+
+		WARN_ON_ONCE(IS_ROOT(dentry));
+
+		spin_lock(&dentry->d_lock);
+		dentry->d_flags &= ~DCACHE_DISCONNECTED;
+		spin_unlock(&dentry->d_lock);
 
 		dput(dentry);
 		dentry = parent;
 	}
-	return dentry;
+	dput(dentry);
+}
+
+/*
+ * Reconnect a directory dentry with its parent.
+ *
+ * This can return a dentry, or NULL, or an error.
+ *
+ * In the first case the returned dentry is the parent of the given
+ * dentry, and may itself need to be reconnected to its parent.
+ *
+ * In the NULL case, a concurrent VFS operation has either renamed or
+ * removed this directory.  The concurrent operation has reconnected our
+ * dentry, so we no longer need to.
+ */
+static struct dentry *reconnect_one(struct vfsmount *mnt,
+		struct dentry *dentry, char *nbuf)
+{
+	struct dentry *parent;
+	struct dentry *tmp;
+	int err;
+
+	parent = ERR_PTR(-EACCES);
+	mutex_lock(&dentry->d_inode->i_mutex);
+	if (mnt->mnt_sb->s_export_op->get_parent)
+		parent = mnt->mnt_sb->s_export_op->get_parent(dentry);
+	mutex_unlock(&dentry->d_inode->i_mutex);
+
+	if (IS_ERR(parent)) {
+		dprintk("%s: get_parent of %ld failed, err %d\n",
+			__func__, dentry->d_inode->i_ino, PTR_ERR(parent));
+		return parent;
+	}
+
+	dprintk("%s: find name of %lu in %lu\n", __func__,
+		dentry->d_inode->i_ino, parent->d_inode->i_ino);
+	err = exportfs_get_name(mnt, parent, nbuf, dentry);
+	if (err == -ENOENT)
+		goto out_reconnected;
+	if (err)
+		goto out_err;
+	dprintk("%s: found name: %s\n", __func__, nbuf);
+	mutex_lock(&parent->d_inode->i_mutex);
+	tmp = lookup_one_len(nbuf, parent, strlen(nbuf));
+	mutex_unlock(&parent->d_inode->i_mutex);
+	if (IS_ERR(tmp)) {
+		dprintk("%s: lookup failed: %d\n", __func__, PTR_ERR(tmp));
+		goto out_err;
+	}
+	if (tmp != dentry) {
+		dput(tmp);
+		goto out_reconnected;
+	}
+	dput(tmp);
+	if (IS_ROOT(dentry)) {
+		err = -ESTALE;
+		goto out_err;
+	}
+	return parent;
+
+out_err:
+	dput(parent);
+	return ERR_PTR(err);
+out_reconnected:
+	dput(parent);
+	/*
+	 * Someone must have renamed our entry into another parent, in
+	 * which case it has been reconnected by the rename.
+	 *
+	 * Or someone removed it entirely, in which case filehandle
+	 * lookup will succeed but the directory is now IS_DEAD and
+	 * subsequent operations on it will fail.
+	 *
+	 * Alternatively, maybe there was no race at all, and the
+	 * filesystem is just corrupt and gave us a parent that doesn't
+	 * actually contain any entry pointing to this inode.  So,
+	 * double check that this worked and return -ESTALE if not:
+	 */
+	if (!dentry_connected(dentry))
+		return ERR_PTR(-ESTALE);
+	return NULL;
 }
 
 /*
  * Make sure target_dir is fully connected to the dentry tree.
  *
- * It may already be, as the flag isn't always updated when connection happens.
+ * On successful return, DCACHE_DISCONNECTED will be cleared on
+ * target_dir, and target_dir->d_parent->...->d_parent will reach the
+ * root of the filesystem.
+ *
+ * Whenever DCACHE_DISCONNECTED is unset, target_dir is fully connected.
+ * But the converse is not true: target_dir may have DCACHE_DISCONNECTED
+ * set but already be connected.  In that case we'll verify the
+ * connection to root and then clear the flag.
+ *
+ * Note that target_dir could be removed by a concurrent operation.  In
+ * that case reconnect_path may still succeed with target_dir fully
+ * connected, but further operations using the filehandle will fail when
+ * necessary (due to S_DEAD being set on the directory).
  */
 static int
 reconnect_path(struct vfsmount *mnt, struct dentry *target_dir, char *nbuf)
 {
-	int noprogress = 0;
-	int err = -ESTALE;
+	struct dentry *dentry, *parent;
 
-	/*
-	 * It is possible that a confused file system might not let us complete
-	 * the path to the root.  For example, if get_parent returns a directory
-	 * in which we cannot find a name for the child.  While this implies a
-	 * very sick filesystem we don't want it to cause knfsd to spin.  Hence
-	 * the noprogress counter.  If we go through the loop 10 times (2 is
-	 * probably enough) without getting anywhere, we just give up
-	 */
-	while (target_dir->d_flags & DCACHE_DISCONNECTED && noprogress++ < 10) {
-		struct dentry *pd = find_disconnected_root(target_dir);
-
-		if (!IS_ROOT(pd)) {
-			/* must have found a connected parent - great */
-			spin_lock(&pd->d_lock);
-			pd->d_flags &= ~DCACHE_DISCONNECTED;
-			spin_unlock(&pd->d_lock);
-			noprogress = 0;
-		} else if (pd == mnt->mnt_sb->s_root) {
-			printk(KERN_ERR "export: Eeek filesystem root is not connected, impossible\n");
-			spin_lock(&pd->d_lock);
-			pd->d_flags &= ~DCACHE_DISCONNECTED;
-			spin_unlock(&pd->d_lock);
-			noprogress = 0;
-		} else {
-			/*
-			 * We have hit the top of a disconnected path, try to
-			 * find parent and connect.
-			 *
-			 * Racing with some other process renaming a directory
-			 * isn't much of a problem here.  If someone renames
-			 * the directory, it will end up properly connected,
-			 * which is what we want
-			 *
-			 * Getting the parent can't be supported generically,
-			 * the locking is too icky.
-			 *
-			 * Instead we just return EACCES.  If server reboots
-			 * or inodes get flushed, you lose
-			 */
-			struct dentry *ppd = ERR_PTR(-EACCES);
-			struct dentry *npd;
-
-			mutex_lock(&pd->d_inode->i_mutex);
-			if (mnt->mnt_sb->s_export_op->get_parent)
-				ppd = mnt->mnt_sb->s_export_op->get_parent(pd);
-			mutex_unlock(&pd->d_inode->i_mutex);
-
-			if (IS_ERR(ppd)) {
-				err = PTR_ERR(ppd);
-				dprintk("%s: get_parent of %ld failed, err %d\n",
-					__func__, pd->d_inode->i_ino, err);
-				dput(pd);
-				break;
-			}
+	dentry = dget(target_dir);
 
-			dprintk("%s: find name of %lu in %lu\n", __func__,
-				pd->d_inode->i_ino, ppd->d_inode->i_ino);
-			err = exportfs_get_name(mnt, ppd, nbuf, pd);
-			if (err) {
-				dput(ppd);
-				dput(pd);
-				if (err == -ENOENT)
-					/* some race between get_parent and
-					 * get_name?  just try again
-					 */
-					continue;
-				break;
-			}
-			dprintk("%s: found name: %s\n", __func__, nbuf);
-			mutex_lock(&ppd->d_inode->i_mutex);
-			npd = lookup_one_len(nbuf, ppd, strlen(nbuf));
-			mutex_unlock(&ppd->d_inode->i_mutex);
-			if (IS_ERR(npd)) {
-				err = PTR_ERR(npd);
-				dprintk("%s: lookup failed: %d\n",
-					__func__, err);
-				dput(ppd);
-				dput(pd);
-				break;
-			}
-			/* we didn't really want npd, we really wanted
-			 * a side-effect of the lookup.
-			 * hopefully, npd == pd, though it isn't really
-			 * a problem if it isn't
-			 */
-			if (npd == pd)
-				noprogress = 0;
-			else
-				printk("%s: npd != pd\n", __func__);
-			dput(npd);
-			dput(ppd);
-			if (IS_ROOT(pd)) {
-				/* something went wrong, we have to give up */
-				dput(pd);
-				break;
-			}
-		}
-		dput(pd);
-	}
+	while (dentry->d_flags & DCACHE_DISCONNECTED) {
+		BUG_ON(dentry == mnt->mnt_sb->s_root);
 
-	if (target_dir->d_flags & DCACHE_DISCONNECTED) {
-		/* something went wrong - oh-well */
-		if (!err)
-			err = -ESTALE;
-		return err;
-	}
+		if (IS_ROOT(dentry))
+			parent = reconnect_one(mnt, dentry, nbuf);
+		else
+			parent = dget_parent(dentry);
 
+		if (!parent)
+			break;
+		dput(dentry);
+		if (IS_ERR(parent))
+			return PTR_ERR(parent);
+		dentry = parent;
+	}
+	dput(dentry);
+	clear_disconnected(target_dir);
 	return 0;
 }
 
@@ -215,7 +232,7 @@ struct getdents_callback {
 	struct dir_context ctx;
 	char *name;		/* name that was found. It already points to a
 				   buffer NAME_MAX+1 is size */
-	unsigned long ino;	/* the inum we are looking for */
+	u64 ino;		/* the inum we are looking for */
 	int found;		/* inode matched? */
 	int sequence;		/* sequence counter */
 };
@@ -255,10 +272,14 @@ static int get_name(const struct path *path, char *name, struct dentry *child)
 	struct inode *dir = path->dentry->d_inode;
 	int error;
 	struct file *file;
+	struct kstat stat;
+	struct path child_path = {
+		.mnt = path->mnt,
+		.dentry = child,
+	};
 	struct getdents_callback buffer = {
 		.ctx.actor = filldir_one,
 		.name = name,
-		.ino = child->d_inode->i_ino
 	};
 
 	error = -ENOTDIR;
@@ -268,6 +289,16 @@ static int get_name(const struct path *path, char *name, struct dentry *child)
 	if (!dir->i_fop)
 		goto out;
 	/*
+	 * inode->i_ino is unsigned long, kstat->ino is u64, so the
+	 * former would be insufficient on 32-bit hosts when the
+	 * filesystem supports 64-bit inode numbers.  So we need to
+	 * actually call ->getattr, not just read i_ino:
+	 */
+	error = vfs_getattr_nosec(&child_path, &stat);
+	if (error)
+		return error;
+	buffer.ino = stat.ino;
+	/*
 	 * Open the directory ...
 	 */
 	file = dentry_open(path, O_RDONLY, cred);