xfs: fix locking in xfs_iget_cache_hit

The locking in xfs_iget_cache_hit currently has numerous problems:

 - we clear the reclaim tag without i_flags_lock which protects
   modifications to it
 - we call inode_init_always which can sleep with pag_ici_lock
   held (this is oss.sgi.com BZ #819)
 - we acquire and drop i_flags_lock a lot and thus provide no
   consistency between the various flags we set/clear under it

This patch fixes all that with a major revamp of the locking in
the function.  The new version acquires i_flags_lock early and
only drops it once we need to call into inode_init_always or before
calling xfs_ilock.

This patch fixes a bug seen in the wild where we race modifying the
reclaim tag.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Felix Blyakher <felixb@sgi.com>
Reviewed-by: Eric Sandeen <sandeen@sandeen.net>
Signed-off-by: Felix Blyakher <felixb@sgi.com>

fs/xfs/linux-2.6/xfs_sync.c

@@ -708,6 +708,16 @@ xfs_reclaim_inode(
 	return 0;
 }
 
+void
+__xfs_inode_set_reclaim_tag(
+	struct xfs_perag	*pag,
+	struct xfs_inode	*ip)
+{
+	radix_tree_tag_set(&pag->pag_ici_root,
+			   XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
+			   XFS_ICI_RECLAIM_TAG);
+}
+
 /*
  * We set the inode flag atomically with the radix tree tag.
  * Once we get tag lookups on the radix tree, this inode flag
@@ -722,8 +732,7 @@ xfs_inode_set_reclaim_tag(
 
 	read_lock(&pag->pag_ici_lock);
 	spin_lock(&ip->i_flags_lock);
-	radix_tree_tag_set(&pag->pag_ici_root,
-			XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
+	__xfs_inode_set_reclaim_tag(pag, ip);
 	__xfs_iflags_set(ip, XFS_IRECLAIMABLE);
 	spin_unlock(&ip->i_flags_lock);
 	read_unlock(&pag->pag_ici_lock);

fs/xfs/linux-2.6/xfs_sync.h

@@ -48,6 +48,7 @@ int xfs_reclaim_inode(struct xfs_inode *ip, int locked, int sync_mode);
 int xfs_reclaim_inodes(struct xfs_mount *mp, int mode);
 
 void xfs_inode_set_reclaim_tag(struct xfs_inode *ip);
+void __xfs_inode_set_reclaim_tag(struct xfs_perag *pag, struct xfs_inode *ip);
 void xfs_inode_clear_reclaim_tag(struct xfs_inode *ip);
 void __xfs_inode_clear_reclaim_tag(struct xfs_mount *mp, struct xfs_perag *pag,
 				struct xfs_inode *ip);

fs/xfs/xfs_iget.c

@@ -191,79 +191,81 @@ xfs_iget_cache_hit(
 	int			flags,
 	int			lock_flags) __releases(pag->pag_ici_lock)
 {
+	struct inode		*inode = VFS_I(ip);
 	struct xfs_mount	*mp = ip->i_mount;
-	int			error = EAGAIN;
+	int			error;
+
+	spin_lock(&ip->i_flags_lock);
 
 	/*
-	 * If INEW is set this inode is being set up
-	 * If IRECLAIM is set this inode is being torn down
-	 * Pause and try again.
+	 * If we are racing with another cache hit that is currently
+	 * instantiating this inode or currently recycling it out of
+	 * reclaimabe state, wait for the initialisation to complete
+	 * before continuing.
+	 *
+	 * XXX(hch): eventually we should do something equivalent to
+	 *	     wait_on_inode to wait for these flags to be cleared
+	 *	     instead of polling for it.
 	 */
-	if (xfs_iflags_test(ip, (XFS_INEW|XFS_IRECLAIM))) {
+	if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
 		XFS_STATS_INC(xs_ig_frecycle);
+		error = EAGAIN;
 		goto out_error;
 	}
 
-	/* If IRECLAIMABLE is set, we've torn down the vfs inode part */
-	if (xfs_iflags_test(ip, XFS_IRECLAIMABLE)) {
-
 	/*
-		 * If lookup is racing with unlink, then we should return an
-		 * error immediately so we don't remove it from the reclaim
-		 * list and potentially leak the inode.
+	 * If lookup is racing with unlink return an error immediately.
 	 */
-		if ((ip->i_d.di_mode == 0) && !(flags & XFS_IGET_CREATE)) {
+	if (ip->i_d.di_mode == 0 && !(flags & XFS_IGET_CREATE)) {
 		error = ENOENT;
 		goto out_error;
 	}
 
+	/*
+	 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
+	 * Need to carefully get it back into useable state.
+	 */
+	if (ip->i_flags & XFS_IRECLAIMABLE) {
 		xfs_itrace_exit_tag(ip, "xfs_iget.alloc");
 
 		/*
-		 * We need to re-initialise the VFS inode as it has been
-		 * 'freed' by the VFS. Do this here so we can deal with
-		 * errors cleanly, then tag it so it can be set up correctly
-		 * later.
+		 * We need to set XFS_INEW atomically with clearing the
+		 * reclaimable tag so that we do have an indicator of the
+		 * inode still being initialized.
 		 */
-		if (inode_init_always(mp->m_super, VFS_I(ip))) {
-			error = ENOMEM;
-			goto out_error;
-		}
+		ip->i_flags |= XFS_INEW;
+		ip->i_flags &= ~XFS_IRECLAIMABLE;
+		__xfs_inode_clear_reclaim_tag(mp, pag, ip);
+
+		spin_unlock(&ip->i_flags_lock);
+		read_unlock(&pag->pag_ici_lock);
 
+		error = -inode_init_always(mp->m_super, inode);
+		if (error) {
 			/*
-		 * We must set the XFS_INEW flag before clearing the
-		 * XFS_IRECLAIMABLE flag so that if a racing lookup does
-		 * not find the XFS_IRECLAIMABLE above but has the igrab()
-		 * below succeed we can safely check XFS_INEW to detect
-		 * that this inode is still being initialised.
+			 * Re-initializing the inode failed, and we are in deep
+			 * trouble.  Try to re-add it to the reclaim list.
 			 */
-		xfs_iflags_set(ip, XFS_INEW);
-		xfs_iflags_clear(ip, XFS_IRECLAIMABLE);
+			read_lock(&pag->pag_ici_lock);
+			spin_lock(&ip->i_flags_lock);
 
-		/* clear the radix tree reclaim flag as well. */
-		__xfs_inode_clear_reclaim_tag(mp, pag, ip);
-	} else if (!igrab(VFS_I(ip))) {
-		/* If the VFS inode is being torn down, pause and try again. */
-		XFS_STATS_INC(xs_ig_frecycle);
+			ip->i_flags &= ~XFS_INEW;
+			ip->i_flags |= XFS_IRECLAIMABLE;
+			__xfs_inode_set_reclaim_tag(pag, ip);
 			goto out_error;
-	} else if (xfs_iflags_test(ip, XFS_INEW)) {
-		/*
-		 * We are racing with another cache hit that is
-		 * currently recycling this inode out of the XFS_IRECLAIMABLE
-		 * state. Wait for the initialisation to complete before
-		 * continuing.
-		 */
-		wait_on_inode(VFS_I(ip));
 		}
-
-	if (ip->i_d.di_mode == 0 && !(flags & XFS_IGET_CREATE)) {
-		error = ENOENT;
-		iput(VFS_I(ip));
+		inode->i_state = I_LOCK|I_NEW;
+	} else {
+		/* If the VFS inode is being torn down, pause and try again. */
+		if (!igrab(inode)) {
+			error = EAGAIN;
 			goto out_error;
 		}
 
 		/* We've got a live one. */
+		spin_unlock(&ip->i_flags_lock);
 		read_unlock(&pag->pag_ici_lock);
+	}
 
 	if (lock_flags != 0)
 		xfs_ilock(ip, lock_flags);
@@ -274,6 +276,7 @@ xfs_iget_cache_hit(
 	return 0;
 
 out_error:
+	spin_unlock(&ip->i_flags_lock);
 	read_unlock(&pag->pag_ici_lock);
 	return error;
 }