Re: next-20090220: XFS: inconsistent lock state

[Christoph Hellwig <hch@infradead.org>]

On Fri, Feb 20, 2009 at 08:52:59PM +0300, Alexander Beregalov wrote:
> Hi
> 
> [ INFO: inconsistent lock state ]
> 2.6.29-rc5-next-20090220 #2
> ---------------------------------
> inconsistent {RECLAIM_FS-ON-W} -> {IN-RECLAIM_FS-R} usage.
> kswapd0/324 [HC0[0]:SC0[0]:HE1:SE1] takes:
>  (&(&ip->i_lock)->mr_lock){+++++?}, at: [<ffffffff803ca60a>]
> xfs_ilock+0xaa/0x120
> {RECLAIM_FS-ON-W} state was registered at:

That's a false positive.  While the ilock can be taken in reclaim the
allocation here is done before the inode is added to the inode cache.

The patch below should help avoiding the warning:


Index: xfs/fs/xfs/xfs_iget.c
===================================================================
--- xfs.orig/fs/xfs/xfs_iget.c	2009-02-24 20:56:00.716027739 +0100
+++ xfs/fs/xfs/xfs_iget.c	2009-02-24 20:56:46.089031360 +0100
@@ -246,9 +246,6 @@ xfs_iget_cache_miss(
 		goto out_destroy;
 	}
 
-	if (lock_flags)
-		xfs_ilock(ip, lock_flags);
-
 	/*
 	 * Preload the radix tree so we can insert safely under the
 	 * write spinlock. Note that we cannot sleep inside the preload
@@ -259,6 +256,15 @@ xfs_iget_cache_miss(
 		goto out_unlock;
 	}
 
+	/*
+	 * Because the inode hasn't been added to the radix-tree yet it can't
+	 * be found by another thread, so we can do the non-sleeping lock here.
+	 */
+	if (lock_flags) {
+		if (!xfs_ilock_nowait(ip, lock_flags))
+			BUG();
+	}
+
 	mask = ~(((XFS_INODE_CLUSTER_SIZE(mp) >> mp->m_sb.sb_inodelog)) - 1);
 	first_index = agino & mask;
 	write_lock(&pag->pag_ici_lock);

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs