Per cluster-object inode chain locks

Index: 2.6.x-xfs-new/fs/xfs/xfs_iget.c
===================================================================
@@ -367,7 +367,9 @@ finish_inode:
 		if (chl->chl_blkno == ip->i_blkno) {
 
 			/* insert this inode into the doubly-linked list
-			 * where chl points */
+			 * where chl points. lock the chl to protect against
+			 * others traversing the chl list */
+			spin_lock(&chl->chl_lock);
 			if ((iq = chl->chl_ip)) {
 				ip->i_cprev = iq->i_cprev;
 				iq->i_cprev->i_cnext = ip;

@@ -379,6 +381,7 @@ finish_inode:
 			}
 			chl->chl_ip = ip;
 			ip->i_chash = chl;
+			spin_unlock(&chl->chl_lock);
 			break;
 		}
 	}

@@ -392,8 +395,11 @@ finish_inode:
 					kmem_zone_alloc(xfs_chashlist_zone,
 						KM_SLEEP);
 			ASSERT(chlnew != NULL);
+			spin_lock_init(&chlnew->chl_lock);
 			goto chlredo;
 		} else {
+			/* exclusive access to this chl thanks to the ch_lock
+			 * in write mode, so no lock really needed */
 			ip->i_cnext = ip;
 			ip->i_cprev = ip;
 			ip->i_chash = chlnew;

@@ -679,13 +685,21 @@ xfs_iextract(
 	 */
 	mp = ip->i_mount;
 	ch = XFS_CHASH(mp, ip->i_blkno);
-	s = mutex_spinlock(&ch->ch_lock);
-
-	if (ip->i_cnext == ip) {
+	spin_lock(&ip->i_chash->chl_lock);
+	if (unlikely(ip->i_cnext == ip)) {
 		/* Last inode on chashlist */
 		ASSERT(ip->i_cnext == ip && ip->i_cprev == ip);
 		ASSERT(ip->i_chash != NULL);
 		chm=NULL;
+
+		spin_unlock(&ip->i_chash->chl_lock);
+		spin_lock(&ch->ch_lock);
+		spin_lock(&ip->i_chash->chl_lock);
+		if (ip->i_cnext != ip) {
+			spin_unlock(&ch->ch_lock);
+			goto delete;
+		}
+		spin_unlock(&ip->i_chash->chl_lock);
 		chl = ip->i_chash;
 		if (chl->chl_prev)
 			chl->chl_prev->chl_next = chl->chl_next;

@@ -693,20 +707,22 @@ xfs_iextract(
 			ch->ch_list = chl->chl_next;
 		if (chl->chl_next)
 			chl->chl_next->chl_prev = chl->chl_prev;
+		spin_unlock(&ch->ch_lock);
 		kmem_zone_free(xfs_chashlist_zone, chl);
 	} else {
 		/* delete one inode from a non-empty list */
+delete:
 		iq = ip->i_cnext;
 		iq->i_cprev = ip->i_cprev;
 		ip->i_cprev->i_cnext = iq;
 		if (ip->i_chash->chl_ip == ip) {
 			ip->i_chash->chl_ip = iq;
 		}
+		spin_unlock(&ip->i_chash->chl_lock);
 		ip->i_chash = __return_address;
 		ip->i_cprev = __return_address;
 		ip->i_cnext = __return_address;
 	}
-	mutex_spinunlock(&ch->ch_lock, s);
 
 	/*
 	 * Remove from mount's inode list.

Index: 2.6.x-xfs-new/fs/xfs/xfs_inode.c
===================================================================
@@ -3004,7 +3004,6 @@ xfs_iflush(
 	xfs_mount_t		*mp;
 	int			error;
 	/* REFERENCED */
-	xfs_chash_t		*ch;
 	xfs_inode_t		*iq;
 	int			clcount;	/* count of inodes clustered */
 	int			bufwasdelwri;

@@ -3123,12 +3122,9 @@ xfs_iflush(
 	 * inode clustering:
 	 * see if other inodes can be gathered into this write
 	 */
-
+	spin_lock(&ip->i_chash->chl_lock);
 	ip->i_chash->chl_buf = bp;
 
-	ch = XFS_CHASH(mp, ip->i_blkno);
-	s = mutex_spinlock(&ch->ch_lock);
-
 	clcount = 0;
 	for (iq = ip->i_cnext; iq != ip; iq = iq->i_cnext) {
 		/*

@@ -3181,7 +3177,7 @@ xfs_iflush(
 			xfs_iunlock(iq, XFS_ILOCK_SHARED);
 		}
 	}
-	mutex_spinunlock(&ch->ch_lock, s);
+	spin_unlock(&ip->i_chash->chl_lock);
 
 	if (clcount) {
 		XFS_STATS_INC(xs_icluster_flushcnt);

@@ -3218,7 +3214,7 @@ cluster_corrupt_out:
 	/* Corruption detected in the clustering loop.  Invalidate the
 	 * inode buffer and shut down the filesystem.
 	 */
-	mutex_spinunlock(&ch->ch_lock, s);
+	spin_unlock(&ip->i_chash->chl_lock);
 
 	/*
 	 * Clean up the buffer.  If it was B_DELWRI, just release it --

Index: 2.6.x-xfs-new/fs/xfs/xfs_inode.h
===================================================================
@@ -194,6 +194,7 @@ typedef struct xfs_chashlist {
 	xfs_daddr_t		chl_blkno;	/* starting block number of
 						 * the cluster */
 	struct xfs_buf		*chl_buf;	/* the inode buffer */
+	lock_t			chl_lock;	/* inode list lock */
 } xfs_chashlist_t;
 
 typedef struct xfs_chash {