Skip to contentSkip to conversation

xfsprogs: duplicate extent btrees in xfs_repair need locking

PATCH In review Dave Chinner 2 participants 2 messages Mar 11, 2010 – Mar 12, 2010

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xfsprogs: duplicate extent btrees in xfs_repair need locking

From: Dave Chinner <[email protected]>

The per-ag duplicate extent btrees can be search concurrently from multiple threads. This occurs when inode extent lists are being processed and inodes with extents in the same AG are checked concurrently. The btrees have an internal traversal cursor, so doing concurrent searches can result in the cursor being corrupted for both searches.

Add an external lock for each duplicate extent tree and use it for searches, inserts and deletes to ensure that we don't trash the state of any operation.

---
 repair/incore_ext.c |   33 +++++++++++++++++++++++++++------
 1 files changed, 27 insertions(+), 6 deletions(-)
repair/incore_ext.c +27 -6
index a362e5a..60dd4c4 100644
@@ -74,6 +74,7 @@ static rt_ext_flist_t rt_ext_flist;
 static avl64tree_desc_t	*rt_ext_tree_ptr;	/* dup extent tree for rt */
 
 static struct btree_root **dup_extent_trees;	/* per ag dup extent trees */
+static pthread_mutex_t *dup_extent_tree_locks;
 
 static avltree_desc_t	**extent_bno_ptrs;	/*
 						 * array of extent tree ptrs
@@ -108,7 +109,9 @@ void
 release_dup_extent_tree(
 	xfs_agnumber_t		agno)
 {
+	pthread_mutex_lock(&dup_extent_tree_locks[agno]);
 	btree_clear(dup_extent_trees[agno]);
+	pthread_mutex_unlock(&dup_extent_tree_locks[agno]);
 }
 
 int
@@ -117,12 +120,16 @@ add_dup_extent(
 	xfs_agblock_t		startblock,
 	xfs_extlen_t		blockcount)
 {
+	int	ret;
 #ifdef XR_DUP_TRACE
 	fprintf(stderr, "Adding dup extent - %d/%d %d\n", agno, startblock,
 		blockcount);
 #endif
-	return btree_insert(dup_extent_trees[agno], startblock,
+	pthread_mutex_lock(&dup_extent_tree_locks[agno]);
+	ret = btree_insert(dup_extent_trees[agno], startblock,
 				(void *)(uintptr_t)(startblock + blockcount));
+	pthread_mutex_unlock(&dup_extent_tree_locks[agno]);
+	return ret;
 }
 
 int
@@ -132,13 +139,22 @@ search_dup_extent(
 	xfs_agblock_t		end_agbno)
 {
 	unsigned long	bno;
+	int		ret;
 
-	if (!btree_find(dup_extent_trees[agno], start_agbno, &bno))
-		return 0;	/* this really shouldn't happen */
-	if (bno < end_agbno)
-		return 1;
-	return (uintptr_t)btree_peek_prev(dup_extent_trees[agno], NULL) >
+	pthread_mutex_lock(&dup_extent_tree_locks[agno]);
+	if (!btree_find(dup_extent_trees[agno], start_agbno, &bno)) {
+		ret = 0;
+		goto out;	/* this really shouldn't happen */
+	}
+	if (bno < end_agbno) {
+		ret = 1;
+		goto out;
+	}
+	ret = (uintptr_t)btree_peek_prev(dup_extent_trees[agno], NULL) >
 								start_agbno;
+out:
+	pthread_mutex_unlock(&dup_extent_tree_locks[agno]);
+	return ret;
 }
 
 
@@ -856,6 +872,10 @@ incore_ext_init(xfs_mount_t *mp)
 	if (!dup_extent_trees)
 		do_error(_("couldn't malloc dup extent tree descriptor table\n"));
 
+	dup_extent_tree_locks = calloc(agcount, sizeof(pthread_mutex_t));
+	if (!dup_extent_tree_locks)
+		do_error(_("couldn't malloc dup extent tree descriptor table\n"));
+
 	if ((extent_bno_ptrs = malloc(agcount *
 					sizeof(avltree_desc_t *))) == NULL)
 		do_error(
@@ -879,6 +899,7 @@ incore_ext_init(xfs_mount_t *mp)
 
 	for (i = 0; i < agcount; i++)  {
 		btree_init(&dup_extent_trees[i]);
+		pthread_mutex_init(&dup_extent_tree_locks[i], NULL);
 		avl_init_tree(extent_bno_ptrs[i], &avl_extent_tree_ops);
 		avl_init_tree(extent_bcnt_ptrs[i], &avl_extent_bcnt_tree_ops);
 	}
Signature · 6 lines
1.6.5

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs
Reviewed-by on v1 1/1
#1 lore raw
Quoted 8 lines · Dave Chinner

From: Dave Chinner <[email protected]>

The per-ag duplicate extent btrees can be search concurrently from multiple threads. This occurs when inode extent lists are being processed and inodes with extents in the same AG are checked concurrently. The btrees have an internal traversal cursor, so doing concurrent searches can result in the cursor being corrupted for both searches.

Quoted 4 lines

Add an external lock for each duplicate extent tree and use it for searches, inserts and deletes to ensure that we don't trash the state of any operation.

Note that all actual content modifications to the tree are done from single threaded code, so strictly speaking we'd only need to lock in search_dup_extent. Anyway, the locking shouldn't add enough overhead to worry about this, so:

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs