Btrfs: fix race during transaction joins
authorChris Mason <chris.mason@oracle.com>
Sun, 6 Nov 2011 08:26:19 +0000 (03:26 -0500)
committerChris Mason <chris.mason@oracle.com>
Sun, 6 Nov 2011 08:26:19 +0000 (03:26 -0500)
While we're allocating ram for a new transaction, we drop our spinlock.
When we get the lock back, we do check to see if a transaction started
while we slept, but we don't check to make sure it isn't blocked
because a commit has already started.

Signed-off-by: Chris Mason <chris.mason@oracle.com>
fs/btrfs/transaction.c

index 29f782cc2cc93af4ad09dd9444de96e72dbdd690..960835eaf4da7acfd4b9f3f11dd0fb16f811b007 100644 (file)
@@ -55,6 +55,7 @@ static noinline int join_transaction(struct btrfs_root *root, int nofail)
        struct btrfs_transaction *cur_trans;
 
        spin_lock(&root->fs_info->trans_lock);
+loop:
        if (root->fs_info->trans_no_join) {
                if (!nofail) {
                        spin_unlock(&root->fs_info->trans_lock);
@@ -75,16 +76,18 @@ static noinline int join_transaction(struct btrfs_root *root, int nofail)
        cur_trans = kmem_cache_alloc(btrfs_transaction_cachep, GFP_NOFS);
        if (!cur_trans)
                return -ENOMEM;
+
        spin_lock(&root->fs_info->trans_lock);
        if (root->fs_info->running_transaction) {
+               /*
+                * someone started a transaction after we unlocked.  Make sure
+                * to redo the trans_no_join checks above
+                */
                kmem_cache_free(btrfs_transaction_cachep, cur_trans);
                cur_trans = root->fs_info->running_transaction;
-               atomic_inc(&cur_trans->use_count);
-               atomic_inc(&cur_trans->num_writers);
-               cur_trans->num_joined++;
-               spin_unlock(&root->fs_info->trans_lock);
-               return 0;
+               goto loop;
        }
+
        atomic_set(&cur_trans->num_writers, 1);
        cur_trans->num_joined = 0;
        init_waitqueue_head(&cur_trans->writer_wait);