NFSv4: nfs_client_return_marked_delegations can't flush data
authorTrond Myklebust <Trond.Myklebust@netapp.com>
Sun, 6 May 2012 23:46:30 +0000 (19:46 -0400)
committerTrond Myklebust <Trond.Myklebust@netapp.com>
Tue, 8 May 2012 16:53:21 +0000 (12:53 -0400)
Since even filemap_flush() needs to lock pages that are dirty, we
cannot risk calling it from the state manager context. Therefore,
we need to move the call to filemap_flush() to
nfs_async_inode_return_delegation().

Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
fs/nfs/delegation.c

index a19cb5ad6b1396a0be667f5876c9773923e1dbc6..bd3a9601d32d9915a70e1888ab5a710671e00aac 100644 (file)
@@ -316,6 +316,10 @@ out:
  * nfs_client_return_marked_delegations - return previously marked delegations
  * @clp: nfs_client to process
  *
+ * Note that this function is designed to be called by the state
+ * manager thread. For this reason, it cannot flush the dirty data,
+ * since that could deadlock in case of a state recovery error.
+ *
  * Returns zero on success, or a negative errno value.
  */
 int nfs_client_return_marked_delegations(struct nfs_client *clp)
@@ -340,11 +344,9 @@ restart:
                                                                server);
                        rcu_read_unlock();
 
-                       if (delegation != NULL) {
-                               filemap_flush(inode->i_mapping);
+                       if (delegation != NULL)
                                err = __nfs_inode_return_delegation(inode,
                                                                delegation, 0);
-                       }
                        iput(inode);
                        if (!err)
                                goto restart;
@@ -542,6 +544,8 @@ int nfs_async_inode_return_delegation(struct inode *inode,
        struct nfs_client *clp = server->nfs_client;
        struct nfs_delegation *delegation;
 
+       filemap_flush(inode->i_mapping);
+
        rcu_read_lock();
        delegation = rcu_dereference(NFS_I(inode)->delegation);