---

yaml
---
r: 34818
b: refs/heads/master
c: f551e44
h: refs/heads/master
v: v3

[refs]

@@ -1,2 +1,2 @@
 ---
-refs/heads/master: a53a3c58fd83e572a7c768d88b4c4e9840a57e82
+refs/heads/master: f551e44ff11d3e2ec8f37907bb88ec2433cc8b74

trunk/fs/nfs/inode.c

@@ -840,6 +840,12 @@ int nfs_refresh_inode(struct inode *inode, struct nfs_fattr *fattr)
  *
  * After an operation that has changed the inode metadata, mark the
  * attribute cache as being invalid, then try to update it.
+ *
+ * NB: if the server didn't return any post op attributes, this
+ * function will force the retrieval of attributes before the next
+ * NFS request.  Thus it should be used only for operations that
+ * are expected to change one or more attributes, to avoid
+ * unnecessary NFS requests and trips through nfs_update_inode().
  */
 int nfs_post_op_update_inode(struct inode *inode, struct nfs_fattr *fattr)
 {

trunk/fs/nfs/write.c

@@ -1253,7 +1253,13 @@ int nfs_writeback_done(struct rpc_task *task, struct nfs_write_data *data)
 	dprintk("NFS: %4d nfs_writeback_done (status %d)\n",
 		task->tk_pid, task->tk_status);
 
-	/* Call the NFS version-specific code */
+	/*
+	 * ->write_done will attempt to use post-op attributes to detect
+	 * conflicting writes by other clients.  A strict interpretation
+	 * of close-to-open would allow us to continue caching even if
+	 * another writer had changed the file, but some applications
+	 * depend on tighter cache coherency when writing.
+	 */
 	status = NFS_PROTO(data->inode)->write_done(task, data);
 	if (status != 0)
 		return status;