Merge pull request #9 from mattf/address-mkdirs-race

Address a potential race in File.mkdirs() with a double exists() check

Author: bradley childs

src/main/java/org/apache/hadoop/fs/glusterfs/GlusterFileSystem.java

@@ -312,10 +312,28 @@ public FSDataOutputStream create(Path path, FsPermission permission,
 
 		parent = path.getParent();
 		fParent = new File((makeAbsolute(parent)).toUri().getPath());
-		if ((parent != null) && (fParent != null) && (!fParent.exists()))
-			if (!fParent.mkdirs())
-				throw new IOException("cannot create parent directory: "
-						+ fParent.getPath());
+		if ((parent != null) && (fParent != null) && (!fParent.exists())) {
+			if (!fParent.mkdirs()) {
+				//
+				// File.mkdirs() is not multi-process safe. It is possible for
+				// a peer who is running mkdirs() to cause us to fail. In such
+				// a case, a rudimentary test is to try our exists() test for a
+				// second time. The isDirectory() protects us from exists()
+				// passing when a file is put in place of the directory we were
+				// trying to create. We can be fooled by a directory, or set of
+				// directories in the path, being owned by another user or with
+				// incompatible permissions.
+				//
+				// This could be slightly improved to retry the mkdirs(), which
+				// would cover races deep within the fParent's path. Each
+				// iteration will address one race.
+				//
+				if (!fParent.exists() || !fParent.isDirectory()) {
+					throw new IOException("cannot create parent directory: "
+							+ fParent.getPath());
+				}
+			}
+		}
 
 		glusterFileStream = new FSDataOutputStream(new GlusterFUSEOutputStream(
 				f.getPath(), false));