Fix a regression in the performance of named pipes by only making the
fd non-blocking when overlapped I/O is requested for the fd (reported
by Ivan Leo Puoti).
diff --git a/server/named_pipe.c b/server/named_pipe.c
index 27cc6c0..bbf011e 100644
--- a/server/named_pipe.c
+++ b/server/named_pipe.c
@@ -616,17 +616,26 @@
client = create_pipe_client( server, req->flags );
if( client )
{
- if( !socketpair( PF_UNIX, SOCK_STREAM, 0, fds ) &&
- (fcntl( fds[0], F_SETFL, O_NONBLOCK ) != -1) &&
- (fcntl( fds[1], F_SETFL, O_NONBLOCK ) != -1))
+ if (!socketpair( PF_UNIX, SOCK_STREAM, 0, fds ))
{
+ int res = 0;
+
assert( !client->fd );
assert( !server->fd );
+
+ /* for performance reasons, only set nonblocking mode when using
+ * overlapped I/O. Otherwise, we will be doing too much busy
+ * looping */
+ if (is_overlapped( req->flags ))
+ res = fcntl( fds[1], F_SETFL, O_NONBLOCK );
+ if ((res != -1) && is_overlapped( server->options ))
+ res = fcntl( fds[0], F_SETFL, O_NONBLOCK );
+
client->fd = create_anonymous_fd( &pipe_client_fd_ops,
fds[1], &client->obj );
server->fd = create_anonymous_fd( &pipe_server_fd_ops,
fds[0], &server->obj );
- if (client->fd && server->fd)
+ if (client->fd && server->fd && res != 1)
{
if( server->state == ps_wait_open )
async_terminate_head( &server->wait_q, STATUS_SUCCESS );
