bufferpool: avoid excessive GstPoll activity

Keep an extra write ref on the control socket. This ensures that we
avoid a read/write on the socket when going from non-empty->empty->not-empty.
We remove the write ref only when we actually are empty and we need to
wait for flushing or a new buffer.

This makes the bufferpool benchmark about 30% faster than the pure
malloc implementation.

diff --git a/gst/gstbufferpool.c b/gst/gstbufferpool.c
index edd9ec9cf408311df32742825b5a943b430d9a50..6e11c72292111eca34661083eb5d37a0ca531a1a 100644 (file)
--- a/gst/gstbufferpool.c
+++ b/gst/gstbufferpool.c
@@ -178,6 +178,9 @@ gst_buffer_pool_init (GstBufferPool * pool)
   gst_allocation_params_init (&priv->params);
   gst_buffer_pool_config_set_allocator (priv->config, priv->allocator,
       &priv->params);
+  /* 1 control write for flushing */
+  gst_poll_write_control (priv->poll);
+  /* 1 control write for marking that we are not waiting for poll */
   gst_poll_write_control (priv->poll);
 
   GST_DEBUG_OBJECT (pool, "created");
@@ -1007,9 +1010,12 @@ default_acquire_buffer (GstBufferPool * pool, GstBuffer ** buffer,
       break;
     }
 
-    /* now wait */
-    GST_LOG_OBJECT (pool, "waiting for free buffers");
+    /* now we release the control socket, we wait for a buffer release or
+     * flushing */
+    gst_poll_read_control (pool->priv->poll);
+    GST_LOG_OBJECT (pool, "waiting for free buffers or flushing");
     gst_poll_wait (priv->poll, GST_CLOCK_TIME_NONE);
+    gst_poll_write_control (pool->priv->poll);
   }
 
   return result;