memleak fix: gst_tensor_meta_info_parse_memory usage

The input GstMemory given to gst_tensor_meta_info_parse_memory
should be deallocated (unref) by the caller.

There are a few elements that do not call unref it, which
often create a huge memory leak with converter/filter.

Leave a Doxygen note that the caller should unref it, too.

Signed-off-by: MyungJoo Ham <myungjoo.ham@samsung.com>

gst/nnstreamer/elements/gsttensor_converter.c

@@ -839,12 +839,12 @@ _gst_tensor_converter_chain_timestamp (GstTensorConverter * self,
 }
 
 /** @brief Chain function's private routine to process octet stream */
-static GstBuffer *
+static void
 _gst_tensor_converter_chain_octet (GstTensorConverter * self, GstBuffer * buf)
 {
-  GstBuffer *buffer = buf;
   GstTensorsInfo *_info = &self->tensors_config.info;
   gboolean multi = (_info->num_tensors > 1);
+  guint nummem = gst_buffer_n_memory (buf);
 
   /* configure multi tensors */
   if (multi || gst_buffer_n_memory (buf) > 1) {
@@ -855,42 +855,38 @@ _gst_tensor_converter_chain_octet (GstTensorConverter * self, GstBuffer * buf)
     g_assert (self->frames_per_tensor == 1);
 
     offset = 0;
-    buffer = gst_buffer_new ();
     mem = gst_buffer_get_all_memory (buf);
 
     if (multi) {
       for (i = 0; i < _info->num_tensors; ++i) {
         size = gst_tensors_info_get_size (_info, i);
-        gst_buffer_append_memory (buffer, gst_memory_share (mem, offset, size));
+        if (nummem > i) {
+          gst_buffer_replace_memory (buf, i,
+              gst_memory_share (mem, offset, size));
+        } else {
+          gst_buffer_append_memory (buf, gst_memory_share (mem, offset, size));
+        }
         offset += size;
       }
 
       gst_memory_unref (mem);
     } else {
-      gst_buffer_append_memory (buffer, mem);
+      gst_buffer_replace_all_memory (buf, mem);
     }
-
-    /* copy timestamps */
-    gst_buffer_copy_into (buffer, buf, GST_BUFFER_COPY_METADATA, 0, -1);
-    gst_buffer_unref (buf);
   }
-
-  return buffer;
 }
 
 /** @brief Chain function's private routine to process flex tensor */
-static GstBuffer *
+static void
 _gst_tensor_converter_chain_flex_tensor (GstTensorConverter * self,
     GstBuffer * buf)
 {
-  GstBuffer *buffer;
-  GstMemory *mem;
+  GstMemory *mem, *newmem;
   GstTensorsInfo *info;
   GstTensorMetaInfo meta;
   guint i;
 
   info = &self->tensors_config.info;
-  buffer = gst_buffer_new ();
 
   for (i = 0; i < info->num_tensors; i++) {
     gst_tensor_info_convert_to_meta (&info->info[i], &meta);
@@ -910,14 +906,11 @@ _gst_tensor_converter_chain_flex_tensor (GstTensorConverter * self,
     }
 
     mem = gst_buffer_peek_memory (buf, i);
-    mem = gst_tensor_meta_info_append_header (&meta, mem);
+    newmem = gst_tensor_meta_info_append_header (&meta, mem);
+    /** The old mem is unref'ed by gst_buffer_unref (buf) a few lines below */
 
-    gst_buffer_append_memory (buffer, mem);
+    gst_buffer_replace_memory (buf, i, newmem);
   }
-
-  gst_buffer_copy_into (buffer, buf, GST_BUFFER_COPY_METADATA, 0, -1);
-  gst_buffer_unref (buf);
-  return buffer;
 }
 
 /** @brief Chain function's private routine to push buffer into src pad */
@@ -926,15 +919,16 @@ _gst_tensor_converter_chain_push (GstTensorConverter * self, GstBuffer * buf)
 {
   GstBuffer *buffer = buf;
 
+
   if (self->in_media_type == _NNS_OCTET) {
     /* configure multi tensors */
-    buffer = _gst_tensor_converter_chain_octet (self, buffer);
+    _gst_tensor_converter_chain_octet (self, buffer);
   }
 
   /* if output is flexible, add header. */
   if (!self->do_not_append_header
       && gst_tensor_pad_caps_is_flexible (self->srcpad)) {
-    buffer = _gst_tensor_converter_chain_flex_tensor (self, buffer);
+    _gst_tensor_converter_chain_flex_tensor (self, buffer);
   }
 
   silent_debug_timestamp (self, buffer);

gst/nnstreamer/elements/gsttensor_mux.c

@@ -442,36 +442,36 @@ gst_tensor_mux_send_segment_event (GstTensorMux * tensor_mux,
 
 /**
  * @brief Process flex tensor.
+ * @param buf The buffer to be updated with flexible-tensor header info
  */
-static GstBuffer *
+static void
 gst_tensor_mux_chain_flex_tensor (GstTensorMux * tensor_mux, GstBuffer * buf)
 {
-  GstBuffer *buffer;
-  GstMemory *mem;
+  GstMemory *mem, *mem_temp;
   GstTensorsInfo *info;
   GstTensorMetaInfo meta;
   guint i;
 
   /* If input is flexible, do nothing. It is already flexible tensor. */
   if (gst_tensors_config_is_flexible (&tensor_mux->tensors_config))
-    return buf;
+    return;
 
   info = &tensor_mux->tensors_config.info;
-  buffer = gst_buffer_new ();
 
   for (i = 0; i < info->num_tensors; i++) {
     mem = gst_buffer_peek_memory (buf, i);
 
     /* append header */
     gst_tensor_info_convert_to_meta (&info->info[i], &meta);
-    mem = gst_tensor_meta_info_append_header (&meta, mem);
+    mem_temp = gst_tensor_meta_info_append_header (&meta, mem);
+    gst_buffer_replace_memory (buf, i, mem_temp);
 
-    gst_buffer_append_memory (buffer, mem);
   }
-
-  gst_buffer_copy_into (buffer, buf, GST_BUFFER_COPY_METADATA, 0, -1);
-  gst_buffer_unref (buf);
-  return buffer;
+  /**  Don't replace buf with another instance of GstBuffer
+    * because the original buf instance is in tensor_mux->collect_pad
+    * which becomes dangling after the replacement.
+    *  So, replace GstMemory of the GstBuffer instance (buf)!
+    */
 }
 
 /**
@@ -531,7 +531,7 @@ gst_tensor_mux_collected (GstCollectPads * pads, GstTensorMux * tensor_mux)
 
   /* add header if output is flexible */
   if (gst_tensor_pad_caps_is_flexible (tensor_mux->srcpad))
-    tensors_buf = gst_tensor_mux_chain_flex_tensor (tensor_mux, tensors_buf);
+    gst_tensor_mux_chain_flex_tensor (tensor_mux, tensors_buf);
 
   ret = gst_pad_push (tensor_mux->srcpad, tensors_buf);
   tensor_mux->need_set_time = TRUE;

gst/nnstreamer/elements/gsttensor_transform.c

@@ -1723,11 +1723,14 @@ gst_tensor_transform_transform (GstBaseTransform * trans,
 
     if (filter->apply && !g_list_find (filter->apply, GINT_TO_POINTER (i))) {
       GstMemory *mem = gst_buffer_peek_memory (inbuf, i);
+      GstMemory *mem_temp;
 
       if (!in_flexible && out_flexible) {
         /* append meta */
         gst_tensor_info_convert_to_meta (out_info, &meta);
-        mem = gst_tensor_meta_info_append_header (&meta, mem);
+        mem_temp = gst_tensor_meta_info_append_header (&meta, mem);
+        gst_memory_unref (mem);
+        mem = mem_temp;
       } else {
         mem = gst_memory_ref (mem);
       }

gst/nnstreamer/nnstreamer_plugin_api_impl.c

@@ -1495,6 +1495,7 @@ gst_tensor_meta_info_parse_memory (GstTensorMetaInfo * meta, GstMemory * mem)
  * @param[in] meta tensor meta structure
  * @param[in] mem pointer to GstMemory
  * @return Newly allocated GstMemory (Caller should free returned memory using gst_memory_unref())
+ * @note The caller has both original GstMemory (mem) and returned GstMemory objects. The caller should unref both.
  */
 GstMemory *
 gst_tensor_meta_info_append_header (GstTensorMetaInfo * meta, GstMemory * mem)

gst/nnstreamer/tensor_filter/tensor_filter.c

@@ -865,6 +865,7 @@ gst_tensor_filter_transform (GstBaseTransform * trans,
       }
 
       gst_buffer_append_memory (outbuf, mem);
+      gst_memory_unref (in_mem[i]);
     }
   }
 
@@ -903,6 +904,8 @@ gst_tensor_filter_transform (GstBaseTransform * trans,
           out_tensors[i].data, out_tensors[i].size, 0,
           out_tensors[i].size, out_tensors[i].data, g_free);
 
+      if (!allocate_in_invoke)
+        gst_allocator_free (out_mem[i]->allocator, out_mem[i]);
       out_mem[i] = gst_tensor_meta_info_append_header (&meta, flex_mem);
       gst_memory_unref (flex_mem);
     } else if (allocate_in_invoke) {