Qt 4.8 endInsert/RemoveRows cause memory leak? - c++

Question

Recently I've made some stress tests on my Qt 4.8 application. I've seen using valgrind massif tool, that it cause heap memory expansion...
Using that tool I found out this memory is allocated with this stack trace (this is other massif call, so values vary from the one on screenshot):
->02.11% (1,133,952B) 0x221FD9EB: ??? (in /usr/lib/x86_64-linux-gnu/qt4/plugins/accessible/libqtaccessiblewidgets.so)
| ->02.11% (1,133,952B) 0x80ADE69: QAccessible::queryAccessibleInterface(QObject*) (in /usr/lib/x86_64-linux-gnu/libQtGui.so.4.8.7)
| ->01.18% (637,824B) 0x80B5156: ??? (in /usr/lib/x86_64-linux-gnu/libQtGui.so.4.8.7)
| | ->01.18% (637,824B) 0x8BA0F6E: QMetaObject::activate(QObject*, QMetaObject const*, int, void**) (in /usr/lib/x86_64-linux-gnu/libQtCore.so.4.8.7)
| | ->01.18% (637,824B) 0x8BF1472: QAbstractItemModel::rowsInserted(QModelIndex const&, int, int) (in /usr/lib/x86_64-linux-gnu/libQtCore.so.4.8.7)
| | ->01.18% (637,824B) 0x8B86510: QAbstractItemModel::endInsertRows() (in /usr/lib/x86_64-linux-gnu/libQtCore.so.4.8.7)
| | ->01.18% (637,824B) in 5 places, all below massif's threshold (1.00%)
The cause of expansions are endInsertRows and endRemoveRows functions. My ModelView implementation looks like this:
void TrainScheduleModelView::addTrain(const model::object::Train &train)
{
if (this->m_rows == TrainScheduleModelView::MAX_TRAIN_SCHEDULE_SIZE)
{
beginRemoveRows(QModelIndex(),
TrainScheduleModelView::MAX_TRAIN_SCHEDULE_SIZE - 1,
TrainScheduleModelView::MAX_TRAIN_SCHEDULE_SIZE - 1);
endRemoveRows();
}
beginInsertRows(QModelIndex(), 0, 0);
this->m_trains[this->m_head].second = train;
this->m_trains[this->m_head].first = true;
if (0 == this->m_head)
{
this->m_head = TrainScheduleModelView::MAX_TRAIN_SCHEDULE_SIZE - 1;
}
else
{
--(this->m_head);
}
if (this->m_rows < TrainScheduleModelView::MAX_TRAIN_SCHEDULE_SIZE)
{
++(this->m_rows);
}
endInsertRows();
}
This model is based on C-arrays intentionaly, every time the new object is added, last one is truncated.
Could anyone tell me is there a bug, or I'm using it wrong?

Answer 1

I had two attempts to solve it:
First of all I tried to eliminate Accesibility module - I hoped it is not necessary. I had to build Qt from source with ./configure ... -no-accessibility... option. The case was resolved, but UI of the application dramatically suffered. I was not able to enter password, because keyboard was blocked. So this solution was not acceptable.
Next workaround was elimination of dynamic row insertion. I've changed approach and instead of emulating row insertion, I return always fixed number of rows (empty rows were displayed, however there were no data inside). When I put some data inside the table I emit dataChanged signal with all rows affected (because each one was shifted). This solution eliminated using queryAccessibleInterface function and there were no leak. So the code looked like this:
void TrainScheduleModelView::addTrain(const model::object::Train &train)
{
this->m_trains[this->m_head].second = train;
this->m_trains[this->m_head].first = true;
if (0 == this->m_head)
{
this->m_head = TrainScheduleModelView::MAX_TRAIN_SCHEDULE_SIZE - 1;
}
else
{
--(this->m_head);
}
if (this->m_rows < TrainScheduleModelView::MAX_TRAIN_SCHEDULE_SIZE)
{
++(this->m_rows);
}
emit dataChanged(this->index(0, 0),
this->index(TrainScheduleModelView::MAX_TRAIN_SCHEDULE_SIZE-1,
TrainScheduleModelView::COLUMNS_AMOUNT - 1));
}
The case is not completely solved. I suppose there might be a problem with porting Qt on Ubuntu or init code of application.
UPDATE
I found leak source. queryAccessibleInterface function returns allocated with new pointer and code loose reference to it.
void QAbstractItemViewPrivate::_q_layoutChanged()
{
doDelayedItemsLayout();
#ifndef QT_NO_ACCESSIBILITY
#ifdef Q_WS_X11
Q_Q(QAbstractItemView);
if (QAccessible::isActive()) {
QAccessible::queryAccessibleInterface(q)->table2Interface()->modelReset();
QAccessible::updateAccessibility(q, 0, QAccessible::TableModelChanged);
}
#endif
#endif
}