rgw: fix RGWSyncTraceNode crash in reload

reload or restart will finalize RGWSyncTraceManager and destroy
complete_nodes, which will trigger RGWSyncTraceNode release twice.

A situation is sync more than 4096 objects and get full complete_nodes,
complete_nodes have a member A who's parent is P which is not finished
and A is the last child of P. In destroy complete_nodes, nodes will be
released one by one.
Release A will trigger release P, and call finish_node(P). there is
trap that circular_buffer won't change size until destroy all items,
so circular_buffer is still full and pop the first item which is A,
and released A again.

fixes: http://tracker.ceph.com/issues/24432

Signed-off-by: Tianshan Qu <tianshan@xsky.com>
(cherry picked from commit 3536c7810ae1c1ed508d4eb62b08b7da0c098c07)

diff --git a/src/rgw/rgw_sync_trace.cc b/src/rgw/rgw_sync_trace.cc
index 38219f9cd6d0c824f138730d68cddd8456daf29f..3015f9c114d7636ba65c10d286fec4598086e73b 100644 (file)
--- a/src/rgw/rgw_sync_trace.cc
+++ b/src/rgw/rgw_sync_trace.cc
@@ -134,6 +134,8 @@ RGWSyncTraceManager::~RGWSyncTraceManager()
 
   service_map_thread->stop();
   delete service_map_thread;
+
+  nodes.clear();
 }
 
 int RGWSyncTraceManager::hook_to_admin_command()