OSD: avoid failure peer info to resent

  maybe_update_heartbeat_peers may remove one peer and never add in.
  if that peer is in failure_pending but that peer is not really has problem,
  failure_pending will hold that peer until it really down or add in again.

  once ms_handle_connect called, these pending failure will be resent again.

  2019-09-12 09:44:47.080933 7f1fad781700 10 osd.13 6175 ms_handle_connect on mon
  2019-09-12 09:44:47.080937 7f1fad781700 10 osd.13 6175 send_alive up_thru currently 6159 want 6155
  2019-09-12 09:44:47.080945 7f1fad781700 10 osd.13 6175 requeue_failures 0 + 1 -> 1

Signed-off-by: 宋顺10180185 <song.shun3@zte.com.cn>

diff --git a/src/osd/OSD.cc b/src/osd/OSD.cc
index d67002909e9ef25188a8be1e0e16bc0f34f334c6..f2b928b9468c1056eaece724e2a3932d1579c612 100644 (file)
--- a/src/osd/OSD.cc
+++ b/src/osd/OSD.cc
@@ -4482,6 +4482,16 @@ void OSD::maybe_update_heartbeat_peers()
   }
 
   dout(10) << "maybe_update_heartbeat_peers " << heartbeat_peers.size() << " peers, extras " << extras << dendl;
+
+  // clean up stale failure pending
+  for (auto it = failure_pending.begin(); it != failure_pending.end();) {
+    if (heartbeat_peers.count(it->first) == 0) {
+      send_still_alive(osdmap->get_epoch(), it->first, it->second.second);
+      failure_pending.erase(it++);
+    } else {
+      it++;
+    }
+  }
 }
 
 void OSD::reset_heartbeat_peers(bool all)