From: xie xingguo <xie.xingguo@zte.com.cn>
Date: Mon, 22 Feb 2016 10:05:39 +0000 (+0800)
Subject: OSD: fix race condition for heartbeat_need_update
X-Git-Tag: v10.1.0~263^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=0914246383f8bfefb15e05a49c67b8b3091d50cf;p=ceph.git

OSD: fix race condition for heartbeat_need_update

The heartbeat_need_update member can be accessed by both OSD tick thread and
PG relevant threads, that is why the heartbeat_update_lock mutex is introduced
to protect against its change. However, in the tick thread we may still be able
to reset heartbeat_need_update to false whithout holding heartbeat_update_lock
in hand, which shall be considered as a race conditon.

This pr solves the above problem by add a new API to clear heartbeat_need_update
atomically, which fix the above potential race condition.

Fixes: #14387
Signed-off-by: xie xingguo <xie.xingguo@zte.com.cn>
---

diff --git a/src/osd/OSD.cc b/src/osd/OSD.cc
index 7345d535301b..c4d8fa52ca2e 100644
--- a/src/osd/OSD.cc
+++ b/src/osd/OSD.cc
@@ -3658,10 +3658,11 @@ void OSD::maybe_update_heartbeat_peers()
     }
   }
 
-  Mutex::Locker l(heartbeat_lock);
   if (!heartbeat_peers_need_update())
     return;
-  heartbeat_need_update = false;
+  heartbeat_clear_peers_need_update();
+
+  Mutex::Locker l(heartbeat_lock);
 
   dout(10) << "maybe_update_heartbeat_peers updating" << dendl;
 
diff --git a/src/osd/OSD.h b/src/osd/OSD.h
index e53fc6b640ca..5c1066a16746 100644
--- a/src/osd/OSD.h
+++ b/src/osd/OSD.h
@@ -1541,6 +1541,10 @@ private:
     Mutex::Locker l(heartbeat_update_lock);
     heartbeat_need_update = true;
   }
+  void heartbeat_clear_peers_need_update() {
+    Mutex::Locker l(heartbeat_update_lock);
+    heartbeat_need_update = false;
+  }
   void heartbeat();
   void heartbeat_check();
   void heartbeat_entry();