osd: cancel pending failure reports on re-activating osd

To reproduce, construct a cluster with 3 hosts, each containing a single
osd only:
- cut off osd.1's cluster network, waiting osd.1 to be marked as down
- cut off both osd.2 & osd.3's cluster network
```
Note that there are two possible outputs for the above step:
1. osd.1's failure reports get ignored by monitor as osd.1 has already been
   marked as down. Osd.2 & osd.3 stay __up__ as a result.
2. osd.1's failure reports are considered as valid. Either osd.2 or osd.3 is
   marked as __down__.
We consider case __2__ only here.
```
- restore osd.1 & osd.2's cluster network
Now you get __3__ up osds.

The root cause is that monitor will simply discard any failure reports from
dead osds, whereas osds never re-send pending failure reports unless they are
re-connecting to monitors.

Fix by cancelling any pending failure reports each time an osd is
transiting from dead to active *again*.

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

diff --git a/src/osd/OSD.cc b/src/osd/OSD.cc
index 8f3bd8e6afc80faf168ef6db023f609f1ea93a33..e239e4961896d580a1b2c01d697a886773e83e33 100644 (file)
--- a/src/osd/OSD.cc
+++ b/src/osd/OSD.cc
@@ -5876,6 +5876,18 @@ void OSD::send_still_alive(epoch_t epoch, int osd, const entity_addrvec_t &addrs
   monc->send_mon_message(m);
 }
 
+void OSD::cancel_pending_failures()
+{
+  std::lock_guard l(heartbeat_lock);
+  auto it = failure_pending.begin();
+  while (it != failure_pending.end()) {
+    dout(10) << __func__ << " canceling in-flight failure report for osd."
+             << it->first << dendl;
+    send_still_alive(osdmap->get_epoch(), it->first, it->second.second);
+    failure_pending.erase(it++);
+  }
+}
+
 void OSD::send_beacon(const ceph::coarse_mono_clock::time_point& now)
 {
   const auto& monmap = monc->monmap;
@@ -7834,6 +7846,7 @@ void OSD::_committed_osd_maps(epoch_t first, epoch_t last, MOSDMap *m)
       // set incarnation so that osd_reqid_t's we generate for our
       // objecter requests are unique across restarts.
       service.objecter->set_client_incarnation(osdmap->get_epoch());
+      cancel_pending_failures();
     }
   }
 
@@ -7975,14 +7988,7 @@ void OSD::_committed_osd_maps(epoch_t first, epoch_t last, MOSDMap *m)
 
   if (do_shutdown) {
     if (network_error) {
-      std::lock_guard l(heartbeat_lock);
-      auto it = failure_pending.begin();
-      while (it != failure_pending.end()) {
-        dout(10) << "handle_osd_ping canceling in-flight failure report for osd."
-                << it->first << dendl;
-        send_still_alive(osdmap->get_epoch(), it->first, it->second.second);
-        failure_pending.erase(it++);
-      }
+      cancel_pending_failures();
     }
     // trigger shutdown in a different thread
     dout(0) << __func__ << " shutdown OSD via async signal" << dendl;

diff --git a/src/osd/OSD.h b/src/osd/OSD.h
index c8e0cea6a81d18f101d271ec908cb9cbfbb7bd90..3f22ffe8dd0a2f04e72825e146047a4108f7d63d 100644 (file)
--- a/src/osd/OSD.h
+++ b/src/osd/OSD.h
@@ -2008,6 +2008,7 @@ protected:
   void requeue_failures();
   void send_failures();
   void send_still_alive(epoch_t epoch, int osd, const entity_addrvec_t &addrs);
+  void cancel_pending_failures();
 
   ceph::coarse_mono_clock::time_point last_sent_beacon;
   Mutex min_last_epoch_clean_lock{"OSD::min_last_epoch_clean_lock"};