mon: prevent osd mark-down with NODOWN flag

If the NODOWN osdmap flag is set,

 - ignore osd failure reports
 - do not mark osds down due to lack of osd/pg stats

We *do* still allow explicit admin 'ceph osd down N' commands, and a
booting OSD to mark the previous instance of itself down.

Signed-off-by: Sage Weil <sage@newdream.net>

diff --git a/src/mon/OSDMonitor.cc b/src/mon/OSDMonitor.cc
index dbc4e92fd98cd07f4b61022e5b2fb7858332cec7..381c93df8b3d38d12854d0ded76665a04f3602bc 100644 (file)
--- a/src/mon/OSDMonitor.cc
+++ b/src/mon/OSDMonitor.cc
@@ -468,6 +468,7 @@ bool OSDMonitor::preprocess_failure(MOSDFailure *m)
       send_incremental(m, m->get_epoch()+1);
     goto didit;
   }
+
   // already reported?
   if (osdmap.is_down(badboy)) {
     dout(5) << "preprocess_failure dup: " << m->get_target() << ", from " << m->get_orig_source_inst() << dendl;
@@ -476,6 +477,12 @@ bool OSDMonitor::preprocess_failure(MOSDFailure *m)
     goto didit;
   }
 
+  // NODOWN?
+  if (osdmap.test_flag(CEPH_OSDMAP_NODOWN)) {
+    dout(5) << "preprocess_failure NODOWN flag set, ignoring report of " << m->get_target() << " from " << m->get_orig_source_inst() << dendl;
+    goto didit;
+  }
+
   dout(10) << "preprocess_failure new: " << m->get_target() << ", from " << m->get_orig_source_inst() << dendl;
   return false;
 
@@ -1203,7 +1210,7 @@ void OSDMonitor::handle_osd_timeouts(const utime_t &now,
     if (t == last_osd_report.end()) {
       // it wasn't in the map; start the timer.
       last_osd_report[i] = now;
-    } else {
+    } else if (!osdmap.test_flag(CEPH_OSDMAP_NODOWN)) {
       utime_t diff = now - t->second;
       if (diff > timeo) {
        derr << "no osd or pg stats from osd." << i << " since " << t->second << ", " << diff