mgr/ClusterState: make pg stat filtering less fragile

We want to drop updates for pgs for pools that don't exist.  Keep an
updated set of those pools instead of relying on the previous PGMap
having them instantiated.  (The previous map may drift due to bugs.)

Signed-off-by: Sage Weil <sage@redhat.com>

1 files changed, 17 insertions, 7 deletions

diff --git a/src/mgr/ClusterState.cc b/src/mgr/ClusterState.cc
index b8d35d30453..3ca97499cd9 100644
--- a/src/mgr/ClusterState.cc
+++ b/src/mgr/ClusterState.cc
@@ -68,17 +68,20 @@ void ClusterState::ingest_pgstats(MPGStats *stats)
 
     // In case we're hearing about a PG that according to last
     // OSDMap update should not exist
-    if (pg_map.pg_stat.count(pgid) == 0) {
-      dout(15) << " got " << pgid << " reported at " << pg_stats.reported_epoch << ":"
+    if (existing_pools.count(pgid.pool()) == 0) {
+      dout(15) << " got " << pgid
+	       << " reported at " << pg_stats.reported_epoch << ":"
                << pg_stats.reported_seq
                << " state " << pg_state_string(pg_stats.state)
-               << " but DNE in pg_map; pool was probably deleted."
+               << " but pool not in " << existing_pools
                << dendl;
       continue;
-     // In case we already heard about more recent stats from this PG
-     // from another OSD
-    } else if (pg_map.pg_stat[pgid].get_version_pair() > pg_stats.get_version_pair()) {
-      dout(15) << " had " << pgid << " from " << pg_map.pg_stat[pgid].reported_epoch << ":"
+    }
+    // In case we already heard about more recent stats from this PG
+    // from another OSD
+    if (pg_map.pg_stat[pgid].get_version_pair() > pg_stats.get_version_pair()) {
+      dout(15) << " had " << pgid << " from "
+	       << pg_map.pg_stat[pgid].reported_epoch << ":"
                << pg_map.pg_stat[pgid].reported_seq << dendl;
       continue;
     }
@@ -105,6 +108,13 @@ void ClusterState::notify_osdmap(const OSDMap &osd_map)
 
   PGMapUpdater::check_osd_map(g_ceph_context, osd_map, pg_map, &pending_inc);
 
+  // update our list of pools that exist, so that we can filter pg_map updates
+  // in synchrony with this OSDMap.
+  existing_pools.clear();
+  for (auto& p : osd_map.get_pools()) {
+    existing_pools.insert(p.first);
+  }
+
   // brute force this for now (don't bother being clever by only
   // checking osds that went up/down)
   set<int> need_check_down_pg_osds;