OSD, PG: ignore peering messages from before the last peering restart

Check them before entering the state machine so we can
safely enter the Crashed state on unexpected messages
from the current interval.

Signed-off-by: Josh Durgin <josh.durgin@dreamhost.com>

diff --git a/src/osd/OSD.cc b/src/osd/OSD.cc
index cd6dee2ab2032e28ee45b52193f29f95abc99488..369e138d553b9d1363ec81a2d99561addcf16f49 100644 (file)
--- a/src/osd/OSD.cc
+++ b/src/osd/OSD.cc
@@ -3979,6 +3979,12 @@ void OSD::handle_pg_notify(MOSDPGNotify *m)
     if (!pg)
       continue;
 
+    if (pg->old_peering_msg(m->get_epoch())) {
+      dout(10) << "ignoring old peering message " << *m << dendl;
+      pg->unlock();
+      continue;
+    }
+
     PG::RecoveryCtx rctx(&query_map, &info_map, 0, &fin->contexts, t);
     pg->handle_notify(from, *it, &rctx);
 
@@ -4018,6 +4024,12 @@ void OSD::handle_pg_log(MOSDPGLog *m)
     return;
   }
 
+  if (pg->old_peering_msg(m->get_epoch())) {
+    dout(10) << "ignoring old peering message " << *m << dendl;
+    pg->unlock();
+    return;
+  }
+
   map< int, map<pg_t,PG::Query> > query_map;
   map< int, MOSDPGInfo* > info_map;
   PG::RecoveryCtx rctx(&query_map, &info_map, 0, &fin->contexts, t);
@@ -4057,6 +4069,13 @@ void OSD::handle_pg_info(MOSDPGInfo *m)
     PG::RecoveryCtx rctx(0, &info_map, 0, &fin->contexts, t);
     if (!pg)
       continue;
+
+    if (pg->old_peering_msg(m->get_epoch())) {
+      dout(10) << "ignoring old peering message " << *m << dendl;
+      pg->unlock();
+      continue;
+    }
+
     pg->handle_info(from, *p, &rctx);
 
     int tr = store->queue_transaction(&pg->osr, t, new ObjectStore::C_DeleteTransaction(t), fin);
@@ -4207,6 +4226,12 @@ void OSD::handle_pg_query(MOSDPGQuery *m)
       continue;
     }
 
+    if (pg->old_peering_msg(m->get_epoch())) {
+      dout(10) << "ignoring old peering message " << *m << dendl;
+      pg->unlock();
+      continue;
+    }
+
     if (pg->deleting) {
       /*
        * We cancel deletion on pg change.  And the primary will never

diff --git a/src/osd/PG.cc b/src/osd/PG.cc
index 0c40157c0a72656ed21cf58067cb80b0f5f83fbe..975cb73692ce12357632c8c7db3b78b2c86ca955 100644 (file)
--- a/src/osd/PG.cc
+++ b/src/osd/PG.cc
@@ -3425,6 +3425,11 @@ bool PG::acting_up_affected(const vector<int>& newup, const vector<int>& newacti
   }
 }
 
+bool PG::old_peering_msg(const epoch_t &msg_epoch)
+{
+  return (last_warm_restart > msg_epoch);
+}
+
 /* Called before initializing peering during advance_map */
 void PG::warm_restart(const OSDMap& lastmap, const vector<int>& newup, const vector<int>& newacting)
 {
@@ -3433,6 +3438,8 @@ void PG::warm_restart(const OSDMap& lastmap, const vector<int>& newup, const vec
   // -- there was a change! --
   kick();
 
+  last_warm_restart = osdmap.get_epoch();
+
   vector<int> oldacting, oldup;
   int oldrole = get_role();
   int oldprimary = get_primary();
@@ -4613,7 +4620,7 @@ void PG::RecoveryState::RecoveryMachine::log_exit(const char *state_name, utime_
 }
 
 
-/*----RecoverState Methods-----*/
+/*----RecoveryState Methods-----*/
 #undef dout_prefix
 #define dout_prefix *_dout << machine.pg->gen_prefix() 
 

diff --git a/src/osd/PG.h b/src/osd/PG.h
index ffa847703ce6ef3fbc684ae5a8d64ab76ab86d98..1045519c1a3bb91e307fa3cbe12ab713f8a95dcc 100644 (file)
--- a/src/osd/PG.h
+++ b/src/osd/PG.h
@@ -1277,6 +1277,9 @@ protected:
   // primary-only, recovery-only state
   set<int>             might_have_unfound;  // These osds might have objects on them
                                            // which are unfound on the primary
+
+  epoch_t last_warm_restart;
+
   friend class OSD;
 
 
@@ -1481,6 +1484,7 @@ public:
     have_master_log(true),
     recovery_state(this),
     need_up_thru(false),
+    last_warm_restart(0),
     pg_stats_lock("PG::pg_stats_lock"),
     pg_stats_valid(false),
     finish_sync_event(NULL),
@@ -1563,7 +1567,8 @@ public:
                    pair<int, Info> &notify_info);
   void fulfill_log(int from, const Query &query);
   bool acting_up_affected(const vector<int>& newup, const vector<int>& newacting);
-    
+  bool old_peering_msg(const epoch_t &msg_epoch);
+
   // recovery bits
   void handle_notify(int from, PG::Info& i, RecoveryCtx *rctx) {
     recovery_state.handle_notify(from, i, rctx);