osd: guard rep_ops based on map epoch. fix DOWN state. prefer prior_osds for newest...

diff --git a/src/osd/OSD.cc b/src/osd/OSD.cc
index 47b3584307d74c8fdff41946eb0905a4d243873f..0878c1dde7567f1f2c86f20d620635567f02bbf7 100644 (file)
--- a/src/osd/OSD.cc
+++ b/src/osd/OSD.cc
@@ -1704,6 +1704,7 @@ void OSD::advance_map(ObjectStore::Transaction& t)
     
     // deactivate.
     pg->state_clear(PG_STATE_ACTIVE);
+    pg->state_clear(PG_STATE_DOWN);
     
     // reset primary state?
     if (oldrole == 0 || pg->get_role() == 0)

diff --git a/src/osd/PG.cc b/src/osd/PG.cc
index 5decfcb44f0ea51413fd19190f5f2f76e836544d..2ef8ac3d8db9689cf00b335f8a7c3ec9514d8c89 100644 (file)
--- a/src/osd/PG.cc
+++ b/src/osd/PG.cc
@@ -676,7 +676,8 @@ void PG::build_prior()
     }
   }
 
-  if (!any_up) {
+  if (info.history.last_epoch_started < info.history.same_since &&
+      !any_up) {
     dout(10) << "build_prior  no osds are up from the last epoch started, PG is down for now." << dendl;
     state_set(PG_STATE_DOWN);
   }
@@ -763,7 +764,9 @@ void PG::peer(ObjectStore::Transaction& t,
   for (map<int,Info>::iterator it = peer_info.begin();
        it != peer_info.end();
        it++) {
-    if (it->second.last_update > newest_update) {
+    if (it->second.last_update > newest_update ||
+       (it->second.last_update == newest_update &&    // prefer osds in the prior set
+        prior_set.count(newest_update_osd) == 0)) {
       newest_update = it->second.last_update;
       newest_update_osd = it->first;
     }
@@ -774,6 +777,8 @@ void PG::peer(ObjectStore::Transaction& t,
         peers_complete_thru = it->second.last_complete;
     }    
   }
+  if (newest_update == info.last_update)   // or just me, if nobody better.
+    newest_update_osd = osd->whoami;
 
   // gather log(+missing) from that person!
   if (newest_update_osd != osd->whoami) {
@@ -943,6 +948,7 @@ void PG::activate(ObjectStore::Transaction& t,
   // twiddle pg state
   state_set(PG_STATE_ACTIVE);
   state_clear(PG_STATE_STRAY);
+  state_clear(PG_STATE_DOWN);
   if (is_crashed()) {
     //assert(is_replay());      // HELP.. not on replica?
     state_clear(PG_STATE_CRASHED);

diff --git a/src/osd/PG.h b/src/osd/PG.h
index c8d1fda99d5e5cbb07b84b43de4557cd925bdb18..37b7097f15e927d70ccc196414e73621c07b2964 100644 (file)
--- a/src/osd/PG.h
+++ b/src/osd/PG.h
@@ -645,6 +645,7 @@ public:
   void       set_role(int r) { role = r; }
 
   bool       is_primary() const { return role == PG_ROLE_HEAD; }
+  bool       is_replica() const { return role > 0; }
   bool       is_acker() const { 
     if (g_conf.osd_rep == OSD_REP_PRIMARY)
       return is_primary();
@@ -662,6 +663,7 @@ public:
 
   bool is_complete() const { return info.last_complete == info.last_update; }
 
+  int get_state() const { return state; }
   bool       is_active() const { return state_test(PG_STATE_ACTIVE); }
   bool       is_crashed() const { return state_test(PG_STATE_CRASHED); }
   bool       is_down() const { return state_test(PG_STATE_DOWN); }
@@ -794,11 +796,9 @@ inline ostream& operator<<(ostream& out, const PG& pg)
     out << " pct " << pg.peers_complete_thru;
     if (!pg.have_master_log) out << " !hml";
   }
-  if (pg.is_active()) out << " active";
-  if (pg.is_crashed()) out << " crashed";
-  if (pg.is_replay()) out << " replay";
-  if (pg.is_clean()) out << " clean";
-  if (pg.is_stray()) out << " stray";
+
+  out << " " << pg_state_string(pg.get_state());
+
   //out << " (" << pg.log.bottom << "," << pg.log.top << "]";
   if (pg.missing.num_missing()) out << " m=" << pg.missing.num_missing();
   if (pg.missing.num_lost()) out << " l=" << pg.missing.num_lost();

diff --git a/src/osd/ReplicatedPG.cc b/src/osd/ReplicatedPG.cc
index d54927ec618ec850fa715bf97f51decf56876a14..69e636fa7b156057f38a548c130ab007040278f4 100644 (file)
--- a/src/osd/ReplicatedPG.cc
+++ b/src/osd/ReplicatedPG.cc
@@ -1285,6 +1285,20 @@ void ReplicatedPG::sub_op_modify(MOSDSubOp *op)
            << " v " << nv 
            << " " << op->get_offset() << "~" << op->get_length()
            << dendl;  
+
+  // sanity checks
+  if (op->get_map_epoch() < info.history.same_primary_since) {
+    dout(10) << "sub_op_modify discarding old sub_op from "
+            << op->get_map_epoch() << " < " << info.history.same_primary_since << dendl;
+    delete op;
+    return;
+  }
+  if (!is_active()) {
+    dout(10) << "sub_op_modify not active" << dendl;
+    delete op;
+    return;
+  }
+  assert(is_replica());
   
   // note peer's stat
   int fromosd = op->get_source().num();
@@ -1478,37 +1492,20 @@ void ReplicatedPG::sub_op_pull(MOSDSubOp *op)
 {
   const pobject_t poid = op->get_poid();
   const eversion_t v = op->get_version();
-  int from = op->get_source().num();
 
   dout(7) << "op_pull " << poid << " v " << op->get_version()
           << " from " << op->get_source()
           << dendl;
 
-  // is a replica asking?  are they missing it?
-  if (is_primary()) {
-    // primary
-    assert(peer_missing.count(from));  // we had better know this, from the peering process.
+  if (op->get_map_epoch() < info.history.same_primary_since) {
+    dout(10) << "sub_op_pull discarding old sub_op from "
+            << op->get_map_epoch() << " < " << info.history.same_primary_since << dendl;
+    delete op;
+    return;
+  }
 
-    if (!peer_missing[from].is_missing(poid.oid)) {
-      dout(7) << "op_pull replica isn't actually missing it, we must have already pushed to them" << dendl;
-      delete op;
-      return;
-    }
+  assert(!is_primary());  // we should be a replica or stray.
 
-    // do we have it yet?
-    if (is_missing_object(poid.oid)) {
-      wait_for_missing_object(poid.oid, op);
-      return;
-    }
-  } else {
-    // non-primary
-    if (missing.is_missing(poid.oid)) {
-      dout(7) << "op_pull not primary, and missing " << poid << ", ignoring" << dendl;
-      delete op;
-      return;
-    }
-  }
-    
   // push it back!
   push(poid, op->get_source().num());
 }
@@ -1522,6 +1519,30 @@ void ReplicatedPG::sub_op_push(MOSDSubOp *op)
   pobject_t poid = op->get_poid();
   eversion_t v = op->get_version();
 
+  dout(7) << "op_push " 
+          << poid 
+          << " v " << v 
+          << " size " << op->get_length() << " " << op->get_data().length()
+          << dendl;
+
+  if (is_replica()) {
+    // replica should only accept pushes from the current primary.
+    if (op->get_map_epoch() < info.history.same_primary_since) {
+      dout(10) << "sub_op_push discarding old sub_op from "
+              << op->get_map_epoch() << " < " << info.history.same_primary_since << dendl;
+      delete op;
+      return;
+    }
+    // FIXME: actually, no, what i really want here is a personal "same_role_since"
+    if (!is_active()) {
+      dout(10) << "sub_op_push not active" << dendl;
+      delete op;
+      return;
+    }
+  } else {
+    // primary will accept pushes anytime.
+  }
+
   // are we missing (this specific version)?
   //  (if version is wrong, it is either old (we don't want it) or 
   //   newer (peering is buggy))
@@ -1531,12 +1552,6 @@ void ReplicatedPG::sub_op_push(MOSDSubOp *op)
     return;
   }
   
-  dout(7) << "op_push " 
-          << poid 
-          << " v " << v 
-          << " size " << op->get_length() << " " << op->get_data().length()
-          << dendl;
-
   assert(op->get_data().length() == op->get_length());
   
   // write object and add it to the PG