From 984b07dabc319b8d1b881e8601f0e0214273300e Mon Sep 17 00:00:00 2001
From: Sage Weil <sage@newdream.net>
Date: Tue, 9 Dec 2008 14:57:19 -0800
Subject: [PATCH] osd: drive primary recovery via missing map, not log

---
 src/osd/PG.cc           | 33 +++++++--------------------------
 src/osd/PG.h            |  6 ++----
 src/osd/ReplicatedPG.cc | 24 +++++++++++-------------
 3 files changed, 20 insertions(+), 43 deletions(-)

diff --git a/src/osd/PG.cc b/src/osd/PG.cc
index e014454738a33..8e99007fe43d0 100644
--- a/src/osd/PG.cc
+++ b/src/osd/PG.cc
@@ -120,17 +120,13 @@ void PG::IndexedLog::trim(ObjectStore::Transaction& t, eversion_t s)
   if (backlog && s < bottom)
     s = bottom;
 
-  assert(complete_to == log.end() &&
-	 requested_to == log.end());
+  assert(complete_to == log.end());
 
   while (!log.empty()) {
     Entry &e = *log.begin();
 
     if (e.version > s) break;
 
-    assert(complete_to != log.begin());
-    assert(requested_to != log.begin());
-
     // remove from index,
     unindex(e);
 
@@ -1174,18 +1170,6 @@ void PG::peer(ObjectStore::Transaction& t,
 }
 
 
-void PG::init_recovery_pointers()
-{
-  dout(10) << "init_recovery_pointers" << dendl;
-  log.complete_to = log.log.begin();
-  while (log.complete_to->version < info.last_complete)
-    log.complete_to++;
-  assert(log.complete_to != log.log.end());
-  
-  if (is_primary())
-    log.requested_to = log.complete_to;
-}
-
 void PG::activate(ObjectStore::Transaction& t,
 		  map<int, MOSDPGInfo*> *activator_map)
 {
@@ -1239,9 +1223,11 @@ void PG::activate(ObjectStore::Transaction& t,
     log.reset_recovery_pointers();
   } else {
     dout(10) << "activate - not complete, " << missing << dendl;
-
-    init_recovery_pointers();
-
+    log.complete_to = log.log.begin();
+    while (log.complete_to->version < info.last_complete)
+      log.complete_to++;
+    assert(log.complete_to != log.log.end());
+    log.last_requested = object_t();
     dout(10) << "activate -     complete_to = " << log.complete_to->version << dendl;
     if (is_primary()) {
       dout(10) << "activate - starting recovery" << dendl;
@@ -1908,12 +1894,7 @@ void PG::repair_object(ScrubMap::object *po, int bad_peer, int ok_peer)
   } else {
     missing.add(po->poid.oid, v, eversion_t());
     missing_loc[po->poid.oid].insert(ok_peer);
-
-    // primary recovery is log driven
-    if (v < info.last_complete) {
-      info.last_complete = v;
-      init_recovery_pointers();
-    }
+    log.last_requested = object_t();
   }
   uptodate_set.erase(bad_peer);
   osd->queue_for_recovery(this);
diff --git a/src/osd/PG.h b/src/osd/PG.h
index 4949e57602ed9..8e7367e7b427c 100644
--- a/src/osd/PG.h
+++ b/src/osd/PG.h
@@ -294,8 +294,8 @@ public:
     hash_set<osd_reqid_t>     caller_ops;
 
     // recovery pointers
-    list<Entry>::iterator requested_to; // not inclusive of referenced item
     list<Entry>::iterator complete_to;  // not inclusive of referenced item
+    object_t last_requested;            // last object requested by primary
 
     /****/
     IndexedLog() {}
@@ -307,8 +307,8 @@ public:
       reset_recovery_pointers();
     }
     void reset_recovery_pointers() {
-      requested_to = log.end();
       complete_to = log.end();
+      last_requested = object_t();
     }
 
     bool logged_object(object_t oid) const {
@@ -705,8 +705,6 @@ public:
   virtual void cancel_recovery() = 0;
   virtual int start_recovery_ops(int max) = 0;
 
-  void init_recovery_pointers();
-
   void purge_strays();
 
 
diff --git a/src/osd/ReplicatedPG.cc b/src/osd/ReplicatedPG.cc
index 67b081f8e7ac3..6edbe4a11f655 100644
--- a/src/osd/ReplicatedPG.cc
+++ b/src/osd/ReplicatedPG.cc
@@ -2679,24 +2679,23 @@ int ReplicatedPG::recover_primary(int max)
   int started = 0;
   int skipped = 0;
 
-  list<Log::Entry>::iterator p = log.requested_to;
-
-  while (p != log.log.end()) {
-    assert(log.objects.count(p->oid));
-    latest = log.objects[p->oid];
+  map<object_t, Missing::item>::iterator p = missing.missing.lower_bound(log.last_requested);
+  while (p != missing.missing.end()) {
+    assert(log.objects.count(p->first));
+    latest = log.objects[p->first];
     assert(latest);
 
     dout(10) << "recover_primary "
-             << *p
+             << *latest
 	     << (latest->is_update() ? " (update)":"")
 	     << (missing.is_missing(latest->oid) ? " (missing)":"")
              << (pulling.count(latest->oid) ? " (pulling)":"")
 	     << (waiting_for_head.count(latest->oid) ? " (waiting for head)":"")
              << dendl;
+    
+    assert(latest->is_update());
 
-    if (latest->is_update() &&
-        !pulling.count(latest->oid) &&
-        missing.is_missing(latest->oid)) {
+    if (!pulling.count(latest->oid)) {
       if (waiting_for_head.count(latest->oid)) {
 	++skipped;
       } else {
@@ -2732,9 +2731,9 @@ int ReplicatedPG::recover_primary(int max)
     
     p++;
 
-    // only advance requested_to if we haven't skipped anything
+    // only advance last_requested if we haven't skipped anything
     if (!skipped)
-      log.requested_to = p;
+      log.last_requested = latest->oid;
   }
 
   // done?
@@ -2753,8 +2752,7 @@ int ReplicatedPG::recover_primary(int max)
     info.last_complete = info.last_update;
   }
 
-  log.complete_to = log.log.end();
-  log.requested_to = log.log.end();
+  log.reset_recovery_pointers();
 
   uptodate_set.insert(osd->whoami);
   if (is_all_uptodate()) {
-- 
2.39.5