osd: maintain a cache of past osd maps during repeering

It's expensive and stupid to load and reparse them for each PG.

diff --git a/src/osd/OSD.cc b/src/osd/OSD.cc
index e7b82e72330269d546e9cead8a668e0eed752182..cec7888a91ea2b9ed62f6d21b8c0ab123be100c7 100644 (file)
--- a/src/osd/OSD.cc
+++ b/src/osd/OSD.cc
@@ -270,6 +270,8 @@ OSD::OSD(int id, Messenger *m, MonMap *mm, const char *dev) :
   qlen_calc(3),
   iat_averager(g_conf.osd_flash_crowd_iat_alpha),
   finished_lock("OSD::finished_lock"),
+  osdmap(NULL),
+  map_cache_lock("OSD::map_cache_lock"),
   snap_trimmer_lock("OSD::snap_trimmer_lock"),
   snap_trimmer_thread(this),
   pg_stat_queue_lock("OSD::pg_stat_queue_lock"),
@@ -732,10 +734,9 @@ void OSD::calc_priors_during(pg_t pgid, epoch_t start, epoch_t end, set<int>& ps
   dout(15) << "calc_priors_during " << pgid << " [" << start << "," << end << ")" << dendl;
   
   for (epoch_t e = start; e < end; e++) {
-    OSDMap oldmap;
-    get_map(e, oldmap);
+    OSDMap *oldmap = get_map(e);
     vector<int> acting;
-    oldmap.pg_to_acting_osds(pgid, acting);
+    oldmap->pg_to_acting_osds(pgid, acting);
     dout(20) << "  " << pgid << " in epoch " << e << " was " << acting << dendl;
     int added = 0;
     for (unsigned i=0; i<acting.size(); i++)
@@ -772,11 +773,10 @@ void OSD::project_pg_history(pg_t pgid, PG::Info::History& h, epoch_t from,
        e > from;
        e--) {
     // verify during intermediate epoch (e-1)
-    OSDMap oldmap;
-    get_map(e-1, oldmap);
+    OSDMap *oldmap = get_map(e-1);
 
     vector<int> acting;
-    oldmap.pg_to_acting_osds(pgid, acting);
+    oldmap->pg_to_acting_osds(pgid, acting);
 
     // acting set change?
     if (acting != last && 
@@ -937,8 +937,7 @@ void OSD::heartbeat()
                            messenger->get_myinst());
     return;
   }
-    
-
+  
   // get CPU load avg
   ifstream in("/proc/loadavg");
   if (in.is_open()) {
@@ -1900,6 +1899,8 @@ void OSD::activate_map(ObjectStore::Transaction& t)
   map< int, map<pg_t,PG::Query> > query_map;    // peer -> PG -> get_summary_since
   map<int,MOSDPGInfo*> info_map;  // peer -> message
 
+  clear_map_cache();  // we're done with it
+
   // scan pg's
   for (hash_map<pg_t,PG*>::iterator it = pg_map.begin();
        it != pg_map.end();
@@ -1979,8 +1980,15 @@ bool OSD::get_inc_map_bl(epoch_t e, bufferlist& bl)
   return store->read(0, get_inc_osdmap_pobject_name(e), 0, 0, bl) >= 0;
 }
 
-void OSD::get_map(epoch_t epoch, OSDMap &m)
+OSDMap *OSD::get_map(epoch_t epoch)
 {
+  Mutex::Locker l(map_cache_lock);
+
+  if (map_cache.count(epoch))
+    return map_cache[epoch];
+
+  OSDMap *map = new OSDMap;
+
   // find a complete map
   list<OSDMap::Incremental> incs;
   epoch_t e;
@@ -1988,7 +1996,7 @@ void OSD::get_map(epoch_t epoch, OSDMap &m)
     bufferlist bl;
     if (get_map_bl(e, bl)) {
       //dout(10) << "get_map " << epoch << " full " << e << dendl;
-      m.decode(bl);
+      map->decode(bl);
       break;
     } else {
       OSDMap::Incremental inc;
@@ -2002,11 +2010,23 @@ void OSD::get_map(epoch_t epoch, OSDMap &m)
   // apply incrementals
   for (e++; e <= epoch; e++) {
     //dout(10) << "get_map " << epoch << " inc " << e << dendl;
-    m.apply_incremental( incs.front() );
+    map->apply_incremental( incs.front() );
     incs.pop_front();
   }
+
+  map_cache[epoch] = map;
+  return map;
 }
 
+void OSD::clear_map_cache()
+{
+  Mutex::Locker l(map_cache_lock);
+  for (map<epoch_t,OSDMap*>::iterator p = map_cache.begin();
+       p != map_cache.end();
+       p++)
+    delete p->second;
+  map_cache.clear();
+}
 
 bool OSD::get_inc_map(epoch_t e, OSDMap::Incremental &inc)
 {

diff --git a/src/osd/OSD.h b/src/osd/OSD.h
index b4ac9ed33827744dc57a60a4123188dc041706b9..aa8ed2a37d378b2b2e4b1e29435be1905dca3730 100644 (file)
--- a/src/osd/OSD.h
+++ b/src/osd/OSD.h
@@ -263,7 +263,13 @@ private:
   void advance_map(ObjectStore::Transaction& t, interval_set<snapid_t>& removed_snaps);
   void activate_map(ObjectStore::Transaction& t);
 
-  void get_map(epoch_t e, OSDMap &m);
+  // osd map cache (past osd maps)
+  map<epoch_t,OSDMap*> map_cache;
+  Mutex map_cache_lock;
+
+  OSDMap* get_map(epoch_t e);
+  void clear_map_cache();
+
   bool get_map_bl(epoch_t e, bufferlist& bl);
   bool get_inc_map_bl(epoch_t e, bufferlist& bl);
   bool get_inc_map(epoch_t e, OSDMap::Incremental &inc);

diff --git a/src/osd/PG.cc b/src/osd/PG.cc
index 78f7432770e075071ee40df969c5f2182cead435..a999081e0c5879545edf48d0f4467a01e002aa18 100644 (file)
--- a/src/osd/PG.cc
+++ b/src/osd/PG.cc
@@ -657,9 +657,8 @@ void PG::build_prior()
   epoch_t stop = MAX(1, info.history.last_epoch_started);
 
   dout(10) << "build_prior considering interval " << first_epoch << " down to " << stop << dendl;
-  OSDMap *nextmap = new OSDMap;
-  osd->get_map(last_epoch, *nextmap);
-
+  OSDMap *nextmap = osd->get_map(last_epoch);
+  
   for (; last_epoch >= stop; last_epoch = first_epoch-1) {
     OSDMap *lastmap = nextmap;
     assert(last_epoch == lastmap->get_epoch());
@@ -668,9 +667,8 @@ void PG::build_prior()
     lastmap->pg_to_acting_osds(get_pgid(), acting);
     
     // calc first_epoch, first_map
-    nextmap = new OSDMap;
     for (first_epoch = last_epoch; first_epoch > stop; first_epoch--) {
-      osd->get_map(first_epoch-1, *nextmap);
+      nextmap = osd->get_map(first_epoch-1);
       vector<int> t;
       nextmap->pg_to_acting_osds(get_pgid(), t);
       if (t != acting)