osd: track per-pg epochs, min

Add some simple tracking so that we can quickly determine what the min
pg osdmap epoch is.

Signed-off-by: Sage Weil <sage@inktank.com>
(cherry picked from commit 81e4c47722255ac3d46f701a80e104cc390e766c)

diff --git a/src/osd/OSD.cc b/src/osd/OSD.cc
index 51fdfa94d900c7b915b37aabbcc5a7f94c130fb1..747894acf6dcff156388a1a8685a6fc577a64f41 100644 (file)
--- a/src/osd/OSD.cc
+++ b/src/osd/OSD.cc
@@ -191,6 +191,7 @@ OSDService::OSDService(OSD *osd) :
   push_wq("push_wq", cct->_conf->osd_recovery_thread_timeout, &osd->recovery_tp),
   gen_wq("gen_wq", cct->_conf->osd_recovery_thread_timeout, &osd->recovery_tp),
   class_handler(osd->class_handler),
+  pg_epoch_lock("OSDService::pg_epoch_lock"),
   publish_lock("OSDService::publish_lock"),
   pre_publish_lock("OSDService::pre_publish_lock"),
   sched_scrub_lock("OSDService::sched_scrub_lock"), scrubs_pending(0),
@@ -1854,6 +1855,8 @@ PG *OSD::_open_lock_pg(
 
   pg_map[pgid] = pg;
 
+  service.pg_add_epoch(pg->info.pgid, createmap->get_epoch());
+
   pg->lock(no_lockdep_check);
   pg->get("PGMap");  // because it's in pg_map
   return pg;
@@ -1885,6 +1888,7 @@ void OSD::add_newly_split_pg(PG *pg, PG::RecoveryCtx *rctx)
   epoch_t e(service.get_osdmap()->get_epoch());
   pg->get("PGMap");  // For pg_map
   pg_map[pg->info.pgid] = pg;
+  service.pg_add_epoch(pg->info.pgid, pg->get_osdmap()->get_epoch());
   dout(10) << "Adding newly split pg " << *pg << dendl;
   vector<int> up, acting;
   pg->get_osdmap()->pg_to_up_acting_osds(pg->info.pgid.pgid, up, acting);
@@ -5795,6 +5799,7 @@ void OSD::advance_pg(
     lastmap = nextmap;
     handle.reset_tp_timeout();
   }
+  service.pg_update_epoch(pg->info.pgid, lastmap->get_epoch());
   pg->handle_activate_map(rctx);
 }
 
@@ -7178,6 +7183,8 @@ void OSD::_remove_pg(PG *pg)
     );
   remove_wq.queue(make_pair(PGRef(pg), deleting));
 
+  service.pg_remove_epoch(pg->info.pgid);
+
   // remove from map
   pg_map.erase(pg->info.pgid);
   pg->put("PGMap"); // since we've taken it out of map

diff --git a/src/osd/OSD.h b/src/osd/OSD.h
index d652db7da56963b4c75e483a390868d85f55f248..a8a1d08625f1a16a4bb7f7a7511d4c3e9e2a9f58 100644 (file)
--- a/src/osd/OSD.h
+++ b/src/osd/OSD.h
@@ -334,6 +334,42 @@ public:
 
   void dequeue_pg(PG *pg, list<OpRequestRef> *dequeued);
 
+  // -- map epoch lower bound --
+  Mutex pg_epoch_lock;
+  multiset<epoch_t> pg_epochs;
+  map<spg_t,epoch_t> pg_epoch;
+
+  void pg_add_epoch(spg_t pgid, epoch_t epoch) {
+    Mutex::Locker l(pg_epoch_lock);
+    map<spg_t,epoch_t>::iterator t = pg_epoch.find(pgid);
+    assert(t == pg_epoch.end());
+    pg_epoch[pgid] = epoch;
+    pg_epochs.insert(epoch);
+  }
+  void pg_update_epoch(spg_t pgid, epoch_t epoch) {
+    Mutex::Locker l(pg_epoch_lock);
+    map<spg_t,epoch_t>::iterator t = pg_epoch.find(pgid);
+    assert(t != pg_epoch.end());
+    pg_epochs.erase(t->second);
+    t->second = epoch;
+    pg_epochs.insert(epoch);
+  }
+  void pg_remove_epoch(spg_t pgid) {
+    Mutex::Locker l(pg_epoch_lock);
+    map<spg_t,epoch_t>::iterator t = pg_epoch.find(pgid);
+    if (t != pg_epoch.end()) {
+      pg_epochs.erase(t->second);
+      pg_epoch.erase(t);
+    }
+  }
+  epoch_t get_min_pg_epoch() {
+    Mutex::Locker l(pg_epoch_lock);
+    if (pg_epochs.empty())
+      return 0;
+    else
+      return *pg_epochs.begin();
+  }
+
   // -- superblock --
   Mutex publish_lock, pre_publish_lock; // pre-publish orders before publish
   OSDSuperblock superblock;