osd: use atomic for pg_map_size

This avoids the need for pg_map_lock in the max pg check.

Signed-off-by: Sage Weil <sage@redhat.com>

diff --git a/src/osd/OSD.cc b/src/osd/OSD.cc
index 67a17acc7c6ac8372ff7a3e7ce93b7520988c604..261173c7cb82c9b28ecfe78564a05d7f38c05d58 100644 (file)
--- a/src/osd/OSD.cc
+++ b/src/osd/OSD.cc
@@ -2119,10 +2119,7 @@ bool OSD::asok_command(std::string_view admin_command, const cmdmap_t& cmdmap,
     f->dump_string("state", get_state_name(get_state()));
     f->dump_unsigned("oldest_map", superblock.oldest_map);
     f->dump_unsigned("newest_map", superblock.newest_map);
-    {
-      RWLock::RLocker l(pg_map_lock);
-      f->dump_unsigned("num_pgs", pg_map.size());
-    }
+    f->dump_unsigned("num_pgs", pg_map_size);
     f->close_section();
   } else if (admin_command == "flush_journal") {
     store->flush_journal();
@@ -3802,6 +3799,7 @@ PGRef OSD::_open_pg(
   {
     RWLock::WLocker l(pg_map_lock);
     pg_map[pgid] = pg;
+    pg_map_size = pg_map.size();
     pg->get("PGMap");  // because it's in pg_map
     service.pg_add_epoch(pg->pg_id, createmap->get_epoch());
     service.init_splits_between(pgid, createmap, servicemap);
@@ -3909,6 +3907,7 @@ PGRef OSD::_lookup_pg(spg_t pgid)
        return pg;
       }
       pg_map.erase(p);
+      pg_map_size = pg_map.size();
       pg->put("PGMap");
     }
     return nullptr;
@@ -4199,10 +4198,10 @@ bool OSD::maybe_wait_for_max_pg(spg_t pgid, bool is_mon_create)
     (cct->_conf->get_val<uint64_t>("mon_max_pg_per_osd") *
      cct->_conf->get_val<double>("osd_max_pg_per_osd_hard_ratio"));
 
-  RWLock::RLocker pg_map_locker{pg_map_lock};
-  if (pg_map.size() < max_pgs_per_osd) {
+  if (pg_map_size < max_pgs_per_osd) {
     return false;
   }
+
   [[gnu::unused]] auto&& pending_creates_locker = guardedly_lock(pending_creates_lock);
   if (is_mon_create) {
     pending_creates_from_mon++;
@@ -4211,7 +4210,7 @@ bool OSD::maybe_wait_for_max_pg(spg_t pgid, bool is_mon_create)
     pending_creates_from_osd.emplace(pgid.pgid, is_primary);
   }
   dout(1) << __func__ << " withhold creation of pg " << pgid
-         << ": " << pg_map.size() << " >= "<< max_pgs_per_osd << dendl;
+         << ": " << pg_map_size << " >= "<< max_pgs_per_osd << dendl;
   return true;
 }
 
@@ -7938,6 +7937,7 @@ void OSD::_finish_splits(set<PGRef>& pgs)
     pg_map_lock.get_write();
     pg->get("PGMap");  // For pg_map
     pg_map[pg->get_pgid()] = pg;
+    pg_map_size = pg_map.size();
     service.complete_split(pg->get_pgid());
     service.pg_add_epoch(pg->pg_id, e);
     pg_map_lock.put_write();
@@ -9068,6 +9068,7 @@ void OSD::dequeue_peering_evt(
          p->second == pg) {
        dout(20) << __func__ << " removed pg " << pg << " from pg_map" << dendl;
        pg_map.erase(p);
+       pg_map_size = pg_map.size();
        pg->put("PGMap");
       } else {
        dout(20) << __func__ << " failed to remove pg " << pg << " from pg_map" << dendl;

diff --git a/src/osd/OSD.h b/src/osd/OSD.h
index a9b7ecc5548f332b3bc723e592590a29a9461f48..258656837e8af626756e7d66a4c9854f63e1a361 100644 (file)
--- a/src/osd/OSD.h
+++ b/src/osd/OSD.h
@@ -1851,6 +1851,7 @@ protected:
   // -- placement groups --
   RWLock pg_map_lock; // this lock orders *above* individual PG _locks
   ceph::unordered_map<spg_t, PG*> pg_map; // protected by pg_map lock
+  std::atomic<size_t> pg_map_size = {0};
 
   std::mutex pending_creates_lock;
   using create_from_osd_t = std::pair<pg_t, bool /* is primary*/>;