osd/PG: clean up fastinfo key when last_update does not increase

If the last_update field goes backwards, we write a full info instead of
the fastinfo key. However, a stale fastinfo key may still be present, and
it may have a last_update that looks valid to a restarting OSD.

Fixes: https://tracker.ceph.com/issues/43580
Reported-by: song <jinglesong0814@163.com>
Signed-off-by: Sage Weil <sage@redhat.com>

src/osd/PG.cc

@@ -949,10 +949,12 @@ void PG::prepare_write(
   info.stats.stats.add(unstable_stats);
   unstable_stats.clear();
   map<string,bufferlist> km;
+  set<string> kd;
   if (dirty_big_info || dirty_info) {
     int ret = prepare_info_keymap(
       cct,
       &km,
+      &kd,
       get_osdmap_epoch(),
       info,
       last_written_info,
@@ -968,6 +970,8 @@ void PG::prepare_write(
     t, &km, coll, pgmeta_oid, pool.info.require_rollback());
   if (!km.empty())
     t.omap_setkeys(coll, pgmeta_oid, km);
+  if (!kd.empty())
+    t.omap_rmkeys(coll, pgmeta_oid, kd);
 }
 
 #pragma GCC diagnostic ignored "-Wpragmas"

src/osd/osd_types.cc

@@ -6840,6 +6840,7 @@ void OSDOp::clear_data(vector<OSDOp>& ops)
 int prepare_info_keymap(
   CephContext* cct,
   map<string,bufferlist> *km,
+  set<string> *kd,
   epoch_t epoch,
   pg_info_t &info,
   pg_info_t &last_written_info,
@@ -6885,6 +6886,10 @@ int prepare_info_keymap(
       }
       *_dout << dendl;
     }
+  } else if (info.last_update <= last_written_info.last_update) {
+    // clean up any potentially stale fastinfo key resulting from last_update
+    // not moving forwards (e.g., a backwards jump during peering)
+    kd->insert(string(fastinfo_key));
   }
 
   last_written_info = info;

src/osd/osd_types.h

@@ -6229,6 +6229,7 @@ static const __u8 pg_compat_struct_v = 10;
 int prepare_info_keymap(
   CephContext* cct,
   map<string,bufferlist> *km,
+  set<string> *kd,
   epoch_t epoch,
   pg_info_t &info,
   pg_info_t &last_written_info,