DAOS-15968 container: Fix CAPA fetch on NS master (#14511)

Apparently, when an I/O request fetches a container handle on the engine
who is the PS leader, if the ds_cont_hdl object is absent (because the
engine has restarted), then the cont_iv code won't create a ds_cont_hdl
object, even though the handle is found in the RDB. This patch adds a
ds_cont_tgt_open call to cont_iv_ent_fetch, so that the flow is similar
to cont_iv_ent_update.

Signed-off-by: Li Wei <wei.g.li@intel.com>

src/container/container_iv.c

@@ -497,8 +497,15 @@ cont_iv_ent_fetch(struct ds_iv_entry *entry, struct ds_iv_key *key,
 					}
 					prop_entry = daos_prop_entry_get(prop, DAOS_PROP_CO_STATUS);
 					D_ASSERT(prop_entry != NULL);
-
 					daos_prop_val_2_co_status(prop_entry->dpe_val, &stat);
+
+					rc = ds_cont_tgt_open(entry->ns->iv_pool_uuid,
+							      civ_key->cont_uuid, chdl.ch_cont,
+							      chdl.ch_flags, chdl.ch_sec_capas,
+							      stat.dcs_pm_ver);
+					if (rc != 0)
+						D_GOTO(out, rc);
+
 					iv_entry.iv_capa.status_pm_ver = stat.dcs_pm_ver;
 					daos_prop_free(prop);
 					/* Only happens on xstream 0 */
@@ -510,6 +517,11 @@ cont_iv_ent_fetch(struct ds_iv_entry *entry, struct ds_iv_key *key,
 					rc = dbtree_update(root_hdl, &key_iov, &val_iov);
 					if (rc == 0)
 						goto again;
+					/*
+					 * It seems that not rolling back the ds_cont_tgt_open call
+					 * above is harmless. Also, an error from the dbtree_update
+					 * call should be rare.
+					 */
 				} else {
 					rc = -DER_NONEXIST;
 				}