Merge PR ceph#53839 into main

* refs/pull/53839/head:
	qa: enhance test cases
	mds: erase clients getting evicted from laggy_clients
	mds: report clients laggy due laggy OSDs only after checking any OSD is laggy

Reviewed-by: Venky Shankar <[email protected]>

Author: vshankar

qa/tasks/cephfs/test_client_recovery.py

@@ -13,6 +13,7 @@
 import string
 import os
 
+from teuthology import contextutil
 from teuthology.orchestra import run
 from teuthology.exceptions import CommandFailedError
 from tasks.cephfs.fuse_mount import FuseMount
@@ -808,24 +809,27 @@ def test_client_eviction_if_config_is_set(self):
             # it takes time to have laggy clients entries in cluster log,
             # wait for 6 minutes to see if it is visible, finally restart
             # the client
-            tries = 6
-            while True:
-                try:
-                    with self.assert_cluster_log("1 client(s) laggy due to laggy OSDs",
-                                                 timeout=55):
-                        # make sure clients weren't evicted
-                        self.assert_session_count(2)
-                        break
-                except AssertionError:
-                    tries -= 1
-                    if tries:
-                        continue
-                    raise
+            with contextutil.safe_while(sleep=5, tries=6) as proceed:
+                while proceed():
+                    try:
+                        with self.assert_cluster_log("1 client(s) laggy due to"
+                                                     " laggy OSDs",
+                                                     timeout=55):
+                            # make sure clients weren't evicted
+                            self.assert_session_count(2)
+                            break
+                    except (AssertionError, CommandFailedError) as e:
+                        log.debug(f'{e}, retrying')
+
+            # clear lagginess, expect to get the warning cleared and make sure
+            # client gets evicted
+            self.clear_laggy_params(osd)
+            self.wait_for_health_clear(60)
+            self.assert_session_count(1)
         finally:
             self.mount_a.kill_cleanup()
             self.mount_a.mount_wait()
             self.mount_a.create_destroy()
-            self.clear_laggy_params(osd)
 
     def test_client_eviction_if_config_is_unset(self):
         """
@@ -857,6 +861,11 @@ def test_client_eviction_if_config_is_unset(self):
 
             time.sleep(session_timeout)
             self.assert_session_count(1)
+
+            # make sure warning wasn't seen in cluster log
+            with self.assert_cluster_log("laggy due to laggy OSDs",
+                                         timeout=120, present=False):
+                pass
         finally:
             self.mount_a.kill_cleanup()
             self.mount_a.mount_wait()

src/mds/Beacon.cc

@@ -500,13 +500,17 @@ void Beacon::notify_health(MDSRank const *mds)
 
   // Report laggy client(s) due to laggy OSDs
   {
+    bool defer_client_eviction =
+    g_conf().get_val<bool>("defer_client_eviction_on_laggy_osds")
+    && mds->objecter->with_osdmap([](const OSDMap &map) {
+      return map.any_osd_laggy(); });
     auto&& laggy_clients = mds->server->get_laggy_clients();
-    if (!laggy_clients.empty()) {
+    if (defer_client_eviction && !laggy_clients.empty()) {
       std::vector<MDSHealthMetric> laggy_clients_metrics;
       for (const auto& laggy_client: laggy_clients) {
         CachedStackStringStream css;
         *css << "Client " << laggy_client << " is laggy; not evicted"
-            << " because some OSD(s) is/are laggy";
+             << " because some OSD(s) is/are laggy";
         MDSHealthMetric m(MDS_HEALTH_CLIENTS_LAGGY, HEALTH_WARN, css->strv());
         laggy_clients_metrics.emplace_back(std::move(m));
       }

src/mds/Server.cc

@@ -1277,6 +1277,8 @@ void Server::find_idle_sessions()
       kill_session(session, NULL);
     }
   }
+  // clear as there's no use to keep the evicted clients in laggy_clients
+  clear_laggy_clients();
 }
 
 void Server::evict_cap_revoke_non_responders() {