From: Venky Shankar <vshankar@redhat.com>
Date: Wed, 3 Jun 2020 07:17:38 +0000 (-0400)
Subject: client: introduce timeout for client shutdown
X-Git-Tag: v14.2.11~19^2
X-Git-Url: http://git.apps.os.sepia.ceph.com/?a=commitdiff_plain;h=a58e642a167e3ea8f2eef03044d031ce3422b4ff;p=ceph.git

client: introduce timeout for client shutdown

Client::shutdown() could indefinitely wait when tearing down
MDS sessions if an MDS is unreachable during shutdown, but a
valid session existed, i.e., Client::mount() was successfull.

These failures were initially observed in mgr/volumes tests in
teuthology and lately during rados/mgr selftests where manager
plugins do not respawn as volumes plugin waits for cleaning up
libcephfs handles via Client::shutdown().

Fixes: http://tracker.ceph.com/issues/44276
Signed-off-by: Venky Shankar <vshankar@redhat.com>
(cherry picked from commit 7c716be4b94d7640d051d427f5d7e97fefe5baf2)

 Conflicts:
	src/client/Client.cc
	src/client/Client.h
	src/common/options.cc

Condition variables in nautilus use WaitInterval() rather
than wait_for(). Client::_closed_mds_session() does not accept
errno parameter in nautilus -- so adjust for that. Also, fixup
config option conflict.
---

diff --git a/src/client/Client.cc b/src/client/Client.cc
index 3a02b72c2ba29..5a4d3135b4c8a 100644
--- a/src/client/Client.cc
+++ b/src/client/Client.cc
@@ -2083,6 +2083,7 @@ void Client::_closed_mds_session(MetaSession *s)
   mount_cond.Signal();
   remove_session_caps(s);
   kick_requests_closed(s);
+  mds_ranks_closing.erase(s->mds_num);
   mds_sessions.erase(s->mds_num);
 }
 
@@ -6006,12 +6007,34 @@ void Client::_close_sessions()
     for (auto &p : mds_sessions) {
       if (p.second.state != MetaSession::STATE_CLOSING) {
 	_close_mds_session(&p.second);
+	mds_ranks_closing.insert(p.first);
       }
     }
 
     // wait for sessions to close
-    ldout(cct, 2) << "waiting for " << mds_sessions.size() << " mds sessions to close" << dendl;
-    mount_cond.Wait(client_lock);
+    double timo = cct->_conf.get_val<std::chrono::seconds>("client_shutdown_timeout").count();
+    ldout(cct, 2) << "waiting for " << mds_ranks_closing.size() << " mds session(s) to close (timeout: "
+                  << timo << "s)" << dendl;
+    if (!timo) {
+      mount_cond.Wait(client_lock);
+    } else {
+      int r = 0;
+      utime_t t;
+      t.set_from_double(timo);
+      while (!mds_ranks_closing.empty() && r == 0) {
+        r = mount_cond.WaitInterval(client_lock, t);
+      }
+      if (r != 0) {
+        ldout(cct, 1) << mds_ranks_closing.size() << " mds(s) did not respond to session close -- timing out." << dendl;
+        while (!mds_ranks_closing.empty()) {
+          auto session = mds_sessions.at(*mds_ranks_closing.begin());
+          // this prunes entry from mds_sessions and mds_ranks_closing
+          _closed_mds_session(&session);
+        }
+      }
+    }
+
+    mds_ranks_closing.clear();
   }
 }
 
diff --git a/src/client/Client.h b/src/client/Client.h
index 6e34e4ba1bbb8..0747baab91863 100644
--- a/src/client/Client.h
+++ b/src/client/Client.h
@@ -1210,6 +1210,8 @@ private:
 
   // mds sessions
   map<mds_rank_t, MetaSession> mds_sessions;  // mds -> push seq
+
+  std::set<mds_rank_t> mds_ranks_closing;  // mds ranks currently tearing down sessions
   list<Cond*> waiting_for_mdsmap;
 
   // FSMap, for when using mds_command
diff --git a/src/common/options.cc b/src/common/options.cc
index 149d14db7c8fa..43f10eccec474 100644
--- a/src/common/options.cc
+++ b/src/common/options.cc
@@ -8441,6 +8441,14 @@ std::vector<Option> get_mds_client_options() {
     Option("debug_allow_any_pool_priority", Option::TYPE_BOOL, Option::LEVEL_DEV)
     .set_default(false)
     .set_description("Allow any pool priority to be set to test conversion to new range"),
+
+    Option("client_shutdown_timeout", Option::TYPE_SECS, Option::LEVEL_ADVANCED)
+    .set_flag(Option::FLAG_RUNTIME)
+    .set_default(30)
+    .set_min(0)
+    .set_description("timeout for shutting down CephFS")
+    .set_long_description("Timeout for shutting down CephFS via unmount or shutdown.")
+    .add_tag("client")
   });
 }