From: Patrick Donnelly <pdonnell@redhat.com>
Date: Wed, 3 Mar 2021 03:39:09 +0000 (-0800)
Subject: mds,qa: exit instead of respawn under valgrind
X-Git-Tag: v16.2.0~30^2~1
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=da9ed46cdcdca2b351553955b8452afd28489318;p=ceph.git

mds,qa: exit instead of respawn under valgrind

valgrind can't handle execve of /proc/self/exe:

    2021-02-27T05:52:37.813 INFO:tasks.ceph.mds.d.smithi073.stderr:==00:01:03:20.556 41218== execve(0x18546740(/proc/self/exe), 0x18546670, 0x133ef310) failed, errno 2
    2021-02-27T05:52:37.813 INFO:tasks.ceph.mds.d.smithi073.stderr:==00:01:03:20.556 41218== EXEC FAILED: I can't recover from execve() failing, so I'm dying.
    2021-02-27T05:52:37.813 INFO:tasks.ceph.mds.d.smithi073.stderr:==00:01:03:20.556 41218== Add more stringent tests in PRE(sys_execve), or work out how to recover.

So configure the MDS to just exit so it can be restarted by QA infra (the
daemon watchdog).

Signed-off-by: Patrick Donnelly <pdonnell@redhat.com>
(cherry picked from commit 5faf0ee0f367f66e25a3927e6ac4c40fb83568ce)
---

diff --git a/qa/suites/fs/verify/validater/valgrind.yaml b/qa/suites/fs/verify/validater/valgrind.yaml
index 14c06ae5e1d3..930872fc7aaa 100644
--- a/qa/suites/fs/verify/validater/valgrind.yaml
+++ b/qa/suites/fs/verify/validater/valgrind.yaml
@@ -13,6 +13,8 @@ overrides:
         osd heartbeat grace: 60
         mds heartbeat grace: 60
         mds beacon grace: 60
+      mds:
+        mds valgrind exit: true
       mon:
         mon osd crush smoke test: false
       osd:
@@ -20,6 +22,8 @@ overrides:
     valgrind:
       mon: [--tool=memcheck, --leak-check=full, --show-reachable=yes]
       mds: [--tool=memcheck]
+    watchdog:
+      daemon_restart: normal
   ceph-fuse:
     client.0:
       valgrind: [--tool=memcheck, --leak-check=full, --show-reachable=yes]
diff --git a/qa/tasks/daemonwatchdog.py b/qa/tasks/daemonwatchdog.py
index b95c29920ddd..f72ccd7cef38 100644
--- a/qa/tasks/daemonwatchdog.py
+++ b/qa/tasks/daemonwatchdog.py
@@ -17,14 +17,19 @@ class DaemonWatchdog(Greenlet):
     SIGTERM to all daemons. The duration of an extended failure is configurable
     with watchdog_daemon_timeout.
 
-    watchdog_daemon_timeout [default: 300]: number of seconds a daemon
-        is allowed to be failed before the watchdog will bark.
+    ceph:
+      watchdog:
+        daemon_restart [default: no]: restart daemon if "normal" exit (status==0).
+
+        daemon_timeout [default: 300]: number of seconds a daemon
+                                              is allowed to be failed before the
+                                              watchdog will bark.
     """
 
     def __init__(self, ctx, config, thrashers):
         super(DaemonWatchdog, self).__init__()
+        self.config = ctx.config.get('watchdog', {})
         self.ctx = ctx
-        self.config = config
         self.e = None
         self.logger = log.getChild('daemon_watchdog')
         self.cluster = config.get('cluster', 'ceph')
@@ -70,7 +75,8 @@ class DaemonWatchdog(Greenlet):
 
     def watch(self):
         self.log("watchdog starting")
-        daemon_timeout = int(self.config.get('watchdog_daemon_timeout', 300))
+        daemon_timeout = int(self.config.get('daemon_timeout', 300))
+        daemon_restart = self.config.get('daemon_restart', False)
         daemon_failure_time = {}
         while not self.stopping.is_set():
             bark = False
@@ -97,6 +103,9 @@ class DaemonWatchdog(Greenlet):
                 self.log("daemon {name} is failed for ~{t:.0f}s".format(name=name, t=delta))
                 if delta > daemon_timeout:
                     bark = True
+                if daemon_restart == 'normal' and daemon.proc.exitstatus == 0:
+                    self.log(f"attempting to restart daemon {name}")
+                    daemon.restart()
 
             # If a daemon is no longer failed, remove it from tracking:
             for name in list(daemon_failure_time.keys()):
diff --git a/src/common/options.cc b/src/common/options.cc
index e315b267a2a1..aebaea279622 100644
--- a/src/common/options.cc
+++ b/src/common/options.cc
@@ -8030,6 +8030,10 @@ std::vector<Option> get_mds_options() {
     .set_flag(Option::FLAG_RUNTIME)
     .set_description("set the maximum length of alternate names for dentries"),
 
+    Option("mds_valgrind_exit", Option::TYPE_BOOL, Option::LEVEL_DEV)
+    .set_default(false)
+    .set_flag(Option::FLAG_RUNTIME),
+
     Option("mds_numa_node", Option::TYPE_INT, Option::LEVEL_ADVANCED)
     .set_default(-1)
     .set_flag(Option::FLAG_STARTUP)
diff --git a/src/mds/MDSDaemon.cc b/src/mds/MDSDaemon.cc
index 85a70412f1d2..b18a6b757367 100644
--- a/src/mds/MDSDaemon.cc
+++ b/src/mds/MDSDaemon.cc
@@ -845,6 +845,11 @@ void MDSDaemon::respawn()
    * be removed from the MDSMap leading to respawn. */
   g_ceph_context->_log->dump_recent();
 
+  /* valgrind can't handle execve; just exit and let QA infra restart */
+  if (g_conf().get_val<bool>("mds_valgrind_exit")) {
+    _exit(0);
+  }
+
   char *new_argv[orig_argc+1];
   dout(1) << " e: '" << orig_argv[0] << "'" << dendl;
   for (int i=0; i<orig_argc; i++) {