From: John Spray <john.spray@redhat.com>
Date: Mon, 16 Mar 2015 00:04:36 +0000 (+0000)
Subject: mds: add damaged() method
X-Git-Tag: v9.0.0~113^2~8
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=c9ff8b4439fcf45b4692a81b70b6e488cd2905c6;p=ceph.git

mds: add damaged() method

A hook to fire off one last DAMAGED beacon
to mon before respawning.

Signed-off-by: John Spray <john.spray@redhat.com>
---

diff --git a/src/common/config_opts.h b/src/common/config_opts.h
index 9441a9ec946f..43a3f20b0a16 100644
--- a/src/common/config_opts.h
+++ b/src/common/config_opts.h
@@ -449,6 +449,7 @@ OPTION(mds_snap_rstat, OPT_BOOL, false) // enable/disbale nested stat for snapsh
 OPTION(mds_verify_backtrace, OPT_U32, 1)
 
 OPTION(mds_action_on_write_error, OPT_U32, 1) // 0: ignore; 1: force readonly; 2: crash
+OPTION(mds_mon_shutdown_timeout, OPT_DOUBLE, 5)
 
 // If true, compact leveldb store on mount
 OPTION(osd_compact_leveldb_on_mount, OPT_BOOL, false)
diff --git a/src/mds/MDS.cc b/src/mds/MDS.cc
index a56d6ba177a1..cf042a951c72 100644
--- a/src/mds/MDS.cc
+++ b/src/mds/MDS.cc
@@ -2361,6 +2361,19 @@ void MDS::handle_signal(int signum)
   mds_lock.Unlock();
 }
 
+void MDS::damaged()
+{
+  set_want_state(MDSMap::STATE_DAMAGED);
+  beacon.notify_health(this);  // Include latest status in our swan song
+  beacon.send_and_wait(g_conf->mds_mon_shutdown_timeout);
+
+  // It's okay if we timed out and the mon didn't get our beacon, because
+  // another daemon (or ourselves after respawn) will eventually take the
+  // rank and report DAMAGED again when it hits same problem we did.
+
+  respawn();  // Respawn into standby in case mon has other work for us
+}
+
 void MDS::suicide()
 {
   assert(mds_lock.is_locked());
diff --git a/src/mds/MDS.h b/src/mds/MDS.h
index dca10c702494..62adf193a292 100644
--- a/src/mds/MDS.h
+++ b/src/mds/MDS.h
@@ -443,6 +443,14 @@ private:
   void handle_mds_recovery(mds_rank_t who);
   void handle_mds_failure(mds_rank_t who);
 
+  /**
+   * Report state DAMAGED to the mon, and then pass on to respawn().  Call
+   * this when an unrecoverable error is encountered while attempting
+   * to load an MDS rank's data structures.  This is *not* for use with
+   * errors affecting normal dirfrag/inode objects -- they should be handled
+   * through cleaner scrub/repair mechanisms.
+   */
+  void damaged();
   void suicide();
   void respawn();
   void handle_write_error(int err);