From: Patrick Donnelly <pdonnell@redhat.com>
Date: Tue, 25 Jul 2023 19:07:21 +0000 (-0400)
Subject: mds: flush monc log before abort
X-Git-Tag: v18.2.1~110^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=d8db3fc5e818046fea864bcb598890c8b9d2c660;p=ceph.git

mds: flush monc log before abort

These are asynchronously flushed periodically. If the MDS plans to abort (and
we know messages are pending), it needs to manually trigger a flush.

Signed-off-by: Patrick Donnelly <pdonnell@redhat.com>
(cherry picked from commit 59c5a50bd9e69d58b4cdb9eff68b55fb2ccb4f29)
---

diff --git a/src/mds/CDentry.cc b/src/mds/CDentry.cc
index 6f20f53d144d..eda37ccb4df7 100644
--- a/src/mds/CDentry.cc
+++ b/src/mds/CDentry.cc
@@ -716,7 +716,7 @@ bool CDentry::check_corruption(bool load)
     }
     if (!load && g_conf().get_val<bool>("mds_abort_on_newly_corrupt_dentry")) {
       dir->mdcache->mds->clog->error() << "MDS abort because newly corrupt dentry to be committed: " << *this;
-      ceph_abort("detected newly corrupt dentry"); /* avoid writing out newly corrupted dn */
+      dir->mdcache->mds->abort("detected newly corrupt dentry"); /* avoid writing out newly corrupted dn */
     }
     return true;
   }
diff --git a/src/mds/MDSRank.cc b/src/mds/MDSRank.cc
index 2bb5b7d7bb8a..13f8e015f3da 100644
--- a/src/mds/MDSRank.cc
+++ b/src/mds/MDSRank.cc
@@ -931,6 +931,12 @@ void MDSRank::respawn()
   }
 }
 
+void MDSRank::abort(std::string_view msg)
+{
+  monc->flush_log();
+  ceph_abort(msg);
+}
+
 void MDSRank::damaged()
 {
   ceph_assert(whoami != MDS_RANK_NONE);
diff --git a/src/mds/MDSRank.h b/src/mds/MDSRank.h
index c52c837083b9..ca61a10b4b12 100644
--- a/src/mds/MDSRank.h
+++ b/src/mds/MDSRank.h
@@ -295,6 +295,13 @@ class MDSRank {
       return count * _heartbeat_reset_grace;
     }
 
+    /**
+     * Abort the MDS and flush any clog messages.
+     *
+     * Callers must already hold mds_lock.
+     */
+    void abort(std::string_view msg);
+
     /**
      * Report state DAMAGED to the mon, and then pass on to respawn().  Call
      * this when an unrecoverable error is encountered while attempting