mds: avoid race between cache expire and MDentryLink

commit 22535340 tried to fix race between cache expire and
MDentryLink. It avoids trimming null dentry whose lock is
not readable. The fix does not handle the case that MDS
first recevies a MDentryUnlink message, then receives a
MDentryLink message.

Signed-off-by: "Yan, Zheng" <zyan@redhat.com>

diff --git a/src/mds/MDCache.cc b/src/mds/MDCache.cc
index 38177c60296699462cc3d40aa186a41d306e0b1d..6aa92064e1ef8c99b56d78981dbaa74e10c646de 100644 (file)
--- a/src/mds/MDCache.cc
+++ b/src/mds/MDCache.cc
@@ -6565,11 +6565,20 @@ bool MDCache::trim_dentry(CDentry *dn, map<mds_rank_t, MCacheExpire*>& expiremap
     assert(dn->is_auth());
   }
 
+  // If replica dentry is not readable, it's likely we will receive
+  // MDentryLink/MDentryUnlink message soon (It's possible we first
+  // receive a MDentryUnlink message, then MDentryLink message)
+  // MDentryLink message only replicates an inode, so we should
+  // avoid trimming the inode's parent dentry. This is because that
+  // unconnected replicas are problematic for subtree migration.
+  if (!dn->is_auth() && !dn->lock.can_read(-1) &&
+      !dn->get_dir()->get_inode()->is_stray())
+    return true;
+
   // adjust the dir state
   // NOTE: we can safely remove a clean, null dentry without effecting
   //       directory completeness.
   // (check this _before_ we unlink the inode, below!)
-  bool null_dentry = false;
   bool clear_complete = false;
   if (!(dnl->is_null() && dn->is_clean()))
     clear_complete = true;
@@ -6586,7 +6595,6 @@ bool MDCache::trim_dentry(CDentry *dn, map<mds_rank_t, MCacheExpire*>& expiremap
       return true; // purging stray instead of trimming
   } else {
     assert(dnl->is_null());
-    null_dentry = true;
   }
 
   if (dn->is_auth()) {
@@ -6595,16 +6603,6 @@ bool MDCache::trim_dentry(CDentry *dn, map<mds_rank_t, MCacheExpire*>& expiremap
     }
   } else {
     // notify dentry authority.
-
-    // If null replica dentry is not readable, it's likely we will
-    // receive a MDentryLink message soon. MDentryLink message only
-    // replicates an inode, so we should avoid trimming the inode's
-    // parent dentry. This is because that unconnected replicas are
-    // problematic for subtree migration.
-    if (null_dentry && !dn->lock.can_read(-1) &&
-       !dn->get_dir()->get_inode()->is_stray())
-      return true;
-
     mds_authority_t auth = dn->authority();
     
     for (int p=0; p<2; p++) {