mds: commit new dirs before first mention is trimmed from journal

src/include/xlist.h

@@ -33,6 +33,7 @@ public:
 
     
     xlist* get_xlist() { return _list; }
+    bool is_on_xlist() { return _list ? true:false; }
     void remove_myself() {
       if (_list) 
 	_list->remove(this);

src/mds/CDir.cc

@@ -137,7 +137,7 @@ ostream& CDir::print_db_line_prefix(ostream& out)
 // CDir
 
 CDir::CDir(CInode *in, frag_t fg, MDCache *mdcache, bool auth) :
-  xlist_dirty(this)
+  xlist_dirty(this), xlist_new(this)
 {
   inode = in;
   frag = fg;
@@ -874,8 +874,13 @@ void CDir::_mark_dirty(LogSegment *ls)
   } else {
     dout(10) << "mark_dirty (already dirty) " << *this << " version " << get_version() << dendl;
   }
-  if (ls) 
+  if (ls) {
     ls->dirty_dirfrags.push_back(&xlist_dirty);
+
+    // if i've never committed, i need to be before _any_ mention of me is trimmed from the journal.
+    if (committed_version == 0 && !xlist_new.is_on_xlist())
+      ls->new_dirfrags.push_back(&xlist_dirty);
+  }
 }
 
 void CDir::mark_clean()
@@ -886,6 +891,7 @@ void CDir::mark_clean()
     put(PIN_DIRTY);
 
     xlist_dirty.remove_myself();
+    xlist_new.remove_myself();
   }
 }
 

src/mds/CDir.h

@@ -159,7 +159,7 @@ protected:
   version_t projected_version;
   list<fnode_t*> projected_fnode;
 
-  xlist<CDir*>::item xlist_dirty;
+  xlist<CDir*>::item xlist_dirty, xlist_new;
 
 
 public:

src/mds/Locker.cc

@@ -2402,7 +2402,7 @@ void Locker::scatter_eval(ScatterLock *lock)
 void Locker::mark_updated_scatterlock(ScatterLock *lock)
 {
   lock->set_updated();
-  if (lock->xlistitem_updated.get_xlist()) {
+  if (lock->xlistitem_updated.is_on_xlist()) {
     dout(10) << "mark_updated_scatterlock " << *lock
 	     << " -- already on list since " << lock->update_stamp << dendl;
   } else {

src/mds/LogSegment.h

@@ -36,7 +36,7 @@ class LogSegment {
   loff_t trimmable_at;
 
   // dirty items
-  xlist<CDir*>    dirty_dirfrags;
+  xlist<CDir*>    dirty_dirfrags, new_dirfrags;
   xlist<CInode*>  dirty_inodes;
   xlist<CDentry*> dirty_dentries;
 

src/mds/Server.cc

@@ -2055,20 +2055,6 @@ void Server::handle_client_mkdir(MDRequest *mdr)
   
   // log + wait
   mdlog->submit_entry(le, new C_MDS_mknod_finish(mds, mdr, dn, newi));
-
-  /* old export heuristic.  pbly need to reimplement this at some point.    
-  if (
-      diri->dir->is_auth() &&
-      diri->dir->is_rep() &&
-      newdir->is_auth() &&
-      !newdir->is_hashing()) {
-    int dest = rand() % mds->mdsmap->get_num_mds();
-    if (dest != whoami) {
-      dout(10) << "exporting new dir " << *newdir << " in replicated parent " << *diri->dir << dendl;
-      mdcache->migrator->export_dir(newdir, dest);
-    }
-  }
-  */
 }
 
 
@@ -4400,7 +4386,7 @@ void Server::_do_open(MDRequest *mdr, CInode *cur)
   reply_request(mdr, reply);
 
   // make sure this inode gets into the journal
-  if (cur->xlist_open_file.get_xlist() == 0) {
+  if (!cur->xlist_open_file.is_on_xlist()) {
     LogSegment *ls = mds->mdlog->get_current_segment();
     EOpen *le = new EOpen(mds->mdlog);
     le->add_clean_inode(cur);

src/mds/journal.cc

@@ -75,6 +75,10 @@ C_Gather *LogSegment::try_to_expire(MDS *mds)
   dout(6) << "LogSegment(" << offset << ").try_to_expire" << dendl;
 
   // commit dirs
+  for (xlist<CDir*>::iterator p = new_dirfrags.begin(); !p.end(); ++p) {
+    assert((*p)->is_auth());
+    commit.insert(*p);
+  }
   for (xlist<CDir*>::iterator p = dirty_dirfrags.begin(); !p.end(); ++p) {
     assert((*p)->is_auth());
     commit.insert(*p);