From ed45b7d600b64d065bcb393edb69cfca1a0c9655 Mon Sep 17 00:00:00 2001
From: sageweil <sageweil@29311d96-e01e-0410-9327-a35deaab8ce9>
Date: Fri, 30 Nov 2007 19:52:48 +0000
Subject: [PATCH] first pass at adding pobject abstraction

git-svn-id: https://ceph.svn.sf.net/svnroot/ceph@2161 29311d96-e01e-0410-9327-a35deaab8ce9
---
 branches/ebofs/config.cc           |   4 +-
 branches/ebofs/dupstore.cc         |   8 +-
 branches/ebofs/ebofs/BufferCache.h |   6 +-
 branches/ebofs/ebofs/Ebofs.cc      | 168 ++++++++++++++---------------
 branches/ebofs/ebofs/Ebofs.h       | 110 +++++++++----------
 branches/ebofs/ebofs/Onode.h       |   4 +-
 branches/ebofs/ebofs/types.h       |   3 +-
 branches/ebofs/include/object.h    |  33 ++----
 branches/ebofs/include/pobject.h   |  68 ++++++++++++
 branches/ebofs/mds/mdstypes.h      |   2 +-
 branches/ebofs/osd/Fake.h          |  26 ++---
 branches/ebofs/osd/FakeStore.cc    | 102 ++++++++++--------
 branches/ebofs/osd/FakeStore.h     |  40 +++----
 branches/ebofs/osd/OSD.cc          |   4 +-
 branches/ebofs/osd/ObjectStore.h   | 111 +++++++++----------
 branches/ebofs/osd/PG.cc           |  15 +--
 branches/ebofs/osd/ReplicatedPG.cc |  15 +--
 branches/ebofs/osd/osd_types.h     |   9 +-
 18 files changed, 402 insertions(+), 326 deletions(-)
 create mode 100644 branches/ebofs/include/pobject.h

diff --git a/branches/ebofs/config.cc b/branches/ebofs/config.cc
index 3dd7dfb79f2a9..b4c108d0b53eb 100644
--- a/branches/ebofs/config.cc
+++ b/branches/ebofs/config.cc
@@ -331,8 +331,8 @@ md_config_t g_conf = {
   ebofs_idle_commit_ms: 0,         // 0 = no idle detection.  UGLY HACK.  use bdev_idle_kick_after_ms instead.
   ebofs_oc_size:        10000,      // onode cache
   ebofs_cc_size:        10000,      // cnode cache
-  ebofs_bc_size:        (50 *256), // 4k blocks, *256 for MB
-  ebofs_bc_max_dirty:   (30 *256), // before write() will block
+  ebofs_bc_size:        (5 *256), // 4k blocks, *256 for MB
+  ebofs_bc_max_dirty:   (3 *256), // before write() will block
   ebofs_max_prefetch: 1000, // 4k blocks
   ebofs_realloc: false,    // hrm, this can cause bad fragmentation, don't use!
   ebofs_verify_csum_on_read: true,
diff --git a/branches/ebofs/dupstore.cc b/branches/ebofs/dupstore.cc
index d43f935cb50cc..0b3b6fa6eb8c2 100644
--- a/branches/ebofs/dupstore.cc
+++ b/branches/ebofs/dupstore.cc
@@ -24,12 +24,12 @@ int dupstore(ObjectStore* src, ObjectStore* dst)
   if (dst->mount() < 0) return 1;
 
   // objects
-  list<object_t> objects;
+  list<pobject_t> objects;
   src->list_objects(objects);
   int num = objects.size();
   cout << num << " objects" << std::endl;
   int i = 1;
-  for (list<object_t>::iterator p = objects.begin(); p != objects.end(); ++p) {
+  for (list<pobject_t>::iterator p = objects.begin(); p != objects.end(); ++p) {
     bufferlist bl;
     src->read(*p, 0, 0, bl);
     cout << "object " << i++ << "/" << num << " " << *p << " = " << bl.length() << " bytes" << std::endl;
@@ -52,10 +52,10 @@ int dupstore(ObjectStore* src, ObjectStore* dst)
     map<string,bufferptr> attrs;
     src->collection_getattrs(*p, attrs);
     dst->collection_setattrs(*p, attrs);
-    list<object_t> o;
+    list<pobject_t> o;
     src->collection_list(*p, o);
     int numo = 0;
-    for (list<object_t>::iterator q = o.begin(); q != o.end(); q++) {
+    for (list<pobject_t>::iterator q = o.begin(); q != o.end(); q++) {
       dst->collection_add(*p, *q, 0);
       numo++;
     }
diff --git a/branches/ebofs/ebofs/BufferCache.h b/branches/ebofs/ebofs/BufferCache.h
index 31575c76f6bfa..fc8424057f1a0 100644
--- a/branches/ebofs/ebofs/BufferCache.h
+++ b/branches/ebofs/ebofs/BufferCache.h
@@ -274,7 +274,7 @@ inline ostream& operator<<(ostream& out, BufferHead& bh)
 
 class ObjectCache {
  public:
-  object_t object_id;
+  pobject_t object_id;
   Onode *on;
   BufferCache *bc;
 
@@ -287,7 +287,7 @@ class ObjectCache {
 
 
  public:
-  ObjectCache(object_t o, Onode *_on, BufferCache *b) : 
+  ObjectCache(pobject_t o, Onode *_on, BufferCache *b) : 
     object_id(o), on(_on), bc(b), ref(0),
     write_count(0) { }
   ~ObjectCache() {
@@ -307,7 +307,7 @@ class ObjectCache {
     return ref; 
   }
   
-  object_t get_object_id() { return object_id; }
+  pobject_t get_object_id() { return object_id; }
 
   void add_bh(BufferHead *bh) {
     // add to my map
diff --git a/branches/ebofs/ebofs/Ebofs.cc b/branches/ebofs/ebofs/Ebofs.cc
index 1e2101f6ebae4..504d8df749d8f 100644
--- a/branches/ebofs/ebofs/Ebofs.cc
+++ b/branches/ebofs/ebofs/Ebofs.cc
@@ -106,14 +106,14 @@ int Ebofs::mount()
   
   // open tables
   dout(3) << "mount opening tables" << dendl;
-  object_tab = new Table<object_t, Extent>( nodepool, sb->object_tab );
+  object_tab = new Table<pobject_t, Extent>( nodepool, sb->object_tab );
   for (int i=0; i<EBOFS_NUM_FREE_BUCKETS; i++)
     free_tab[i] = new Table<block_t, block_t>( nodepool, sb->free_tab[i] );
   limbo_tab = new Table<block_t, block_t>( nodepool, sb->limbo_tab );
   alloc_tab = new Table<block_t, pair<block_t,int> >( nodepool, sb->alloc_tab );
   
   collection_tab = new Table<coll_t, Extent>( nodepool, sb->collection_tab );
-  co_tab = new Table<coll_object_t, bool>( nodepool, sb->co_tab );
+  co_tab = new Table<coll_pobject_t, bool>( nodepool, sb->co_tab );
 
   verify_tables();
 
@@ -223,7 +223,7 @@ int Ebofs::mkfs()
   empty.root.csum = 0;
   empty.depth = 0;
   
-  object_tab = new Table<object_t, Extent>( nodepool, empty );
+  object_tab = new Table<pobject_t, Extent>( nodepool, empty );
   collection_tab = new Table<coll_t, Extent>( nodepool, empty );
   
   for (int i=0; i<EBOFS_NUM_FREE_BUCKETS; i++)
@@ -231,7 +231,7 @@ int Ebofs::mkfs()
   limbo_tab = new Table<block_t,block_t>( nodepool, empty );
   alloc_tab = new Table<block_t,pair<block_t,int> >( nodepool, empty );
   
-  co_tab = new Table<coll_object_t, bool>( nodepool, empty );
+  co_tab = new Table<coll_pobject_t, bool>( nodepool, empty );
 
   // add free space
   Extent left;
@@ -339,7 +339,7 @@ int Ebofs::umount()
   trim_bc(0);
   trim_inodes(0);
 
-  for (hash_map<object_t,Onode*>::iterator i = onode_map.begin();
+  for (hash_map<pobject_t,Onode*>::iterator i = onode_map.begin();
        i != onode_map.end();
        i++) {
     dout(0) << "umount *** leftover: " << i->first << "   " << *(i->second) << dendl;
@@ -661,7 +661,7 @@ void *Ebofs::finisher_thread_entry()
 
 // *** onodes ***
 
-Onode* Ebofs::new_onode(object_t oid)
+Onode* Ebofs::new_onode(pobject_t oid)
 {
   Onode* on = new Onode(oid);
 
@@ -683,7 +683,7 @@ Onode* Ebofs::new_onode(object_t oid)
 }
 
 
-Onode* Ebofs::get_onode(object_t oid)
+Onode* Ebofs::get_onode(pobject_t oid)
 {
   while (1) {
     // in cache?
@@ -954,7 +954,7 @@ void Ebofs::remove_onode(Onode *on)
   for (set<coll_t>::iterator i = on->collections.begin();
        i != on->collections.end();
        i++) {
-    co_tab->remove(coll_object_t(*i,on->object_id));
+    co_tab->remove(coll_pobject_t(*i,on->object_id));
   }
   on->collections.clear();
 
@@ -2154,7 +2154,7 @@ int Ebofs::attempt_read(Onode *on, off_t off, size_t len, bufferlist& bl,
  * return value of -1 if onode isn't loaded.  otherwise, the number
  * of extents that need to be read (i.e. # of seeks)  
  */
-int Ebofs::is_cached(object_t oid, off_t off, size_t len)
+int Ebofs::is_cached(pobject_t oid, off_t off, size_t len)
 {
   ebofs_lock.Lock();
   int r = _is_cached(oid, off, len);
@@ -2162,7 +2162,7 @@ int Ebofs::is_cached(object_t oid, off_t off, size_t len)
   return r;
 }
 
-int Ebofs::_is_cached(object_t oid, off_t off, size_t len)
+int Ebofs::_is_cached(pobject_t oid, off_t off, size_t len)
 {
   if (!have_onode(oid)) {
     dout(7) << "_is_cached " << oid << " " << off << "~" << len << " ... onode  " << dendl;
@@ -2205,14 +2205,14 @@ int Ebofs::_is_cached(object_t oid, off_t off, size_t len)
   */
 }
 
-void Ebofs::trim_from_cache(object_t oid, off_t off, size_t len)
+void Ebofs::trim_from_cache(pobject_t oid, off_t off, size_t len)
 {
   ebofs_lock.Lock();
   _trim_from_cache(oid, off, len);
   ebofs_lock.Unlock();
 }
 
-void Ebofs::_trim_from_cache(object_t oid, off_t off, size_t len)
+void Ebofs::_trim_from_cache(pobject_t oid, off_t off, size_t len)
 {
   // be careful not to load it if we don't have it
   if (!have_onode(oid)) {
@@ -2237,7 +2237,7 @@ void Ebofs::_trim_from_cache(object_t oid, off_t off, size_t len)
 }
 
 
-int Ebofs::read(object_t oid, 
+int Ebofs::read(pobject_t oid, 
                 off_t off, size_t len,
                 bufferlist& bl)
 {
@@ -2247,7 +2247,7 @@ int Ebofs::read(object_t oid,
   return r;
 }
 
-int Ebofs::_read(object_t oid, off_t off, size_t len, bufferlist& bl)
+int Ebofs::_read(pobject_t oid, off_t off, size_t len, bufferlist& bl)
 {
   dout(7) << "_read " << oid << " " << off << "~" << len << dendl;
 
@@ -2348,7 +2348,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
     switch (op) {
     case Transaction::OP_READ:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
         off_t offset, len;
 	t.get_length(offset);
@@ -2364,7 +2364,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
 
     case Transaction::OP_STAT:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
         struct stat *st;
 	t.get_pstat(st);
@@ -2377,7 +2377,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
 
     case Transaction::OP_GETATTR:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
 	const char *attrname;
 	t.get_attrname(attrname);
@@ -2392,7 +2392,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
 
     case Transaction::OP_GETATTRS:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
         map<string,bufferptr> *pset;
 	t.get_pattrset(pset);
@@ -2406,7 +2406,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
 
     case Transaction::OP_WRITE:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
         off_t offset, len;
 	t.get_length(offset);
@@ -2422,7 +2422,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
 
     case Transaction::OP_TRIMCACHE:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
         off_t offset, len;
 	t.get_length(offset);
@@ -2433,7 +2433,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
 
     case Transaction::OP_TRUNCATE:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
         off_t len;
 	t.get_length(len);
@@ -2446,7 +2446,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
 
     case Transaction::OP_REMOVE:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
         if (_remove(oid) < 0) {
           dout(7) << "apply_transaction fail on _remove" << dendl;
@@ -2457,7 +2457,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
       
     case Transaction::OP_SETATTR:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
 	const char *attrname;
 	t.get_attrname(attrname);
@@ -2472,7 +2472,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
 
     case Transaction::OP_SETATTRS:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
         map<string,bufferptr> *pattrset;
 	t.get_pattrset(pattrset);
@@ -2485,7 +2485,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
       
     case Transaction::OP_RMATTR:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
 	const char *attrname;
 	t.get_attrname(attrname);
@@ -2498,9 +2498,9 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
 
     case Transaction::OP_CLONE:
       {
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
-        object_t noid;
+        pobject_t noid;
 	t.get_oid(noid);
 	if (_clone(oid, noid) < 0) {
 	  dout(7) << "apply_transaction fail on _clone" << dendl;
@@ -2535,7 +2535,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
       {
         coll_t cid;
 	t.get_cid(cid);
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
         if (_collection_add(cid, oid) < 0) {
           //dout(7) << "apply_transaction fail on _collection_add" << dendl;
@@ -2548,7 +2548,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
       {
         coll_t cid;
 	t.get_cid(cid);
-        object_t oid;
+        pobject_t oid;
 	t.get_oid(oid);
         if (_collection_remove(cid, oid) < 0) {
           dout(7) << "apply_transaction fail on _collection_remove" << dendl;
@@ -2600,7 +2600,7 @@ unsigned Ebofs::_apply_transaction(Transaction& t)
 
 
 
-int Ebofs::_write(object_t oid, off_t offset, size_t length, const bufferlist& bl)
+int Ebofs::_write(pobject_t oid, off_t offset, size_t length, const bufferlist& bl)
 {
   dout(7) << "_write " << oid << " " << offset << "~" << length << dendl;
   assert(bl.length() == length);
@@ -2662,7 +2662,7 @@ int Ebofs::_write(object_t oid, off_t offset, size_t length, const bufferlist& b
   return length;
 }
 
-int Ebofs::_zero(object_t oid, off_t offset, size_t length)
+int Ebofs::_zero(pobject_t oid, off_t offset, size_t length)
 {
   dout(7) << "_zero " << oid << " " << offset << "~" << length << dendl;
 
@@ -2692,7 +2692,7 @@ int Ebofs::_zero(object_t oid, off_t offset, size_t length)
 }
 
 
-int Ebofs::write(object_t oid, 
+int Ebofs::write(pobject_t oid, 
                  off_t off, size_t len,
                  const bufferlist& bl, Context *onsafe)
 {
@@ -2723,7 +2723,7 @@ int Ebofs::write(object_t oid,
   return r;
 }
 
-int Ebofs::zero(object_t oid, off_t off, size_t len, Context *onsafe)
+int Ebofs::zero(pobject_t oid, off_t off, size_t len, Context *onsafe)
 {
   ebofs_lock.Lock();
   
@@ -2753,7 +2753,7 @@ int Ebofs::zero(object_t oid, off_t off, size_t len, Context *onsafe)
 }
 
 
-int Ebofs::_remove(object_t oid)
+int Ebofs::_remove(pobject_t oid)
 {
   dout(7) << "_remove " << oid << dendl;
 
@@ -2768,7 +2768,7 @@ int Ebofs::_remove(object_t oid)
 }
 
 
-int Ebofs::remove(object_t oid, Context *onsafe)
+int Ebofs::remove(pobject_t oid, Context *onsafe)
 {
   ebofs_lock.Lock();
 
@@ -2796,7 +2796,7 @@ int Ebofs::remove(object_t oid, Context *onsafe)
   return r;
 }
 
-int Ebofs::_truncate(object_t oid, off_t size)
+int Ebofs::_truncate(pobject_t oid, off_t size)
 {
   dout(7) << "_truncate " << oid << " size " << size << dendl;
 
@@ -2855,7 +2855,7 @@ int Ebofs::_truncate(object_t oid, off_t size)
 }
 
 
-int Ebofs::truncate(object_t oid, off_t size, Context *onsafe)
+int Ebofs::truncate(pobject_t oid, off_t size, Context *onsafe)
 {
   ebofs_lock.Lock();
   
@@ -2884,7 +2884,7 @@ int Ebofs::truncate(object_t oid, off_t size, Context *onsafe)
 
 
 
-int Ebofs::clone(object_t from, object_t to, Context *onsafe)
+int Ebofs::clone(pobject_t from, pobject_t to, Context *onsafe)
 {
   ebofs_lock.Lock();
   
@@ -2911,7 +2911,7 @@ int Ebofs::clone(object_t from, object_t to, Context *onsafe)
   return r;
 }
 
-int Ebofs::_clone(object_t from, object_t to)
+int Ebofs::_clone(pobject_t from, pobject_t to)
 {
   dout(7) << "_clone " << from << " -> " << to << dendl;
 
@@ -2971,34 +2971,34 @@ int Ebofs::_clone(object_t from, object_t to)
  *  (oid.rev is a noninclusive upper bound.)
  *
  */
-int Ebofs::pick_object_revision_lt(object_t& oid)
+int Ebofs::pick_object_revision_lt(pobject_t& oid)
 {
-  assert(oid.rev > 0);   // this is only useful for non-zero oid.rev
+  assert(oid.oid.rev > 0);   // this is only useful for non-zero oid.rev
 
   int r = -EEXIST;             // return code
   ebofs_lock.Lock();
   {
-    object_t orig = oid;
-    object_t live = oid;
-    live.rev = 0;
+    pobject_t orig = oid;
+    pobject_t live = oid;
+    live.oid.rev = 0;
     
     if (object_tab->get_num_keys() > 0) {
-      Table<object_t, Extent>::Cursor cursor(object_tab);
+      Table<pobject_t, Extent>::Cursor cursor(object_tab);
       
       object_tab->find(oid, cursor);  // this will be just _past_ highest eligible rev
       if (cursor.move_left() > 0) {
 	bool firstpass = true;
 	while (1) {
-	  object_t t = cursor.current().key;
-	  if (t.ino != oid.ino || 
-	      t.bno != oid.bno)                 // passed to previous object
+	  pobject_t t = cursor.current().key;
+	  if (t.oid.ino != oid.oid.ino || 
+	      t.oid.bno != oid.oid.bno)                 // passed to previous object
 	    break;
-	  if (oid.rev < t.rev) {                // rev < desired.  possible match.
+	  if (oid.oid.rev < t.oid.rev) {                // rev < desired.  possible match.
 	    r = 0;
 	    oid = t;
 	    break;
 	  }
-	  if (firstpass && oid.rev >= t.rev) {  // there is no old rev < desired.  try live.
+	  if (firstpass && oid.oid.rev >= t.oid.rev) {  // there is no old rev < desired.  try live.
 	    r = 0;
 	    oid = live;
 	    break;
@@ -3019,7 +3019,7 @@ int Ebofs::pick_object_revision_lt(object_t& oid)
 
 
 
-bool Ebofs::exists(object_t oid)
+bool Ebofs::exists(pobject_t oid)
 {
   ebofs_lock.Lock();
   dout(8) << "exists " << oid << dendl;
@@ -3028,7 +3028,7 @@ bool Ebofs::exists(object_t oid)
   return e;
 }
 
-int Ebofs::stat(object_t oid, struct stat *st)
+int Ebofs::stat(pobject_t oid, struct stat *st)
 {
   ebofs_lock.Lock();
   int r = _stat(oid,st);
@@ -3036,7 +3036,7 @@ int Ebofs::stat(object_t oid, struct stat *st)
   return r;
 }
 
-int Ebofs::_stat(object_t oid, struct stat *st)
+int Ebofs::_stat(pobject_t oid, struct stat *st)
 {
   dout(7) << "_stat " << oid << dendl;
 
@@ -3051,7 +3051,7 @@ int Ebofs::_stat(object_t oid, struct stat *st)
 }
 
 
-int Ebofs::_setattr(object_t oid, const char *name, const void *value, size_t size) 
+int Ebofs::_setattr(pobject_t oid, const char *name, const void *value, size_t size) 
 {
   dout(8) << "setattr " << oid << " '" << name << "' len " << size << dendl;
 
@@ -3072,7 +3072,7 @@ int Ebofs::_setattr(object_t oid, const char *name, const void *value, size_t si
   return 0;
 }
 
-int Ebofs::setattr(object_t oid, const char *name, const void *value, size_t size, Context *onsafe)
+int Ebofs::setattr(pobject_t oid, const char *name, const void *value, size_t size, Context *onsafe)
 {
   ebofs_lock.Lock();
   int r = _setattr(oid, name, value, size);
@@ -3098,7 +3098,7 @@ int Ebofs::setattr(object_t oid, const char *name, const void *value, size_t siz
   return r;
 }
 
-int Ebofs::_setattrs(object_t oid, map<string,bufferptr>& attrset)
+int Ebofs::_setattrs(pobject_t oid, map<string,bufferptr>& attrset)
 {
   dout(8) << "setattrs " << oid << dendl;
 
@@ -3115,7 +3115,7 @@ int Ebofs::_setattrs(object_t oid, map<string,bufferptr>& attrset)
   return 0;
 }
 
-int Ebofs::setattrs(object_t oid, map<string,bufferptr>& attrset, Context *onsafe)
+int Ebofs::setattrs(pobject_t oid, map<string,bufferptr>& attrset, Context *onsafe)
 {
   ebofs_lock.Lock();
   int r = _setattrs(oid, attrset);
@@ -3142,7 +3142,7 @@ int Ebofs::setattrs(object_t oid, map<string,bufferptr>& attrset, Context *onsaf
 }
 
 
-int Ebofs::get_object_collections(object_t oid, set<coll_t>& ls)
+int Ebofs::get_object_collections(pobject_t oid, set<coll_t>& ls)
 {
   ebofs_lock.Lock();
   int r = _get_object_collections(oid, ls);
@@ -3150,7 +3150,7 @@ int Ebofs::get_object_collections(object_t oid, set<coll_t>& ls)
   return r;
 }
 
-int Ebofs::_get_object_collections(object_t oid, set<coll_t>& ls)
+int Ebofs::_get_object_collections(pobject_t oid, set<coll_t>& ls)
 {
   dout(8) << "_get_object_collections " << oid << dendl;
 
@@ -3161,7 +3161,7 @@ int Ebofs::_get_object_collections(object_t oid, set<coll_t>& ls)
   return 0;
 }
 
-int Ebofs::getattr(object_t oid, const char *name, void *value, size_t size)
+int Ebofs::getattr(pobject_t oid, const char *name, void *value, size_t size)
 {
   ebofs_lock.Lock();
   int r = _getattr(oid, name, value, size);
@@ -3169,7 +3169,7 @@ int Ebofs::getattr(object_t oid, const char *name, void *value, size_t size)
   return r;
 }
 
-int Ebofs::_getattr(object_t oid, const char *name, void *value, size_t size)
+int Ebofs::_getattr(pobject_t oid, const char *name, void *value, size_t size)
 {
   dout(8) << "_getattr " << oid << " '" << name << "' maxlen " << size << dendl;
 
@@ -3190,7 +3190,7 @@ int Ebofs::_getattr(object_t oid, const char *name, void *value, size_t size)
   return r;
 }
 
-int Ebofs::getattrs(object_t oid, map<string,bufferptr> &aset)
+int Ebofs::getattrs(pobject_t oid, map<string,bufferptr> &aset)
 {
   ebofs_lock.Lock();
   int r = _getattrs(oid, aset);
@@ -3198,7 +3198,7 @@ int Ebofs::getattrs(object_t oid, map<string,bufferptr> &aset)
   return r;
 }
 
-int Ebofs::_getattrs(object_t oid, map<string,bufferptr> &aset)
+int Ebofs::_getattrs(pobject_t oid, map<string,bufferptr> &aset)
 {
   dout(8) << "_getattrs " << oid << dendl;
 
@@ -3211,7 +3211,7 @@ int Ebofs::_getattrs(object_t oid, map<string,bufferptr> &aset)
 
 
 
-int Ebofs::_rmattr(object_t oid, const char *name) 
+int Ebofs::_rmattr(pobject_t oid, const char *name) 
 {
   dout(8) << "_rmattr " << oid << " '" << name << "'" << dendl;
 
@@ -3229,7 +3229,7 @@ int Ebofs::_rmattr(object_t oid, const char *name)
   return 0;
 }
 
-int Ebofs::rmattr(object_t oid, const char *name, Context *onsafe) 
+int Ebofs::rmattr(pobject_t oid, const char *name, Context *onsafe) 
 {
   ebofs_lock.Lock();
 
@@ -3256,7 +3256,7 @@ int Ebofs::rmattr(object_t oid, const char *name, Context *onsafe)
   return r;
 }
 
-int Ebofs::listattr(object_t oid, vector<string>& attrs)
+int Ebofs::listattr(pobject_t oid, vector<string>& attrs)
 {
   ebofs_lock.Lock();
   dout(8) << "listattr " << oid << dendl;
@@ -3279,15 +3279,15 @@ int Ebofs::listattr(object_t oid, vector<string>& attrs)
   return 0;
 }
 
-int Ebofs::list_objects(list<object_t>& ls)
+int Ebofs::list_objects(list<pobject_t>& ls)
 {
   ebofs_lock.Lock();
   dout(9) << "list_objects " << dendl;
 
-  Table<object_t, Extent>::Cursor cursor(object_tab);
+  Table<pobject_t, Extent>::Cursor cursor(object_tab);
 
   int num = 0;
-  if (object_tab->find(object_t(), cursor) >= 0) {
+  if (object_tab->find(pobject_t(), cursor) >= 0) {
     while (1) {
       ls.push_back(cursor.current().key);
       num++;
@@ -3373,12 +3373,12 @@ int Ebofs::_destroy_collection(coll_t cid)
   assert(cn);
 
   // hose mappings
-  list<object_t> objects;
+  list<pobject_t> objects;
   collection_list(cid, objects);
-  for (list<object_t>::iterator i = objects.begin(); 
+  for (list<pobject_t>::iterator i = objects.begin(); 
        i != objects.end();
        i++) {
-    co_tab->remove(coll_object_t(cid,*i));
+    co_tab->remove(coll_pobject_t(cid,*i));
 
     Onode *on = get_onode(*i);
     if (on) {
@@ -3432,7 +3432,7 @@ bool Ebofs::_collection_exists(coll_t cid)
   return (collection_tab->lookup(cid) == 0);
 }
 
-int Ebofs::_collection_add(coll_t cid, object_t oid)
+int Ebofs::_collection_add(coll_t cid, pobject_t oid)
 {
   dout(9) << "_collection_add " << hex << cid << " object " << oid << dec << dendl;
 
@@ -3447,7 +3447,7 @@ int Ebofs::_collection_add(coll_t cid, object_t oid)
   if (on->collections.count(cid) == 0) {
     on->collections.insert(cid);
     dirty_onode(on);
-    co_tab->insert(coll_object_t(cid,oid), true);
+    co_tab->insert(coll_pobject_t(cid,oid), true);
   } else {
     r = -ENOENT;  // FIXME?  already in collection.
   }
@@ -3456,7 +3456,7 @@ int Ebofs::_collection_add(coll_t cid, object_t oid)
   return r;
 }
 
-int Ebofs::collection_add(coll_t cid, object_t oid, Context *onsafe)
+int Ebofs::collection_add(coll_t cid, pobject_t oid, Context *onsafe)
 {
   ebofs_lock.Lock();
 
@@ -3483,7 +3483,7 @@ int Ebofs::collection_add(coll_t cid, object_t oid, Context *onsafe)
   return 0;
 }
 
-int Ebofs::_collection_remove(coll_t cid, object_t oid)
+int Ebofs::_collection_remove(coll_t cid, pobject_t oid)
 {
   dout(9) << "_collection_remove " << hex << cid << " object " << oid << dec << dendl;
 
@@ -3498,7 +3498,7 @@ int Ebofs::_collection_remove(coll_t cid, object_t oid)
   if (on->collections.count(cid)) {
     on->collections.erase(cid);
     dirty_onode(on);
-    co_tab->remove(coll_object_t(cid,oid));
+    co_tab->remove(coll_pobject_t(cid,oid));
   } else {
     r = -ENOENT;  // FIXME?
   } 
@@ -3507,7 +3507,7 @@ int Ebofs::_collection_remove(coll_t cid, object_t oid)
   return r;
 }
 
-int Ebofs::collection_remove(coll_t cid, object_t oid, Context *onsafe)
+int Ebofs::collection_remove(coll_t cid, pobject_t oid, Context *onsafe)
 {
   ebofs_lock.Lock();
 
@@ -3534,7 +3534,7 @@ int Ebofs::collection_remove(coll_t cid, object_t oid, Context *onsafe)
   return 0;
 }
 
-int Ebofs::collection_list(coll_t cid, list<object_t>& ls)
+int Ebofs::collection_list(coll_t cid, list<pobject_t>& ls)
 {
   ebofs_lock.Lock();
   dout(9) << "collection_list " << hex << cid << dec << dendl;
@@ -3544,13 +3544,13 @@ int Ebofs::collection_list(coll_t cid, list<object_t>& ls)
     return -ENOENT;
   }
   
-  Table<coll_object_t, bool>::Cursor cursor(co_tab);
+  Table<coll_pobject_t, bool>::Cursor cursor(co_tab);
 
   int num = 0;
-  if (co_tab->find(coll_object_t(cid,object_t()), cursor) >= 0) {
+  if (co_tab->find(coll_pobject_t(cid,pobject_t()), cursor) >= 0) {
     while (1) {
       const coll_t c = cursor.current().key.first;
-      const object_t o = cursor.current().key.second;
+      const pobject_t o = cursor.current().key.second;
       if (c != cid) break;   // end!
       dout(10) << "collection_list  " << hex << cid << " includes " << o << dec << dendl;
       ls.push_back(o);
@@ -3835,8 +3835,8 @@ void Ebofs::_get_frag_stat(FragmentationStat& st)
   st.avg_extent_per_object = 0;
   st.avg_extent_jump = 0;
 
-  Table<object_t,Extent>::Cursor cursor(object_tab);
-  object_tab->find(object_t(), cursor);
+  Table<pobject_t,Extent>::Cursor cursor(object_tab);
+  object_tab->find(pobject_t(), cursor);
   int nobj = 0;
   int njump = 0;
   while (object_tab->get_num_keys() > 0) {
diff --git a/branches/ebofs/ebofs/Ebofs.h b/branches/ebofs/ebofs/Ebofs.h
index b8bdb9b45b528..bbd618d8f9d61 100644
--- a/branches/ebofs/ebofs/Ebofs.h
+++ b/branches/ebofs/ebofs/Ebofs.h
@@ -38,7 +38,7 @@ using namespace __gnu_cxx;
 #include "osd/ObjectStore.h"
 
 //typedef pair<object_t,coll_t> object_coll_t;
-typedef pair<coll_t,object_t> coll_object_t;
+typedef pair<coll_t,pobject_t> coll_pobject_t;
 
 
 class Ebofs : public ObjectStore {
@@ -107,30 +107,30 @@ protected:
   NodePool     nodepool;   // for all tables...
 
   // tables
-  Table<object_t, Extent> *object_tab;
+  Table<pobject_t, Extent> *object_tab;
   Table<block_t,block_t>  *free_tab[EBOFS_NUM_FREE_BUCKETS];
   Table<block_t,block_t>  *limbo_tab;
   Table<block_t,pair<block_t,int> > *alloc_tab;
 
   // collections
   Table<coll_t, Extent>  *collection_tab;
-  Table<coll_object_t, bool>  *co_tab;
+  Table<coll_pobject_t, bool>  *co_tab;
 
   void close_tables();
   void verify_tables();
 
 
   // ** onodes **
-  hash_map<object_t, Onode*>  onode_map;  // onode cache
+  hash_map<pobject_t, Onode*>  onode_map;  // onode cache
   LRU                         onode_lru;
   set<Onode*>                 dirty_onodes;
-  map<object_t, list<Cond*> > waitfor_onode;
+  map<pobject_t, list<Cond*> > waitfor_onode;
 
-  Onode* new_onode(object_t oid);     // make new onode.  ref++.
-  bool have_onode(object_t oid) {
+  Onode* new_onode(pobject_t oid);     // make new onode.  ref++.
+  bool have_onode(pobject_t oid) {
     return onode_map.count(oid);
   }
-  Onode* get_onode(object_t oid);     // get cached onode, or read from disk.  ref++.
+  Onode* get_onode(pobject_t oid);     // get cached onode, or read from disk.  ref++.
   void remove_onode(Onode *on);
   void put_onode(Onode* o);         // put it back down.  ref--.
   void dirty_onode(Onode* o);
@@ -270,35 +270,35 @@ protected:
   // atomic transaction
   unsigned apply_transaction(Transaction& t, Context *onsafe=0);
 
-  int pick_object_revision_lt(object_t& oid);
+  int pick_object_revision_lt(pobject_t& oid);
 
   // object interface
-  bool exists(object_t);
-  int stat(object_t, struct stat*);
-  int read(object_t, off_t off, size_t len, bufferlist& bl);
-  int is_cached(object_t oid, off_t off, size_t len);
-
-  int write(object_t oid, off_t off, size_t len, const bufferlist& bl, Context *onsafe);
-  int zero(object_t oid, off_t off, size_t len, Context *onsafe);
-  int truncate(object_t oid, off_t size, Context *onsafe=0);
-  int remove(object_t oid, Context *onsafe=0);
+  bool exists(pobject_t);
+  int stat(pobject_t, struct stat*);
+  int read(pobject_t, off_t off, size_t len, bufferlist& bl);
+  int is_cached(pobject_t oid, off_t off, size_t len);
+
+  int write(pobject_t oid, off_t off, size_t len, const bufferlist& bl, Context *onsafe);
+  int zero(pobject_t oid, off_t off, size_t len, Context *onsafe);
+  int truncate(pobject_t oid, off_t size, Context *onsafe=0);
+  int remove(pobject_t oid, Context *onsafe=0);
   bool write_will_block();
-  void trim_from_cache(object_t oid, off_t off, size_t len);
+  void trim_from_cache(pobject_t oid, off_t off, size_t len);
 
-  int rename(object_t from, object_t to);
-  int clone(object_t from, object_t to, Context *onsafe);
+  int rename(pobject_t from, pobject_t to);
+  int clone(pobject_t from, pobject_t to, Context *onsafe);
 
-  int list_objects(list<object_t>& ls);
+  int list_objects(list<pobject_t>& ls);
 
   // object attr
-  int setattr(object_t oid, const char *name, const void *value, size_t size, Context *onsafe=0);
-  int setattrs(object_t oid, map<string,bufferptr>& attrset, Context *onsafe=0);
-  int getattr(object_t oid, const char *name, void *value, size_t size);
-  int getattrs(object_t oid, map<string,bufferptr> &aset);
-  int rmattr(object_t oid, const char *name, Context *onsafe=0);
-  int listattr(object_t oid, vector<string>& attrs);
+  int setattr(pobject_t oid, const char *name, const void *value, size_t size, Context *onsafe=0);
+  int setattrs(pobject_t oid, map<string,bufferptr>& attrset, Context *onsafe=0);
+  int getattr(pobject_t oid, const char *name, void *value, size_t size);
+  int getattrs(pobject_t oid, map<string,bufferptr> &aset);
+  int rmattr(pobject_t oid, const char *name, Context *onsafe=0);
+  int listattr(pobject_t oid, vector<string>& attrs);
 
-  int get_object_collections(object_t oid, set<coll_t>& ls);
+  int get_object_collections(pobject_t oid, set<coll_t>& ls);
 
   // collections
   int list_collections(list<coll_t>& ls);
@@ -306,10 +306,10 @@ protected:
 
   int create_collection(coll_t c, Context *onsafe);
   int destroy_collection(coll_t c, Context *onsafe);
-  int collection_add(coll_t c, object_t o, Context *onsafe);
-  int collection_remove(coll_t c, object_t o, Context *onsafe);
+  int collection_add(coll_t c, pobject_t o, Context *onsafe);
+  int collection_remove(coll_t c, pobject_t o, Context *onsafe);
 
-  int collection_list(coll_t c, list<object_t>& o);
+  int collection_list(coll_t c, list<pobject_t>& o);
   
   int collection_setattr(coll_t cid, const char *name, const void *value, size_t size, Context *onsafe);
   int collection_setattrs(coll_t cid, map<string,bufferptr> &aset);
@@ -319,12 +319,12 @@ protected:
   int collection_listattr(coll_t oid, vector<string>& attrs);
   
   // maps
-  int map_lookup(object_t o, bufferlist& key, bufferlist& val);
-  int map_insert(object_t o, bufferlist& key, bufferlist& val);
-  int map_remove(object_t o, bufferlist& key);
-  int map_list(object_t o, list<bufferlist>& keys);
-  int map_list(object_t o, map<bufferlist,bufferlist>& vals);
-  int map_list(object_t o, 
+  int map_lookup(pobject_t o, bufferlist& key, bufferlist& val);
+  int map_insert(pobject_t o, bufferlist& key, bufferlist& val);
+  int map_remove(pobject_t o, bufferlist& key);
+  int map_list(pobject_t o, list<bufferlist>& keys);
+  int map_list(pobject_t o, map<bufferlist,bufferlist>& vals);
+  int map_list(pobject_t o, 
 	       bufferlist& start, bufferlist& end,
 	       map<bufferlist,bufferlist>& vals);
 
@@ -340,28 +340,28 @@ private:
   // private interface -- use if caller already holds lock
   unsigned _apply_transaction(Transaction& t);
 
-  int _read(object_t oid, off_t off, size_t len, bufferlist& bl);
-  int _is_cached(object_t oid, off_t off, size_t len);
-  int _stat(object_t oid, struct stat *st);
-  int _getattr(object_t oid, const char *name, void *value, size_t size);
-  int _getattrs(object_t oid, map<string,bufferptr> &aset);
-  int _get_object_collections(object_t oid, set<coll_t>& ls);
+  int _read(pobject_t oid, off_t off, size_t len, bufferlist& bl);
+  int _is_cached(pobject_t oid, off_t off, size_t len);
+  int _stat(pobject_t oid, struct stat *st);
+  int _getattr(pobject_t oid, const char *name, void *value, size_t size);
+  int _getattrs(pobject_t oid, map<string,bufferptr> &aset);
+  int _get_object_collections(pobject_t oid, set<coll_t>& ls);
 
   bool _write_will_block();
-  int _write(object_t oid, off_t off, size_t len, const bufferlist& bl);
-  void _trim_from_cache(object_t oid, off_t off, size_t len);
-  int _truncate(object_t oid, off_t size);
-  int _zero(object_t oid, off_t offset, size_t length);
-  int _remove(object_t oid);
-  int _clone(object_t from, object_t to);
-  int _setattr(object_t oid, const char *name, const void *value, size_t size);
-  int _setattrs(object_t oid, map<string,bufferptr>& attrset);
-  int _rmattr(object_t oid, const char *name);
+  int _write(pobject_t oid, off_t off, size_t len, const bufferlist& bl);
+  void _trim_from_cache(pobject_t oid, off_t off, size_t len);
+  int _truncate(pobject_t oid, off_t size);
+  int _zero(pobject_t oid, off_t offset, size_t length);
+  int _remove(pobject_t oid);
+  int _clone(pobject_t from, pobject_t to);
+  int _setattr(pobject_t oid, const char *name, const void *value, size_t size);
+  int _setattrs(pobject_t oid, map<string,bufferptr>& attrset);
+  int _rmattr(pobject_t oid, const char *name);
   bool _collection_exists(coll_t c);
   int _create_collection(coll_t c);
   int _destroy_collection(coll_t c);
-  int _collection_add(coll_t c, object_t o);
-  int _collection_remove(coll_t c, object_t o);
+  int _collection_add(coll_t c, pobject_t o);
+  int _collection_remove(coll_t c, pobject_t o);
   int _collection_getattrs(coll_t oid, map<string,bufferptr> &aset);
   int _collection_setattr(coll_t oid, const char *name, const void *value, size_t size);
   int _collection_setattrs(coll_t oid, map<string,bufferptr> &aset);
diff --git a/branches/ebofs/ebofs/Onode.h b/branches/ebofs/ebofs/Onode.h
index 58ba445cacfaf..fd78b4f788a85 100644
--- a/branches/ebofs/ebofs/Onode.h
+++ b/branches/ebofs/ebofs/Onode.h
@@ -69,7 +69,7 @@ private:
   int ref;
 
 public:
-  object_t object_id;
+  pobject_t object_id;
   version_t version;      // incremented on each modify.
 
   // data
@@ -97,7 +97,7 @@ public:
   list<Context*>   commit_waiters;
 
  public:
-  Onode(object_t oid) : ref(0), object_id(oid), version(0),
+  Onode(pobject_t oid) : ref(0), object_id(oid), version(0),
 			object_size(0), alloc_blocks(0), last_block(0), data_csum(0),
 			readonly(0),
 			oc(0),
diff --git a/branches/ebofs/ebofs/types.h b/branches/ebofs/ebofs/types.h
index 638cad437fe85..6645881e355e5 100644
--- a/branches/ebofs/ebofs/types.h
+++ b/branches/ebofs/ebofs/types.h
@@ -19,6 +19,7 @@
 #include <cassert>
 #include "include/buffer.h"
 #include "include/Context.h"
+#include "include/pobject.h"
 #include "common/Cond.h"
 
 #include <ext/hash_map>
@@ -71,7 +72,7 @@ struct ebofs_onode {
   csum_t data_csum;
 
   Extent onode_loc;       /* this is actually the block we live in */
-  object_t object_id;       /* for kicks */
+  pobject_t object_id;       /* for kicks */
   __u8 readonly;
 
   __s64 object_size;     /* file size in bytes.  should this be 64-bit? */
diff --git a/branches/ebofs/include/object.h b/branches/ebofs/include/object.h
index 3b8ac05a86b38..39e67212be23c 100644
--- a/branches/ebofs/include/object.h
+++ b/branches/ebofs/include/object.h
@@ -31,8 +31,8 @@ typedef uint32_t objectrev_t;
 struct object_t {
   static const uint32_t MAXREV = 0xffffffffU;
 
-  uint64_t ino;  // "file" identifier
-  uint32_t bno;  // "block" in that "file"
+  uint64_t ino;    // "file" identifier
+  uint32_t bno;    // "block" in that "file"
   objectrev_t rev; // revision.  normally ctime (as epoch).
 
   object_t() : ino(0), bno(0), rev(0) {}
@@ -40,34 +40,23 @@ struct object_t {
   object_t(uint64_t i, uint32_t b, uint32_t r) : ino(i), bno(b), rev(r) {}
 };
 
-
 inline bool operator==(const object_t l, const object_t r) {
-  return (l.ino == r.ino) && (l.bno == r.bno) && (l.rev == r.rev);
+  return memcmp(&l, &r, sizeof(l)) == 0;
 }
 inline bool operator!=(const object_t l, const object_t r) {
-  return (l.ino != r.ino) || (l.bno != r.bno) || (l.rev != r.rev);
+  return memcmp(&l, &r, sizeof(l)) != 0;
 }
 inline bool operator>(const object_t l, const object_t r) {
-  if (l.ino > r.ino) return true;
-  if (l.ino < r.ino) return false;
-  if (l.bno > r.bno) return true;
-  if (l.bno < r.bno) return false;
-  if (l.rev > r.rev) return true;
-  return false;
+  return memcmp(&l, &r, sizeof(l)) > 0;
 }
 inline bool operator<(const object_t l, const object_t r) {
-  if (l.ino < r.ino) return true;
-  if (l.ino > r.ino) return false;
-  if (l.bno < r.bno) return true;
-  if (l.bno > r.bno) return false;
-  if (l.rev < r.rev) return true;
-  return false;
+  return memcmp(&l, &r, sizeof(l)) < 0;
 }
 inline bool operator>=(const object_t l, const object_t r) { 
-  return !(l < r);
+  return memcmp(&l, &r, sizeof(l)) >= 0;
 }
 inline bool operator<=(const object_t l, const object_t r) {
-  return !(l > r);
+  return memcmp(&l, &r, sizeof(l)) <= 0;
 }
 inline ostream& operator<<(ostream& out, const object_t o) {
   out << hex << o.ino << '.';
@@ -80,17 +69,11 @@ inline ostream& operator<<(ostream& out, const object_t o) {
   return out;
 }
 
-
-
-
-
 namespace __gnu_cxx {
   template<> struct hash<object_t> {
     size_t operator()(const object_t &r) const { 
       static rjhash<uint64_t> H;
       static rjhash<uint32_t> I;
-      //static hash<uint64_t> H;
-      //static hash<uint32_t> I;
       return H(r.ino) ^ I(r.bno) ^ I(r.rev);
     }
   };
diff --git a/branches/ebofs/include/pobject.h b/branches/ebofs/include/pobject.h
new file mode 100644
index 0000000000000..a334524483719
--- /dev/null
+++ b/branches/ebofs/include/pobject.h
@@ -0,0 +1,68 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- 
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software 
+ * Foundation.  See file COPYING.
+ * 
+ */
+
+#ifndef __CEPH_POBJECT_H
+#define __CEPH_POBJECT_H
+
+#include "object.h"
+
+/*
+ * "physical" object stored in an individual OSD's object store.
+ * includes fields to describe which volume the logical object_t
+ * belongs to, and/or a specific part of the object (if striped
+ * or encoded for redundancy, etc.).
+ */
+struct pobject_t {
+  uint32_t volume;     // "volume"
+  uint32_t rank;       // rank/stripe id (e.g. for parity encoding)
+  object_t oid;        // logical object
+  pobject_t() : volume(0), rank(0) {}
+  pobject_t(object_t o) : volume(0), rank(0), oid(o) {}  // this should go away eventually
+  pobject_t(uint16_t v, uint16_t r, object_t o) : volume(v), rank(r), oid(o) {}
+};
+
+inline ostream& operator<<(ostream& out, const pobject_t o) {
+  return out << o.volume << '/' << o.rank << '/' << o.oid;
+}
+
+inline bool operator==(const pobject_t l, const pobject_t r) {
+  return memcmp(&l, &r, sizeof(l)) == 0;
+}
+inline bool operator!=(const pobject_t l, const pobject_t r) {
+  return memcmp(&l, &r, sizeof(l)) != 0;
+}
+inline bool operator>(const pobject_t l, const pobject_t r) {
+  return memcmp(&l, &r, sizeof(l)) > 0;
+}
+inline bool operator>=(const pobject_t l, const pobject_t r) {
+  return memcmp(&l, &r, sizeof(l)) >= 0;
+}
+inline bool operator<(const pobject_t l, const pobject_t r) {
+  return memcmp(&l, &r, sizeof(l)) < 0;
+}
+inline bool operator<=(const pobject_t l, const pobject_t r) {
+  return memcmp(&l, &r, sizeof(l)) <= 0;
+}
+
+namespace __gnu_cxx {
+  template<> struct hash<pobject_t> {
+    size_t operator()(const pobject_t &r) const { 
+      static rjhash<uint64_t> H;
+      static rjhash<uint32_t> I;
+      return I(r.volume) ^ I(r.rank) ^ H(r.oid.ino) ^ I(r.oid.bno) ^ I(r.oid.rev);
+    }
+  };
+}
+
+#endif
diff --git a/branches/ebofs/mds/mdstypes.h b/branches/ebofs/mds/mdstypes.h
index c185909d89fae..bda0b5591ae5f 100644
--- a/branches/ebofs/mds/mdstypes.h
+++ b/branches/ebofs/mds/mdstypes.h
@@ -32,7 +32,7 @@ using namespace std;
 #define MDS_INO_ROOT              1
 #define MDS_INO_PGTABLE           2
 #define MDS_INO_ANCHORTABLE       3
-#define MDS_INO_PG                4       // *** WARNING: this should match osd/osd_types.h PG_INO ***
+
 #define MDS_INO_LOG_OFFSET        (1*MAX_MDS)
 #define MDS_INO_IDS_OFFSET        (2*MAX_MDS)
 #define MDS_INO_CLIENTMAP_OFFSET  (3*MAX_MDS)
diff --git a/branches/ebofs/osd/Fake.h b/branches/ebofs/osd/Fake.h
index 342c153c25cfd..f364438d3b088 100644
--- a/branches/ebofs/osd/Fake.h
+++ b/branches/ebofs/osd/Fake.h
@@ -28,7 +28,7 @@ class FakeStoreCollections {
  private:
   Mutex faker_lock;
   ObjectStore *store;
-  hash_map<coll_t, set<object_t> > fakecollections;
+  hash_map<coll_t, set<pobject_t> > fakecollections;
 
  public:
   FakeStoreCollections(ObjectStore *s) : store(s) {}
@@ -37,7 +37,7 @@ class FakeStoreCollections {
   int list_collections(list<coll_t>& ls) {
     faker_lock.Lock();
     int r = 0;
-    for (hash_map< coll_t, set<object_t> >::iterator p = fakecollections.begin();
+    for (hash_map< coll_t, set<pobject_t> >::iterator p = fakecollections.begin();
          p != fakecollections.end();
          p++) {
       r++;
@@ -81,7 +81,7 @@ class FakeStoreCollections {
     return r;
   }
 
-  int collection_add(coll_t c, object_t o,
+  int collection_add(coll_t c, pobject_t o,
                      Context *onsafe=0) {
     faker_lock.Lock();
     fakecollections[c].insert(o);
@@ -90,7 +90,7 @@ class FakeStoreCollections {
     return 0;
   }
 
-  int collection_remove(coll_t c, object_t o,
+  int collection_remove(coll_t c, pobject_t o,
                         Context *onsafe=0) {
     faker_lock.Lock();
     fakecollections[c].erase(o);
@@ -99,10 +99,10 @@ class FakeStoreCollections {
     return 0;
   }
 
-  int collection_list(coll_t c, list<object_t>& o) {
+  int collection_list(coll_t c, list<pobject_t>& o) {
     faker_lock.Lock();
     int r = 0;
-    for (set<object_t>::iterator p = fakecollections[c].begin();
+    for (set<pobject_t>::iterator p = fakecollections[c].begin();
          p != fakecollections[c].end();
          p++) {
       o.push_back(*p);
@@ -164,13 +164,13 @@ class FakeStoreAttrs {
 
   Mutex faker_lock;
   ObjectStore *store;
-  hash_map<object_t, FakeAttrSet> fakeoattrs;
+  hash_map<pobject_t, FakeAttrSet> fakeoattrs;
   hash_map<coll_t, FakeAttrSet> fakecattrs;
 
  public:
   FakeStoreAttrs(ObjectStore *s) : store(s) {}
 
-  int setattr(object_t oid, const char *name,
+  int setattr(pobject_t oid, const char *name,
               const void *value, size_t size,
               Context *onsafe=0) {
     faker_lock.Lock();
@@ -179,26 +179,26 @@ class FakeStoreAttrs {
     faker_lock.Unlock();
     return r;
   }
-  int setattrs(object_t oid, map<string,bufferptr>& aset) {
+  int setattrs(pobject_t oid, map<string,bufferptr>& aset) {
     faker_lock.Lock();
     int r = fakeoattrs[oid].setattrs(aset);
     faker_lock.Unlock();
     return r;
   }
-  int getattr(object_t oid, const char *name,
+  int getattr(pobject_t oid, const char *name,
               void *value, size_t size) {
     faker_lock.Lock();
     int r = fakeoattrs[oid].getattr(name, value, size);
     faker_lock.Unlock();
     return r;
   }
-  int getattrs(object_t oid, map<string,bufferptr>& aset) {
+  int getattrs(pobject_t oid, map<string,bufferptr>& aset) {
     faker_lock.Lock();
     int r = fakeoattrs[oid].getattrs(aset);
     faker_lock.Unlock();
     return r;
   }
-  int rmattr(object_t oid, const char *name,
+  int rmattr(pobject_t oid, const char *name,
              Context *onsafe=0) {
     faker_lock.Lock();
     int r = fakeoattrs[oid].rmattr(name);
@@ -207,7 +207,7 @@ class FakeStoreAttrs {
     return r;
   }
 
-  int listattr(object_t oid, char *attrs, size_t size) {
+  int listattr(pobject_t oid, char *attrs, size_t size) {
     faker_lock.Lock();
     int r = fakeoattrs[oid].listattr(attrs,size);
     faker_lock.Unlock();
diff --git a/branches/ebofs/osd/FakeStore.cc b/branches/ebofs/osd/FakeStore.cc
index e7c77f3eab558..6f50e16e4a074 100644
--- a/branches/ebofs/osd/FakeStore.cc
+++ b/branches/ebofs/osd/FakeStore.cc
@@ -64,25 +64,48 @@ int FakeStore::statfs(struct statfs *buf)
 
 
 /* 
- * sorry, these are sentitive to the object_t and coll_t typing.
+ * sorry, these are sentitive to the pobject_t and coll_t typing.
  */ 
-void FakeStore::get_oname(object_t oid, char *s) 
+void FakeStore::get_oname(pobject_t oid, char *s) 
 {
-  //static hash<object_t> H;
-  assert(sizeof(oid) == 16);
+  //static hash<pobject_t> H;
+  assert(sizeof(oid) == 24);
 #ifdef __LP64__
   //sprintf(s, "%s/objects/%02lx/%016lx.%016lx", basedir.c_str(), H(oid) & HASH_MASK, 
-  sprintf(s, "%s/objects/%016lx.%016lx", basedir.c_str(), 
-	  *((uint64_t*)&oid),
-	  *(((uint64_t*)&oid) + 1));
+  sprintf(s, "%s/objects/%04x.%04x.%016lx.%016lx", basedir.c_str(), 
+	  oid.volume, oid.rank,
+	  *((uint64_t*)&oid.oid),
+	  *(((uint64_t*)&oid.oid) + 1));
 #else
   //sprintf(s, "%s/objects/%02x/%016llx.%016llx", basedir.c_str(), H(oid) & HASH_MASK, 
-  sprintf(s, "%s/objects/%016llx.%016llx", basedir.c_str(), 
+  sprintf(s, "%s/objects/%04x.%04x.%016llx.%016llx", basedir.c_str(), 
+	  oid.volume, oisd.rank,
 	  *((uint64_t*)&oid),
 	  *(((uint64_t*)&oid) + 1));
 #endif
 }
 
+pobject_t FakeStore::parse_object(char *s)
+{
+  pobject_t o;
+  assert(sizeof(o) == 24);
+  //cout << "  got object " << de->d_name << std::endl;
+  o.volume = strtoll(s, 0, 16);
+  assert(s[4] == '.');
+  o.rank = strtoll(s+5, 0, 16);
+  assert(s[9] == '.');
+  *(((uint64_t*)&o.oid) + 0) = strtoll(s+10, 0, 16);
+  assert(s[26] == '.');
+  *(((uint64_t*)&o.oid) + 1) = strtoll(s+27, 0, 16);
+  dout(0) << " got " << o << " errno " << errno << " on " << s << dendl;
+  return o;
+}
+
+coll_t FakeStore::parse_coll(char *s)
+{
+  return strtoll(s, 0, 16);
+}
+
 void FakeStore::get_cdir(coll_t cid, char *s) 
 {
   assert(sizeof(cid) == 8);
@@ -95,15 +118,17 @@ void FakeStore::get_cdir(coll_t cid, char *s)
 #endif
 }
 
-void FakeStore::get_coname(coll_t cid, object_t oid, char *s) 
+void FakeStore::get_coname(coll_t cid, pobject_t oid, char *s) 
 {
-  assert(sizeof(oid) == 16);
+  assert(sizeof(oid) == 24);
 #ifdef __LP64__
-  sprintf(s, "%s/collections/%016lx/%016lx.%016lx", basedir.c_str(), cid, 
-	  *((uint64_t*)&oid),
-	  *(((uint64_t*)&oid) + 1));
+  sprintf(s, "%s/collections/%016lx/%04x.%04x.%016lx.%016lx", basedir.c_str(), cid, 
+	  oid.volume, oid.rank,
+	  *((uint64_t*)&oid.oid),
+	  *(((uint64_t*)&oid.oid) + 1));
 #else
-  sprintf(s, "%s/collections/%016llx/%016llx.%016llx", basedir.c_str(), cid, 
+  sprintf(s, "%s/collections/%016llx/%04x.%04x.%016llx.%016llx", basedir.c_str(), cid, 
+	  oid.volume, oid.rank,
 	  *((uint64_t*)&oid),
 	  *(((uint64_t*)&oid) + 1));
 #endif
@@ -226,7 +251,7 @@ int FakeStore::umount()
 // objects
 
 
-bool FakeStore::exists(object_t oid)
+bool FakeStore::exists(pobject_t oid)
 {
   struct stat st;
   if (stat(oid, &st) == 0)
@@ -236,7 +261,7 @@ bool FakeStore::exists(object_t oid)
 }
 
   
-int FakeStore::stat(object_t oid,
+int FakeStore::stat(pobject_t oid,
                     struct stat *st)
 {
   dout(20) << "stat " << oid << dendl;
@@ -248,7 +273,7 @@ int FakeStore::stat(object_t oid,
  
  
 
-int FakeStore::remove(object_t oid, Context *onsafe) 
+int FakeStore::remove(pobject_t oid, Context *onsafe) 
 {
   dout(20) << "remove " << oid << dendl;
   char fn[200];
@@ -258,7 +283,7 @@ int FakeStore::remove(object_t oid, Context *onsafe)
   return r;
 }
 
-int FakeStore::truncate(object_t oid, off_t size, Context *onsafe)
+int FakeStore::truncate(pobject_t oid, off_t size, Context *onsafe)
 {
   dout(20) << "truncate " << oid << " size " << size << dendl;
 
@@ -269,7 +294,7 @@ int FakeStore::truncate(object_t oid, off_t size, Context *onsafe)
   return r;
 }
 
-int FakeStore::read(object_t oid, 
+int FakeStore::read(pobject_t oid, 
                     off_t offset, size_t len,
                     bufferlist& bl) {
   dout(20) << "read " << oid << " len " << len << " off " << offset << dendl;
@@ -305,7 +330,7 @@ int FakeStore::read(object_t oid,
 }
 
 
-int FakeStore::write(object_t oid, 
+int FakeStore::write(pobject_t oid, 
                      off_t offset, size_t len,
                      const bufferlist& bl, 
                      Context *onsafe)
@@ -410,7 +435,7 @@ void FakeStore::sync(Context *onsafe)
 
 // objects
 
-int FakeStore::setattr(object_t oid, const char *name,
+int FakeStore::setattr(pobject_t oid, const char *name,
 		       const void *value, size_t size,
 		       Context *onsafe) 
 {
@@ -425,7 +450,7 @@ int FakeStore::setattr(object_t oid, const char *name,
   return r;
 }
 
-int FakeStore::setattrs(object_t oid, map<string,bufferptr>& aset) 
+int FakeStore::setattrs(pobject_t oid, map<string,bufferptr>& aset) 
 {
   if (fake_attrs) return attrs.setattrs(oid, aset);
 
@@ -446,7 +471,7 @@ int FakeStore::setattrs(object_t oid, map<string,bufferptr>& aset)
   return r;
 }
 
-int FakeStore::getattr(object_t oid, const char *name,
+int FakeStore::getattr(pobject_t oid, const char *name,
 		       void *value, size_t size) 
 {
   if (fake_attrs) return attrs.getattr(oid, name, value, size);
@@ -459,7 +484,7 @@ int FakeStore::getattr(object_t oid, const char *name,
   return r;
 }
 
-int FakeStore::getattrs(object_t oid, map<string,bufferptr>& aset) 
+int FakeStore::getattrs(pobject_t oid, map<string,bufferptr>& aset) 
 {
   if (fake_attrs) return attrs.getattrs(oid, aset);
 
@@ -483,7 +508,7 @@ int FakeStore::getattrs(object_t oid, map<string,bufferptr>& aset)
   return 0;
 }
 
-int FakeStore::rmattr(object_t oid, const char *name, Context *onsafe) 
+int FakeStore::rmattr(pobject_t oid, const char *name, Context *onsafe) 
 {
   if (fake_attrs) return attrs.rmattr(oid, name, onsafe);
   int r = 0;
@@ -496,7 +521,7 @@ int FakeStore::rmattr(object_t oid, const char *name, Context *onsafe)
 }
 
 /*
-int FakeStore::listattr(object_t oid, char *attrls, size_t size) 
+int FakeStore::listattr(pobject_t oid, char *attrls, size_t size) 
 {
   if (fake_attrs) return attrs.listattr(oid, attrls, size);
   char fn[100];
@@ -582,7 +607,7 @@ int FakeStore::collection_listattr(coll_t c, char *attrs, size_t size)
 */
 
 
-int FakeStore::list_objects(list<object_t>& ls) 
+int FakeStore::list_objects(list<pobject_t>& ls) 
 {
   char fn[200];
   sprintf(fn, "%s/objects", basedir.c_str());
@@ -594,13 +619,7 @@ int FakeStore::list_objects(list<object_t>& ls)
   while ((de = ::readdir(dir)) != 0) {
     if (de->d_name[0] == '.') continue;
     // parse
-    object_t o;
-    assert(sizeof(o) == 16);
-    //cout << "  got object " << de->d_name << std::endl;
-    *(((uint64_t*)&o) + 0) = strtoll(de->d_name, 0, 16);
-    assert(de->d_name[16] == '.');
-    *(((uint64_t*)&o) + 1) = strtoll(de->d_name+17, 0, 16);
-    //dout(0) << " got " << o << " errno " << errno << " on " << de->d_name << dendl;
+    pobject_t o = parse_object(de->d_name);
     if (errno) continue;
     ls.push_back(o);
   }
@@ -627,7 +646,7 @@ int FakeStore::list_collections(list<coll_t>& ls)
   while ((de = ::readdir(dir)) != 0) {
     // parse
     errno = 0;
-    coll_t c = strtoll(de->d_name, 0, 16);
+    coll_t c = parse_coll(de->d_name);
     if (c) ls.push_back(c);
   }
   
@@ -682,7 +701,7 @@ bool FakeStore::collection_exists(coll_t c)
 }
 
 
-int FakeStore::collection_add(coll_t c, object_t o,
+int FakeStore::collection_add(coll_t c, pobject_t o,
 			      Context *onsafe) 
 {
   if (fake_collections) return collections.collection_add(c, o, onsafe);
@@ -697,7 +716,7 @@ int FakeStore::collection_add(coll_t c, object_t o,
   return r;
 }
 
-int FakeStore::collection_remove(coll_t c, object_t o,
+int FakeStore::collection_remove(coll_t c, pobject_t o,
 				 Context *onsafe) 
 {
   if (fake_collections) return collections.collection_remove(c, o, onsafe);
@@ -710,7 +729,7 @@ int FakeStore::collection_remove(coll_t c, object_t o,
   return r;
 }
 
-int FakeStore::collection_list(coll_t c, list<object_t>& ls) 
+int FakeStore::collection_list(coll_t c, list<pobject_t>& ls) 
 {  
   if (fake_collections) return collections.collection_list(c, ls);
 
@@ -725,12 +744,7 @@ int FakeStore::collection_list(coll_t c, list<object_t>& ls)
     // parse
     if (de->d_name[0] == '.') continue;
     //cout << "  got object " << de->d_name << std::endl;
-    object_t o;
-    assert(sizeof(o) == 16);
-    *(((uint64_t*)&o) + 0) = strtoll(de->d_name, 0, 16);
-    assert(de->d_name[16] == '.');
-    *(((uint64_t*)&o) + 1) = strtoll(de->d_name+17, 0, 16);
-    dout(0) << " got " << o << " errno " << errno << " on " << de->d_name << dendl;
+    pobject_t o = parse_object(de->d_name);
     if (errno) continue;
     ls.push_back(o);
   }
diff --git a/branches/ebofs/osd/FakeStore.h b/branches/ebofs/osd/FakeStore.h
index 5828c27c14d96..bbd7d82d3978f 100644
--- a/branches/ebofs/osd/FakeStore.h
+++ b/branches/ebofs/osd/FakeStore.h
@@ -49,9 +49,11 @@ class FakeStore : public ObjectStore {
   bool fake_collections;
   
   // helper fns
-  void get_oname(object_t oid, char *s);
+  void get_oname(pobject_t oid, char *s);
   void get_cdir(coll_t cid, char *s);
-  void get_coname(coll_t cid, object_t oid, char *s);
+  void get_coname(coll_t cid, pobject_t oid, char *s);
+  pobject_t parse_object(char *s);
+  coll_t parse_coll(char *s);
 
  public:
   FakeStore(char *base) : 
@@ -68,28 +70,28 @@ class FakeStore : public ObjectStore {
 
   // ------------------
   // objects
-  int pick_object_revision_lt(object_t& oid) {
+  int pick_object_revision_lt(pobject_t& oid) {
     return 0;
   }
-  bool exists(object_t oid);
-  int stat(object_t oid, struct stat *st);
-  int remove(object_t oid, Context *onsafe);
-  int truncate(object_t oid, off_t size, Context *onsafe);
-  int read(object_t oid, off_t offset, size_t len, bufferlist& bl);
-  int write(object_t oid, off_t offset, size_t len, const bufferlist& bl, Context *onsafe);
+  bool exists(pobject_t oid);
+  int stat(pobject_t oid, struct stat *st);
+  int remove(pobject_t oid, Context *onsafe);
+  int truncate(pobject_t oid, off_t size, Context *onsafe);
+  int read(pobject_t oid, off_t offset, size_t len, bufferlist& bl);
+  int write(pobject_t oid, off_t offset, size_t len, const bufferlist& bl, Context *onsafe);
 
   void sync();
   void sync(Context *onsafe);
 
-  int list_objects(list<object_t>& ls);
+  int list_objects(list<pobject_t>& ls);
 
   // attrs
-  int setattr(object_t oid, const char *name, const void *value, size_t size, Context *onsafe=0);
-  int setattrs(object_t oid, map<string,bufferptr>& aset);
-  int getattr(object_t oid, const char *name, void *value, size_t size);
-  int getattrs(object_t oid, map<string,bufferptr>& aset);
-  int rmattr(object_t oid, const char *name, Context *onsafe=0);
-  //int listattr(object_t oid, char *attrs, size_t size);
+  int setattr(pobject_t oid, const char *name, const void *value, size_t size, Context *onsafe=0);
+  int setattrs(pobject_t oid, map<string,bufferptr>& aset);
+  int getattr(pobject_t oid, const char *name, void *value, size_t size);
+  int getattrs(pobject_t oid, map<string,bufferptr>& aset);
+  int rmattr(pobject_t oid, const char *name, Context *onsafe=0);
+  //int listattr(pobject_t oid, char *attrs, size_t size);
   int collection_setattr(coll_t c, const char *name, void *value, size_t size, Context *onsafe=0);
   int collection_rmattr(coll_t c, const char *name, Context *onsafe=0);
   int collection_getattr(coll_t c, const char *name, void *value, size_t size);
@@ -103,9 +105,9 @@ class FakeStore : public ObjectStore {
   int destroy_collection(coll_t c, Context *onsafe=0);
   int collection_stat(coll_t c, struct stat *st);
   bool collection_exists(coll_t c);
-  int collection_add(coll_t c, object_t o, Context *onsafe=0);
-  int collection_remove(coll_t c, object_t o, Context *onsafe=0);
-  int collection_list(coll_t c, list<object_t>& o);
+  int collection_add(coll_t c, pobject_t o, Context *onsafe=0);
+  int collection_remove(coll_t c, pobject_t o, Context *onsafe=0);
+  int collection_list(coll_t c, list<pobject_t>& o);
 
 
 
diff --git a/branches/ebofs/osd/OSD.cc b/branches/ebofs/osd/OSD.cc
index 9e5210a4bc862..0512d57f13437 100644
--- a/branches/ebofs/osd/OSD.cc
+++ b/branches/ebofs/osd/OSD.cc
@@ -459,12 +459,12 @@ void OSD::_remove_unlock_pg(PG *pg)
   dout(10) << "_remove_unlock_pg " << pgid << dendl;
 
   // remove from store
-  list<object_t> olist;
+  list<pobject_t> olist;
   store->collection_list(pgid, olist);
   
   ObjectStore::Transaction t;
   {
-    for (list<object_t>::iterator p = olist.begin();
+    for (list<pobject_t>::iterator p = olist.begin();
 	 p != olist.end();
 	 p++)
       t.remove(*p);
diff --git a/branches/ebofs/osd/ObjectStore.h b/branches/ebofs/osd/ObjectStore.h
index d683120933b59..4902d94925c69 100644
--- a/branches/ebofs/osd/ObjectStore.h
+++ b/branches/ebofs/osd/ObjectStore.h
@@ -20,6 +20,7 @@
 #include "osd_types.h"
 #include "include/Context.h"
 #include "include/buffer.h"
+#include "include/pobject.h"
 
 #include "include/Distribution.h"
 
@@ -99,7 +100,7 @@ public:
   private:
     list<int8_t> ops;
     list<bufferlist> bls;
-    list<object_t> oids;
+    list<pobject_t> oids;
     list<coll_t> cids;
     list<int64_t> lengths;
     list<const char*> attrnames;
@@ -125,7 +126,7 @@ public:
       bl.claim(bls.front());
       bls.pop_front();
     }
-    void get_oid(object_t& oid) {
+    void get_oid(pobject_t& oid) {
       oid = oids.front();
       oids.pop_front();
     }
@@ -159,7 +160,7 @@ public:
     }
       
 
-    void read(object_t oid, off_t off, size_t len, bufferlist *pbl) {
+    void read(pobject_t oid, off_t off, size_t len, bufferlist *pbl) {
       int op = OP_READ;
       ops.push_back(op);
       oids.push_back(oid);
@@ -167,27 +168,27 @@ public:
       lengths.push_back(len);
       pbls.push_back(pbl);
     }
-    void stat(object_t oid, struct stat *st) {
+    void stat(pobject_t oid, struct stat *st) {
       int op = OP_STAT;
       ops.push_back(op);
       oids.push_back(oid);
       psts.push_back(st);
     }
-    void getattr(object_t oid, const char* name, void* val, int *plen) {
+    void getattr(pobject_t oid, const char* name, void* val, int *plen) {
       int op = OP_GETATTR;
       ops.push_back(op);
       oids.push_back(oid);
       attrnames.push_back(name);
       pattrvals.push_back(pair<void*,int*>(val,plen));
     }
-    void getattrs(object_t oid, map<string,bufferptr>& aset) {
+    void getattrs(pobject_t oid, map<string,bufferptr>& aset) {
       int op = OP_GETATTRS;
       ops.push_back(op);
       oids.push_back(oid);
       pattrsets.push_back(&aset);
     }
 
-    void write(object_t oid, off_t off, size_t len, const bufferlist& bl) {
+    void write(pobject_t oid, off_t off, size_t len, const bufferlist& bl) {
       int op = OP_WRITE;
       ops.push_back(op);
       oids.push_back(oid);
@@ -195,32 +196,32 @@ public:
       lengths.push_back(len);
       bls.push_back(bl);
     }
-    void zero(object_t oid, off_t off, size_t len) {
+    void zero(pobject_t oid, off_t off, size_t len) {
       int op = OP_ZERO;
       ops.push_back(op);
       oids.push_back(oid);
       lengths.push_back(off);
       lengths.push_back(len);
     }
-    void trim_from_cache(object_t oid, off_t off, size_t len) {
+    void trim_from_cache(pobject_t oid, off_t off, size_t len) {
       int op = OP_TRIMCACHE;
       ops.push_back(op);
       oids.push_back(oid);
       lengths.push_back(off);
       lengths.push_back(len);
     }
-    void truncate(object_t oid, off_t off) {
+    void truncate(pobject_t oid, off_t off) {
       int op = OP_TRUNCATE;
       ops.push_back(op);
       oids.push_back(oid);
       lengths.push_back(off);
     }
-    void remove(object_t oid) {
+    void remove(pobject_t oid) {
       int op = OP_REMOVE;
       ops.push_back(op);
       oids.push_back(oid);
     }
-    void setattr(object_t oid, const char* name, const void* val, int len) {
+    void setattr(pobject_t oid, const char* name, const void* val, int len) {
       int op = OP_SETATTR;
       ops.push_back(op);
       oids.push_back(oid);
@@ -230,19 +231,19 @@ public:
       bl.append((char*)val,len);
       bls.push_back(bl);
     }
-    void setattrs(object_t oid, map<string,bufferptr>& attrset) {
+    void setattrs(pobject_t oid, map<string,bufferptr>& attrset) {
       int op = OP_SETATTRS;
       ops.push_back(op);
       oids.push_back(oid);
       pattrsets.push_back(&attrset);
     }
-    void rmattr(object_t oid, const char* name) {
+    void rmattr(pobject_t oid, const char* name) {
       int op = OP_RMATTR;
       ops.push_back(op);
       oids.push_back(oid);
       attrnames.push_back(name);
     }
-    void clone(object_t oid, object_t noid) {
+    void clone(pobject_t oid, pobject_t noid) {
       int op = OP_CLONE;
       ops.push_back(op);
       oids.push_back(oid);
@@ -258,13 +259,13 @@ public:
       ops.push_back(op);
       cids.push_back(cid);
     }
-    void collection_add(coll_t cid, object_t oid) {
+    void collection_add(coll_t cid, pobject_t oid) {
       int op = OP_COLL_ADD;
       ops.push_back(op);
       cids.push_back(cid);
       oids.push_back(oid);
     }
-    void collection_remove(coll_t cid, object_t oid) {
+    void collection_remove(coll_t cid, pobject_t oid) {
       int op = OP_COLL_REMOVE;
       ops.push_back(op);
       cids.push_back(cid);
@@ -322,7 +323,7 @@ public:
       switch (op) {
       case Transaction::OP_READ:
         {
-          object_t oid;
+          pobject_t oid;
           off_t offset, len;
 	  t.get_oid(oid);
 	  t.get_length(offset);
@@ -334,7 +335,7 @@ public:
         break;
       case Transaction::OP_STAT:
         {
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           struct stat *st;
 	  t.get_pstat(st);
@@ -343,7 +344,7 @@ public:
         break;
       case Transaction::OP_GETATTR:
         {
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           const char *attrname;
 	  t.get_attrname(attrname);
@@ -354,7 +355,7 @@ public:
         break;
       case Transaction::OP_GETATTRS:
         {
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           map<string,bufferptr> *pset;
 	  t.get_pattrset(pset);
@@ -364,7 +365,7 @@ public:
 
       case Transaction::OP_WRITE:
         {
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           off_t offset, len;
 	  t.get_length(offset);
@@ -377,7 +378,7 @@ public:
 
       case Transaction::OP_ZERO:
         {
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           off_t offset, len;
 	  t.get_length(offset);
@@ -388,7 +389,7 @@ public:
 
       case Transaction::OP_TRIMCACHE:
         {
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           off_t offset, len;
 	  t.get_length(offset);
@@ -399,7 +400,7 @@ public:
 
       case Transaction::OP_TRUNCATE:
         {
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           off_t len;
 	  t.get_length(len);
@@ -409,7 +410,7 @@ public:
 
       case Transaction::OP_REMOVE:
         {
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           remove(oid, 0);
         }
@@ -417,7 +418,7 @@ public:
 
       case Transaction::OP_SETATTR:
         {
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           const char *attrname;
 	  t.get_attrname(attrname);
@@ -428,7 +429,7 @@ public:
         break;
       case Transaction::OP_SETATTRS:
         {
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           map<string,bufferptr> *pattrset;
 	  t.get_pattrset(pattrset);
@@ -438,7 +439,7 @@ public:
 
       case Transaction::OP_RMATTR:
         {
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           const char *attrname;
 	  t.get_attrname(attrname);
@@ -448,9 +449,9 @@ public:
 
       case Transaction::OP_CLONE:
 	{
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
-          object_t noid;
+          pobject_t noid;
 	  t.get_oid(noid);
 	  clone(oid, noid);
 	}
@@ -476,7 +477,7 @@ public:
         {
           coll_t cid;
 	  t.get_cid(cid);
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           collection_add(cid, oid, 0);
         }
@@ -486,7 +487,7 @@ public:
         {
           coll_t cid;
 	  t.get_cid(cid);
-          object_t oid;
+          pobject_t oid;
 	  t.get_oid(oid);
           collection_remove(cid, oid, 0);
         }
@@ -542,25 +543,25 @@ public:
   virtual int statfs(struct statfs *buf) = 0;
 
   // objects
-  virtual int pick_object_revision_lt(object_t& oid) = 0;
+  virtual int pick_object_revision_lt(pobject_t& oid) = 0;
 
-  virtual bool exists(object_t oid) = 0;                   // useful?
-  virtual int stat(object_t oid, struct stat *st) = 0;     // struct stat?
+  virtual bool exists(pobject_t oid) = 0;                   // useful?
+  virtual int stat(pobject_t oid, struct stat *st) = 0;     // struct stat?
 
-  virtual int remove(object_t oid,
+  virtual int remove(pobject_t oid,
                      Context *onsafe=0) = 0;
 
-  virtual int truncate(object_t oid, off_t size,
+  virtual int truncate(pobject_t oid, off_t size,
                        Context *onsafe=0) = 0;
   
-  virtual int read(object_t oid, 
+  virtual int read(pobject_t oid, 
                    off_t offset, size_t len,
                    bufferlist& bl) = 0;
-  virtual int write(object_t oid, 
+  virtual int write(pobject_t oid, 
                     off_t offset, size_t len,
                     const bufferlist& bl, 
                     Context *onsafe) = 0;//{ return -1; }
-  virtual int zero(object_t oid, 
+  virtual int zero(pobject_t oid, 
 		   off_t offset, size_t len,
 		   Context *onsafe) {
     // write zeros.. yuck!
@@ -569,33 +570,33 @@ public:
     bl.push_back(bp);
     return write(oid, offset, len, bl, onsafe);
   }
-  virtual void trim_from_cache(object_t oid, 
+  virtual void trim_from_cache(pobject_t oid, 
 			       off_t offset, size_t len) { }
-  virtual int is_cached(object_t oid, 
+  virtual int is_cached(pobject_t oid, 
 			     off_t offset, 
                              size_t len) { return -1; }
 
-  virtual int setattr(object_t oid, const char *name,
+  virtual int setattr(pobject_t oid, const char *name,
                       const void *value, size_t size,
                       Context *onsafe=0) {return 0;} //= 0;
-  virtual int setattrs(object_t oid, map<string,bufferptr>& aset,
+  virtual int setattrs(pobject_t oid, map<string,bufferptr>& aset,
                       Context *onsafe=0) {return 0;} //= 0;
-  virtual int getattr(object_t oid, const char *name,
+  virtual int getattr(pobject_t oid, const char *name,
                       void *value, size_t size) {return 0;} //= 0;
-  virtual int getattrs(object_t oid, map<string,bufferptr>& aset) {return 0;};
+  virtual int getattrs(pobject_t oid, map<string,bufferptr>& aset) {return 0;};
 
-  virtual int rmattr(object_t oid, const char *name,
+  virtual int rmattr(pobject_t oid, const char *name,
                      Context *onsafe=0) {return 0;}
 
-  virtual int clone(object_t oid, object_t noid) {
+  virtual int clone(pobject_t oid, pobject_t noid) {
     return -1; 
   }
   
-  virtual int list_objects(list<object_t>& ls) = 0;//{ return -1; }
+  virtual int list_objects(list<pobject_t>& ls) = 0;//{ return -1; }
 
-  virtual int get_object_collections(object_t oid, set<coll_t>& ls) { return -1; }
+  virtual int get_object_collections(pobject_t oid, set<coll_t>& ls) { return -1; }
 
-  //virtual int listattr(object_t oid, char *attrs, size_t size) {return 0;} //= 0;
+  //virtual int listattr(pobject_t oid, char *attrs, size_t size) {return 0;} //= 0;
   
   // collections
   virtual int list_collections(list<coll_t>& ls) {return 0;}//= 0;
@@ -605,11 +606,11 @@ public:
                                  Context *onsafe=0) {return 0;}//= 0;
   virtual bool collection_exists(coll_t c) {return 0;}
   virtual int collection_stat(coll_t c, struct stat *st) {return 0;}//= 0;
-  virtual int collection_add(coll_t c, object_t o,
+  virtual int collection_add(coll_t c, pobject_t o,
                              Context *onsafe=0) {return 0;}//= 0;
-  virtual int collection_remove(coll_t c, object_t o,
+  virtual int collection_remove(coll_t c, pobject_t o,
                                 Context *onsafe=0) {return 0;}// = 0;
-  virtual int collection_list(coll_t c, list<object_t>& o) {return 0;}//= 0;
+  virtual int collection_list(coll_t c, list<pobject_t>& o) {return 0;}//= 0;
 
   virtual int collection_setattr(coll_t cid, const char *name,
                                  const void *value, size_t size,
diff --git a/branches/ebofs/osd/PG.cc b/branches/ebofs/osd/PG.cc
index 99b81f2f0dc96..cb2703d101ce8 100644
--- a/branches/ebofs/osd/PG.cc
+++ b/branches/ebofs/osd/PG.cc
@@ -430,23 +430,24 @@ void PG::generate_backlog()
   assert(!log.backlog);
   log.backlog = true;
 
-  list<object_t> olist;
+  list<pobject_t> olist;
   osd->store->collection_list(info.pgid, olist);
   
   int local = 0;
   map<eversion_t,Log::Entry> add;
-  for (list<object_t>::iterator it = olist.begin();
+  for (list<pobject_t>::iterator it = olist.begin();
        it != olist.end();
        it++) {
     local++;
+    object_t oid = it->oid;
 
-    if (log.logged_object(*it)) continue; // already have it logged.
+    if (log.logged_object(oid)) continue; // already have it logged.
     
     // add entry
     Log::Entry e;
     e.op = Log::Entry::MODIFY;           // FIXME when we do smarter op codes!
-    e.oid = *it;
-    osd->store->getattr(*it, 
+    e.oid = oid;
+    osd->store->getattr(pobject_t(0,0,oid), 
                         "version",
                         &e.version, sizeof(e.version));
     add[e.version] = e;
@@ -1266,7 +1267,7 @@ bool PG::pick_missing_object_rev(object_t& oid)
 
 bool PG::pick_object_rev(object_t& oid)
 {
-  object_t t = oid;
+  pobject_t t = oid;
 
   if (!osd->store->pick_object_revision_lt(t))
     return false; // we have no revisions of this object!
@@ -1276,7 +1277,7 @@ bool PG::pick_object_rev(object_t& oid)
   assert(r >= 0);
   if (crev <= oid.rev) {
     dout(10) << "pick_object_rev choosing " << t << " crev " << crev << " for " << oid << dendl;
-    oid = t;
+    oid = t.oid;
     return true;
   }
 
diff --git a/branches/ebofs/osd/ReplicatedPG.cc b/branches/ebofs/osd/ReplicatedPG.cc
index 07ab10fadbd20..592c7ca20d1e5 100644
--- a/branches/ebofs/osd/ReplicatedPG.cc
+++ b/branches/ebofs/osd/ReplicatedPG.cc
@@ -150,7 +150,7 @@ bool ReplicatedPG::preprocess_op(MOSDOp *op, utime_t now)
       bool is_balanced = false;
       bool b;
       // *** FIXME *** this may block, and we're in the fast path! ***
-      if (osd->store->getattr(oid, "balance-reads", &b, 1) >= 0)
+      if (osd->store->getattr(pobject_t(0,0,oid), "balance-reads", &b, 1) >= 0)
 	is_balanced = true;
       
       if (!is_balanced && should_balance &&
@@ -321,13 +321,13 @@ bool ReplicatedPG::preprocess_op(MOSDOp *op, utime_t now)
   // -- fastpath read?
   // if this is a read and the data is in the cache, do an immediate read.. 
   if ( g_conf.osd_immediate_read_from_cache ) {
-    if (osd->store->is_cached( oid , 
+    if (osd->store->is_cached( pobject_t(0,0,oid) , 
 			       op->get_offset(), 
 			       op->get_length() ) == 0) {
       if (!is_primary() && !op->get_source().is_osd()) {
 	// am i allowed?
 	bool v;
-	if (osd->store->getattr(oid, "balance-reads", &v, 1) < 0) {
+	if (osd->store->getattr(pobject_t(0,0,oid), "balance-reads", &v, 1) < 0) {
 	  dout(-10) << "preprocess_op in-cache but no balance-reads on " << oid
 		    << ", fwd to primary" << dendl;
 	  osd->messenger->send_message(op, osd->osdmap->get_inst(get_primary()));
@@ -1731,15 +1731,18 @@ void ReplicatedPG::clean_up_local(ObjectStore::Transaction& t)
   assert(info.last_update >= log.bottom);  // otherwise we need some help!
 
   if (log.backlog) {
+
+    // FIXME: sloppy pobject vs object conversions abound!  ***
+    
     // be thorough.
-    list<object_t> ls;
+    list<pobject_t> ls;
     osd->store->collection_list(info.pgid, ls);
     set<object_t> s;
     
-    for (list<object_t>::iterator i = ls.begin();
+    for (list<pobject_t>::iterator i = ls.begin();
          i != ls.end();
          i++) 
-      s.insert(*i);
+      s.insert(i->oid); 
 
     set<object_t> did;
     for (list<Log::Entry>::reverse_iterator p = log.log.rbegin();
diff --git a/branches/ebofs/osd/osd_types.h b/branches/ebofs/osd/osd_types.h
index abfc7ad7da1e5..babc9fe4c8d2e 100644
--- a/branches/ebofs/osd/osd_types.h
+++ b/branches/ebofs/osd/osd_types.h
@@ -17,6 +17,7 @@
 
 #include "msg/msg_types.h"
 #include "include/types.h"
+#include "include/pobject.h"
 
 /* osdreqid_t - caller name + incarnation# + tid to unique identify this request
  * use for metadata and osd ops.
@@ -67,8 +68,6 @@ typedef uint64_t coll_t;        // collection id
 
 // pg stuff
 
-#define PG_INO 4    // this should match mds/mdstypes.h MDS_INO_PG
-
 typedef uint16_t ps_t;
 typedef uint8_t pruleset_t;
 
@@ -119,7 +118,11 @@ public:
   */
   operator uint64_t() const { return u.pg64; }
 
-  object_t to_object() const { return object_t(PG_INO, u.pg64 >> 32, u.pg64 & 0xffffffff); }
+  pobject_t to_object() const { 
+    return pobject_t(1,  // volume 1 == osd metadata, for now
+		     0,
+		     object_t(u.pg64, 0));
+  }
 };
 
 inline ostream& operator<<(ostream& out, pg_t pg) 
-- 
2.39.5