Whamcloud - gitweb
- changes with names on exports and another fields in llite and mds. lov_exp is
[fs/lustre-release.git] / lustre / mds / mds_open.c
index 80728da..2209c17 100644 (file)
@@ -28,6 +28,7 @@
 #endif
 #define DEBUG_SUBSYSTEM S_MDS
 
+#include <linux/version.h>
 #include <linux/module.h>
 #include <linux/init.h>
 #include <linux/version.h>
@@ -39,7 +40,6 @@
 #endif
 
 #include <linux/obd_class.h>
-#include <linux/lustre_mds.h>
 #include <linux/obd_lov.h>
 #include <linux/lustre_fsfilt.h>
 #include <linux/lprocfs_status.h>
@@ -232,7 +232,7 @@ int mds_query_write_access(struct inode *inode)
         RETURN(atomic_read(&inode->i_writecount));
 }
 
-/* This replaces the VFS mds_dentry_open, it manages mfd and writecount */
+/* This replaces the VFS dentry_open, it manages mfd and writecount */
 static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
                                              struct vfsmount *mnt, int flags,
                                              struct ptlrpc_request *req)
@@ -267,7 +267,10 @@ static struct mds_file_data *mds_dentry_open(struct dentry *dentry,
         dget(dentry);
 
         /* Mark the file as open to handle open-unlink. */
-        mds_open_orphan_inc(dentry->d_inode);
+
+        DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
+        mds_orphan_open_inc(dentry->d_inode);
+        UP_WRITE_I_ALLOC_SEM(dentry->d_inode);
 
         mfd->mfd_mode = flags;
         mfd->mfd_dentry = dentry;
@@ -303,13 +306,15 @@ static void mds_objids_from_lmm(obd_id *ids, struct lov_mds_md *lmm,
 static int mds_create_objects(struct ptlrpc_request *req, int offset,
                               struct mds_update_record *rec,
                               struct mds_obd *mds, struct obd_device *obd,
-                              struct inode *inode, void **handle, obd_id **ids)
+                              struct dentry *dchild, void **handle, 
+                              obd_id **ids)
 {
-        struct obdo *oa;
+        struct obdo *oa = NULL;
         struct obd_trans_info oti = { 0 };
         struct mds_body *body;
         struct lov_stripe_md *lsm = NULL;
         struct lov_mds_md *lmm = NULL;
+        struct inode *inode = dchild->d_inode;
         void *lmm_buf;
         int rc, lmm_bufsize, lmm_size;
         ENTRY;
@@ -325,27 +330,28 @@ static int mds_create_objects(struct ptlrpc_request *req, int offset,
         if (body->valid & OBD_MD_FLEASIZE)
                 RETURN(0);
 
-        OBD_ALLOC(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
+        OBD_ALLOC(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
         if (*ids == NULL)
                 RETURN(-ENOMEM);
         oti.oti_objid = *ids;
 
-        if (*handle == NULL)
-                *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
-        if (IS_ERR(*handle)) {
-                rc = PTR_ERR(*handle);
-                *handle = NULL;
-                GOTO(out_ids, rc);
-        }
-
         /* replay case */
-        if (rec->ur_fid2->id) {
+        if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
+                LASSERT(id_ino(rec->ur_id2));
                 body->valid |= OBD_MD_FLBLKSZ | OBD_MD_FLEASIZE;
                 lmm_size = rec->ur_eadatalen;
                 lmm = rec->ur_eadata;
                 LASSERT(lmm);
 
-                mds_objids_from_lmm(*ids, lmm, &mds->mds_lov_desc);
+                if (*handle == NULL)
+                        *handle = fsfilt_start(obd,inode,FSFILT_OP_CREATE,NULL);
+                if (IS_ERR(*handle)) {
+                        rc = PTR_ERR(*handle);
+                        *handle = NULL;
+                        GOTO(out_oa, rc);
+                }
+
+                mds_objids_from_lmm(*ids, lmm, &mds->mds_dt_desc);
 
                 lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
                 lmm_bufsize = req->rq_repmsg->buflens[offset];
@@ -358,31 +364,40 @@ static int mds_create_objects(struct ptlrpc_request *req, int offset,
                 RETURN(0);
         }
 
+        if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_ALLOC_OBDO))
+                GOTO(out_ids, rc = -ENOMEM);
+
         oa = obdo_alloc();
         if (oa == NULL)
                 GOTO(out_ids, rc = -ENOMEM);
         oa->o_mode = S_IFREG | 0600;
         oa->o_id = inode->i_ino;
+        oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
         oa->o_generation = inode->i_generation;
         oa->o_uid = 0; /* must have 0 uid / gid on OST */
         oa->o_gid = 0;
         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FLTYPE |
-                OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID;
+                OBD_MD_FLMODE | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGROUP;
         oa->o_size = 0;
 
         obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|OBD_MD_FLMTIME|
                         OBD_MD_FLCTIME);
 
         if (!(rec->ur_flags & MDS_OPEN_HAS_OBJS)) {
-                /* check if things like lstripe/lfs stripe are sending us the ea */
+                /* check if things like lfs setstripe are sending us the ea */
                 if (rec->ur_flags & MDS_OPEN_HAS_EA) {
-                        rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE, 
-                                           mds->mds_osc_exp,
+                        rc = obd_iocontrol(OBD_IOC_LOV_SETSTRIPE,
+                                           mds->mds_dt_exp,
                                            0, &lsm, rec->ur_eadata);
                         if (rc)
                                 GOTO(out_oa, rc);
-                }
-                rc = obd_create(mds->mds_osc_exp, oa, &lsm, &oti);
+                } else {
+                        /* Per-directory striping default code removed, because
+                         * it uses the same unnamed EA storage as the directory
+                         * striping for CMD. -p */
+                } 
+                LASSERT(oa->o_gr >= FILTER_GROUP_FIRST_MDS);
+                rc = obd_create(mds->mds_dt_exp, oa, &lsm, &oti);
                 if (rc) {
                         int level = D_ERROR;
                         if (rc == -ENOSPC)
@@ -398,18 +413,19 @@ static int mds_create_objects(struct ptlrpc_request *req, int offset,
                         GOTO(out_oa, rc);
                 }
         } else {
-                rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_osc_exp,
+                rc = obd_iocontrol(OBD_IOC_LOV_SETEA, mds->mds_dt_exp,
                                    0, &lsm, rec->ur_eadata);
                 if (rc) {
                         GOTO(out_oa, rc);
                 }
                 lsm->lsm_object_id = oa->o_id;
+                lsm->lsm_object_gr = oa->o_gr;
         }
         if (inode->i_size) {
                 oa->o_size = inode->i_size;
                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE|OBD_MD_FLATIME|
                                 OBD_MD_FLMTIME| OBD_MD_FLCTIME| OBD_MD_FLSIZE);
-                rc = obd_setattr(mds->mds_osc_exp, oa, lsm, &oti);
+                rc = obd_setattr(mds->mds_dt_exp, oa, lsm, &oti);
                 if (rc) {
                         CERROR("error setting attrs for inode %lu: rc %d\n",
                                inode->i_ino, rc);
@@ -426,12 +442,21 @@ static int mds_create_objects(struct ptlrpc_request *req, int offset,
 
         LASSERT(lsm && lsm->lsm_object_id);
         lmm = NULL;
-        rc = obd_packmd(mds->mds_osc_exp, &lmm, lsm);
-        if (!rec->ur_fid2->id)
-                obd_free_memmd(mds->mds_osc_exp, &lsm);
+        rc = obd_packmd(mds->mds_dt_exp, &lmm, lsm);
+        if (!id_ino(rec->ur_id2))
+                obd_free_memmd(mds->mds_dt_exp, &lsm);
         LASSERT(rc >= 0);
         lmm_size = rc;
         body->eadatasize = rc;
+
+        if (*handle == NULL)
+                *handle = fsfilt_start(obd, inode, FSFILT_OP_CREATE, NULL);
+        if (IS_ERR(*handle)) {
+                rc = PTR_ERR(*handle);
+                *handle = NULL;
+                GOTO(out_ids, rc);
+        }
+
         rc = fsfilt_set_md(obd, inode, *handle, lmm, lmm_size);
         lmm_buf = lustre_msg_buf(req->rq_repmsg, offset, 0);
         lmm_bufsize = req->rq_repmsg->buflens[offset];
@@ -439,13 +464,13 @@ static int mds_create_objects(struct ptlrpc_request *req, int offset,
         LASSERT(lmm_bufsize >= lmm_size);
 
         memcpy(lmm_buf, lmm, lmm_size);
-        obd_free_diskmd(mds->mds_osc_exp, &lmm);
+        obd_free_diskmd(mds->mds_dt_exp, &lmm);
  out_oa:
         oti_free_cookies(&oti);
         obdo_free(oa);
  out_ids:
         if (rc) {
-                OBD_FREE(*ids, mds->mds_lov_desc.ld_tgt_count * sizeof(**ids));
+                OBD_FREE(*ids, mds->mds_dt_desc.ld_tgt_count * sizeof(**ids));
                 *ids = NULL;
         }
         RETURN(rc);
@@ -455,21 +480,20 @@ static void reconstruct_open(struct mds_update_record *rec, int offset,
                              struct ptlrpc_request *req,
                              struct lustre_handle *child_lockh)
 {
-        struct ptlrpc_request *oldreq = req->rq_export->exp_outstanding_reply;
         struct mds_export_data *med = &req->rq_export->exp_mds_data;
         struct mds_client_data *mcd = med->med_mcd;
         struct mds_obd *mds = mds_req2mds(req);
         struct mds_file_data *mfd;
         struct obd_device *obd = req->rq_export->exp_obd;
-        struct dentry *parent, *child;
+        struct dentry *parent, *dchild;
         struct ldlm_reply *rep;
         struct mds_body *body;
-        int rc;
         struct list_head *t;
         int put_child = 1;
+        int rc;
         ENTRY;
 
-        LASSERT(offset == 2);                  /* only called via intent */
+        LASSERT(offset == 3); /* only called via intent */
         rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
         body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
 
@@ -483,32 +507,30 @@ static void reconstruct_open(struct mds_update_record *rec, int offset,
                 return; /* error looking up parent or child */
         }
 
-        parent = mds_fid2dentry(mds, rec->ur_fid1, NULL);
-        LASSERT(!IS_ERR(parent));
+        parent = mds_id2dentry(obd, rec->ur_id1, NULL);
+        LASSERTF(!IS_ERR(parent), "lid "DLID4" rc %ld\n", 
+                 OLID4(rec->ur_id1), PTR_ERR(parent));
 
-        child = ll_lookup_one_len(rec->ur_name, parent, rec->ur_namelen - 1);
-        LASSERT(!IS_ERR(child));
+        dchild = ll_lookup_one_len(rec->ur_name, parent, 
+                                   rec->ur_namelen - 1);
+        LASSERTF(!IS_ERR(dchild), "parent "DLID4" child %s rc %ld\n",
+                 OLID4(rec->ur_id1), rec->ur_name, PTR_ERR(dchild));
 
-        if (!child->d_inode) {
+        if (!dchild->d_inode)
                 GOTO(out_dput, 0); /* child not present to open */
-        }
 
         /* At this point, we know we have a child. We'll send
-         * it back _unless_ it not created and open failed.
-         */
+         * it back _unless_ it not created and open failed.  */
         if (intent_disposition(rep, DISP_OPEN_OPEN) &&
             !intent_disposition(rep, DISP_OPEN_CREATE) &&
-            req->rq_status) {
+            req->rq_status)
                 GOTO(out_dput, 0);
-        }
 
         /* get lock (write for O_CREAT, read otherwise) */
-
-        mds_pack_inode2fid(&body->fid1, child->d_inode);
-        mds_pack_inode2body(body, child->d_inode);
-        if (S_ISREG(child->d_inode->i_mode)) {
+        mds_pack_inode2body(obd, body, dchild->d_inode, 0);
+        if (S_ISREG(dchild->d_inode->i_mode)) {
                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
-                                 child->d_inode, 1);
+                                 dchild->d_inode, 1);
 
                 if (rc)
                         LASSERT(rc == req->rq_status);
@@ -517,13 +539,8 @@ static void reconstruct_open(struct mds_update_record *rec, int offset,
                 if (!(body->valid & OBD_MD_FLEASIZE))
                         body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
                                         OBD_MD_FLATIME | OBD_MD_FLMTIME);
-        } else {
-                /* XXX need to check this case */
         }
-
-        /* If we're opening a file without an EA, change to a write
-           lock (unless we already have one). */
-
+        
         /* If we have -EEXIST as the status, and we were asked to create
          * exclusively, we can tell we failed because the file already existed.
          */
@@ -553,24 +570,29 @@ static void reconstruct_open(struct mds_update_record *rec, int offset,
                 mfd = NULL;
         }
 
-        if (oldreq != NULL) {
-                /* if we're not recovering, it had better be found */
-                LASSERT(mfd != NULL);
-        } else if (mfd == NULL) {
+        /* #warning "XXX fixme" bug 2991 */
+        /* Here it used to LASSERT(mfd) if exp_outstanding_reply != NULL.
+         * Now that exp_outstanding_reply is a list, it's just using mfd != NULL
+         * to detect a re-open */
+        if (mfd == NULL) {
                 mntget(mds->mds_vfsmnt);
                 CERROR("Re-opened file \n");
-                mfd = mds_dentry_open(child, mds->mds_vfsmnt,
+                mfd = mds_dentry_open(dchild, mds->mds_vfsmnt,
                                       rec->ur_flags & ~MDS_OPEN_TRUNC, req);
                 if (!mfd) {
                         CERROR("mds: out of memory\n");
                         GOTO(out_dput, req->rq_status = -ENOMEM);
                 }
                 put_child = 0;
+        } else {
+                body->handle.cookie = mfd->mfd_handle.h_cookie;
+                CDEBUG(D_INODE, "resend mfd %p, cookie "LPX64"\n", mfd,
+                       mfd->mfd_handle.h_cookie);
         }
 
  out_dput:
         if (put_child)
-                l_dput(child);
+                l_dput(dchild);
         l_dput(parent);
         EXIT;
 }
@@ -598,13 +620,15 @@ static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
         struct obd_device *obd = req->rq_export->exp_obd;
         struct mds_file_data *mfd = NULL;
         obd_id *ids = NULL; /* object IDs created */
+        unsigned mode;
         int rc = 0;
         ENTRY;
 
         /* atomically create objects if necessary */
         down(&dchild->d_inode->i_sem);
-        if (S_ISREG(dchild->d_inode->i_mode) &&
-            !(body->valid & OBD_MD_FLEASIZE)) {
+        mode = dchild->d_inode->i_mode;
+        if ((S_ISREG(mode) && !(body->valid & OBD_MD_FLEASIZE)) || 
+            (S_ISDIR(mode) && !(body->valid & OBD_MD_FLDIREA))) {
                 rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
                                  dchild->d_inode, 0);
                 if (rc) {
@@ -614,15 +638,35 @@ static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
         }
         if (rec != NULL) {
                 /* no EA: create objects */
-                rc = mds_create_objects(req, 2, rec, mds, obd,
-                                        dchild->d_inode, handle, &ids);
-                if (rc) {
-                        CERROR("mds_create_objects: rc = %d\n", rc);
+                if ((body->valid & OBD_MD_FLEASIZE) &&
+                    (rec->ur_flags & MDS_OPEN_HAS_EA)) {
                         up(&dchild->d_inode->i_sem);
-                        RETURN(rc);
+                        RETURN(-EEXIST);
+                }
+                if (!(body->valid & OBD_MD_FLEASIZE)) {
+                        /* no EA: create objects */
+                        rc = mds_create_objects(req, 2, rec, mds, obd,
+                                                dchild, handle, &ids);
+                        if (rc) {
+                                CERROR("mds_create_objects: rc = %d\n", rc);
+                                up(&dchild->d_inode->i_sem);
+                                RETURN(rc);
+                        }
+                }
+                if (S_ISREG(dchild->d_inode->i_mode) &&
+                    (body->valid & OBD_MD_FLEASIZE)) {
+                        rc = mds_revalidate_lov_ea(obd, dchild->d_inode,
+                                                   req->rq_repmsg, 2);
+                        if (!rc)
+                                rc = mds_pack_md(obd, req->rq_repmsg, 2, body,
+                                                 dchild->d_inode, 0);
+                        if (rc) {
+                                up(&dchild->d_inode->i_sem);
+                                RETURN(rc);
+                        }
                 }
         }
-        /* If the inode has EA data, then OSTs hold size, mtime */
+        /* If the inode has no EA data, then MDSs hold size, mtime */
         if (S_ISREG(dchild->d_inode->i_mode) &&
             !(body->valid & OBD_MD_FLEASIZE)) {
                 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
@@ -638,58 +682,64 @@ static int mds_finish_open(struct ptlrpc_request *req, struct dentry *dchild,
         CDEBUG(D_INODE, "mfd %p, cookie "LPX64"\n", mfd,
                mfd->mfd_handle.h_cookie);
         if (ids != NULL) {
-                mds_lov_update_objids(obd, ids);
-                OBD_FREE(ids, sizeof(*ids) * mds->mds_lov_desc.ld_tgt_count);
+                mds_dt_update_objids(obd, ids);
+                OBD_FREE(ids, sizeof(*ids) * mds->mds_dt_desc.ld_tgt_count);
         }
         //if (rc)
         //        mds_mfd_destroy(mfd);
         RETURN(rc);
 }
 
-static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
-                           struct mds_body *body, int flags,
-                           struct mds_update_record *rec,struct ldlm_reply *rep)
+static int mds_open_by_id(struct ptlrpc_request *req,
+                          struct lustre_id *id,
+                          struct mds_body *body, int flags,
+                          struct mds_update_record *rec,
+                          struct ldlm_reply *rep)
 {
         struct mds_obd *mds = mds_req2mds(req);
         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
         struct dentry *dchild;
-        char fidname[LL_FID_NAMELEN];
-        int fidlen = 0, rc;
+        char idname[LL_ID_NAMELEN];
+        int idlen = 0, rc;
         void *handle = NULL;
         ENTRY;
 
         down(&pending_dir->i_sem);
-        fidlen = ll_fid2str(fidname, fid->id, fid->generation);
-        dchild = lookup_one_len(fidname, mds->mds_pending_dir, fidlen);
+        
+        idlen = ll_id2str(idname, id_ino(id), id_gen(id));
+        
+        dchild = lookup_one_len(idname, mds->mds_pending_dir,
+                                idlen);
         if (IS_ERR(dchild)) {
                 up(&pending_dir->i_sem);
                 rc = PTR_ERR(dchild);
-                CERROR("error looking up %s in PENDING: rc = %d\n",fidname, rc);
+                CERROR("error looking up %s in PENDING: rc = %d\n", 
+                       idname, rc);
                 RETURN(rc);
         }
 
         if (dchild->d_inode != NULL) {
                 up(&pending_dir->i_sem);
                 mds_inode_set_orphan(dchild->d_inode);
-                mds_pack_inode2fid(&body->fid1, dchild->d_inode);
-                mds_pack_inode2body(body, dchild->d_inode);
+                mds_pack_inode2body(req2obd(req), body, dchild->d_inode, 0);
                 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
                 intent_set_disposition(rep, DISP_LOOKUP_POS);
                 CWARN("Orphan %s found and opened in PENDING directory\n",
-                       fidname);
+                       idname);
                 goto open;
         }
         dput(dchild);
         up(&pending_dir->i_sem);
 
-        /* We didn't find it in PENDING so it isn't an orphan.  See
-         * if it was a regular inode that was previously created. */
-        dchild = mds_fid2dentry(mds, fid, NULL);
+        /*
+         * we didn't find it in PENDING so it isn't an orphan.  See if it was a
+         * regular inode that was previously created.
+         */
+        dchild = mds_id2dentry(req2obd(req), id, NULL);
         if (IS_ERR(dchild))
                 RETURN(PTR_ERR(dchild));
 
-        mds_pack_inode2fid(&body->fid1, dchild->d_inode);
-        mds_pack_inode2body(body, dchild->d_inode);
+        mds_pack_inode2body(req2obd(req), body, dchild->d_inode, 0);
         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
         intent_set_disposition(rep, DISP_LOOKUP_POS);
 
@@ -703,47 +753,63 @@ static int mds_open_by_fid(struct ptlrpc_request *req, struct ll_fid *fid,
         RETURN(rc);
 }
 
-int mds_pin(struct ptlrpc_request *req)
+int mds_pin(struct ptlrpc_request *req, int offset)
 {
         struct obd_device *obd = req->rq_export->exp_obd;
         struct mds_body *request_body, *reply_body;
-        struct obd_run_ctxt saved;
+        struct lvfs_run_ctxt saved;
         int rc, size = sizeof(*reply_body);
         ENTRY;
 
-        request_body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*request_body));
+        request_body = lustre_msg_buf(req->rq_reqmsg, offset,
+                                      sizeof(*request_body));
 
         rc = lustre_pack_reply(req, 1, &size, NULL);
         if (rc)
                 RETURN(rc);
         reply_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply_body));
 
-        push_ctxt(&saved, &obd->obd_ctxt, NULL);
-        rc = mds_open_by_fid(req, &request_body->fid1, reply_body,
-                             request_body->flags, NULL, NULL);
-        pop_ctxt(&saved, &obd->obd_ctxt, NULL);
+        push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
+        rc = mds_open_by_id(req, &request_body->id1, reply_body,
+                            request_body->flags, NULL, NULL);
+        pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 
         RETURN(rc);
 }
 
-/*  Get a lock on the ino to sync with creation WRT inode reuse (bug 2029).
- *  If child_lockh is NULL we just get the lock as a barrier to wait for
- *  other holders of this lock, and drop it right away again. */
+/*
+ * get a lock on the ino to sync with creation WRT inode reuse (bug 2029). If
+ * child_lockh is NULL we just get the lock as a barrier to wait for other
+ * holders of this lock, and drop it right away again.
+ */
 int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
                        struct lustre_handle *child_lockh)
 {
-        struct ldlm_res_id child_res_id = { .name = { inode->i_ino, 0, 1, 0 } };
+        struct ldlm_res_id child_res_id = { .name = { 0, 0, 0, 0 } };
         struct lustre_handle lockh;
+        struct lustre_id sid;
         int lock_flags = 0;
         int rc;
 
+        down(&inode->i_sem);
+        rc = mds_read_inode_sid(obd, inode, &sid);
+        up(&inode->i_sem);
+        if (rc) {
+                CERROR("Can't read inode self id, err = %d\n", rc);
+                RETURN(rc);
+        }
+
+        child_res_id.name[0] = id_fid(&sid);
+        child_res_id.name[1] = id_group(&sid);
+        child_res_id.name[2] = 1;
+
         if (child_lockh == NULL)
                 child_lockh = &lockh;
 
-        rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
-                              child_res_id, LDLM_PLAIN, NULL, 0,
-                              LCK_EX, &lock_flags, ldlm_completion_ast,
-                              mds_blocking_ast, NULL, child_lockh);
+        rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, child_res_id,
+                              LDLM_PLAIN, NULL, LCK_EX, &lock_flags,
+                              mds_blocking_ast, ldlm_completion_ast, NULL,
+                              NULL, NULL, 0, NULL, child_lockh);
         if (rc != ELDLM_OK)
                 CERROR("ldlm_cli_enqueue: %d\n", rc);
         else if (child_lockh == &lockh)
@@ -752,6 +818,21 @@ int mds_lock_new_child(struct obd_device *obd, struct inode *inode,
         RETURN(rc);
 }
 
+static int is_mount_object(struct dentry *dparent)
+{
+        struct dentry *dchild;
+
+        if (!(dparent->d_inode->i_mode & S_ISUID))
+                return 0;
+
+        dchild = lookup_one_len(".mntinfo", dparent, strlen(".mntinfo"));
+        if (IS_ERR(dchild) || dchild == NULL)
+                return 0;
+
+        dput(dchild);
+        return 1;
+}
+
 int mds_open(struct mds_update_record *rec, int offset,
              struct ptlrpc_request *req, struct lustre_handle *child_lockh)
 {
@@ -762,17 +843,23 @@ int mds_open(struct mds_update_record *rec, int offset,
         struct mds_body *body = NULL;
         struct dentry *dchild = NULL, *dparent = NULL;
         struct mds_export_data *med;
-        struct lustre_handle parent_lockh;
+        struct lustre_handle parent_lockh[2] = {{0}, {0}};
         int rc = 0, cleanup_phase = 0, acc_mode, created = 0;
         int parent_mode = LCK_PR;
         void *handle = NULL;
         struct dentry_params dp;
+        struct mea *mea = NULL;
+        int mea_size, update_mode;
         ENTRY;
 
-        if (offset == 2) { /* intent */
+        DEBUG_REQ(D_INODE, req, "parent "DLID4" name %*s mode %o",
+                  OLID4(rec->ur_id1), rec->ur_namelen - 1, rec->ur_name,
+                  rec->ur_mode);
+
+        if (offset == 3) { /* intent */
                 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
                 body = lustre_msg_buf(req->rq_repmsg, 1, sizeof (*body));
-        } else if (offset == 0) { /* non-intent reint */
+        } else if (offset == 1) { /* non-intent reint */
                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
         } else {
                 body = NULL;
@@ -781,19 +868,31 @@ int mds_open(struct mds_update_record *rec, int offset,
 
         MDS_CHECK_RESENT(req, reconstruct_open(rec, offset, req, child_lockh));
 
-        /* Step 0: If we are passed a fid, then we assume the client already
-         * opened this file and is only replaying the RPC, so we open the
-         * inode by fid (at some large expense in security). */
-        if (rec->ur_fid2->id) {
-                rc = mds_open_by_fid(req, rec->ur_fid2, body, rec->ur_flags,
-                                     rec, rep);
+        /*
+         * Step 0: If we are passed a id, then we assume the client already
+         * opened this file and is only replaying the RPC, so we open the inode
+         * by fid (at some large expense in security).
+         */
+        if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
+                CDEBUG(D_HA, "open obj "DLID4" name %*s mode %o\n",
+                       OLID4(rec->ur_id2), rec->ur_namelen - 1, 
+                       rec->ur_name, rec->ur_mode);
+
+                LASSERT(id_ino(rec->ur_id2));
+
+                rc = mds_open_by_id(req, rec->ur_id2, body, rec->ur_flags,
+                                    rec, rep);
                 if (rc != -ENOENT)
                         RETURN(rc);
+
                 /* We didn't find the correct inode on disk either, so we
                  * need to re-create it via a regular replay. */
                 LASSERT(rec->ur_flags & MDS_OPEN_CREAT);
+        } else {
+                LASSERT(!id_ino(rec->ur_id2));
         }
-        LASSERT(offset == 2); /* If we got here, we must be called via intent */
+
+        LASSERT(offset == 3); /* If we got here, we must be called via intent */
 
         med = &req->rq_export->exp_mds_data;
         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
@@ -804,20 +903,62 @@ int mds_open(struct mds_update_record *rec, int offset,
         acc_mode = accmode(rec->ur_flags);
 
         /* Step 1: Find and lock the parent */
-        if (rec->ur_flags & O_CREAT)
+        if (rec->ur_flags & O_CREAT) {
+                /* XXX Well, in fact we only need this lock mode change if
+                   in addition to O_CREAT, the file does not exist.
+                   But we do not know if it exists or not yet */
                 parent_mode = LCK_PW;
-        dparent = mds_fid2locked_dentry(obd, rec->ur_fid1, NULL, parent_mode,
-                                        &parent_lockh, rec->ur_name,
-                                        rec->ur_namelen - 1);
+        }
+        
+        if (rec->ur_namelen == 1) {
+                /* client (LMV) wants to open the file by id */
+                CDEBUG(D_OTHER, "OPEN by id "DLID4"\n", OLID4(rec->ur_id1));
+                dchild = mds_id2dentry(obd, rec->ur_id1, NULL);
+                if (IS_ERR(dchild)) {
+                        rc = PTR_ERR(dparent);
+                        CERROR("child lookup by an id error %d\n", rc);
+                        GOTO(cleanup, rc);
+                }
+                goto got_child;
+        }
+        
+        dparent = mds_id2locked_dentry(obd, rec->ur_id1, NULL, parent_mode,
+                                       parent_lockh, &update_mode, rec->ur_name,
+                                       rec->ur_namelen - 1, MDS_INODELOCK_UPDATE);
         if (IS_ERR(dparent)) {
                 rc = PTR_ERR(dparent);
-                CERROR("parent lookup error %d\n", rc);
+                if (rc != -ENOENT)
+                        CERROR("parent lookup for "DLID4" failed, error %d\n",
+                               OLID4(rec->ur_id1), rc);
                 GOTO(cleanup, rc);
         }
         LASSERT(dparent->d_inode != NULL);
 
         cleanup_phase = 1; /* parent dentry and lock */
 
+        /* try to retrieve MEA data for this dir */
+        rc = mds_md_get_attr(obd, dparent->d_inode, &mea, &mea_size);
+        if (rc)
+                GOTO(cleanup, rc);
+       
+        if (mea != NULL) {
+                /*
+                 * dir is already splitted, check is requested filename should *
+                 * live at this MDS or at another one.
+                 */
+                int i;
+                i = mea_name2idx(mea, rec->ur_name, rec->ur_namelen - 1);
+                if (mea->mea_master != id_group(&mea->mea_ids[i])) {
+                        CDEBUG(D_OTHER,
+                               "%s: inapropriate MDS(%d) for %lu/%u:%s."
+                               " should be %lu(%d)\n", obd->obd_name,
+                               mea->mea_master, dparent->d_inode->i_ino,
+                               dparent->d_inode->i_generation, rec->ur_name,
+                               (unsigned long)id_group(&mea->mea_ids[i]), i);
+                        GOTO(cleanup, rc = -ERESTART);
+                }
+        }
+
         /* Step 2: Lookup the child */
         dchild = ll_lookup_one_len(rec->ur_name, dparent, rec->ur_namelen - 1);
         if (IS_ERR(dchild)) {
@@ -826,22 +967,76 @@ int mds_open(struct mds_update_record *rec, int offset,
                 GOTO(cleanup, rc);
         }
 
+got_child:
         cleanup_phase = 2; /* child dentry */
 
         intent_set_disposition(rep, DISP_LOOKUP_EXECD);
+
+        if (dchild->d_flags & DCACHE_CROSS_REF) {
+                struct ldlm_res_id res_id = { . name = {0} };
+                ldlm_policy_data_t policy;
+                int flags = 0;
+
+                mds_pack_dentry2body(obd, body, dchild, 1);
+                intent_set_disposition(rep, DISP_LOOKUP_POS);
+
+                CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
+                       OLID4(&body->id1));
+                
+                res_id.name[0] = dchild->d_fid;
+                res_id.name[1] = dchild->d_mdsnum;
+                
+                policy.l_inodebits.bits = MDS_INODELOCK_LOOKUP;
+                rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
+                                      res_id, LDLM_IBITS, &policy,
+                                      LCK_PR, &flags, mds_blocking_ast,
+                                      ldlm_completion_ast, NULL, NULL,
+                                      NULL, 0, NULL, child_lockh);
+#ifdef S_PDIROPS
+                if (parent_lockh[1].cookie != 0)
+                        ldlm_lock_decref(parent_lockh + 1, update_mode);
+#endif
+                ldlm_lock_decref(parent_lockh, parent_mode);
+                if (mea)
+                        OBD_FREE(mea, mea_size);
+                l_dput(dchild);
+                l_dput(dparent);
+                RETURN(rc);
+        }
+        
         if (dchild->d_inode)
                 intent_set_disposition(rep, DISP_LOOKUP_POS);
         else
                 intent_set_disposition(rep, DISP_LOOKUP_NEG);
 
-        /*Step 3: If the child was negative, and we're supposed to, create it.*/
+        /* Step 3: If the child was negative, and we're supposed to, create it.*/
         if (dchild->d_inode == NULL) {
-                unsigned long ino = rec->ur_fid2->id;
+                unsigned long ino;
                 struct iattr iattr;
                 struct inode *inode;
+                
+                ino = (unsigned long)id_ino(rec->ur_id2);
+                
+                rc = mds_try_to_split_dir(obd, dparent, &mea, 0,
+                                          update_mode);
+                
+                CDEBUG(D_OTHER, "%s: splitted %lu/%u - %d\n",
+                       obd->obd_name, dparent->d_inode->i_ino,
+                       dparent->d_inode->i_generation, rc);
+
+                if (rc > 0) {
+                        /* dir got splitted */
+                        GOTO(cleanup, rc = -ERESTART);
+                } else if (rc < 0) {
+                        /* error happened during spitting */
+                        GOTO(cleanup, rc);
+                }
 
                 if (!(rec->ur_flags & MDS_OPEN_CREAT)) {
-                        /* It's negative and we weren't supposed to create it */
+                        /*
+                         * dentry is negative and we weren't supposed to create
+                         * object, get out of here.
+                         */
                         GOTO(cleanup, rc = -ENOENT);
                 }
 
@@ -857,7 +1052,7 @@ int mds_open(struct mds_update_record *rec, int offset,
                 dp.p_ptr = req;
                 dp.p_inum = ino;
 
-                rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode,NULL);
+                rc = ll_vfs_create(dparent->d_inode, dchild, rec->ur_mode, NULL);
                 if (dchild->d_fsdata == (void *)(unsigned long)ino)
                         dchild->d_fsdata = NULL;
 
@@ -868,8 +1063,9 @@ int mds_open(struct mds_update_record *rec, int offset,
                 inode = dchild->d_inode;
                 if (ino) {
                         LASSERT(ino == inode->i_ino);
-                        /* Written as part of setattr */
-                        inode->i_generation = rec->ur_fid2->generation;
+                        
+                        /* written as part of setattr */
+                        inode->i_generation = id_gen(rec->ur_id2);
                         CDEBUG(D_HA, "recreated ino %lu with gen %u\n",
                                inode->i_ino, inode->i_generation);
                 }
@@ -897,14 +1093,53 @@ int mds_open(struct mds_update_record *rec, int offset,
                 rc = fsfilt_setattr(obd, dparent, handle, &iattr, 0);
                 if (rc)
                         CERROR("error on parent setattr: rc = %d\n", rc);
+                else {
+                        MD_COUNTER_INCREMENT(obd, create);
+                }
 
-                acc_mode = 0;                  /* Don't check for permissions */
-        }
-
-        LASSERT(!mds_inode_is_orphan(dchild->d_inode));
+                down(&dchild->d_inode->i_sem);
+                if (ino) {
+                        rc = mds_update_inode_sid(obd, dchild->d_inode,
+                                                  handle, rec->ur_id2);
+                        if (rc) {
+                                CERROR("mds_update_inode_sid() failed, "
+                                       "rc = %d\n", rc);
+                        }
+                        id_assign_fid(&body->id1, rec->ur_id2);
+
+                        /* 
+                         * make sure, that fid is up-to-date.
+                         */
+                        mds_set_last_fid(obd, id_fid(rec->ur_id2));
+                } else {
+                        rc = mds_alloc_inode_sid(obd, dchild->d_inode,
+                                                 handle, &body->id1);
+                        if (rc) {
+                                CERROR("mds_alloc_inode_sid() failed, "
+                                       "rc = %d\n", rc);
+                        }
+                }
+                up(&dchild->d_inode->i_sem);
+                
+                if (!(rec->ur_flags & O_EXCL)) { /* bug 3313 */
+                        rc = fsfilt_commit(obd, dchild->d_inode->i_sb,
+                                           dchild->d_inode, handle, 0);
+                        handle = NULL;
+                }
 
-        mds_pack_inode2fid(&body->fid1, dchild->d_inode);
-        mds_pack_inode2body(body, dchild->d_inode);
+                acc_mode = 0;           /* Don't check for permissions */
+                
+                /* 
+                 * we do not read fid from EA here, because it is already
+                 * updated and thus we avoid not needed IO, locking, etc.
+                 */
+                mds_pack_inode2body(obd, body, dchild->d_inode, 0);
+        } else {
+                mds_pack_inode2body(obd, body, dchild->d_inode, 1);
+        }
+        LASSERTF(!mds_inode_is_orphan(dchild->d_inode),
+                 "dchild %*s (%p) inode %p\n", dchild->d_name.len,
+                 dchild->d_name.name, dchild, dchild->d_inode);
 
         if (S_ISREG(dchild->d_inode->i_mode)) {
                 /* Check permissions etc */
@@ -939,9 +1174,26 @@ int mds_open(struct mds_update_record *rec, int offset,
             !S_ISDIR(dchild->d_inode->i_mode))
                 GOTO(cleanup, rc = -ENOTDIR);
 
+        if (S_ISDIR(dchild->d_inode->i_mode)) {
+                if (rec->ur_flags & MDS_OPEN_CREAT ||
+                    rec->ur_flags & FMODE_WRITE) {
+                        /*we are trying to create or write a exist dir*/
+                        GOTO(cleanup, rc = -EISDIR);
+                }
+                if (ll_permission(dchild->d_inode, acc_mode, NULL))
+                        GOTO(cleanup, rc = -EACCES);
+
+                if (is_mount_object(dchild)) {
+                        CERROR("Found possible GNS mount object %*s; not "
+                               "opening.\n", dchild->d_name.len,
+                               dchild->d_name.name);
+                        GOTO(cleanup, rc = 0); // success, but don't really open
+                }
+        }
+
         /* Step 5: mds_open it */
-        rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle, rec,
-                             rep);
+        rc = mds_finish_open(req, dchild, body, rec->ur_flags, &handle,
+                             rec, rep);
         GOTO(cleanup, rc);
 
  cleanup:
@@ -966,18 +1218,22 @@ int mds_open(struct mds_update_record *rec, int offset,
                         break;
 
                 l_dput(dparent);
+#ifdef S_PDIROPS
+                if (parent_lockh[1].cookie != 0)
+                        ldlm_lock_decref(parent_lockh + 1, update_mode);
+#endif
                 if (rc)
-                        ldlm_lock_decref(&parent_lockh, parent_mode);
+                        ldlm_lock_decref(parent_lockh, parent_mode);
                 else
-                        ldlm_put_lock_into_req(req, &parent_lockh, parent_mode);
+                        ptlrpc_save_lock (req, parent_lockh, parent_mode);
         }
-        if (rc == 0)
-                atomic_inc(&mds->mds_open_count);
+        if (mea)
+                OBD_FREE(mea, mea_size);
         RETURN(rc);
 }
 
 /* Close a "file descriptor" and possibly unlink an orphan from the
- * PENDING directory.
+ * PENDING directory.  Caller must hold child->i_sem, this drops it. 
  *
  * If we are being called from mds_disconnect() because the client has
  * disappeared, then req == NULL and we do not update last_rcvd because
@@ -985,31 +1241,37 @@ int mds_open(struct mds_update_record *rec, int offset,
  * (it will not even _have_ an entry in last_rcvd anymore).
  *
  * Returns EAGAIN if the client needs to get more data and re-close. */
-int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
-                  struct mds_file_data *mfd, int unlink_orphan)
+int mds_mfd_close(struct ptlrpc_request *req, int offset,
+                  struct obd_device *obd, struct mds_file_data *mfd,
+                  int unlink_orphan)
 {
         struct inode *inode = mfd->mfd_dentry->d_inode;
-        char fidname[LL_FID_NAMELEN];
-        int last_orphan, fidlen, rc = 0, cleanup_phase = 0;
+        char idname[LL_ID_NAMELEN];
+        int last_orphan, idlen, rc = 0, cleanup_phase = 0;
         struct dentry *pending_child = NULL;
         struct mds_obd *mds = &obd->u.mds;
         struct inode *pending_dir = mds->mds_pending_dir->d_inode;
         void *handle = NULL;
         struct mds_body *request_body = NULL, *reply_body = NULL;
         struct dentry_params dp;
+        struct iattr iattr = { 0 };
+        struct llog_create_locks *lcl = NULL;
         ENTRY;
 
-        if (req != NULL) {
-                request_body = lustre_msg_buf(req->rq_reqmsg, 0,
+        if (req && req->rq_reqmsg != NULL)
+                request_body = lustre_msg_buf(req->rq_reqmsg, offset,
                                               sizeof(*request_body));
+        if (req && req->rq_repmsg != NULL)
                 reply_body = lustre_msg_buf(req->rq_repmsg, 0,
                                             sizeof(*reply_body));
-        }
 
-        fidlen = ll_fid2str(fidname, inode->i_ino, inode->i_generation);
+        idlen = ll_id2str(idname, inode->i_ino, inode->i_generation);
+        CDEBUG(D_INODE, "inode %p ino %s nlink %d orphan %d\n", inode, 
+               idname, inode->i_nlink, mds_orphan_open_count(inode));
 
-        last_orphan = mds_open_orphan_dec_test(inode) &&
+        last_orphan = mds_orphan_open_dec_test(inode) &&
                 mds_inode_is_orphan(inode);
+        UP_WRITE_I_ALLOC_SEM(inode);
 
         /* this is half of the actual "close" */
         if (mfd->mfd_mode & FMODE_WRITE) {
@@ -1020,38 +1282,50 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
         }
 
         if (last_orphan && unlink_orphan) {
+                struct lov_mds_md *lmm = NULL;
+                int stripe_count = 0;
                 LASSERT(rc == 0); /* mds_put_write_access must have succeeded */
 
-                CDEBUG(D_HA, "destroying orphan object %s\n", fidname);
+                if (obd->obd_recovering) {
+                        CDEBUG(D_HA, "not remove orphan %s until recovery"
+                               " is over\n", idname);
+                        GOTO(out, rc);
+                }
 
-                /* Sadly, there is no easy way to save pending_child from
-                 * mds_reint_unlink() into mfd, so we need to re-lookup,
-                 * but normally it will still be in the dcache. */
-                pending_dir = mds->mds_pending_dir->d_inode;
+                CDEBUG(D_HA, "destroying orphan object %s\n", idname);
+                
+                if ((S_ISREG(inode->i_mode) && inode->i_nlink != 1) ||
+                    (S_ISDIR(inode->i_mode) && inode->i_nlink != 2))
+                        CERROR("found \"orphan\" %s %s with link count %d\n",
+                               S_ISREG(inode->i_mode) ? "file" : "dir",
+                               idname, inode->i_nlink);
+                /*
+                 * sadly, there is no easy way to save pending_child from
+                 * mds_reint_unlink() into mfd, so we need to re-lookup, but
+                 * normally it will still be in the dcache.
+                 */
                 down(&pending_dir->i_sem);
                 cleanup_phase = 1; /* up(i_sem) when finished */
-                pending_child = lookup_one_len(fidname, mds->mds_pending_dir,
-                                               fidlen);
+                pending_child = lookup_one_len(idname, mds->mds_pending_dir,
+                                               idlen);
                 if (IS_ERR(pending_child))
                         GOTO(cleanup, rc = PTR_ERR(pending_child));
                 LASSERT(pending_child->d_inode != NULL);
 
                 cleanup_phase = 2; /* dput(pending_child) when finished */
-                handle = fsfilt_start(obd, pending_dir, FSFILT_OP_UNLINK_LOG,
-                                      NULL);
+                if (req != NULL && req->rq_repmsg != NULL) {
+                        lmm = lustre_msg_buf(req->rq_repmsg, 1, 0);
+                        stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
+                }
+
+                handle = fsfilt_start_log(obd, pending_dir, FSFILT_OP_UNLINK,
+                                          NULL, stripe_count);
                 if (IS_ERR(handle)) {
                         rc = PTR_ERR(handle);
                         handle = NULL;
                         GOTO(cleanup, rc);
                 }
 
-                if (req != NULL &&
-                    (reply_body->valid & OBD_MD_FLEASIZE) &&
-                    mds_log_op_unlink(obd, pending_child->d_inode,
-                                      req->rq_repmsg, 1) > 0) {
-                        reply_body->valid |= OBD_MD_FLCOOKIE;
-                }
-
                 pending_child->d_fsdata = (void *) &dp;
                 dp.p_inum = 0;
                 dp.p_ptr = req;
@@ -1060,33 +1334,75 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
                 else
                         rc = vfs_unlink(pending_dir, pending_child);
                 if (rc)
-                        CERROR("error unlinking orphan %s: rc %d\n",fidname,rc);
-        } else if (mfd->mfd_mode & FMODE_WRITE && rc == 0) {
+                        CERROR("error unlinking orphan %s: rc %d\n",
+                               idname, rc);
+
+                if (req != NULL && req->rq_repmsg != NULL &&
+                    (reply_body->valid & OBD_MD_FLEASIZE) &&
+                    mds_log_op_unlink(obd, pending_child->d_inode,
+                                      lmm, req->rq_repmsg->buflens[1],
+                                      lustre_msg_buf(req->rq_repmsg, 2, 0),
+                                      req->rq_repmsg->buflens[2], &lcl) > 0) {
+                        reply_body->valid |= OBD_MD_FLCOOKIE;
+                }
+
+                goto out; /* Don't bother updating attrs on unlinked inode */
+        }
+
+#if 0
+        if (request_body != NULL && mfd->mfd_mode & FMODE_WRITE && rc == 0) {
                 /* Update the on-disk attributes if this was the last write
                  * close, and all information was provided (i.e., rc == 0)
                  *
                  * XXX this should probably be abstracted with mds_reint_setattr
                  */
-#if 0
-                struct iattr iattr;
+
+                if (request_body->valid & OBD_MD_FLMTIME &&
+                    LTIME_S(request_body->mtime) > LTIME_S(inode->i_mtime)) {
+                        LTIME_S(iattr.ia_mtime) = LTIME_S(request_body->mtime);
+                        iattr.ia_valid |= ATTR_MTIME;
+                }
+                if (request_body->valid & OBD_MD_FLCTIME &&
+                    LTIME_S(request_body->ctime) > LTIME_S(inode->i_ctime)) {
+                        LTIME_S(iattr.ia_ctime) = LTIME_S(request_body->ctime);
+                        iattr.ia_valid |= ATTR_CTIME;
+                }
 
                 /* XXX can't set block count with fsfilt_setattr (!) */
-                iattr.ia_valid = ATTR_CTIME | ATTR_ATIME |
-                        ATTR_MTIME | ATTR_SIZE;
-                iattr.ia_atime = request_body->atime;
-                iattr.ia_ctime = request_body->ctime;
-                iattr.ia_mtime = request_body->mtime;
-                iattr.ia_size = request_body->size;
-                /* iattr.ia_blocks = request_body->blocks */
+                if (request_body->valid & OBD_MD_FLSIZE) {
+                        iattr.ia_valid |= ATTR_SIZE;
+                        iattr.ia_size = request_body->size;
+                }
+                /* if (request_body->valid & OBD_MD_FLBLOCKS) {
+                        iattr.ia_valid |= ATTR_BLOCKS;
+                        iattr.ia_blocks = request_body->blocks
+                } */
 
+        }
+#endif
+        if (request_body != NULL && request_body->valid & OBD_MD_FLATIME) {
+                /* Only start a transaction to write out only the atime if
+                 * it is more out-of-date than the specified limit.  If we
+                 * are already going to write out the atime then do it anyway.
+                 * */
+                if ((request_body->atime >
+                     LTIME_S(inode->i_atime) + MAX_ATIME_DIFF) ||
+                    (iattr.ia_valid != 0 &&
+                     request_body->atime > LTIME_S(inode->i_atime))) {
+                        LTIME_S(iattr.ia_atime) = request_body->atime;
+                        iattr.ia_valid |= ATTR_ATIME;
+                }
+        }
+
+        if (iattr.ia_valid != 0) {
                 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
                 if (IS_ERR(handle))
                         GOTO(cleanup, rc = PTR_ERR(handle));
                 rc = fsfilt_setattr(obd, mfd->mfd_dentry, handle, &iattr, 0);
                 if (rc)
-                        CERROR("error in setattr(%s): rc %d\n", fidname, rc);
-#endif
+                        CERROR("error in setattr(%s): rc %d\n", idname, rc);
         }
+out:
         /* If other clients have this file open for write, rc will be > 0 */
         if (rc > 0)
                 rc = 0;
@@ -1094,11 +1410,10 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
         mds_mfd_destroy(mfd);
 
  cleanup:
-        atomic_dec(&mds->mds_open_count);
-        if (req) {
+        if (req != NULL && reply_body != NULL) {
                 rc = mds_finish_transno(mds, pending_dir, handle, req, rc, 0);
         } else if (handle) {
-                int err = fsfilt_commit(obd, pending_dir, handle, 0);
+                int err = fsfilt_commit(obd, mds->mds_sb, pending_dir, handle, 0);
                 if (err) {
                         CERROR("error committing close: %d\n", err);
                         if (!rc)
@@ -1108,6 +1423,8 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
 
         switch (cleanup_phase) {
         case 2:
+                if (lcl != NULL)
+                        ptlrpc_save_llog_lock(req, lcl);
                 dput(pending_child);
         case 1:
                 up(&pending_dir->i_sem);
@@ -1115,18 +1432,19 @@ int mds_mfd_close(struct ptlrpc_request *req, struct obd_device *obd,
         RETURN(rc);
 }
 
-int mds_close(struct ptlrpc_request *req)
+int mds_close(struct ptlrpc_request *req, int offset)
 {
         struct mds_export_data *med = &req->rq_export->exp_mds_data;
         struct obd_device *obd = req->rq_export->exp_obd;
         struct mds_body *body;
         struct mds_file_data *mfd;
-        struct obd_run_ctxt saved;
+        struct lvfs_run_ctxt saved;
         struct inode *inode;
         int rc, repsize[3] = {sizeof(struct mds_body),
                               obd->u.mds.mds_max_mdsize,
                               obd->u.mds.mds_max_cookiesize};
         ENTRY;
+        MD_COUNTER_INCREMENT(obd, close);
 
         rc = lustre_pack_reply(req, 3, repsize, NULL);
         if (rc) {
@@ -1136,7 +1454,16 @@ int mds_close(struct ptlrpc_request *req)
                 MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
         }
 
-        body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
+        if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
+                DEBUG_REQ(D_HA, req, "close replay\n");
+                memcpy(lustre_msg_buf(req->rq_repmsg, 2, 0),
+                       lustre_msg_buf(req->rq_reqmsg, offset + 1, 0),
+                       req->rq_repmsg->buflens[2]);
+        }
+
+
+        body = lustre_swab_reqbuf(req, offset, sizeof(*body),
+                                  lustre_swab_mds_body);
         if (body == NULL) {
                 CERROR("Can't unpack body\n");
                 req->rq_status = -EFAULT;
@@ -1149,27 +1476,29 @@ int mds_close(struct ptlrpc_request *req)
         mfd = mds_handle2mfd(&body->handle);
         if (mfd == NULL) {
                 DEBUG_REQ(D_ERROR, req, "no handle for file close ino "LPD64
-                          ": cookie "LPX64, body->fid1.id, body->handle.cookie);
+                          ": cookie "LPX64, body->id1.li_stc.u.e3s.l3s_ino,
+                          body->handle.cookie);
                 req->rq_status = -ESTALE;
                 RETURN(-ESTALE);
         }
 
         inode = mfd->mfd_dentry->d_inode;
-        if (mds_inode_is_orphan(inode) && mds_open_orphan_count(inode) == 1) {
+        /* child i_alloc_sem protects orphan_dec_test && is_orphan race */
+        DOWN_WRITE_I_ALLOC_SEM(inode); /* mds_mfd_close drops this */
+        if (mds_inode_is_orphan(inode) && mds_orphan_open_count(inode) == 1) {
                 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
                 LASSERT(body != NULL);
 
-                mds_pack_inode2fid(&body->fid1, inode);
-                mds_pack_inode2body(body, inode);
-                mds_pack_md(obd, req->rq_repmsg, 1, body, inode, 1);
+                mds_pack_inode2body(obd, body, inode, 0);
+                mds_pack_md(obd, req->rq_repmsg, 1, body, inode, MDS_PACK_MD_LOCK);
         }
         spin_lock(&med->med_open_lock);
         list_del(&mfd->mfd_list);
         spin_unlock(&med->med_open_lock);
 
-        push_ctxt(&saved, &obd->obd_ctxt, NULL);
-        req->rq_status = mds_mfd_close(rc ? NULL : req, obd, mfd, 1);
-        pop_ctxt(&saved, &obd->obd_ctxt, NULL);
+        push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
+        req->rq_status = mds_mfd_close(req, offset, obd, mfd, 1);
+        pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
 
         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
                 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
@@ -1182,7 +1511,7 @@ int mds_close(struct ptlrpc_request *req)
         RETURN(0);
 }
 
-int mds_done_writing(struct ptlrpc_request *req)
+int mds_done_writing(struct ptlrpc_request *req, int offset)
 {
         struct mds_body *body;
         int rc, size = sizeof(struct mds_body);
@@ -1190,7 +1519,8 @@ int mds_done_writing(struct ptlrpc_request *req)
 
         MDS_CHECK_RESENT(req, mds_reconstruct_generic(req));
 
-        body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_mds_body);
+        body = lustre_swab_reqbuf(req, offset, sizeof(*body),
+                                  lustre_swab_mds_body);
         if (body == NULL) {
                 CERROR("Can't unpack body\n");
                 req->rq_status = -EFAULT;