X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmds%2Fmds_lmv.c;h=8d5a43148eea6e0a901a5dd5b4c323e158a9a1d3;hp=c4ea3af242053938e5996def6ec7396e79885d50;hb=926c6309185a25a8ac1541cfa67910325ed8626f;hpb=b5f041057359892d0a77d72639882c0caecc5b24;ds=sidebyside

diff --git a/lustre/mds/mds_lmv.c b/lustre/mds/mds_lmv.c
index c4ea3af..8d5a431 100644
--- a/lustre/mds/mds_lmv.c
+++ b/lustre/mds/mds_lmv.c
@@ -47,9 +47,8 @@
 int mds_lmv_connect(struct obd_device *obd, char * lmv_name)
 {
         struct mds_obd *mds = &obd->u.mds;
-        struct lustre_handle conn = {0,};
-        int valsize, mdsize;
-        int rc;
+        struct lustre_handle conn = {0};
+        int rc, valsize, value;
         ENTRY;
 
         if (IS_ERR(mds->mds_lmv_obd))
@@ -58,20 +57,26 @@ int mds_lmv_connect(struct obd_device *obd, char * lmv_name)
         if (mds->mds_lmv_obd)
                 RETURN(0);
 
+        down(&mds->mds_lmv_sem);
+        if (mds->mds_lmv_obd) {
+                up(&mds->mds_lmv_sem);
+                RETURN(0);
+        }
+
         mds->mds_lmv_obd = class_name2obd(lmv_name);
         if (!mds->mds_lmv_obd) {
                 CERROR("MDS cannot locate LMV %s\n",
                        lmv_name);
                 mds->mds_lmv_obd = ERR_PTR(-ENOTCONN);
-                RETURN(-ENOTCONN);
+                GOTO(err_last, rc = -ENOTCONN);
         }
 
-        rc = obd_connect(&conn, mds->mds_lmv_obd, &obd->obd_uuid);
+        rc = obd_connect(&conn, mds->mds_lmv_obd, &obd->obd_uuid, OBD_OPT_MDS_CONNECTION);
         if (rc) {
                 CERROR("MDS cannot connect to LMV %s (%d)\n",
                        lmv_name, rc);
                 mds->mds_lmv_obd = ERR_PTR(rc);
-                RETURN(rc);
+                GOTO(err_last, rc);
         }
         mds->mds_lmv_exp = class_conn2export(&conn);
         if (mds->mds_lmv_exp == NULL)
@@ -79,47 +84,58 @@ int mds_lmv_connect(struct obd_device *obd, char * lmv_name)
 
         rc = obd_register_observer(mds->mds_lmv_obd, obd);
         if (rc) {
-                CERROR("MDS cannot register as observer of LMV %s (%d)\n",
-                       lmv_name, rc);
+                CERROR("MDS cannot register as observer of LMV %s, "
+                       "rc = %d\n", lmv_name, rc);
                 GOTO(err_discon, rc);
         }
 
         /* retrieve size of EA */
         rc = obd_get_info(mds->mds_lmv_exp, strlen("mdsize"), "mdsize", 
-                          &valsize, &mdsize);
+                          &valsize, &value);
         if (rc) 
                 GOTO(err_reg, rc);
-        if (mdsize > mds->mds_max_mdsize)
-                mds->mds_max_mdsize = mdsize;
+
+        if (value > mds->mds_max_mdsize)
+                mds->mds_max_mdsize = value;
 
         /* find our number in LMV cluster */
         rc = obd_get_info(mds->mds_lmv_exp, strlen("mdsnum"), "mdsnum", 
-                          &valsize, &mdsize);
+                          &valsize, &value);
         if (rc) 
                 GOTO(err_reg, rc);
-        mds->mds_num = mdsize;
+        
+        mds->mds_num = value;
 
+        rc = obd_set_info(mds->mds_lmv_exp, strlen("inter_mds"),
+                          "inter_mds", 0, NULL);
+        if (rc)
+                GOTO(err_reg, rc);
+
+        up(&mds->mds_lmv_sem);
 	RETURN(0);
 
 err_reg:
-        RETURN(rc);
-
+        obd_register_observer(mds->mds_lmv_obd, NULL);
 err_discon:
-        /* FIXME: cleanups here! */
         obd_disconnect(mds->mds_lmv_exp, 0);
         mds->mds_lmv_exp = NULL;
         mds->mds_lmv_obd = ERR_PTR(rc);
+err_last:
+        up(&mds->mds_lmv_sem);
         RETURN(rc);
 }
 
 int mds_lmv_postsetup(struct obd_device *obd)
 {
         struct mds_obd *mds = &obd->u.mds;
+        int rc = 0;
         ENTRY;
+
         if (mds->mds_lmv_exp)
-                obd_init_ea_size(mds->mds_lmv_exp, mds->mds_max_mdsize,
-                                 mds->mds_max_cookiesize);
-        RETURN(0);
+                rc = obd_init_ea_size(mds->mds_lmv_exp, mds->mds_max_mdsize,
+                                      mds->mds_max_cookiesize);
+        
+        RETURN(rc);
 }
 
 int mds_lmv_disconnect(struct obd_device *obd, int flags)
@@ -128,26 +144,27 @@ int mds_lmv_disconnect(struct obd_device *obd, int flags)
         int rc = 0;
         ENTRY;
 
+        down(&mds->mds_lmv_sem);
         if (!IS_ERR(mds->mds_lmv_obd) && mds->mds_lmv_exp != NULL) {
-
                 obd_register_observer(mds->mds_lmv_obd, NULL);
 
+                /* if obd_disconnect fails (probably because the export was
+                 * disconnected by class_disconnect_exports) then we just need
+                 * to drop our ref. */
                 rc = obd_disconnect(mds->mds_lmv_exp, flags);
-                /* if obd_disconnect fails (probably because the
-                 * export was disconnected by class_disconnect_exports)
-                 * then we just need to drop our ref. */
-                if (rc != 0)
+                if (rc)
                         class_export_put(mds->mds_lmv_exp);
+                
                 mds->mds_lmv_exp = NULL;
                 mds->mds_lmv_obd = NULL;
         }
+        up(&mds->mds_lmv_sem);
 
         RETURN(rc);
 }
 
-
 int mds_get_lmv_attr(struct obd_device *obd, struct inode *inode,
-			struct mea **mea, int *mea_size)
+                     struct mea **mea, int *mea_size)
 {
         struct mds_obd *mds = &obd->u.mds;
 	int rc;
@@ -155,23 +172,23 @@ int mds_get_lmv_attr(struct obd_device *obd, struct inode *inode,
 
 	if (!mds->mds_lmv_obd)
 		RETURN(0);
+        if (!S_ISDIR(inode->i_mode))
+                RETURN(0);
 
 	/* first calculate mea size */
         *mea_size = obd_alloc_diskmd(mds->mds_lmv_exp,
-                                     (struct lov_mds_md **) mea);
-	/* FIXME: error handling here */
-	LASSERT(*mea != NULL);
-
-	down(&inode->i_sem);
-	rc = fsfilt_get_md(obd, inode, *mea, *mea_size);
-	up(&inode->i_sem);
-	/* FIXME: error handling here */
+                                     (struct lov_mds_md **)mea);
+        if (*mea_size < 0 || *mea == NULL)
+                return *mea_size < 0 ? *mea_size : -EINVAL;
+
+        rc = mds_get_md(obd, inode, *mea, mea_size, 1);
+
 	if (rc <= 0) {
 		OBD_FREE(*mea, *mea_size);
 		*mea = NULL;
-	}
-        if (rc > 0)
+	} else
                 rc = 0;
+
 	RETURN(rc);
 }
 
@@ -226,6 +243,7 @@ static int dc_new_page_to_cache(struct dir_cache * dirc)
 
 static int retrieve_generation_numbers(struct dirsplit_control *dc, void *buf)
 {
+        struct mds_obd *mds = &dc->obd->u.mds;
         struct dir_entry *de;
         struct dentry *dentry;
         char * end;
@@ -233,20 +251,29 @@ static int retrieve_generation_numbers(struct dirsplit_control *dc, void *buf)
         end = buf + PAGE_SIZE;
         de = (struct dir_entry *) buf;
         while ((char *) de < end && de->namelen) {
-                LASSERT(de->namelen <= 255);
                 /* lookup an inode */
+                LASSERT(de->namelen <= 255);
                 dentry = ll_lookup_one_len(de->name, dc->dentry, de->namelen);
                 if (IS_ERR(dentry)) {
-                        CERROR("can't lookup '%*s'/%u in %lu: %d\n",
-                                (int) de->namelen, de->name,
-                                (unsigned) de->namelen,
-                                (unsigned long) dc->dentry->d_inode->i_ino,
-                                (int) PTR_ERR(dentry));
+                        CERROR("can't lookup %*s: %d\n", de->namelen,
+                               de->name, (int) PTR_ERR(dentry));
+                        goto next;
+                }
+                if (dentry->d_inode != NULL) {
+                        de->mds = mds->mds_num;
+                        de->ino = dentry->d_inode->i_ino;
+                        de->generation = dentry->d_inode->i_generation;
+                } else if (dentry->d_flags & DCACHE_CROSS_REF) {
+                        de->mds = dentry->d_mdsnum;
+                        de->ino = dentry->d_inum;
+                        de->generation = dentry->d_generation;
+                } else {
+                        CERROR("can't lookup %*s\n", de->namelen, de->name);
+                        goto next;
                 }
-                LASSERT(!IS_ERR(dentry));
-                LASSERT(dentry->d_inode != NULL);
-                de->generation = dentry->d_inode->i_generation;
                 l_dput(dentry);
+
+next:
                 de = (struct dir_entry *)
                         ((char *) de + DIR_REC_LEN(de->namelen));
         }
@@ -256,8 +283,9 @@ static int retrieve_generation_numbers(struct dirsplit_control *dc, void *buf)
 static int flush_buffer_onto_mds(struct dirsplit_control *dc, int mdsnum)
 {
         struct mds_obd *mds = &dc->obd->u.mds;
-        struct dir_cache *ca;
         struct list_head *cur, *tmp;
+        struct dir_cache *ca;
+        int rc;
         ENTRY; 
         ca = dc->cache + mdsnum;
 
@@ -280,12 +308,51 @@ static int flush_buffer_onto_mds(struct dirsplit_control *dc, int mdsnum)
                 ca->brwc.count = PAGE_SIZE;
                 ca->brwc.flag = 0;
                 ca->oa.o_mds = mdsnum;
-                obd_brw(OBD_BRW_WRITE, mds->mds_lmv_exp, &ca->oa,
-                                (struct lov_stripe_md *) dc->mea,
-                                1, &ca->brwc, NULL);
+                rc = obd_brw(OBD_BRW_WRITE, mds->mds_lmv_exp, &ca->oa,
+                             (struct lov_stripe_md *) dc->mea,
+                             1, &ca->brwc, NULL);
+                if (rc)
+                        RETURN(rc);
 
-                list_del(&page->list);
-                __free_page(page);
+        }
+        RETURN(0);
+}
+
+static int remove_entries_from_orig_dir(struct dirsplit_control *dc, int mdsnum)
+{
+        struct list_head *cur, *tmp;
+        struct dentry *dentry;
+        struct dir_cache *ca;
+        struct dir_entry *de;
+        struct page *page;
+        char *buf, *end;
+        int rc;
+        ENTRY; 
+
+        ca = dc->cache + mdsnum;
+        list_for_each_safe(cur, tmp, &ca->list) {
+                page = list_entry(cur, struct page, list);
+                buf = page_address(page);
+                end = buf + PAGE_SIZE;
+
+                de = (struct dir_entry *) buf;
+                while ((char *) de < end && de->namelen) {
+                        /* lookup an inode */
+                        LASSERT(de->namelen <= 255);
+
+                        dentry = ll_lookup_one_len(de->name, dc->dentry,
+                                                   de->namelen);
+                        if (IS_ERR(dentry)) {
+                                CERROR("can't lookup %*s: %d\n", de->namelen,
+                                                de->name, (int) PTR_ERR(dentry));
+                                goto next;
+                        }
+                        rc = fsfilt_del_dir_entry(dc->obd, dentry);
+                        l_dput(dentry);
+next:
+                        de = (struct dir_entry *)
+                                ((char *) de + DIR_REC_LEN(de->namelen));
+                }
         }
         RETURN(0);
 }
@@ -381,27 +448,99 @@ int scan_and_distribute(struct obd_device *obd, struct dentry *dentry,
         }
 
         err = vfs_readdir(file, filldir, &dc);
-        
         filp_close(file, 0);
+        if (err)
+                GOTO(cleanup, err);
 
         for (i = 0; i < mea->mea_count; i++) {
-                if (dc.cache[i].cached)
-                        flush_buffer_onto_mds(&dc, i);
+                if (!dc.cache[i].cached)
+                        continue;
+                err = flush_buffer_onto_mds(&dc, i);
+                if (err)
+                        GOTO(cleanup, err);
         }
 
+        for (i = 0; i < mea->mea_count; i++) {
+                if (!dc.cache[i].cached)
+                        continue;
+                err = remove_entries_from_orig_dir(&dc, i);
+                if (err)
+                        GOTO(cleanup, err);
+        }
+
+cleanup:
+        for (i = 0; i < mea->mea_count; i++) {
+                struct list_head *cur, *tmp;
+                if (!dc.cache[i].cached)
+                        continue;
+                list_for_each_safe(cur, tmp, &dc.cache[i].list) {
+                        struct page *page;
+                        page = list_entry(cur, struct page, list);
+                        list_del(&page->list);
+                        __free_page(page);
+                }
+        }
         OBD_FREE(dc.cache, sizeof(struct dir_cache) * mea->mea_count);
         OBD_FREE(file_name, nlen);
 
-        return 0;
+        RETURN(err);
 }
 
 #define MAX_DIR_SIZE    (64 * 1024)
 
+#define I_NON_SPLITTABLE        256
+
+int mds_splitting_expected(struct obd_device *obd, struct dentry *dentry)
+{
+        struct mds_obd *mds = &obd->u.mds;
+        struct mea *mea = NULL;
+        int rc, size;
+
+	/* clustered MD ? */
+	if (!mds->mds_lmv_obd)
+		RETURN(MDS_NO_SPLITTABLE);
+
+        /* inode exist? */
+        if (dentry->d_inode == NULL)
+                return MDS_NO_SPLITTABLE;
+
+        /* a dir can be splitted only */
+        if (!S_ISDIR(dentry->d_inode->i_mode))
+                return MDS_NO_SPLITTABLE;
+
+        /* already splittied or slave directory (part of splitted dir) */
+        if (dentry->d_inode->i_flags & I_NON_SPLITTABLE)
+                return MDS_NO_SPLITTABLE;
+
+        /* don't split root directory */
+        if (dentry->d_inode->i_ino == mds->mds_rootfid.id)
+                return MDS_NO_SPLITTABLE;
+
+        /* large enough to be splitted? */
+        if (dentry->d_inode->i_size < MAX_DIR_SIZE)
+                return MDS_NO_SPLIT_EXPECTED;
+
+        mds_get_lmv_attr(obd, dentry->d_inode, &mea, &size);
+        if (mea) {
+                /* already splitted or slave object: shouldn't be splitted */
+                rc = MDS_NO_SPLITTABLE;
+                /* mark to skip subsequent checks */
+                dentry->d_inode->i_flags |= I_NON_SPLITTABLE;
+        } else {
+                /* may be splitted */
+                rc = MDS_EXPECT_SPLIT;
+        }
+
+        if (mea)
+                OBD_FREE(mea, size);
+        RETURN(rc);
+}
+
 /*
- * must not be called on already splitted directories
+ * must not be called on already splitted directories.
  */
-int mds_try_to_split_dir(struct obd_device *obd,
-                         struct dentry *dentry, struct mea **mea, int nstripes)
+int mds_try_to_split_dir(struct obd_device *obd, struct dentry *dentry,
+                         struct mea **mea, int nstripes, int update_mode)
 {
         struct inode *dir = dentry->d_inode;
         struct mds_obd *mds = &obd->u.mds;
@@ -411,25 +550,22 @@ int mds_try_to_split_dir(struct obd_device *obd,
 	void *handle;
 	ENTRY;
 
-	/* clustered MD ? */
-	if (!mds->mds_lmv_obd)
-		RETURN(0);
-
-        /* don't split root directory */
-        if (dentry->d_inode->i_ino == mds->mds_rootfid.id)
-                RETURN(0);
-
-        /* we want to split only large dirs. this may be already
-         * splitted dir or a slave dir created during splitting */
-        if (dir->i_size < MAX_DIR_SIZE)
-                RETURN(0);
-
-        /* check is directory marked non-splittable */
-        if (mea && *mea)
-                RETURN(0);
+        if (update_mode != LCK_EX)
+                return 0;
+        /* TODO: optimization possible - we already may have mea here */
+        rc = mds_splitting_expected(obd, dentry);
+        if (rc == MDS_NO_SPLITTABLE)
+                return 0;
+        if (rc == MDS_NO_SPLIT_EXPECTED && nstripes == 0)
+                return 0;
+        if (nstripes && nstripes == 1)
+                return 0;
+        
+        LASSERT(mea == NULL || *mea == NULL);
 
-        CDEBUG(D_OTHER, "%s: split directory %lu/%lu\n",
-               obd->obd_name, dir->i_ino, (unsigned long) dir->i_generation);
+        CDEBUG(D_OTHER, "%s: split directory %u/%lu/%lu\n",
+               obd->obd_name, mds->mds_num, dir->i_ino,
+               (unsigned long) dir->i_generation);
 
         if (mea == NULL)
                 mea = &tmea;
@@ -439,54 +575,84 @@ int mds_try_to_split_dir(struct obd_device *obd,
          * necessary amount of stripes, but on the other hand with this
          * approach of allocating maximal possible amount of MDS slots,
          * it would be easier to split the dir over more MDSes */
-        rc = obd_alloc_diskmd(mds->mds_lmv_exp, (void *) mea);
-        if (!(*mea))
-                RETURN(-ENOMEM);
+        rc = obd_alloc_diskmd(mds->mds_lmv_exp, (void *)mea);
+        if (rc < 0) {
+                CERROR("obd_alloc_diskmd() failed, error %d.\n", rc);
+                RETURN(rc);
+        }
+        if (*mea == NULL)
+                RETURN(-EINVAL);
+
         (*mea)->mea_count = nstripes;
        
-#warning "we have to take EX lock on a dir for splitting"
-        
 	/* 1) create directory objects on slave MDS'es */
 	/* FIXME: should this be OBD method? */
         oa = obdo_alloc();
-        /* FIXME: error handling here */
-        LASSERT(oa != NULL);
+        if (!oa)
+                RETURN(-ENOMEM);
+
 	oa->o_id = dir->i_ino;
 	oa->o_generation = dir->i_generation;
-	obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
+
+        obdo_from_inode(oa, dir, OBD_MD_FLTYPE | OBD_MD_FLATIME |
 			OBD_MD_FLMTIME | OBD_MD_FLCTIME |
                         OBD_MD_FLUID | OBD_MD_FLGID);
+
         oa->o_gr = FILTER_GROUP_FIRST_MDS + mds->mds_num;
         oa->o_valid |= OBD_MD_FLID | OBD_MD_FLFLAGS | OBD_MD_FLGROUP;
         oa->o_mode = dir->i_mode;
+
         CDEBUG(D_OTHER, "%s: create subdirs with mode %o, uid %u, gid %u\n",
                         obd->obd_name, dir->i_mode, dir->i_uid, dir->i_gid);
                         
         rc = obd_create(mds->mds_lmv_exp, oa,
-                        (struct lov_stripe_md **) mea, NULL);
-        /* FIXME: error handling here */
-	LASSERT(rc == 0);
-        CDEBUG(D_OTHER, "%d dirobjects created\n",
-               (int) (*mea)->mea_count);
+                        (struct lov_stripe_md **)mea, NULL);
+        if (rc)
+                GOTO(err_oa, rc);
+
+        CDEBUG(D_OTHER, "%d dirobjects created\n", (int)(*mea)->mea_count);
 
 	/* 2) update dir attribute */
         down(&dir->i_sem);
+        
         handle = fsfilt_start(obd, dir, FSFILT_OP_SETATTR, NULL);
-        LASSERT(!IS_ERR(handle));
+        if (IS_ERR(handle)) {
+                up(&dir->i_sem);
+                CERROR("fsfilt_start() failed: %d\n", (int) PTR_ERR(handle));
+                GOTO(err_oa, rc = PTR_ERR(handle));
+        }
+        
 	rc = fsfilt_set_md(obd, dir, handle, *mea, mea_size);
-        LASSERT(rc == 0);
-        fsfilt_commit(obd, dir, handle, 0);
-        LASSERT(rc == 0);
+        if (rc) {
+                up(&dir->i_sem);
+                CERROR("fsfilt_set_md() failed, error %d.\n", rc);
+                GOTO(err_oa, rc);
+        }
+        
+        rc = fsfilt_commit(obd, mds->mds_sb, dir, handle, 0);
+        if (rc) {
+                up(&dir->i_sem);
+                CERROR("fsfilt_commit() failed, error %d.\n", rc);
+                GOTO(err_oa, rc);
+        }
+        
 	up(&dir->i_sem);
 	obdo_free(oa);
 
 	/* 3) read through the dir and distribute it over objects */
-        scan_and_distribute(obd, dentry, *mea);
-
+        rc = scan_and_distribute(obd, dentry, *mea);
 	if (mea == &tmea)
-                obd_free_diskmd(mds->mds_lmv_exp,
-                                (struct lov_mds_md **) mea);
+                obd_free_diskmd(mds->mds_lmv_exp, (struct lov_mds_md **)mea);
+        if (rc) {
+                CERROR("scan_and_distribute() failed, error %d.\n", rc);
+                RETURN(rc);
+        }
+
 	RETURN(1);
+
+err_oa:
+	obdo_free(oa);
+        RETURN(rc);
 }
 
 static int filter_start_page_write(struct inode *inode,
@@ -507,7 +673,6 @@ static int filter_start_page_write(struct inode *inode,
 struct dentry *filter_fid2dentry(struct obd_device *obd,
                                  struct dentry *dir_dentry,
                                  obd_gr group, obd_id id);
-void f_dput(struct dentry *dentry);
 
 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
                 int objcount, struct obd_ioobj *obj,
@@ -531,12 +696,16 @@ int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
         fid.id = obj->ioo_id;
         fid.generation = obj->ioo_gr;
         dentry = mds_fid2dentry(mds, &fid, NULL);
-        LASSERT(!IS_ERR(dentry));
+        if (IS_ERR(dentry)) {
+                CERROR("can't get dentry for "LPU64"/%u: %d\n",
+                       fid.id, fid.generation, (int) PTR_ERR(dentry));
+                GOTO(cleanup, rc = (int) PTR_ERR(dentry));
+        }
 
         if (dentry->d_inode == NULL) {
                 CERROR("trying to BRW to non-existent file "LPU64"\n",
                        obj->ioo_id);
-                f_dput(dentry);
+                l_dput(dentry);
                 GOTO(cleanup, rc = -ENOENT);
         }
 
@@ -560,7 +729,7 @@ int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
                                i, obj->ioo_bufcnt, dentry, rc);
                         while (lnb-- > res)
                                 __free_pages(lnb->page, 0);
-                        f_dput(dentry);
+                        l_dput(dentry);
                         GOTO(cleanup, rc);
                 }
                 tot_bytes += lnb->len;
@@ -606,30 +775,349 @@ int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
                         err = fsfilt_add_dir_entry(obd, res->dentry, de->name,
                                                    de->namelen, de->ino,
                                                    de->generation, de->mds);
-                        /* FIXME: remove entries from the original dir */
-#warning "removing entries from the original dir"
+                        if (err) {
+                                CERROR("can't add dir entry %*s->%u/%u/%u"
+                                       " to %lu/%u: %d\n",
+                                       de->namelen, de->name,
+                                       de->mds, (unsigned) de->ino,
+                                       (unsigned) de->generation,
+                                       res->dentry->d_inode->i_ino,
+                                       res->dentry->d_inode->i_generation,
+                                       err);
+                                rc = err;
+                                break;
+                        }
                         LASSERT(err == 0);
                         de = (struct dir_entry *)
                                 ((char *) de + DIR_REC_LEN(de->namelen));
                         entries++;
                 }
-                kunmap(buf);
+                kunmap(lnb->page);
         }
 
         for (i = 0, lnb = res; i < obj->ioo_bufcnt; i++, lnb++)
                 __free_page(lnb->page);
-        f_dput(res->dentry);
+        l_dput(res->dentry);
 
         RETURN(rc);
 }
 
-int mds_choose_mdsnum(struct obd_device *obd, const char *name, int len)
+int mds_choose_mdsnum(struct obd_device *obd, const char *name, int len, int flags)
 {
+        struct lmv_obd *lmv;
         struct mds_obd *mds = &obd->u.mds;
-        struct lmv_obd *lmv = &mds->mds_lmv_exp->exp_obd->u.lmv;
-        int i;
+        int i = mds->mds_num;
 
-        i = raw_name2idx(lmv->count, name, len);
+        if (flags & REC_REINT_CREATE) { 
+                i = mds->mds_num;
+        } else if (mds->mds_lmv_exp) {
+                lmv = &mds->mds_lmv_exp->exp_obd->u.lmv;
+                i = raw_name2idx(MEA_MAGIC_LAST_CHAR, lmv->desc.ld_tgt_count, name, len);
+        }
         RETURN(i);
 }
 
+int mds_lock_slave_objs(struct obd_device *obd, struct dentry *dentry,
+                        struct lustre_handle **rlockh)
+{
+        struct mds_obd *mds = &obd->u.mds;
+        struct mdc_op_data op_data;
+        struct lookup_intent it;
+        struct mea *mea = NULL;
+        int mea_size, rc;
+
+        LASSERT(rlockh != NULL);
+        LASSERT(dentry != NULL);
+        LASSERT(dentry->d_inode != NULL);
+
+	/* clustered MD ? */
+	if (!mds->mds_lmv_obd)
+	        return 0;
+
+        /* a dir can be splitted only */
+        if (!S_ISDIR(dentry->d_inode->i_mode))
+                return 0;
+
+        rc = mds_get_lmv_attr(obd, dentry->d_inode, &mea, &mea_size);
+        if (rc)
+                return rc;
+
+        if (mea == NULL)
+                return 0;
+        if (mea->mea_count == 0) {
+                /* this is slave object */
+                GOTO(cleanup, rc = 0);
+        }
+                
+        CDEBUG(D_OTHER, "%s: lock slaves for %lu/%lu\n", obd->obd_name,
+               (unsigned long) dentry->d_inode->i_ino,
+               (unsigned long) dentry->d_inode->i_generation);
+
+        OBD_ALLOC(*rlockh, sizeof(struct lustre_handle) * mea->mea_count);
+        if (*rlockh == NULL)
+                GOTO(cleanup, rc = -ENOMEM);
+        memset(*rlockh, 0, sizeof(struct lustre_handle) * mea->mea_count);
+
+        memset(&op_data, 0, sizeof(op_data));
+        op_data.mea1 = mea;
+        it.it_op = IT_UNLINK;
+        rc = md_enqueue(mds->mds_lmv_exp, LDLM_IBITS, &it, LCK_EX, &op_data,
+                        *rlockh, NULL, 0, ldlm_completion_ast, mds_blocking_ast,
+                        NULL);
+cleanup:
+        OBD_FREE(mea, mea_size);
+        RETURN(rc);
+}
+
+void mds_unlock_slave_objs(struct obd_device *obd, struct dentry *dentry,
+                        struct lustre_handle *lockh)
+{
+        struct mds_obd *mds = &obd->u.mds;
+        struct mea *mea = NULL;
+        int mea_size, rc, i;
+
+        if (lockh == NULL)
+                return;
+
+	LASSERT(mds->mds_lmv_obd != NULL);
+        LASSERT(S_ISDIR(dentry->d_inode->i_mode));
+
+        rc = mds_get_lmv_attr(obd, dentry->d_inode, &mea, &mea_size);
+        if (rc) {
+                CERROR("locks are leaked\n");
+                return;
+        }
+        LASSERT(mea_size != 0);
+        LASSERT(mea != NULL);
+        LASSERT(mea->mea_count != 0);
+
+        CDEBUG(D_OTHER, "%s: unlock slaves for %lu/%lu\n", obd->obd_name,
+               (unsigned long) dentry->d_inode->i_ino,
+               (unsigned long) dentry->d_inode->i_generation);
+
+        for (i = 0; i < mea->mea_count; i++) {
+                if (lockh[i].cookie != 0)
+                        ldlm_lock_decref(lockh + i, LCK_EX);
+        }
+
+        OBD_FREE(lockh, sizeof(struct lustre_handle) * mea->mea_count);
+        OBD_FREE(mea, mea_size);
+        return;
+}
+
+int mds_unlink_slave_objs(struct obd_device *obd, struct dentry *dentry)
+{
+        struct mds_obd *mds = &obd->u.mds;
+        struct ptlrpc_request *req = NULL;
+        struct mdc_op_data op_data;
+        struct mea *mea = NULL;
+        int mea_size, rc;
+
+	/* clustered MD ? */
+	if (!mds->mds_lmv_obd)
+	        return 0;
+
+        /* a dir can be splitted only */
+        if (!S_ISDIR(dentry->d_inode->i_mode))
+                RETURN(0);
+
+        rc = mds_get_lmv_attr(obd, dentry->d_inode, &mea, &mea_size);
+        if (rc)
+                RETURN(rc);
+
+        if (mea == NULL)
+                return 0;
+        if (mea->mea_count == 0)
+                GOTO(cleanup, rc = 0);
+
+        CDEBUG(D_OTHER, "%s: unlink slaves for %lu/%lu\n", obd->obd_name,
+               (unsigned long) dentry->d_inode->i_ino,
+               (unsigned long) dentry->d_inode->i_generation);
+
+        memset(&op_data, 0, sizeof(op_data));
+        op_data.mea1 = mea;
+        rc = md_unlink(mds->mds_lmv_exp, &op_data, &req);
+        LASSERT(req == NULL);
+cleanup:
+        OBD_FREE(mea, mea_size);
+        RETURN(rc);
+}
+
+struct ide_tracking {
+        int entries;
+        int empty;
+};
+
+int mds_ide_filldir(void *__buf, const char *name, int namelen,
+                    loff_t offset, ino_t ino, unsigned int d_type)
+{
+        struct ide_tracking *it = __buf;
+
+        if (ino == 0)
+                return 0;
+
+        it->entries++;
+        if (it->entries > 2)
+                goto noempty;
+        if (namelen > 2)
+                goto noempty;
+        if (name[0] == '.' && namelen == 1)
+                return 0;
+        if (name[0] == '.' && name[1] == '.' && namelen == 2)
+                return 0;
+noempty:
+        it->empty = 0;
+        return -ENOTEMPTY;
+}
+
+int mds_is_dir_empty(struct obd_device *obd, struct dentry *dentry)
+{
+        struct ide_tracking it;
+        struct file * file;
+        char *file_name;
+        int nlen, i, rc;
+        
+        it.entries = 0;
+        it.empty = 1;
+
+        nlen = strlen("__iopen__/") + 10 + 1;
+        OBD_ALLOC(file_name, nlen);
+        if (!file_name)
+                RETURN(-ENOMEM);
+        i = sprintf(file_name, "__iopen__/0x%lx", dentry->d_inode->i_ino);
+
+        file = filp_open(file_name, O_RDONLY, 0);
+        if (IS_ERR(file)) {
+                CERROR("can't open directory %s: %d\n",
+                       file_name, (int) PTR_ERR(file));
+                GOTO(cleanup, rc = PTR_ERR(file));
+        }
+
+        rc = vfs_readdir(file, mds_ide_filldir, &it);
+        filp_close(file, 0);
+
+        if (it.empty && rc == 0)
+                rc = 1;
+        else
+                rc = 0;
+
+cleanup:
+        OBD_FREE(file_name, nlen);
+        return rc;
+}
+
+int mds_lock_and_check_slave(int offset, struct ptlrpc_request *req,
+                             struct lustre_handle *lockh)
+{
+        struct obd_device *obd = req->rq_export->exp_obd;
+        struct dentry *dentry = NULL;
+        struct lvfs_run_ctxt saved;
+        int cleanup_phase = 0;
+        struct mds_body *body;
+        struct lvfs_ucred uc;
+        int rc, update_mode;
+        ENTRY;
+
+        body = lustre_swab_reqbuf(req, offset, sizeof(*body),
+                                  lustre_swab_mds_body);
+        if (body == NULL) {
+                CERROR("Can't swab mds_body\n");
+                GOTO(cleanup, rc = -EFAULT);
+        }
+        CDEBUG(D_OTHER, "%s: check slave %lu/%lu\n", obd->obd_name,
+               (unsigned long) body->fid1.id,
+               (unsigned long) body->fid1.generation);
+        dentry = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_EX, lockh,
+                                       &update_mode, NULL, 0,
+                                       MDS_INODELOCK_UPDATE);
+        if (IS_ERR(dentry)) {
+                CERROR("can't find inode: %d\n", (int) PTR_ERR(dentry));
+                GOTO(cleanup, rc = PTR_ERR(dentry));
+        }
+        cleanup_phase = 1;
+
+        LASSERT(S_ISDIR(dentry->d_inode->i_mode));
+
+        uc.luc_fsuid = body->fsuid;
+        uc.luc_fsgid = body->fsgid;
+        uc.luc_cap = body->capability;
+        uc.luc_suppgid1 = body->suppgid;
+        uc.luc_suppgid2 = -1;
+        push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
+
+        rc = 0;
+        if (!mds_is_dir_empty(obd, dentry))
+                rc = -ENOTEMPTY;
+
+cleanup:
+        switch(cleanup_phase) {
+        case 1:
+                if (rc)
+                        ldlm_lock_decref(lockh, LCK_EX);
+                l_dput(dentry);
+                pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
+        default:
+                break;
+        }
+        RETURN(rc);
+}
+
+int mds_convert_mea_ea(struct obd_device *obd, struct inode *inode,
+                       struct lov_mds_md *lmm, int lmmsize)
+{
+        int i, rc, err, size;
+        struct mea_old *old;
+        struct mea *mea;
+        struct mea *new;
+        void *handle;
+        ENTRY;
+
+        mea = (struct mea *) lmm;
+        if (mea->mea_magic == MEA_MAGIC_LAST_CHAR ||
+                mea->mea_magic == MEA_MAGIC_ALL_CHARS)
+                RETURN(0);
+
+        old = (struct mea_old *) lmm;
+        rc = sizeof(struct ll_fid) * old->mea_count + sizeof(struct mea_old);
+        if (old->mea_count > 256 || old->mea_master > 256 || lmmsize < rc
+                        || old->mea_master > old->mea_count) {
+                CWARN("unknown MEA format, dont convert it\n");
+                CWARN("  count %u, master %u, size %u\n",
+                      old->mea_count, old->mea_master, rc);
+                RETURN(0);
+        }
+                
+        CWARN("converting MEA EA on %lu/%u from V0 to V1 (%u/%u)\n",
+              inode->i_ino, inode->i_generation, old->mea_count, old->mea_master);
+
+        size = sizeof(struct ll_fid) * old->mea_count + sizeof(struct mea);
+        OBD_ALLOC(new, size);
+        if (new == NULL)
+                RETURN(-ENOMEM);
+
+        new->mea_magic = MEA_MAGIC_LAST_CHAR;
+        new->mea_count = old->mea_count;
+        new->mea_master = old->mea_master;
+        for (i = 0; i < new->mea_count; i++)
+                new->mea_fids[i] = old->mea_fids[i];
+
+        handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
+        if (IS_ERR(handle)) {
+                rc = PTR_ERR(handle);
+                GOTO(conv_free, rc);
+        }
+
+        rc = fsfilt_set_md(obd, inode, handle, (struct lov_mds_md *) new, size);
+        if (rc > lmmsize)
+                size = lmmsize;
+        memcpy(lmm, new, size);
+
+        err = fsfilt_commit(obd, obd->u.mds.mds_sb, inode, handle, 0);
+        if (!rc)
+                rc = err ? err : size;
+        GOTO(conv_free, rc);
+conv_free:
+        OBD_FREE(new, size);
+        return rc;
+}
+