Whamcloud - gitweb
Lproc-snmp code drop
[fs/lustre-release.git] / lustre / mds / handler.c
index c64b045..6bd5e0c 100644 (file)
 #include <linux/lustre_dlm.h>
 #include <linux/init.h>
 #include <linux/obd_class.h>
+#include <linux/random.h>
+#include <linux/locks.h>
+#if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
+#include <linux/buffer_head.h>
+#endif
+#include <linux/lprocfs_status.h>
+
+static kmem_cache_t *mds_file_cache;
 
-extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
+extern int mds_get_lovtgts(struct mds_obd *obd, int tgt_count,
                            obd_uuid_t *uuidarray);
-extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
+extern int mds_get_lovdesc(struct mds_obd  *obd, struct lov_desc *desc);
 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
                                 struct ptlrpc_request *req);
 static int mds_cleanup(struct obd_device * obddev);
 
+extern lprocfs_vars_t status_var_nm_1[];
+extern lprocfs_vars_t status_class_var[];
+
 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
 {
         return &req->rq_export->exp_obd->u.mds;
@@ -133,7 +144,7 @@ struct dentry *mds_name2locked_dentry(struct obd_device *obd,
                                       int dir_lock_mode)
 {
         struct dentry *dchild;
-        int flags, rc;
+        int flags = 0, rc;
         __u64 res_id[3] = {0};
         ENTRY;
 
@@ -154,6 +165,7 @@ struct dentry *mds_name2locked_dentry(struct obd_device *obd,
                 RETURN(dchild);
 
         res_id[0] = dchild->d_inode->i_ino;
+        res_id[1] = dchild->d_inode->i_generation;
         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
                                    &flags, ldlm_completion_ast,
@@ -173,7 +185,7 @@ struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
 {
         struct mds_obd *mds = &obd->u.mds;
         struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
-        int flags, rc;
+        int flags = 0, rc;
         __u64 res_id[3] = {0};
         ENTRY;
 
@@ -181,6 +193,7 @@ struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
                 RETURN(de);
 
         res_id[0] = de->d_inode->i_ino;
+        res_id[1] = de->d_inode->i_generation;
         rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
                                    res_id, LDLM_PLAIN, NULL, 0, lock_mode,
                                    &flags, ldlm_completion_ast,
@@ -193,6 +206,10 @@ struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
         RETURN(retval);
 }
 
+#ifndef DCACHE_DISCONNECTED
+#define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
+#endif
+
 /* Look up an entry by inode number. */
 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
                               struct vfsmount **mnt)
@@ -232,7 +249,7 @@ struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
         spin_lock(&dcache_lock);
         list_for_each(lp, &inode->i_dentry) {
                 result = list_entry(lp, struct dentry, d_alias);
-                if (!(result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
+                if (!(result->d_flags & DCACHE_DISCONNECTED)) {
                         dget_locked(result);
                         result->d_vfs_flags |= DCACHE_REFERENCED;
                         spin_unlock(&dcache_lock);
@@ -250,7 +267,7 @@ struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
         }
         if (mnt)
                 mntget(*mnt);
-        result->d_flags |= DCACHE_NFSD_DISCONNECTED;
+        result->d_flags |= DCACHE_DISCONNECTED;
         return result;
 }
 
@@ -261,9 +278,11 @@ struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
  * on the server, etc.
  */
 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
-                       obd_uuid_t cluuid)
+                       obd_uuid_t cluuid, struct recovd_obd *recovd,
+                       ptlrpc_recovery_cb_t recover)
 {
         struct obd_export *exp;
+        struct mds_export_data *med;
         struct mds_client_data *mcd;
         struct list_head *p;
         int rc;
@@ -284,6 +303,7 @@ static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
                         if (!list_empty(&exp->exp_conn_chain)) {
                                 CERROR("existing uuid/export, list not empty!\n");
                                 spin_unlock(&obd->obd_dev_lock);
+                                /* XXX should we MOD_DEC_USE_COUNT; here? */
                                 RETURN(-EALREADY);
                         }
                         conn->addr = (__u64) (unsigned long)exp;
@@ -293,6 +313,7 @@ static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
                                cluuid, exp);
                         CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
                                (long long)conn->addr, (long long)conn->cookie);
+                        MOD_DEC_USE_COUNT;
                         RETURN(0);
                 }
         }
@@ -306,15 +327,21 @@ static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
                 GOTO(out_dec, rc);
         exp = class_conn2export(conn);
         LASSERT(exp);
+        med = &exp->exp_mds_data;
 
         OBD_ALLOC(mcd, sizeof(*mcd));
         if (!mcd) {
                 CERROR("mds: out of memory for client data\n");
                 GOTO(out_export, rc = -ENOMEM);
         }
+
         memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
-        exp->exp_mds_data.med_mcd = mcd;
-        rc = mds_client_add(&exp->exp_mds_data, -1);
+        med->med_mcd = mcd;
+
+        INIT_LIST_HEAD(&med->med_open_head);
+        spin_lock_init(&med->med_open_lock);
+
+        rc = mds_client_add(med, -1);
         if (rc)
                 GOTO(out_mdc, rc);
 
@@ -330,10 +357,41 @@ out_dec:
         return rc;
 }
 
+/* Call with med->med_open_lock held, please. */
+inline int mds_close_mfd(struct mds_file_data *mfd, struct mds_export_data *med)
+{
+        struct file *file = mfd->mfd_file;
+        LASSERT(file->private_data == mfd);
+
+        list_del(&mfd->mfd_list);
+        mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
+        kmem_cache_free(mds_file_cache, mfd);
+
+        return filp_close(file, 0);
+}
+
 static int mds_disconnect(struct lustre_handle *conn)
 {
-        int rc;
         struct obd_export *export = class_conn2export(conn);
+        struct list_head *tmp, *n;
+        struct mds_export_data *med = &export->exp_mds_data;
+        int rc;
+        ENTRY;
+
+        /*
+         * Close any open files.
+         */
+        spin_lock(&med->med_open_lock);
+        list_for_each_safe(tmp, n, &med->med_open_head) {
+                struct mds_file_data *mfd = 
+                        list_entry(tmp, struct mds_file_data, mfd_list);
+                rc = mds_close_mfd(mfd, med);
+                if (rc) {
+                        /* XXX better diagnostics, with file path and stuff */
+                        CDEBUG(D_INODE, "Error %d closing mfd %p\n", rc, mfd);
+                }
+        }
+        spin_unlock(&med->med_open_lock);
 
         ldlm_cancel_locks_for_export(export);
         mds_client_free(export);
@@ -342,7 +400,24 @@ static int mds_disconnect(struct lustre_handle *conn)
         if (!rc)
                 MOD_DEC_USE_COUNT;
 
-        return rc;
+        RETURN(rc);
+}
+
+/*
+ * XXX This is NOT guaranteed to flush all transactions to disk (even though
+ *     it is equivalent to calling sync()) because it only _starts_ the flush
+ *     and does not wait for completion.  It's better than nothing though.
+ *     What we really want is a mild form of fsync_dev_lockfs(), but it is
+ *     non-standard, or enabling do_sync_supers in ext3, just for this call.
+ */
+static void mds_fsync_super(struct super_block *sb)
+{
+        lock_kernel();
+        lock_super(sb);
+        if (sb->s_dirt && sb->s_op && sb->s_op->write_super)
+                sb->s_op->write_super(sb);
+        unlock_super(sb);
+        unlock_kernel();
 }
 
 static int mds_getstatus(struct ptlrpc_request *req)
@@ -359,7 +434,13 @@ static int mds_getstatus(struct ptlrpc_request *req)
                 RETURN(0);
         }
 
-        /* Anything we need to do here with the client's trans no or so? */
+        /* Flush any outstanding transactions to disk so the client will
+         * get the latest last_committed value and can drop their local
+         * requests if they have any.  This would be fsync_super() if it
+         * was exported.
+         */
+        mds_fsync_super(mds->mds_sb);
+
         body = lustre_msg_buf(req->rq_repmsg, 0);
         memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
 
@@ -391,8 +472,8 @@ static int mds_getlovinfo(struct ptlrpc_request *req)
         }
 
         desc = lustre_msg_buf(req->rq_repmsg, 0);
-        rc = mds_get_lovdesc(req->rq_obd, desc);
-        if (rc != 0 ) {
+        rc = mds_get_lovdesc(mds, desc);
+        if (rc) {
                 CERROR("mds_get_lovdesc error %d", rc);
                 req->rq_status = rc;
                 RETURN(0);
@@ -407,7 +488,7 @@ static int mds_getlovinfo(struct ptlrpc_request *req)
 
         mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
                 tgt_count * sizeof(struct lov_object_id);
-        rc = mds_get_lovtgts(req->rq_obd, tgt_count,
+        rc = mds_get_lovtgts(mds, tgt_count,
                              lustre_msg_buf(req->rq_repmsg, 1));
         if (rc) {
                 CERROR("get_lovtgts error %d\n", rc);
@@ -428,6 +509,7 @@ int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
                 RETURN(0);
         }
 
+        /* XXX layering violation!  -phil */
         l_lock(&lock->l_resource->lr_namespace->ns_lock);
         lock->l_flags |= LDLM_FL_CBPENDING;
         do_ast = (!lock->l_readers && !lock->l_writers);
@@ -505,7 +587,7 @@ static int mds_getattr_name(int offset, struct ptlrpc_request *req)
         struct inode *dir;
         struct lustre_handle lockh;
         char *name;
-        int namelen, flags, lock_mode, rc = 0;
+        int namelen, flags = 0, lock_mode, rc = 0;
         struct obd_ucred uc;
         __u64 res_id[3] = {0, 0, 0};
         ENTRY;
@@ -538,6 +620,7 @@ static int mds_getattr_name(int offset, struct ptlrpc_request *req)
 
         lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
         res_id[0] = dir->i_ino;
+        res_id[1] = dir->i_generation;
 
         rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
                              NULL, 0, lock_mode, &lockh);
@@ -595,8 +678,7 @@ static int mds_getattr(int offset, struct ptlrpc_request *req)
         push_ctxt(&saved, &mds->mds_ctxt, &uc);
         de = mds_fid2dentry(mds, &body->fid1, NULL);
         if (IS_ERR(de)) {
-                req->rq_status = -ENOENT;
-                rc = 0;
+                rc = req->rq_status = -ENOENT;
                 GOTO(out_pop, PTR_ERR(de));
         }
 
@@ -606,7 +688,9 @@ static int mds_getattr(int offset, struct ptlrpc_request *req)
                 size[1] = mds->mds_max_mdsize;
         } else if (body->valid & OBD_MD_LINKNAME) {
                 bufcount = 2;
-                size[1] = inode->i_size + 1;
+                size[1] = MIN(inode->i_size + 1, body->size);
+                CDEBUG(D_INODE, "symlink size: %d, reply space: %d\n",
+                       inode->i_size + 1, body->size);
         }
 
         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
@@ -662,43 +746,99 @@ out:
         RETURN(0);
 }
 
+static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
+{
+        struct mds_file_data *mfd = NULL;
+
+        if (!handle || !handle->addr)
+                RETURN(NULL);
+
+        mfd = (struct mds_file_data *)(unsigned long)(handle->addr);
+        if (!kmem_cache_validate(mds_file_cache, mfd))
+                RETURN(NULL);
+
+        if (mfd->mfd_servercookie != handle->cookie)
+                RETURN(NULL);
+
+        return mfd;
+}
+
+static int mds_store_ea(struct mds_obd *mds, struct ptlrpc_request *req,
+                        struct mds_body *body, struct dentry *de,
+                        struct lov_mds_md *lmm)
+{
+        struct obd_run_ctxt saved;
+        struct obd_ucred uc;
+        void *handle;
+        int rc, rc2;
+
+        uc.ouc_fsuid = body->fsuid;
+        uc.ouc_fsgid = body->fsgid;
+        push_ctxt(&saved, &mds->mds_ctxt, &uc);
+        handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
+        if (!handle)
+                GOTO(out_ea, rc = -ENOMEM);
+
+        rc = mds_fs_set_md(mds, de->d_inode, handle, lmm);
+        if (!rc)
+                rc = mds_update_last_rcvd(mds, handle, req);
+
+        rc2 = mds_fs_commit(mds, de->d_inode, handle);
+        if (rc2 && !rc)
+                rc = rc2;
+out_ea:
+        pop_ctxt(&saved);
+
+        return rc;
+}
+
 static int mds_open(struct ptlrpc_request *req)
 {
-        struct dentry *de;
+        struct mds_obd *mds = mds_req2mds(req);
         struct mds_body *body;
+        struct mds_export_data *med;
+        struct mds_file_data *mfd;
+        struct dentry *de;
         struct file *file;
         struct vfsmount *mnt;
-        struct mds_obd *mds = mds_req2mds(req);
-        struct mds_export_data *med;
         __u32 flags;
         struct list_head *tmp;
-        struct mds_file_data *mfd;
         int rc, size = sizeof(*body);
         ENTRY;
 
-        rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
-        if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
-                CERROR("mds: out of memory\n");
+        if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
+                CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
                 req->rq_status = -ENOMEM;
-                RETURN(0);
+                RETURN(-ENOMEM);
+        }
+
+        rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
+        if (rc) {
+                CERROR("mds: pack error: rc = %d\n", rc);
+                req->rq_status = rc;
+                RETURN(rc);
         }
 
         body = lustre_msg_buf(req->rq_reqmsg, 0);
 
-        /* was this animal open already? */
-        /* XXX we should only check on re-open, or do a refcount... */
+        /* was this animal open already and the client lost the reply? */
+        /* XXX need some way to detect a reopen, to avoid locked list walks */
         med = &req->rq_export->exp_mds_data;
+        spin_lock(&med->med_open_lock);
         list_for_each(tmp, &med->med_open_head) {
-                struct mds_file_data *fd;
-                fd = list_entry(tmp, struct mds_file_data, mfd_list);
-                if (body->extra == fd->mfd_clientfd &&
-                    body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
+                mfd = list_entry(tmp, typeof(*mfd), mfd_list);
+                if (!memcmp(&mfd->mfd_clienthandle, &body->handle,
+                            sizeof(mfd->mfd_clienthandle)) &&
+                    body->fid1.id == mfd->mfd_file->f_dentry->d_inode->i_ino) {
+                        de = mfd->mfd_file->f_dentry;
+                        spin_unlock(&med->med_open_lock);
                         CERROR("Re opening "LPD64"\n", body->fid1.id);
-                        RETURN(0);
+                        GOTO(out_pack, rc = 0);
                 }
         }
+        spin_unlock(&med->med_open_lock);
 
-        OBD_ALLOC(mfd, sizeof(*mfd));
+        mfd = kmem_cache_alloc(mds_file_cache, GFP_KERNEL);
         if (!mfd) {
                 CERROR("mds: out of memory\n");
                 req->rq_status = -ENOMEM;
@@ -706,105 +846,87 @@ static int mds_open(struct ptlrpc_request *req)
         }
 
         de = mds_fid2dentry(mds, &body->fid1, &mnt);
-        if (IS_ERR(de)) {
-                req->rq_status = -ENOENT;
-                RETURN(0);
-        }
+        if (IS_ERR(de))
+                GOTO(out_free, rc = PTR_ERR(de));
 
         /* check if this inode has seen a delayed object creation */
-        if (req->rq_reqmsg->bufcount > 1) {
-                void *handle;
-                struct lov_mds_md *lmm;
-                struct inode *inode = de->d_inode;
-                int rc;
+        if (lustre_msg_get_op_flags(req->rq_reqmsg) & MDS_OPEN_HAS_EA) {
+                struct lov_mds_md *lmm = lustre_msg_buf(req->rq_reqmsg, 1);
 
-                lmm = lustre_msg_buf(req->rq_reqmsg, 1);
-
-                handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
-                if (!handle) {
-                        req->rq_status = -ENOMEM;
-                        RETURN(0);
-                }
-
-                /* XXX error handling */
-                rc = mds_fs_set_md(mds, inode, handle, lmm);
-                if (!rc) {
-                        struct obd_run_ctxt saved;
-                        struct obd_ucred uc;
-                        uc.ouc_fsuid = body->fsuid;
-                        uc.ouc_fsgid = body->fsgid;
-                        push_ctxt(&saved, &mds->mds_ctxt, &uc);
-                        rc = mds_update_last_rcvd(mds, handle, req);
-                        pop_ctxt(&saved);
-                } else {
-                        req->rq_status = rc;
-                        RETURN(0);
-                }
-                /* FIXME: need to return last_rcvd, last_committed */
-
-                /* FIXME: keep rc intact */
-                rc = mds_fs_commit(mds, de->d_inode, handle);
+                rc = mds_store_ea(mds, req, body, de, lmm);
                 if (rc) {
-                        req->rq_status = rc;
-                        RETURN(0);
+                        l_dput(de);
+                        mntput(mnt);
+                        GOTO(out_free, rc);
                 }
         }
 
         flags = body->flags;
+        /* dentry_open does a dput(de) and mntput(mnt) on error */
         file = dentry_open(de, mnt, flags & ~O_DIRECT);
-        if (!file || IS_ERR(file)) {
-                req->rq_status = -EINVAL;
-                OBD_FREE(mfd, sizeof(*mfd));
-                RETURN(0);
+        if (IS_ERR(file)) {
+                rc = PTR_ERR(file);
+                GOTO(out_free, 0);
         }
 
         file->private_data = mfd;
         mfd->mfd_file = file;
-        mfd->mfd_clientfd = body->extra;
+        memcpy(&mfd->mfd_clienthandle, &body->handle, sizeof(body->handle));
+        get_random_bytes(&mfd->mfd_servercookie, sizeof(mfd->mfd_servercookie));
+        spin_lock(&med->med_open_lock);
         list_add(&mfd->mfd_list, &med->med_open_head);
+        spin_unlock(&med->med_open_lock);
 
+out_pack:
         body = lustre_msg_buf(req->rq_repmsg, 0);
-        /* FIXME: need to have cookies involved here */
-        body->extra = (__u64) (unsigned long)file;
+        mds_pack_inode2fid(&body->fid1, de->d_inode);
+        mds_pack_inode2body(body, de->d_inode);
+        body->handle.addr = (__u64)(unsigned long)mfd;
+        body->handle.cookie = mfd->mfd_servercookie;
+        CDEBUG(D_INODE, "llite file "LPX64": addr %p, cookie "LPX64"\n",
+               mfd->mfd_clienthandle.addr, mfd, mfd->mfd_servercookie);
+        RETURN(0);
+
+out_free:
+        mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
+        kmem_cache_free(mds_file_cache, mfd);
+        req->rq_status = rc;
         RETURN(0);
 }
 
 static int mds_close(struct ptlrpc_request *req)
 {
-        struct dentry *de;
+        struct mds_export_data *med = &req->rq_export->exp_mds_data;
         struct mds_body *body;
-        struct file *file;
-        struct mds_obd *mds = mds_req2mds(req);
-        struct vfsmount *mnt;
         struct mds_file_data *mfd;
         int rc;
         ENTRY;
 
-        rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
-        if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
-                CERROR("mds: out of memory\n");
-                req->rq_status = -ENOMEM;
-                RETURN(0);
-        }
-
         body = lustre_msg_buf(req->rq_reqmsg, 0);
-        de = mds_fid2dentry(mds, &body->fid1, &mnt);
-        if (IS_ERR(de)) {
-                req->rq_status = -ENOENT;
-                RETURN(0);
+
+        mfd = mds_handle2mfd(&body->handle);
+        if (!mfd) {
+                CERROR("no handle for file close "LPD64
+                       ": addr "LPX64", cookie "LPX64"\n",
+                       body->fid1.id, body->handle.addr, body->handle.cookie);
+                RETURN(-ESTALE);
         }
 
-        /* FIXME: need to have cookies involved here */
-        file = (struct file *)(unsigned long)body->extra;
-        if (!file->f_dentry)
-                LBUG();
-        mfd = (struct mds_file_data *)file->private_data;
-        list_del(&mfd->mfd_list);
-        OBD_FREE(mfd, sizeof(*mfd));
+        spin_lock(&med->med_open_lock);
+        req->rq_status = mds_close_mfd(mfd, med);
+        spin_unlock(&med->med_open_lock);
 
-        req->rq_status = filp_close(file, 0);
-        l_dput(de);
-        mntput(mnt);
+        if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
+                CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
+                req->rq_status = -ENOMEM;
+                RETURN(-ENOMEM);
+        }
+
+        rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
+        if (rc) {
+                CERROR("mds: lustre_pack_msg: rc = %d\n", rc);
+                req->rq_status = rc;
+        }
 
         RETURN(0);
 }
@@ -890,8 +1012,7 @@ int mds_handle(struct ptlrpc_request *req)
         if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
                 GOTO(out, rc = -ENOTCONN);
 
-        if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
-                GOTO(out, rc = -EINVAL);
+        LASSERT(!strcmp(req->rq_obd->obd_type->typ_name, LUSTRE_MDT_NAME));
 
         switch (req->rq_reqmsg->opc) {
         case MDS_CONNECT:
@@ -973,7 +1094,6 @@ int mds_handle(struct ptlrpc_request *req)
                 if (rc)
                         break;
                 RETURN(0);
-
         case LDLM_CONVERT:
                 CDEBUG(D_INODE, "convert\n");
                 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
@@ -981,14 +1101,6 @@ int mds_handle(struct ptlrpc_request *req)
                 if (rc)
                         break;
                 RETURN(0);
-
-        case LDLM_CANCEL:
-                CDEBUG(D_INODE, "cancel\n");
-                OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
-                rc = ldlm_handle_cancel(req);
-                if (rc)
-                        break;
-                RETURN(0);
         case LDLM_BL_CALLBACK:
         case LDLM_CP_CALLBACK:
                 CDEBUG(D_INODE, "callback\n");
@@ -1018,7 +1130,8 @@ int mds_handle(struct ptlrpc_request *req)
         }
  out:
         if (rc) {
-                CERROR("mds: processing error %d\n", rc);
+                CERROR("mds: processing error (opcode %d): %d\n",
+                       req->rq_reqmsg->opc, rc);
                 ptlrpc_error(req->rq_svc, req);
         } else {
                 CDEBUG(D_NET, "sending reply\n");
@@ -1057,7 +1170,11 @@ int mds_update_server_data(struct mds_obd *mds)
                         RETURN(-EIO);
                 RETURN(rc);
         }
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
         rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
+#else
+        rc = file_fsync(filp,  filp->f_dentry, 1);
+#endif
         if (rc)
                 CERROR("error flushing MDS server data: rc = %d\n", rc);
 
@@ -1080,11 +1197,9 @@ static int mds_recover(struct obd_device *obddev)
         return rc;
 }
 
-#define MDS_NUM_THREADS 8
 /* mount the file system (secretly) */
 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
 {
-        int i;
         struct obd_ioctl_data* data = buf;
         struct mds_obd *mds = &obddev->u.mds;
         struct vfsmount *mnt;
@@ -1107,6 +1222,7 @@ static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
                 GOTO(err_kfree, rc);
         }
 
+        CERROR("%s: mnt is %p\n", data->ioc_inlbuf1, mnt);
         mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
         if (!mds->mds_sb)
                 GOTO(err_put, rc = -ENODEV);
@@ -1119,45 +1235,23 @@ static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
                 GOTO(err_put, rc);
         }
 
-        mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
-                                           MDS_BUFSIZE, MDS_MAXREQSIZE,
-                                           MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL, 
-                                           "self", mds_handle, "mds");
-        if (!mds->mds_service) {
-                CERROR("failed to start service\n");
-                GOTO(err_fs, rc = -EINVAL);
-        }
-
         obddev->obd_namespace =
                 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
         if (obddev->obd_namespace == NULL) {
                 mds_cleanup(obddev);
-                GOTO(err_svc, rc = -ENOMEM);
+                GOTO(err_fs, rc = -ENOMEM);
         }
 
-        for (i = 0; i < MDS_NUM_THREADS; i++) {
-                char name[32];
-                sprintf(name, "lustre_MDS_%02d", i);
-                rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
-                if (rc) {
-                        CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
-                        GOTO(err_thread, rc);
-                }
-        }
 
         rc = mds_recover(obddev);
         if (rc)
-                GOTO(err_thread, rc);
+                GOTO(err_fs, rc);
 
         ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
                            "mds_ldlm_client", &obddev->obd_ldlm_client);
 
         RETURN(0);
 
-err_thread:
-        ptlrpc_stop_all_threads(mds->mds_service);
-err_svc:
-        ptlrpc_unregister_service(mds->mds_service);
 err_fs:
         mds_fs_cleanup(obddev);
 err_put:
@@ -1169,7 +1263,7 @@ err_kfree:
         kfree(mds->mds_fstype);
 err_dec:
         MOD_DEC_USE_COUNT;
-        return rc;
+        RETURN(rc);
 }
 
 static int mds_cleanup(struct obd_device *obddev)
@@ -1179,9 +1273,6 @@ static int mds_cleanup(struct obd_device *obddev)
         struct obd_run_ctxt saved;
         ENTRY;
 
-        ptlrpc_stop_all_threads(mds->mds_service);
-        ptlrpc_unregister_service(mds->mds_service);
-
         sb = mds->mds_sb;
         if (!mds->mds_sb)
                 RETURN(0);
@@ -1216,7 +1307,7 @@ static int mds_cleanup(struct obd_device *obddev)
 }
 
 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
-                              ldlm_mode_t mode, void *data)
+                              ldlm_mode_t mode, int flags, void *data)
 {
         struct ptlrpc_request *req = req_cookie;
         int rc = 0;
@@ -1329,9 +1420,14 @@ static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
                         }
                 }
 
+                if (flags & LDLM_FL_INTENT_ONLY) {
+                        LDLM_DEBUG(lock, "INTENT_ONLY, aborting lock");
+                        RETURN(ELDLM_LOCK_ABORTED);
+                }
                 /* Give the client a lock on the child object, instead of the
                  * parent that it requested. */
                 new_resid[0] = NTOH__u32(mds_rep->ino);
+                new_resid[1] = NTOH__u32(mds_rep->generation);
                 if (new_resid[0] == 0)
                         LBUG();
                 old_res = lock->l_resource->lr_name[0];
@@ -1356,12 +1452,82 @@ static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
         RETURN(rc);
 }
 
+int mds_attach(struct obd_device *dev, 
+               obd_count len, void *data)
+{
+        int rc;
+        rc = lprocfs_reg_obd(dev, (lprocfs_vars_t*)status_var_nm_1, (void*)dev);
+        return rc; 
+}
+
+int mds_detach(struct obd_device *dev)
+{
+        int rc;
+        rc = lprocfs_dereg_obd(dev);
+        return rc;
+
+}
+
+#define MDT_NUM_THREADS 8
+static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
+{
+        int i;
+        //        struct obd_ioctl_data* data = buf;
+        struct mds_obd *mds = &obddev->u.mds;
+        int rc = 0;
+        ENTRY;
+
+        MOD_INC_USE_COUNT;
+
+        mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
+                                           MDS_BUFSIZE, MDS_MAXREQSIZE,
+                                           MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
+                                           "self", mds_handle, "mds");
+        if (!mds->mds_service) {
+                CERROR("failed to start service\n");
+                GOTO(err_dec, rc = -EINVAL);
+        }
+
+        for (i = 0; i < MDT_NUM_THREADS; i++) {
+                char name[32];
+                sprintf(name, "lustre_MDT_%02d", i);
+                rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
+                if (rc) {
+                        CERROR("cannot start MDT thread #%d: rc %d\n", i, rc);
+                        GOTO(err_thread, rc);
+                }
+        }
+
+        RETURN(0);
+
+err_thread:
+        ptlrpc_stop_all_threads(mds->mds_service);
+        ptlrpc_unregister_service(mds->mds_service);
+err_dec:
+        MOD_DEC_USE_COUNT;
+        RETURN(rc);
+}
+
+
+static int mdt_cleanup(struct obd_device *obddev)
+{
+        struct mds_obd *mds = &obddev->u.mds;
+        ENTRY;
+
+        ptlrpc_stop_all_threads(mds->mds_service);
+        ptlrpc_unregister_service(mds->mds_service);
+
+        MOD_DEC_USE_COUNT;
+        RETURN(0);
+}
 
 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
                          int len, void *karg, void *uarg);
 
 /* use obd ops to offer management infrastructure */
 static struct obd_ops mds_obd_ops = {
+        o_attach:      mds_attach,
+        o_detach:      mds_detach,
         o_connect:     mds_connect,
         o_disconnect:  mds_disconnect,
         o_setup:       mds_setup,
@@ -1369,17 +1535,41 @@ static struct obd_ops mds_obd_ops = {
         o_iocontrol:   mds_iocontrol
 };
 
+static struct obd_ops mdt_obd_ops = {
+        o_setup:       mdt_setup,
+        o_cleanup:     mdt_cleanup,
+};
+
+
 static int __init mds_init(void)
 {
-        class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
+     
+        mds_file_cache = kmem_cache_create("ll_mds_file_data",
+                                           sizeof(struct mds_file_data),
+                                           0, 0, NULL, NULL);
+        if (mds_file_cache == NULL)
+                return -ENOMEM;
+
+        class_register_type(&mds_obd_ops, (lprocfs_vars_t*)status_class_var, 
+                            LUSTRE_MDS_NAME);
+        class_register_type(&mdt_obd_ops, 0, LUSTRE_MDT_NAME);
+
         ldlm_register_intent(ldlm_intent_policy);
+        
         return 0;
+        
 }
 
 static void __exit mds_exit(void)
 {
+        
+        
         ldlm_unregister_intent();
         class_unregister_type(LUSTRE_MDS_NAME);
+        class_unregister_type(LUSTRE_MDT_NAME);
+        if (kmem_cache_destroy(mds_file_cache))
+                CERROR("couldn't free MDS file cache\n");
+        
 }
 
 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");