1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * Lustre Metadata Server (mds) request handler
8 * Copyright (C) 2001, 2002 Cluster File Systems, Inc.
10 * This code is issued under the GNU General Public License.
11 * See the file COPYING in this distribution
13 * by Peter Braam <braam@clusterfs.com> &
14 * Andreas Dilger <braam@clusterfs.com>
16 * This server is single threaded at present (but can easily be multi threaded)
21 #define DEBUG_SUBSYSTEM S_MDS
23 #include <linux/module.h>
24 #include <linux/lustre_mds.h>
25 #include <linux/lustre_dlm.h>
26 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
27 struct ptlrpc_request *req);
28 static int mds_cleanup(struct obd_device * obddev);
30 /* Assumes caller has already pushed into the kernel filesystem context */
31 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
35 struct mds_obd *mds = &req->rq_obd->u.mds;
36 struct ptlrpc_bulk_desc *desc;
37 struct ptlrpc_bulk_page *bulk;
41 desc = ptlrpc_prep_bulk(req->rq_connection);
43 GOTO(out, rc = -ENOMEM);
45 bulk = ptlrpc_prep_bulk_page(desc);
47 GOTO(cleanup_bulk, rc = -ENOMEM);
49 OBD_ALLOC(buf, PAGE_SIZE);
51 GOTO(cleanup_bulk, rc = -ENOMEM);
53 rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
56 GOTO(cleanup_buf, rc = -EIO);
58 bulk->b_xid = req->rq_xid;
60 bulk->b_buflen = PAGE_SIZE;
61 desc->b_portal = MDS_BULK_PORTAL;
63 rc = ptlrpc_send_bulk(desc);
65 GOTO(cleanup_buf, rc);
67 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
68 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
69 OBD_FAIL_MDS_SENDPAGE, rc);
70 ptlrpc_abort_bulk(desc);
71 GOTO(cleanup_buf, rc);
74 wait_event_interruptible(desc->b_waitq, ptlrpc_check_bulk_sent(desc));
75 if (desc->b_flags & PTL_RPC_FL_INTR)
76 GOTO(cleanup_buf, rc = -EINTR);
80 OBD_FREE(buf, PAGE_SIZE);
82 ptlrpc_free_bulk(desc);
87 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
88 struct vfsmount **mnt)
91 struct super_block *sb = mds->mds_sb;
92 unsigned long ino = fid->id;
93 __u32 generation = fid->generation;
96 struct dentry *result;
99 RETURN(ERR_PTR(-ESTALE));
101 inode = iget(sb, ino);
103 RETURN(ERR_PTR(-ENOMEM));
105 CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
107 if (is_bad_inode(inode) ||
108 (generation && inode->i_generation != generation)) {
109 /* we didn't find the right inode.. */
110 CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
111 inode->i_ino, inode->i_nlink,
112 atomic_read(&inode->i_count), inode->i_generation,
116 RETURN(ERR_PTR(-ESTALE));
119 /* now to find a dentry.
120 * If possible, get a well-connected one
123 *mnt = mds->mds_vfsmnt;
124 spin_lock(&dcache_lock);
125 for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
126 result = list_entry(lp,struct dentry, d_alias);
127 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
129 result->d_vfs_flags |= DCACHE_REFERENCED;
130 spin_unlock(&dcache_lock);
137 spin_unlock(&dcache_lock);
138 result = d_alloc_root(inode);
139 if (result == NULL) {
141 return ERR_PTR(-ENOMEM);
145 result->d_flags |= DCACHE_NFSD_DISCONNECTED;
149 static int mds_connect(struct ptlrpc_request *req, struct mds_obd **mdsp)
156 uuid = lustre_msg_buf(req->rq_reqmsg, 0);
157 if (req->rq_reqmsg->buflens[0] > 37) {
159 req->rq_status = -EINVAL;
163 i = obd_class_name2dev(uuid);
165 req->rq_status = -ENODEV;
169 *mdsp = mds = &(obd_dev[i].u.mds);
170 if (mds != &(req->rq_obd->u.mds)) {
171 CERROR("device %d (%s) is not an mds\n", i, uuid);
172 req->rq_status = -ENODEV;
176 CDEBUG(D_INFO, "MDS connect from UUID '%s'\n",
177 ptlrpc_req_to_uuid(req));
178 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
179 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CONNECT_PACK)) {
180 req->rq_status = -ENOMEM;
183 req->rq_repmsg->target_id = i;
188 /* FIXME: the error cases need fixing to avoid leaks */
189 static int mds_getstatus(struct mds_obd *mds, struct ptlrpc_request *req)
191 struct mds_body *body;
192 struct mds_client_info *mci;
193 struct mds_client_data *mcd;
194 int rc, size = sizeof(*body);
197 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
198 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
199 CERROR("mds: out of memory for message: size=%d\n", size);
200 req->rq_status = -ENOMEM;
204 body = lustre_msg_buf(req->rq_reqmsg, 0);
205 mds_unpack_body(body);
206 /* Anything we need to do here with the client's trans no or so? */
208 body = lustre_msg_buf(req->rq_repmsg, 0);
209 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
211 mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
213 /* We don't have any old connection data for this client */
216 CDEBUG(D_INFO, "allocating new client data for UUID '%s'",
217 ptlrpc_req_to_uuid(req));
219 OBD_ALLOC(mcd, sizeof(*mcd));
221 CERROR("mds: out of memory for client data\n");
222 req->rq_status = -ENOMEM;
225 memcpy(mcd->mcd_uuid, ptlrpc_req_to_uuid(req),
226 sizeof(mcd->mcd_uuid));
227 rc = mds_client_add(mds, mcd, -1);
233 /* We have old connection data for this client... */
235 CDEBUG(D_INFO, "found existing data for UUID '%s' at #%d\n",
236 mcd->mcd_uuid, mci->mci_off);
238 /* mcd_last_xid is is stored in little endian on the disk and
239 mds_pack_rep_body converts it to network order */
240 body->last_xid = le32_to_cpu(mcd->mcd_last_xid);
241 mds_pack_rep_body(req);
245 static int mds_disconnect(struct mds_obd *mds, struct ptlrpc_request *req)
247 struct mds_body *body;
251 body = lustre_msg_buf(req->rq_reqmsg, 0);
253 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
260 int mds_lock_callback(struct lustre_handle *lockh, struct ldlm_lock_desc *desc,
261 void *data, int data_len, struct ptlrpc_request **reqp)
266 /* Completion AST. Do nothing */
270 if (ldlm_cli_cancel(lockh) < 0)
275 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
277 struct mds_obd *mds = &req->rq_obd->u.mds;
278 struct obd_run_ctxt saved;
279 struct mds_body *body;
280 struct dentry *de = NULL, *dchild = NULL;
282 struct lustre_handle lockh;
284 int namelen, flags, lock_mode, rc = 0;
285 __u64 res_id[3] = {0, 0, 0};
288 if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
291 if (req->rq_reqmsg->bufcount <= offset + 1) {
293 GOTO(out_pre_de, rc = -EINVAL);
296 body = lustre_msg_buf(req->rq_reqmsg, offset);
297 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
298 namelen = req->rq_reqmsg->buflens[offset + 1];
299 /* requests were at offset 2, replies go back at 1 */
303 push_ctxt(&saved, &mds->mds_ctxt);
304 de = mds_fid2dentry(mds, &body->fid1, NULL);
307 GOTO(out_pre_de, rc = -ESTALE);
311 CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
313 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
314 res_id[0] = dir->i_ino;
316 rc = ldlm_lock_match(mds->mds_local_namespace, res_id, LDLM_PLAIN,
317 NULL, 0, lock_mode, &lockh);
319 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
320 rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
321 NULL, mds->mds_local_namespace, NULL,
322 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
323 &flags, (void *)mds_lock_callback,
325 if (rc != ELDLM_OK) {
326 CERROR("lock enqueue: err: %d\n", rc);
327 GOTO(out_create_de, rc = -EIO);
330 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
333 dchild = lookup_one_len(name, de, namelen - 1);
334 if (IS_ERR(dchild)) {
335 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
338 GOTO(out_create_dchild, rc = -ESTALE);
341 if (dchild->d_inode) {
342 struct mds_body *body;
344 struct inode *inode = dchild->d_inode;
345 CERROR("child exists (dir %ld, name %s, ino %ld)\n",
346 dir->i_ino, name, dchild->d_inode->i_ino);
348 body = lustre_msg_buf(req->rq_repmsg, offset);
349 mds_pack_inode2fid(&body->fid1, inode);
350 mds_pack_inode2body(body, inode);
351 if (S_ISREG(inode->i_mode)) {
352 obdo = lustre_msg_buf(req->rq_repmsg, offset + 1);
353 mds_fs_get_obdo(mds, inode, obdo);
355 /* now a normal case for intent locking */
365 ldlm_lock_decref(&lockh, lock_mode);
375 static int mds_getattr(int offset, struct ptlrpc_request *req)
377 struct mds_obd *mds = &req->rq_obd->u.mds;
378 struct obd_run_ctxt saved;
381 struct mds_body *body;
382 int rc, size[2] = {sizeof(*body)}, bufcount = 1;
385 body = lustre_msg_buf(req->rq_reqmsg, offset);
386 push_ctxt(&saved, &mds->mds_ctxt);
387 de = mds_fid2dentry(mds, &body->fid1, NULL);
389 GOTO(out_pop, rc = -ENOENT);
393 if (S_ISREG(body->fid1.f_type)) {
395 size[1] = sizeof(struct obdo);
396 } else if (body->valid & OBD_MD_LINKNAME) {
398 size[1] = inode->i_size;
401 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
403 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
404 CERROR("mds: out of memory\n");
408 if (body->valid & OBD_MD_LINKNAME) {
409 char *tmp = lustre_msg_buf(req->rq_repmsg, 1);
411 rc = inode->i_op->readlink(de, tmp, size[1]);
414 CERROR("readlink failed: %d\n", rc);
419 body = lustre_msg_buf(req->rq_repmsg, 0);
420 body->ino = inode->i_ino;
421 body->generation = inode->i_generation;
422 body->atime = inode->i_atime;
423 body->ctime = inode->i_ctime;
424 body->mtime = inode->i_mtime;
425 body->uid = inode->i_uid;
426 body->gid = inode->i_gid;
427 body->size = inode->i_size;
428 body->mode = inode->i_mode;
429 body->nlink = inode->i_nlink;
430 body->valid = ~0; /* FIXME: should be more selective */
432 if (S_ISREG(inode->i_mode)) {
433 rc = mds_fs_get_obdo(mds, inode,
434 lustre_msg_buf(req->rq_repmsg, 1));
436 CERROR("mds_fs_get_obdo failed: %d\n", rc);
448 static int mds_statfs(struct ptlrpc_request *req)
450 struct mds_obd *mds = &req->rq_obd->u.mds;
451 struct obd_statfs *osfs;
453 int rc, size = sizeof(*osfs);
456 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
458 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
459 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
463 rc = vfs_statfs(mds->mds_sb, &sfs);
465 CERROR("mds: statfs failed: rc %d\n", rc);
468 osfs = lustre_msg_buf(req->rq_repmsg, 0);
469 memset(osfs, 0, size);
470 /* FIXME: hack until the MDS gets the OST data, next month */
474 obd_statfs_pack(osfs, &sfs);
481 static int mds_open(struct ptlrpc_request *req)
484 struct mds_body *body;
486 struct vfsmount *mnt;
487 struct mds_obd *mds = &req->rq_obd->u.mds;
488 struct mds_client_info *mci;
490 struct list_head *tmp;
491 struct mds_file_data *mfd;
492 int rc, size = sizeof(*body);
495 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
496 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
497 CERROR("mds: out of memory\n");
498 req->rq_status = -ENOMEM;
502 mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
504 CERROR("mds: no mci!\n");
505 req->rq_status = -ENOTCONN;
509 body = lustre_msg_buf(req->rq_reqmsg, 0);
511 /* was this animal open already? */
512 /* XXX we chould only check on re-open, or do a refcount... */
513 list_for_each(tmp, &mci->mci_open_head) {
514 struct mds_file_data *fd;
515 fd = list_entry(tmp, struct mds_file_data, mfd_list);
516 if (body->extra == fd->mfd_clientfd &&
517 body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
518 CERROR("Re opening %Ld\n", body->fid1.id);
523 OBD_ALLOC(mfd, sizeof(*mfd));
525 CERROR("mds: out of memory\n");
526 req->rq_status = -ENOMEM;
530 de = mds_fid2dentry(mds, &body->fid1, &mnt);
532 req->rq_status = -ENOENT;
536 /* check if this inode has seen a delayed object creation */
537 if (req->rq_reqmsg->bufcount > 1) {
539 struct inode *inode = de->d_inode;
540 //struct iattr iattr;
544 obdo = lustre_msg_buf(req->rq_reqmsg, 1);
545 //iattr.ia_valid = ATTR_MODE;
546 //iattr.ia_mode = inode->i_mode;
548 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
550 req->rq_status = -ENOMEM;
554 /* XXX error handling */
555 rc = mds_fs_set_obdo(mds, inode, handle, obdo);
556 // rc = mds_fs_setattr(mds, de, handle, &iattr);
558 rc = mds_update_last_rcvd(mds, handle, req);
563 /* FIXME: need to return last_rcvd, last_committed */
565 /* FIXME: keep rc intact */
566 rc = mds_fs_commit(mds, de->d_inode, handle);
574 file = dentry_open(de, mnt, flags & ~O_DIRECT);
575 if (!file || IS_ERR(file)) {
576 req->rq_status = -EINVAL;
577 OBD_FREE(mfd, sizeof(*mfd));
581 file->private_data = mfd;
582 mfd->mfd_file = file;
583 mfd->mfd_clientfd = body->extra;
584 list_add(&mfd->mfd_list, &mci->mci_open_head);
586 body = lustre_msg_buf(req->rq_repmsg, 0);
587 body->extra = (__u64) (unsigned long)file;
592 int mds_close(struct ptlrpc_request *req)
595 struct mds_body *body;
597 struct mds_obd *mds = &req->rq_obd->u.mds;
598 struct vfsmount *mnt;
599 struct mds_file_data *mfd;
603 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
604 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
605 CERROR("mds: out of memory\n");
606 req->rq_status = -ENOMEM;
610 body = lustre_msg_buf(req->rq_reqmsg, 0);
611 de = mds_fid2dentry(mds, &body->fid1, &mnt);
613 req->rq_status = -ENOENT;
617 file = (struct file *)(unsigned long)body->extra;
620 mfd = (struct mds_file_data *)file->private_data;
621 list_del(&mfd->mfd_list);
622 OBD_FREE(mfd, sizeof(*mfd));
624 req->rq_status = filp_close(file, 0);
631 static int mds_readpage(struct mds_obd *mds, struct ptlrpc_request *req)
633 struct vfsmount *mnt;
636 struct mds_body *body;
637 struct obd_run_ctxt saved;
638 int rc, size = sizeof(*body);
641 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
642 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
643 CERROR("mds: out of memory\n");
644 GOTO(out, rc = -ENOMEM);
647 body = lustre_msg_buf(req->rq_reqmsg, 0);
648 push_ctxt(&saved, &mds->mds_ctxt);
649 de = mds_fid2dentry(mds, &body->fid1, &mnt);
651 GOTO(out_pop, rc = PTR_ERR(de));
653 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
655 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
656 /* note: in case of an error, dentry_open puts dentry */
658 GOTO(out_pop, rc = PTR_ERR(file));
660 /* to make this asynchronous make sure that the handling function
661 doesn't send a reply when this function completes. Instead a
662 callback function would send the reply */
663 rc = mds_sendpage(req, file, body->size);
673 int mds_reint(int offset, struct ptlrpc_request *req)
676 struct mds_update_record rec;
678 rc = mds_update_unpack(req, offset, &rec);
679 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
680 CERROR("invalid record\n");
681 req->rq_status = -EINVAL;
684 /* rc will be used to interrupt a for loop over multiple records */
685 rc = mds_reint_rec(&rec, offset, req);
689 int mds_handle(struct obd_device *dev, struct ptlrpc_service *svc,
690 struct ptlrpc_request *req)
692 struct mds_obd *mds = NULL;
696 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
697 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
698 CERROR("lustre_mds: Invalid request\n");
702 if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
703 CERROR("lustre_mds: wrong packet type sent %d\n",
704 req->rq_reqmsg->type);
705 GOTO(out, rc = -EINVAL);
708 if (req->rq_reqmsg->opc != MDS_CONNECT) {
709 int id = req->rq_reqmsg->target_id;
710 struct obd_device *obddev;
711 if (id < 0 || id > MAX_OBD_DEVICES)
712 GOTO(out, rc = -ENODEV);
713 obddev = &obd_dev[id];
714 if (strcmp(obddev->obd_type->typ_name, "mds") != 0)
715 GOTO(out, rc = -EINVAL);
716 mds = &obddev->u.mds;
717 req->rq_obd = obddev;
720 switch (req->rq_reqmsg->opc) {
722 CDEBUG(D_INODE, "connect\n");
723 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
724 rc = mds_connect(req, &mds);
728 CDEBUG(D_INODE, "disconnect\n");
729 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
730 rc = mds_disconnect(mds, req);
734 CDEBUG(D_INODE, "getstatus\n");
735 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
736 rc = mds_getstatus(mds, req);
740 CDEBUG(D_INODE, "getattr\n");
741 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
742 rc = mds_getattr(0, req);
746 CDEBUG(D_INODE, "statfs\n");
747 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
748 rc = mds_statfs(req);
752 CDEBUG(D_INODE, "readpage\n");
753 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
754 rc = mds_readpage(mds, req);
756 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
761 int size = sizeof(struct mds_body);
762 CDEBUG(D_INODE, "reint\n");
763 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
765 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
768 rc = req->rq_status = -ENOMEM;
771 rc = mds_reint(0, req);
772 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
777 CDEBUG(D_INODE, "open\n");
778 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
783 CDEBUG(D_INODE, "close\n");
784 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
789 rc = ptlrpc_error(svc, req);
795 /* Still not 100% sure whether we should reply with the server
796 * last_rcvd or that of this client. I'm not sure it even makes
797 * a difference on a per-client basis, because last_rcvd is global
798 * and we are not supposed to allow transactions while in recovery.
801 ptlrpc_error(svc, req);
803 req->rq_repmsg->last_rcvd = HTON__u64(mds->mds_last_rcvd);
804 req->rq_repmsg->last_committed =
805 HTON__u64(mds->mds_last_committed);
806 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
807 (unsigned long long)mds->mds_last_rcvd,
808 (unsigned long long)mds->mds_last_committed,
809 cpu_to_le32(req->rq_xid));
810 CDEBUG(D_NET, "sending reply\n");
811 ptlrpc_reply(svc, req);
816 /* Update the server data on disk. This stores the new mount_count and
817 * also the last_rcvd value to disk. If we don't have a clean shutdown,
818 * then the server last_rcvd value may be less than that of the clients.
819 * This will alert us that we may need to do client recovery.
822 int mds_update_server_data(struct mds_obd *mds)
824 struct obd_run_ctxt saved;
825 struct mds_server_data *msd = mds->mds_server_data;
826 struct file *filp = mds->mds_rcvd_filp;
830 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
831 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
833 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
834 (unsigned long long)mds->mds_mount_count,
835 (unsigned long long)mds->mds_last_rcvd);
836 push_ctxt(&saved, &mds->mds_ctxt);
837 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
838 if (rc != sizeof(*msd)) {
839 CERROR("error writing MDS server data: rc = %d\n", rc);
844 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
847 CERROR("error flushing MDS server data: rc = %d\n", rc);
852 /* Do recovery actions for the MDS */
853 static int mds_recover(struct obd_device *obddev)
855 struct mds_obd *mds = &obddev->u.mds;
858 /* This happens at the end when recovery is complete */
859 ++mds->mds_mount_count;
860 rc = mds_update_server_data(mds);
866 /* mount the file system (secretly) */
867 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
869 struct obd_ioctl_data* data = buf;
870 struct mds_obd *mds = &obddev->u.mds;
871 struct vfsmount *mnt;
876 #ifdef CONFIG_DEV_RDONLY
879 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
880 GOTO(err_dec, rc = -EINVAL);
882 mds->mds_fstype = strdup(data->ioc_inlbuf2);
884 mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
887 CERROR("do_kern_mount failed: rc = %d\n", rc);
891 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
893 GOTO(err_put, rc = -ENODEV);
895 rc = mds_fs_setup(mds, mnt);
897 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
901 mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
902 MDC_REPLY_PORTAL, "self",mds_handle);
903 if (!mds->mds_service) {
904 CERROR("failed to start service\n");
905 GOTO(err_fs, rc = -EINVAL);
908 rc = ptlrpc_start_thread(obddev, mds->mds_service, "lustre_mds");
910 CERROR("cannot start thread: rc = %d\n", rc);
915 mds->mds_ldlm_conn = ptlrpc_uuid_to_connection("self");
916 if (!mds->mds_ldlm_conn) {
918 GOTO(err_thread, rc);
921 obddev->obd_namespace =
922 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
923 if (obddev->obd_namespace == NULL) {
926 GOTO(err_thread, rc);
929 mds->mds_local_namespace =
930 ldlm_namespace_new("mds_client", LDLM_NAMESPACE_CLIENT);
931 if (mds->mds_local_namespace == NULL) {
934 GOTO(err_thread, rc);
937 OBD_ALLOC(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
938 if (mds->mds_ldlm_client == NULL) {
941 GOTO(err_thread, rc);
943 ptlrpc_init_client(NULL, NULL, LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
944 mds->mds_ldlm_client);
945 mds->mds_ldlm_client->cli_target_devno = obddev->obd_minor;
946 mds->mds_ldlm_client->cli_name = "mds ldlm";
948 rc = mds_recover(obddev);
950 GOTO(err_thread, rc);
955 ptlrpc_stop_all_threads(mds->mds_service);
957 ptlrpc_unregister_service(mds->mds_service);
962 mntput(mds->mds_vfsmnt);
966 kfree(mds->mds_fstype);
972 static int mds_cleanup(struct obd_device * obddev)
974 struct super_block *sb;
975 struct mds_obd *mds = &obddev->u.mds;
979 if ( !list_empty(&obddev->obd_gen_clients) ) {
980 CERROR("still has clients!\n");
984 ptlrpc_stop_all_threads(mds->mds_service);
985 ptlrpc_unregister_service(mds->mds_service);
991 mds_update_server_data(mds);
993 if (mds->mds_rcvd_filp) {
994 int rc = filp_close(mds->mds_rcvd_filp, 0);
995 mds->mds_rcvd_filp = NULL;
998 CERROR("last_rcvd file won't close, rc=%d\n", rc);
1002 mntput(mds->mds_vfsmnt);
1004 kfree(mds->mds_fstype);
1006 ldlm_namespace_free(mds->mds_local_namespace);
1007 ldlm_namespace_free(obddev->obd_namespace);
1009 if (mds->mds_ldlm_conn != NULL)
1010 ptlrpc_put_connection(mds->mds_ldlm_conn);
1012 OBD_FREE(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
1015 #ifdef CONFIG_DEV_RDONLY
1016 dev_clear_rdonly(2);
1018 mds_fs_cleanup(mds);
1024 /* use obd ops to offer management infrastructure */
1025 static struct obd_ops mds_obd_ops = {
1027 o_cleanup: mds_cleanup,
1030 static int __init mds_init(void)
1032 inter_module_register("mds_reint", THIS_MODULE, &mds_reint);
1033 inter_module_register("mds_getattr_name", THIS_MODULE,
1035 obd_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1039 static void __exit mds_exit(void)
1041 inter_module_unregister("mds_reint");
1042 inter_module_unregister("mds_getattr_name");
1043 obd_unregister_type(LUSTRE_MDS_NAME);
1046 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1047 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1048 MODULE_LICENSE("GPL");
1050 module_init(mds_init);
1051 module_exit(mds_exit);