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 #include <linux/init.h>
28 extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
30 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
31 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
32 struct ptlrpc_request *req);
33 static int mds_cleanup(struct obd_device * obddev);
35 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
37 return &req->rq_export->exp_obd->u.mds;
40 /* Assumes caller has already pushed into the kernel filesystem context */
41 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
45 struct mds_obd *mds = mds_req2mds(req);
46 struct ptlrpc_bulk_desc *desc;
47 struct ptlrpc_bulk_page *bulk;
51 desc = ptlrpc_prep_bulk(req->rq_connection);
53 GOTO(out, rc = -ENOMEM);
55 bulk = ptlrpc_prep_bulk_page(desc);
57 GOTO(cleanup_bulk, rc = -ENOMEM);
59 OBD_ALLOC(buf, PAGE_SIZE);
61 GOTO(cleanup_bulk, rc = -ENOMEM);
63 rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
66 GOTO(cleanup_buf, rc = -EIO);
68 bulk->b_xid = req->rq_xid;
70 bulk->b_buflen = PAGE_SIZE;
71 desc->b_portal = MDS_BULK_PORTAL;
73 rc = ptlrpc_send_bulk(desc);
75 GOTO(cleanup_buf, rc);
77 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
78 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
79 OBD_FAIL_MDS_SENDPAGE, rc);
80 ptlrpc_abort_bulk(desc);
81 GOTO(cleanup_buf, rc);
84 wait_event(desc->b_waitq, ptlrpc_check_bulk_sent(desc));
85 if (desc->b_flags & PTL_RPC_FL_INTR)
86 GOTO(cleanup_buf, rc = -EINTR);
90 OBD_FREE(buf, PAGE_SIZE);
92 ptlrpc_free_bulk(desc);
97 /* 'dir' is a inode for which a lock has already been taken */
98 struct dentry *mds_name2locked_dentry(struct mds_obd *mds, struct dentry *dir,
99 struct vfsmount **mnt, char *name,
100 int namelen, int lock_mode,
101 struct lustre_handle *lockh,
104 struct dentry *dchild;
106 __u64 res_id[3] = {0};
109 down(&dir->d_inode->i_sem);
110 dchild = lookup_one_len(name, dir, namelen);
111 if (IS_ERR(dchild)) {
112 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
113 up(&dir->d_inode->i_sem);
117 if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
118 up(&dir->d_inode->i_sem);
119 ldlm_lock_decref(lockh, dir_lock_mode);
122 if (lock_mode == 0 || !dchild->d_inode)
125 res_id[0] = dchild->d_inode->i_ino;
126 rc = ldlm_match_or_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
127 (struct lustre_handle *)&mds->mds_connh,
128 NULL, mds->mds_local_namespace, NULL,
129 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
130 &flags, (void *)mds_lock_callback, NULL,
132 if (rc != ELDLM_OK) {
134 up(&dir->d_inode->i_sem);
135 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
141 struct dentry *mds_fid2locked_dentry(struct mds_obd *mds, struct ll_fid *fid,
142 struct vfsmount **mnt, int lock_mode,
143 struct lustre_handle *lockh)
145 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
147 __u64 res_id[3] = {0};
153 res_id[0] = de->d_inode->i_ino;
154 rc = ldlm_match_or_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
155 (struct lustre_handle *)&mds->mds_connh,
156 NULL, mds->mds_local_namespace, NULL,
157 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
158 &flags, (void *)mds_lock_callback, NULL,
160 if (rc != ELDLM_OK) {
162 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
168 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
169 struct vfsmount **mnt)
171 /* stolen from NFS */
172 struct super_block *sb = mds->mds_sb;
173 unsigned long ino = fid->id;
174 __u32 generation = fid->generation;
176 struct list_head *lp;
177 struct dentry *result;
180 RETURN(ERR_PTR(-ESTALE));
182 inode = iget(sb, ino);
184 RETURN(ERR_PTR(-ENOMEM));
186 CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
188 if (is_bad_inode(inode) ||
189 (generation && inode->i_generation != generation)) {
190 /* we didn't find the right inode.. */
191 CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
192 inode->i_ino, inode->i_nlink,
193 atomic_read(&inode->i_count), inode->i_generation,
197 RETURN(ERR_PTR(-ESTALE));
200 /* now to find a dentry.
201 * If possible, get a well-connected one
204 *mnt = mds->mds_vfsmnt;
205 spin_lock(&dcache_lock);
206 for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
207 result = list_entry(lp,struct dentry, d_alias);
208 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
210 result->d_vfs_flags |= DCACHE_REFERENCED;
211 spin_unlock(&dcache_lock);
218 spin_unlock(&dcache_lock);
219 result = d_alloc_root(inode);
220 if (result == NULL) {
222 return ERR_PTR(-ENOMEM);
226 result->d_flags |= DCACHE_NFSD_DISCONNECTED;
230 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd)
235 rc = class_connect(conn, obd);
243 static int mds_disconnect(struct lustre_handle *conn)
247 rc = class_disconnect(conn);
254 /* FIXME: the error cases need fixing to avoid leaks */
255 static int mds_getstatus(struct ptlrpc_request *req)
257 struct mds_obd *mds = mds_req2mds(req);
258 struct mds_body *body;
259 struct mds_client_info *mci;
260 struct mds_client_data *mcd;
261 int rc, size = sizeof(*body);
264 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
265 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
266 CERROR("mds: out of memory for message: size=%d\n", size);
267 req->rq_status = -ENOMEM;
271 body = lustre_msg_buf(req->rq_reqmsg, 0);
272 mds_unpack_body(body);
274 /* Anything we need to do here with the client's trans no or so? */
275 body = lustre_msg_buf(req->rq_repmsg, 0);
276 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
278 mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
280 /* We don't have any old connection data for this client */
283 CDEBUG(D_INFO, "allocating new client data for UUID '%s'",
284 ptlrpc_req_to_uuid(req));
286 OBD_ALLOC(mcd, sizeof(*mcd));
288 CERROR("mds: out of memory for client data\n");
289 req->rq_status = -ENOMEM;
292 memcpy(mcd->mcd_uuid, ptlrpc_req_to_uuid(req),
293 sizeof(mcd->mcd_uuid));
294 rc = mds_client_add(mds, mcd, -1);
300 /* We have old connection data for this client... */
302 CDEBUG(D_INFO, "found existing data for UUID '%s' at #%d\n",
303 mcd->mcd_uuid, mci->mci_off);
305 /* mcd_last_xid is is stored in little endian on the disk and
306 mds_pack_rep_body converts it to network order */
307 body->last_xid = le32_to_cpu(mcd->mcd_last_xid);
308 mds_pack_rep_body(req);
312 static int mds_getlovinfo(struct ptlrpc_request *req)
314 struct mds_obd *mds = mds_req2mds(req);
315 struct mds_status_req *streq;
316 struct lov_desc *desc;
318 int rc, size[2] = {sizeof(*desc)};
321 streq = lustre_msg_buf(req->rq_reqmsg, 0);
322 streq->flags = NTOH__u32(streq->flags);
323 streq->repbuf = NTOH__u32(streq->repbuf);
324 size[1] = streq->repbuf;
326 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
328 CERROR("mds: out of memory for message: size=%d\n", size[1]);
329 req->rq_status = -ENOMEM;
333 desc = lustre_msg_buf(req->rq_repmsg, 0);
334 rc = mds_get_lovdesc(req->rq_obd, desc);
336 CERROR("mds_get_lovdesc error %d", rc);
341 tgt_count = NTOH__u32(desc->ld_tgt_count);
342 if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
343 CERROR("too many targets, enlarge client buffers\n");
344 req->rq_status = -ENOSPC;
348 mds->mds_max_mdsize = sizeof(desc) + tgt_count * sizeof(uuid_t);
349 rc = mds_get_lovtgts(req->rq_obd, tgt_count,
350 lustre_msg_buf(req->rq_repmsg, 1));
352 CERROR("get_lovtgts error %d\n", rc);
359 int mds_lock_callback(struct lustre_handle *lockh, struct ldlm_lock_desc *desc,
360 void *data, int data_len, struct ptlrpc_request **reqp)
365 /* Completion AST. Do nothing */
369 if (ldlm_cli_cancel(lockh) < 0)
374 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
376 struct mds_obd *mds = mds_req2mds(req);
377 struct obd_run_ctxt saved;
378 struct mds_body *body;
379 struct dentry *de = NULL, *dchild = NULL;
381 struct lustre_handle lockh;
383 int namelen, flags, lock_mode, rc = 0;
384 __u64 res_id[3] = {0, 0, 0};
387 if (strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds") != 0)
390 if (req->rq_reqmsg->bufcount <= offset + 1) {
392 GOTO(out_pre_de, rc = -EINVAL);
395 body = lustre_msg_buf(req->rq_reqmsg, offset);
396 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
397 namelen = req->rq_reqmsg->buflens[offset + 1];
398 /* requests were at offset 2, replies go back at 1 */
402 push_ctxt(&saved, &mds->mds_ctxt);
403 de = mds_fid2dentry(mds, &body->fid1, NULL);
406 GOTO(out_pre_de, rc = -ESTALE);
410 CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
412 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
413 res_id[0] = dir->i_ino;
415 rc = ldlm_lock_match(mds->mds_local_namespace, res_id, LDLM_PLAIN,
416 NULL, 0, lock_mode, &lockh);
418 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
419 rc = ldlm_cli_enqueue(mds->mds_ldlm_client, mds->mds_ldlm_conn,
420 (struct lustre_handle *)&mds->mds_connh,
421 NULL, mds->mds_local_namespace, NULL,
422 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
423 &flags, (void *)mds_lock_callback,
425 if (rc != ELDLM_OK) {
426 CERROR("lock enqueue: err: %d\n", rc);
427 GOTO(out_create_de, rc = -EIO);
430 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
433 dchild = lookup_one_len(name, de, namelen - 1);
434 if (IS_ERR(dchild)) {
435 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
438 GOTO(out_create_dchild, rc = -ESTALE);
441 if (dchild->d_inode) {
442 struct mds_body *body;
443 struct inode *inode = dchild->d_inode;
444 CDEBUG(D_INODE, "child exists (dir %ld, name %s, ino %ld)\n",
445 dir->i_ino, name, dchild->d_inode->i_ino);
447 body = lustre_msg_buf(req->rq_repmsg, offset);
448 mds_pack_inode2fid(&body->fid1, inode);
449 mds_pack_inode2body(body, inode);
450 if (S_ISREG(inode->i_mode)) {
451 struct lov_stripe_md *md;
452 md = lustre_msg_buf(req->rq_repmsg, offset + 1);
453 md->lmd_size = mds->mds_max_mdsize;
454 mds_fs_get_md(mds, inode, md);
456 /* now a normal case for intent locking */
465 ldlm_lock_decref(&lockh, lock_mode);
475 static int mds_getattr(int offset, struct ptlrpc_request *req)
477 struct mds_obd *mds = mds_req2mds(req);
478 struct obd_run_ctxt saved;
481 struct mds_body *body;
482 int rc, size[2] = {sizeof(*body)}, bufcount = 1;
485 body = lustre_msg_buf(req->rq_reqmsg, offset);
486 push_ctxt(&saved, &mds->mds_ctxt);
487 de = mds_fid2dentry(mds, &body->fid1, NULL);
489 GOTO(out_pop, rc = -ENOENT);
493 if (S_ISREG(body->fid1.f_type)) {
495 size[1] = mds->mds_max_mdsize;
496 } else if (body->valid & OBD_MD_LINKNAME) {
498 size[1] = inode->i_size;
501 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
503 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
504 CERROR("mds: out of memory\n");
508 if (body->valid & OBD_MD_LINKNAME) {
509 char *tmp = lustre_msg_buf(req->rq_repmsg, 1);
511 rc = inode->i_op->readlink(de, tmp, size[1]);
514 CERROR("readlink failed: %d\n", rc);
519 body = lustre_msg_buf(req->rq_repmsg, 0);
520 body->ino = inode->i_ino;
521 body->generation = inode->i_generation;
522 body->atime = inode->i_atime;
523 body->ctime = inode->i_ctime;
524 body->mtime = inode->i_mtime;
525 body->uid = inode->i_uid;
526 body->gid = inode->i_gid;
527 body->size = inode->i_size;
528 body->mode = inode->i_mode;
529 body->nlink = inode->i_nlink;
530 body->valid = ~0; /* FIXME: should be more selective */
532 if (S_ISREG(inode->i_mode)) {
533 rc = mds_fs_get_md(mds, inode,
534 lustre_msg_buf(req->rq_repmsg, 1));
536 CERROR("mds_fs_get_md failed: %d\n", rc);
548 static int mds_statfs(struct ptlrpc_request *req)
550 struct mds_obd *mds = mds_req2mds(req);
551 struct obd_statfs *osfs;
553 int rc, size = sizeof(*osfs);
556 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
558 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
559 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
563 rc = vfs_statfs(mds->mds_sb, &sfs);
565 CERROR("mds: statfs failed: rc %d\n", rc);
568 osfs = lustre_msg_buf(req->rq_repmsg, 0);
569 memset(osfs, 0, size);
570 obd_statfs_pack(osfs, &sfs);
577 static int mds_open(struct ptlrpc_request *req)
580 struct mds_body *body;
582 struct vfsmount *mnt;
583 struct mds_obd *mds = mds_req2mds(req);
584 struct mds_client_info *mci;
586 struct list_head *tmp;
587 struct mds_file_data *mfd;
588 int rc, size = sizeof(*body);
591 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
592 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
593 CERROR("mds: out of memory\n");
594 req->rq_status = -ENOMEM;
598 mci = mds_uuid_to_mci(mds, ptlrpc_req_to_uuid(req));
600 CERROR("mds: no mci!\n");
601 req->rq_status = -ENOTCONN;
605 body = lustre_msg_buf(req->rq_reqmsg, 0);
607 /* was this animal open already? */
608 /* XXX we should only check on re-open, or do a refcount... */
609 list_for_each(tmp, &mci->mci_open_head) {
610 struct mds_file_data *fd;
611 fd = list_entry(tmp, struct mds_file_data, mfd_list);
612 if (body->extra == fd->mfd_clientfd &&
613 body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
614 CERROR("Re opening %Ld\n", body->fid1.id);
619 OBD_ALLOC(mfd, sizeof(*mfd));
621 CERROR("mds: out of memory\n");
622 req->rq_status = -ENOMEM;
626 de = mds_fid2dentry(mds, &body->fid1, &mnt);
628 req->rq_status = -ENOENT;
632 /* check if this inode has seen a delayed object creation */
633 if (req->rq_reqmsg->bufcount > 1) {
635 struct lov_stripe_md *md;
636 struct inode *inode = de->d_inode;
637 //struct iattr iattr;
640 md = lustre_msg_buf(req->rq_reqmsg, 1);
641 //iattr.ia_mode = inode->i_mode;
643 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
645 req->rq_status = -ENOMEM;
649 /* XXX error handling */
650 rc = mds_fs_set_md(mds, inode, handle, md);
651 // rc = mds_fs_setattr(mds, de, handle, &iattr);
653 struct obd_run_ctxt saved;
654 push_ctxt(&saved, &mds->mds_ctxt);
655 rc = mds_update_last_rcvd(mds, handle, req);
661 /* FIXME: need to return last_rcvd, last_committed */
663 /* FIXME: keep rc intact */
664 rc = mds_fs_commit(mds, de->d_inode, handle);
672 file = dentry_open(de, mnt, flags & ~O_DIRECT);
673 if (!file || IS_ERR(file)) {
674 req->rq_status = -EINVAL;
675 OBD_FREE(mfd, sizeof(*mfd));
679 file->private_data = mfd;
680 mfd->mfd_file = file;
681 mfd->mfd_clientfd = body->extra;
682 list_add(&mfd->mfd_list, &mci->mci_open_head);
684 body = lustre_msg_buf(req->rq_repmsg, 0);
685 body->extra = (__u64) (unsigned long)file;
689 static int mds_close(struct ptlrpc_request *req)
692 struct mds_body *body;
694 struct mds_obd *mds = mds_req2mds(req);
695 struct vfsmount *mnt;
696 struct mds_file_data *mfd;
700 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
701 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
702 CERROR("mds: out of memory\n");
703 req->rq_status = -ENOMEM;
707 body = lustre_msg_buf(req->rq_reqmsg, 0);
708 de = mds_fid2dentry(mds, &body->fid1, &mnt);
710 req->rq_status = -ENOENT;
714 file = (struct file *)(unsigned long)body->extra;
717 mfd = (struct mds_file_data *)file->private_data;
718 list_del(&mfd->mfd_list);
719 OBD_FREE(mfd, sizeof(*mfd));
721 req->rq_status = filp_close(file, 0);
728 static int mds_readpage(struct ptlrpc_request *req)
730 struct mds_obd *mds = mds_req2mds(req);
731 struct vfsmount *mnt;
734 struct mds_body *body;
735 struct obd_run_ctxt saved;
736 int rc, size = sizeof(*body);
739 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
740 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
741 CERROR("mds: out of memory\n");
742 GOTO(out, rc = -ENOMEM);
745 body = lustre_msg_buf(req->rq_reqmsg, 0);
746 push_ctxt(&saved, &mds->mds_ctxt);
747 de = mds_fid2dentry(mds, &body->fid1, &mnt);
749 GOTO(out_pop, rc = PTR_ERR(de));
751 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
753 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
754 /* note: in case of an error, dentry_open puts dentry */
756 GOTO(out_pop, rc = PTR_ERR(file));
758 /* to make this asynchronous make sure that the handling function
759 doesn't send a reply when this function completes. Instead a
760 callback function would send the reply */
761 rc = mds_sendpage(req, file, body->size);
771 int mds_reint(int offset, struct ptlrpc_request *req)
774 struct mds_update_record rec;
776 rc = mds_update_unpack(req, offset, &rec);
777 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
778 CERROR("invalid record\n");
779 req->rq_status = -EINVAL;
782 /* rc will be used to interrupt a for loop over multiple records */
783 rc = mds_reint_rec(&rec, offset, req);
787 int mds_handle(struct ptlrpc_request *req)
792 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
793 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
794 CERROR("lustre_mds: Invalid request\n");
798 if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
799 CERROR("lustre_mds: wrong packet type sent %d\n",
800 req->rq_reqmsg->type);
801 GOTO(out, rc = -EINVAL);
804 if (req->rq_reqmsg->opc != MDS_CONNECT &&
805 req->rq_export == NULL)
806 GOTO(out, rc = -ENOTCONN);
808 if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
809 GOTO(out, rc = -EINVAL);
811 switch (req->rq_reqmsg->opc) {
813 CDEBUG(D_INODE, "connect\n");
814 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
815 rc = target_handle_connect(req);
819 CDEBUG(D_INODE, "disconnect\n");
820 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
821 rc = target_handle_disconnect(req);
825 CDEBUG(D_INODE, "getstatus\n");
826 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
827 rc = mds_getstatus(req);
831 CDEBUG(D_INODE, "getlovinfo\n");
832 rc = mds_getlovinfo(req);
836 CDEBUG(D_INODE, "getattr\n");
837 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
838 rc = mds_getattr(0, req);
842 CDEBUG(D_INODE, "statfs\n");
843 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
844 rc = mds_statfs(req);
848 CDEBUG(D_INODE, "readpage\n");
849 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
850 rc = mds_readpage(req);
852 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
857 int size = sizeof(struct mds_body);
858 CDEBUG(D_INODE, "reint\n");
859 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
861 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
864 rc = req->rq_status = -ENOMEM;
867 rc = mds_reint(0, req);
868 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
873 CDEBUG(D_INODE, "open\n");
874 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
879 CDEBUG(D_INODE, "close\n");
880 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
885 rc = ptlrpc_error(req->rq_svc, req);
892 struct mds_obd *mds = mds_req2mds(req);
893 req->rq_repmsg->last_rcvd = HTON__u64(mds->mds_last_rcvd);
894 req->rq_repmsg->last_committed =
895 HTON__u64(mds->mds_last_committed);
896 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
897 (unsigned long long)mds->mds_last_rcvd,
898 (unsigned long long)mds->mds_last_committed,
899 cpu_to_le32(req->rq_xid));
902 /* Still not 100% sure whether we should reply with the server
903 * last_rcvd or that of this client. I'm not sure it even makes
904 * a difference on a per-client basis, because last_rcvd is global
905 * and we are not supposed to allow transactions while in recovery.
908 CERROR("mds: processing error %d\n", rc);
909 ptlrpc_error(req->rq_svc, req);
911 CDEBUG(D_NET, "sending reply\n");
912 ptlrpc_reply(req->rq_svc, req);
917 /* Update the server data on disk. This stores the new mount_count and
918 * also the last_rcvd value to disk. If we don't have a clean shutdown,
919 * then the server last_rcvd value may be less than that of the clients.
920 * This will alert us that we may need to do client recovery.
923 int mds_update_server_data(struct mds_obd *mds)
925 struct obd_run_ctxt saved;
926 struct mds_server_data *msd = mds->mds_server_data;
927 struct file *filp = mds->mds_rcvd_filp;
931 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
932 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
934 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
935 (unsigned long long)mds->mds_mount_count,
936 (unsigned long long)mds->mds_last_rcvd);
937 push_ctxt(&saved, &mds->mds_ctxt);
938 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
939 if (rc != sizeof(*msd)) {
940 CERROR("error writing MDS server data: rc = %d\n", rc);
945 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
948 CERROR("error flushing MDS server data: rc = %d\n", rc);
953 /* Do recovery actions for the MDS */
954 static int mds_recover(struct obd_device *obddev)
956 struct mds_obd *mds = &obddev->u.mds;
959 /* This happens at the end when recovery is complete */
960 ++mds->mds_mount_count;
961 rc = mds_update_server_data(mds);
967 /* mount the file system (secretly) */
968 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
970 struct obd_ioctl_data* data = buf;
971 struct obd_export *export;
972 struct mds_obd *mds = &obddev->u.mds;
973 struct vfsmount *mnt;
978 #ifdef CONFIG_DEV_RDONLY
981 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
982 GOTO(err_dec, rc = -EINVAL);
984 mds->mds_fstype = strdup(data->ioc_inlbuf2);
986 mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
989 CERROR("do_kern_mount failed: rc = %d\n", rc);
993 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
995 GOTO(err_put, rc = -ENODEV);
997 mds->mds_max_mdsize = sizeof(struct lov_stripe_md);
998 rc = mds_fs_setup(mds, mnt);
1000 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1004 mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
1005 MDC_REPLY_PORTAL, "self",mds_handle);
1006 if (!mds->mds_service) {
1007 CERROR("failed to start service\n");
1008 GOTO(err_fs, rc = -EINVAL);
1011 rc = ptlrpc_start_thread(obddev, mds->mds_service, "lustre_mds");
1013 CERROR("cannot start thread: rc = %d\n", rc);
1018 mds->mds_ldlm_conn = ptlrpc_uuid_to_connection("self");
1019 if (!mds->mds_ldlm_conn) {
1020 mds_cleanup(obddev);
1021 GOTO(err_thread, rc);
1024 obddev->obd_namespace =
1025 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1026 if (obddev->obd_namespace == NULL) {
1028 mds_cleanup(obddev);
1029 GOTO(err_thread, rc);
1032 mds->mds_local_namespace =
1033 ldlm_namespace_new("mds_client", LDLM_NAMESPACE_CLIENT);
1034 if (mds->mds_local_namespace == NULL) {
1036 mds_cleanup(obddev);
1037 GOTO(err_thread, rc);
1040 OBD_ALLOC(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
1041 if (mds->mds_ldlm_client == NULL) {
1043 mds_cleanup(obddev);
1044 GOTO(err_thread, rc);
1046 ptlrpc_init_client(NULL, NULL, LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1047 mds->mds_ldlm_client);
1048 mds->mds_ldlm_client->cli_target_devno = obddev->obd_minor;
1049 mds->mds_ldlm_client->cli_name = "mds ldlm";
1051 rc = mds_recover(obddev);
1053 GOTO(err_thread, rc);
1055 rc = class_connect(&mds->mds_connh, obddev);
1057 GOTO(err_thread, rc);
1058 export = class_conn2export(&mds->mds_connh);
1061 export->exp_connection = mds->mds_ldlm_conn;
1068 ptlrpc_stop_all_threads(mds->mds_service);
1070 ptlrpc_unregister_service(mds->mds_service);
1072 mds_fs_cleanup(mds);
1075 mntput(mds->mds_vfsmnt);
1079 kfree(mds->mds_fstype);
1085 static int mds_cleanup(struct obd_device * obddev)
1087 struct super_block *sb;
1088 struct mds_obd *mds = &obddev->u.mds;
1091 class_disconnect(&mds->mds_connh);
1094 if ( !list_empty(&obddev->obd_exports) ) {
1095 CERROR("still has exports!\n");
1099 ptlrpc_stop_all_threads(mds->mds_service);
1100 ptlrpc_unregister_service(mds->mds_service);
1106 mds_update_server_data(mds);
1108 if (mds->mds_rcvd_filp) {
1109 int rc = filp_close(mds->mds_rcvd_filp, 0);
1110 mds->mds_rcvd_filp = NULL;
1113 CERROR("last_rcvd file won't close, rc=%d\n", rc);
1117 mntput(mds->mds_vfsmnt);
1119 kfree(mds->mds_fstype);
1121 ldlm_namespace_free(mds->mds_local_namespace);
1122 ldlm_namespace_free(obddev->obd_namespace);
1124 if (mds->mds_ldlm_conn != NULL)
1125 ptlrpc_put_connection(mds->mds_ldlm_conn);
1127 OBD_FREE(mds->mds_ldlm_client, sizeof(*mds->mds_ldlm_client));
1130 #ifdef CONFIG_DEV_RDONLY
1131 dev_clear_rdonly(2);
1133 mds_fs_cleanup(mds);
1139 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1140 int len, void *karg, void *uarg);
1142 /* use obd ops to offer management infrastructure */
1143 static struct obd_ops mds_obd_ops = {
1144 o_connect: mds_connect,
1145 o_disconnect: mds_disconnect,
1147 o_cleanup: mds_cleanup,
1148 o_iocontrol: mds_iocontrol
1151 static int __init mds_init(void)
1153 inter_module_register("mds_reint", THIS_MODULE, &mds_reint);
1154 inter_module_register("mds_getattr_name", THIS_MODULE,
1156 class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1160 static void __exit mds_exit(void)
1162 inter_module_unregister("mds_reint");
1163 inter_module_unregister("mds_getattr_name");
1164 class_unregister_type(LUSTRE_MDS_NAME);
1167 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1168 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1169 MODULE_LICENSE("GPL");
1171 module_init(mds_init);
1172 module_exit(mds_exit);