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>
19 #define DEBUG_SUBSYSTEM S_MDS
21 #include <linux/module.h>
22 #include <linux/lustre_mds.h>
23 #include <linux/lustre_dlm.h>
24 #include <linux/init.h>
26 extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
28 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
29 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
30 struct ptlrpc_request *req);
31 static int mds_cleanup(struct obd_device * obddev);
33 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
35 return &req->rq_export->exp_obd->u.mds;
39 /* Assumes caller has already pushed into the kernel filesystem context */
40 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
44 struct mds_obd *mds = mds_req2mds(req);
45 struct ptlrpc_bulk_desc *desc;
46 struct ptlrpc_bulk_page *bulk;
50 desc = ptlrpc_prep_bulk(req->rq_connection);
52 GOTO(out, rc = -ENOMEM);
54 bulk = ptlrpc_prep_bulk_page(desc);
56 GOTO(cleanup_bulk, rc = -ENOMEM);
58 OBD_ALLOC(buf, PAGE_SIZE);
60 GOTO(cleanup_bulk, rc = -ENOMEM);
62 rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
65 GOTO(cleanup_buf, rc = -EIO);
67 bulk->b_xid = req->rq_xid;
69 bulk->b_buflen = PAGE_SIZE;
70 desc->b_portal = MDS_BULK_PORTAL;
72 rc = ptlrpc_send_bulk(desc);
74 GOTO(cleanup_buf, rc);
76 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
77 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
78 OBD_FAIL_MDS_SENDPAGE, rc);
79 ptlrpc_abort_bulk(desc);
80 GOTO(cleanup_buf, rc);
83 wait_event(desc->b_waitq, ptlrpc_check_bulk_sent(desc));
84 if (desc->b_flags & PTL_RPC_FL_INTR)
85 GOTO(cleanup_buf, rc = -EINTR);
89 OBD_FREE(buf, PAGE_SIZE);
91 ptlrpc_free_bulk(desc);
96 /* 'dir' is a inode for which a lock has already been taken */
97 struct dentry *mds_name2locked_dentry(struct obd_device *obd, struct dentry *dir,
98 struct vfsmount **mnt, char *name,
99 int namelen, int lock_mode,
100 struct lustre_handle *lockh,
103 struct dentry *dchild;
105 __u64 res_id[3] = {0};
108 down(&dir->d_inode->i_sem);
109 dchild = lookup_one_len(name, dir, namelen);
110 if (IS_ERR(dchild)) {
111 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
112 up(&dir->d_inode->i_sem);
116 if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
117 up(&dir->d_inode->i_sem);
118 ldlm_lock_decref(lockh, dir_lock_mode);
121 if (lock_mode == 0 || !dchild->d_inode)
124 res_id[0] = dchild->d_inode->i_ino;
125 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
126 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
127 &flags, ldlm_completion_ast, (void *)mds_lock_callback, NULL,
129 if (rc != ELDLM_OK) {
131 up(&dir->d_inode->i_sem);
132 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
138 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
139 struct vfsmount **mnt, int lock_mode,
140 struct lustre_handle *lockh)
142 struct mds_obd *mds = &obd->u.mds;
143 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
145 __u64 res_id[3] = {0};
151 res_id[0] = de->d_inode->i_ino;
152 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
153 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
154 &flags, ldlm_completion_ast, (void *)mds_lock_callback, NULL,
156 if (rc != ELDLM_OK) {
158 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
164 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
165 struct vfsmount **mnt)
167 /* stolen from NFS */
168 struct super_block *sb = mds->mds_sb;
169 unsigned long ino = fid->id;
170 __u32 generation = fid->generation;
172 struct list_head *lp;
173 struct dentry *result;
176 RETURN(ERR_PTR(-ESTALE));
178 inode = iget(sb, ino);
180 RETURN(ERR_PTR(-ENOMEM));
182 CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
184 if (is_bad_inode(inode) ||
185 (generation && inode->i_generation != generation)) {
186 /* we didn't find the right inode.. */
187 CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
188 inode->i_ino, inode->i_nlink,
189 atomic_read(&inode->i_count), inode->i_generation,
193 RETURN(ERR_PTR(-ESTALE));
196 /* now to find a dentry. If possible, get a well-connected one */
198 *mnt = mds->mds_vfsmnt;
199 spin_lock(&dcache_lock);
200 for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
201 result = list_entry(lp,struct dentry, d_alias);
202 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
204 result->d_vfs_flags |= DCACHE_REFERENCED;
205 spin_unlock(&dcache_lock);
212 spin_unlock(&dcache_lock);
213 result = d_alloc_root(inode);
214 if (result == NULL) {
216 return ERR_PTR(-ENOMEM);
220 result->d_flags |= DCACHE_NFSD_DISCONNECTED;
224 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd)
228 #warning shaver: find existing export if there is one.
230 rc = class_connect(conn, obd);
238 static int mds_disconnect(struct lustre_handle *conn)
242 rc = class_disconnect(conn);
249 /* FIXME: the error cases need fixing to avoid leaks */
250 static int mds_getstatus(struct ptlrpc_request *req)
252 struct mds_obd *mds = mds_req2mds(req);
253 struct mds_body *body;
254 struct mds_client_data *mcd;
255 struct mds_export_data *med = &req->rq_export->exp_mds_data;
256 int rc, size = sizeof(*body);
259 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
260 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
261 CERROR("mds: out of memory for message: size=%d\n", size);
262 req->rq_status = -ENOMEM;
266 body = lustre_msg_buf(req->rq_reqmsg, 0);
267 mds_unpack_body(body);
269 /* Anything we need to do here with the client's trans no or so? */
270 body = lustre_msg_buf(req->rq_repmsg, 0);
271 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
277 CDEBUG(D_INFO, "allocating new client data for UUID '%s'\n",
278 ptlrpc_req_to_uuid(req));
279 OBD_ALLOC(mcd, sizeof(*mcd));
281 CERROR("mds: out of memory for client data\n");
282 req->rq_status = -ENOMEM;
285 memcpy(mcd->mcd_uuid, ptlrpc_req_to_uuid(req),
286 sizeof(mcd->mcd_uuid));
287 rc = mds_client_add(mds, med, -1);
290 OBD_FREE(mcd, sizeof(*mcd));
295 CDEBUG(D_INFO, "found existing data for UUID '%s' at #%d\n",
296 mcd->mcd_uuid, med->med_off);
298 /* mcd_last_xid is is stored in little endian on the disk and
299 mds_pack_rep_body converts it to network order */
300 body->last_xid = le32_to_cpu(mcd->mcd_last_xid);
301 mds_pack_rep_body(req);
305 static int mds_getlovinfo(struct ptlrpc_request *req)
307 struct mds_obd *mds = mds_req2mds(req);
308 struct mds_status_req *streq;
309 struct lov_desc *desc;
311 int rc, size[2] = {sizeof(*desc)};
314 streq = lustre_msg_buf(req->rq_reqmsg, 0);
315 streq->flags = NTOH__u32(streq->flags);
316 streq->repbuf = NTOH__u32(streq->repbuf);
317 size[1] = streq->repbuf;
319 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
321 CERROR("mds: out of memory for message: size=%d\n", size[1]);
322 req->rq_status = -ENOMEM;
326 desc = lustre_msg_buf(req->rq_repmsg, 0);
327 rc = mds_get_lovdesc(req->rq_obd, desc);
329 CERROR("mds_get_lovdesc error %d", rc);
334 tgt_count = NTOH__u32(desc->ld_tgt_count);
335 if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
336 CERROR("too many targets, enlarge client buffers\n");
337 req->rq_status = -ENOSPC;
341 mds->mds_max_mdsize = sizeof(desc) + tgt_count * sizeof(uuid_t);
342 rc = mds_get_lovtgts(req->rq_obd, tgt_count,
343 lustre_msg_buf(req->rq_repmsg, 1));
345 CERROR("get_lovtgts error %d\n", rc);
352 int mds_lock_callback(struct lustre_handle *lockh, struct ldlm_lock_desc *desc,
353 void *data, int data_len, struct ptlrpc_request **reqp)
358 /* Completion AST. Do nothing */
362 if (ldlm_cli_cancel(lockh) < 0)
367 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
369 struct mds_obd *mds = mds_req2mds(req);
370 struct obd_device *obd = req->rq_export->exp_obd;
371 struct obd_run_ctxt saved;
372 struct mds_body *body;
373 struct dentry *de = NULL, *dchild = NULL;
375 struct lustre_handle lockh;
377 int namelen, flags, lock_mode, rc = 0;
378 __u64 res_id[3] = {0, 0, 0};
381 if (strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds") != 0)
384 if (req->rq_reqmsg->bufcount <= offset + 1) {
386 GOTO(out_pre_de, rc = -EINVAL);
389 body = lustre_msg_buf(req->rq_reqmsg, offset);
390 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
391 namelen = req->rq_reqmsg->buflens[offset + 1];
392 /* requests were at offset 2, replies go back at 1 */
396 push_ctxt(&saved, &mds->mds_ctxt);
397 de = mds_fid2dentry(mds, &body->fid1, NULL);
400 GOTO(out_pre_de, rc = -ESTALE);
404 CDEBUG(D_INODE, "parent ino %ld\n", dir->i_ino);
406 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
407 res_id[0] = dir->i_ino;
409 rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
410 NULL, 0, lock_mode, &lockh);
412 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
413 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
414 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
415 &flags, ldlm_completion_ast, (void *)mds_lock_callback,
417 if (rc != ELDLM_OK) {
418 CERROR("lock enqueue: err: %d\n", rc);
419 GOTO(out_create_de, rc = -EIO);
422 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
425 dchild = lookup_one_len(name, de, namelen - 1);
426 if (IS_ERR(dchild)) {
427 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
430 GOTO(out_create_dchild, rc = -ESTALE);
433 if (dchild->d_inode) {
434 struct mds_body *body;
435 struct inode *inode = dchild->d_inode;
436 CDEBUG(D_INODE, "child exists (dir %ld, name %s, ino %ld)\n",
437 dir->i_ino, name, dchild->d_inode->i_ino);
439 body = lustre_msg_buf(req->rq_repmsg, offset);
440 mds_pack_inode2fid(&body->fid1, inode);
441 mds_pack_inode2body(body, inode);
442 if (S_ISREG(inode->i_mode)) {
443 struct lov_stripe_md *md;
444 md = lustre_msg_buf(req->rq_repmsg, offset + 1);
445 md->lmd_easize = mds->mds_max_mdsize;
446 mds_fs_get_md(mds, inode, md);
448 /* now a normal case for intent locking */
457 ldlm_lock_decref(&lockh, lock_mode);
467 static int mds_getattr(int offset, struct ptlrpc_request *req)
469 struct mds_obd *mds = mds_req2mds(req);
470 struct obd_run_ctxt saved;
473 struct mds_body *body;
474 int rc, size[2] = {sizeof(*body)}, bufcount = 1;
477 body = lustre_msg_buf(req->rq_reqmsg, offset);
478 push_ctxt(&saved, &mds->mds_ctxt);
479 de = mds_fid2dentry(mds, &body->fid1, NULL);
481 GOTO(out_pop, rc = -ENOENT);
485 if (S_ISREG(body->fid1.f_type)) {
487 size[1] = mds->mds_max_mdsize;
488 } else if (body->valid & OBD_MD_LINKNAME) {
490 size[1] = inode->i_size;
493 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
495 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
496 CERROR("mds: out of memory\n");
500 if (body->valid & OBD_MD_LINKNAME) {
501 char *tmp = lustre_msg_buf(req->rq_repmsg, 1);
503 rc = inode->i_op->readlink(de, tmp, size[1]);
506 CERROR("readlink failed: %d\n", rc);
511 body = lustre_msg_buf(req->rq_repmsg, 0);
512 body->ino = inode->i_ino;
513 body->generation = inode->i_generation;
514 body->atime = inode->i_atime;
515 body->ctime = inode->i_ctime;
516 body->mtime = inode->i_mtime;
517 body->uid = inode->i_uid;
518 body->gid = inode->i_gid;
519 body->size = inode->i_size;
520 body->mode = inode->i_mode;
521 body->nlink = inode->i_nlink;
522 body->valid = (OBD_MD_FLID | OBD_MD_FLATIME | OBD_MD_FLCTIME |
523 OBD_MD_FLMTIME | OBD_MD_FLSIZE | OBD_MD_FLUID |
524 OBD_MD_FLGID | OBD_MD_FLNLINK | OBD_MD_FLGENER |
527 if (S_ISREG(inode->i_mode)) {
528 rc = mds_fs_get_md(mds, inode,
529 lustre_msg_buf(req->rq_repmsg, 1));
531 CERROR("mds_fs_get_md failed: %d\n", rc);
534 body->valid |= OBD_MD_FLEASIZE;
535 } else if (S_ISLNK(inode->i_mode))
536 body->valid |= OBD_MD_LINKNAME;
546 static int mds_statfs(struct ptlrpc_request *req)
548 struct mds_obd *mds = mds_req2mds(req);
549 struct obd_statfs *osfs;
551 int rc, size = sizeof(*osfs);
554 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
556 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
557 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
561 rc = vfs_statfs(mds->mds_sb, &sfs);
563 CERROR("mds: statfs failed: rc %d\n", rc);
566 osfs = lustre_msg_buf(req->rq_repmsg, 0);
567 memset(osfs, 0, size);
568 obd_statfs_pack(osfs, &sfs);
575 static int mds_open(struct ptlrpc_request *req)
578 struct mds_body *body;
580 struct vfsmount *mnt;
581 struct mds_obd *mds = mds_req2mds(req);
582 struct mds_export_data *med;
584 struct list_head *tmp;
585 struct mds_file_data *mfd;
586 int rc, size = sizeof(*body);
589 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
590 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
591 CERROR("mds: out of memory\n");
592 req->rq_status = -ENOMEM;
596 body = lustre_msg_buf(req->rq_reqmsg, 0);
598 /* was this animal open already? */
599 /* XXX we should only check on re-open, or do a refcount... */
600 med = &req->rq_export->exp_mds_data;
601 list_for_each(tmp, &med->med_open_head) {
602 struct mds_file_data *fd;
603 fd = list_entry(tmp, struct mds_file_data, mfd_list);
604 if (body->extra == fd->mfd_clientfd &&
605 body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
606 CERROR("Re opening %Ld\n", body->fid1.id);
611 OBD_ALLOC(mfd, sizeof(*mfd));
613 CERROR("mds: out of memory\n");
614 req->rq_status = -ENOMEM;
618 de = mds_fid2dentry(mds, &body->fid1, &mnt);
620 req->rq_status = -ENOENT;
624 /* check if this inode has seen a delayed object creation */
625 if (req->rq_reqmsg->bufcount > 1) {
627 struct lov_stripe_md *md;
628 struct inode *inode = de->d_inode;
631 md = lustre_msg_buf(req->rq_reqmsg, 1);
633 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
635 req->rq_status = -ENOMEM;
639 /* XXX error handling */
640 rc = mds_fs_set_md(mds, inode, handle, md);
642 struct obd_run_ctxt saved;
643 push_ctxt(&saved, &mds->mds_ctxt);
644 rc = mds_update_last_rcvd(mds, handle, req);
650 /* FIXME: need to return last_rcvd, last_committed */
652 /* FIXME: keep rc intact */
653 rc = mds_fs_commit(mds, de->d_inode, handle);
661 file = dentry_open(de, mnt, flags & ~O_DIRECT);
662 if (!file || IS_ERR(file)) {
663 req->rq_status = -EINVAL;
664 OBD_FREE(mfd, sizeof(*mfd));
668 file->private_data = mfd;
669 mfd->mfd_file = file;
670 mfd->mfd_clientfd = body->extra;
671 list_add(&mfd->mfd_list, &med->med_open_head);
673 body = lustre_msg_buf(req->rq_repmsg, 0);
674 body->extra = (__u64) (unsigned long)file;
678 static int mds_close(struct ptlrpc_request *req)
681 struct mds_body *body;
683 struct mds_obd *mds = mds_req2mds(req);
684 struct vfsmount *mnt;
685 struct mds_file_data *mfd;
689 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
690 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
691 CERROR("mds: out of memory\n");
692 req->rq_status = -ENOMEM;
696 body = lustre_msg_buf(req->rq_reqmsg, 0);
697 de = mds_fid2dentry(mds, &body->fid1, &mnt);
699 req->rq_status = -ENOENT;
703 file = (struct file *)(unsigned long)body->extra;
706 mfd = (struct mds_file_data *)file->private_data;
707 list_del(&mfd->mfd_list);
708 OBD_FREE(mfd, sizeof(*mfd));
710 req->rq_status = filp_close(file, 0);
717 static int mds_readpage(struct ptlrpc_request *req)
719 struct mds_obd *mds = mds_req2mds(req);
720 struct vfsmount *mnt;
723 struct mds_body *body;
724 struct obd_run_ctxt saved;
725 int rc, size = sizeof(*body);
728 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
729 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
730 CERROR("mds: out of memory\n");
731 GOTO(out, rc = -ENOMEM);
734 body = lustre_msg_buf(req->rq_reqmsg, 0);
735 push_ctxt(&saved, &mds->mds_ctxt);
736 de = mds_fid2dentry(mds, &body->fid1, &mnt);
738 GOTO(out_pop, rc = PTR_ERR(de));
740 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
742 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
743 /* note: in case of an error, dentry_open puts dentry */
745 GOTO(out_pop, rc = PTR_ERR(file));
747 /* to make this asynchronous make sure that the handling function
748 doesn't send a reply when this function completes. Instead a
749 callback function would send the reply */
750 rc = mds_sendpage(req, file, body->size);
760 int mds_reint(int offset, struct ptlrpc_request *req)
763 struct mds_update_record rec;
765 rc = mds_update_unpack(req, offset, &rec);
766 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
767 CERROR("invalid record\n");
768 req->rq_status = -EINVAL;
771 /* rc will be used to interrupt a for loop over multiple records */
772 rc = mds_reint_rec(&rec, offset, req);
776 int mds_handle(struct ptlrpc_request *req)
781 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
782 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
783 CERROR("lustre_mds: Invalid request\n");
787 if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
788 CERROR("lustre_mds: wrong packet type sent %d\n",
789 req->rq_reqmsg->type);
790 GOTO(out, rc = -EINVAL);
793 if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
794 GOTO(out, rc = -ENOTCONN);
796 if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
797 GOTO(out, rc = -EINVAL);
799 switch (req->rq_reqmsg->opc) {
801 CDEBUG(D_INODE, "connect\n");
802 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
803 rc = target_handle_connect(req);
807 CDEBUG(D_INODE, "disconnect\n");
808 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
809 rc = target_handle_disconnect(req);
813 CDEBUG(D_INODE, "getstatus\n");
814 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
815 rc = mds_getstatus(req);
819 CDEBUG(D_INODE, "getlovinfo\n");
820 rc = mds_getlovinfo(req);
824 CDEBUG(D_INODE, "getattr\n");
825 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
826 rc = mds_getattr(0, req);
830 CDEBUG(D_INODE, "statfs\n");
831 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
832 rc = mds_statfs(req);
836 CDEBUG(D_INODE, "readpage\n");
837 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
838 rc = mds_readpage(req);
840 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
845 int size = sizeof(struct mds_body);
846 CDEBUG(D_INODE, "reint\n");
847 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
849 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
852 rc = req->rq_status = -ENOMEM;
855 rc = mds_reint(0, req);
856 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
861 CDEBUG(D_INODE, "open\n");
862 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
867 CDEBUG(D_INODE, "close\n");
868 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
873 CDEBUG(D_INODE, "enqueue\n");
874 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
875 rc = ldlm_handle_enqueue(req);
881 CDEBUG(D_INODE, "convert\n");
882 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
883 rc = ldlm_handle_convert(req);
889 CDEBUG(D_INODE, "cancel\n");
890 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
891 rc = ldlm_handle_cancel(req);
896 CDEBUG(D_INODE, "callback\n");
897 CERROR("callbacks should not happen on MDS\n");
899 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CALLBACK, 0);
904 rc = ptlrpc_error(req->rq_svc, req);
911 struct mds_obd *mds = mds_req2mds(req);
912 req->rq_repmsg->last_rcvd = HTON__u64(mds->mds_last_rcvd);
913 req->rq_repmsg->last_committed =
914 HTON__u64(mds->mds_last_committed);
915 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
916 (unsigned long long)mds->mds_last_rcvd,
917 (unsigned long long)mds->mds_last_committed,
918 cpu_to_le32(req->rq_xid));
921 /* Still not 100% sure whether we should reply with the server
922 * last_rcvd or that of this client. I'm not sure it even makes
923 * a difference on a per-client basis, because last_rcvd is global
924 * and we are not supposed to allow transactions while in recovery.
927 CERROR("mds: processing error %d\n", rc);
928 ptlrpc_error(req->rq_svc, req);
930 CDEBUG(D_NET, "sending reply\n");
931 ptlrpc_reply(req->rq_svc, req);
936 /* Update the server data on disk. This stores the new mount_count and
937 * also the last_rcvd value to disk. If we don't have a clean shutdown,
938 * then the server last_rcvd value may be less than that of the clients.
939 * This will alert us that we may need to do client recovery.
942 int mds_update_server_data(struct mds_obd *mds)
944 struct obd_run_ctxt saved;
945 struct mds_server_data *msd = mds->mds_server_data;
946 struct file *filp = mds->mds_rcvd_filp;
950 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
951 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
953 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
954 (unsigned long long)mds->mds_mount_count,
955 (unsigned long long)mds->mds_last_rcvd);
956 push_ctxt(&saved, &mds->mds_ctxt);
957 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
958 if (rc != sizeof(*msd)) {
959 CERROR("error writing MDS server data: rc = %d\n", rc);
964 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
967 CERROR("error flushing MDS server data: rc = %d\n", rc);
972 /* Do recovery actions for the MDS */
973 static int mds_recover(struct obd_device *obddev)
975 struct mds_obd *mds = &obddev->u.mds;
978 /* This happens at the end when recovery is complete */
979 ++mds->mds_mount_count;
980 rc = mds_update_server_data(mds);
985 #define MDS_NUM_THREADS 8
986 /* mount the file system (secretly) */
987 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
990 struct obd_ioctl_data* data = buf;
991 struct mds_obd *mds = &obddev->u.mds;
992 struct vfsmount *mnt;
997 #ifdef CONFIG_DEV_RDONLY
1000 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1001 GOTO(err_dec, rc = -EINVAL);
1003 mds->mds_fstype = strdup(data->ioc_inlbuf2);
1005 mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1008 CERROR("do_kern_mount failed: rc = %d\n", rc);
1009 GOTO(err_kfree, rc);
1012 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1014 GOTO(err_put, rc = -ENODEV);
1016 mds->mds_max_mdsize = sizeof(struct lov_stripe_md);
1017 rc = mds_fs_setup(obddev, mnt);
1019 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1023 mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
1024 MDC_REPLY_PORTAL, "self",mds_handle);
1025 if (!mds->mds_service) {
1026 CERROR("failed to start service\n");
1027 GOTO(err_fs, rc = -EINVAL);
1031 obddev->obd_namespace =
1032 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1033 if (obddev->obd_namespace == NULL) {
1035 mds_cleanup(obddev);
1039 for (i = 0; i < MDS_NUM_THREADS; i++) {
1041 sprintf(name, "lustre_MDS_%2d", i);
1042 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1044 CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1046 GOTO(err_thread, rc);
1050 rc = mds_recover(obddev);
1052 GOTO(err_thread, rc);
1059 ptlrpc_stop_all_threads(mds->mds_service);
1061 ptlrpc_unregister_service(mds->mds_service);
1063 mds_fs_cleanup(obddev);
1066 mntput(mds->mds_vfsmnt);
1070 kfree(mds->mds_fstype);
1076 static int mds_cleanup(struct obd_device * obddev)
1078 struct super_block *sb;
1079 struct mds_obd *mds = &obddev->u.mds;
1083 if ( !list_empty(&obddev->obd_exports) ) {
1084 CERROR("still has exports!\n");
1088 ptlrpc_stop_all_threads(mds->mds_service);
1089 ptlrpc_unregister_service(mds->mds_service);
1095 mds_update_server_data(mds);
1097 if (mds->mds_rcvd_filp) {
1098 int rc = filp_close(mds->mds_rcvd_filp, 0);
1099 mds->mds_rcvd_filp = NULL;
1102 CERROR("last_rcvd file won't close, rc=%d\n", rc);
1106 mntput(mds->mds_vfsmnt);
1108 kfree(mds->mds_fstype);
1110 ldlm_namespace_free(obddev->obd_namespace);
1113 #ifdef CONFIG_DEV_RDONLY
1114 dev_clear_rdonly(2);
1116 mds_fs_cleanup(obddev);
1122 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1123 ldlm_mode_t mode, void *data)
1125 struct ptlrpc_request *req = req_cookie;
1132 if (req->rq_reqmsg->bufcount > 1) {
1133 /* an intent needs to be considered */
1134 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1135 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1136 struct mds_body *mds_rep;
1137 struct ldlm_reply *rep;
1138 __u64 new_resid[3] = {0, 0, 0}, old_res;
1139 int bufcount = -1, rc, size[3] = {sizeof(struct ldlm_reply),
1140 sizeof(struct mds_body),
1141 mds->mds_max_mdsize};
1143 it->opc = NTOH__u64(it->opc);
1145 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
1148 switch((long)it->opc) {
1150 /* Note that in the negative case you may be returning
1151 * a file and its obdo */
1153 case IT_CREAT|IT_OPEN:
1174 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
1177 rc = req->rq_status = -ENOMEM;
1181 rep = lustre_msg_buf(req->rq_repmsg, 0);
1182 rep->lock_policy_res1 = 1;
1184 /* execute policy */
1185 switch ((long)it->opc) {
1187 case IT_CREAT|IT_OPEN:
1195 rc = mds_reint(2, req);
1196 if (rc || req->rq_status != 0) {
1197 rep->lock_policy_res2 = req->rq_status;
1198 RETURN(ELDLM_LOCK_ABORTED);
1208 rc = mds_getattr_name(2, req);
1209 /* FIXME: we need to sit down and decide on who should
1210 * set req->rq_status, who should return negative and
1211 * positive return values, and what they all mean. */
1212 if (rc || req->rq_status != 0) {
1213 rep->lock_policy_res2 = req->rq_status;
1214 RETURN(ELDLM_LOCK_ABORTED);
1217 case IT_READDIR|IT_OPEN:
1221 CERROR("Unhandled intent\n");
1225 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR ||
1226 it->opc == IT_RENAME || it->opc == IT_RENAME2)
1227 RETURN(ELDLM_LOCK_ABORTED);
1229 rep->lock_policy_res2 = req->rq_status;
1230 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1231 new_resid[0] = NTOH__u32(mds_rep->ino);
1232 if (new_resid[0] == 0)
1234 old_res = lock->l_resource->lr_name[0];
1236 CDEBUG(D_INFO, "remote intent: locking %d instead of"
1237 "%ld\n", mds_rep->ino, (long)old_res);
1239 ldlm_lock_change_resource(lock, new_resid);
1240 if (lock->l_resource == NULL) {
1244 LDLM_DEBUG(lock, "intent policy, old res %ld",
1246 RETURN(ELDLM_LOCK_CHANGED);
1248 int size = sizeof(struct ldlm_reply);
1249 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1252 CERROR("out of memory\n");
1261 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1262 int len, void *karg, void *uarg);
1264 /* use obd ops to offer management infrastructure */
1265 static struct obd_ops mds_obd_ops = {
1266 o_connect: mds_connect,
1267 o_disconnect: mds_disconnect,
1269 o_cleanup: mds_cleanup,
1270 o_iocontrol: mds_iocontrol
1273 static int __init mds_init(void)
1275 class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1276 ldlm_register_intent(ldlm_intent_policy);
1280 static void __exit mds_exit(void)
1282 ldlm_unregister_intent();
1283 class_unregister_type(LUSTRE_MDS_NAME);
1286 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1287 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1288 MODULE_LICENSE("GPL");
1290 module_init(mds_init);
1291 module_exit(mds_exit);