1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
5 * Lustre Metadata Server (mds) request handler
7 * Copyright (c) 2001, 2002 Cluster File Systems, Inc.
8 * Author: Peter Braam <braam@clusterfs.com>
9 * Author: Andreas Dilger <adilger@clusterfs.com>
10 * Author: Phil Schwan <phil@clusterfs.com>
12 * This file is part of Lustre, http://www.lustre.org.
14 * Lustre is free software; you can redistribute it and/or
15 * modify it under the terms of version 2 of the GNU General Public
16 * License as published by the Free Software Foundation.
18 * Lustre is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
23 * You should have received a copy of the GNU General Public License
24 * along with Lustre; if not, write to the Free Software
25 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
29 #define DEBUG_SUBSYSTEM S_MDS
31 #include <linux/module.h>
32 #include <linux/lustre_mds.h>
33 #include <linux/lustre_dlm.h>
34 #include <linux/init.h>
35 #include <linux/obd_class.h>
37 extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
39 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
40 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
41 struct ptlrpc_request *req);
42 static int mds_cleanup(struct obd_device * obddev);
44 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
46 return &req->rq_export->exp_obd->u.mds;
49 static int mds_bulk_timeout(void *data)
51 struct ptlrpc_bulk_desc *desc = data;
54 CERROR("(not yet) starting recovery of client %p\n", desc->bd_client);
58 /* Assumes caller has already pushed into the kernel filesystem context */
59 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
63 struct mds_obd *mds = mds_req2mds(req);
64 struct ptlrpc_bulk_desc *desc;
65 struct ptlrpc_bulk_page *bulk;
66 struct l_wait_info lwi;
70 desc = ptlrpc_prep_bulk(req->rq_connection);
72 GOTO(out, rc = -ENOMEM);
74 bulk = ptlrpc_prep_bulk_page(desc);
76 GOTO(cleanup_bulk, rc = -ENOMEM);
78 OBD_ALLOC(buf, PAGE_SIZE);
80 GOTO(cleanup_bulk, rc = -ENOMEM);
82 rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
85 GOTO(cleanup_buf, rc = -EIO);
87 bulk->bp_xid = req->rq_xid;
89 bulk->bp_buflen = PAGE_SIZE;
90 desc->bd_portal = MDS_BULK_PORTAL;
92 rc = ptlrpc_send_bulk(desc);
94 GOTO(cleanup_buf, rc);
96 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
97 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
98 OBD_FAIL_MDS_SENDPAGE, rc);
99 ptlrpc_abort_bulk(desc);
100 GOTO(cleanup_buf, rc);
103 lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
104 rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_SENT, &lwi);
106 if (rc != -ETIMEDOUT)
108 GOTO(cleanup_buf, rc);
113 OBD_FREE(buf, PAGE_SIZE);
115 ptlrpc_free_bulk(desc);
121 * Look up a named entry in a directory, and get an LDLM lock on it.
122 * 'dir' is a inode for which an LDLM lock has already been taken.
124 * If we do not need an exclusive or write lock on this entry (e.g.
125 * a read lock for attribute lookup only) then we do not hold the
126 * directory on return. It is up to the caller to know what type
127 * of lock it is getting, and clean up appropriately.
129 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
130 struct dentry *dir, struct vfsmount **mnt,
131 char *name, int namelen, int lock_mode,
132 struct lustre_handle *lockh,
135 struct dentry *dchild;
137 __u64 res_id[3] = {0};
140 down(&dir->d_inode->i_sem);
141 dchild = lookup_one_len(name, dir, namelen);
142 if (IS_ERR(dchild)) {
143 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
144 up(&dir->d_inode->i_sem);
148 if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
149 up(&dir->d_inode->i_sem);
150 ldlm_lock_decref(lockh, dir_lock_mode);
153 if (lock_mode == 0 || !dchild->d_inode)
156 res_id[0] = dchild->d_inode->i_ino;
157 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
158 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
159 &flags, ldlm_completion_ast,
160 mds_blocking_ast, NULL, 0, lockh);
161 if (rc != ELDLM_OK) {
163 up(&dir->d_inode->i_sem);
164 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
170 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
171 struct vfsmount **mnt, int lock_mode,
172 struct lustre_handle *lockh)
174 struct mds_obd *mds = &obd->u.mds;
175 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
177 __u64 res_id[3] = {0};
183 res_id[0] = de->d_inode->i_ino;
184 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
185 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
186 &flags, ldlm_completion_ast,
187 mds_blocking_ast, NULL, 0, lockh);
188 if (rc != ELDLM_OK) {
190 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
196 /* Look up an entry by inode number. */
197 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
198 struct vfsmount **mnt)
200 /* stolen from NFS */
201 struct super_block *sb = mds->mds_sb;
202 unsigned long ino = fid->id;
203 __u32 generation = fid->generation;
205 struct list_head *lp;
206 struct dentry *result;
209 RETURN(ERR_PTR(-ESTALE));
211 inode = iget(sb, ino);
213 RETURN(ERR_PTR(-ENOMEM));
215 CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
217 if (is_bad_inode(inode) ||
218 (generation && inode->i_generation != generation)) {
219 /* we didn't find the right inode.. */
220 CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
221 inode->i_ino, inode->i_nlink,
222 atomic_read(&inode->i_count), inode->i_generation,
226 RETURN(ERR_PTR(-ESTALE));
229 /* now to find a dentry. If possible, get a well-connected one */
231 *mnt = mds->mds_vfsmnt;
232 spin_lock(&dcache_lock);
233 list_for_each(lp, &inode->i_dentry) {
234 result = list_entry(lp, struct dentry, d_alias);
235 if (!(result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
237 result->d_vfs_flags |= DCACHE_REFERENCED;
238 spin_unlock(&dcache_lock);
245 spin_unlock(&dcache_lock);
246 result = d_alloc_root(inode);
247 if (result == NULL) {
249 return ERR_PTR(-ENOMEM);
253 result->d_flags |= DCACHE_NFSD_DISCONNECTED;
257 /* Establish a connection to the MDS.
259 * This will set up an export structure for the client to hold state data
260 * about that client, like open files, the last operation number it did
261 * on the server, etc.
263 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
266 struct obd_export *exp;
267 struct mds_client_data *mcd;
272 if (!conn || !obd || !cluuid)
277 spin_lock(&obd->obd_dev_lock);
278 list_for_each(p, &obd->obd_exports) {
279 exp = list_entry(p, struct obd_export, exp_obd_chain);
280 mcd = exp->exp_mds_data.med_mcd;
281 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
282 LASSERT(exp->exp_obd == obd);
284 if (!list_empty(&exp->exp_conn_chain)) {
285 CERROR("existing uuid/export, list not empty!\n");
288 conn->addr = (__u64) (unsigned long)exp;
289 conn->cookie = exp->exp_cookie;
290 spin_unlock(&obd->obd_dev_lock);
291 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
293 CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
294 (long long)conn->addr, (long long)conn->cookie);
298 spin_unlock(&obd->obd_dev_lock);
299 /* XXX There is a small race between checking the list and adding a
300 * new connection for the same UUID, but the real threat (list
301 * corruption when multiple different clients connect) is solved.
303 rc = class_connect(conn, obd, cluuid);
306 exp = class_conn2export(conn);
309 OBD_ALLOC(mcd, sizeof(*mcd));
311 CERROR("mds: out of memory for client data\n");
312 GOTO(out_export, rc = -ENOMEM);
314 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
315 exp->exp_mds_data.med_mcd = mcd;
316 rc = mds_client_add(&exp->exp_mds_data, -1);
323 OBD_FREE(mcd, sizeof(*mcd));
325 class_disconnect(conn);
332 static int mds_disconnect(struct lustre_handle *conn)
336 rc = class_disconnect(conn);
343 static int mds_getstatus(struct ptlrpc_request *req)
345 struct mds_obd *mds = mds_req2mds(req);
346 struct mds_body *body;
347 int rc, size = sizeof(*body);
350 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
351 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
352 CERROR("mds: out of memory for message: size=%d\n", size);
353 req->rq_status = -ENOMEM;
357 /* Anything we need to do here with the client's trans no or so? */
358 body = lustre_msg_buf(req->rq_repmsg, 0);
359 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
361 /* the last_committed and last_xid fields are filled in for all
362 * replies already - no need to do so here also.
367 static int mds_getlovinfo(struct ptlrpc_request *req)
369 struct mds_obd *mds = mds_req2mds(req);
370 struct mds_status_req *streq;
371 struct lov_desc *desc;
373 int rc, size[2] = {sizeof(*desc)};
376 streq = lustre_msg_buf(req->rq_reqmsg, 0);
377 streq->flags = NTOH__u32(streq->flags);
378 streq->repbuf = NTOH__u32(streq->repbuf);
379 size[1] = streq->repbuf;
381 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
383 CERROR("mds: out of memory for message: size=%d\n", size[1]);
384 req->rq_status = -ENOMEM;
388 desc = lustre_msg_buf(req->rq_repmsg, 0);
389 rc = mds_get_lovdesc(req->rq_obd, desc);
391 CERROR("mds_get_lovdesc error %d", rc);
396 tgt_count = NTOH__u32(desc->ld_tgt_count);
397 if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
398 CERROR("too many targets, enlarge client buffers\n");
399 req->rq_status = -ENOSPC;
403 mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
404 tgt_count * sizeof(struct lov_object_id);
405 rc = mds_get_lovtgts(req->rq_obd, tgt_count,
406 lustre_msg_buf(req->rq_repmsg, 1));
408 CERROR("get_lovtgts error %d\n", rc);
415 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
416 void *data, __u32 data_len)
421 l_lock(&lock->l_resource->lr_namespace->ns_lock);
422 lock->l_flags |= LDLM_FL_CBPENDING;
423 do_ast = (!lock->l_readers && !lock->l_writers);
424 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
427 struct lustre_handle lockh;
430 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
431 ldlm_lock2handle(lock, &lockh);
432 rc = ldlm_cli_cancel(&lockh);
434 CERROR("ldlm_cli_cancel: %d\n", rc);
436 LDLM_DEBUG(lock, "Lock still has references, will be"
441 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
442 struct ptlrpc_request *req,
443 struct mds_body *reqbody, int reply_off)
445 struct mds_body *body;
446 struct inode *inode = dentry->d_inode;
453 body = lustre_msg_buf(req->rq_repmsg, reply_off);
455 mds_pack_inode2fid(&body->fid1, inode);
456 mds_pack_inode2body(body, inode);
458 if (S_ISREG(inode->i_mode)) {
459 struct lov_mds_md *md;
461 md = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
462 md->lmd_easize = mds->mds_max_mdsize;
463 rc = mds_fs_get_md(mds, inode, md);
468 CERROR("mds_fs_get_md failed: %d\n", rc);
471 body->valid |= OBD_MD_FLEASIZE;
472 } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
473 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
474 int len = req->rq_repmsg->buflens[reply_off + 1];
476 rc = inode->i_op->readlink(dentry, symname, len);
478 CERROR("readlink failed: %d\n", rc);
481 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
483 body->valid |= OBD_MD_LINKNAME;
488 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
490 struct mds_obd *mds = mds_req2mds(req);
491 struct obd_device *obd = req->rq_export->exp_obd;
492 struct obd_run_ctxt saved;
493 struct mds_body *body;
494 struct dentry *de = NULL, *dchild = NULL;
496 struct lustre_handle lockh;
498 int namelen, flags, lock_mode, rc = 0;
499 __u64 res_id[3] = {0, 0, 0};
502 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
504 if (req->rq_reqmsg->bufcount <= offset + 1) {
506 GOTO(out_pre_de, rc = -EINVAL);
509 body = lustre_msg_buf(req->rq_reqmsg, offset);
510 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
511 namelen = req->rq_reqmsg->buflens[offset + 1];
512 /* requests were at offset 2, replies go back at 1 */
516 push_ctxt(&saved, &mds->mds_ctxt);
517 de = mds_fid2dentry(mds, &body->fid1, NULL);
520 GOTO(out_pre_de, rc = -ESTALE);
524 CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
526 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
527 res_id[0] = dir->i_ino;
529 rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
530 NULL, 0, lock_mode, &lockh);
532 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
533 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
534 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
535 &flags, ldlm_completion_ast,
536 mds_blocking_ast, NULL, 0, &lockh);
537 if (rc != ELDLM_OK) {
538 CERROR("lock enqueue: err: %d\n", rc);
539 GOTO(out_create_de, rc = -EIO);
542 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
545 dchild = lookup_one_len(name, de, namelen - 1);
546 if (IS_ERR(dchild)) {
547 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
550 GOTO(out_create_dchild, rc = -ESTALE);
553 rc = mds_getattr_internal(mds, dchild, req, body, offset);
559 ldlm_lock_decref(&lockh, lock_mode);
568 static int mds_getattr(int offset, struct ptlrpc_request *req)
570 struct mds_obd *mds = mds_req2mds(req);
571 struct obd_run_ctxt saved;
574 struct mds_body *body;
575 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
578 body = lustre_msg_buf(req->rq_reqmsg, offset);
579 push_ctxt(&saved, &mds->mds_ctxt);
580 de = mds_fid2dentry(mds, &body->fid1, NULL);
582 req->rq_status = -ENOENT;
583 GOTO(out_pop, rc = -ENOENT);
587 if (S_ISREG(body->fid1.f_type)) {
589 size[1] = mds->mds_max_mdsize;
590 } else if (body->valid & OBD_MD_LINKNAME) {
592 size[1] = inode->i_size + 1;
595 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
597 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
598 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
603 req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
612 static int mds_statfs(struct ptlrpc_request *req)
614 struct mds_obd *mds = mds_req2mds(req);
615 struct obd_statfs *osfs;
617 int rc, size = sizeof(*osfs);
620 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
622 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
623 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
627 rc = vfs_statfs(mds->mds_sb, &sfs);
629 CERROR("mds: statfs failed: rc %d\n", rc);
632 osfs = lustre_msg_buf(req->rq_repmsg, 0);
633 memset(osfs, 0, size);
634 statfs_pack(osfs, &sfs);
635 obd_statfs_pack(osfs, osfs);
642 static int mds_open(struct ptlrpc_request *req)
645 struct mds_body *body;
647 struct vfsmount *mnt;
648 struct mds_obd *mds = mds_req2mds(req);
649 struct mds_export_data *med;
651 struct list_head *tmp;
652 struct mds_file_data *mfd;
653 int rc, size = sizeof(*body);
656 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
657 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
658 CERROR("mds: out of memory\n");
659 req->rq_status = -ENOMEM;
663 body = lustre_msg_buf(req->rq_reqmsg, 0);
665 /* was this animal open already? */
666 /* XXX we should only check on re-open, or do a refcount... */
667 med = &req->rq_export->exp_mds_data;
668 list_for_each(tmp, &med->med_open_head) {
669 struct mds_file_data *fd;
670 fd = list_entry(tmp, struct mds_file_data, mfd_list);
671 if (body->extra == fd->mfd_clientfd &&
672 body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
673 CERROR("Re opening "LPD64"\n", body->fid1.id);
678 OBD_ALLOC(mfd, sizeof(*mfd));
680 CERROR("mds: out of memory\n");
681 req->rq_status = -ENOMEM;
685 de = mds_fid2dentry(mds, &body->fid1, &mnt);
687 req->rq_status = -ENOENT;
691 /* check if this inode has seen a delayed object creation */
692 if (req->rq_reqmsg->bufcount > 1) {
694 struct lov_mds_md *md;
695 struct inode *inode = de->d_inode;
698 md = lustre_msg_buf(req->rq_reqmsg, 1);
700 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
702 req->rq_status = -ENOMEM;
706 /* XXX error handling */
707 rc = mds_fs_set_md(mds, inode, handle, md);
709 struct obd_run_ctxt saved;
710 push_ctxt(&saved, &mds->mds_ctxt);
711 rc = mds_update_last_rcvd(mds, handle, req);
717 /* FIXME: need to return last_rcvd, last_committed */
719 /* FIXME: keep rc intact */
720 rc = mds_fs_commit(mds, de->d_inode, handle);
728 file = dentry_open(de, mnt, flags & ~O_DIRECT);
729 if (!file || IS_ERR(file)) {
730 req->rq_status = -EINVAL;
731 OBD_FREE(mfd, sizeof(*mfd));
735 file->private_data = mfd;
736 mfd->mfd_file = file;
737 mfd->mfd_clientfd = body->extra;
738 list_add(&mfd->mfd_list, &med->med_open_head);
740 body = lustre_msg_buf(req->rq_repmsg, 0);
741 /* FIXME: need to have cookies involved here */
742 body->extra = (__u64) (unsigned long)file;
746 static int mds_close(struct ptlrpc_request *req)
749 struct mds_body *body;
751 struct mds_obd *mds = mds_req2mds(req);
752 struct vfsmount *mnt;
753 struct mds_file_data *mfd;
757 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
758 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
759 CERROR("mds: out of memory\n");
760 req->rq_status = -ENOMEM;
764 body = lustre_msg_buf(req->rq_reqmsg, 0);
765 de = mds_fid2dentry(mds, &body->fid1, &mnt);
767 req->rq_status = -ENOENT;
771 /* FIXME: need to have cookies involved here */
772 file = (struct file *)(unsigned long)body->extra;
775 mfd = (struct mds_file_data *)file->private_data;
776 list_del(&mfd->mfd_list);
777 OBD_FREE(mfd, sizeof(*mfd));
779 req->rq_status = filp_close(file, 0);
786 static int mds_readpage(struct ptlrpc_request *req)
788 struct mds_obd *mds = mds_req2mds(req);
789 struct vfsmount *mnt;
792 struct mds_body *body;
793 struct obd_run_ctxt saved;
794 int rc, size = sizeof(*body);
797 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
798 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
799 CERROR("mds: out of memory\n");
800 GOTO(out, rc = -ENOMEM);
803 body = lustre_msg_buf(req->rq_reqmsg, 0);
804 push_ctxt(&saved, &mds->mds_ctxt);
805 de = mds_fid2dentry(mds, &body->fid1, &mnt);
807 GOTO(out_pop, rc = PTR_ERR(de));
809 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
811 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
812 /* note: in case of an error, dentry_open puts dentry */
814 GOTO(out_pop, rc = PTR_ERR(file));
816 /* to make this asynchronous make sure that the handling function
817 doesn't send a reply when this function completes. Instead a
818 callback function would send the reply */
819 rc = mds_sendpage(req, file, body->size);
829 int mds_reint(int offset, struct ptlrpc_request *req)
832 struct mds_update_record rec;
834 rc = mds_update_unpack(req, offset, &rec);
835 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
836 CERROR("invalid record\n");
837 req->rq_status = -EINVAL;
840 /* rc will be used to interrupt a for loop over multiple records */
841 rc = mds_reint_rec(&rec, offset, req);
845 int mds_handle(struct ptlrpc_request *req)
850 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
851 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
852 CERROR("lustre_mds: Invalid request\n");
856 if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
857 GOTO(out, rc = -ENOTCONN);
859 if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
860 GOTO(out, rc = -EINVAL);
862 switch (req->rq_reqmsg->opc) {
864 CDEBUG(D_INODE, "connect\n");
865 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
866 rc = target_handle_connect(req);
870 CDEBUG(D_INODE, "disconnect\n");
871 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
872 rc = target_handle_disconnect(req);
876 CDEBUG(D_INODE, "getstatus\n");
877 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
878 rc = mds_getstatus(req);
882 CDEBUG(D_INODE, "getlovinfo\n");
883 rc = mds_getlovinfo(req);
887 CDEBUG(D_INODE, "getattr\n");
888 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
889 rc = mds_getattr(0, req);
893 CDEBUG(D_INODE, "statfs\n");
894 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
895 rc = mds_statfs(req);
899 CDEBUG(D_INODE, "readpage\n");
900 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
901 rc = mds_readpage(req);
903 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
908 int size = sizeof(struct mds_body);
909 CDEBUG(D_INODE, "reint\n");
910 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
912 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
918 rc = mds_reint(0, req);
919 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
924 CDEBUG(D_INODE, "open\n");
925 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
930 CDEBUG(D_INODE, "close\n");
931 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
936 CDEBUG(D_INODE, "enqueue\n");
937 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
938 rc = ldlm_handle_enqueue(req);
944 CDEBUG(D_INODE, "convert\n");
945 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
946 rc = ldlm_handle_convert(req);
952 CDEBUG(D_INODE, "cancel\n");
953 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
954 rc = ldlm_handle_cancel(req);
958 case LDLM_BL_CALLBACK:
959 case LDLM_CP_CALLBACK:
960 CDEBUG(D_INODE, "callback\n");
961 CERROR("callbacks should not happen on MDS\n");
963 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
966 rc = ptlrpc_error(req->rq_svc, req);
973 struct mds_export_data *med = &req->rq_export->exp_mds_data;
974 struct mds_obd *mds = mds_req2mds(req);
976 req->rq_repmsg->last_xid =
977 HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
978 req->rq_repmsg->last_committed =
979 HTON__u64(mds->mds_last_committed);
980 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
981 (unsigned long long)mds->mds_last_rcvd,
982 (unsigned long long)mds->mds_last_committed,
983 cpu_to_le32(req->rq_xid));
987 CERROR("mds: processing error %d\n", rc);
988 ptlrpc_error(req->rq_svc, req);
990 CDEBUG(D_NET, "sending reply\n");
991 ptlrpc_reply(req->rq_svc, req);
996 /* Update the server data on disk. This stores the new mount_count and
997 * also the last_rcvd value to disk. If we don't have a clean shutdown,
998 * then the server last_rcvd value may be less than that of the clients.
999 * This will alert us that we may need to do client recovery.
1001 * Assumes we are already in the server filesystem context.
1003 * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1006 int mds_update_server_data(struct mds_obd *mds)
1008 struct mds_server_data *msd = mds->mds_server_data;
1009 struct file *filp = mds->mds_rcvd_filp;
1013 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1014 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1016 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1017 (unsigned long long)mds->mds_mount_count,
1018 (unsigned long long)mds->mds_last_rcvd);
1019 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1020 if (rc != sizeof(*msd)) {
1021 CERROR("error writing MDS server data: rc = %d\n", rc);
1026 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1028 CERROR("error flushing MDS server data: rc = %d\n", rc);
1033 /* Do recovery actions for the MDS */
1034 static int mds_recover(struct obd_device *obddev)
1036 struct mds_obd *mds = &obddev->u.mds;
1037 struct obd_run_ctxt saved;
1040 /* This happens at the end when recovery is complete */
1041 ++mds->mds_mount_count;
1042 push_ctxt(&saved, &mds->mds_ctxt);
1043 rc = mds_update_server_data(mds);
1049 #define MDS_NUM_THREADS 8
1050 /* mount the file system (secretly) */
1051 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1054 struct obd_ioctl_data* data = buf;
1055 struct mds_obd *mds = &obddev->u.mds;
1056 struct vfsmount *mnt;
1061 #ifdef CONFIG_DEV_RDONLY
1062 dev_clear_rdonly(2);
1064 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1065 GOTO(err_dec, rc = -EINVAL);
1067 mds->mds_fstype = strdup(data->ioc_inlbuf2);
1069 mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1072 CERROR("do_kern_mount failed: rc = %d\n", rc);
1073 GOTO(err_kfree, rc);
1076 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1078 GOTO(err_put, rc = -ENODEV);
1080 spin_lock_init(&mds->mds_last_lock);
1081 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1082 rc = mds_fs_setup(obddev, mnt);
1084 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1088 mds->mds_service = ptlrpc_init_svc(1024, 640, MDS_REQUEST_PORTAL,
1089 MDC_REPLY_PORTAL, "self",mds_handle,
1091 if (!mds->mds_service) {
1092 CERROR("failed to start service\n");
1093 GOTO(err_fs, rc = -EINVAL);
1096 obddev->obd_namespace =
1097 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1098 if (obddev->obd_namespace == NULL) {
1099 mds_cleanup(obddev);
1100 GOTO(err_svc, rc = -ENOMEM);
1103 for (i = 0; i < MDS_NUM_THREADS; i++) {
1105 sprintf(name, "lustre_MDS_%02d", i);
1106 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1108 CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1109 GOTO(err_thread, rc);
1113 rc = mds_recover(obddev);
1115 GOTO(err_thread, rc);
1117 mds_destroy_export = mds_client_free;
1119 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1120 "mds_ldlm_client", &obddev->obd_ldlm_client);
1125 ptlrpc_stop_all_threads(mds->mds_service);
1127 ptlrpc_unregister_service(mds->mds_service);
1129 mds_fs_cleanup(obddev);
1132 mntput(mds->mds_vfsmnt);
1136 kfree(mds->mds_fstype);
1142 static int mds_cleanup(struct obd_device *obddev)
1144 struct super_block *sb;
1145 struct mds_obd *mds = &obddev->u.mds;
1146 struct obd_run_ctxt saved;
1149 ptlrpc_stop_all_threads(mds->mds_service);
1150 ptlrpc_unregister_service(mds->mds_service);
1156 push_ctxt(&saved, &mds->mds_ctxt);
1157 mds_update_server_data(mds);
1159 if (mds->mds_rcvd_filp) {
1160 int rc = filp_close(mds->mds_rcvd_filp, 0);
1161 mds->mds_rcvd_filp = NULL;
1164 CERROR("last_rcvd file won't close, rc=%d\n", rc);
1169 mntput(mds->mds_vfsmnt);
1171 kfree(mds->mds_fstype);
1173 ldlm_namespace_free(obddev->obd_namespace);
1176 #ifdef CONFIG_DEV_RDONLY
1177 dev_clear_rdonly(2);
1179 mds_fs_cleanup(obddev);
1185 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1186 ldlm_mode_t mode, void *data)
1188 struct ptlrpc_request *req = req_cookie;
1195 if (req->rq_reqmsg->bufcount > 1) {
1196 /* an intent needs to be considered */
1197 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1198 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1199 struct mds_body *mds_rep;
1200 struct ldlm_reply *rep;
1201 __u64 new_resid[3] = {0, 0, 0}, old_res;
1202 int rc, size[3] = {sizeof(struct ldlm_reply),
1203 sizeof(struct mds_body),
1204 mds->mds_max_mdsize};
1206 it->opc = NTOH__u64(it->opc);
1208 LDLM_DEBUG(lock, "intent policy, opc: %s",
1209 ldlm_it2str(it->opc));
1211 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1214 rc = req->rq_status = -ENOMEM;
1218 rep = lustre_msg_buf(req->rq_repmsg, 0);
1219 rep->lock_policy_res1 = 1;
1221 /* execute policy */
1222 switch ((long)it->opc) {
1223 case IT_CREAT|IT_OPEN:
1224 rc = mds_reint(2, req);
1225 if (rc || (req->rq_status != 0 &&
1226 req->rq_status != -EEXIST)) {
1227 rep->lock_policy_res2 = req->rq_status;
1228 RETURN(ELDLM_LOCK_ABORTED);
1239 rc = mds_reint(2, req);
1240 if (rc || (req->rq_status != 0 &&
1241 req->rq_status != -EISDIR &&
1242 req->rq_status != -ENOTDIR)) {
1243 rep->lock_policy_res2 = req->rq_status;
1244 RETURN(ELDLM_LOCK_ABORTED);
1254 rc = mds_getattr_name(2, req);
1255 /* FIXME: we need to sit down and decide on who should
1256 * set req->rq_status, who should return negative and
1257 * positive return values, and what they all mean. */
1258 if (rc || req->rq_status != 0) {
1259 rep->lock_policy_res2 = req->rq_status;
1260 RETURN(ELDLM_LOCK_ABORTED);
1263 case IT_READDIR|IT_OPEN:
1267 CERROR("Unhandled intent\n");
1271 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR ||
1272 it->opc == IT_RENAME || it->opc == IT_RENAME2)
1273 RETURN(ELDLM_LOCK_ABORTED);
1275 rep->lock_policy_res2 = req->rq_status;
1276 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1278 /* If the client is about to open a file that doesn't have an MD
1279 * stripe record, it's going to need a write lock. */
1280 if (it->opc & IT_OPEN) {
1281 struct lov_mds_md *md =
1282 lustre_msg_buf(req->rq_repmsg, 2);
1283 if (md->lmd_easize == 0) {
1284 LDLM_DEBUG(lock, "open with no EA; returning PW"
1286 lock->l_req_mode = LCK_PW;
1290 /* Give the client a lock on the child object, instead of the
1291 * parent that it requested. */
1292 new_resid[0] = NTOH__u32(mds_rep->ino);
1293 if (new_resid[0] == 0)
1295 old_res = lock->l_resource->lr_name[0];
1297 ldlm_lock_change_resource(lock, new_resid);
1298 if (lock->l_resource == NULL) {
1302 LDLM_DEBUG(lock, "intent policy, old res %ld",
1304 RETURN(ELDLM_LOCK_CHANGED);
1306 int size = sizeof(struct ldlm_reply);
1307 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1318 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1319 int len, void *karg, void *uarg);
1321 /* use obd ops to offer management infrastructure */
1322 static struct obd_ops mds_obd_ops = {
1323 o_connect: mds_connect,
1324 o_disconnect: mds_disconnect,
1326 o_cleanup: mds_cleanup,
1327 o_iocontrol: mds_iocontrol
1330 static int __init mds_init(void)
1332 class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1333 ldlm_register_intent(ldlm_intent_policy);
1337 static void __exit mds_exit(void)
1339 ldlm_unregister_intent();
1340 class_unregister_type(LUSTRE_MDS_NAME);
1343 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1344 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1345 MODULE_LICENSE("GPL");
1347 module_init(mds_init);
1348 module_exit(mds_exit);