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);
97 * Look up a named entry in a directory, and get an LDLM lock on it.
98 * 'dir' is a inode for which an LDLM lock has already been taken.
100 * If we do not need an exclusive or write lock on this entry (e.g.
101 * a read lock for attribute lookup only) then we do not hold the
102 * directory on return. It is up to the caller to know what type
103 * of lock it is getting, and clean up appropriately.
105 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
106 struct dentry *dir, struct vfsmount **mnt,
107 char *name, int namelen, int lock_mode,
108 struct lustre_handle *lockh,
111 struct dentry *dchild;
113 __u64 res_id[3] = {0};
116 down(&dir->d_inode->i_sem);
117 dchild = lookup_one_len(name, dir, namelen);
118 if (IS_ERR(dchild)) {
119 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
120 up(&dir->d_inode->i_sem);
124 if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
125 up(&dir->d_inode->i_sem);
126 ldlm_lock_decref(lockh, dir_lock_mode);
129 if (lock_mode == 0 || !dchild->d_inode)
132 res_id[0] = dchild->d_inode->i_ino;
133 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
134 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
135 &flags, ldlm_completion_ast,
136 mds_blocking_ast, NULL, 0, lockh);
137 if (rc != ELDLM_OK) {
139 up(&dir->d_inode->i_sem);
140 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
146 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
147 struct vfsmount **mnt, int lock_mode,
148 struct lustre_handle *lockh)
150 struct mds_obd *mds = &obd->u.mds;
151 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
153 __u64 res_id[3] = {0};
159 res_id[0] = de->d_inode->i_ino;
160 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
161 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
162 &flags, ldlm_completion_ast,
163 mds_blocking_ast, NULL, 0, lockh);
164 if (rc != ELDLM_OK) {
166 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
172 /* Look up an entry by inode number. */
173 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
174 struct vfsmount **mnt)
176 /* stolen from NFS */
177 struct super_block *sb = mds->mds_sb;
178 unsigned long ino = fid->id;
179 __u32 generation = fid->generation;
181 struct list_head *lp;
182 struct dentry *result;
185 RETURN(ERR_PTR(-ESTALE));
187 inode = iget(sb, ino);
189 RETURN(ERR_PTR(-ENOMEM));
191 CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
193 if (is_bad_inode(inode) ||
194 (generation && inode->i_generation != generation)) {
195 /* we didn't find the right inode.. */
196 CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
197 inode->i_ino, inode->i_nlink,
198 atomic_read(&inode->i_count), inode->i_generation,
202 RETURN(ERR_PTR(-ESTALE));
205 /* now to find a dentry. If possible, get a well-connected one */
207 *mnt = mds->mds_vfsmnt;
208 spin_lock(&dcache_lock);
209 for (lp = inode->i_dentry.next; lp != &inode->i_dentry ; lp=lp->next) {
210 result = list_entry(lp,struct dentry, d_alias);
211 if (! (result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
213 result->d_vfs_flags |= DCACHE_REFERENCED;
214 spin_unlock(&dcache_lock);
221 spin_unlock(&dcache_lock);
222 result = d_alloc_root(inode);
223 if (result == NULL) {
225 return ERR_PTR(-ENOMEM);
229 result->d_flags |= DCACHE_NFSD_DISCONNECTED;
233 /* Establish a connection to the MDS.
235 * This will set up an export structure for the client to hold state data
236 * about that client, like open files, the last operation number it did
237 * on the server, etc.
239 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
242 struct obd_export *exp;
243 struct mds_client_data *mcd;
249 list_for_each(p, &obd->obd_exports) {
250 exp = list_entry(p, struct obd_export, exp_chain);
251 mcd = exp->exp_mds_data.med_mcd;
252 if (mcd && !memcmp(cluuid, mcd->mcd_uuid,
253 sizeof(mcd->mcd_uuid))) {
255 "existing export for UUID '%s' at %p\n",
257 LASSERT(exp->exp_obd == obd);
258 exp->exp_rconnh.addr = conn->addr;
259 exp->exp_rconnh.cookie = conn->cookie;
260 conn->addr = (__u64) (unsigned long)exp;
261 conn->cookie = exp->exp_cookie;
262 CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
263 (long long)conn->addr,
264 (long long)conn->cookie);
269 #warning shaver: we might need a real cluuid here
270 rc = class_connect(conn, obd, NULL);
273 exp = class_conn2export(conn);
276 OBD_ALLOC(mcd, sizeof(*mcd));
278 CERROR("mds: out of memory for client data\n");
279 GOTO(out_export, rc = -ENOMEM);
281 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
282 exp->exp_mds_data.med_mcd = mcd;
283 rc = mds_client_add(&exp->exp_mds_data, -1);
290 OBD_FREE(mcd, sizeof(*mcd));
292 class_disconnect(conn);
299 static int mds_disconnect(struct lustre_handle *conn)
301 struct obd_export *exp;
304 exp = class_conn2export(conn);
308 rc = mds_client_free(&exp->exp_mds_data);
310 CERROR("error freeing client data: rc = %d\n", rc);
312 rc = class_disconnect(conn);
319 static int mds_getstatus(struct ptlrpc_request *req)
321 struct mds_obd *mds = mds_req2mds(req);
322 struct mds_body *body;
323 struct mds_export_data *med = &req->rq_export->exp_mds_data;
324 int rc, size = sizeof(*body);
327 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
328 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
329 CERROR("mds: out of memory for message: size=%d\n", size);
330 req->rq_status = -ENOMEM;
334 body = lustre_msg_buf(req->rq_reqmsg, 0);
335 mds_unpack_body(body);
337 /* Anything we need to do here with the client's trans no or so? */
338 body = lustre_msg_buf(req->rq_repmsg, 0);
339 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
341 LASSERT(med->med_mcd);
343 /* mcd_last_xid is is stored in little endian on the disk and
344 mds_pack_rep_body converts it to network order */
345 req->rq_repmsg->last_xid = le32_to_cpu(med->med_mcd->mcd_last_xid);
346 mds_pack_rep_body(req);
350 static int mds_getlovinfo(struct ptlrpc_request *req)
352 struct mds_obd *mds = mds_req2mds(req);
353 struct mds_status_req *streq;
354 struct lov_desc *desc;
356 int rc, size[2] = {sizeof(*desc)};
359 streq = lustre_msg_buf(req->rq_reqmsg, 0);
360 streq->flags = NTOH__u32(streq->flags);
361 streq->repbuf = NTOH__u32(streq->repbuf);
362 size[1] = streq->repbuf;
364 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
366 CERROR("mds: out of memory for message: size=%d\n", size[1]);
367 req->rq_status = -ENOMEM;
371 desc = lustre_msg_buf(req->rq_repmsg, 0);
372 rc = mds_get_lovdesc(req->rq_obd, desc);
374 CERROR("mds_get_lovdesc error %d", rc);
379 tgt_count = NTOH__u32(desc->ld_tgt_count);
380 if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
381 CERROR("too many targets, enlarge client buffers\n");
382 req->rq_status = -ENOSPC;
386 mds->mds_max_mdsize = sizeof(desc) + tgt_count * sizeof(uuid_t);
387 rc = mds_get_lovtgts(req->rq_obd, tgt_count,
388 lustre_msg_buf(req->rq_repmsg, 1));
390 CERROR("get_lovtgts error %d\n", rc);
397 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
398 void *data, __u32 data_len)
400 struct lustre_handle lockh;
403 ldlm_lock2handle(lock, &lockh);
404 if (ldlm_cli_cancel(&lockh) < 0)
409 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
411 struct mds_obd *mds = mds_req2mds(req);
412 struct obd_device *obd = req->rq_export->exp_obd;
413 struct obd_run_ctxt saved;
414 struct mds_body *body;
415 struct dentry *de = NULL, *dchild = NULL;
417 struct lustre_handle lockh;
419 int namelen, flags, lock_mode, rc = 0;
420 __u64 res_id[3] = {0, 0, 0};
423 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
425 if (req->rq_reqmsg->bufcount <= offset + 1) {
427 GOTO(out_pre_de, rc = -EINVAL);
430 body = lustre_msg_buf(req->rq_reqmsg, offset);
431 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
432 namelen = req->rq_reqmsg->buflens[offset + 1];
433 /* requests were at offset 2, replies go back at 1 */
437 push_ctxt(&saved, &mds->mds_ctxt);
438 de = mds_fid2dentry(mds, &body->fid1, NULL);
441 GOTO(out_pre_de, rc = -ESTALE);
445 CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
447 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
448 res_id[0] = dir->i_ino;
450 rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
451 NULL, 0, lock_mode, &lockh);
453 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
454 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
455 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
456 &flags, ldlm_completion_ast,
457 mds_blocking_ast, NULL, 0, &lockh);
458 if (rc != ELDLM_OK) {
459 CERROR("lock enqueue: err: %d\n", rc);
460 GOTO(out_create_de, rc = -EIO);
463 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
466 dchild = lookup_one_len(name, de, namelen - 1);
467 if (IS_ERR(dchild)) {
468 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
471 GOTO(out_create_dchild, rc = -ESTALE);
474 if (dchild->d_inode) {
475 struct mds_body *body;
476 struct inode *inode = dchild->d_inode;
477 CDEBUG(D_INODE, "child exists (dir %ld, name %s, ino %ld)\n",
478 dir->i_ino, name, dchild->d_inode->i_ino);
480 body = lustre_msg_buf(req->rq_repmsg, offset);
481 mds_pack_inode2fid(&body->fid1, inode);
482 mds_pack_inode2body(body, inode);
483 if (S_ISREG(inode->i_mode)) {
484 struct lov_stripe_md *md;
485 md = lustre_msg_buf(req->rq_repmsg, offset + 1);
486 md->lmd_easize = mds->mds_max_mdsize;
487 mds_fs_get_md(mds, inode, md);
489 /* now a normal case for intent locking */
498 ldlm_lock_decref(&lockh, lock_mode);
508 static int mds_getattr(int offset, struct ptlrpc_request *req)
510 struct mds_obd *mds = mds_req2mds(req);
511 struct obd_run_ctxt saved;
514 struct mds_body *body;
515 int rc, size[2] = {sizeof(*body)}, bufcount = 1;
518 body = lustre_msg_buf(req->rq_reqmsg, offset);
519 push_ctxt(&saved, &mds->mds_ctxt);
520 de = mds_fid2dentry(mds, &body->fid1, NULL);
522 GOTO(out_pop, rc = -ENOENT);
526 if (S_ISREG(body->fid1.f_type)) {
528 size[1] = mds->mds_max_mdsize;
529 } else if (body->valid & OBD_MD_LINKNAME) {
531 size[1] = inode->i_size;
534 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
536 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
537 CERROR("mds: out of memory\n");
541 if (body->valid & OBD_MD_LINKNAME) {
542 char *tmp = lustre_msg_buf(req->rq_repmsg, 1);
544 rc = inode->i_op->readlink(de, tmp, size[1]);
547 CERROR("readlink failed: %d\n", rc);
552 body = lustre_msg_buf(req->rq_repmsg, 0);
553 body->ino = inode->i_ino;
554 body->generation = inode->i_generation;
555 body->atime = inode->i_atime;
556 body->ctime = inode->i_ctime;
557 body->mtime = inode->i_mtime;
558 body->uid = inode->i_uid;
559 body->gid = inode->i_gid;
560 body->size = inode->i_size;
561 body->mode = inode->i_mode;
562 body->nlink = inode->i_nlink;
563 body->valid = (OBD_MD_FLID | OBD_MD_FLATIME | OBD_MD_FLCTIME |
564 OBD_MD_FLMTIME | OBD_MD_FLSIZE | OBD_MD_FLUID |
565 OBD_MD_FLGID | OBD_MD_FLNLINK | OBD_MD_FLGENER |
568 if (S_ISREG(inode->i_mode)) {
569 rc = mds_fs_get_md(mds, inode,
570 lustre_msg_buf(req->rq_repmsg, 1));
572 CERROR("mds_fs_get_md failed: %d\n", rc);
575 body->valid |= OBD_MD_FLEASIZE;
576 } else if (S_ISLNK(inode->i_mode))
577 body->valid |= OBD_MD_LINKNAME;
587 static int mds_statfs(struct ptlrpc_request *req)
589 struct mds_obd *mds = mds_req2mds(req);
590 struct obd_statfs *osfs;
592 int rc, size = sizeof(*osfs);
595 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
597 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
598 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
602 rc = vfs_statfs(mds->mds_sb, &sfs);
604 CERROR("mds: statfs failed: rc %d\n", rc);
607 osfs = lustre_msg_buf(req->rq_repmsg, 0);
608 memset(osfs, 0, size);
609 obd_statfs_pack(osfs, &sfs);
616 static int mds_open(struct ptlrpc_request *req)
619 struct mds_body *body;
621 struct vfsmount *mnt;
622 struct mds_obd *mds = mds_req2mds(req);
623 struct mds_export_data *med;
625 struct list_head *tmp;
626 struct mds_file_data *mfd;
627 int rc, size = sizeof(*body);
630 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
631 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
632 CERROR("mds: out of memory\n");
633 req->rq_status = -ENOMEM;
637 body = lustre_msg_buf(req->rq_reqmsg, 0);
639 /* was this animal open already? */
640 /* XXX we should only check on re-open, or do a refcount... */
641 med = &req->rq_export->exp_mds_data;
642 list_for_each(tmp, &med->med_open_head) {
643 struct mds_file_data *fd;
644 fd = list_entry(tmp, struct mds_file_data, mfd_list);
645 if (body->extra == fd->mfd_clientfd &&
646 body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
647 CERROR("Re opening %Ld\n", body->fid1.id);
652 OBD_ALLOC(mfd, sizeof(*mfd));
654 CERROR("mds: out of memory\n");
655 req->rq_status = -ENOMEM;
659 de = mds_fid2dentry(mds, &body->fid1, &mnt);
661 req->rq_status = -ENOENT;
665 /* check if this inode has seen a delayed object creation */
666 if (req->rq_reqmsg->bufcount > 1) {
668 struct lov_stripe_md *md;
669 struct inode *inode = de->d_inode;
672 md = lustre_msg_buf(req->rq_reqmsg, 1);
674 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
676 req->rq_status = -ENOMEM;
680 /* XXX error handling */
681 rc = mds_fs_set_md(mds, inode, handle, md);
683 struct obd_run_ctxt saved;
684 push_ctxt(&saved, &mds->mds_ctxt);
685 rc = mds_update_last_rcvd(mds, handle, req);
691 /* FIXME: need to return last_rcvd, last_committed */
693 /* FIXME: keep rc intact */
694 rc = mds_fs_commit(mds, de->d_inode, handle);
702 file = dentry_open(de, mnt, flags & ~O_DIRECT);
703 if (!file || IS_ERR(file)) {
704 req->rq_status = -EINVAL;
705 OBD_FREE(mfd, sizeof(*mfd));
709 file->private_data = mfd;
710 mfd->mfd_file = file;
711 mfd->mfd_clientfd = body->extra;
712 list_add(&mfd->mfd_list, &med->med_open_head);
714 body = lustre_msg_buf(req->rq_repmsg, 0);
715 /* FIXME: need to have cookies involved here */
716 body->extra = (__u64) (unsigned long)file;
720 static int mds_close(struct ptlrpc_request *req)
723 struct mds_body *body;
725 struct mds_obd *mds = mds_req2mds(req);
726 struct vfsmount *mnt;
727 struct mds_file_data *mfd;
731 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
732 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
733 CERROR("mds: out of memory\n");
734 req->rq_status = -ENOMEM;
738 body = lustre_msg_buf(req->rq_reqmsg, 0);
739 de = mds_fid2dentry(mds, &body->fid1, &mnt);
741 req->rq_status = -ENOENT;
745 /* FIXME: need to have cookies involved here */
746 file = (struct file *)(unsigned long)body->extra;
749 mfd = (struct mds_file_data *)file->private_data;
750 list_del(&mfd->mfd_list);
751 OBD_FREE(mfd, sizeof(*mfd));
753 req->rq_status = filp_close(file, 0);
760 static int mds_readpage(struct ptlrpc_request *req)
762 struct mds_obd *mds = mds_req2mds(req);
763 struct vfsmount *mnt;
766 struct mds_body *body;
767 struct obd_run_ctxt saved;
768 int rc, size = sizeof(*body);
771 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
772 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
773 CERROR("mds: out of memory\n");
774 GOTO(out, rc = -ENOMEM);
777 body = lustre_msg_buf(req->rq_reqmsg, 0);
778 push_ctxt(&saved, &mds->mds_ctxt);
779 de = mds_fid2dentry(mds, &body->fid1, &mnt);
781 GOTO(out_pop, rc = PTR_ERR(de));
783 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
785 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
786 /* note: in case of an error, dentry_open puts dentry */
788 GOTO(out_pop, rc = PTR_ERR(file));
790 /* to make this asynchronous make sure that the handling function
791 doesn't send a reply when this function completes. Instead a
792 callback function would send the reply */
793 rc = mds_sendpage(req, file, body->size);
803 int mds_reint(int offset, struct ptlrpc_request *req)
806 struct mds_update_record rec;
808 rc = mds_update_unpack(req, offset, &rec);
809 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
810 CERROR("invalid record\n");
811 req->rq_status = -EINVAL;
814 /* rc will be used to interrupt a for loop over multiple records */
815 rc = mds_reint_rec(&rec, offset, req);
819 int mds_handle(struct ptlrpc_request *req)
824 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
825 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
826 CERROR("lustre_mds: Invalid request\n");
830 if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
831 CERROR("lustre_mds: wrong packet type sent %d\n",
832 req->rq_reqmsg->type);
833 GOTO(out, rc = -EINVAL);
836 if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
837 GOTO(out, rc = -ENOTCONN);
839 if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
840 GOTO(out, rc = -EINVAL);
842 switch (req->rq_reqmsg->opc) {
844 CDEBUG(D_INODE, "connect\n");
845 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
846 rc = target_handle_connect(req);
850 CDEBUG(D_INODE, "disconnect\n");
851 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
852 rc = target_handle_disconnect(req);
856 CDEBUG(D_INODE, "getstatus\n");
857 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
858 rc = mds_getstatus(req);
862 CDEBUG(D_INODE, "getlovinfo\n");
863 rc = mds_getlovinfo(req);
867 CDEBUG(D_INODE, "getattr\n");
868 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
869 rc = mds_getattr(0, req);
873 CDEBUG(D_INODE, "statfs\n");
874 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
875 rc = mds_statfs(req);
879 CDEBUG(D_INODE, "readpage\n");
880 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
881 rc = mds_readpage(req);
883 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
888 int size = sizeof(struct mds_body);
889 CDEBUG(D_INODE, "reint\n");
890 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
892 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
898 rc = mds_reint(0, req);
899 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
904 CDEBUG(D_INODE, "open\n");
905 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
910 CDEBUG(D_INODE, "close\n");
911 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
916 CDEBUG(D_INODE, "enqueue\n");
917 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
918 rc = ldlm_handle_enqueue(req);
924 CDEBUG(D_INODE, "convert\n");
925 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
926 rc = ldlm_handle_convert(req);
932 CDEBUG(D_INODE, "cancel\n");
933 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
934 rc = ldlm_handle_cancel(req);
938 case LDLM_BL_CALLBACK:
939 case LDLM_CP_CALLBACK:
940 CDEBUG(D_INODE, "callback\n");
941 CERROR("callbacks should not happen on MDS\n");
943 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
946 rc = ptlrpc_error(req->rq_svc, req);
953 struct mds_obd *mds = mds_req2mds(req);
954 req->rq_repmsg->last_xid = HTON__u64(mds->mds_last_rcvd);
955 req->rq_repmsg->last_committed =
956 HTON__u64(mds->mds_last_committed);
957 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
958 (unsigned long long)mds->mds_last_rcvd,
959 (unsigned long long)mds->mds_last_committed,
960 cpu_to_le32(req->rq_xid));
963 /* Still not 100% sure whether we should reply with the server
964 * last_rcvd or that of this client. I'm not sure it even makes
965 * a difference on a per-client basis, because last_rcvd is global
966 * and we are not supposed to allow transactions while in recovery.
969 CERROR("mds: processing error %d\n", rc);
970 ptlrpc_error(req->rq_svc, req);
972 CDEBUG(D_NET, "sending reply\n");
973 ptlrpc_reply(req->rq_svc, req);
978 /* Update the server data on disk. This stores the new mount_count and
979 * also the last_rcvd value to disk. If we don't have a clean shutdown,
980 * then the server last_rcvd value may be less than that of the clients.
981 * This will alert us that we may need to do client recovery.
983 * Assumes we are already in the server filesystem context.
986 int mds_update_server_data(struct mds_obd *mds)
988 struct mds_server_data *msd = mds->mds_server_data;
989 struct file *filp = mds->mds_rcvd_filp;
993 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
994 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
996 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
997 (unsigned long long)mds->mds_mount_count,
998 (unsigned long long)mds->mds_last_rcvd);
999 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1000 if (rc != sizeof(*msd)) {
1001 CERROR("error writing MDS server data: rc = %d\n", rc);
1006 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1008 CERROR("error flushing MDS server data: rc = %d\n", rc);
1013 /* Do recovery actions for the MDS */
1014 static int mds_recover(struct obd_device *obddev)
1016 struct mds_obd *mds = &obddev->u.mds;
1017 struct obd_run_ctxt saved;
1020 /* This happens at the end when recovery is complete */
1021 ++mds->mds_mount_count;
1022 push_ctxt(&saved, &mds->mds_ctxt);
1023 rc = mds_update_server_data(mds);
1029 #define MDS_NUM_THREADS 8
1030 /* mount the file system (secretly) */
1031 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1034 struct obd_ioctl_data* data = buf;
1035 struct mds_obd *mds = &obddev->u.mds;
1036 struct vfsmount *mnt;
1041 #ifdef CONFIG_DEV_RDONLY
1042 dev_clear_rdonly(2);
1044 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1045 GOTO(err_dec, rc = -EINVAL);
1047 mds->mds_fstype = strdup(data->ioc_inlbuf2);
1049 mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1052 CERROR("do_kern_mount failed: rc = %d\n", rc);
1053 GOTO(err_kfree, rc);
1056 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1058 GOTO(err_put, rc = -ENODEV);
1060 mds->mds_max_mdsize = sizeof(struct lov_stripe_md);
1061 rc = mds_fs_setup(obddev, mnt);
1063 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1067 mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
1068 MDC_REPLY_PORTAL, "self",mds_handle);
1069 if (!mds->mds_service) {
1070 CERROR("failed to start service\n");
1071 GOTO(err_fs, rc = -EINVAL);
1074 obddev->obd_namespace =
1075 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1076 if (obddev->obd_namespace == NULL) {
1077 mds_cleanup(obddev);
1078 GOTO(err_svc, rc = -ENOMEM);
1081 for (i = 0; i < MDS_NUM_THREADS; i++) {
1083 sprintf(name, "lustre_MDS_%02d", i);
1084 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1086 CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1087 GOTO(err_thread, rc);
1091 rc = mds_recover(obddev);
1093 GOTO(err_thread, rc);
1098 ptlrpc_stop_all_threads(mds->mds_service);
1100 ptlrpc_unregister_service(mds->mds_service);
1102 mds_fs_cleanup(obddev);
1105 mntput(mds->mds_vfsmnt);
1109 kfree(mds->mds_fstype);
1115 static int mds_cleanup(struct obd_device *obddev)
1117 struct super_block *sb;
1118 struct mds_obd *mds = &obddev->u.mds;
1119 struct obd_run_ctxt saved;
1122 ptlrpc_stop_all_threads(mds->mds_service);
1123 ptlrpc_unregister_service(mds->mds_service);
1129 push_ctxt(&saved, &mds->mds_ctxt);
1130 mds_update_server_data(mds);
1132 if (mds->mds_rcvd_filp) {
1133 int rc = filp_close(mds->mds_rcvd_filp, 0);
1134 mds->mds_rcvd_filp = NULL;
1137 CERROR("last_rcvd file won't close, rc=%d\n", rc);
1142 mntput(mds->mds_vfsmnt);
1144 kfree(mds->mds_fstype);
1146 ldlm_namespace_free(obddev->obd_namespace);
1149 #ifdef CONFIG_DEV_RDONLY
1150 dev_clear_rdonly(2);
1152 mds_fs_cleanup(obddev);
1158 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1159 ldlm_mode_t mode, void *data)
1161 struct ptlrpc_request *req = req_cookie;
1168 if (req->rq_reqmsg->bufcount > 1) {
1169 /* an intent needs to be considered */
1170 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1171 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1172 struct mds_body *mds_rep;
1173 struct ldlm_reply *rep;
1174 __u64 new_resid[3] = {0, 0, 0}, old_res;
1175 int bufcount = -1, rc, size[3] = {sizeof(struct ldlm_reply),
1176 sizeof(struct mds_body),
1177 mds->mds_max_mdsize};
1179 it->opc = NTOH__u64(it->opc);
1181 LDLM_DEBUG(lock, "intent policy, opc: %s",
1182 ldlm_it2str(it->opc));
1185 switch((long)it->opc) {
1187 /* Note that in the negative case you may be returning
1188 * a file and its obdo */
1190 case IT_CREAT|IT_OPEN:
1211 rc = lustre_pack_msg(bufcount, 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 rep->lock_policy_res2 = req->rq_status;
1242 RETURN(ELDLM_LOCK_ABORTED);
1252 rc = mds_getattr_name(2, req);
1253 /* FIXME: we need to sit down and decide on who should
1254 * set req->rq_status, who should return negative and
1255 * positive return values, and what they all mean. */
1256 if (rc || req->rq_status != 0) {
1257 rep->lock_policy_res2 = req->rq_status;
1258 RETURN(ELDLM_LOCK_ABORTED);
1261 case IT_READDIR|IT_OPEN:
1265 CERROR("Unhandled intent\n");
1269 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR ||
1270 it->opc == IT_RENAME || it->opc == IT_RENAME2)
1271 RETURN(ELDLM_LOCK_ABORTED);
1273 rep->lock_policy_res2 = req->rq_status;
1274 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1275 new_resid[0] = NTOH__u32(mds_rep->ino);
1276 if (new_resid[0] == 0)
1278 old_res = lock->l_resource->lr_name[0];
1280 CDEBUG(D_INFO, "remote intent: locking %d instead of"
1281 "%ld\n", mds_rep->ino, (long)old_res);
1283 ldlm_lock_change_resource(lock, new_resid);
1284 if (lock->l_resource == NULL) {
1288 LDLM_DEBUG(lock, "intent policy, old res %ld",
1290 RETURN(ELDLM_LOCK_CHANGED);
1292 int size = sizeof(struct ldlm_reply);
1293 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1296 CERROR("out of memory\n");
1305 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1306 int len, void *karg, void *uarg);
1308 /* use obd ops to offer management infrastructure */
1309 static struct obd_ops mds_obd_ops = {
1310 o_connect: mds_connect,
1311 o_disconnect: mds_disconnect,
1313 o_cleanup: mds_cleanup,
1314 o_iocontrol: mds_iocontrol
1317 static int __init mds_init(void)
1319 class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1320 ldlm_register_intent(ldlm_intent_policy);
1324 static void __exit mds_exit(void)
1326 ldlm_unregister_intent();
1327 class_unregister_type(LUSTRE_MDS_NAME);
1330 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1331 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1332 MODULE_LICENSE("GPL");
1334 module_init(mds_init);
1335 module_exit(mds_exit);