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,
38 obd_uuid_t *uuidarray);
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");
286 spin_unlock(&obd->obd_dev_lock);
289 conn->addr = (__u64) (unsigned long)exp;
290 conn->cookie = exp->exp_cookie;
291 spin_unlock(&obd->obd_dev_lock);
292 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
294 CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
295 (long long)conn->addr, (long long)conn->cookie);
299 spin_unlock(&obd->obd_dev_lock);
300 /* XXX There is a small race between checking the list and adding a
301 * new connection for the same UUID, but the real threat (list
302 * corruption when multiple different clients connect) is solved.
304 rc = class_connect(conn, obd, cluuid);
307 exp = class_conn2export(conn);
310 OBD_ALLOC(mcd, sizeof(*mcd));
312 CERROR("mds: out of memory for client data\n");
313 GOTO(out_export, rc = -ENOMEM);
315 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
316 exp->exp_mds_data.med_mcd = mcd;
317 rc = mds_client_add(&exp->exp_mds_data, -1);
324 OBD_FREE(mcd, sizeof(*mcd));
326 class_disconnect(conn);
333 static int mds_disconnect(struct lustre_handle *conn)
336 struct obd_export *export = class_conn2export(conn);
338 ldlm_cancel_locks_for_export(export);
339 mds_client_free(export);
341 rc = class_disconnect(conn);
348 static int mds_getstatus(struct ptlrpc_request *req)
350 struct mds_obd *mds = mds_req2mds(req);
351 struct mds_body *body;
352 int rc, size = sizeof(*body);
355 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
356 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
357 CERROR("mds: out of memory for message: size=%d\n", size);
358 req->rq_status = -ENOMEM;
362 /* Anything we need to do here with the client's trans no or so? */
363 body = lustre_msg_buf(req->rq_repmsg, 0);
364 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
366 /* the last_committed and last_xid fields are filled in for all
367 * replies already - no need to do so here also.
372 static int mds_getlovinfo(struct ptlrpc_request *req)
374 struct mds_obd *mds = mds_req2mds(req);
375 struct mds_status_req *streq;
376 struct lov_desc *desc;
378 int rc, size[2] = {sizeof(*desc)};
381 streq = lustre_msg_buf(req->rq_reqmsg, 0);
382 streq->flags = NTOH__u32(streq->flags);
383 streq->repbuf = NTOH__u32(streq->repbuf);
384 size[1] = streq->repbuf;
386 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
388 CERROR("mds: out of memory for message: size=%d\n", size[1]);
389 req->rq_status = -ENOMEM;
393 desc = lustre_msg_buf(req->rq_repmsg, 0);
394 rc = mds_get_lovdesc(req->rq_obd, desc);
396 CERROR("mds_get_lovdesc error %d", rc);
401 tgt_count = le32_to_cpu(desc->ld_tgt_count);
402 if (tgt_count * sizeof(obd_uuid_t) > streq->repbuf) {
403 CERROR("too many targets, enlarge client buffers\n");
404 req->rq_status = -ENOSPC;
408 mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
409 tgt_count * sizeof(struct lov_object_id);
410 rc = mds_get_lovtgts(req->rq_obd, tgt_count,
411 lustre_msg_buf(req->rq_repmsg, 1));
413 CERROR("get_lovtgts error %d\n", rc);
420 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
421 void *data, __u32 data_len, int flag)
426 if (flag == LDLM_CB_CANCELING) {
427 /* Don't need to do anything here. */
431 l_lock(&lock->l_resource->lr_namespace->ns_lock);
432 lock->l_flags |= LDLM_FL_CBPENDING;
433 do_ast = (!lock->l_readers && !lock->l_writers);
434 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
437 struct lustre_handle lockh;
440 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
441 ldlm_lock2handle(lock, &lockh);
442 rc = ldlm_cli_cancel(&lockh);
444 CERROR("ldlm_cli_cancel: %d\n", rc);
446 LDLM_DEBUG(lock, "Lock still has references, will be"
451 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
452 struct ptlrpc_request *req,
453 struct mds_body *reqbody, int reply_off)
455 struct mds_body *body;
456 struct inode *inode = dentry->d_inode;
463 body = lustre_msg_buf(req->rq_repmsg, reply_off);
465 mds_pack_inode2fid(&body->fid1, inode);
466 mds_pack_inode2body(body, inode);
468 if (S_ISREG(inode->i_mode)) {
469 struct lov_mds_md *lmm;
471 lmm = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
472 lmm->lmm_easize = mds->mds_max_mdsize;
473 rc = mds_fs_get_md(mds, inode, lmm);
478 CERROR("mds_fs_get_md failed: %d\n", rc);
481 body->valid |= OBD_MD_FLEASIZE;
482 } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
483 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
484 int len = req->rq_repmsg->buflens[reply_off + 1];
486 rc = inode->i_op->readlink(dentry, symname, len);
488 CERROR("readlink failed: %d\n", rc);
491 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
493 body->valid |= OBD_MD_LINKNAME;
498 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
500 struct mds_obd *mds = mds_req2mds(req);
501 struct obd_device *obd = req->rq_export->exp_obd;
502 struct obd_run_ctxt saved;
503 struct mds_body *body;
504 struct dentry *de = NULL, *dchild = NULL;
506 struct lustre_handle lockh;
508 int namelen, flags, lock_mode, rc = 0;
510 __u64 res_id[3] = {0, 0, 0};
513 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
515 if (req->rq_reqmsg->bufcount <= offset + 1) {
517 GOTO(out_pre_de, rc = -EINVAL);
520 body = lustre_msg_buf(req->rq_reqmsg, offset);
521 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
522 namelen = req->rq_reqmsg->buflens[offset + 1];
523 /* requests were at offset 2, replies go back at 1 */
527 uc.ouc_fsuid = body->fsuid;
528 uc.ouc_fsgid = body->fsgid;
529 push_ctxt(&saved, &mds->mds_ctxt, &uc);
530 de = mds_fid2dentry(mds, &body->fid1, NULL);
533 GOTO(out_pre_de, rc = -ESTALE);
537 CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
539 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
540 res_id[0] = dir->i_ino;
542 rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
543 NULL, 0, lock_mode, &lockh);
545 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
546 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
547 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
548 &flags, ldlm_completion_ast,
549 mds_blocking_ast, NULL, 0, &lockh);
550 if (rc != ELDLM_OK) {
551 CERROR("lock enqueue: err: %d\n", rc);
552 GOTO(out_create_de, rc = -EIO);
555 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
558 dchild = lookup_one_len(name, de, namelen - 1);
559 if (IS_ERR(dchild)) {
560 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
563 GOTO(out_create_dchild, rc = -ESTALE);
566 rc = mds_getattr_internal(mds, dchild, req, body, offset);
572 ldlm_lock_decref(&lockh, lock_mode);
581 static int mds_getattr(int offset, struct ptlrpc_request *req)
583 struct mds_obd *mds = mds_req2mds(req);
584 struct obd_run_ctxt saved;
587 struct mds_body *body;
589 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
592 body = lustre_msg_buf(req->rq_reqmsg, offset);
593 uc.ouc_fsuid = body->fsuid;
594 uc.ouc_fsgid = body->fsgid;
595 push_ctxt(&saved, &mds->mds_ctxt, &uc);
596 de = mds_fid2dentry(mds, &body->fid1, NULL);
598 req->rq_status = -ENOENT;
600 GOTO(out_pop, PTR_ERR(de));
604 if (S_ISREG(body->fid1.f_type)) {
606 size[1] = mds->mds_max_mdsize;
607 } else if (body->valid & OBD_MD_LINKNAME) {
609 size[1] = MIN(inode->i_size + 1, body->size);
610 CDEBUG(D_INODE, "symlink size: %d, reply space: %d\n",
611 inode->i_size + 1, body->size);
614 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
615 CERROR("failed GETATTR_PACK test\n");
616 req->rq_status = -ENOMEM;
617 GOTO(out, rc = -ENOMEM);
620 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
623 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
628 req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
637 static int mds_statfs(struct ptlrpc_request *req)
639 struct mds_obd *mds = mds_req2mds(req);
640 struct obd_statfs *osfs;
642 int rc, size = sizeof(*osfs);
645 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
647 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
648 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
652 rc = vfs_statfs(mds->mds_sb, &sfs);
654 CERROR("mds: statfs failed: rc %d\n", rc);
657 osfs = lustre_msg_buf(req->rq_repmsg, 0);
658 memset(osfs, 0, size);
659 statfs_pack(osfs, &sfs);
660 obd_statfs_pack(osfs, osfs);
667 static int mds_open(struct ptlrpc_request *req)
670 struct mds_body *body;
672 struct vfsmount *mnt;
673 struct mds_obd *mds = mds_req2mds(req);
674 struct mds_export_data *med;
676 struct list_head *tmp;
677 struct mds_file_data *mfd;
678 int rc, size = sizeof(*body);
681 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
682 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
683 CERROR("mds: out of memory\n");
684 req->rq_status = -ENOMEM;
688 body = lustre_msg_buf(req->rq_reqmsg, 0);
690 /* was this animal open already? */
691 /* XXX we should only check on re-open, or do a refcount... */
692 med = &req->rq_export->exp_mds_data;
693 list_for_each(tmp, &med->med_open_head) {
694 struct mds_file_data *fd;
695 fd = list_entry(tmp, struct mds_file_data, mfd_list);
696 if (body->extra == fd->mfd_clientfd &&
697 body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
698 CERROR("Re opening "LPD64"\n", body->fid1.id);
703 OBD_ALLOC(mfd, sizeof(*mfd));
705 CERROR("mds: out of memory\n");
706 req->rq_status = -ENOMEM;
710 de = mds_fid2dentry(mds, &body->fid1, &mnt);
712 req->rq_status = -ENOENT;
716 /* check if this inode has seen a delayed object creation */
717 if (req->rq_reqmsg->bufcount > 1) {
719 struct lov_mds_md *lmm;
720 struct inode *inode = de->d_inode;
723 lmm = lustre_msg_buf(req->rq_reqmsg, 1);
725 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
727 req->rq_status = -ENOMEM;
731 /* XXX error handling */
732 rc = mds_fs_set_md(mds, inode, handle, lmm);
734 struct obd_run_ctxt saved;
736 uc.ouc_fsuid = body->fsuid;
737 uc.ouc_fsgid = body->fsgid;
738 push_ctxt(&saved, &mds->mds_ctxt, &uc);
739 rc = mds_update_last_rcvd(mds, handle, req);
745 /* FIXME: need to return last_rcvd, last_committed */
747 /* FIXME: keep rc intact */
748 rc = mds_fs_commit(mds, de->d_inode, handle);
756 file = dentry_open(de, mnt, flags & ~O_DIRECT);
757 if (!file || IS_ERR(file)) {
758 req->rq_status = -EINVAL;
759 OBD_FREE(mfd, sizeof(*mfd));
763 file->private_data = mfd;
764 mfd->mfd_file = file;
765 mfd->mfd_clientfd = body->extra;
766 list_add(&mfd->mfd_list, &med->med_open_head);
768 body = lustre_msg_buf(req->rq_repmsg, 0);
769 /* FIXME: need to have cookies involved here */
770 body->extra = (__u64) (unsigned long)file;
774 static int mds_close(struct ptlrpc_request *req)
777 struct mds_body *body;
779 struct mds_obd *mds = mds_req2mds(req);
780 struct vfsmount *mnt;
781 struct mds_file_data *mfd;
785 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
786 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
787 CERROR("mds: out of memory\n");
788 req->rq_status = -ENOMEM;
792 body = lustre_msg_buf(req->rq_reqmsg, 0);
793 de = mds_fid2dentry(mds, &body->fid1, &mnt);
795 req->rq_status = -ENOENT;
799 /* FIXME: need to have cookies involved here */
800 file = (struct file *)(unsigned long)body->extra;
803 mfd = (struct mds_file_data *)file->private_data;
804 list_del(&mfd->mfd_list);
805 OBD_FREE(mfd, sizeof(*mfd));
807 req->rq_status = filp_close(file, 0);
814 static int mds_readpage(struct ptlrpc_request *req)
816 struct mds_obd *mds = mds_req2mds(req);
817 struct vfsmount *mnt;
820 struct mds_body *body, *repbody;
821 struct obd_run_ctxt saved;
822 int rc, size = sizeof(*body);
826 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
827 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
828 CERROR("mds: out of memory\n");
829 GOTO(out, rc = -ENOMEM);
832 body = lustre_msg_buf(req->rq_reqmsg, 0);
833 uc.ouc_fsuid = body->fsuid;
834 uc.ouc_fsgid = body->fsgid;
835 push_ctxt(&saved, &mds->mds_ctxt, &uc);
836 de = mds_fid2dentry(mds, &body->fid1, &mnt);
838 GOTO(out_pop, rc = PTR_ERR(de));
840 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
842 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
843 /* note: in case of an error, dentry_open puts dentry */
845 GOTO(out_pop, rc = PTR_ERR(file));
847 repbody = lustre_msg_buf(req->rq_repmsg, 0);
848 repbody->size = file->f_dentry->d_inode->i_size;
849 repbody->valid = OBD_MD_FLSIZE;
851 /* to make this asynchronous make sure that the handling function
852 doesn't send a reply when this function completes. Instead a
853 callback function would send the reply */
854 /* note: in case of an error, dentry_open puts dentry */
855 rc = mds_sendpage(req, file, body->size);
865 int mds_reint(int offset, struct ptlrpc_request *req)
868 struct mds_update_record rec;
870 rc = mds_update_unpack(req, offset, &rec);
871 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
872 CERROR("invalid record\n");
873 req->rq_status = -EINVAL;
876 /* rc will be used to interrupt a for loop over multiple records */
877 rc = mds_reint_rec(&rec, offset, req);
881 int mds_handle(struct ptlrpc_request *req)
886 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
887 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
888 CERROR("lustre_mds: Invalid request\n");
892 if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
893 GOTO(out, rc = -ENOTCONN);
895 if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
896 GOTO(out, rc = -EINVAL);
898 switch (req->rq_reqmsg->opc) {
900 CDEBUG(D_INODE, "connect\n");
901 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
902 rc = target_handle_connect(req);
906 CDEBUG(D_INODE, "disconnect\n");
907 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
908 rc = target_handle_disconnect(req);
912 CDEBUG(D_INODE, "getstatus\n");
913 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
914 rc = mds_getstatus(req);
918 CDEBUG(D_INODE, "getlovinfo\n");
919 rc = mds_getlovinfo(req);
923 CDEBUG(D_INODE, "getattr\n");
924 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
925 rc = mds_getattr(0, req);
929 CDEBUG(D_INODE, "statfs\n");
930 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
931 rc = mds_statfs(req);
935 CDEBUG(D_INODE, "readpage\n");
936 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
937 rc = mds_readpage(req);
939 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
944 int size = sizeof(struct mds_body);
945 CDEBUG(D_INODE, "reint\n");
946 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
948 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
954 rc = mds_reint(0, req);
955 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
960 CDEBUG(D_INODE, "open\n");
961 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
966 CDEBUG(D_INODE, "close\n");
967 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
972 CDEBUG(D_INODE, "enqueue\n");
973 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
974 rc = ldlm_handle_enqueue(req);
980 CDEBUG(D_INODE, "convert\n");
981 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
982 rc = ldlm_handle_convert(req);
988 CDEBUG(D_INODE, "cancel\n");
989 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
990 rc = ldlm_handle_cancel(req);
994 case LDLM_BL_CALLBACK:
995 case LDLM_CP_CALLBACK:
996 CDEBUG(D_INODE, "callback\n");
997 CERROR("callbacks should not happen on MDS\n");
999 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1002 rc = ptlrpc_error(req->rq_svc, req);
1009 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1010 struct mds_obd *mds = mds_req2mds(req);
1012 req->rq_repmsg->last_xid =
1013 HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
1014 req->rq_repmsg->last_committed =
1015 HTON__u64(mds->mds_last_committed);
1016 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
1017 (unsigned long long)mds->mds_last_rcvd,
1018 (unsigned long long)mds->mds_last_committed,
1019 cpu_to_le32(req->rq_xid));
1023 CERROR("mds: processing error %d\n", rc);
1024 ptlrpc_error(req->rq_svc, req);
1026 CDEBUG(D_NET, "sending reply\n");
1027 ptlrpc_reply(req->rq_svc, req);
1032 /* Update the server data on disk. This stores the new mount_count and
1033 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1034 * then the server last_rcvd value may be less than that of the clients.
1035 * This will alert us that we may need to do client recovery.
1037 * Assumes we are already in the server filesystem context.
1039 * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1042 int mds_update_server_data(struct mds_obd *mds)
1044 struct mds_server_data *msd = mds->mds_server_data;
1045 struct file *filp = mds->mds_rcvd_filp;
1049 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1050 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1052 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1053 (unsigned long long)mds->mds_mount_count,
1054 (unsigned long long)mds->mds_last_rcvd);
1055 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1056 if (rc != sizeof(*msd)) {
1057 CERROR("error writing MDS server data: rc = %d\n", rc);
1062 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1064 CERROR("error flushing MDS server data: rc = %d\n", rc);
1069 /* Do recovery actions for the MDS */
1070 static int mds_recover(struct obd_device *obddev)
1072 struct mds_obd *mds = &obddev->u.mds;
1073 struct obd_run_ctxt saved;
1076 /* This happens at the end when recovery is complete */
1077 ++mds->mds_mount_count;
1078 push_ctxt(&saved, &mds->mds_ctxt, NULL);
1079 rc = mds_update_server_data(mds);
1085 #define MDS_NUM_THREADS 8
1086 /* mount the file system (secretly) */
1087 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1090 struct obd_ioctl_data* data = buf;
1091 struct mds_obd *mds = &obddev->u.mds;
1092 struct vfsmount *mnt;
1097 #ifdef CONFIG_DEV_RDONLY
1098 dev_clear_rdonly(2);
1100 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1101 GOTO(err_dec, rc = -EINVAL);
1103 mds->mds_fstype = strdup(data->ioc_inlbuf2);
1105 mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1108 CERROR("do_kern_mount failed: rc = %d\n", rc);
1109 GOTO(err_kfree, rc);
1112 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1114 GOTO(err_put, rc = -ENODEV);
1116 spin_lock_init(&mds->mds_last_lock);
1117 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1118 rc = mds_fs_setup(obddev, mnt);
1120 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1124 mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1125 MDS_BUFSIZE, MDS_MAXREQSIZE,
1126 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1127 "self", mds_handle, "mds");
1128 if (!mds->mds_service) {
1129 CERROR("failed to start service\n");
1130 GOTO(err_fs, rc = -EINVAL);
1133 obddev->obd_namespace =
1134 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1135 if (obddev->obd_namespace == NULL) {
1136 mds_cleanup(obddev);
1137 GOTO(err_svc, rc = -ENOMEM);
1140 for (i = 0; i < MDS_NUM_THREADS; i++) {
1142 sprintf(name, "lustre_MDS_%02d", i);
1143 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1145 CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1146 GOTO(err_thread, rc);
1150 rc = mds_recover(obddev);
1152 GOTO(err_thread, rc);
1154 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1155 "mds_ldlm_client", &obddev->obd_ldlm_client);
1160 ptlrpc_stop_all_threads(mds->mds_service);
1162 ptlrpc_unregister_service(mds->mds_service);
1164 mds_fs_cleanup(obddev);
1167 mntput(mds->mds_vfsmnt);
1171 kfree(mds->mds_fstype);
1177 static int mds_cleanup(struct obd_device *obddev)
1179 struct super_block *sb;
1180 struct mds_obd *mds = &obddev->u.mds;
1181 struct obd_run_ctxt saved;
1184 ptlrpc_stop_all_threads(mds->mds_service);
1185 ptlrpc_unregister_service(mds->mds_service);
1191 push_ctxt(&saved, &mds->mds_ctxt, NULL);
1192 mds_update_server_data(mds);
1194 if (mds->mds_rcvd_filp) {
1195 int rc = filp_close(mds->mds_rcvd_filp, 0);
1196 mds->mds_rcvd_filp = NULL;
1199 CERROR("last_rcvd file won't close, rc=%d\n", rc);
1204 mntput(mds->mds_vfsmnt);
1206 kfree(mds->mds_fstype);
1208 ldlm_namespace_free(obddev->obd_namespace);
1211 #ifdef CONFIG_DEV_RDONLY
1212 dev_clear_rdonly(2);
1214 mds_fs_cleanup(obddev);
1220 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1221 ldlm_mode_t mode, void *data)
1223 struct ptlrpc_request *req = req_cookie;
1230 if (req->rq_reqmsg->bufcount > 1) {
1231 /* an intent needs to be considered */
1232 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1233 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1234 struct mds_body *mds_rep;
1235 struct ldlm_reply *rep;
1236 __u64 new_resid[3] = {0, 0, 0}, old_res;
1237 int rc, size[3] = {sizeof(struct ldlm_reply),
1238 sizeof(struct mds_body),
1239 mds->mds_max_mdsize};
1241 it->opc = NTOH__u64(it->opc);
1243 LDLM_DEBUG(lock, "intent policy, opc: %s",
1244 ldlm_it2str(it->opc));
1246 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1249 rc = req->rq_status = -ENOMEM;
1253 rep = lustre_msg_buf(req->rq_repmsg, 0);
1254 rep->lock_policy_res1 = 1;
1256 /* execute policy */
1257 switch ((long)it->opc) {
1258 case IT_CREAT|IT_OPEN:
1259 rc = mds_reint(2, req);
1260 if (rc || (req->rq_status != 0 &&
1261 req->rq_status != -EEXIST)) {
1262 rep->lock_policy_res2 = req->rq_status;
1263 RETURN(ELDLM_LOCK_ABORTED);
1274 rc = mds_reint(2, req);
1275 if (rc || (req->rq_status != 0 &&
1276 req->rq_status != -EISDIR &&
1277 req->rq_status != -ENOTDIR)) {
1278 rep->lock_policy_res2 = req->rq_status;
1279 RETURN(ELDLM_LOCK_ABORTED);
1290 rc = mds_getattr_name(2, req);
1291 /* FIXME: we need to sit down and decide on who should
1292 * set req->rq_status, who should return negative and
1293 * positive return values, and what they all mean. */
1294 if (rc || req->rq_status != 0) {
1295 rep->lock_policy_res2 = req->rq_status;
1296 RETURN(ELDLM_LOCK_ABORTED);
1299 case IT_READDIR|IT_OPEN:
1303 CERROR("Unhandled intent\n");
1307 /* We don't bother returning a lock to the client for a file
1308 * or directory we are removing.
1310 * As for link and rename, there is no reason for the client
1311 * to get a lock on the target at this point. If they are
1312 * going to modify the file/directory later they will get a
1313 * lock at that time.
1315 if (it->opc & (IT_UNLINK | IT_RMDIR | IT_LINK | IT_LINK2 |
1316 IT_RENAME | IT_RENAME2))
1317 RETURN(ELDLM_LOCK_ABORTED);
1319 rep->lock_policy_res2 = req->rq_status;
1320 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1322 /* If the client is about to open a file that doesn't have an MD
1323 * stripe record, it's going to need a write lock. */
1324 if (it->opc & IT_OPEN) {
1325 struct lov_mds_md *lmm =
1326 lustre_msg_buf(req->rq_repmsg, 2);
1327 if (lmm->lmm_easize == 0) {
1328 LDLM_DEBUG(lock, "open with no EA; returning PW"
1330 lock->l_req_mode = LCK_PW;
1334 /* Give the client a lock on the child object, instead of the
1335 * parent that it requested. */
1336 new_resid[0] = NTOH__u32(mds_rep->ino);
1337 if (new_resid[0] == 0)
1339 old_res = lock->l_resource->lr_name[0];
1341 ldlm_lock_change_resource(lock, new_resid);
1342 if (lock->l_resource == NULL) {
1346 LDLM_DEBUG(lock, "intent policy, old res %ld",
1348 RETURN(ELDLM_LOCK_CHANGED);
1350 int size = sizeof(struct ldlm_reply);
1351 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1362 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1363 int len, void *karg, void *uarg);
1365 /* use obd ops to offer management infrastructure */
1366 static struct obd_ops mds_obd_ops = {
1367 o_connect: mds_connect,
1368 o_disconnect: mds_disconnect,
1370 o_cleanup: mds_cleanup,
1371 o_iocontrol: mds_iocontrol
1374 static int __init mds_init(void)
1376 class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1377 ldlm_register_intent(ldlm_intent_policy);
1381 static void __exit mds_exit(void)
1383 ldlm_unregister_intent();
1384 class_unregister_type(LUSTRE_MDS_NAME);
1387 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1388 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1389 MODULE_LICENSE("GPL");
1391 module_init(mds_init);
1392 module_exit(mds_exit);