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->b_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->b_xid = req->rq_xid;
89 bulk->b_buflen = PAGE_SIZE;
90 desc->b_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->b_waitq, desc->b_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 list_for_each(p, &obd->obd_exports) {
278 exp = list_entry(p, struct obd_export, exp_chain);
279 mcd = exp->exp_mds_data.med_mcd;
280 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
281 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
283 LASSERT(exp->exp_obd == obd);
285 exp->exp_rconnh.addr = conn->addr;
286 exp->exp_rconnh.cookie = conn->cookie;
287 conn->addr = (__u64) (unsigned long)exp;
288 conn->cookie = exp->exp_cookie;
289 CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
290 (long long)conn->addr, (long long)conn->cookie);
294 rc = class_connect(conn, obd, cluuid);
297 exp = class_conn2export(conn);
300 OBD_ALLOC(mcd, sizeof(*mcd));
302 CERROR("mds: out of memory for client data\n");
303 GOTO(out_export, rc = -ENOMEM);
305 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
306 exp->exp_mds_data.med_mcd = mcd;
307 rc = mds_client_add(&exp->exp_mds_data, -1);
314 OBD_FREE(mcd, sizeof(*mcd));
316 class_disconnect(conn);
323 static int mds_disconnect(struct lustre_handle *conn)
325 struct obd_export *exp;
328 exp = class_conn2export(conn);
332 rc = mds_client_free(&exp->exp_mds_data);
334 CERROR("error freeing client data: rc = %d\n", rc);
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 struct mds_export_data *med = &req->rq_export->exp_mds_data;
348 int rc, size = sizeof(*body);
351 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
352 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
353 CERROR("mds: out of memory for message: size=%d\n", size);
354 req->rq_status = -ENOMEM;
358 body = lustre_msg_buf(req->rq_reqmsg, 0);
359 mds_unpack_body(body);
361 /* Anything we need to do here with the client's trans no or so? */
362 body = lustre_msg_buf(req->rq_repmsg, 0);
363 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
365 LASSERT(med->med_mcd);
367 /* mcd_last_xid is is stored in little endian on the disk and
368 mds_pack_rep_body converts it to network order */
369 req->rq_repmsg->last_xid = le32_to_cpu(med->med_mcd->mcd_last_xid);
370 mds_pack_rep_body(req);
374 static int mds_getlovinfo(struct ptlrpc_request *req)
376 struct mds_obd *mds = mds_req2mds(req);
377 struct mds_status_req *streq;
378 struct lov_desc *desc;
380 int rc, size[2] = {sizeof(*desc)};
383 streq = lustre_msg_buf(req->rq_reqmsg, 0);
384 streq->flags = NTOH__u32(streq->flags);
385 streq->repbuf = NTOH__u32(streq->repbuf);
386 size[1] = streq->repbuf;
388 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
390 CERROR("mds: out of memory for message: size=%d\n", size[1]);
391 req->rq_status = -ENOMEM;
395 desc = lustre_msg_buf(req->rq_repmsg, 0);
396 rc = mds_get_lovdesc(req->rq_obd, desc);
398 CERROR("mds_get_lovdesc error %d", rc);
403 tgt_count = NTOH__u32(desc->ld_tgt_count);
404 if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
405 CERROR("too many targets, enlarge client buffers\n");
406 req->rq_status = -ENOSPC;
410 mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
411 tgt_count * sizeof(struct lov_object_id);
412 rc = mds_get_lovtgts(req->rq_obd, tgt_count,
413 lustre_msg_buf(req->rq_repmsg, 1));
415 CERROR("get_lovtgts error %d\n", rc);
422 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
423 void *data, __u32 data_len)
428 l_lock(&lock->l_resource->lr_namespace->ns_lock);
429 lock->l_flags |= LDLM_FL_CBPENDING;
430 do_ast = (!lock->l_readers && !lock->l_writers);
431 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
434 struct lustre_handle lockh;
437 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
438 ldlm_lock2handle(lock, &lockh);
439 rc = ldlm_cli_cancel(&lockh);
441 CERROR("ldlm_cli_cancel: %d\n", rc);
443 LDLM_DEBUG(lock, "Lock still has references, will be"
448 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
449 struct ptlrpc_request *req,
450 int request_off, int reply_off)
452 struct mds_body *request_body, *body;
453 struct inode *inode = dentry->d_inode;
460 /* Did the client request the link name? */
461 request_body = lustre_msg_buf(req->rq_reqmsg, request_off);
462 body = lustre_msg_buf(req->rq_repmsg, reply_off);
463 if ((body->valid & OBD_MD_LINKNAME) && S_ISLNK(inode->i_mode)) {
464 char *tmp = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
466 rc = inode->i_op->readlink(dentry, tmp, req->rq_repmsg->
467 buflens[reply_off + 1]);
469 CERROR("readlink failed: %d\n", rc);
473 body->valid |= OBD_MD_LINKNAME;
476 mds_pack_inode2fid(&body->fid1, inode);
477 mds_pack_inode2body(body, inode);
478 if (S_ISREG(inode->i_mode)) {
479 struct lov_mds_md *md;
481 md = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
482 md->lmd_easize = mds->mds_max_mdsize;
483 rc = mds_fs_get_md(mds, inode, md);
488 CERROR("mds_fs_get_md failed: %d\n", rc);
491 body->valid |= OBD_MD_FLEASIZE;
496 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
498 struct mds_obd *mds = mds_req2mds(req);
499 struct obd_device *obd = req->rq_export->exp_obd;
500 struct obd_run_ctxt saved;
501 struct mds_body *body;
502 struct dentry *de = NULL, *dchild = NULL;
504 struct lustre_handle lockh;
506 int namelen, flags, lock_mode, rc = 0, old_offset = offset;
507 __u64 res_id[3] = {0, 0, 0};
510 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
512 if (req->rq_reqmsg->bufcount <= offset + 1) {
514 GOTO(out_pre_de, rc = -EINVAL);
517 body = lustre_msg_buf(req->rq_reqmsg, offset);
518 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
519 namelen = req->rq_reqmsg->buflens[offset + 1];
520 /* requests were at offset 2, replies go back at 1 */
524 push_ctxt(&saved, &mds->mds_ctxt);
525 de = mds_fid2dentry(mds, &body->fid1, NULL);
528 GOTO(out_pre_de, rc = -ESTALE);
532 CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
534 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
535 res_id[0] = dir->i_ino;
537 rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
538 NULL, 0, lock_mode, &lockh);
540 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
541 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
542 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
543 &flags, ldlm_completion_ast,
544 mds_blocking_ast, NULL, 0, &lockh);
545 if (rc != ELDLM_OK) {
546 CERROR("lock enqueue: err: %d\n", rc);
547 GOTO(out_create_de, rc = -EIO);
550 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
553 dchild = lookup_one_len(name, de, namelen - 1);
554 if (IS_ERR(dchild)) {
555 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
558 GOTO(out_create_dchild, rc = -ESTALE);
561 rc = mds_getattr_internal(mds, dchild, req, old_offset, offset);
567 ldlm_lock_decref(&lockh, lock_mode);
576 static int mds_getattr(int offset, struct ptlrpc_request *req)
578 struct mds_obd *mds = mds_req2mds(req);
579 struct obd_run_ctxt saved;
582 struct mds_body *body;
583 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
586 body = lustre_msg_buf(req->rq_reqmsg, offset);
587 push_ctxt(&saved, &mds->mds_ctxt);
588 de = mds_fid2dentry(mds, &body->fid1, NULL);
590 req->rq_status = -ENOENT;
591 GOTO(out_pop, rc = -ENOENT);
595 if (S_ISREG(body->fid1.f_type)) {
597 size[1] = mds->mds_max_mdsize;
598 } else if (body->valid & OBD_MD_LINKNAME) {
600 size[1] = inode->i_size;
603 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
605 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
606 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
611 req->rq_status = mds_getattr_internal(mds, de, req, offset, 0);
620 static int mds_statfs(struct ptlrpc_request *req)
622 struct mds_obd *mds = mds_req2mds(req);
623 struct obd_statfs *osfs;
625 int rc, size = sizeof(*osfs);
628 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
630 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
631 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
635 rc = vfs_statfs(mds->mds_sb, &sfs);
637 CERROR("mds: statfs failed: rc %d\n", rc);
640 osfs = lustre_msg_buf(req->rq_repmsg, 0);
641 memset(osfs, 0, size);
642 obd_statfs_pack(osfs, &sfs);
649 static int mds_open(struct ptlrpc_request *req)
652 struct mds_body *body;
654 struct vfsmount *mnt;
655 struct mds_obd *mds = mds_req2mds(req);
656 struct mds_export_data *med;
658 struct list_head *tmp;
659 struct mds_file_data *mfd;
660 int rc, size = sizeof(*body);
663 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
664 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
665 CERROR("mds: out of memory\n");
666 req->rq_status = -ENOMEM;
670 body = lustre_msg_buf(req->rq_reqmsg, 0);
672 /* was this animal open already? */
673 /* XXX we should only check on re-open, or do a refcount... */
674 med = &req->rq_export->exp_mds_data;
675 list_for_each(tmp, &med->med_open_head) {
676 struct mds_file_data *fd;
677 fd = list_entry(tmp, struct mds_file_data, mfd_list);
678 if (body->extra == fd->mfd_clientfd &&
679 body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
680 CERROR("Re opening %Ld\n", body->fid1.id);
685 OBD_ALLOC(mfd, sizeof(*mfd));
687 CERROR("mds: out of memory\n");
688 req->rq_status = -ENOMEM;
692 de = mds_fid2dentry(mds, &body->fid1, &mnt);
694 req->rq_status = -ENOENT;
698 /* check if this inode has seen a delayed object creation */
699 if (req->rq_reqmsg->bufcount > 1) {
701 struct lov_mds_md *md;
702 struct inode *inode = de->d_inode;
705 md = lustre_msg_buf(req->rq_reqmsg, 1);
707 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
709 req->rq_status = -ENOMEM;
713 /* XXX error handling */
714 rc = mds_fs_set_md(mds, inode, handle, md);
716 struct obd_run_ctxt saved;
717 push_ctxt(&saved, &mds->mds_ctxt);
718 rc = mds_update_last_rcvd(mds, handle, req);
724 /* FIXME: need to return last_rcvd, last_committed */
726 /* FIXME: keep rc intact */
727 rc = mds_fs_commit(mds, de->d_inode, handle);
735 file = dentry_open(de, mnt, flags & ~O_DIRECT);
736 if (!file || IS_ERR(file)) {
737 req->rq_status = -EINVAL;
738 OBD_FREE(mfd, sizeof(*mfd));
742 file->private_data = mfd;
743 mfd->mfd_file = file;
744 mfd->mfd_clientfd = body->extra;
745 list_add(&mfd->mfd_list, &med->med_open_head);
747 body = lustre_msg_buf(req->rq_repmsg, 0);
748 /* FIXME: need to have cookies involved here */
749 body->extra = (__u64) (unsigned long)file;
753 static int mds_close(struct ptlrpc_request *req)
756 struct mds_body *body;
758 struct mds_obd *mds = mds_req2mds(req);
759 struct vfsmount *mnt;
760 struct mds_file_data *mfd;
764 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
765 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
766 CERROR("mds: out of memory\n");
767 req->rq_status = -ENOMEM;
771 body = lustre_msg_buf(req->rq_reqmsg, 0);
772 de = mds_fid2dentry(mds, &body->fid1, &mnt);
774 req->rq_status = -ENOENT;
778 /* FIXME: need to have cookies involved here */
779 file = (struct file *)(unsigned long)body->extra;
782 mfd = (struct mds_file_data *)file->private_data;
783 list_del(&mfd->mfd_list);
784 OBD_FREE(mfd, sizeof(*mfd));
786 req->rq_status = filp_close(file, 0);
793 static int mds_readpage(struct ptlrpc_request *req)
795 struct mds_obd *mds = mds_req2mds(req);
796 struct vfsmount *mnt;
799 struct mds_body *body;
800 struct obd_run_ctxt saved;
801 int rc, size = sizeof(*body);
804 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
805 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
806 CERROR("mds: out of memory\n");
807 GOTO(out, rc = -ENOMEM);
810 body = lustre_msg_buf(req->rq_reqmsg, 0);
811 push_ctxt(&saved, &mds->mds_ctxt);
812 de = mds_fid2dentry(mds, &body->fid1, &mnt);
814 GOTO(out_pop, rc = PTR_ERR(de));
816 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
818 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
819 /* note: in case of an error, dentry_open puts dentry */
821 GOTO(out_pop, rc = PTR_ERR(file));
823 /* to make this asynchronous make sure that the handling function
824 doesn't send a reply when this function completes. Instead a
825 callback function would send the reply */
826 rc = mds_sendpage(req, file, body->size);
836 int mds_reint(int offset, struct ptlrpc_request *req)
839 struct mds_update_record rec;
841 rc = mds_update_unpack(req, offset, &rec);
842 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
843 CERROR("invalid record\n");
844 req->rq_status = -EINVAL;
847 /* rc will be used to interrupt a for loop over multiple records */
848 rc = mds_reint_rec(&rec, offset, req);
852 int mds_handle(struct ptlrpc_request *req)
857 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
858 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
859 CERROR("lustre_mds: Invalid request\n");
863 if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
864 CERROR("lustre_mds: wrong packet type sent %d\n",
865 req->rq_reqmsg->type);
866 GOTO(out, rc = -EINVAL);
869 if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
870 GOTO(out, rc = -ENOTCONN);
872 if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
873 GOTO(out, rc = -EINVAL);
875 switch (req->rq_reqmsg->opc) {
877 CDEBUG(D_INODE, "connect\n");
878 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
879 rc = target_handle_connect(req);
883 CDEBUG(D_INODE, "disconnect\n");
884 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
885 rc = target_handle_disconnect(req);
889 CDEBUG(D_INODE, "getstatus\n");
890 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
891 rc = mds_getstatus(req);
895 CDEBUG(D_INODE, "getlovinfo\n");
896 rc = mds_getlovinfo(req);
900 CDEBUG(D_INODE, "getattr\n");
901 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
902 rc = mds_getattr(0, req);
906 CDEBUG(D_INODE, "statfs\n");
907 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
908 rc = mds_statfs(req);
912 CDEBUG(D_INODE, "readpage\n");
913 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
914 rc = mds_readpage(req);
916 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
921 int size = sizeof(struct mds_body);
922 CDEBUG(D_INODE, "reint\n");
923 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
925 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
931 rc = mds_reint(0, req);
932 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
937 CDEBUG(D_INODE, "open\n");
938 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
943 CDEBUG(D_INODE, "close\n");
944 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
949 CDEBUG(D_INODE, "enqueue\n");
950 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
951 rc = ldlm_handle_enqueue(req);
957 CDEBUG(D_INODE, "convert\n");
958 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
959 rc = ldlm_handle_convert(req);
965 CDEBUG(D_INODE, "cancel\n");
966 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
967 rc = ldlm_handle_cancel(req);
971 case LDLM_BL_CALLBACK:
972 case LDLM_CP_CALLBACK:
973 CDEBUG(D_INODE, "callback\n");
974 CERROR("callbacks should not happen on MDS\n");
976 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
979 rc = ptlrpc_error(req->rq_svc, req);
986 struct mds_obd *mds = mds_req2mds(req);
987 req->rq_repmsg->last_xid = HTON__u64(mds->mds_last_rcvd);
988 req->rq_repmsg->last_committed =
989 HTON__u64(mds->mds_last_committed);
990 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
991 (unsigned long long)mds->mds_last_rcvd,
992 (unsigned long long)mds->mds_last_committed,
993 cpu_to_le32(req->rq_xid));
996 /* Still not 100% sure whether we should reply with the server
997 * last_rcvd or that of this client. I'm not sure it even makes
998 * a difference on a per-client basis, because last_rcvd is global
999 * and we are not supposed to allow transactions while in recovery.
1002 CERROR("mds: processing error %d\n", rc);
1003 ptlrpc_error(req->rq_svc, req);
1005 CDEBUG(D_NET, "sending reply\n");
1006 ptlrpc_reply(req->rq_svc, req);
1011 /* Update the server data on disk. This stores the new mount_count and
1012 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1013 * then the server last_rcvd value may be less than that of the clients.
1014 * This will alert us that we may need to do client recovery.
1016 * Assumes we are already in the server filesystem context.
1019 int mds_update_server_data(struct mds_obd *mds)
1021 struct mds_server_data *msd = mds->mds_server_data;
1022 struct file *filp = mds->mds_rcvd_filp;
1026 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1027 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1029 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1030 (unsigned long long)mds->mds_mount_count,
1031 (unsigned long long)mds->mds_last_rcvd);
1032 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1033 if (rc != sizeof(*msd)) {
1034 CERROR("error writing MDS server data: rc = %d\n", rc);
1039 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1041 CERROR("error flushing MDS server data: rc = %d\n", rc);
1046 /* Do recovery actions for the MDS */
1047 static int mds_recover(struct obd_device *obddev)
1049 struct mds_obd *mds = &obddev->u.mds;
1050 struct obd_run_ctxt saved;
1053 /* This happens at the end when recovery is complete */
1054 ++mds->mds_mount_count;
1055 push_ctxt(&saved, &mds->mds_ctxt);
1056 rc = mds_update_server_data(mds);
1062 #define MDS_NUM_THREADS 8
1063 /* mount the file system (secretly) */
1064 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1067 struct obd_ioctl_data* data = buf;
1068 struct mds_obd *mds = &obddev->u.mds;
1069 struct vfsmount *mnt;
1074 #ifdef CONFIG_DEV_RDONLY
1075 dev_clear_rdonly(2);
1077 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1078 GOTO(err_dec, rc = -EINVAL);
1080 mds->mds_fstype = strdup(data->ioc_inlbuf2);
1082 mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1085 CERROR("do_kern_mount failed: rc = %d\n", rc);
1086 GOTO(err_kfree, rc);
1089 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1091 GOTO(err_put, rc = -ENODEV);
1093 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1094 rc = mds_fs_setup(obddev, mnt);
1096 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1100 mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
1101 MDC_REPLY_PORTAL, "self",mds_handle,
1103 if (!mds->mds_service) {
1104 CERROR("failed to start service\n");
1105 GOTO(err_fs, rc = -EINVAL);
1108 obddev->obd_namespace =
1109 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1110 if (obddev->obd_namespace == NULL) {
1111 mds_cleanup(obddev);
1112 GOTO(err_svc, rc = -ENOMEM);
1115 for (i = 0; i < MDS_NUM_THREADS; i++) {
1117 sprintf(name, "lustre_MDS_%02d", i);
1118 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1120 CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1121 GOTO(err_thread, rc);
1125 rc = mds_recover(obddev);
1127 GOTO(err_thread, rc);
1132 ptlrpc_stop_all_threads(mds->mds_service);
1134 ptlrpc_unregister_service(mds->mds_service);
1136 mds_fs_cleanup(obddev);
1139 mntput(mds->mds_vfsmnt);
1143 kfree(mds->mds_fstype);
1149 static int mds_cleanup(struct obd_device *obddev)
1151 struct super_block *sb;
1152 struct mds_obd *mds = &obddev->u.mds;
1153 struct obd_run_ctxt saved;
1156 ptlrpc_stop_all_threads(mds->mds_service);
1157 ptlrpc_unregister_service(mds->mds_service);
1163 push_ctxt(&saved, &mds->mds_ctxt);
1164 mds_update_server_data(mds);
1166 if (mds->mds_rcvd_filp) {
1167 int rc = filp_close(mds->mds_rcvd_filp, 0);
1168 mds->mds_rcvd_filp = NULL;
1171 CERROR("last_rcvd file won't close, rc=%d\n", rc);
1176 mntput(mds->mds_vfsmnt);
1178 kfree(mds->mds_fstype);
1180 ldlm_namespace_free(obddev->obd_namespace);
1183 #ifdef CONFIG_DEV_RDONLY
1184 dev_clear_rdonly(2);
1186 mds_fs_cleanup(obddev);
1192 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1193 ldlm_mode_t mode, void *data)
1195 struct ptlrpc_request *req = req_cookie;
1202 if (req->rq_reqmsg->bufcount > 1) {
1203 /* an intent needs to be considered */
1204 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1205 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1206 struct mds_body *mds_rep;
1207 struct ldlm_reply *rep;
1208 __u64 new_resid[3] = {0, 0, 0}, old_res;
1209 int rc, size[3] = {sizeof(struct ldlm_reply),
1210 sizeof(struct mds_body),
1211 mds->mds_max_mdsize};
1213 it->opc = NTOH__u64(it->opc);
1215 LDLM_DEBUG(lock, "intent policy, opc: %s",
1216 ldlm_it2str(it->opc));
1218 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1221 rc = req->rq_status = -ENOMEM;
1225 rep = lustre_msg_buf(req->rq_repmsg, 0);
1226 rep->lock_policy_res1 = 1;
1228 /* execute policy */
1229 switch ((long)it->opc) {
1230 case IT_CREAT|IT_OPEN:
1231 rc = mds_reint(2, req);
1232 if (rc || (req->rq_status != 0 &&
1233 req->rq_status != -EEXIST)) {
1234 rep->lock_policy_res2 = req->rq_status;
1235 RETURN(ELDLM_LOCK_ABORTED);
1246 rc = mds_reint(2, req);
1247 if (rc || (req->rq_status != 0 &&
1248 req->rq_status != -EISDIR &&
1249 req->rq_status != -ENOTDIR)) {
1250 rep->lock_policy_res2 = req->rq_status;
1251 RETURN(ELDLM_LOCK_ABORTED);
1261 rc = mds_getattr_name(2, req);
1262 /* FIXME: we need to sit down and decide on who should
1263 * set req->rq_status, who should return negative and
1264 * positive return values, and what they all mean. */
1265 if (rc || req->rq_status != 0) {
1266 rep->lock_policy_res2 = req->rq_status;
1267 RETURN(ELDLM_LOCK_ABORTED);
1270 case IT_READDIR|IT_OPEN:
1274 CERROR("Unhandled intent\n");
1278 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR ||
1279 it->opc == IT_RENAME || it->opc == IT_RENAME2)
1280 RETURN(ELDLM_LOCK_ABORTED);
1282 rep->lock_policy_res2 = req->rq_status;
1283 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1285 /* If the client is about to open a file that doesn't have an MD
1286 * stripe record, it's going to need a write lock. */
1287 if (it->opc & IT_OPEN) {
1288 struct lov_mds_md *md =
1289 lustre_msg_buf(req->rq_repmsg, 2);
1290 if (md->lmd_easize == 0) {
1291 LDLM_DEBUG(lock, "open with no EA; returning PW"
1293 lock->l_req_mode = LCK_PW;
1297 /* Give the client a lock on the child object, instead of the
1298 * parent that it requested. */
1299 new_resid[0] = NTOH__u32(mds_rep->ino);
1300 if (new_resid[0] == 0)
1302 old_res = lock->l_resource->lr_name[0];
1304 ldlm_lock_change_resource(lock, new_resid);
1305 if (lock->l_resource == NULL) {
1309 LDLM_DEBUG(lock, "intent policy, old res %ld",
1311 RETURN(ELDLM_LOCK_CHANGED);
1313 int size = sizeof(struct ldlm_reply);
1314 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1325 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1326 int len, void *karg, void *uarg);
1328 /* use obd ops to offer management infrastructure */
1329 static struct obd_ops mds_obd_ops = {
1330 o_connect: mds_connect,
1331 o_disconnect: mds_disconnect,
1333 o_cleanup: mds_cleanup,
1334 o_iocontrol: mds_iocontrol
1337 static int __init mds_init(void)
1339 class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1340 ldlm_register_intent(ldlm_intent_policy);
1344 static void __exit mds_exit(void)
1346 ldlm_unregister_intent();
1347 class_unregister_type(LUSTRE_MDS_NAME);
1350 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1351 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1352 MODULE_LICENSE("GPL");
1354 module_init(mds_init);
1355 module_exit(mds_exit);