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>
36 extern int mds_get_lovtgts(struct obd_device *obd, int tgt_count,
38 extern int mds_get_lovdesc(struct obd_device *obd, struct lov_desc *desc);
39 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
40 struct ptlrpc_request *req);
41 static int mds_cleanup(struct obd_device * obddev);
43 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
45 return &req->rq_export->exp_obd->u.mds;
48 static int mds_bulk_timeout(void *data)
50 struct ptlrpc_bulk_desc *desc = data;
53 CERROR("(not yet) starting recovery of client %p\n", desc->b_client);
57 /* Assumes caller has already pushed into the kernel filesystem context */
58 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
62 struct mds_obd *mds = mds_req2mds(req);
63 struct ptlrpc_bulk_desc *desc;
64 struct ptlrpc_bulk_page *bulk;
65 struct l_wait_info lwi;
69 desc = ptlrpc_prep_bulk(req->rq_connection);
71 GOTO(out, rc = -ENOMEM);
73 bulk = ptlrpc_prep_bulk_page(desc);
75 GOTO(cleanup_bulk, rc = -ENOMEM);
77 OBD_ALLOC(buf, PAGE_SIZE);
79 GOTO(cleanup_bulk, rc = -ENOMEM);
81 rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
84 GOTO(cleanup_buf, rc = -EIO);
86 bulk->b_xid = req->rq_xid;
88 bulk->b_buflen = PAGE_SIZE;
89 desc->b_portal = MDS_BULK_PORTAL;
91 rc = ptlrpc_send_bulk(desc);
93 GOTO(cleanup_buf, rc);
95 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
96 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
97 OBD_FAIL_MDS_SENDPAGE, rc);
98 ptlrpc_abort_bulk(desc);
99 GOTO(cleanup_buf, rc);
102 lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
103 rc = l_wait_event(desc->b_waitq, desc->b_flags & PTL_BULK_FL_SENT, &lwi);
105 if (rc != -ETIMEDOUT)
107 GOTO(cleanup_buf, rc);
112 OBD_FREE(buf, PAGE_SIZE);
114 ptlrpc_free_bulk(desc);
120 * Look up a named entry in a directory, and get an LDLM lock on it.
121 * 'dir' is a inode for which an LDLM lock has already been taken.
123 * If we do not need an exclusive or write lock on this entry (e.g.
124 * a read lock for attribute lookup only) then we do not hold the
125 * directory on return. It is up to the caller to know what type
126 * of lock it is getting, and clean up appropriately.
128 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
129 struct dentry *dir, struct vfsmount **mnt,
130 char *name, int namelen, int lock_mode,
131 struct lustre_handle *lockh,
134 struct dentry *dchild;
136 __u64 res_id[3] = {0};
139 down(&dir->d_inode->i_sem);
140 dchild = lookup_one_len(name, dir, namelen);
141 if (IS_ERR(dchild)) {
142 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
143 up(&dir->d_inode->i_sem);
147 if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
148 up(&dir->d_inode->i_sem);
149 ldlm_lock_decref(lockh, dir_lock_mode);
152 if (lock_mode == 0 || !dchild->d_inode)
155 res_id[0] = dchild->d_inode->i_ino;
156 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
157 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
158 &flags, ldlm_completion_ast,
159 mds_blocking_ast, NULL, 0, lockh);
160 if (rc != ELDLM_OK) {
162 up(&dir->d_inode->i_sem);
163 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
169 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
170 struct vfsmount **mnt, int lock_mode,
171 struct lustre_handle *lockh)
173 struct mds_obd *mds = &obd->u.mds;
174 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
176 __u64 res_id[3] = {0};
182 res_id[0] = de->d_inode->i_ino;
183 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
184 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
185 &flags, ldlm_completion_ast,
186 mds_blocking_ast, NULL, 0, lockh);
187 if (rc != ELDLM_OK) {
189 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
195 /* Look up an entry by inode number. */
196 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
197 struct vfsmount **mnt)
199 /* stolen from NFS */
200 struct super_block *sb = mds->mds_sb;
201 unsigned long ino = fid->id;
202 __u32 generation = fid->generation;
204 struct list_head *lp;
205 struct dentry *result;
208 RETURN(ERR_PTR(-ESTALE));
210 inode = iget(sb, ino);
212 RETURN(ERR_PTR(-ENOMEM));
214 CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
216 if (is_bad_inode(inode) ||
217 (generation && inode->i_generation != generation)) {
218 /* we didn't find the right inode.. */
219 CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
220 inode->i_ino, inode->i_nlink,
221 atomic_read(&inode->i_count), inode->i_generation,
225 RETURN(ERR_PTR(-ESTALE));
228 /* now to find a dentry. If possible, get a well-connected one */
230 *mnt = mds->mds_vfsmnt;
231 spin_lock(&dcache_lock);
232 list_for_each(lp, &inode->i_dentry) {
233 result = list_entry(lp, struct dentry, d_alias);
234 if (!(result->d_flags & DCACHE_NFSD_DISCONNECTED)) {
236 result->d_vfs_flags |= DCACHE_REFERENCED;
237 spin_unlock(&dcache_lock);
244 spin_unlock(&dcache_lock);
245 result = d_alloc_root(inode);
246 if (result == NULL) {
248 return ERR_PTR(-ENOMEM);
252 result->d_flags |= DCACHE_NFSD_DISCONNECTED;
256 /* Establish a connection to the MDS.
258 * This will set up an export structure for the client to hold state data
259 * about that client, like open files, the last operation number it did
260 * on the server, etc.
262 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
265 struct obd_export *exp;
266 struct mds_client_data *mcd;
271 if (!conn || !obd || !cluuid)
276 list_for_each(p, &obd->obd_exports) {
277 exp = list_entry(p, struct obd_export, exp_chain);
278 mcd = exp->exp_mds_data.med_mcd;
279 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
280 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
282 LASSERT(exp->exp_obd == obd);
284 exp->exp_rconnh.addr = conn->addr;
285 exp->exp_rconnh.cookie = conn->cookie;
286 conn->addr = (__u64) (unsigned long)exp;
287 conn->cookie = exp->exp_cookie;
288 CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
289 (long long)conn->addr, (long long)conn->cookie);
293 rc = class_connect(conn, obd, cluuid);
296 exp = class_conn2export(conn);
299 OBD_ALLOC(mcd, sizeof(*mcd));
301 CERROR("mds: out of memory for client data\n");
302 GOTO(out_export, rc = -ENOMEM);
304 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
305 exp->exp_mds_data.med_mcd = mcd;
306 rc = mds_client_add(&exp->exp_mds_data, -1);
313 OBD_FREE(mcd, sizeof(*mcd));
315 class_disconnect(conn);
322 static int mds_disconnect(struct lustre_handle *conn)
324 struct obd_export *exp;
327 exp = class_conn2export(conn);
331 rc = mds_client_free(&exp->exp_mds_data);
333 CERROR("error freeing client data: rc = %d\n", rc);
335 rc = class_disconnect(conn);
342 static int mds_getstatus(struct ptlrpc_request *req)
344 struct mds_obd *mds = mds_req2mds(req);
345 struct mds_body *body;
346 struct mds_export_data *med = &req->rq_export->exp_mds_data;
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 body = lustre_msg_buf(req->rq_reqmsg, 0);
358 mds_unpack_body(body);
360 /* Anything we need to do here with the client's trans no or so? */
361 body = lustre_msg_buf(req->rq_repmsg, 0);
362 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
364 LASSERT(med->med_mcd);
366 /* mcd_last_xid is is stored in little endian on the disk and
367 mds_pack_rep_body converts it to network order */
368 req->rq_repmsg->last_xid = le32_to_cpu(med->med_mcd->mcd_last_xid);
369 mds_pack_rep_body(req);
373 static int mds_getlovinfo(struct ptlrpc_request *req)
375 struct mds_obd *mds = mds_req2mds(req);
376 struct mds_status_req *streq;
377 struct lov_desc *desc;
379 int rc, size[2] = {sizeof(*desc)};
382 streq = lustre_msg_buf(req->rq_reqmsg, 0);
383 streq->flags = NTOH__u32(streq->flags);
384 streq->repbuf = NTOH__u32(streq->repbuf);
385 size[1] = streq->repbuf;
387 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
389 CERROR("mds: out of memory for message: size=%d\n", size[1]);
390 req->rq_status = -ENOMEM;
394 desc = lustre_msg_buf(req->rq_repmsg, 0);
395 rc = mds_get_lovdesc(req->rq_obd, desc);
397 CERROR("mds_get_lovdesc error %d", rc);
402 tgt_count = NTOH__u32(desc->ld_tgt_count);
403 if (tgt_count * sizeof(uuid_t) > streq->repbuf) {
404 CERROR("too many targets, enlarge client buffers\n");
405 req->rq_status = -ENOSPC;
409 mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
410 tgt_count * sizeof(struct lov_object_id);
411 rc = mds_get_lovtgts(req->rq_obd, tgt_count,
412 lustre_msg_buf(req->rq_repmsg, 1));
414 CERROR("get_lovtgts error %d\n", rc);
421 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
422 void *data, __u32 data_len)
427 l_lock(&lock->l_resource->lr_namespace->ns_lock);
428 lock->l_flags |= LDLM_FL_CBPENDING;
429 do_ast = (!lock->l_readers && !lock->l_writers);
430 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
433 struct lustre_handle lockh;
436 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
437 ldlm_lock2handle(lock, &lockh);
438 rc = ldlm_cli_cancel(&lockh);
440 CERROR("ldlm_cli_cancel: %d\n", rc);
442 LDLM_DEBUG(lock, "Lock still has references, will be"
447 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
448 struct ptlrpc_request *req,
449 int request_off, int reply_off)
451 struct mds_body *request_body, *body;
452 struct inode *inode = dentry->d_inode;
459 /* Did the client request the link name? */
460 request_body = lustre_msg_buf(req->rq_reqmsg, request_off);
461 body = lustre_msg_buf(req->rq_repmsg, reply_off);
462 if ((body->valid & OBD_MD_LINKNAME) && S_ISLNK(inode->i_mode)) {
463 char *tmp = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
465 rc = inode->i_op->readlink(dentry, tmp, req->rq_repmsg->
466 buflens[reply_off + 1]);
468 CERROR("readlink failed: %d\n", rc);
472 body->valid |= OBD_MD_LINKNAME;
475 mds_pack_inode2fid(&body->fid1, inode);
476 mds_pack_inode2body(body, inode);
477 if (S_ISREG(inode->i_mode)) {
478 struct lov_mds_md *md;
480 md = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
481 md->lmd_easize = mds->mds_max_mdsize;
482 rc = mds_fs_get_md(mds, inode, md);
487 CERROR("mds_fs_get_md failed: %d\n", rc);
490 body->valid |= OBD_MD_FLEASIZE;
495 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
497 struct mds_obd *mds = mds_req2mds(req);
498 struct obd_device *obd = req->rq_export->exp_obd;
499 struct obd_run_ctxt saved;
500 struct mds_body *body;
501 struct dentry *de = NULL, *dchild = NULL;
503 struct lustre_handle lockh;
505 int namelen, flags, lock_mode, rc = 0, old_offset = offset;
506 __u64 res_id[3] = {0, 0, 0};
509 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
511 if (req->rq_reqmsg->bufcount <= offset + 1) {
513 GOTO(out_pre_de, rc = -EINVAL);
516 body = lustre_msg_buf(req->rq_reqmsg, offset);
517 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
518 namelen = req->rq_reqmsg->buflens[offset + 1];
519 /* requests were at offset 2, replies go back at 1 */
523 push_ctxt(&saved, &mds->mds_ctxt);
524 de = mds_fid2dentry(mds, &body->fid1, NULL);
527 GOTO(out_pre_de, rc = -ESTALE);
531 CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
533 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
534 res_id[0] = dir->i_ino;
536 rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
537 NULL, 0, lock_mode, &lockh);
539 LDLM_DEBUG_NOLOCK("enqueue res %Lu", res_id[0]);
540 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
541 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
542 &flags, ldlm_completion_ast,
543 mds_blocking_ast, NULL, 0, &lockh);
544 if (rc != ELDLM_OK) {
545 CERROR("lock enqueue: err: %d\n", rc);
546 GOTO(out_create_de, rc = -EIO);
549 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
552 dchild = lookup_one_len(name, de, namelen - 1);
553 if (IS_ERR(dchild)) {
554 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
557 GOTO(out_create_dchild, rc = -ESTALE);
560 rc = mds_getattr_internal(mds, dchild, req, old_offset, offset);
566 ldlm_lock_decref(&lockh, lock_mode);
575 static int mds_getattr(int offset, struct ptlrpc_request *req)
577 struct mds_obd *mds = mds_req2mds(req);
578 struct obd_run_ctxt saved;
581 struct mds_body *body;
582 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
585 body = lustre_msg_buf(req->rq_reqmsg, offset);
586 push_ctxt(&saved, &mds->mds_ctxt);
587 de = mds_fid2dentry(mds, &body->fid1, NULL);
589 req->rq_status = -ENOENT;
590 GOTO(out_pop, rc = -ENOENT);
594 if (S_ISREG(body->fid1.f_type)) {
596 size[1] = mds->mds_max_mdsize;
597 } else if (body->valid & OBD_MD_LINKNAME) {
599 size[1] = inode->i_size;
602 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
604 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
605 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
610 req->rq_status = mds_getattr_internal(mds, de, req, offset, 0);
619 static int mds_statfs(struct ptlrpc_request *req)
621 struct mds_obd *mds = mds_req2mds(req);
622 struct obd_statfs *osfs;
624 int rc, size = sizeof(*osfs);
627 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
629 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
630 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
634 rc = vfs_statfs(mds->mds_sb, &sfs);
636 CERROR("mds: statfs failed: rc %d\n", rc);
639 osfs = lustre_msg_buf(req->rq_repmsg, 0);
640 memset(osfs, 0, size);
641 obd_statfs_pack(osfs, &sfs);
648 static int mds_open(struct ptlrpc_request *req)
651 struct mds_body *body;
653 struct vfsmount *mnt;
654 struct mds_obd *mds = mds_req2mds(req);
655 struct mds_export_data *med;
657 struct list_head *tmp;
658 struct mds_file_data *mfd;
659 int rc, size = sizeof(*body);
662 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
663 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
664 CERROR("mds: out of memory\n");
665 req->rq_status = -ENOMEM;
669 body = lustre_msg_buf(req->rq_reqmsg, 0);
671 /* was this animal open already? */
672 /* XXX we should only check on re-open, or do a refcount... */
673 med = &req->rq_export->exp_mds_data;
674 list_for_each(tmp, &med->med_open_head) {
675 struct mds_file_data *fd;
676 fd = list_entry(tmp, struct mds_file_data, mfd_list);
677 if (body->extra == fd->mfd_clientfd &&
678 body->fid1.id == fd->mfd_file->f_dentry->d_inode->i_ino) {
679 CERROR("Re opening %Ld\n", body->fid1.id);
684 OBD_ALLOC(mfd, sizeof(*mfd));
686 CERROR("mds: out of memory\n");
687 req->rq_status = -ENOMEM;
691 de = mds_fid2dentry(mds, &body->fid1, &mnt);
693 req->rq_status = -ENOENT;
697 /* check if this inode has seen a delayed object creation */
698 if (req->rq_reqmsg->bufcount > 1) {
700 struct lov_mds_md *md;
701 struct inode *inode = de->d_inode;
704 md = lustre_msg_buf(req->rq_reqmsg, 1);
706 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
708 req->rq_status = -ENOMEM;
712 /* XXX error handling */
713 rc = mds_fs_set_md(mds, inode, handle, md);
715 struct obd_run_ctxt saved;
716 push_ctxt(&saved, &mds->mds_ctxt);
717 rc = mds_update_last_rcvd(mds, handle, req);
723 /* FIXME: need to return last_rcvd, last_committed */
725 /* FIXME: keep rc intact */
726 rc = mds_fs_commit(mds, de->d_inode, handle);
734 file = dentry_open(de, mnt, flags & ~O_DIRECT);
735 if (!file || IS_ERR(file)) {
736 req->rq_status = -EINVAL;
737 OBD_FREE(mfd, sizeof(*mfd));
741 file->private_data = mfd;
742 mfd->mfd_file = file;
743 mfd->mfd_clientfd = body->extra;
744 list_add(&mfd->mfd_list, &med->med_open_head);
746 body = lustre_msg_buf(req->rq_repmsg, 0);
747 /* FIXME: need to have cookies involved here */
748 body->extra = (__u64) (unsigned long)file;
752 static int mds_close(struct ptlrpc_request *req)
755 struct mds_body *body;
757 struct mds_obd *mds = mds_req2mds(req);
758 struct vfsmount *mnt;
759 struct mds_file_data *mfd;
763 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
764 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
765 CERROR("mds: out of memory\n");
766 req->rq_status = -ENOMEM;
770 body = lustre_msg_buf(req->rq_reqmsg, 0);
771 de = mds_fid2dentry(mds, &body->fid1, &mnt);
773 req->rq_status = -ENOENT;
777 /* FIXME: need to have cookies involved here */
778 file = (struct file *)(unsigned long)body->extra;
781 mfd = (struct mds_file_data *)file->private_data;
782 list_del(&mfd->mfd_list);
783 OBD_FREE(mfd, sizeof(*mfd));
785 req->rq_status = filp_close(file, 0);
792 static int mds_readpage(struct ptlrpc_request *req)
794 struct mds_obd *mds = mds_req2mds(req);
795 struct vfsmount *mnt;
798 struct mds_body *body;
799 struct obd_run_ctxt saved;
800 int rc, size = sizeof(*body);
803 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
804 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
805 CERROR("mds: out of memory\n");
806 GOTO(out, rc = -ENOMEM);
809 body = lustre_msg_buf(req->rq_reqmsg, 0);
810 push_ctxt(&saved, &mds->mds_ctxt);
811 de = mds_fid2dentry(mds, &body->fid1, &mnt);
813 GOTO(out_pop, rc = PTR_ERR(de));
815 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
817 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
818 /* note: in case of an error, dentry_open puts dentry */
820 GOTO(out_pop, rc = PTR_ERR(file));
822 /* to make this asynchronous make sure that the handling function
823 doesn't send a reply when this function completes. Instead a
824 callback function would send the reply */
825 rc = mds_sendpage(req, file, body->size);
835 int mds_reint(int offset, struct ptlrpc_request *req)
838 struct mds_update_record rec;
840 rc = mds_update_unpack(req, offset, &rec);
841 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
842 CERROR("invalid record\n");
843 req->rq_status = -EINVAL;
846 /* rc will be used to interrupt a for loop over multiple records */
847 rc = mds_reint_rec(&rec, offset, req);
851 int mds_handle(struct ptlrpc_request *req)
856 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
857 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
858 CERROR("lustre_mds: Invalid request\n");
862 if (req->rq_reqmsg->type != PTL_RPC_MSG_REQUEST) {
863 CERROR("lustre_mds: wrong packet type sent %d\n",
864 req->rq_reqmsg->type);
865 GOTO(out, rc = -EINVAL);
868 if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
869 GOTO(out, rc = -ENOTCONN);
871 if (strcmp(req->rq_obd->obd_type->typ_name, "mds") != 0)
872 GOTO(out, rc = -EINVAL);
874 switch (req->rq_reqmsg->opc) {
876 CDEBUG(D_INODE, "connect\n");
877 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
878 rc = target_handle_connect(req);
882 CDEBUG(D_INODE, "disconnect\n");
883 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
884 rc = target_handle_disconnect(req);
888 CDEBUG(D_INODE, "getstatus\n");
889 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
890 rc = mds_getstatus(req);
894 CDEBUG(D_INODE, "getlovinfo\n");
895 rc = mds_getlovinfo(req);
899 CDEBUG(D_INODE, "getattr\n");
900 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
901 rc = mds_getattr(0, req);
905 CDEBUG(D_INODE, "statfs\n");
906 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
907 rc = mds_statfs(req);
911 CDEBUG(D_INODE, "readpage\n");
912 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
913 rc = mds_readpage(req);
915 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
920 int size = sizeof(struct mds_body);
921 CDEBUG(D_INODE, "reint\n");
922 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
924 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
930 rc = mds_reint(0, req);
931 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
936 CDEBUG(D_INODE, "open\n");
937 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
942 CDEBUG(D_INODE, "close\n");
943 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
948 CDEBUG(D_INODE, "enqueue\n");
949 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
950 rc = ldlm_handle_enqueue(req);
956 CDEBUG(D_INODE, "convert\n");
957 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
958 rc = ldlm_handle_convert(req);
964 CDEBUG(D_INODE, "cancel\n");
965 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CANCEL, 0);
966 rc = ldlm_handle_cancel(req);
970 case LDLM_BL_CALLBACK:
971 case LDLM_CP_CALLBACK:
972 CDEBUG(D_INODE, "callback\n");
973 CERROR("callbacks should not happen on MDS\n");
975 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
978 rc = ptlrpc_error(req->rq_svc, req);
985 struct mds_obd *mds = mds_req2mds(req);
986 req->rq_repmsg->last_xid = HTON__u64(mds->mds_last_rcvd);
987 req->rq_repmsg->last_committed =
988 HTON__u64(mds->mds_last_committed);
989 CDEBUG(D_INFO, "last_rcvd %Lu, last_committed %Lu, xid %d\n",
990 (unsigned long long)mds->mds_last_rcvd,
991 (unsigned long long)mds->mds_last_committed,
992 cpu_to_le32(req->rq_xid));
995 /* Still not 100% sure whether we should reply with the server
996 * last_rcvd or that of this client. I'm not sure it even makes
997 * a difference on a per-client basis, because last_rcvd is global
998 * and we are not supposed to allow transactions while in recovery.
1001 CERROR("mds: processing error %d\n", rc);
1002 ptlrpc_error(req->rq_svc, req);
1004 CDEBUG(D_NET, "sending reply\n");
1005 ptlrpc_reply(req->rq_svc, req);
1010 /* Update the server data on disk. This stores the new mount_count and
1011 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1012 * then the server last_rcvd value may be less than that of the clients.
1013 * This will alert us that we may need to do client recovery.
1015 * Assumes we are already in the server filesystem context.
1018 int mds_update_server_data(struct mds_obd *mds)
1020 struct mds_server_data *msd = mds->mds_server_data;
1021 struct file *filp = mds->mds_rcvd_filp;
1025 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1026 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1028 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1029 (unsigned long long)mds->mds_mount_count,
1030 (unsigned long long)mds->mds_last_rcvd);
1031 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1032 if (rc != sizeof(*msd)) {
1033 CERROR("error writing MDS server data: rc = %d\n", rc);
1038 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1040 CERROR("error flushing MDS server data: rc = %d\n", rc);
1045 /* Do recovery actions for the MDS */
1046 static int mds_recover(struct obd_device *obddev)
1048 struct mds_obd *mds = &obddev->u.mds;
1049 struct obd_run_ctxt saved;
1052 /* This happens at the end when recovery is complete */
1053 ++mds->mds_mount_count;
1054 push_ctxt(&saved, &mds->mds_ctxt);
1055 rc = mds_update_server_data(mds);
1061 #define MDS_NUM_THREADS 8
1062 /* mount the file system (secretly) */
1063 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1066 struct obd_ioctl_data* data = buf;
1067 struct mds_obd *mds = &obddev->u.mds;
1068 struct vfsmount *mnt;
1073 #ifdef CONFIG_DEV_RDONLY
1074 dev_clear_rdonly(2);
1076 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1077 GOTO(err_dec, rc = -EINVAL);
1079 mds->mds_fstype = strdup(data->ioc_inlbuf2);
1081 mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1084 CERROR("do_kern_mount failed: rc = %d\n", rc);
1085 GOTO(err_kfree, rc);
1088 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1090 GOTO(err_put, rc = -ENODEV);
1092 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1093 rc = mds_fs_setup(obddev, mnt);
1095 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1099 mds->mds_service = ptlrpc_init_svc(64 * 1024, MDS_REQUEST_PORTAL,
1100 MDC_REPLY_PORTAL, "self",mds_handle,
1102 if (!mds->mds_service) {
1103 CERROR("failed to start service\n");
1104 GOTO(err_fs, rc = -EINVAL);
1107 obddev->obd_namespace =
1108 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1109 if (obddev->obd_namespace == NULL) {
1110 mds_cleanup(obddev);
1111 GOTO(err_svc, rc = -ENOMEM);
1114 for (i = 0; i < MDS_NUM_THREADS; i++) {
1116 sprintf(name, "lustre_MDS_%02d", i);
1117 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1119 CERROR("cannot start MDS thread #%d: rc %d\n", i, rc);
1120 GOTO(err_thread, rc);
1124 rc = mds_recover(obddev);
1126 GOTO(err_thread, rc);
1131 ptlrpc_stop_all_threads(mds->mds_service);
1133 ptlrpc_unregister_service(mds->mds_service);
1135 mds_fs_cleanup(obddev);
1138 mntput(mds->mds_vfsmnt);
1142 kfree(mds->mds_fstype);
1148 static int mds_cleanup(struct obd_device *obddev)
1150 struct super_block *sb;
1151 struct mds_obd *mds = &obddev->u.mds;
1152 struct obd_run_ctxt saved;
1155 ptlrpc_stop_all_threads(mds->mds_service);
1156 ptlrpc_unregister_service(mds->mds_service);
1162 push_ctxt(&saved, &mds->mds_ctxt);
1163 mds_update_server_data(mds);
1165 if (mds->mds_rcvd_filp) {
1166 int rc = filp_close(mds->mds_rcvd_filp, 0);
1167 mds->mds_rcvd_filp = NULL;
1170 CERROR("last_rcvd file won't close, rc=%d\n", rc);
1175 mntput(mds->mds_vfsmnt);
1177 kfree(mds->mds_fstype);
1179 ldlm_namespace_free(obddev->obd_namespace);
1182 #ifdef CONFIG_DEV_RDONLY
1183 dev_clear_rdonly(2);
1185 mds_fs_cleanup(obddev);
1191 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1192 ldlm_mode_t mode, void *data)
1194 struct ptlrpc_request *req = req_cookie;
1201 if (req->rq_reqmsg->bufcount > 1) {
1202 /* an intent needs to be considered */
1203 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1204 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1205 struct mds_body *mds_rep;
1206 struct ldlm_reply *rep;
1207 __u64 new_resid[3] = {0, 0, 0}, old_res;
1208 int rc, size[3] = {sizeof(struct ldlm_reply),
1209 sizeof(struct mds_body),
1210 mds->mds_max_mdsize};
1212 it->opc = NTOH__u64(it->opc);
1214 LDLM_DEBUG(lock, "intent policy, opc: %s",
1215 ldlm_it2str(it->opc));
1217 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1220 rc = req->rq_status = -ENOMEM;
1224 rep = lustre_msg_buf(req->rq_repmsg, 0);
1225 rep->lock_policy_res1 = 1;
1227 /* execute policy */
1228 switch ((long)it->opc) {
1229 case IT_CREAT|IT_OPEN:
1230 rc = mds_reint(2, req);
1231 if (rc || (req->rq_status != 0 &&
1232 req->rq_status != -EEXIST)) {
1233 rep->lock_policy_res2 = req->rq_status;
1234 RETURN(ELDLM_LOCK_ABORTED);
1245 rc = mds_reint(2, req);
1246 if (rc || (req->rq_status != 0 &&
1247 req->rq_status != -EISDIR &&
1248 req->rq_status != -ENOTDIR)) {
1249 rep->lock_policy_res2 = req->rq_status;
1250 RETURN(ELDLM_LOCK_ABORTED);
1260 rc = mds_getattr_name(2, req);
1261 /* FIXME: we need to sit down and decide on who should
1262 * set req->rq_status, who should return negative and
1263 * positive return values, and what they all mean. */
1264 if (rc || req->rq_status != 0) {
1265 rep->lock_policy_res2 = req->rq_status;
1266 RETURN(ELDLM_LOCK_ABORTED);
1269 case IT_READDIR|IT_OPEN:
1273 CERROR("Unhandled intent\n");
1277 if (it->opc == IT_UNLINK || it->opc == IT_RMDIR ||
1278 it->opc == IT_RENAME || it->opc == IT_RENAME2)
1279 RETURN(ELDLM_LOCK_ABORTED);
1281 rep->lock_policy_res2 = req->rq_status;
1282 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1284 /* If the client is about to open a file that doesn't have an MD
1285 * stripe record, it's going to need a write lock. */
1286 if (it->opc & IT_OPEN) {
1287 struct lov_mds_md *md =
1288 lustre_msg_buf(req->rq_repmsg, 2);
1289 if (md->lmd_easize == 0) {
1290 LDLM_DEBUG(lock, "open with no EA; returning PW"
1292 lock->l_req_mode = LCK_PW;
1296 /* Give the client a lock on the child object, instead of the
1297 * parent that it requested. */
1298 new_resid[0] = NTOH__u32(mds_rep->ino);
1299 if (new_resid[0] == 0)
1301 old_res = lock->l_resource->lr_name[0];
1303 ldlm_lock_change_resource(lock, new_resid);
1304 if (lock->l_resource == NULL) {
1308 LDLM_DEBUG(lock, "intent policy, old res %ld",
1310 RETURN(ELDLM_LOCK_CHANGED);
1312 int size = sizeof(struct ldlm_reply);
1313 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1324 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1325 int len, void *karg, void *uarg);
1327 /* use obd ops to offer management infrastructure */
1328 static struct obd_ops mds_obd_ops = {
1329 o_connect: mds_connect,
1330 o_disconnect: mds_disconnect,
1332 o_cleanup: mds_cleanup,
1333 o_iocontrol: mds_iocontrol
1336 static int __init mds_init(void)
1338 class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1339 ldlm_register_intent(ldlm_intent_policy);
1343 static void __exit mds_exit(void)
1345 ldlm_unregister_intent();
1346 class_unregister_type(LUSTRE_MDS_NAME);
1349 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1350 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1351 MODULE_LICENSE("GPL");
1353 module_init(mds_init);
1354 module_exit(mds_exit);