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>
36 #include <linux/random.h>
37 #include <linux/locks.h>
38 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
39 #include <linux/buffer_head.h>
42 static kmem_cache_t *mds_file_cache;
44 extern int mds_get_lovtgts(struct mds_obd *obd, int tgt_count,
45 obd_uuid_t *uuidarray);
46 extern int mds_get_lovdesc(struct mds_obd *obd, struct lov_desc *desc);
47 extern int mds_update_last_rcvd(struct mds_obd *mds, void *handle,
48 struct ptlrpc_request *req);
49 static int mds_cleanup(struct obd_device * obddev);
51 inline struct mds_obd *mds_req2mds(struct ptlrpc_request *req)
53 return &req->rq_export->exp_obd->u.mds;
56 static int mds_bulk_timeout(void *data)
58 struct ptlrpc_bulk_desc *desc = data;
61 CERROR("(not yet) starting recovery of client %p\n", desc->bd_client);
65 /* Assumes caller has already pushed into the kernel filesystem context */
66 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
70 struct mds_obd *mds = mds_req2mds(req);
71 struct ptlrpc_bulk_desc *desc;
72 struct ptlrpc_bulk_page *bulk;
73 struct l_wait_info lwi;
77 desc = ptlrpc_prep_bulk(req->rq_connection);
79 GOTO(out, rc = -ENOMEM);
81 bulk = ptlrpc_prep_bulk_page(desc);
83 GOTO(cleanup_bulk, rc = -ENOMEM);
85 OBD_ALLOC(buf, PAGE_SIZE);
87 GOTO(cleanup_bulk, rc = -ENOMEM);
89 rc = mds_fs_readpage(mds, file, buf, PAGE_SIZE, (loff_t *)&offset);
92 GOTO(cleanup_buf, rc = -EIO);
94 bulk->bp_xid = req->rq_xid;
96 bulk->bp_buflen = PAGE_SIZE;
97 desc->bd_portal = MDS_BULK_PORTAL;
99 rc = ptlrpc_send_bulk(desc);
101 GOTO(cleanup_buf, rc);
103 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
104 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
105 OBD_FAIL_MDS_SENDPAGE, rc);
106 ptlrpc_abort_bulk(desc);
107 GOTO(cleanup_buf, rc);
110 lwi = LWI_TIMEOUT(obd_timeout * HZ, mds_bulk_timeout, desc);
111 rc = l_wait_event(desc->bd_waitq, desc->bd_flags & PTL_BULK_FL_SENT, &lwi);
113 if (rc != -ETIMEDOUT)
115 GOTO(cleanup_buf, rc);
120 OBD_FREE(buf, PAGE_SIZE);
122 ptlrpc_free_bulk(desc);
128 * Look up a named entry in a directory, and get an LDLM lock on it.
129 * 'dir' is a inode for which an LDLM lock has already been taken.
131 * If we do not need an exclusive or write lock on this entry (e.g.
132 * a read lock for attribute lookup only) then we do not hold the
133 * directory on return. It is up to the caller to know what type
134 * of lock it is getting, and clean up appropriately.
136 struct dentry *mds_name2locked_dentry(struct obd_device *obd,
137 struct dentry *dir, struct vfsmount **mnt,
138 char *name, int namelen, int lock_mode,
139 struct lustre_handle *lockh,
142 struct dentry *dchild;
144 __u64 res_id[3] = {0};
147 down(&dir->d_inode->i_sem);
148 dchild = lookup_one_len(name, dir, namelen);
149 if (IS_ERR(dchild)) {
150 CERROR("child lookup error %ld\n", PTR_ERR(dchild));
151 up(&dir->d_inode->i_sem);
155 if (dir_lock_mode != LCK_EX && dir_lock_mode != LCK_PW) {
156 up(&dir->d_inode->i_sem);
157 ldlm_lock_decref(lockh, dir_lock_mode);
160 if (lock_mode == 0 || !dchild->d_inode)
163 res_id[0] = dchild->d_inode->i_ino;
164 res_id[1] = dchild->d_inode->i_generation;
165 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
166 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
167 &flags, ldlm_completion_ast,
168 mds_blocking_ast, NULL, 0, lockh);
169 if (rc != ELDLM_OK) {
171 up(&dir->d_inode->i_sem);
172 RETURN(ERR_PTR(-ENOLCK)); /* XXX translate ldlm code */
178 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
179 struct vfsmount **mnt, int lock_mode,
180 struct lustre_handle *lockh)
182 struct mds_obd *mds = &obd->u.mds;
183 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
185 __u64 res_id[3] = {0};
191 res_id[0] = de->d_inode->i_ino;
192 res_id[1] = de->d_inode->i_generation;
193 rc = ldlm_match_or_enqueue(NULL, NULL, obd->obd_namespace, NULL,
194 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
195 &flags, ldlm_completion_ast,
196 mds_blocking_ast, NULL, 0, lockh);
197 if (rc != ELDLM_OK) {
199 retval = ERR_PTR(-ENOLCK); /* XXX translate ldlm code */
205 #ifndef DCACHE_DISCONNECTED
206 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
209 /* Look up an entry by inode number. */
210 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
211 struct vfsmount **mnt)
213 /* stolen from NFS */
214 struct super_block *sb = mds->mds_sb;
215 unsigned long ino = fid->id;
216 __u32 generation = fid->generation;
218 struct list_head *lp;
219 struct dentry *result;
222 RETURN(ERR_PTR(-ESTALE));
224 inode = iget(sb, ino);
226 RETURN(ERR_PTR(-ENOMEM));
228 CDEBUG(D_DENTRY, "--> mds_fid2dentry: sb %p\n", inode->i_sb);
230 if (is_bad_inode(inode) ||
231 (generation && inode->i_generation != generation)) {
232 /* we didn't find the right inode.. */
233 CERROR("bad inode %lu, link: %d ct: %d or version %u/%u\n",
234 inode->i_ino, inode->i_nlink,
235 atomic_read(&inode->i_count), inode->i_generation,
239 RETURN(ERR_PTR(-ESTALE));
242 /* now to find a dentry. If possible, get a well-connected one */
244 *mnt = mds->mds_vfsmnt;
245 spin_lock(&dcache_lock);
246 list_for_each(lp, &inode->i_dentry) {
247 result = list_entry(lp, struct dentry, d_alias);
248 if (!(result->d_flags & DCACHE_DISCONNECTED)) {
250 result->d_vfs_flags |= DCACHE_REFERENCED;
251 spin_unlock(&dcache_lock);
258 spin_unlock(&dcache_lock);
259 result = d_alloc_root(inode);
260 if (result == NULL) {
262 return ERR_PTR(-ENOMEM);
266 result->d_flags |= DCACHE_DISCONNECTED;
270 /* Establish a connection to the MDS.
272 * This will set up an export structure for the client to hold state data
273 * about that client, like open files, the last operation number it did
274 * on the server, etc.
276 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
277 obd_uuid_t cluuid, struct recovd_obd *recovd,
278 ptlrpc_recovery_cb_t recover)
280 struct obd_export *exp;
281 struct mds_export_data *med;
282 struct mds_client_data *mcd;
287 if (!conn || !obd || !cluuid)
292 spin_lock(&obd->obd_dev_lock);
293 list_for_each(p, &obd->obd_exports) {
294 exp = list_entry(p, struct obd_export, exp_obd_chain);
295 mcd = exp->exp_mds_data.med_mcd;
296 if (!memcmp(cluuid, mcd->mcd_uuid, sizeof(mcd->mcd_uuid))) {
297 LASSERT(exp->exp_obd == obd);
299 if (!list_empty(&exp->exp_conn_chain)) {
300 CERROR("existing uuid/export, list not empty!\n");
301 spin_unlock(&obd->obd_dev_lock);
302 /* XXX should we MOD_DEC_USE_COUNT; here? */
305 conn->addr = (__u64) (unsigned long)exp;
306 conn->cookie = exp->exp_cookie;
307 spin_unlock(&obd->obd_dev_lock);
308 CDEBUG(D_INFO, "existing export for UUID '%s' at %p\n",
310 CDEBUG(D_IOCTL,"connect: addr %Lx cookie %Lx\n",
311 (long long)conn->addr, (long long)conn->cookie);
316 spin_unlock(&obd->obd_dev_lock);
317 /* XXX There is a small race between checking the list and adding a
318 * new connection for the same UUID, but the real threat (list
319 * corruption when multiple different clients connect) is solved.
321 rc = class_connect(conn, obd, cluuid);
324 exp = class_conn2export(conn);
326 med = &exp->exp_mds_data;
328 OBD_ALLOC(mcd, sizeof(*mcd));
330 CERROR("mds: out of memory for client data\n");
331 GOTO(out_export, rc = -ENOMEM);
334 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
337 INIT_LIST_HEAD(&med->med_open_head);
338 spin_lock_init(&med->med_open_lock);
340 rc = mds_client_add(med, -1);
347 OBD_FREE(mcd, sizeof(*mcd));
349 class_disconnect(conn);
356 /* Call with med->med_open_lock held, please. */
357 inline int mds_close_mfd(struct mds_file_data *mfd, struct mds_export_data *med)
359 struct file *file = mfd->mfd_file;
360 LASSERT(file->private_data == mfd);
362 list_del(&mfd->mfd_list);
363 mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
364 kmem_cache_free(mds_file_cache, mfd);
366 return filp_close(file, 0);
369 static int mds_disconnect(struct lustre_handle *conn)
371 struct obd_export *export = class_conn2export(conn);
372 struct list_head *tmp, *n;
373 struct mds_export_data *med = &export->exp_mds_data;
378 * Close any open files.
380 spin_lock(&med->med_open_lock);
381 list_for_each_safe(tmp, n, &med->med_open_head) {
382 struct mds_file_data *mfd =
383 list_entry(tmp, struct mds_file_data, mfd_list);
384 rc = mds_close_mfd(mfd, med);
386 /* XXX better diagnostics, with file path and stuff */
387 CDEBUG(D_INODE, "Error %d closing mfd %p\n", rc, mfd);
390 spin_unlock(&med->med_open_lock);
392 ldlm_cancel_locks_for_export(export);
393 mds_client_free(export);
395 rc = class_disconnect(conn);
403 * XXX This is NOT guaranteed to flush all transactions to disk (even though
404 * it is equivalent to calling sync()) because it only _starts_ the flush
405 * and does not wait for completion. It's better than nothing though.
406 * What we really want is a mild form of fsync_dev_lockfs(), but it is
407 * non-standard, or enabling do_sync_supers in ext3, just for this call.
409 static void mds_fsync_super(struct super_block *sb)
413 if (sb->s_dirt && sb->s_op && sb->s_op->write_super)
414 sb->s_op->write_super(sb);
419 static int mds_getstatus(struct ptlrpc_request *req)
421 struct mds_obd *mds = mds_req2mds(req);
422 struct mds_body *body;
423 int rc, size = sizeof(*body);
426 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
427 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
428 CERROR("mds: out of memory for message: size=%d\n", size);
429 req->rq_status = -ENOMEM;
433 /* Flush any outstanding transactions to disk so the client will
434 * get the latest last_committed value and can drop their local
435 * requests if they have any. This would be fsync_super() if it
438 mds_fsync_super(mds->mds_sb);
440 body = lustre_msg_buf(req->rq_repmsg, 0);
441 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
443 /* the last_committed and last_xid fields are filled in for all
444 * replies already - no need to do so here also.
449 static int mds_getlovinfo(struct ptlrpc_request *req)
451 struct mds_obd *mds = mds_req2mds(req);
452 struct mds_status_req *streq;
453 struct lov_desc *desc;
455 int rc, size[2] = {sizeof(*desc)};
458 streq = lustre_msg_buf(req->rq_reqmsg, 0);
459 streq->flags = NTOH__u32(streq->flags);
460 streq->repbuf = NTOH__u32(streq->repbuf);
461 size[1] = streq->repbuf;
463 rc = lustre_pack_msg(2, size, NULL, &req->rq_replen, &req->rq_repmsg);
465 CERROR("mds: out of memory for message: size=%d\n", size[1]);
466 req->rq_status = -ENOMEM;
470 desc = lustre_msg_buf(req->rq_repmsg, 0);
471 rc = mds_get_lovdesc(mds, desc);
473 CERROR("mds_get_lovdesc error %d", rc);
478 tgt_count = le32_to_cpu(desc->ld_tgt_count);
479 if (tgt_count * sizeof(obd_uuid_t) > streq->repbuf) {
480 CERROR("too many targets, enlarge client buffers\n");
481 req->rq_status = -ENOSPC;
485 mds->mds_max_mdsize = sizeof(struct lov_mds_md) +
486 tgt_count * sizeof(struct lov_object_id);
487 rc = mds_get_lovtgts(mds, tgt_count,
488 lustre_msg_buf(req->rq_repmsg, 1));
490 CERROR("get_lovtgts error %d\n", rc);
497 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
498 void *data, __u32 data_len, int flag)
503 if (flag == LDLM_CB_CANCELING) {
504 /* Don't need to do anything here. */
508 /* XXX layering violation! -phil */
509 l_lock(&lock->l_resource->lr_namespace->ns_lock);
510 lock->l_flags |= LDLM_FL_CBPENDING;
511 do_ast = (!lock->l_readers && !lock->l_writers);
512 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
515 struct lustre_handle lockh;
518 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
519 ldlm_lock2handle(lock, &lockh);
520 rc = ldlm_cli_cancel(&lockh);
522 CERROR("ldlm_cli_cancel: %d\n", rc);
524 LDLM_DEBUG(lock, "Lock still has references, will be"
529 static int mds_getattr_internal(struct mds_obd *mds, struct dentry *dentry,
530 struct ptlrpc_request *req,
531 struct mds_body *reqbody, int reply_off)
533 struct mds_body *body;
534 struct inode *inode = dentry->d_inode;
541 body = lustre_msg_buf(req->rq_repmsg, reply_off);
543 mds_pack_inode2fid(&body->fid1, inode);
544 mds_pack_inode2body(body, inode);
546 if (S_ISREG(inode->i_mode)) {
547 struct lov_mds_md *lmm;
549 lmm = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
550 lmm->lmm_easize = mds->mds_max_mdsize;
551 rc = mds_fs_get_md(mds, inode, lmm);
556 CERROR("mds_fs_get_md failed: %d\n", rc);
559 body->valid |= OBD_MD_FLEASIZE;
560 } else if (S_ISLNK(inode->i_mode) && reqbody->valid & OBD_MD_LINKNAME) {
561 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1);
562 int len = req->rq_repmsg->buflens[reply_off + 1];
564 rc = inode->i_op->readlink(dentry, symname, len);
566 CERROR("readlink failed: %d\n", rc);
569 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
571 body->valid |= OBD_MD_LINKNAME;
576 static int mds_getattr_name(int offset, struct ptlrpc_request *req)
578 struct mds_obd *mds = mds_req2mds(req);
579 struct obd_device *obd = req->rq_export->exp_obd;
580 struct obd_run_ctxt saved;
581 struct mds_body *body;
582 struct dentry *de = NULL, *dchild = NULL;
584 struct lustre_handle lockh;
586 int namelen, flags = 0, lock_mode, rc = 0;
588 __u64 res_id[3] = {0, 0, 0};
591 LASSERT(!strcmp(req->rq_export->exp_obd->obd_type->typ_name, "mds"));
593 if (req->rq_reqmsg->bufcount <= offset + 1) {
595 GOTO(out_pre_de, rc = -EINVAL);
598 body = lustre_msg_buf(req->rq_reqmsg, offset);
599 name = lustre_msg_buf(req->rq_reqmsg, offset + 1);
600 namelen = req->rq_reqmsg->buflens[offset + 1];
601 /* requests were at offset 2, replies go back at 1 */
605 uc.ouc_fsuid = body->fsuid;
606 uc.ouc_fsgid = body->fsgid;
607 push_ctxt(&saved, &mds->mds_ctxt, &uc);
608 de = mds_fid2dentry(mds, &body->fid1, NULL);
611 GOTO(out_pre_de, rc = -ESTALE);
615 CDEBUG(D_INODE, "parent ino %ld, name %*s\n", dir->i_ino,namelen,name);
617 lock_mode = (req->rq_reqmsg->opc == MDS_REINT) ? LCK_CW : LCK_PW;
618 res_id[0] = dir->i_ino;
619 res_id[1] = dir->i_generation;
621 rc = ldlm_lock_match(obd->obd_namespace, res_id, LDLM_PLAIN,
622 NULL, 0, lock_mode, &lockh);
624 LDLM_DEBUG_NOLOCK("enqueue res "LPU64, res_id[0]);
625 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, NULL,
626 res_id, LDLM_PLAIN, NULL, 0, lock_mode,
627 &flags, ldlm_completion_ast,
628 mds_blocking_ast, NULL, 0, &lockh);
629 if (rc != ELDLM_OK) {
630 CERROR("lock enqueue: err: %d\n", rc);
631 GOTO(out_create_de, rc = -EIO);
634 ldlm_lock_dump((void *)(unsigned long)lockh.addr);
637 dchild = lookup_one_len(name, de, namelen - 1);
638 if (IS_ERR(dchild)) {
639 CDEBUG(D_INODE, "child lookup error %ld\n", PTR_ERR(dchild));
642 GOTO(out_create_dchild, rc = -ESTALE);
645 rc = mds_getattr_internal(mds, dchild, req, body, offset);
651 ldlm_lock_decref(&lockh, lock_mode);
660 static int mds_getattr(int offset, struct ptlrpc_request *req)
662 struct mds_obd *mds = mds_req2mds(req);
663 struct obd_run_ctxt saved;
666 struct mds_body *body;
668 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
671 body = lustre_msg_buf(req->rq_reqmsg, offset);
672 uc.ouc_fsuid = body->fsuid;
673 uc.ouc_fsgid = body->fsgid;
674 push_ctxt(&saved, &mds->mds_ctxt, &uc);
675 de = mds_fid2dentry(mds, &body->fid1, NULL);
677 rc = req->rq_status = -ENOENT;
678 GOTO(out_pop, PTR_ERR(de));
682 if (S_ISREG(body->fid1.f_type)) {
684 size[1] = mds->mds_max_mdsize;
685 } else if (body->valid & OBD_MD_LINKNAME) {
687 size[1] = MIN(inode->i_size + 1, body->size);
688 CDEBUG(D_INODE, "symlink size: %d, reply space: %d\n",
689 inode->i_size + 1, body->size);
692 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
693 CERROR("failed GETATTR_PACK test\n");
694 req->rq_status = -ENOMEM;
695 GOTO(out, rc = -ENOMEM);
698 rc = lustre_pack_msg(bufcount, size, NULL, &req->rq_replen,
701 CERROR("out of memory or FAIL_MDS_GETATTR_PACK\n");
706 req->rq_status = mds_getattr_internal(mds, de, req, body, 0);
715 static int mds_statfs(struct ptlrpc_request *req)
717 struct mds_obd *mds = mds_req2mds(req);
718 struct obd_statfs *osfs;
720 int rc, size = sizeof(*osfs);
723 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
725 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
726 CERROR("mds: statfs lustre_pack_msg failed: rc = %d\n", rc);
730 rc = vfs_statfs(mds->mds_sb, &sfs);
732 CERROR("mds: statfs failed: rc %d\n", rc);
735 osfs = lustre_msg_buf(req->rq_repmsg, 0);
736 memset(osfs, 0, size);
737 statfs_pack(osfs, &sfs);
738 obd_statfs_pack(osfs, osfs);
745 static struct mds_file_data *mds_handle2mfd(struct lustre_handle *handle)
747 struct mds_file_data *mfd = NULL;
749 if (!handle || !handle->addr)
752 mfd = (struct mds_file_data *)(unsigned long)(handle->addr);
753 if (!kmem_cache_validate(mds_file_cache, mfd))
756 if (mfd->mfd_servercookie != handle->cookie)
762 static int mds_store_ea(struct mds_obd *mds, struct ptlrpc_request *req,
763 struct mds_body *body, struct dentry *de,
764 struct lov_mds_md *lmm)
766 struct obd_run_ctxt saved;
771 uc.ouc_fsuid = body->fsuid;
772 uc.ouc_fsgid = body->fsgid;
773 push_ctxt(&saved, &mds->mds_ctxt, &uc);
774 handle = mds_fs_start(mds, de->d_inode, MDS_FSOP_SETATTR);
776 GOTO(out_ea, rc = -ENOMEM);
778 rc = mds_fs_set_md(mds, de->d_inode, handle, lmm);
780 rc = mds_update_last_rcvd(mds, handle, req);
782 rc2 = mds_fs_commit(mds, de->d_inode, handle);
791 static int mds_open(struct ptlrpc_request *req)
793 struct mds_obd *mds = mds_req2mds(req);
794 struct mds_body *body;
795 struct mds_export_data *med;
796 struct mds_file_data *mfd;
799 struct vfsmount *mnt;
801 struct list_head *tmp;
802 int rc, size = sizeof(*body);
805 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK)) {
806 CERROR("test case OBD_FAIL_MDS_OPEN_PACK\n");
807 req->rq_status = -ENOMEM;
811 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
813 CERROR("mds: pack error: rc = %d\n", rc);
818 body = lustre_msg_buf(req->rq_reqmsg, 0);
820 /* was this animal open already and the client lost the reply? */
821 /* XXX need some way to detect a reopen, to avoid locked list walks */
822 med = &req->rq_export->exp_mds_data;
823 spin_lock(&med->med_open_lock);
824 list_for_each(tmp, &med->med_open_head) {
825 mfd = list_entry(tmp, typeof(*mfd), mfd_list);
826 if (!memcmp(&mfd->mfd_clienthandle, &body->handle,
827 sizeof(mfd->mfd_clienthandle)) &&
828 body->fid1.id == mfd->mfd_file->f_dentry->d_inode->i_ino) {
829 de = mfd->mfd_file->f_dentry;
830 spin_unlock(&med->med_open_lock);
831 CERROR("Re opening "LPD64"\n", body->fid1.id);
832 GOTO(out_pack, rc = 0);
835 spin_unlock(&med->med_open_lock);
837 mfd = kmem_cache_alloc(mds_file_cache, GFP_KERNEL);
839 CERROR("mds: out of memory\n");
840 req->rq_status = -ENOMEM;
844 de = mds_fid2dentry(mds, &body->fid1, &mnt);
846 GOTO(out_free, rc = PTR_ERR(de));
848 /* check if this inode has seen a delayed object creation */
849 if (lustre_msg_get_op_flags(req->rq_reqmsg) & MDS_OPEN_HAS_EA) {
850 struct lov_mds_md *lmm = lustre_msg_buf(req->rq_reqmsg, 1);
852 rc = mds_store_ea(mds, req, body, de, lmm);
861 /* dentry_open does a dput(de) and mntput(mnt) on error */
862 file = dentry_open(de, mnt, flags & ~O_DIRECT);
868 file->private_data = mfd;
869 mfd->mfd_file = file;
870 memcpy(&mfd->mfd_clienthandle, &body->handle, sizeof(body->handle));
871 get_random_bytes(&mfd->mfd_servercookie, sizeof(mfd->mfd_servercookie));
872 spin_lock(&med->med_open_lock);
873 list_add(&mfd->mfd_list, &med->med_open_head);
874 spin_unlock(&med->med_open_lock);
877 body = lustre_msg_buf(req->rq_repmsg, 0);
878 mds_pack_inode2fid(&body->fid1, de->d_inode);
879 mds_pack_inode2body(body, de->d_inode);
880 body->handle.addr = (__u64)(unsigned long)mfd;
881 body->handle.cookie = mfd->mfd_servercookie;
882 CDEBUG(D_INODE, "llite file "LPX64": addr %p, cookie "LPX64"\n",
883 mfd->mfd_clienthandle.addr, mfd, mfd->mfd_servercookie);
887 mfd->mfd_servercookie = DEAD_HANDLE_MAGIC;
888 kmem_cache_free(mds_file_cache, mfd);
893 static int mds_close(struct ptlrpc_request *req)
895 struct mds_export_data *med = &req->rq_export->exp_mds_data;
896 struct mds_body *body;
897 struct mds_file_data *mfd;
901 body = lustre_msg_buf(req->rq_reqmsg, 0);
903 mfd = mds_handle2mfd(&body->handle);
905 CERROR("no handle for file close "LPD64
906 ": addr "LPX64", cookie "LPX64"\n",
907 body->fid1.id, body->handle.addr, body->handle.cookie);
911 spin_lock(&med->med_open_lock);
912 req->rq_status = mds_close_mfd(mfd, med);
913 spin_unlock(&med->med_open_lock);
915 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK)) {
916 CERROR("test case OBD_FAIL_MDS_CLOSE_PACK\n");
917 req->rq_status = -ENOMEM;
921 rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
923 CERROR("mds: lustre_pack_msg: rc = %d\n", rc);
930 static int mds_readpage(struct ptlrpc_request *req)
932 struct mds_obd *mds = mds_req2mds(req);
933 struct vfsmount *mnt;
936 struct mds_body *body, *repbody;
937 struct obd_run_ctxt saved;
938 int rc, size = sizeof(*body);
942 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen, &req->rq_repmsg);
943 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
944 CERROR("mds: out of memory\n");
945 GOTO(out, rc = -ENOMEM);
948 body = lustre_msg_buf(req->rq_reqmsg, 0);
949 uc.ouc_fsuid = body->fsuid;
950 uc.ouc_fsgid = body->fsgid;
951 push_ctxt(&saved, &mds->mds_ctxt, &uc);
952 de = mds_fid2dentry(mds, &body->fid1, &mnt);
954 GOTO(out_pop, rc = PTR_ERR(de));
956 CDEBUG(D_INODE, "ino %ld\n", de->d_inode->i_ino);
958 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
959 /* note: in case of an error, dentry_open puts dentry */
961 GOTO(out_pop, rc = PTR_ERR(file));
963 repbody = lustre_msg_buf(req->rq_repmsg, 0);
964 repbody->size = file->f_dentry->d_inode->i_size;
965 repbody->valid = OBD_MD_FLSIZE;
967 /* to make this asynchronous make sure that the handling function
968 doesn't send a reply when this function completes. Instead a
969 callback function would send the reply */
970 /* note: in case of an error, dentry_open puts dentry */
971 rc = mds_sendpage(req, file, body->size);
981 int mds_reint(int offset, struct ptlrpc_request *req)
984 struct mds_update_record rec;
986 rc = mds_update_unpack(req, offset, &rec);
987 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
988 CERROR("invalid record\n");
989 req->rq_status = -EINVAL;
992 /* rc will be used to interrupt a for loop over multiple records */
993 rc = mds_reint_rec(&rec, offset, req);
997 int mds_handle(struct ptlrpc_request *req)
1002 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1003 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_HANDLE_UNPACK)) {
1004 CERROR("lustre_mds: Invalid request\n");
1008 if (req->rq_reqmsg->opc != MDS_CONNECT && req->rq_export == NULL)
1009 GOTO(out, rc = -ENOTCONN);
1011 LASSERT(!strcmp(req->rq_obd->obd_type->typ_name, LUSTRE_MDT_NAME));
1013 switch (req->rq_reqmsg->opc) {
1015 CDEBUG(D_INODE, "connect\n");
1016 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1017 rc = target_handle_connect(req);
1020 case MDS_DISCONNECT:
1021 CDEBUG(D_INODE, "disconnect\n");
1022 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1023 rc = target_handle_disconnect(req);
1027 CDEBUG(D_INODE, "getstatus\n");
1028 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1029 rc = mds_getstatus(req);
1032 case MDS_GETLOVINFO:
1033 CDEBUG(D_INODE, "getlovinfo\n");
1034 rc = mds_getlovinfo(req);
1038 CDEBUG(D_INODE, "getattr\n");
1039 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1040 rc = mds_getattr(0, req);
1044 CDEBUG(D_INODE, "statfs\n");
1045 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1046 rc = mds_statfs(req);
1050 CDEBUG(D_INODE, "readpage\n");
1051 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1052 rc = mds_readpage(req);
1054 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE))
1059 int size = sizeof(struct mds_body);
1060 CDEBUG(D_INODE, "reint\n");
1061 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1063 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1066 req->rq_status = rc;
1069 rc = mds_reint(0, req);
1070 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET_REP, 0);
1075 CDEBUG(D_INODE, "open\n");
1076 OBD_FAIL_RETURN(OBD_FAIL_MDS_OPEN_NET, 0);
1081 CDEBUG(D_INODE, "close\n");
1082 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1083 rc = mds_close(req);
1087 CDEBUG(D_INODE, "enqueue\n");
1088 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1089 rc = ldlm_handle_enqueue(req);
1094 CDEBUG(D_INODE, "convert\n");
1095 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1096 rc = ldlm_handle_convert(req);
1100 case LDLM_BL_CALLBACK:
1101 case LDLM_CP_CALLBACK:
1102 CDEBUG(D_INODE, "callback\n");
1103 CERROR("callbacks should not happen on MDS\n");
1105 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1108 rc = ptlrpc_error(req->rq_svc, req);
1115 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1116 struct mds_obd *mds = mds_req2mds(req);
1118 req->rq_repmsg->last_xid =
1119 HTON__u64(le64_to_cpu(med->med_mcd->mcd_last_xid));
1120 req->rq_repmsg->last_committed =
1121 HTON__u64(mds->mds_last_committed);
1122 CDEBUG(D_INFO, "last_rcvd ~%Lu, last_committed %Lu, xid %d\n",
1123 (unsigned long long)mds->mds_last_rcvd,
1124 (unsigned long long)mds->mds_last_committed,
1125 cpu_to_le32(req->rq_xid));
1129 CERROR("mds: processing error (opcode %d): %d\n",
1130 req->rq_reqmsg->opc, rc);
1131 ptlrpc_error(req->rq_svc, req);
1133 CDEBUG(D_NET, "sending reply\n");
1134 ptlrpc_reply(req->rq_svc, req);
1139 /* Update the server data on disk. This stores the new mount_count and
1140 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1141 * then the server last_rcvd value may be less than that of the clients.
1142 * This will alert us that we may need to do client recovery.
1144 * Assumes we are already in the server filesystem context.
1146 * Also assumes for mds_last_rcvd that we are not modifying it (no locking).
1149 int mds_update_server_data(struct mds_obd *mds)
1151 struct mds_server_data *msd = mds->mds_server_data;
1152 struct file *filp = mds->mds_rcvd_filp;
1156 msd->msd_last_rcvd = cpu_to_le64(mds->mds_last_rcvd);
1157 msd->msd_mount_count = cpu_to_le64(mds->mds_mount_count);
1159 CDEBUG(D_SUPER, "MDS mount_count is %Lu, last_rcvd is %Lu\n",
1160 (unsigned long long)mds->mds_mount_count,
1161 (unsigned long long)mds->mds_last_rcvd);
1162 rc = lustre_fwrite(filp, (char *)msd, sizeof(*msd), &off);
1163 if (rc != sizeof(*msd)) {
1164 CERROR("error writing MDS server data: rc = %d\n", rc);
1169 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
1170 rc = fsync_dev(filp->f_dentry->d_inode->i_rdev);
1172 rc = file_fsync(filp, filp->f_dentry, 1);
1175 CERROR("error flushing MDS server data: rc = %d\n", rc);
1180 /* Do recovery actions for the MDS */
1181 static int mds_recover(struct obd_device *obddev)
1183 struct mds_obd *mds = &obddev->u.mds;
1184 struct obd_run_ctxt saved;
1187 /* This happens at the end when recovery is complete */
1188 ++mds->mds_mount_count;
1189 push_ctxt(&saved, &mds->mds_ctxt, NULL);
1190 rc = mds_update_server_data(mds);
1196 /* mount the file system (secretly) */
1197 static int mds_setup(struct obd_device *obddev, obd_count len, void *buf)
1199 struct obd_ioctl_data* data = buf;
1200 struct mds_obd *mds = &obddev->u.mds;
1201 struct vfsmount *mnt;
1206 #ifdef CONFIG_DEV_RDONLY
1207 dev_clear_rdonly(2);
1209 if (!data->ioc_inlbuf1 || !data->ioc_inlbuf2)
1210 GOTO(err_dec, rc = -EINVAL);
1212 mds->mds_fstype = strdup(data->ioc_inlbuf2);
1214 mnt = do_kern_mount(mds->mds_fstype, 0, data->ioc_inlbuf1, NULL);
1217 CERROR("do_kern_mount failed: rc = %d\n", rc);
1218 GOTO(err_kfree, rc);
1221 CERROR("%s: mnt is %p\n", data->ioc_inlbuf1, mnt);
1222 mds->mds_sb = mnt->mnt_root->d_inode->i_sb;
1224 GOTO(err_put, rc = -ENODEV);
1226 spin_lock_init(&mds->mds_last_lock);
1227 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1228 rc = mds_fs_setup(obddev, mnt);
1230 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
1234 obddev->obd_namespace =
1235 ldlm_namespace_new("mds_server", LDLM_NAMESPACE_SERVER);
1236 if (obddev->obd_namespace == NULL) {
1237 mds_cleanup(obddev);
1238 GOTO(err_fs, rc = -ENOMEM);
1242 rc = mds_recover(obddev);
1246 ptlrpc_init_client(LDLM_REQUEST_PORTAL, LDLM_REPLY_PORTAL,
1247 "mds_ldlm_client", &obddev->obd_ldlm_client);
1252 mds_fs_cleanup(obddev);
1255 mntput(mds->mds_vfsmnt);
1259 kfree(mds->mds_fstype);
1265 static int mds_cleanup(struct obd_device *obddev)
1267 struct super_block *sb;
1268 struct mds_obd *mds = &obddev->u.mds;
1269 struct obd_run_ctxt saved;
1276 push_ctxt(&saved, &mds->mds_ctxt, NULL);
1277 mds_update_server_data(mds);
1279 if (mds->mds_rcvd_filp) {
1280 int rc = filp_close(mds->mds_rcvd_filp, 0);
1281 mds->mds_rcvd_filp = NULL;
1284 CERROR("last_rcvd file won't close, rc=%d\n", rc);
1289 mntput(mds->mds_vfsmnt);
1291 kfree(mds->mds_fstype);
1293 ldlm_namespace_free(obddev->obd_namespace);
1296 #ifdef CONFIG_DEV_RDONLY
1297 dev_clear_rdonly(2);
1299 mds_fs_cleanup(obddev);
1305 static int ldlm_intent_policy(struct ldlm_lock *lock, void *req_cookie,
1306 ldlm_mode_t mode, int flags, void *data)
1308 struct ptlrpc_request *req = req_cookie;
1315 if (req->rq_reqmsg->bufcount > 1) {
1316 /* an intent needs to be considered */
1317 struct ldlm_intent *it = lustre_msg_buf(req->rq_reqmsg, 1);
1318 struct mds_obd *mds= &req->rq_export->exp_obd->u.mds;
1319 struct mds_body *mds_rep;
1320 struct ldlm_reply *rep;
1321 __u64 new_resid[3] = {0, 0, 0}, old_res;
1322 int rc, size[3] = {sizeof(struct ldlm_reply),
1323 sizeof(struct mds_body),
1324 mds->mds_max_mdsize};
1326 it->opc = NTOH__u64(it->opc);
1328 LDLM_DEBUG(lock, "intent policy, opc: %s",
1329 ldlm_it2str(it->opc));
1331 rc = lustre_pack_msg(3, size, NULL, &req->rq_replen,
1334 rc = req->rq_status = -ENOMEM;
1338 rep = lustre_msg_buf(req->rq_repmsg, 0);
1339 rep->lock_policy_res1 = 1;
1341 /* execute policy */
1342 switch ((long)it->opc) {
1343 case IT_CREAT|IT_OPEN:
1344 rc = mds_reint(2, req);
1345 if (rc || (req->rq_status != 0 &&
1346 req->rq_status != -EEXIST)) {
1347 rep->lock_policy_res2 = req->rq_status;
1348 RETURN(ELDLM_LOCK_ABORTED);
1359 rc = mds_reint(2, req);
1360 if (rc || (req->rq_status != 0 &&
1361 req->rq_status != -EISDIR &&
1362 req->rq_status != -ENOTDIR)) {
1363 rep->lock_policy_res2 = req->rq_status;
1364 RETURN(ELDLM_LOCK_ABORTED);
1375 rc = mds_getattr_name(2, req);
1376 /* FIXME: we need to sit down and decide on who should
1377 * set req->rq_status, who should return negative and
1378 * positive return values, and what they all mean. */
1379 if (rc || req->rq_status != 0) {
1380 rep->lock_policy_res2 = req->rq_status;
1381 RETURN(ELDLM_LOCK_ABORTED);
1384 case IT_READDIR|IT_OPEN:
1388 CERROR("Unhandled intent\n");
1392 /* We don't bother returning a lock to the client for a file
1393 * or directory we are removing.
1395 * As for link and rename, there is no reason for the client
1396 * to get a lock on the target at this point. If they are
1397 * going to modify the file/directory later they will get a
1398 * lock at that time.
1400 if (it->opc & (IT_UNLINK | IT_RMDIR | IT_LINK | IT_LINK2 |
1401 IT_RENAME | IT_RENAME2))
1402 RETURN(ELDLM_LOCK_ABORTED);
1404 rep->lock_policy_res2 = req->rq_status;
1405 mds_rep = lustre_msg_buf(req->rq_repmsg, 1);
1407 /* If the client is about to open a file that doesn't have an MD
1408 * stripe record, it's going to need a write lock. */
1409 if (it->opc & IT_OPEN) {
1410 struct lov_mds_md *lmm =
1411 lustre_msg_buf(req->rq_repmsg, 2);
1412 if (lmm->lmm_easize == 0) {
1413 LDLM_DEBUG(lock, "open with no EA; returning PW"
1415 lock->l_req_mode = LCK_PW;
1419 if (flags & LDLM_FL_INTENT_ONLY) {
1420 LDLM_DEBUG(lock, "INTENT_ONLY, aborting lock");
1421 RETURN(ELDLM_LOCK_ABORTED);
1423 /* Give the client a lock on the child object, instead of the
1424 * parent that it requested. */
1425 new_resid[0] = NTOH__u32(mds_rep->ino);
1426 new_resid[1] = NTOH__u32(mds_rep->generation);
1427 if (new_resid[0] == 0)
1429 old_res = lock->l_resource->lr_name[0];
1431 ldlm_lock_change_resource(lock, new_resid);
1432 if (lock->l_resource == NULL) {
1436 LDLM_DEBUG(lock, "intent policy, old res %ld",
1438 RETURN(ELDLM_LOCK_CHANGED);
1440 int size = sizeof(struct ldlm_reply);
1441 rc = lustre_pack_msg(1, &size, NULL, &req->rq_replen,
1452 #define MDT_NUM_THREADS 8
1453 static int mdt_setup(struct obd_device *obddev, obd_count len, void *buf)
1456 // struct obd_ioctl_data* data = buf;
1457 struct mds_obd *mds = &obddev->u.mds;
1463 mds->mds_service = ptlrpc_init_svc(MDS_NEVENTS, MDS_NBUFS,
1464 MDS_BUFSIZE, MDS_MAXREQSIZE,
1465 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
1466 "self", mds_handle, "mds");
1467 if (!mds->mds_service) {
1468 CERROR("failed to start service\n");
1469 GOTO(err_dec, rc = -EINVAL);
1472 for (i = 0; i < MDT_NUM_THREADS; i++) {
1474 sprintf(name, "lustre_MDT_%02d", i);
1475 rc = ptlrpc_start_thread(obddev, mds->mds_service, name);
1477 CERROR("cannot start MDT thread #%d: rc %d\n", i, rc);
1478 GOTO(err_thread, rc);
1485 ptlrpc_stop_all_threads(mds->mds_service);
1486 ptlrpc_unregister_service(mds->mds_service);
1493 static int mdt_cleanup(struct obd_device *obddev)
1495 struct mds_obd *mds = &obddev->u.mds;
1498 ptlrpc_stop_all_threads(mds->mds_service);
1499 ptlrpc_unregister_service(mds->mds_service);
1505 extern int mds_iocontrol(long cmd, struct lustre_handle *conn,
1506 int len, void *karg, void *uarg);
1508 /* use obd ops to offer management infrastructure */
1509 static struct obd_ops mds_obd_ops = {
1510 o_connect: mds_connect,
1511 o_disconnect: mds_disconnect,
1513 o_cleanup: mds_cleanup,
1514 o_iocontrol: mds_iocontrol
1517 static struct obd_ops mdt_obd_ops = {
1519 o_cleanup: mdt_cleanup,
1523 static int __init mds_init(void)
1525 mds_file_cache = kmem_cache_create("ll_mds_file_data",
1526 sizeof(struct mds_file_data),
1528 if (mds_file_cache == NULL)
1531 class_register_type(&mds_obd_ops, LUSTRE_MDS_NAME);
1532 class_register_type(&mdt_obd_ops, LUSTRE_MDT_NAME);
1533 ldlm_register_intent(ldlm_intent_policy);
1537 static void __exit mds_exit(void)
1539 ldlm_unregister_intent();
1540 class_unregister_type(LUSTRE_MDS_NAME);
1541 class_unregister_type(LUSTRE_MDT_NAME);
1542 if (kmem_cache_destroy(mds_file_cache))
1543 CERROR("couldn't free MDS file cache\n");
1546 MODULE_AUTHOR("Cluster File Systems <info@clusterfs.com>");
1547 MODULE_DESCRIPTION("Lustre Metadata Server (MDS) v0.01");
1548 MODULE_LICENSE("GPL");
1550 module_init(mds_init);
1551 module_exit(mds_exit);