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) 2001-2003 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>
11 * Author: Mike Shaver <shaver@clusterfs.com>
13 * This file is part of Lustre, http://www.lustre.org.
15 * Lustre is free software; you can redistribute it and/or
16 * modify it under the terms of version 2 of the GNU General Public
17 * License as published by the Free Software Foundation.
19 * Lustre is distributed in the hope that it will be useful,
20 * but WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * GNU General Public License for more details.
24 * You should have received a copy of the GNU General Public License
25 * along with Lustre; if not, write to the Free Software
26 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
30 # define EXPORT_SYMTAB
32 #define DEBUG_SUBSYSTEM S_MDS
34 #include <linux/module.h>
35 #include <linux/lustre_mds.h>
36 #include <linux/lustre_dlm.h>
37 #include <linux/init.h>
38 #include <linux/obd_class.h>
39 #include <linux/random.h>
41 #include <linux/jbd.h>
42 #include <linux/namei.h>
43 #include <linux/ext3_fs.h>
44 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
45 # include <linux/smp_lock.h>
46 # include <linux/buffer_head.h>
47 # include <linux/workqueue.h>
48 # include <linux/mount.h>
50 # include <linux/locks.h>
52 #include <linux/obd_lov.h>
53 #include <linux/obd_ost.h>
54 #include <linux/lustre_mds.h>
55 #include <linux/lustre_fsfilt.h>
56 #include <linux/lprocfs_status.h>
57 #include <linux/lustre_commit_confd.h>
58 #include <linux/lustre_acl.h>
59 #include <linux/lustre_gs.h>
60 #include "mds_internal.h"
61 #include <linux/lustre_sec.h>
62 #include <linux/lustre_audit.h>
64 extern int mds_audit_auth(struct ptlrpc_request *, struct lvfs_ucred *,
65 audit_op, struct lustre_id *,
67 extern int mds_audit_stat(struct ptlrpc_request *, struct lustre_id *,
68 struct dentry *, int);
69 extern int mds_audit_open(struct ptlrpc_request *, struct mds_update_record *,
72 static int mds_intent_policy(struct ldlm_namespace *ns,
73 struct ldlm_lock **lockp, void *req_cookie,
74 ldlm_mode_t mode, int flags, void *data);
75 static int mds_postsetup(struct obd_device *obd);
76 static int mds_cleanup(struct obd_device *obd, int flags);
79 /* Assumes caller has already pushed into the kernel filesystem context */
80 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
81 loff_t offset, int count)
83 struct ptlrpc_bulk_desc *desc;
84 struct l_wait_info lwi;
86 int rc = 0, npages, i, tmpcount, tmpsize = 0;
89 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
91 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
92 OBD_ALLOC(pages, sizeof(*pages) * npages);
94 GOTO(out, rc = -ENOMEM);
96 desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
99 GOTO(out_free, rc = -ENOMEM);
101 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
102 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
104 pages[i] = alloc_pages(GFP_KERNEL, 0);
105 if (pages[i] == NULL)
106 GOTO(cleanup_buf, rc = -ENOMEM);
108 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
111 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
112 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
113 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
114 tmpsize, offset, file->f_dentry->d_inode->i_ino,
115 file->f_dentry->d_inode->i_size);
117 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
118 kmap(pages[i]), tmpsize, &offset);
122 GOTO(cleanup_buf, rc = -EIO);
125 LASSERT(desc->bd_nob == count);
127 rc = ptlrpc_start_bulk_transfer(desc);
129 GOTO(cleanup_buf, rc);
131 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
132 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
133 OBD_FAIL_MDS_SENDPAGE, rc = -EIO);
134 GOTO(abort_bulk, rc);
137 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
138 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
139 LASSERT (rc == 0 || rc == -ETIMEDOUT);
142 if (desc->bd_success &&
143 desc->bd_nob_transferred == count)
144 GOTO(cleanup_buf, rc);
146 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
149 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
150 (rc == -ETIMEDOUT) ? "timeout" : "network error",
151 desc->bd_nob_transferred, count,
152 req->rq_export->exp_client_uuid.uuid,
153 req->rq_export->exp_connection->c_remote_uuid.uuid);
155 ptlrpc_fail_export(req->rq_export);
159 ptlrpc_abort_bulk (desc);
161 for (i = 0; i < npages; i++)
163 __free_pages(pages[i], 0);
165 ptlrpc_free_bulk(desc);
167 OBD_FREE(pages, sizeof(*pages) * npages);
172 extern char *ldlm_lockname[];
174 int mds_lock_mode_for_dir(struct obd_device *obd,
175 struct dentry *dentry, int mode)
177 int ret_mode = 0, split;
179 /* any dir access needs couple locks:
180 * 1) on part of dir we gonna lookup/modify in
181 * 2) on a whole dir to protect it from concurrent splitting
182 * and to flush client's cache for readdir()
183 * so, for a given mode and dentry this routine decides what
184 * lock mode to use for lock #2:
185 * 1) if caller's gonna lookup in dir then we need to protect
186 * dir from being splitted only - LCK_CR
187 * 2) if caller's gonna modify dir then we need to protect
188 * dir from being splitted and to flush cache - LCK_CW
189 * 3) if caller's gonna modify dir and that dir seems ready
190 * for splitting then we need to protect it from any
191 * type of access (lookup/modify/split) - LCK_EX -bzzz */
193 split = mds_splitting_expected(obd, dentry);
196 * it is important to check here only for MDS_NO_SPLITTABLE. The reason
197 * is that MDS_NO_SPLITTABLE means dir is not splittable in principle
198 * and another thread will not split it on the quiet. But if we have
199 * MDS_NO_SPLIT_EXPECTED, this means, that dir may be splitted anytime,
200 * but not now (for current thread) and we should consider that it can
201 * happen soon and go that branch which can yield LCK_EX to protect from
202 * possible splitting.
204 if (split == MDS_NO_SPLITTABLE) {
206 * this inode won't be splitted. so we need not to protect from
207 * just flush client's cache on modification.
214 if (mode == LCK_EX) {
216 } else if (mode == LCK_PR) {
218 } else if (mode == LCK_PW) {
220 * caller gonna modify directory. We use concurrent
221 * write lock here to retract client's cache for
224 if (split == MDS_EXPECT_SPLIT) {
226 * splitting possible. serialize any access the
227 * idea is that first one seen dir is splittable
228 * is given exclusive lock and split
229 * directory. caller passes lock mode to
230 * mds_try_to_split_dir() and splitting would be
231 * done with exclusive lock only -bzzz.
233 CDEBUG(D_OTHER, "%s: gonna split %lu/%lu\n",
235 (unsigned long)dentry->d_inode->i_ino,
236 (unsigned long)dentry->d_inode->i_generation);
247 /* only valid locked dentries or errors should be returned */
248 struct dentry *mds_id2locked_dentry(struct obd_device *obd, struct lustre_id *id,
249 struct vfsmount **mnt, int lock_mode,
250 struct lustre_handle *lockh, int *mode,
251 char *name, int namelen, __u64 lockpart)
253 struct dentry *de = mds_id2dentry(obd, id, mnt), *retval = de;
254 ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
255 struct ldlm_res_id res_id = { .name = {0} };
256 int flags = LDLM_FL_ATOMIC_CB, rc;
263 res_id.name[0] = id_fid(id);
264 res_id.name[1] = id_group(id);
267 if (name && IS_PDIROPS(de->d_inode)) {
268 ldlm_policy_data_t cpolicy =
269 { .l_inodebits = { MDS_INODELOCK_UPDATE } };
270 LASSERT(mode != NULL);
271 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
273 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
275 &cpolicy, *mode, &flags,
277 ldlm_completion_ast, NULL, NULL,
278 NULL, 0, NULL, lockh + 1);
279 if (rc != ELDLM_OK) {
281 RETURN(ERR_PTR(-ENOLCK));
284 flags = LDLM_FL_ATOMIC_CB;
286 res_id.name[2] = full_name_hash((unsigned char *)name, namelen);
288 CDEBUG(D_INFO, "take lock on "DLID4":"LPX64"\n",
289 OLID4(id), res_id.name[2]);
292 #warning "No PDIROPS support in the kernel"
294 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
295 LDLM_IBITS, &policy, lock_mode, &flags,
296 mds_blocking_ast, ldlm_completion_ast,
297 NULL, NULL, NULL, 0, NULL, lockh);
298 if (rc != ELDLM_OK) {
300 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
303 ldlm_lock_decref(lockh + 1, *mode);
305 } else if (de->d_inode && de->d_inode->i_nlink == 0) {
306 /* as sometimes we lookup inode by ino/generation through
307 iopen mechanism, it's possible to find already unlinked
308 inode with nlink == 0. let's interpretate the case as
310 CWARN("found already unlinked inode %lu/%u\n",
311 de->d_inode->i_ino, de->d_inode->i_generation);
313 retval = ERR_PTR(-ENOENT);
314 ldlm_lock_decref(lockh, lock_mode);
317 ldlm_lock_decref(lockh + 1, *mode);
324 #ifndef DCACHE_DISCONNECTED
325 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
328 /* Look up an entry by inode number. This function ONLY returns valid dget'd
329 * dentries with an initialized inode or errors */
330 struct dentry *mds_id2dentry(struct obd_device *obd, struct lustre_id *id,
331 struct vfsmount **mnt)
333 unsigned long ino = (unsigned long)id_ino(id);
334 __u32 generation = (__u32)id_gen(id);
335 struct mds_obd *mds = &obd->u.mds;
336 struct dentry *result;
341 RETURN(ERR_PTR(-ESTALE));
343 snprintf(idname, sizeof(idname), "0x%lx", ino);
345 CDEBUG(D_DENTRY, "--> mds_id2dentry: ino/gen %lu/%u, sb %p\n",
346 ino, generation, mds->mds_sb);
348 /* under ext3 this is neither supposed to return bad inodes nor NULL
350 result = ll_lookup_one_len(idname, mds->mds_id_de,
355 inode = result->d_inode;
357 RETURN(ERR_PTR(-ENOENT));
359 if (is_bad_inode(inode)) {
360 CERROR("bad inode returned %lu/%u\n",
361 inode->i_ino, inode->i_generation);
363 RETURN(ERR_PTR(-ENOENT));
366 /* here we disabled generation check, as root inode i_generation
367 * of cache mds and real mds are different. */
368 if (inode->i_ino != id_ino(&mds->mds_rootid) && generation &&
369 inode->i_generation != generation) {
370 /* we didn't find the right inode.. */
371 if (id_group(id) != mds->mds_num) {
372 CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
373 "%u != %u, mds %u != %u, request to wrong MDS?\n",
374 inode->i_ino, (unsigned long)inode->i_nlink,
375 atomic_read(&inode->i_count), inode->i_generation,
376 generation, mds->mds_num, (unsigned)id_group(id));
378 CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
379 "%u != %u, inode is recreated while request handled?\n",
380 inode->i_ino, (unsigned long)inode->i_nlink,
381 atomic_read(&inode->i_count), inode->i_generation,
385 RETURN(ERR_PTR(-ENOENT));
389 *mnt = mds->mds_vfsmnt;
397 int mds_req_add_idmapping(struct ptlrpc_request *req,
398 struct mds_export_data *med)
400 struct mds_req_sec_desc *rsd;
401 struct lustre_sec_desc *lsd;
404 if (!med->med_remote)
407 /* maybe we should do it more completely: invalidate the gss ctxt? */
408 if (req->rq_mapped_uid == MDS_IDMAP_NOTFOUND) {
409 CWARN("didn't find mapped uid\n");
413 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
415 CERROR("Can't unpack security desc\n");
419 lsd = mds_get_lsd(req->rq_mapped_uid);
421 CERROR("can't get LSD(%u), no mapping added\n",
426 rc = mds_idmap_add(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
427 rsd->rsd_gid, lsd->lsd_gid);
433 int mds_req_del_idmapping(struct ptlrpc_request *req,
434 struct mds_export_data *med)
436 struct mds_req_sec_desc *rsd;
437 struct lustre_sec_desc *lsd;
440 if (!med->med_remote)
443 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
445 CERROR("Can't unpack security desc\n");
449 LASSERT(req->rq_mapped_uid != -1);
450 lsd = mds_get_lsd(req->rq_mapped_uid);
452 CERROR("can't get LSD(%u), no idmapping deleted\n",
457 rc = mds_idmap_del(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
458 rsd->rsd_gid, lsd->lsd_gid);
463 static int mds_init_export_data(struct ptlrpc_request *req,
464 struct mds_export_data *med)
466 struct obd_connect_data *data, *reply;
467 int ask_remote, ask_local;
470 data = lustre_msg_buf(req->rq_reqmsg, 5, sizeof(*data));
471 reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*data));
472 LASSERT(data && reply);
474 if (med->med_initialized) {
475 CDEBUG(D_SEC, "med already initialized, reconnect?\n");
479 ask_remote = data->ocd_connect_flags & OBD_CONNECT_REMOTE;
480 ask_local = data->ocd_connect_flags & OBD_CONNECT_LOCAL;
482 /* currently the policy is simple: satisfy client as possible
485 if (req->rq_auth_uid == -1) {
487 CWARN("null sec is used, force to be local\n");
491 if (!req->rq_remote_realm)
492 CWARN("local realm asked to be remote\n");
494 } else if (ask_local) {
495 if (req->rq_remote_realm)
496 CWARN("remote realm asked to be local\n");
499 med->med_remote = (req->rq_remote_realm != 0);
502 med->med_nllu = data->ocd_nllu[0];
503 med->med_nllg = data->ocd_nllu[1];
505 med->med_initialized = 1;
507 reply->ocd_connect_flags &= ~(OBD_CONNECT_REMOTE | OBD_CONNECT_LOCAL);
508 if (med->med_remote) {
510 med->med_idmap = mds_idmap_alloc();
513 CERROR("Failed to alloc idmap, following request from "
514 "this client will be refused\n");
516 reply->ocd_connect_flags |= OBD_CONNECT_REMOTE;
517 CDEBUG(D_SEC, "set client as remote\n");
519 reply->ocd_connect_flags |= OBD_CONNECT_LOCAL;
520 CDEBUG(D_SEC, "set client as local\n");
526 static void mds_free_export_data(struct mds_export_data *med)
531 LASSERT(med->med_remote);
532 mds_idmap_free(med->med_idmap);
533 med->med_idmap = NULL;
536 /* Establish a connection to the MDS.
538 * This will set up an export structure for the client to hold state data about
539 * that client, like open files, the last operation number it did on the server,
542 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
543 struct obd_uuid *cluuid, struct obd_connect_data *data,
546 struct mds_export_data *med;
547 struct mds_client_data *mcd;
548 struct obd_export *exp;
552 if (!conn || !obd || !cluuid)
555 /* XXX There is a small race between checking the list and adding a new
556 * connection for the same UUID, but the real threat (list corruption
557 * when multiple different clients connect) is solved.
559 * There is a second race between adding the export to the list, and
560 * filling in the client data below. Hence skipping the case of NULL
561 * mcd above. We should already be controlling multiple connects at the
562 * client, and we can't hold the spinlock over memory allocations
563 * without risk of deadlocking.
565 rc = class_connect(conn, obd, cluuid);
568 exp = class_conn2export(conn);
570 LASSERT(exp != NULL);
571 med = &exp->exp_mds_data;
573 OBD_ALLOC(mcd, sizeof(*mcd));
575 CERROR("%s: out of memory for client data.\n",
577 GOTO(out, rc = -ENOMEM);
580 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
583 rc = mds_client_add(obd, &obd->u.mds, med, -1);
591 OBD_FREE(mcd, sizeof(*mcd));
592 class_disconnect(exp, 0);
594 class_export_put(exp);
599 static int mds_connect_post(struct obd_export *exp, unsigned initial,
602 struct obd_device *obd = exp->exp_obd;
603 struct mds_obd *mds = &obd->u.mds;
604 struct mds_export_data *med;
605 struct mds_client_data *mcd;
609 med = &exp->exp_mds_data;
613 /* some one reconnect initially, we have to reset
614 * data existing export can have. bug 6102 */
615 if (mcd->mcd_last_xid != 0)
616 CDEBUG(D_HA, "initial reconnect to existing export\n");
617 mcd->mcd_last_transno = 0;
618 mcd->mcd_last_xid = 0;
619 mcd->mcd_last_close_xid = 0;
620 mcd->mcd_last_result = 0;
621 mcd->mcd_last_data = 0;
624 if (!(flags & OBD_OPT_MDS_CONNECTION)) {
625 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
626 atomic_inc(&mds->mds_real_clients);
627 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
628 obd->obd_name, exp->exp_client_uuid.uuid,
629 atomic_read(&mds->mds_real_clients));
630 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
632 if (mds->mds_md_name)
633 rc = mds_md_connect(obd, mds->mds_md_name);
638 static int mds_init_export(struct obd_export *exp)
640 struct mds_export_data *med = &exp->exp_mds_data;
642 INIT_LIST_HEAD(&med->med_open_head);
643 spin_lock_init(&med->med_open_lock);
647 static int mds_destroy_export(struct obd_export *export)
649 struct obd_device *obd = export->exp_obd;
650 struct mds_export_data *med = &export->exp_mds_data;
651 struct lvfs_run_ctxt saved;
655 mds_free_export_data(med);
656 target_destroy_export(export);
658 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
661 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
663 /* Close any open files (which may also cause orphan unlinking). */
664 spin_lock(&med->med_open_lock);
665 while (!list_empty(&med->med_open_head)) {
666 struct list_head *tmp = med->med_open_head.next;
667 struct mds_file_data *mfd =
668 list_entry(tmp, struct mds_file_data, mfd_list);
669 struct lustre_id sid;
671 BDEVNAME_DECLARE_STORAGE(btmp);
673 /* bug 1579: fix force-closing for 2.5 */
674 struct dentry *dentry = mfd->mfd_dentry;
676 list_del(&mfd->mfd_list);
677 spin_unlock(&med->med_open_lock);
679 down(&dentry->d_inode->i_sem);
680 rc = mds_read_inode_sid(obd, dentry->d_inode, &sid);
681 up(&dentry->d_inode->i_sem);
683 CERROR("Can't read inode self id, inode %lu, "
684 "rc %d\n", dentry->d_inode->i_ino, rc);
685 memset(&sid, 0, sizeof(sid));
688 /* If you change this message, be sure to update
689 * replay_single:test_46 */
690 CERROR("force closing client file handle for %.*s (%s:"
691 DLID4")\n", dentry->d_name.len, dentry->d_name.name,
692 ll_bdevname(dentry->d_inode->i_sb, btmp),
695 /* child inode->i_alloc_sem protects orphan_dec_test and
696 * is_orphan race, mds_mfd_close drops it */
697 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
698 rc = mds_mfd_close(NULL, 0, obd, mfd,
699 !(export->exp_flags & OBD_OPT_FAILOVER));
701 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
702 spin_lock(&med->med_open_lock);
704 spin_unlock(&med->med_open_lock);
705 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
709 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
713 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
715 unsigned long irqflags;
716 struct obd_device *obd;
721 LASSERT(exp != NULL);
722 obd = class_exp2obd(exp);
724 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
725 exp->exp_handle.h_cookie);
731 * suppress any inter-mds requests durring disconnecting lmv if this is
732 * detected --force mode. This is needed to avoid endless recovery.
734 if (atomic_read(&mds->mds_real_clients) > 0 &&
735 !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
736 flags |= OBD_OPT_FORCE;
738 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
739 && !atomic_read(&mds->mds_real_clients)) {
740 /* there was no client at all */
741 mds_md_disconnect(obd, flags);
744 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
745 && atomic_dec_and_test(&mds->mds_real_clients)) {
746 /* time to drop LMV connections */
747 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
748 "Disconnnect from LMV now\n",
749 obd->obd_name, exp->exp_client_uuid.uuid);
750 mds_md_disconnect(obd, flags);
753 spin_lock_irqsave(&exp->exp_lock, irqflags);
754 exp->exp_flags = flags;
755 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
757 /* disconnect early so that clients can't keep using export */
758 rc = class_disconnect(exp, flags);
759 ldlm_cancel_locks_for_export(exp);
761 /* complete all outstanding replies */
762 spin_lock_irqsave(&exp->exp_lock, irqflags);
763 while (!list_empty(&exp->exp_outstanding_replies)) {
764 struct ptlrpc_reply_state *rs =
765 list_entry(exp->exp_outstanding_replies.next,
766 struct ptlrpc_reply_state, rs_exp_list);
767 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
769 spin_lock(&svc->srv_lock);
770 list_del_init(&rs->rs_exp_list);
771 ptlrpc_schedule_difficult_reply(rs);
772 spin_unlock(&svc->srv_lock);
774 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
778 static int mds_getstatus(struct ptlrpc_request *req)
780 struct mds_obd *mds = mds_req2mds(req);
781 struct mds_body *body;
785 size = sizeof(*body);
787 rc = lustre_pack_reply(req, 1, &size, NULL);
788 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
789 CERROR("mds: out of memory for message: size=%d\n", size);
790 req->rq_status = -ENOMEM; /* superfluous? */
794 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
795 body->valid |= OBD_MD_FID;
797 memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
800 * the last_committed and last_xid fields are filled in for all replies
801 * already - no need to do so here also.
806 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
807 void *data, int flag)
812 if (flag == LDLM_CB_CANCELING) {
813 /* Don't need to do anything here. */
817 /* XXX layering violation! -phil */
818 lock_res_and_lock(lock);
821 * get this: if mds_blocking_ast is racing with mds_intent_policy, such
822 * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
823 * then by the time we get the lock, we might not be the correct
824 * blocking function anymore. So check, and return early, if so.
826 if (lock->l_blocking_ast != mds_blocking_ast) {
827 unlock_res_and_lock(lock);
831 lock->l_flags |= LDLM_FL_CBPENDING;
832 do_ast = (!lock->l_readers && !lock->l_writers);
833 unlock_res_and_lock(lock);
836 struct lustre_handle lockh;
839 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
840 ldlm_lock2handle(lock, &lockh);
841 rc = ldlm_cli_cancel(&lockh);
843 CERROR("ldlm_cli_cancel: %d\n", rc);
845 LDLM_DEBUG(lock, "Lock still has references, will be "
851 static int mds_convert_md(struct obd_device *obd, struct inode *inode,
852 void *md, int size, int mea)
856 if (S_ISREG(inode->i_mode)) {
857 rc = mds_convert_lov_ea(obd, inode, md, size);
858 } else if (S_ISDIR(inode->i_mode)) {
860 rc = mds_convert_mea_ea(obd, inode, md, size);
862 rc = mds_convert_lov_ea(obd, inode, md, size);
865 CERROR("Invalid EA format (nor LOV or MEA) "
866 "is detected. Inode %lu/%u\n",
867 inode->i_ino, inode->i_generation);
873 int mds_get_md(struct obd_device *obd, struct inode *inode,
874 void *md, int *size, int lock, int mea)
883 rc = fsfilt_get_md(obd, inode, md, *size,
884 (mea ? EA_MEA : EA_LOV));
886 CERROR("Error %d reading eadata for ino %lu\n",
890 rc = mds_convert_md(obd, inode, md,
905 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
906 * Call with lock=0 if the caller has already taken the i_sem. */
907 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
908 struct mds_body *body, struct inode *inode, int lock, int mea)
910 struct mds_obd *mds = &obd->u.mds;
915 lmm = lustre_msg_buf(msg, offset, 0);
917 /* Some problem with getting eadata when I sized the reply
919 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
923 lmm_size = msg->buflens[offset];
925 /* I don't really like this, but it is a sanity check on the client
926 * MD request. However, if the client doesn't know how much space
927 * to reserve for the MD, it shouldn't be bad to have too much space.
929 if (lmm_size > mds->mds_max_mdsize) {
930 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
931 inode->i_ino, lmm_size, mds->mds_max_mdsize);
935 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, mea);
937 body->valid |= S_ISDIR(inode->i_mode) ?
938 OBD_MD_FLDIREA : OBD_MD_FLEASIZE;
941 body->valid |= OBD_MD_MEA;
943 body->eadatasize = lmm_size;
950 int mds_pack_link(struct dentry *dentry, struct ptlrpc_request *req,
951 struct mds_body *repbody, int reply_off)
953 struct inode *inode = dentry->d_inode;
958 symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
959 LASSERT(symname != NULL);
960 len = req->rq_repmsg->buflens[reply_off + 1];
962 rc = inode->i_op->readlink(dentry, symname, len);
964 CERROR("readlink failed: %d\n", rc);
965 } else if (rc != len - 1) {
966 CERROR ("Unexpected readlink rc %d: expecting %d\n",
970 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
971 repbody->valid |= OBD_MD_LINKNAME;
972 repbody->eadatasize = rc + 1;
973 symname[rc] = 0; /* NULL terminate */
980 int mds_pack_xattr(struct dentry *dentry, struct ptlrpc_request *req,
981 struct mds_body *repbody, int req_off, int reply_off)
983 struct inode *inode = dentry->d_inode;
989 ea_name = lustre_msg_string(req->rq_reqmsg, req_off + 1, 0);
990 len = req->rq_repmsg->buflens[reply_off + 1];
992 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
996 if (!strcmp(ea_name, XATTR_NAME_LUSTRE_ACL)) {
997 struct rmtacl_upcall_desc desc;
999 if (len != LUSTRE_ACL_SIZE_MAX || !value) {
1000 CERROR("no reply buffer prepared\n");
1004 memset(&desc, 0, sizeof(desc));
1006 desc.cmd = lustre_msg_string(req->rq_reqmsg, req_off + 2, 0);
1007 desc.cmdlen = req->rq_reqmsg->buflens[req_off + 2];
1008 desc.res = (char *) value;
1009 desc.reslen = LUSTRE_ACL_SIZE_MAX;
1011 mds_do_remote_acl_upcall(&desc);
1013 if (desc.upcall_status)
1014 RETURN(desc.upcall_status);
1016 if (desc.reslen > LUSTRE_ACL_SIZE_MAX) {
1017 CERROR("downcall claim reslen %u\n", desc.reslen);
1020 /* like remote setfacl, steal "flags" in mds_body as the
1021 * exececution status
1023 repbody->flags = desc.status;
1024 repbody->valid |= OBD_MD_FLXATTR;
1025 repbody->eadatasize = desc.reslen;
1030 if (inode->i_op && inode->i_op->getxattr)
1031 rc = inode->i_op->getxattr(dentry, ea_name, value, len);
1034 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1035 CERROR("getxattr failed: %d", rc);
1037 repbody->valid |= OBD_MD_FLXATTR;
1038 repbody->eadatasize = rc;
1045 int mds_pack_xattr_list(struct dentry *dentry, struct ptlrpc_request *req,
1046 struct mds_body *repbody, int reply_off)
1048 struct inode *inode = dentry->d_inode;
1053 len = req->rq_repmsg->buflens[reply_off + 1];
1055 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
1058 if (inode->i_op && inode->i_op->getxattr)
1059 rc = inode->i_op->listxattr(dentry, value, len);
1062 CERROR("listxattr failed: %d", rc);
1064 repbody->valid |= OBD_MD_FLXATTRLIST;
1065 repbody->eadatasize = rc;
1072 int mds_pack_posix_acl(struct lustre_msg *repmsg, int *offset,
1073 struct mds_body *body, struct inode *inode)
1075 struct dentry de = { .d_inode = inode };
1076 __u32 buflen, *sizep;
1078 int size, pack_off = *offset;
1081 sizep = lustre_msg_buf(repmsg, pack_off++, 4);
1083 CERROR("can't locate returned acl size buf\n");
1087 if (!inode->i_op->getxattr)
1090 buflen = repmsg->buflens[pack_off];
1091 buf = lustre_msg_buf(repmsg, pack_off++, buflen);
1093 size = inode->i_op->getxattr(&de, XATTR_NAME_ACL_ACCESS, buf, buflen);
1094 if (size == -ENODATA || size == -EOPNOTSUPP)
1099 body->valid |= OBD_MD_FLACL;
1101 *sizep = cpu_to_le32(size);
1107 int mds_pack_remote_perm(struct ptlrpc_request *req, int *reply_off,
1108 struct mds_body *body, struct inode *inode)
1110 struct lustre_sec_desc *lsd;
1111 struct mds_remote_perm *perm;
1112 int pack_off = *reply_off;
1115 LASSERT(inode->i_op);
1116 LASSERT(inode->i_op->permission);
1117 LASSERT(req->rq_export->exp_mds_data.med_remote);
1119 perm = (struct mds_remote_perm *)
1120 lustre_msg_buf(req->rq_repmsg, pack_off++, sizeof(perm));
1124 memset(perm, 0, sizeof(*perm));
1126 /* obtain authenticated uid/gid and LSD permissions, which
1127 * might be different from current process context, from LSD
1129 lsd = mds_get_lsd(current->uid);
1131 CWARN("can't LSD of uid %u\n", current->uid);
1135 perm->mrp_auth_uid = lsd->lsd_uid;
1136 perm->mrp_auth_gid = lsd->lsd_gid;
1138 lsd_perms = mds_lsd_get_perms(lsd, 1, 0, req->rq_peer.peer_id.nid);
1139 if (lsd_perms & LSD_PERM_SETUID)
1140 perm->mrp_allow_setuid = 1;
1141 if (lsd_perms & LSD_PERM_SETGID)
1142 perm->mrp_allow_setgid = 1;
1146 /* permission bits of current user
1147 * XXX this is low efficient, could we do it in one blow?
1149 if (inode->i_op->permission(inode, MAY_EXEC, NULL) == 0)
1150 perm->mrp_perm |= MAY_EXEC;
1151 if (inode->i_op->permission(inode, MAY_WRITE, NULL) == 0)
1152 perm->mrp_perm |= MAY_WRITE;
1153 if (inode->i_op->permission(inode, MAY_READ, NULL) == 0)
1154 perm->mrp_perm |= MAY_READ;
1156 body->valid |= (OBD_MD_FLACL | OBD_MD_FLRMTACL);
1158 *reply_off = pack_off;
1163 int mds_pack_acl(struct ptlrpc_request *req, int *reply_off,
1164 struct mds_body *body, struct inode *inode)
1168 if (!req->rq_export->exp_mds_data.med_remote)
1169 rc = mds_pack_posix_acl(req->rq_repmsg, reply_off, body, inode);
1171 rc = mds_pack_remote_perm(req, reply_off, body, inode);
1176 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
1177 struct ptlrpc_request *req, int req_off,
1178 struct mds_body *reqbody, int reply_off,
1179 struct mds_req_sec_desc *rsd)
1181 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1182 struct inode *inode = dentry->d_inode;
1183 struct mds_body *body;
1184 int rc = 0, offset = 0;
1187 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
1190 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
1191 LASSERT(body != NULL); /* caller prepped reply */
1193 if (dentry->d_flags & DCACHE_CROSS_REF) {
1194 mds_pack_dentry2body(obd, body, dentry,
1195 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1196 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
1201 mds_pack_inode2body(obd, body, inode,
1202 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1204 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
1205 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
1207 /* guessing what kind og attribute do we need. */
1208 int is_mea = (S_ISDIR(inode->i_mode) &&
1209 (reqbody->valid & OBD_MD_MEA) != 0);
1211 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1,
1212 body, inode, 1, is_mea);
1214 /* if we have LOV EA data, the OST holds size, atime, mtime. */
1215 if (!(body->valid & OBD_MD_FLEASIZE) &&
1216 !(body->valid & OBD_MD_FLDIREA))
1217 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1218 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1219 } else if (S_ISLNK(inode->i_mode) &&
1220 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
1221 rc = mds_pack_link(dentry, req, body, reply_off);
1222 } else if (reqbody->valid & OBD_MD_FLXATTR) {
1223 rc = mds_pack_xattr(dentry, req, body, req_off, reply_off);
1224 } else if (reqbody->valid & OBD_MD_FLXATTRLIST) {
1225 rc = mds_pack_xattr_list(dentry, req, body, reply_off);
1228 offset = reply_off + ((reqbody->valid & OBD_MD_FLEASIZE) ? 2 : 1);
1229 if (reqbody->valid & OBD_MD_FLACL) {
1230 rc = mds_pack_acl(req, &offset, body, inode);
1233 if (reqbody->valid & OBD_MD_FLKEY) {
1234 rc = mds_pack_gskey(obd, req->rq_repmsg, &offset,
1238 mds_pack_audit(obd, inode, body);
1241 mds_body_do_reverse_map(med, body);
1246 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
1247 struct dentry *dentry,
1250 int rc = 0, size[1] = {sizeof(struct mds_body)};
1253 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1254 CERROR("failed MDS_GETATTR_PACK test\n");
1255 req->rq_status = -ENOMEM;
1259 rc = lustre_pack_reply(req, 1, size, NULL);
1261 CERROR("lustre_pack_reply failed: rc %d\n", rc);
1262 GOTO(out, req->rq_status = rc);
1270 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct dentry *de,
1273 struct inode *inode = de->d_inode;
1274 struct mds_obd *mds = mds_req2mds(req);
1275 struct mds_body *body;
1276 int rc = 0, size[4] = {sizeof(*body)}, bufcount = 1;
1279 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1280 LASSERT(body != NULL); /* checked by caller */
1281 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
1283 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
1284 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
1287 down(&inode->i_sem);
1288 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
1289 ((body->valid & OBD_MD_MEA) ? EA_MEA : EA_LOV));
1292 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1293 CERROR("error getting inode %lu MD: rc = %d\n",
1296 } else if (rc > mds->mds_max_mdsize) {
1298 CERROR("MD size %d larger than maximum possible %u\n",
1299 rc, mds->mds_max_mdsize);
1301 size[bufcount] = rc;
1304 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
1305 if (inode->i_size + 1 != body->eadatasize)
1306 CERROR("symlink size: %Lu, reply space: %d\n",
1307 inode->i_size + 1, body->eadatasize);
1308 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
1310 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
1311 inode->i_size + 1, body->eadatasize);
1312 } else if ((body->valid & OBD_MD_FLXATTR)) {
1313 char *ea_name = lustre_msg_string(req->rq_reqmsg,
1317 if (!strcmp(ea_name, XATTR_NAME_LUSTRE_ACL)) {
1318 size[bufcount] = LUSTRE_ACL_SIZE_MAX;
1320 if (inode->i_op && inode->i_op->getxattr)
1321 rc = inode->i_op->getxattr(de, ea_name,
1325 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1326 CERROR("error get inode %lu EA: %d\n",
1330 size[bufcount] = min_t(int,
1331 body->eadatasize, rc);
1335 } else if (body->valid & OBD_MD_FLXATTRLIST) {
1337 if (inode->i_op && inode->i_op->getxattr)
1338 rc = inode->i_op->listxattr(de, NULL, 0);
1341 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1342 CERROR("error getting inode %lu EA: rc = %d\n",
1346 size[bufcount] = min_t(int, body->eadatasize, rc);
1351 /* may co-exist with OBD_MD_FLEASIZE */
1352 if (body->valid & OBD_MD_FLACL) {
1353 if (req->rq_export->exp_mds_data.med_remote) {
1354 size[bufcount++] = sizeof(struct mds_remote_perm);
1356 size[bufcount++] = sizeof(int);
1357 size[bufcount++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
1361 if (body->valid & OBD_MD_FLKEY) {
1362 size[bufcount++] = sizeof(int);
1363 size[bufcount++] = sizeof(struct crypto_key);
1366 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1367 CERROR("failed MDS_GETATTR_PACK test\n");
1368 req->rq_status = -ENOMEM;
1369 GOTO(out, rc = -ENOMEM);
1372 rc = lustre_pack_reply(req, bufcount, size, NULL);
1374 CERROR("out of memory\n");
1375 GOTO(out, req->rq_status = rc);
1383 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
1384 char *name, int namelen)
1386 struct mea *mea = NULL;
1387 int mea_size, rc = 0;
1390 rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
1393 if (mea != NULL && mea->mea_count) {
1395 * dir is already splitted, check if requested filename should
1396 * live at this MDS or at another one.
1398 int i = mea_name2idx(mea, name, namelen - 1);
1399 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1401 "inapropriate MDS(%d) for %s. should be "
1402 "%lu(%d)\n", mea->mea_master, name,
1403 (unsigned long)id_group(&mea->mea_ids[i]), i);
1409 OBD_FREE(mea, mea_size);
1413 int mds_getattr_size(struct obd_device *obd, struct dentry *dentry,
1414 struct ptlrpc_request *req, struct mds_body *body)
1416 struct inode *inode = dentry->d_inode;
1419 LASSERT(body != NULL);
1421 if (dentry->d_inode == NULL || !S_ISREG(inode->i_mode))
1424 if (obd->obd_recovering) {
1425 CDEBUG(D_INODE, "size for "DLID4" is unknown yet (recovering)\n",
1430 if (atomic_read(&inode->i_writecount)) {
1431 /* some one has opened the file for write.
1432 * mds doesn't know actual size */
1433 CDEBUG(D_INODE, "MDS doesn't know actual size for "DLID4"\n",
1437 CDEBUG(D_INODE, "MDS returns "LPD64"/"LPD64" for"DLID4"\n",
1438 body->size, body->blocks, OLID4(&body->id1));
1439 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1443 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
1444 struct lustre_handle *child_lockh, int child_part)
1446 struct obd_device *obd = req->rq_export->exp_obd;
1447 struct mds_obd *mds = &obd->u.mds;
1448 struct ldlm_reply *rep = NULL;
1449 struct lvfs_run_ctxt saved;
1450 struct mds_req_sec_desc *rsd;
1451 struct mds_body *body;
1452 struct dentry *dparent = NULL, *dchild = NULL;
1453 struct lvfs_ucred uc = {NULL, NULL,};
1454 struct lustre_handle parent_lockh[2] = {{0}, {0}};
1455 unsigned int namesize;
1456 int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
1460 LASSERT(!strcmp(obd->obd_type->typ_name, OBD_MDS_DEVICENAME));
1461 MD_COUNTER_INCREMENT(obd, getattr_lock);
1463 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1465 CERROR("Can't unpack security desc\n");
1469 /* swab now, before anyone looks inside the request. */
1470 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1471 lustre_swab_mds_body);
1473 CERROR("Can't swab mds_body\n");
1474 GOTO(cleanup, rc = -EFAULT);
1477 LASSERT_REQSWAB(req, offset + 1);
1478 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
1480 CERROR("Can't unpack name\n");
1481 GOTO(cleanup, rc = -EFAULT);
1483 namesize = req->rq_reqmsg->buflens[offset + 1];
1485 /* namesize less than 2 means we have empty name, probably came from
1486 revalidate by cfid, so no point in having name to be set */
1490 LASSERT (offset == 1 || offset == 3);
1492 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1498 rc = mds_init_ucred(&uc, req, rsd);
1500 if (child_lockh->cookie == 0)
1501 mds_audit_auth(req, &uc, AUDIT_STAT, &body->id1,
1506 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1507 cleanup_phase = 1; /* kernel context */
1508 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1510 LASSERT(namesize > 0);
1511 if (child_lockh->cookie != 0) {
1512 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1516 if (body->valid == OBD_MD_FLID) {
1517 struct mds_body *mds_reply;
1518 int size = sizeof(*mds_reply);
1522 dparent = mds_id2dentry(obd, &body->id1, NULL);
1523 if (IS_ERR(dparent)) {
1524 rc = PTR_ERR(dparent);
1528 * the user requested ONLY the inode number, so do a raw lookup.
1530 rc = lustre_pack_reply(req, 1, &size, NULL);
1532 CERROR("out of memory\n");
1536 dir = dparent->d_inode;
1537 LASSERT(dir->i_op->lookup_raw != NULL);
1538 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1540 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1541 sizeof(*mds_reply));
1543 id_ino(&mds_reply->id1) = inum;
1544 mds_reply->valid = OBD_MD_FLID;
1548 if (resent_req == 0) {
1549 LASSERT(id_fid(&body->id1) != 0);
1551 rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1552 parent_lockh, &dparent,
1554 MDS_INODELOCK_UPDATE,
1557 child_lockh, &dchild,
1558 LCK_PR, child_part);
1562 cleanup_phase = 2; /* dchild, dparent, locks */
1565 * let's make sure this name should leave on this mds
1568 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1572 /* we have no dentry here, drop LOOKUP bit */
1573 /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1574 child_part &= ~MDS_INODELOCK_LOOKUP;
1575 CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1576 obd->obd_name, OLID4(&body->id1));
1578 dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1579 LCK_PR, parent_lockh,
1582 MDS_INODELOCK_UPDATE);
1583 if (IS_ERR(dchild)) {
1584 CERROR("can't find inode with id "DLID4", err = %d\n",
1585 OLID4(&body->id1), (int)PTR_ERR(dchild));
1586 GOTO(cleanup, rc = PTR_ERR(dchild));
1588 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1591 struct ldlm_lock *granted_lock;
1593 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1594 granted_lock = ldlm_handle2lock(child_lockh);
1596 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1597 id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1598 child_lockh->cookie);
1601 /* usual named request */
1602 dparent = mds_id2dentry(obd, &body->id1, NULL);
1603 LASSERT(!IS_ERR(dparent));
1604 dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1605 LASSERT(!IS_ERR(dchild));
1607 /* client wants to get attr. by id */
1608 dchild = mds_id2dentry(obd, &body->id1, NULL);
1609 LASSERT(!IS_ERR(dchild));
1611 LDLM_LOCK_PUT(granted_lock);
1614 cleanup_phase = 2; /* dchild, dparent, locks */
1616 if (!DENTRY_VALID(dchild)) {
1617 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1619 * in the intent case, the policy clears this error: the
1620 * disposition is enough.
1625 intent_set_disposition(rep, DISP_LOOKUP_POS);
1628 if (req->rq_repmsg == NULL) {
1629 if (dchild->d_flags & DCACHE_CROSS_REF)
1630 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1632 rc = mds_getattr_pack_msg(req, dchild, offset);
1634 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1639 rc = mds_getattr_internal(obd, dchild, req, offset, body,
1642 GOTO(cleanup, rc); /* returns the lock to the client */
1644 /* probably MDS knows actual size? */
1645 body = lustre_msg_buf(req->rq_repmsg, reply_offset, sizeof(*body));
1646 LASSERT(body != NULL);
1647 mds_getattr_size(obd, dchild, req, body);
1652 switch (cleanup_phase) {
1654 if (resent_req == 0) {
1655 if (rc && DENTRY_VALID(dchild))
1656 ldlm_lock_decref(child_lockh, LCK_PR);
1658 ldlm_lock_decref(parent_lockh, LCK_PR);
1660 if (parent_lockh[1].cookie != 0)
1661 ldlm_lock_decref(parent_lockh + 1, update_mode);
1666 /* audit stuff for getattr */
1667 if (dchild->d_inode)
1668 mds_audit_stat(req, &body->id1, dchild, rc);
1672 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1674 mds_exit_ucred(&uc);
1679 static int mds_getattr(struct ptlrpc_request *req, int offset)
1681 struct obd_device *obd = req->rq_export->exp_obd;
1682 struct lvfs_run_ctxt saved;
1684 struct mds_req_sec_desc *rsd;
1685 struct mds_body *body;
1686 struct lvfs_ucred uc = {NULL, NULL,};
1690 MD_COUNTER_INCREMENT(obd, getattr);
1692 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1694 CERROR("Can't unpack security desc\n");
1698 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1699 lustre_swab_mds_body);
1701 CERROR ("Can't unpack body\n");
1705 rc = mds_init_ucred(&uc, req, rsd);
1707 mds_exit_ucred(&uc);
1711 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1712 de = mds_id2dentry(obd, &body->id1, NULL);
1714 rc = req->rq_status = PTR_ERR(de);
1718 rc = mds_getattr_pack_msg(req, de, offset);
1720 CERROR("mds_getattr_pack_msg: %d\n", rc);
1724 req->rq_status = mds_getattr_internal(obd, de, req, offset, body,
1730 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1731 mds_exit_ucred(&uc);
1734 static int mds_access_check(struct ptlrpc_request *req, int offset)
1736 struct obd_device *obd = req->rq_export->exp_obd;
1737 struct lvfs_run_ctxt saved;
1739 struct mds_req_sec_desc *rsd;
1740 struct mds_body *body;
1741 struct lvfs_ucred uc;
1742 int rep_size[2] = {sizeof(*body),
1743 sizeof(struct mds_remote_perm)};
1744 int rc = 0, rep_offset;
1747 if (!req->rq_export->exp_mds_data.med_remote) {
1748 CERROR("from local client "LPU64"\n", req->rq_peer.peer_id.nid);
1752 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1754 CERROR("Can't unpack security desc\n");
1758 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1759 lustre_swab_mds_body);
1761 CERROR ("Can't unpack body\n");
1765 MD_COUNTER_INCREMENT(obd, access_check);
1767 rc = mds_init_ucred(&uc, req, rsd);
1769 CERROR("init ucred error: %d\n", rc);
1772 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1774 de = mds_id2dentry(obd, &body->id1, NULL);
1776 CERROR("grab ino "LPU64": err %ld\n",
1777 body->id1.li_stc.u.e3s.l3s_ino, PTR_ERR(de));
1778 GOTO(out_pop, rc = PTR_ERR(de));
1781 rc = lustre_pack_reply(req, 2, rep_size, NULL);
1783 CERROR("pack reply error: %d\n", rc);
1784 GOTO(out_dput, rc = -EINVAL);
1787 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1791 rc = mds_pack_remote_perm(req, &rep_offset, body, de->d_inode);
1798 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1799 mds_exit_ucred(&uc);
1803 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1804 unsigned long max_age)
1809 spin_lock(&obd->obd_osfs_lock);
1810 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1812 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1813 spin_unlock(&obd->obd_osfs_lock);
1818 static int mds_statfs(struct ptlrpc_request *req)
1820 struct obd_device *obd = req->rq_export->exp_obd;
1821 int rc, size = sizeof(struct obd_statfs);
1824 /* This will trigger a watchdog timeout */
1825 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1826 (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1828 rc = lustre_pack_reply(req, 1, &size, NULL);
1829 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1830 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1834 OBD_COUNTER_INCREMENT(obd, statfs);
1836 /* We call this so that we can cache a bit - 1 jiffie worth */
1837 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1840 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1846 req->rq_status = rc;
1850 static int mds_sync(struct ptlrpc_request *req, int offset)
1852 struct obd_device *obd = req->rq_export->exp_obd;
1853 struct mds_obd *mds = &obd->u.mds;
1854 struct mds_body *body;
1855 int rc, size = sizeof(*body);
1858 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1859 lustre_swab_mds_body);
1861 GOTO(out, rc = -EPROTO);
1863 rc = lustre_pack_reply(req, 1, &size, NULL);
1864 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1865 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1869 if (id_ino(&body->id1) == 0) {
1870 /* an id of zero is taken to mean "sync whole filesystem" */
1871 rc = fsfilt_sync(obd, mds->mds_sb);
1875 /* just any file to grab fsync method - "file" arg unused */
1876 struct file *file = mds->mds_rcvd_filp;
1877 struct mds_body *rep_body;
1880 de = mds_id2dentry(obd, &body->id1, NULL);
1882 GOTO(out, rc = PTR_ERR(de));
1884 rc = file->f_op->fsync(NULL, de, 1);
1888 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1889 mds_pack_inode2body(obd, rep_body, de->d_inode,
1890 (body->valid & OBD_MD_FID) ? 1 : 0);
1896 req->rq_status = rc;
1900 /* mds_readpage does not take a DLM lock on the inode, because the client must
1901 * already have a PR lock.
1903 * If we were to take another one here, a deadlock will result, if another
1904 * thread is already waiting for a PW lock. */
1905 static int mds_readpage(struct ptlrpc_request *req, int offset)
1907 struct obd_device *obd = req->rq_export->exp_obd;
1908 struct vfsmount *mnt;
1911 struct mds_req_sec_desc *rsd;
1912 struct mds_body *body, *repbody;
1913 struct lvfs_run_ctxt saved;
1914 int rc, size = sizeof(*repbody);
1915 struct lvfs_ucred uc = {NULL, NULL,};
1918 rc = lustre_pack_reply(req, 1, &size, NULL);
1919 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1920 CERROR("mds: out of memory\n");
1921 GOTO(out, rc = -ENOMEM);
1924 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1926 CERROR("Can't unpack security desc\n");
1927 GOTO (out, rc = -EFAULT);
1930 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1931 lustre_swab_mds_body);
1933 CERROR("Can't unpack body\n");
1934 GOTO (out, rc = -EFAULT);
1937 rc = mds_init_ucred(&uc, req, rsd);
1939 mds_audit_auth(req, &uc, AUDIT_READDIR, &body->id1,
1944 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1945 de = mds_id2dentry(obd, &body->id1, &mnt);
1947 GOTO(out_pop, rc = PTR_ERR(de));
1949 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1951 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1952 /* note: in case of an error, dentry_open puts dentry */
1954 GOTO(out_pop, rc = PTR_ERR(file));
1956 /* body->size is actually the offset -eeb */
1957 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1958 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1959 body->size, de->d_inode->i_blksize);
1960 GOTO(out_file, rc = -EFAULT);
1963 /* body->nlink is actually the #bytes to read -eeb */
1964 if (body->nlink & (de->d_inode->i_blksize - 1)) {
1965 CERROR("size %u is not multiple of blocksize %lu\n",
1966 body->nlink, de->d_inode->i_blksize);
1967 GOTO(out_file, rc = -EFAULT);
1970 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1971 repbody->size = file->f_dentry->d_inode->i_size;
1972 repbody->valid = OBD_MD_FLSIZE;
1974 /* to make this asynchronous make sure that the handling function
1975 doesn't send a reply when this function completes. Instead a
1976 callback function would send the reply */
1977 /* body->size is actually the offset -eeb */
1978 rc = mds_sendpage(req, file, body->size, body->nlink);
1982 filp_close(file, 0);
1984 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1986 mds_exit_ucred(&uc);
1987 req->rq_status = rc;
1991 /* update master MDS ID, which is stored in local inode EA. */
1992 int mds_update_mid(struct obd_device *obd, struct lustre_id *id,
1993 void *data, int data_len)
1995 struct mds_obd *mds = &obd->u.mds;
1996 struct dentry *dentry;
2004 dentry = mds_id2dentry(obd, id, NULL);
2006 GOTO(out, rc = PTR_ERR(dentry));
2008 if (!dentry->d_inode) {
2009 CERROR("Can't find object "DLID4".\n",
2011 GOTO(out_dentry, rc = -EINVAL);
2014 handle = fsfilt_start(obd, dentry->d_inode,
2015 FSFILT_OP_SETATTR, NULL);
2017 GOTO(out_dentry, rc = PTR_ERR(handle));
2019 rc = mds_update_inode_mid(obd, dentry->d_inode, handle,
2020 (struct lustre_id *)data);
2022 CERROR("Can't update inode "DLID4" master id, "
2023 "error = %d.\n", OLID4(id), rc);
2024 GOTO(out_commit, rc);
2029 fsfilt_commit(obd, mds->mds_sb, dentry->d_inode,
2036 EXPORT_SYMBOL(mds_update_mid);
2038 /* read master MDS ID, which is stored in local inode EA. */
2039 int mds_read_mid(struct obd_device *obd, struct lustre_id *id,
2040 void *data, int data_len)
2042 struct dentry *dentry;
2049 dentry = mds_id2dentry(obd, id, NULL);
2051 GOTO(out, rc = PTR_ERR(dentry));
2053 if (!dentry->d_inode) {
2054 CERROR("Can't find object "DLID4".\n",
2056 GOTO(out_dentry, rc = -EINVAL);
2059 down(&dentry->d_inode->i_sem);
2060 rc = mds_read_inode_mid(obd, dentry->d_inode,
2061 (struct lustre_id *)data);
2062 up(&dentry->d_inode->i_sem);
2064 CERROR("Can't read inode "DLID4" master id, "
2065 "error = %d.\n", OLID4(id), rc);
2066 GOTO(out_dentry, rc);
2075 EXPORT_SYMBOL(mds_read_mid);
2077 int mds_read_md(struct obd_device *obd, struct lustre_id *id,
2078 char **data, int *datalen)
2080 struct dentry *dentry;
2081 struct mds_obd *mds = &obd->u.mds;
2082 int rc = 0, mea = 0;
2089 dentry = mds_id2dentry(obd, id, NULL);
2091 GOTO(out, rc = PTR_ERR(dentry));
2093 if (!dentry->d_inode) {
2094 CERROR("Can't find object "DLID4".\n",
2096 GOTO(out_dentry, rc = -EINVAL);
2098 if (S_ISDIR(dentry->d_inode->i_mode)) {
2099 *datalen = obd_packmd(mds->mds_md_exp, NULL, NULL);
2102 *datalen = obd_packmd(mds->mds_dt_exp, NULL, NULL);
2105 OBD_ALLOC(ea, *datalen);
2108 GOTO(out_dentry, rc = PTR_ERR(dentry));
2111 down(&dentry->d_inode->i_sem);
2112 rc = fsfilt_get_md(obd, dentry->d_inode, *data, *datalen,
2113 (mea ? EA_MEA : EA_LOV));
2114 up(&dentry->d_inode->i_sem);
2117 CERROR("Error %d reading eadata for ino %lu\n",
2118 rc, dentry->d_inode->i_ino);
2124 EXPORT_SYMBOL(mds_read_md);
2126 int mds_reint(struct ptlrpc_request *req, int offset,
2127 struct lustre_handle *lockh)
2129 struct mds_update_record *rec;
2130 struct mds_req_sec_desc *rsd;
2134 OBD_ALLOC(rec, sizeof(*rec));
2138 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
2140 CERROR("Can't unpack security desc\n");
2141 GOTO(out, rc = -EFAULT);
2144 rc = mds_update_unpack(req, offset, rec);
2145 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
2146 CERROR("invalid record\n");
2147 GOTO(out, req->rq_status = -EINVAL);
2150 rc = mds_init_ucred(&rec->ur_uc, req, rsd);
2152 audit_op code = AUDIT_NONE;
2153 char * au_name = NULL;
2155 switch (rec->ur_opcode) {
2157 code = AUDIT_SETATTR;
2160 code = AUDIT_CREATE;
2161 au_name = rec->ur_name;
2162 au_len = rec->ur_namelen;
2168 code = AUDIT_UNLINK;
2171 code = AUDIT_RENAME;
2174 au_name = rec->ur_name;
2175 au_len = rec->ur_namelen;
2179 CERROR("Wrong opcode in reint\n");
2183 mds_audit_auth(req, &rec->ur_uc, code, rec->ur_id1,
2188 /* rc will be used to interrupt a for loop over multiple records */
2189 rc = mds_reint_rec(rec, offset, req, lockh);
2191 /* audit stuff for OPEN */
2192 if (offset == 3 && rec->ur_opcode == REINT_OPEN)
2193 mds_audit_open(req, rec, rc);
2196 mds_exit_ucred(&rec->ur_uc);
2197 OBD_FREE(rec, sizeof(*rec));
2201 static int mds_filter_recovery_request(struct ptlrpc_request *req,
2202 struct obd_device *obd, int *process)
2204 switch (req->rq_reqmsg->opc) {
2205 case MDS_CONNECT: /* This will never get here, but for completeness. */
2206 case OST_CONNECT: /* This will never get here, but for completeness. */
2207 case MDS_DISCONNECT:
2208 case OST_DISCONNECT:
2213 case MDS_SYNC: /* used in unmounting */
2218 *process = target_queue_recovery_request(req, obd);
2222 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
2224 /* XXX what should we set rq_status to here? */
2225 req->rq_status = -EAGAIN;
2226 RETURN(ptlrpc_error(req));
2230 static char *reint_names[] = {
2231 [REINT_SETATTR] "setattr",
2232 [REINT_CREATE] "create",
2233 [REINT_LINK] "link",
2234 [REINT_UNLINK] "unlink",
2235 [REINT_RENAME] "rename",
2236 [REINT_OPEN] "open",
2239 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER | \
2240 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
2241 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
2244 static void reconstruct_create(struct ptlrpc_request *req)
2246 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2247 struct mds_client_data *mcd = med->med_mcd;
2248 struct dentry *dentry;
2249 struct ost_body *body;
2250 struct lustre_id id;
2254 /* copy rc, transno and disp; steal locks */
2255 mds_req_from_mcd(req, mcd);
2256 if (req->rq_status) {
2264 id_ino(&id) = mcd->mcd_last_data;
2265 LASSERT(id_ino(&id) != 0);
2267 dentry = mds_id2dentry(req2obd(req), &id, NULL);
2268 if (IS_ERR(dentry)) {
2269 CERROR("can't find inode "LPU64"\n", id_ino(&id));
2270 req->rq_status = PTR_ERR(dentry);
2275 CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
2276 req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
2277 dentry->d_inode->i_generation);
2279 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2280 obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
2281 body->oa.o_id = dentry->d_inode->i_ino;
2282 body->oa.o_generation = dentry->d_inode->i_generation;
2283 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2285 down(&dentry->d_inode->i_sem);
2286 rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
2287 up(&dentry->d_inode->i_sem);
2289 CERROR("Can't read inode self id, inode %lu, "
2290 "rc %d\n", dentry->d_inode->i_ino, rc);
2294 body->oa.o_fid = id_fid(&id);
2295 body->oa.o_mds = id_group(&id);
2301 static int mds_inode_init_acl(struct obd_device *obd, void *handle,
2302 struct dentry *de, void *xattr, int xattr_size)
2304 struct inode *inode = de->d_inode;
2305 struct posix_acl *acl;
2312 LASSERT(xattr_size > 0);
2314 if (!inode->i_op->getxattr || !inode->i_op->setxattr) {
2315 CERROR("backend fs dosen't support xattr\n");
2319 /* set default acl */
2320 if (S_ISDIR(inode->i_mode)) {
2321 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_DEFAULT,
2322 xattr, xattr_size, 0);
2324 CERROR("set default acl err: %d\n", rc);
2329 /* set access acl */
2330 acl = posix_acl_from_xattr(xattr, xattr_size);
2331 if (acl == NULL || IS_ERR(acl)) {
2332 CERROR("insane attr data\n");
2333 return PTR_ERR(acl);
2336 if (posix_acl_valid(acl)) {
2337 CERROR("default acl not valid: %d\n", rc);
2342 mode = inode->i_mode;
2343 rc = posix_acl_create_masq(acl, &mode);
2345 CERROR("create masq err %d\n", rc);
2349 if (inode->i_mode != mode) {
2350 struct iattr iattr = { .ia_valid = ATTR_MODE,
2354 rc2 = fsfilt_setattr(obd, de, handle, &iattr, 0);
2356 CERROR("setattr mode err: %d\n", rc2);
2363 /* we didn't change acl except mode bits of some
2364 * entries, so should be fit into original size.
2366 rc = posix_acl_to_xattr(acl, xattr, xattr_size);
2369 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_ACCESS,
2370 xattr, xattr_size, 0);
2372 CERROR("set access acl err: %d\n", rc);
2375 posix_acl_release(acl);
2379 static int mdt_obj_create(struct ptlrpc_request *req)
2381 struct obd_device *obd = req->rq_export->exp_obd;
2382 struct mds_obd *mds = &obd->u.mds;
2383 struct ost_body *body, *repbody;
2386 char idname[LL_ID_NAMELEN];
2387 int size = sizeof(*repbody);
2388 struct inode *parent_inode;
2389 struct lvfs_run_ctxt saved;
2390 int rc, cleanup_phase = 0;
2391 struct dentry *new = NULL;
2392 struct dentry_params dp;
2393 int mealen, flags = 0;
2394 struct lvfs_ucred uc;
2395 struct lustre_id id;
2397 void *handle = NULL;
2398 unsigned long cr_inum = 0;
2402 DEBUG_REQ(D_HA, req, "create remote object");
2403 parent_inode = mds->mds_unnamed_dir->d_inode;
2405 body = lustre_swab_reqbuf(req, 0, sizeof(*body),
2406 lustre_swab_ost_body);
2410 /* acl data is packed transparently, no swab here */
2411 LASSERT(req->rq_reqmsg->bufcount >= 2);
2412 acl_size = req->rq_reqmsg->buflens[1];
2414 acl = lustre_msg_buf(req->rq_reqmsg, 1, acl_size);
2416 CERROR("No default acl buf?\n");
2421 rc = lustre_pack_reply(req, 1, &size, NULL);
2425 MDS_CHECK_RESENT(req, reconstruct_create(req));
2428 uc.luc_ginfo = NULL;
2429 uc.luc_uid = body->oa.o_uid;
2430 uc.luc_gid = body->oa.o_gid;
2431 uc.luc_fsuid = body->oa.o_uid;
2432 uc.luc_fsgid = body->oa.o_gid;
2434 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2435 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2437 /* in REPLAY case inum should be given (client or other MDS fills it) */
2438 if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2439 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2441 * this is re-create request from MDS holding directory name.
2442 * we have to lookup given ino/gen first. if it exists (good
2443 * case) then there is nothing to do. if it does not then we
2444 * have to recreate it.
2446 id_ino(&id) = body->oa.o_id;
2447 id_gen(&id) = body->oa.o_generation;
2449 new = mds_id2dentry(obd, &id, NULL);
2450 if (!IS_ERR(new) && new->d_inode) {
2451 struct lustre_id sid;
2453 CDEBUG(D_OTHER, "mkdir repairing %lu/%lu\n",
2454 (unsigned long)id_ino(&id),
2455 (unsigned long)id_gen(&id));
2457 obdo_from_inode(&repbody->oa, new->d_inode,
2458 FILTER_VALID_FLAGS);
2460 repbody->oa.o_id = new->d_inode->i_ino;
2461 repbody->oa.o_generation = new->d_inode->i_generation;
2462 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2465 down(&new->d_inode->i_sem);
2466 rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2467 up(&new->d_inode->i_sem);
2469 CERROR("Can't read inode self id "
2470 "inode %lu, rc %d.\n",
2471 new->d_inode->i_ino, rc);
2475 repbody->oa.o_fid = id_fid(&sid);
2476 repbody->oa.o_mds = id_group(&sid);
2477 LASSERT(id_fid(&sid) != 0);
2480 * here we could use fid passed in body->oa.o_fid and
2481 * thus avoid mds_read_inode_sid().
2483 cr_inum = new->d_inode->i_ino;
2484 GOTO(cleanup, rc = 0);
2488 down(&parent_inode->i_sem);
2489 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2490 if (IS_ERR(handle)) {
2491 up(&parent_inode->i_sem);
2492 CERROR("fsfilt_start() failed, rc = %d\n",
2493 (int)PTR_ERR(handle));
2494 GOTO(cleanup, rc = PTR_ERR(handle));
2496 cleanup_phase = 1; /* transaction */
2499 rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2500 new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2502 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2503 obd->obd_name, idname, (int) PTR_ERR(new));
2504 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2505 up(&parent_inode->i_sem);
2506 RETURN(PTR_ERR(new));
2507 } else if (new->d_inode) {
2508 CERROR("%s: name exists. repeat\n", obd->obd_name);
2511 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2512 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2513 fid = body->oa.o_fid;
2515 fid = mds_alloc_fid(obd);
2517 new->d_fsdata = (void *)&dp;
2521 dp.p_group = mds->mds_num;
2523 if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2524 (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2525 LASSERT(body->oa.o_id != 0);
2526 dp.p_inum = body->oa.o_id;
2527 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2528 (unsigned long)body->oa.o_id,
2529 (unsigned long)body->oa.o_generation);
2532 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2535 rc = mds_inode_init_acl(obd, handle, new,
2538 up(&parent_inode->i_sem);
2542 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2543 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2544 new->d_inode->i_generation = body->oa.o_generation;
2545 mark_inode_dirty(new->d_inode);
2548 * avoiding asserts in cache flush case, as
2549 * @body->oa.o_id should be zero.
2551 if (body->oa.o_id) {
2552 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2553 "BUG 3550: failed to recreate obj "
2554 LPU64" -> %lu\n", body->oa.o_id,
2555 new->d_inode->i_ino);
2557 LASSERTF(body->oa.o_generation ==
2558 new->d_inode->i_generation,
2559 "BUG 3550: failed to recreate obj/gen "
2560 LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2561 body->oa.o_generation,
2562 new->d_inode->i_ino,
2563 new->d_inode->i_generation);
2567 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2568 repbody->oa.o_id = new->d_inode->i_ino;
2569 repbody->oa.o_generation = new->d_inode->i_generation;
2570 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2572 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2573 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2574 LASSERT(body->oa.o_id != 0);
2575 LASSERT(body->oa.o_fid != 0);
2578 mds_inode2id(obd, &id, new->d_inode, fid);
2579 mds_update_inode_ids(obd, new->d_inode, handle, &id,
2580 body->oa.o_valid & OBD_MD_FLID ?
2581 NULL : obdo_id(&body->oa));
2583 /* initializing o_fid after it is allocated. */
2584 repbody->oa.o_fid = id_fid(&id);
2585 repbody->oa.o_mds = id_group(&id);
2587 rc = fsfilt_del_dir_entry(obd, new);
2588 up(&parent_inode->i_sem);
2590 CERROR("can't remove name for object: %d\n", rc);
2594 cleanup_phase = 2; /* created directory object */
2596 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2597 (unsigned long)new->d_inode->i_ino,
2598 (unsigned long)new->d_inode->i_generation,
2599 (unsigned)new->d_inode->i_mode);
2600 cr_inum = new->d_inode->i_ino;
2602 up(&parent_inode->i_sem);
2603 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2607 if (body->oa.o_valid & OBD_MD_FLID) {
2608 /* this is new object for splitted dir. We have to prevent
2609 * recursive splitting on it -bzzz */
2610 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2612 OBD_ALLOC(mea, mealen);
2614 GOTO(cleanup, rc = -ENOMEM);
2616 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2617 mea->mea_master = body->oa.o_mds; /* master mds num */
2620 obdo2id(&mea->mea_ids[body->oa.o_mds], &body->oa);
2622 down(&new->d_inode->i_sem);
2623 rc = fsfilt_set_md(obd, new->d_inode, handle,
2624 mea, mealen, EA_MEA);
2625 up(&new->d_inode->i_sem);
2627 CERROR("fsfilt_set_md() failed, "
2630 OBD_FREE(mea, mealen);
2632 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2633 obd->obd_name, new->d_inode->i_ino,
2634 new->d_inode->i_generation, flags);
2635 } else if (body->oa.o_easize) {
2636 /* we pass LCK_EX to split routine to signal that we have
2637 * exclusive access to the directory. simple because nobody
2638 * knows it already exists -bzzz */
2639 rc = mds_try_to_split_dir(obd, new, NULL,
2640 body->oa.o_easize, LCK_EX);
2642 CERROR("Can't split directory %lu, error = %d.\n",
2643 new->d_inode->i_ino, rc);
2651 switch (cleanup_phase) {
2652 case 2: /* object has been created, but we'll may want to replay it later */
2654 ptlrpc_require_repack(req);
2655 case 1: /* transaction */
2656 rc = mds_finish_transno(mds, parent_inode, handle,
2661 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2665 static int mdt_get_info(struct ptlrpc_request *req)
2667 struct obd_export *exp = req->rq_export;
2672 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2674 DEBUG_REQ(D_HA, req, "no get_info key");
2677 keylen = req->rq_reqmsg->buflens[0];
2679 if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2680 (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2681 (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2684 if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2685 struct lustre_id *reply;
2686 int size = sizeof(*reply);
2688 rc = lustre_pack_reply(req, 1, &size, NULL);
2692 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2693 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2696 int size = sizeof(*reply);
2698 rc = lustre_pack_reply(req, 1, &size, NULL);
2702 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2703 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2706 req->rq_repmsg->status = 0;
2710 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2711 void *key, __u32 vallen, void *val)
2713 struct obd_device *obd;
2714 struct mds_obd *mds;
2718 obd = class_exp2obd(exp);
2720 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2721 exp->exp_handle.h_cookie);
2726 if (keylen >= strlen("mds_type") &&
2727 memcmp(key, "mds_type", keylen) == 0) {
2731 CDEBUG(D_IOCTL, "set mds type to %x\n", *(int*)val);
2733 mds->mds_obd_type = *(int*)val;
2734 group = FILTER_GROUP_FIRST_MDS + mds->mds_obd_type;
2735 valsize = sizeof(group);
2737 /* mds number has been changed, so the corresponding obdfilter
2738 * exp need to be changed too. */
2739 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
2740 "mds_conn", valsize, &group);
2742 } else if (keylen == 5 && memcmp(key, "audit", 5) == 0) {
2743 rc = mds_set_audit(obd, val);
2745 } else if (keylen >= strlen("ids") && memcmp(key, "ids", keylen) == 0) {
2746 struct lustre_id *ids = (struct lustre_id *)val;
2748 struct inode *inode;
2752 de = mds_id2dentry(obd, ids, NULL);
2755 CERROR("lookup by an id error %d\n", rc);
2758 inode = de->d_inode;
2760 GOTO(out_put, rc = -ENOENT);
2762 down(&inode->i_sem);
2763 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
2764 if (IS_ERR(handle)) {
2766 GOTO(out_put, rc = PTR_ERR(handle));
2769 rc = mds_update_inode_ids(obd, inode, handle, NULL, ids + 1);
2771 err = fsfilt_commit(obd, mds->mds_sb, inode, handle,
2774 CERROR("error committing transaction: %d\n", err);
2783 if (keylen >= strlen("crypto_type") &&
2784 memcmp(key, "crypto_type", keylen) == 0) {
2785 rc = mds_set_crypto_type(obd, val, vallen);
2789 CDEBUG(D_IOCTL, "invalid key\n");
2793 static int mdt_set_info(struct ptlrpc_request *req)
2796 struct obd_export *exp = req->rq_export;
2797 int keylen, rc = 0, vallen = 0;
2800 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2802 DEBUG_REQ(D_HA, req, "no set_info key");
2805 keylen = req->rq_reqmsg->buflens[0];
2807 if ((keylen == strlen("mds_type") &&
2808 memcmp(key, "mds_type", keylen) == 0) ||
2809 (keylen == strlen("crypto_type") &&
2810 memcmp(key, "crypto_type", keylen) == 0)) {
2811 rc = lustre_pack_reply(req, 0, NULL, NULL);
2815 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2816 vallen = req->rq_reqmsg->buflens[1];
2818 rc = obd_set_info(exp, keylen, key, vallen, val);
2819 req->rq_repmsg->status = 0;
2821 } else if (keylen == 5 && memcmp(key, "audit", 5) == 0) {
2822 struct audit_attr_msg msg, *p;
2825 rc = lustre_pack_reply(req, 0, NULL, NULL);
2829 p = lustre_swab_reqbuf(req, 1, sizeof(msg),
2830 lustre_swab_audit_attr);
2833 CDEBUG(D_INFO, "Get new audit setting 0x%x\n", (__u32)msg.attr);
2834 rc = obd_set_info(exp, keylen, key, sizeof(msg), &msg);
2836 req->rq_repmsg->status = rc;
2838 } else if (keylen == strlen("ids") &&
2839 memcmp(key, "ids", keylen) == 0) {
2840 struct lustre_id *id, ids[2];
2842 rc = lustre_pack_reply(req, 0, NULL, NULL);
2845 id = lustre_swab_reqbuf(req, 1, sizeof(struct lustre_id),
2846 lustre_swab_lustre_id);
2848 id = lustre_swab_reqbuf(req, 2, sizeof(struct lustre_id),
2849 lustre_swab_lustre_id);
2852 rc = obd_set_info(exp, keylen, key, vallen, ids);
2853 req->rq_repmsg->status = rc;
2857 CDEBUG(D_IOCTL, "invalid key\n");
2861 static void mds_revoke_export_locks(struct obd_export *exp)
2863 struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2864 struct list_head rpc_list;
2865 struct ldlm_lock *lock, *next;
2866 struct ldlm_lock_desc desc;
2868 /* don't do this for local client */
2869 if (!exp->u.eu_mds_data.med_remote)
2872 /* don't revoke locks during recovery */
2873 if (exp->exp_obd->obd_recovering)
2877 INIT_LIST_HEAD(&rpc_list);
2879 spin_lock(&exp->exp_ldlm_data.led_lock);
2880 list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2882 lock_res_and_lock(lock);
2883 if (lock->l_req_mode != lock->l_granted_mode) {
2884 unlock_res_and_lock(lock);
2888 LASSERT(lock->l_resource);
2889 if (lock->l_resource->lr_type != LDLM_IBITS &&
2890 lock->l_resource->lr_type != LDLM_PLAIN) {
2891 unlock_res_and_lock(lock);
2895 if (lock->l_flags & LDLM_FL_AST_SENT) {
2896 unlock_res_and_lock(lock);
2900 LASSERT(lock->l_blocking_ast);
2901 LASSERT(!lock->l_blocking_lock);
2903 lock->l_flags |= LDLM_FL_AST_SENT;
2904 unlock_res_and_lock(lock);
2906 list_move(&lock->l_export_chain, &rpc_list);
2908 spin_unlock(&exp->exp_ldlm_data.led_lock);
2910 while (!list_empty(&rpc_list)) {
2911 lock = list_entry(rpc_list.next, struct ldlm_lock,
2913 list_del_init(&lock->l_export_chain);
2915 /* the desc just pretend to exclusive */
2916 ldlm_lock2desc(lock, &desc);
2917 desc.l_req_mode = LCK_EX;
2918 desc.l_granted_mode = 0;
2920 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2925 static int mds_msg_check_version(struct lustre_msg *msg)
2931 case MDS_DISCONNECT:
2933 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2935 CERROR("bad opc %u version %08x, expecting %08x\n",
2936 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2941 case MDS_GETATTR_LOCK:
2942 case MDS_ACCESS_CHECK:
2946 case MDS_DONE_WRITING:
2949 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2951 CERROR("bad opc %u version %08x, expecting %08x\n",
2952 msg->opc, msg->version, LUSTRE_MDS_VERSION);
2956 case LDLM_BL_CALLBACK:
2957 case LDLM_CP_CALLBACK:
2958 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2960 CERROR("bad opc %u version %08x, expecting %08x\n",
2961 msg->opc, msg->version, LUSTRE_DLM_VERSION);
2963 case OBD_LOG_CANCEL:
2964 case LLOG_ORIGIN_HANDLE_OPEN:
2965 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2966 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2967 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2968 case LLOG_ORIGIN_HANDLE_CLOSE:
2970 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2972 CERROR("bad opc %u version %08x, expecting %08x\n",
2973 msg->opc, msg->version, LUSTRE_LOG_VERSION);
2979 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2981 CERROR("bad opc %u version %08x, expecting %08x\n",
2982 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2985 case SEC_INIT_CONTINUE:
2990 CERROR("MDS unknown opcode %d\n", msg->opc);
2998 int mds_handle(struct ptlrpc_request *req)
3000 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
3001 struct obd_device *obd = NULL;
3002 struct mds_obd *mds = NULL; /* quell gcc overwarning */
3006 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
3008 rc = mds_msg_check_version(req->rq_reqmsg);
3010 CERROR("MDS drop mal-formed request\n");
3014 /* Security opc should NOT trigger any recovery events */
3015 if (req->rq_reqmsg->opc == SEC_INIT ||
3016 req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
3017 if (req->rq_export) {
3018 mds_req_add_idmapping(req,
3019 &req->rq_export->exp_mds_data);
3020 mds_revoke_export_locks(req->rq_export);
3023 } else if (req->rq_reqmsg->opc == SEC_FINI) {
3024 if (req->rq_export) {
3025 mds_req_del_idmapping(req,
3026 &req->rq_export->exp_mds_data);
3027 mds_revoke_export_locks(req->rq_export);
3032 LASSERT(current->journal_info == NULL);
3033 /* XXX identical to OST */
3034 if (req->rq_reqmsg->opc != MDS_CONNECT) {
3035 struct mds_export_data *med;
3038 if (req->rq_export == NULL) {
3039 CERROR("operation %d on unconnected MDS from %s\n",
3040 req->rq_reqmsg->opc,
3042 req->rq_status = -ENOTCONN;
3043 GOTO(out, rc = -ENOTCONN);
3046 med = &req->rq_export->exp_mds_data;
3047 obd = req->rq_export->exp_obd;
3050 /* sanity check: if the xid matches, the request must
3051 * be marked as a resent or replayed */
3052 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
3053 req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
3054 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
3055 (MSG_RESENT | MSG_REPLAY),
3056 "rq_xid "LPU64" matches last_xid, "
3057 "expected RESENT flag\n",
3060 /* else: note the opposite is not always true; a
3061 * RESENT req after a failover will usually not match
3062 * the last_xid, since it was likely never
3063 * committed. A REPLAYed request will almost never
3064 * match the last xid, however it could for a
3065 * committed, but still retained, open. */
3067 spin_lock_bh(&obd->obd_processing_task_lock);
3068 recovering = obd->obd_recovering;
3069 spin_unlock_bh(&obd->obd_processing_task_lock);
3071 rc = mds_filter_recovery_request(req, obd,
3073 if (rc || should_process == 0) {
3075 } else if (should_process < 0) {
3076 req->rq_status = should_process;
3077 rc = ptlrpc_error(req);
3083 switch (req->rq_reqmsg->opc) {
3085 DEBUG_REQ(D_INODE, req, "connect");
3086 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
3087 rc = target_handle_connect(req);
3089 struct mds_export_data *med;
3091 LASSERT(req->rq_export);
3092 med = &req->rq_export->u.eu_mds_data;
3093 mds_init_export_data(req, med);
3094 mds_req_add_idmapping(req, med);
3096 /* Now that we have an export, set mds. */
3097 obd = req->rq_export->exp_obd;
3098 mds = mds_req2mds(req);
3102 case MDS_DISCONNECT:
3103 DEBUG_REQ(D_INODE, req, "disconnect");
3104 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
3105 rc = target_handle_disconnect(req);
3106 req->rq_status = rc; /* superfluous? */
3110 DEBUG_REQ(D_INODE, req, "getstatus");
3111 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
3112 rc = mds_getstatus(req);
3116 DEBUG_REQ(D_INODE, req, "getattr");
3117 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
3118 rc = mds_getattr(req, MDS_REQ_REC_OFF);
3121 case MDS_ACCESS_CHECK:
3122 DEBUG_REQ(D_INODE, req, "access_check");
3123 OBD_FAIL_RETURN(OBD_FAIL_MDS_ACCESS_CHECK_NET, 0);
3124 rc = mds_access_check(req, MDS_REQ_REC_OFF);
3127 case MDS_GETATTR_LOCK: {
3128 struct lustre_handle lockh;
3129 DEBUG_REQ(D_INODE, req, "getattr_lock");
3130 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
3132 /* If this request gets a reconstructed reply, we won't be
3133 * acquiring any new locks in mds_getattr_lock, so we don't
3137 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
3138 MDS_INODELOCK_UPDATE);
3139 /* this non-intent call (from an ioctl) is special */
3140 req->rq_status = rc;
3141 if (rc == 0 && lockh.cookie)
3142 ldlm_lock_decref(&lockh, LCK_PR);
3146 DEBUG_REQ(D_INODE, req, "statfs");
3147 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
3148 rc = mds_statfs(req);
3152 DEBUG_REQ(D_INODE, req, "readpage");
3153 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
3154 rc = mds_readpage(req, MDS_REQ_REC_OFF);
3156 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
3157 if (req->rq_reply_state) {
3158 lustre_free_reply_state (req->rq_reply_state);
3159 req->rq_reply_state = NULL;
3166 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
3169 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
3170 mds->mds_max_cookiesize};
3173 /* NB only peek inside req now; mds_reint() will swab it */
3175 CERROR ("Can't inspect opcode\n");
3180 if (lustre_msg_swabbed (req->rq_reqmsg))
3183 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
3184 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
3185 reint_names[opc] == NULL) ? reint_names[opc] :
3188 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
3190 if (opc == REINT_UNLINK || opc == REINT_RENAME)
3192 else if (opc == REINT_OPEN)
3197 /* for SETATTR: I have different reply setting for
3198 * remote setfacl, so delay the reply buffer allocation.
3200 if (opc != REINT_SETATTR) {
3201 rc = lustre_pack_reply(req, bufcount, size, NULL);
3206 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
3207 fail = OBD_FAIL_MDS_REINT_NET_REP;
3212 DEBUG_REQ(D_INODE, req, "close");
3213 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
3214 rc = mds_close(req, MDS_REQ_REC_OFF);
3217 case MDS_DONE_WRITING:
3218 DEBUG_REQ(D_INODE, req, "done_writing");
3219 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
3220 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
3224 DEBUG_REQ(D_INODE, req, "pin");
3225 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
3226 rc = mds_pin(req, MDS_REQ_REC_OFF);
3230 DEBUG_REQ(D_INODE, req, "sync");
3231 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
3232 rc = mds_sync(req, MDS_REQ_REC_OFF);
3235 DEBUG_REQ(D_INODE, req, "parseid");
3236 rc = mds_parse_id(req);
3239 DEBUG_REQ(D_INODE, req, "ping");
3240 rc = target_handle_ping(req);
3243 case OBD_LOG_CANCEL:
3244 CDEBUG(D_INODE, "log cancel\n");
3245 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
3246 rc = -ENOTSUPP; /* la la la */
3250 DEBUG_REQ(D_INODE, req, "enqueue");
3251 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
3252 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
3253 ldlm_server_blocking_ast, NULL);
3254 fail = OBD_FAIL_LDLM_REPLY;
3257 DEBUG_REQ(D_INODE, req, "convert");
3258 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
3259 rc = ldlm_handle_convert(req);
3261 case LDLM_BL_CALLBACK:
3262 case LDLM_CP_CALLBACK:
3263 DEBUG_REQ(D_INODE, req, "callback");
3264 CERROR("callbacks should not happen on MDS\n");
3266 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
3268 case LLOG_ORIGIN_HANDLE_OPEN:
3269 DEBUG_REQ(D_INODE, req, "llog_init");
3270 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3271 rc = llog_origin_handle_open(req);
3273 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
3274 DEBUG_REQ(D_INODE, req, "llog next block");
3275 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3276 rc = llog_origin_handle_next_block(req);
3278 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
3279 DEBUG_REQ(D_INODE, req, "llog prev block");
3280 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3281 rc = llog_origin_handle_prev_block(req);
3283 case LLOG_ORIGIN_HANDLE_READ_HEADER:
3284 DEBUG_REQ(D_INODE, req, "llog read header");
3285 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3286 rc = llog_origin_handle_read_header(req);
3288 case LLOG_ORIGIN_HANDLE_CLOSE:
3289 DEBUG_REQ(D_INODE, req, "llog close");
3290 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3291 rc = llog_origin_handle_close(req);
3294 DEBUG_REQ(D_INODE, req, "ost_create");
3295 rc = mdt_obj_create(req);
3298 DEBUG_REQ(D_INODE, req, "get_info");
3299 rc = mdt_get_info(req);
3302 DEBUG_REQ(D_INODE, req, "set_info");
3303 rc = mdt_set_info(req);
3306 CDEBUG(D_INODE, "write\n");
3307 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
3308 rc = ost_brw_write(req, NULL);
3309 LASSERT(current->journal_info == NULL);
3310 /* mdt_brw sends its own replies */
3314 DEBUG_REQ(D_INODE, req, "llog catinfo");
3315 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3316 rc = llog_catinfo(req);
3319 req->rq_status = -ENOTSUPP;
3320 rc = ptlrpc_error(req);
3324 LASSERT(current->journal_info == NULL);
3328 /* If we're DISCONNECTing, the mds_export_data is already freed */
3329 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
3330 struct mds_export_data *med = &req->rq_export->exp_mds_data;
3331 struct obd_device *obd = list_entry(mds, struct obd_device,
3333 req->rq_repmsg->last_xid =
3334 le64_to_cpu(med->med_mcd->mcd_last_xid);
3336 if (!obd->obd_no_transno) {
3337 req->rq_repmsg->last_committed =
3338 obd->obd_last_committed;
3340 DEBUG_REQ(D_IOCTL, req,
3341 "not sending last_committed update");
3343 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
3345 mds->mds_last_transno, obd->obd_last_committed,
3351 target_send_reply(req, rc, fail);
3355 /* Update the server data on disk. This stores the new mount_count and also the
3356 * last_rcvd value to disk. If we don't have a clean shutdown, then the server
3357 * last_rcvd value may be less than that of the clients. This will alert us
3358 * that we may need to do client recovery.
3360 * Also assumes for mds_last_transno that we are not modifying it (no locking).
3362 int mds_update_server_data(struct obd_device *obd, int force_sync)
3364 struct mds_obd *mds = &obd->u.mds;
3365 struct mds_server_data *msd = mds->mds_server_data;
3366 struct file *filp = mds->mds_rcvd_filp;
3367 struct lvfs_run_ctxt saved;
3372 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3373 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
3375 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
3376 mds->mds_mount_count, mds->mds_last_transno);
3377 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
3379 CERROR("error writing MDS server data: rc = %d\n", rc);
3380 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3385 /* saves last allocated fid counter to file. */
3386 int mds_update_last_fid(struct obd_device *obd, void *handle,
3389 struct mds_obd *mds = &obd->u.mds;
3390 struct file *filp = mds->mds_fid_filp;
3391 struct lvfs_run_ctxt saved;
3397 spin_lock(&mds->mds_last_fid_lock);
3398 last_fid = mds->mds_last_fid;
3399 spin_unlock(&mds->mds_last_fid_lock);
3401 CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
3405 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
3406 handle, mds_commit_last_fid_cb,
3410 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3411 rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
3413 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3416 CERROR("error writing MDS last_fid #"LPU64
3417 ", err = %d\n", last_fid, rc);
3421 CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
3422 "%llu: err = %d\n", last_fid, off, rc);
3427 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
3429 struct mds_obd *mds = &obd->u.mds;
3431 spin_lock(&mds->mds_last_fid_lock);
3432 if (fid > mds->mds_last_fid)
3433 mds->mds_last_fid = fid;
3434 spin_unlock(&mds->mds_last_fid_lock);
3437 void mds_commit_last_transno_cb(struct obd_device *obd,
3438 __u64 transno, void *data,
3441 obd_transno_commit_cb(obd, transno, error);
3444 void mds_commit_last_fid_cb(struct obd_device *obd,
3445 __u64 fid, void *data,
3449 CERROR("%s: fid "LPD64" commit error: %d\n",
3450 obd->obd_name, fid, error);
3454 CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
3455 obd->obd_name, fid);
3458 __u64 mds_alloc_fid(struct obd_device *obd)
3460 struct mds_obd *mds = &obd->u.mds;
3463 spin_lock(&mds->mds_last_fid_lock);
3464 fid = ++mds->mds_last_fid;
3465 spin_unlock(&mds->mds_last_fid_lock);
3471 * reads inode self id from inode EA. Probably later this should be replaced by
3472 * caching inode self id to avoid raeding it every time it is needed.
3474 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
3475 struct lustre_id *id)
3480 LASSERT(id != NULL);
3481 LASSERT(obd != NULL);
3482 LASSERT(inode != NULL);
3484 rc = fsfilt_get_md(obd, inode, &id->li_fid,
3485 sizeof(id->li_fid), EA_SID);
3487 CERROR("fsfilt_get_md() failed, "
3496 id_ino(id) = inode->i_ino;
3497 id_gen(id) = inode->i_generation;
3498 id_type(id) = S_IFMT & inode->i_mode;
3503 int mds_read_inode_pid(struct obd_device *obd, struct inode *inode,
3504 struct lustre_id *id)
3509 LASSERT(inode && id);
3511 rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_PID);
3513 CERROR("get parent id from EA failed, rc=%d\n", rc);
3522 /* updates inode self id in EA. */
3523 int mds_update_inode_ids(struct obd_device *obd, struct inode *inode,
3524 void *handle, struct lustre_id *id,
3525 struct lustre_id *pid)
3531 LASSERT(id == NULL || id_fid(id) != 0);
3532 LASSERT(pid == NULL || id_fid(pid) != 0);
3533 LASSERT(obd != NULL);
3534 LASSERT(inode != NULL);
3537 mds_set_last_fid(obd, id_fid(id));
3538 rc = fsfilt_set_md(obd, inode, handle, &id->li_fid,
3539 sizeof(id->li_fid), EA_SID);
3540 LASSERTF(rc == 0, "failed to update fid: %d\n", rc);
3543 rc = fsfilt_set_md(obd, inode, handle, pid,
3544 sizeof(*pid), EA_PID);
3545 LASSERTF(rc == 0, "failed to update parent fid: %d\n", rc);
3552 * reads inode id on master MDS. This is usualy done by CMOBD to update requests
3553 * to master MDS by correct store cookie, needed to find inode on master MDS
3556 int mds_read_inode_mid(struct obd_device *obd, struct inode *inode,
3557 struct lustre_id *id)
3562 LASSERT(id != NULL);
3563 LASSERT(obd != NULL);
3564 LASSERT(inode != NULL);
3566 rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_MID);
3568 CERROR("fsfilt_get_md() failed, rc = %d\n", rc);
3581 * updates master inode id. Usualy this is done by CMOBD after an inode is
3582 * created and relationship between cache MDS and master one should be
3585 int mds_update_inode_mid(struct obd_device *obd, struct inode *inode,
3586 void *handle, struct lustre_id *id)
3591 LASSERT(id != NULL);
3592 LASSERT(obd != NULL);
3593 LASSERT(inode != NULL);
3595 rc = fsfilt_set_md(obd, inode, handle, id,
3596 sizeof(*id), EA_MID);
3598 CERROR("fsfilt_set_md() failed, "
3606 /* mount the file system (secretly) */
3607 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3609 struct lustre_cfg* lcfg = buf;
3610 struct mds_obd *mds = &obd->u.mds;
3611 struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3612 char *options = NULL;
3613 struct vfsmount *mnt;
3619 if (lcfg->lcfg_bufcount < 3)
3620 RETURN(rc = -EINVAL);
3622 if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
3623 RETURN(rc = -EINVAL);
3625 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
3626 if (IS_ERR(obd->obd_fsops))
3627 RETURN(rc = PTR_ERR(obd->obd_fsops));
3629 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3631 page = get_zeroed_page(GFP_KERNEL);
3635 options = (char *)page;
3638 * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3639 * and the rest of options are passed by mount options. Probably this
3640 * should be moved to somewhere else like startup scripts or lconf. */
3641 sprintf(options, "iopen_nopriv");
3643 if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0 && lustre_cfg_buf(lcfg, 4))
3644 sprintf(options + strlen(options), ",%s",
3645 lustre_cfg_string(lcfg, 4));
3647 /* we have to know mdsnum before touching underlying fs -bzzz */
3648 atomic_set(&mds->mds_open_count, 0);
3649 sema_init(&mds->mds_md_sem, 1);
3650 sema_init(&mds->mds_create_sem, 1);
3651 mds->mds_md_connected = 0;
3652 mds->mds_md_name = NULL;
3654 if (LUSTRE_CFG_BUFLEN(lcfg, 5) > 0 && lustre_cfg_buf(lcfg, 5) &&
3655 strncmp(lustre_cfg_string(lcfg, 5), "dumb", LUSTRE_CFG_BUFLEN(lcfg, 5))) {
3658 generate_random_uuid(uuid);
3659 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3661 OBD_ALLOC(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3662 if (mds->mds_md_name == NULL)
3663 RETURN(rc = -ENOMEM);
3665 memcpy(mds->mds_md_name, lustre_cfg_buf(lcfg, 5),
3666 LUSTRE_CFG_BUFLEN(lcfg, 5));
3668 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3669 obd->obd_name, mds->mds_md_name);
3671 rc = mds_md_connect(obd, mds->mds_md_name);
3673 OBD_FREE(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3678 mds->mds_obd_type = MDS_MASTER_OBD;
3680 if (LUSTRE_CFG_BUFLEN(lcfg, 6) > 0 && lustre_cfg_buf(lcfg, 6) &&
3681 strncmp(lustre_cfg_string(lcfg, 6), "dumb",
3682 LUSTRE_CFG_BUFLEN(lcfg, 6))) {
3683 if (!memcmp(lustre_cfg_string(lcfg, 6), "master",
3684 strlen("master"))) {
3685 mds->mds_obd_type = MDS_MASTER_OBD;
3686 } else if (!memcmp(lustre_cfg_string(lcfg, 6), "cache",
3688 mds->mds_obd_type = MDS_CACHE_OBD;
3692 rc = lvfs_mount_fs(lustre_cfg_string(lcfg, 1),
3693 lustre_cfg_string(lcfg, 2),
3694 options, 0, &lvfs_ctxt);
3698 if (rc || !lvfs_ctxt) {
3699 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3703 mnt = lvfs_ctxt->loc_mnt;
3704 mds->mds_lvfs_ctxt = lvfs_ctxt;
3705 ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3707 CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
3709 sema_init(&mds->mds_epoch_sem, 1);
3710 atomic_set(&mds->mds_real_clients, 0);
3711 spin_lock_init(&mds->mds_transno_lock);
3712 spin_lock_init(&mds->mds_last_fid_lock);
3713 sema_init(&mds->mds_orphan_recovery_sem, 1);
3714 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3716 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3717 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3719 if (obd->obd_namespace == NULL) {
3720 mds_cleanup(obd, 0);
3721 GOTO(err_put, rc = -ENOMEM);
3723 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3725 rc = mds_fs_setup(obd, mnt);
3727 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3732 rc = llog_start_commit_thread();
3738 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 && lustre_cfg_buf(lcfg, 3) &&
3739 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3740 LUSTRE_CFG_BUFLEN(lcfg, 3))) {
3743 generate_random_uuid(uuid);
3744 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3746 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
3747 if (mds->mds_profile == NULL)
3748 GOTO(err_fs, rc = -ENOMEM);
3750 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
3751 LUSTRE_CFG_BUFLEN(lcfg, 3));
3755 * setup root dir and files ID dir if lmv already connected, or there is
3758 if (mds->mds_md_exp || (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 &&
3759 lustre_cfg_buf(lcfg, 3) &&
3760 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3761 LUSTRE_CFG_BUFLEN(lcfg, 3)))) {
3762 rc = mds_fs_setup_rootid(obd);
3766 rc = mds_fs_setup_virtid(obd);
3772 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3773 "mds_ldlm_client", &obd->obd_ldlm_client);
3774 obd->obd_replayable = 1;
3776 mds->mds_crypto_type = NO_CRYPTO;
3778 rc = mds_postsetup(obd);
3785 /* No extra cleanup needed for llog_init_commit_thread() */
3786 mds_fs_cleanup(obd, 0);
3788 ldlm_namespace_free(obd->obd_namespace, 0);
3789 obd->obd_namespace = NULL;
3792 lvfs_umount_fs(mds->mds_lvfs_ctxt);
3796 fsfilt_put_ops(obd->obd_fsops);
3800 static int mds_fs_post_setup(struct obd_device *obd)
3802 struct mds_obd *mds = &obd->u.mds;
3803 struct dentry *dentry;
3807 dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
3808 if (IS_ERR(dentry)) {
3809 CERROR("Can't find ROOT, err = %d\n",
3810 (int)PTR_ERR(dentry));
3811 RETURN(PTR_ERR(dentry));
3813 rc = fsfilt_post_setup(obd, dentry);
3814 //set id2name function handler
3815 fsfilt_set_info(obd, mds->mds_sb, NULL, 7, "id2name",
3816 sizeof(mds_audit_id2name), mds_audit_id2name);
3822 static int mds_postsetup(struct obd_device *obd)
3824 struct mds_obd *mds = &obd->u.mds;
3828 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3829 obd, 0, NULL, &llog_lvfs_ops);
3833 if (mds->mds_profile) {
3834 struct llog_ctxt *lgctxt;
3835 struct lvfs_run_ctxt saved;
3836 struct lustre_profile *lprof;
3837 struct config_llog_instance cfg;
3839 cfg.cfg_instance = NULL;
3840 cfg.cfg_uuid = mds->mds_dt_uuid;
3841 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3843 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3845 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3846 GOTO(err_llog, rc = -EINVAL);
3849 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3850 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3855 lprof = class_get_profile(mds->mds_profile);
3856 if (lprof == NULL) {
3857 CERROR("No profile found: %s\n", mds->mds_profile);
3858 GOTO(err_cleanup, rc = -ENOENT);
3860 rc = mds_dt_connect(obd, lprof->lp_lov);
3862 GOTO(err_cleanup, rc);
3864 rc = mds_md_postsetup(obd);
3866 GOTO(err_cleanup, rc);
3868 rc = mds_fs_post_setup(obd);
3870 CERROR("can not post setup fsfilt\n");
3875 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3876 LLOG_CONFIG_ORIG_CTXT));
3880 int mds_postrecov_common(struct obd_device *obd)
3882 struct mds_obd *mds = &obd->u.mds;
3883 struct llog_ctxt *ctxt;
3884 int rc, item = 0, valsize;
3888 LASSERT(!obd->obd_recovering);
3889 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3890 LASSERT(ctxt != NULL);
3892 /* clean PENDING dir */
3893 rc = mds_cleanup_orphans(obd);
3898 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3899 valsize = sizeof(group);
3900 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
3901 "mds_conn", valsize, &group);
3905 rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3908 CERROR("%s: failed at llog_origin_connect: %d\n",
3913 /* remove the orphaned precreated objects */
3914 rc = mds_dt_clear_orphans(mds, NULL /* all OSTs */);
3919 RETURN(rc < 0 ? rc : item);
3922 /* cleanup all llogging subsystems */
3923 rc = obd_llog_finish(obd, &obd->obd_llogs,
3924 mds->mds_dt_desc.ld_tgt_count);
3926 CERROR("%s: failed to cleanup llogging subsystems\n",
3931 int mds_postrecov(struct obd_device *obd)
3935 rc = mds_postrecov_common(obd);
3937 rc = mds_md_reconnect(obd);
3941 int mds_dt_clean(struct obd_device *obd)
3943 struct mds_obd *mds = &obd->u.mds;
3946 if (mds->mds_profile) {
3948 struct llog_ctxt *llctx;
3949 struct lvfs_run_ctxt saved;
3950 struct config_llog_instance cfg;
3951 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3953 OBD_ALLOC(cln_prof, len);
3954 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3956 cfg.cfg_instance = NULL;
3957 cfg.cfg_uuid = mds->mds_dt_uuid;
3959 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3960 llctx = llog_get_context(&obd->obd_llogs,
3961 LLOG_CONFIG_ORIG_CTXT);
3962 class_config_process_llog(llctx, cln_prof, &cfg);
3963 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3965 OBD_FREE(cln_prof, len);
3966 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3967 mds->mds_profile = NULL;
3972 int mds_md_clean(struct obd_device *obd)
3974 struct mds_obd *mds = &obd->u.mds;
3977 if (mds->mds_md_name) {
3978 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3979 mds->mds_md_name = NULL;
3984 static int mds_precleanup(struct obd_device *obd, int flags)
3990 mds_dt_disconnect(obd, flags);
3992 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3996 extern void lgss_svc_cache_purge_all(void);
3997 static int mds_cleanup(struct obd_device *obd, int flags)
3999 struct mds_obd *mds = &obd->u.mds;
4002 if (mds->mds_sb == NULL)
4005 mds_update_server_data(obd, 1);
4006 mds_update_last_fid(obd, NULL, 1);
4008 if (mds->mds_dt_objids != NULL) {
4009 int size = mds->mds_dt_desc.ld_tgt_count *
4011 OBD_FREE(mds->mds_dt_objids, size);
4013 mds_fs_cleanup(obd, flags);
4017 /* 2 seems normal on mds, (may_umount() also expects 2
4018 fwiw), but we only see 1 at this point in obdfilter. */
4019 lvfs_umount_fs(mds->mds_lvfs_ctxt);
4023 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
4025 spin_lock_bh(&obd->obd_processing_task_lock);
4026 if (obd->obd_recovering) {
4027 target_cancel_recovery_timer(obd);
4028 obd->obd_recovering = 0;
4030 spin_unlock_bh(&obd->obd_processing_task_lock);
4033 fsfilt_put_ops(obd->obd_fsops);
4037 lgss_svc_cache_purge_all();
4040 spin_lock(&mds->mds_denylist_lock);
4041 while (!list_empty( &mds->mds_denylist ) ) {
4042 deny_sec_t *p_deny_sec = list_entry(mds->mds_denylist.next,
4044 list_del(&p_deny_sec->list);
4045 OBD_FREE(p_deny_sec, sizeof(*p_deny_sec));
4047 spin_unlock(&mds->mds_denylist_lock);
4052 static int set_security(const char *value, char **sec)
4054 if (!strcmp(value, "null"))
4056 else if (!strcmp(value, "krb5i"))
4058 else if (!strcmp(value, "krb5p"))
4061 CERROR("Unrecognized security flavor %s\n", value);
4068 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
4070 struct lustre_cfg *lcfg = buf;
4071 struct mds_obd *mds = &obd->u.mds;
4075 switch(lcfg->lcfg_command) {
4076 case LCFG_SET_SECURITY: {
4077 if ((LUSTRE_CFG_BUFLEN(lcfg, 1) == 0) ||
4078 (LUSTRE_CFG_BUFLEN(lcfg, 2) == 0))
4079 GOTO(out, rc = -EINVAL);
4081 if (!strcmp(lustre_cfg_string(lcfg, 1), "mds_sec"))
4082 rc = set_security(lustre_cfg_string(lcfg, 2),
4084 else if (!strcmp(lustre_cfg_string(lcfg, 1), "oss_sec"))
4085 rc = set_security(lustre_cfg_string(lcfg, 2),
4087 else if (!strcmp(lustre_cfg_string(lcfg, 1), "deny_sec")){
4088 spin_lock(&mds->mds_denylist_lock);
4089 rc = add_deny_security(lustre_cfg_string(lcfg, 2),
4090 &mds->mds_denylist);
4091 spin_unlock(&mds->mds_denylist_lock);
4093 CERROR("Unrecognized key\n");
4099 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
4100 GOTO(out, rc = -EINVAL);
4106 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
4108 struct ldlm_lock *new_lock,
4109 struct ldlm_lock **old_lock,
4110 struct lustre_handle *lockh)
4112 struct obd_export *exp = req->rq_export;
4113 struct obd_device *obd = exp->exp_obd;
4114 struct ldlm_request *dlmreq =
4115 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
4116 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
4117 struct list_head *iter;
4119 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
4122 spin_lock(&obd->obd_namespace->ns_hash_lock);
4123 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
4124 struct ldlm_lock *lock;
4125 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
4126 if (lock == new_lock)
4128 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
4129 lockh->cookie = lock->l_handle.h_cookie;
4130 LDLM_DEBUG(lock, "restoring lock cookie");
4131 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
4134 *old_lock = LDLM_LOCK_GET(lock);
4135 spin_unlock(&obd->obd_namespace->ns_hash_lock);
4139 spin_unlock(&obd->obd_namespace->ns_hash_lock);
4141 /* If the xid matches, then we know this is a resent request,
4142 * and allow it. (It's probably an OPEN, for which we don't
4145 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
4149 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
4152 /* This remote handle isn't enqueued, so we never received or
4153 * processed this request. Clear MSG_RESENT, because it can
4154 * be handled like any normal request now. */
4156 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
4158 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
4162 int intent_disposition(struct ldlm_reply *rep, int flag)
4166 return (rep->lock_policy_res1 & flag);
4169 void intent_set_disposition(struct ldlm_reply *rep, int flag)
4173 rep->lock_policy_res1 |= flag;
4176 static int mds_intent_prepare_reply_buffers(struct ptlrpc_request *req,
4177 struct ldlm_intent *it)
4179 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
4180 int rc, reply_buffers;
4181 int repsize[5] = {sizeof(struct ldlm_reply),
4182 sizeof(struct mds_body),
4183 mds->mds_max_mdsize};
4187 if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
4188 if (req->rq_export->exp_mds_data.med_remote) {
4189 repsize[reply_buffers++] =
4190 sizeof(struct mds_remote_perm);
4192 repsize[reply_buffers++] = sizeof(int);
4193 repsize[reply_buffers++] =
4194 xattr_acl_size(LL_ACL_MAX_ENTRIES);
4196 /*FIXME: ugly here, should be optimize for there
4197 * is no crypto key*/
4198 repsize[reply_buffers++] = sizeof(int);
4199 repsize[reply_buffers++] = sizeof(struct crypto_key);
4202 rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
4207 static int mds_intent_policy(struct ldlm_namespace *ns,
4208 struct ldlm_lock **lockp, void *req_cookie,
4209 ldlm_mode_t mode, int flags, void *data)
4211 struct ptlrpc_request *req = req_cookie;
4212 struct ldlm_lock *lock = *lockp;
4213 struct ldlm_intent *it;
4214 struct ldlm_reply *rep;
4215 struct lustre_handle lockh[2] = {{0}, {0}};
4216 struct ldlm_lock *new_lock = NULL;
4217 int getattr_part = MDS_INODELOCK_UPDATE;
4220 int offset = MDS_REQ_INTENT_REC_OFF;
4223 LASSERT(req != NULL);
4224 MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
4226 if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
4227 /* No intent was provided */
4228 int size = sizeof(struct ldlm_reply);
4229 rc = lustre_pack_reply(req, 1, &size, NULL);
4234 it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
4235 lustre_swab_ldlm_intent);
4237 CERROR("Intent missing\n");
4238 RETURN(req->rq_status = -EFAULT);
4241 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
4243 rc = mds_intent_prepare_reply_buffers(req, it);
4246 RETURN(req->rq_status = rc);
4248 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
4249 LASSERT(rep != NULL);
4251 intent_set_disposition(rep, DISP_IT_EXECD);
4253 /* execute policy */
4254 switch ((long)it->opc) {
4256 case IT_CREAT|IT_OPEN:
4257 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4259 /* XXX swab here to assert that an mds_open reint
4260 * packet is following */
4261 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4263 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
4265 if (rep->lock_policy_res2) {
4267 * mds_open() returns ENOLCK where it should return
4268 * zero, but it has no lock to return.
4270 if (rep->lock_policy_res2 == ENOLCK)
4271 rep->lock_policy_res2 = 0;
4273 RETURN(ELDLM_LOCK_ABORTED);
4277 * IT_OPEN may return lock on cross-node dentry that we want to
4278 * hold during attr retrival -bzzz
4280 if (lockh[0].cookie == 0)
4281 RETURN(ELDLM_LOCK_ABORTED);
4285 getattr_part = MDS_INODELOCK_LOOKUP;
4288 getattr_part |= MDS_INODELOCK_LOOKUP;
4290 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4291 lock, &new_lock, lockh);
4292 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
4294 /* FIXME: LDLM can set req->rq_status. MDS sets
4295 policy_res{1,2} with disposition and status.
4296 - replay: returns 0 & req->status is old status
4297 - otherwise: returns req->status */
4298 if (intent_disposition(rep, DISP_LOOKUP_NEG))
4299 rep->lock_policy_res2 = 0;
4300 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
4301 rep->lock_policy_res2)
4302 RETURN(ELDLM_LOCK_ABORTED);
4303 if (req->rq_status != 0) {
4305 rep->lock_policy_res2 = req->rq_status;
4306 RETURN(ELDLM_LOCK_ABORTED);
4310 rc = mds_lock_and_check_slave(offset, req, lockh);
4311 if ((rep->lock_policy_res2 = rc)) {
4313 rep->lock_policy_res2 = 0;
4314 RETURN(ELDLM_LOCK_ABORTED);
4318 CERROR("Unhandled intent "LPD64"\n", it->opc);
4322 /* By this point, whatever function we called above must have either
4323 * filled in 'lockh', been an intent replay, or returned an error. We
4324 * want to allow replayed RPCs to not get a lock, since we would just
4325 * drop it below anyways because lock replay is done separately by the
4326 * client afterwards. For regular RPCs we want to give the new lock to
4327 * the client instead of whatever lock it was about to get. */
4328 if (new_lock == NULL)
4329 new_lock = ldlm_handle2lock(&lockh[0]);
4330 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
4333 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
4334 it->opc, lockh[0].cookie);
4336 /* If we've already given this lock to a client once, then we should
4337 * have no readers or writers. Otherwise, we should have one reader
4338 * _or_ writer ref (which will be zeroed below) before returning the
4339 * lock to a client. */
4340 if (new_lock->l_export == req->rq_export) {
4341 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
4343 LASSERT(new_lock->l_export == NULL);
4344 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
4349 if (new_lock->l_export == req->rq_export) {
4350 /* Already gave this to the client, which means that we
4351 * reconstructed a reply. */
4352 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
4354 RETURN(ELDLM_LOCK_REPLACED);
4357 /* Fixup the lock to be given to the client */
4358 lock_res_and_lock(new_lock);
4359 new_lock->l_readers = 0;
4360 new_lock->l_writers = 0;
4362 new_lock->l_export = class_export_get(req->rq_export);
4364 spin_lock(&new_lock->l_export->exp_ldlm_data.led_lock);
4365 list_add(&new_lock->l_export_chain,
4366 &new_lock->l_export->exp_ldlm_data.led_held_locks);
4367 spin_unlock(&new_lock->l_export->exp_ldlm_data.led_lock);
4369 new_lock->l_blocking_ast = lock->l_blocking_ast;
4370 new_lock->l_completion_ast = lock->l_completion_ast;
4372 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
4373 sizeof(lock->l_remote_handle));
4375 new_lock->l_flags &= ~LDLM_FL_LOCAL;
4377 unlock_res_and_lock(new_lock);
4378 LDLM_LOCK_PUT(new_lock);
4380 RETURN(ELDLM_LOCK_REPLACED);
4383 int mds_attach(struct obd_device *dev, obd_count len, void *data)
4385 struct lprocfs_static_vars lvars;
4387 struct mds_obd *mds = &dev->u.mds;
4389 spin_lock_init(&mds->mds_denylist_lock);
4390 INIT_LIST_HEAD(&mds->mds_denylist);
4392 lprocfs_init_multi_vars(0, &lvars);
4394 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
4398 return lprocfs_alloc_md_stats(dev, 0);
4401 int mds_detach(struct obd_device *dev)
4403 lprocfs_free_md_stats(dev);
4404 return lprocfs_obd_detach(dev);
4407 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
4409 struct lprocfs_static_vars lvars;
4411 lprocfs_init_multi_vars(1, &lvars);
4412 return lprocfs_obd_attach(dev, lvars.obd_vars);
4415 int mdt_detach(struct obd_device *dev)
4417 return lprocfs_obd_detach(dev);
4420 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
4422 struct mds_obd *mds = &obd->u.mds;
4427 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4428 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
4429 MDS_SERVICE_WATCHDOG_TIMEOUT,
4430 mds_handle, "mds", obd->obd_proc_entry);
4432 if (!mds->mds_service) {
4433 CERROR("failed to start service\n");
4437 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
4440 GOTO(err_thread, rc);
4442 mds->mds_setattr_service =
4443 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4444 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
4445 MDS_SERVICE_WATCHDOG_TIMEOUT,
4446 mds_handle, "mds_setattr",
4447 obd->obd_proc_entry);
4448 if (!mds->mds_setattr_service) {
4449 CERROR("failed to start getattr service\n");
4450 GOTO(err_thread, rc = -ENOMEM);
4453 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
4454 MDT_NUM_THREADS, "ll_mdt_attr");
4456 GOTO(err_thread2, rc);
4458 mds->mds_readpage_service =
4459 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4460 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
4461 MDS_SERVICE_WATCHDOG_TIMEOUT,
4462 mds_handle, "mds_readpage",
4463 obd->obd_proc_entry);
4464 if (!mds->mds_readpage_service) {
4465 CERROR("failed to start readpage service\n");
4466 GOTO(err_thread2, rc = -ENOMEM);
4469 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
4470 MDT_NUM_THREADS, "ll_mdt_rdpg");
4473 GOTO(err_thread3, rc);
4475 mds->mds_close_service =
4476 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4477 MDS_CLOSE_PORTAL, MDC_REPLY_PORTAL,
4478 MDS_SERVICE_WATCHDOG_TIMEOUT,
4479 mds_handle, "mds_close",
4480 obd->obd_proc_entry);
4481 if (!mds->mds_close_service) {
4482 CERROR("failed to start close service\n");
4483 GOTO(err_thread3, rc = -ENOMEM);
4486 rc = ptlrpc_start_n_threads(obd, mds->mds_close_service,
4487 MDT_NUM_THREADS, "ll_mdt_clos");
4490 GOTO(err_thread4, rc);
4494 ptlrpc_unregister_service(mds->mds_close_service);
4496 ptlrpc_unregister_service(mds->mds_readpage_service);
4498 ptlrpc_unregister_service(mds->mds_setattr_service);
4500 ptlrpc_unregister_service(mds->mds_service);
4504 static int mdt_cleanup(struct obd_device *obd, int flags)
4506 struct mds_obd *mds = &obd->u.mds;
4509 ptlrpc_stop_all_threads(mds->mds_close_service);
4510 ptlrpc_unregister_service(mds->mds_close_service);
4512 ptlrpc_stop_all_threads(mds->mds_readpage_service);
4513 ptlrpc_unregister_service(mds->mds_readpage_service);
4515 ptlrpc_stop_all_threads(mds->mds_setattr_service);
4516 ptlrpc_unregister_service(mds->mds_setattr_service);
4518 ptlrpc_stop_all_threads(mds->mds_service);
4519 ptlrpc_unregister_service(mds->mds_service);
4524 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
4525 __u64 gr, void *data)
4527 struct lustre_id id;
4528 struct obd_device *obd = data;
4532 return mds_id2dentry(obd, &id, NULL);
4535 static int mds_get_info(struct obd_export *exp, __u32 keylen,
4536 void *key, __u32 *valsize, void *val)
4538 struct obd_device *obd;
4539 struct mds_obd *mds;
4543 obd = class_exp2obd(exp);
4547 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
4548 exp->exp_handle.h_cookie);
4552 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
4553 /* get log_context handle. */
4554 struct llog_ctxt *ctxt;
4555 unsigned long *llh_handle = val;
4556 *valsize = sizeof(unsigned long);
4557 ctxt = llog_get_context(&obd->obd_llogs, LLOG_REINT_ORIG_CTXT);
4559 CERROR("Cannot get REINT llog context\n");
4562 *llh_handle = (unsigned long)ctxt;
4565 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
4566 /* get log_context handle. */
4567 unsigned long *sb = val;
4568 *valsize = sizeof(unsigned long);
4569 *sb = (unsigned long)obd->u.mds.mds_sb;
4573 if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
4574 __u32 *mdsize = val;
4575 *valsize = sizeof(*mdsize);
4576 *mdsize = mds->mds_max_mdsize;
4580 if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
4581 __u32 *mdsnum = val;
4582 *valsize = sizeof(*mdsnum);
4583 *mdsnum = mds->mds_num;
4587 if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
4588 struct lustre_id *rootid = val;
4589 *valsize = sizeof(*rootid);
4590 *rootid = mds->mds_rootid;
4594 if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
4595 struct lov_desc *desc = val;
4596 *valsize = sizeof(*desc);
4597 *desc = mds->mds_dt_desc;
4601 rc = fsfilt_get_info(obd, mds->mds_sb, NULL, keylen, key, valsize, val);
4603 CDEBUG(D_IOCTL, "invalid key\n");
4608 struct lvfs_callback_ops mds_lvfs_ops = {
4609 l_id2dentry: mds_lvfs_id2dentry,
4612 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
4613 int objcount, struct obd_ioobj *obj,
4614 int niocount, struct niobuf_remote *nb,
4615 struct niobuf_local *res,
4616 struct obd_trans_info *oti);
4618 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
4619 int objcount, struct obd_ioobj *obj, int niocount,
4620 struct niobuf_local *res, struct obd_trans_info *oti,
4623 /* use obd ops to offer management infrastructure */
4624 static struct obd_ops mds_obd_ops = {
4625 .o_owner = THIS_MODULE,
4626 .o_attach = mds_attach,
4627 .o_detach = mds_detach,
4628 .o_connect = mds_connect,
4629 .o_connect_post = mds_connect_post,
4630 .o_init_export = mds_init_export,
4631 .o_destroy_export = mds_destroy_export,
4632 .o_disconnect = mds_disconnect,
4633 .o_setup = mds_setup,
4634 .o_precleanup = mds_precleanup,
4635 .o_cleanup = mds_cleanup,
4636 .o_process_config = mds_process_config,
4637 .o_postrecov = mds_postrecov,
4638 .o_statfs = mds_obd_statfs,
4639 .o_iocontrol = mds_iocontrol,
4640 .o_create = mds_obd_create,
4641 .o_destroy = mds_obd_destroy,
4642 .o_llog_init = mds_llog_init,
4643 .o_llog_finish = mds_llog_finish,
4644 .o_notify = mds_notify,
4645 .o_get_info = mds_get_info,
4646 .o_set_info = mds_set_info,
4647 .o_preprw = mds_preprw,
4648 .o_commitrw = mds_commitrw,
4651 static struct obd_ops mdt_obd_ops = {
4652 .o_owner = THIS_MODULE,
4653 .o_attach = mdt_attach,
4654 .o_detach = mdt_detach,
4655 .o_setup = mdt_setup,
4656 .o_cleanup = mdt_cleanup,
4659 static int __init mds_init(void)
4661 struct lprocfs_static_vars lvars;
4663 mds_init_lsd_cache();
4664 mds_init_rmtacl_upcall_cache();
4666 lprocfs_init_multi_vars(0, &lvars);
4667 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4668 OBD_MDS_DEVICENAME);
4669 lprocfs_init_multi_vars(1, &lvars);
4670 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4671 OBD_MDT_DEVICENAME);
4676 static void /*__exit*/ mds_exit(void)
4678 mds_cleanup_rmtacl_upcall_cache();
4679 mds_cleanup_lsd_cache();
4681 class_unregister_type(OBD_MDS_DEVICENAME);
4682 class_unregister_type(OBD_MDT_DEVICENAME);
4685 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4686 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4687 MODULE_LICENSE("GPL");
4689 module_init(mds_init);
4690 module_exit(mds_exit);