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_sec.h>
60 #include <linux/lustre_gs.h>
61 #include <linux/lustre_audit.h>
63 #include "mds_internal.h"
65 extern int mds_audit_auth(struct ptlrpc_request *, struct lvfs_ucred *,
66 audit_op, struct lustre_id *,
68 extern int mds_audit_stat(struct ptlrpc_request *, struct lustre_id *,
69 struct dentry *, int);
70 extern int mds_audit_open(struct ptlrpc_request *, struct mds_update_record *,
73 static int mds_intent_policy(struct ldlm_namespace *ns,
74 struct ldlm_lock **lockp, void *req_cookie,
75 ldlm_mode_t mode, int flags, void *data);
76 static int mds_postsetup(struct obd_device *obd);
77 static int mds_cleanup(struct obd_device *obd, int flags);
80 /* Assumes caller has already pushed into the kernel filesystem context */
81 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
82 loff_t offset, int count)
84 struct ptlrpc_bulk_desc *desc;
85 struct l_wait_info lwi;
87 int rc = 0, npages, i, tmpcount, tmpsize = 0;
90 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
92 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
93 OBD_ALLOC(pages, sizeof(*pages) * npages);
95 GOTO(out, rc = -ENOMEM);
97 desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
100 GOTO(out_free, rc = -ENOMEM);
102 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
103 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
105 pages[i] = alloc_pages(GFP_KERNEL, 0);
106 if (pages[i] == NULL)
107 GOTO(cleanup_buf, rc = -ENOMEM);
109 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
112 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
113 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
114 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
115 tmpsize, offset, file->f_dentry->d_inode->i_ino,
116 file->f_dentry->d_inode->i_size);
118 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
119 kmap(pages[i]), tmpsize, &offset);
123 GOTO(cleanup_buf, rc = -EIO);
126 LASSERT(desc->bd_nob == count);
128 rc = ptlrpc_start_bulk_transfer(desc);
130 GOTO(cleanup_buf, rc);
132 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
133 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
134 OBD_FAIL_MDS_SENDPAGE, rc = -EIO);
135 GOTO(abort_bulk, rc);
138 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
139 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
140 LASSERT (rc == 0 || rc == -ETIMEDOUT);
143 if (desc->bd_success &&
144 desc->bd_nob_transferred == count)
145 GOTO(cleanup_buf, rc);
147 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
150 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
151 (rc == -ETIMEDOUT) ? "timeout" : "network error",
152 desc->bd_nob_transferred, count,
153 req->rq_export->exp_client_uuid.uuid,
154 req->rq_export->exp_connection->c_remote_uuid.uuid);
156 ptlrpc_fail_export(req->rq_export);
160 ptlrpc_abort_bulk (desc);
162 for (i = 0; i < npages; i++)
164 __free_pages(pages[i], 0);
166 ptlrpc_free_bulk(desc);
168 OBD_FREE(pages, sizeof(*pages) * npages);
173 extern char *ldlm_lockname[];
175 int mds_lock_mode_for_dir(struct obd_device *obd,
176 struct dentry *dentry, int mode)
178 int ret_mode = 0, split;
180 /* any dir access needs couple locks:
181 * 1) on part of dir we gonna lookup/modify in
182 * 2) on a whole dir to protect it from concurrent splitting
183 * and to flush client's cache for readdir()
184 * so, for a given mode and dentry this routine decides what
185 * lock mode to use for lock #2:
186 * 1) if caller's gonna lookup in dir then we need to protect
187 * dir from being splitted only - LCK_CR
188 * 2) if caller's gonna modify dir then we need to protect
189 * dir from being splitted and to flush cache - LCK_CW
190 * 3) if caller's gonna modify dir and that dir seems ready
191 * for splitting then we need to protect it from any
192 * type of access (lookup/modify/split) - LCK_EX -bzzz */
194 split = mds_splitting_expected(obd, dentry);
197 * it is important to check here only for MDS_NO_SPLITTABLE. The reason
198 * is that MDS_NO_SPLITTABLE means dir is not splittable in principle
199 * and another thread will not split it on the quiet. But if we have
200 * MDS_NO_SPLIT_EXPECTED, this means, that dir may be splitted anytime,
201 * but not now (for current thread) and we should consider that it can
202 * happen soon and go that branch which can yield LCK_EX to protect from
203 * possible splitting.
205 if (split == MDS_NO_SPLITTABLE) {
207 * this inode won't be splitted. so we need not to protect from
208 * just flush client's cache on modification.
215 if (mode == LCK_EX) {
217 } else if (mode == LCK_PR) {
219 } else if (mode == LCK_PW) {
221 * caller gonna modify directory. We use concurrent
222 * write lock here to retract client's cache for
225 if (split == MDS_EXPECT_SPLIT) {
227 * splitting possible. serialize any access the
228 * idea is that first one seen dir is splittable
229 * is given exclusive lock and split
230 * directory. caller passes lock mode to
231 * mds_try_to_split_dir() and splitting would be
232 * done with exclusive lock only -bzzz.
234 CDEBUG(D_OTHER, "%s: gonna split %lu/%lu\n",
236 (unsigned long)dentry->d_inode->i_ino,
237 (unsigned long)dentry->d_inode->i_generation);
248 /* only valid locked dentries or errors should be returned */
249 struct dentry *mds_id2locked_dentry(struct obd_device *obd, struct lustre_id *id,
250 struct vfsmount **mnt, int lock_mode,
251 struct lustre_handle *lockh, int *mode,
252 char *name, int namelen, __u64 lockpart)
254 struct dentry *de = mds_id2dentry(obd, id, mnt), *retval = de;
255 ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
256 struct ldlm_res_id res_id = { .name = {0} };
257 int flags = LDLM_FL_ATOMIC_CB, rc;
264 res_id.name[0] = id_fid(id);
265 res_id.name[1] = id_group(id);
268 if (name && IS_PDIROPS(de->d_inode)) {
269 ldlm_policy_data_t cpolicy =
270 { .l_inodebits = { MDS_INODELOCK_UPDATE } };
271 LASSERT(mode != NULL);
272 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
274 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
276 &cpolicy, *mode, &flags,
278 ldlm_completion_ast, NULL, NULL,
279 NULL, 0, NULL, lockh + 1);
280 if (rc != ELDLM_OK) {
282 RETURN(ERR_PTR(-ENOLCK));
285 flags = LDLM_FL_ATOMIC_CB;
287 res_id.name[2] = full_name_hash((unsigned char *)name, namelen);
289 CDEBUG(D_INFO, "take lock on "DLID4":"LPX64"\n",
290 OLID4(id), res_id.name[2]);
293 #warning "No PDIROPS support in the kernel"
295 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
296 LDLM_IBITS, &policy, lock_mode, &flags,
297 mds_blocking_ast, ldlm_completion_ast,
298 NULL, NULL, NULL, 0, NULL, lockh);
299 if (rc != ELDLM_OK) {
301 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
304 ldlm_lock_decref(lockh + 1, *mode);
306 } else if (de->d_inode && de->d_inode->i_nlink == 0) {
307 /* as sometimes we lookup inode by ino/generation through
308 iopen mechanism, it's possible to find already unlinked
309 inode with nlink == 0. let's interpretate the case as
311 CWARN("found already unlinked inode %lu/%u\n",
312 de->d_inode->i_ino, de->d_inode->i_generation);
314 retval = ERR_PTR(-ENOENT);
315 ldlm_lock_decref(lockh, lock_mode);
318 ldlm_lock_decref(lockh + 1, *mode);
325 #ifndef DCACHE_DISCONNECTED
326 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
329 /* Look up an entry by inode number. This function ONLY returns valid dget'd
330 * dentries with an initialized inode or errors */
331 struct dentry *mds_id2dentry(struct obd_device *obd, struct lustre_id *id,
332 struct vfsmount **mnt)
334 struct mds_obd *mds = &obd->u.mds;
335 struct dentry *result;
337 unsigned long ino = 0;
341 if (!id_ino(id) && id_fid(id)) {
342 struct lustre_id *lid;
344 /* if this is reint case we should use fidmap for resolving
345 * correct local store cookie. */
346 lid = mds_fidmap_lookup(obd, id);
348 RETURN(ERR_PTR(-ENOENT));
350 ino = (unsigned long)id_ino(lid);
351 generation = (__u32)id_gen(lid);
353 CDEBUG(D_DENTRY, "fidmap resolved "DLID4"->"DLID4"\n",
354 OLID4(id), OLID4(lid));
356 } else if (id_ino(id)) {
357 ino = (unsigned long)id_ino(id);
358 generation = (__u32)id_gen(id);
360 CERROR("invalid id for lookup "
361 DLID4"\n", OLID4(id));
365 RETURN(ERR_PTR(-ESTALE));
367 snprintf(idname, sizeof(idname), "0x%lx", ino);
369 CDEBUG(D_DENTRY, "--> mds_id2dentry: ino/gen %lu/%u, sb %p\n",
370 ino, generation, mds->mds_sb);
372 /* under ext3 this is neither supposed to return bad inodes nor NULL
374 result = ll_lookup_one_len(idname, mds->mds_id_de,
379 inode = result->d_inode;
381 RETURN(ERR_PTR(-ENOENT));
383 if (is_bad_inode(inode)) {
384 CERROR("bad inode returned %lu/%u\n",
385 inode->i_ino, inode->i_generation);
387 RETURN(ERR_PTR(-ENOENT));
390 /* here we disabled generation check, as root inode i_generation
391 * of cache mds and real mds are different. */
392 if (id_fid(id) != id_fid(&mds->mds_rootid) && generation != 0 &&
393 inode->i_generation != generation) {
394 /* we didn't find the right inode.. */
395 if (id_group(id) != mds->mds_num) {
396 CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
397 "%u != %u, mds %u != %u, request to wrong MDS?\n",
398 inode->i_ino, (unsigned long)inode->i_nlink,
399 atomic_read(&inode->i_count), inode->i_generation,
400 generation, mds->mds_num, (unsigned)id_group(id));
402 CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
403 "%u != %u, inode is recreated while request handled?\n",
404 inode->i_ino, (unsigned long)inode->i_nlink,
405 atomic_read(&inode->i_count), inode->i_generation,
409 RETURN(ERR_PTR(-ENOENT));
413 *mnt = mds->mds_vfsmnt;
421 int mds_req_add_idmapping(struct ptlrpc_request *req,
422 struct mds_export_data *med)
424 struct mds_req_sec_desc *rsd;
425 struct lustre_sec_desc *lsd;
428 if (!med->med_remote)
431 /* maybe we should do it more completely: invalidate the gss ctxt? */
432 if (req->rq_mapped_uid == MDS_IDMAP_NOTFOUND) {
433 CWARN("didn't find mapped uid\n");
437 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
439 CERROR("Can't unpack security desc\n");
443 lsd = mds_get_lsd(req->rq_mapped_uid);
445 CERROR("can't get LSD(%u), no mapping added\n",
450 rc = mds_idmap_add(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
451 rsd->rsd_gid, lsd->lsd_gid);
457 int mds_req_del_idmapping(struct ptlrpc_request *req,
458 struct mds_export_data *med)
460 struct mds_req_sec_desc *rsd;
461 struct lustre_sec_desc *lsd;
464 if (!med->med_remote)
467 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
469 CERROR("Can't unpack security desc\n");
473 LASSERT(req->rq_mapped_uid != -1);
474 lsd = mds_get_lsd(req->rq_mapped_uid);
476 CERROR("can't get LSD(%u), no idmapping deleted\n",
481 rc = mds_idmap_del(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
482 rsd->rsd_gid, lsd->lsd_gid);
487 static int mds_init_export_data(struct ptlrpc_request *req,
488 struct mds_export_data *med)
490 struct obd_connect_data *data, *reply;
491 int ask_remote, ask_local;
494 data = lustre_msg_buf(req->rq_reqmsg, 5, sizeof(*data));
495 reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*data));
496 LASSERT(data && reply);
498 if (med->med_initialized) {
499 CDEBUG(D_SEC, "med already initialized, reconnect?\n");
503 ask_remote = data->ocd_connect_flags & OBD_CONNECT_REMOTE;
504 ask_local = data->ocd_connect_flags & OBD_CONNECT_LOCAL;
506 /* currently the policy is simple: satisfy client as possible
509 if (req->rq_auth_uid == -1) {
511 CWARN("null sec is used, force to be local\n");
515 if (!req->rq_remote_realm)
516 CWARN("local realm asked to be remote\n");
518 } else if (ask_local) {
519 if (req->rq_remote_realm)
520 CWARN("remote realm asked to be local\n");
523 med->med_remote = (req->rq_remote_realm != 0);
526 med->med_nllu = data->ocd_nllu[0];
527 med->med_nllg = data->ocd_nllu[1];
529 med->med_initialized = 1;
531 reply->ocd_connect_flags &= ~(OBD_CONNECT_REMOTE | OBD_CONNECT_LOCAL);
532 if (med->med_remote) {
534 med->med_idmap = mds_idmap_alloc();
537 CERROR("Failed to alloc idmap, following request from "
538 "this client will be refused\n");
540 reply->ocd_connect_flags |= OBD_CONNECT_REMOTE;
541 CDEBUG(D_SEC, "set client as remote\n");
543 reply->ocd_connect_flags |= OBD_CONNECT_LOCAL;
544 CDEBUG(D_SEC, "set client as local\n");
550 static void mds_free_export_data(struct mds_export_data *med)
555 LASSERT(med->med_remote);
556 mds_idmap_free(med->med_idmap);
557 med->med_idmap = NULL;
560 /* Establish a connection to the MDS.
562 * This will set up an export structure for the client to hold state data about
563 * that client, like open files, the last operation number it did on the server,
566 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
567 struct obd_uuid *cluuid, struct obd_connect_data *data,
570 struct mds_export_data *med;
571 struct mds_client_data *mcd;
572 struct obd_export *exp;
576 if (!conn || !obd || !cluuid)
579 /* XXX There is a small race between checking the list and adding a new
580 * connection for the same UUID, but the real threat (list corruption
581 * when multiple different clients connect) is solved.
583 * There is a second race between adding the export to the list, and
584 * filling in the client data below. Hence skipping the case of NULL
585 * mcd above. We should already be controlling multiple connects at the
586 * client, and we can't hold the spinlock over memory allocations
587 * without risk of deadlocking.
589 rc = class_connect(conn, obd, cluuid);
592 exp = class_conn2export(conn);
594 LASSERT(exp != NULL);
595 med = &exp->exp_mds_data;
597 OBD_ALLOC(mcd, sizeof(*mcd));
599 CERROR("%s: out of memory for client data.\n",
601 GOTO(out, rc = -ENOMEM);
604 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
607 rc = mds_client_add(obd, &obd->u.mds, med, -1);
615 OBD_FREE(mcd, sizeof(*mcd));
616 class_disconnect(exp, 0);
618 class_export_put(exp);
623 static int mds_connect_post(struct obd_export *exp, unsigned initial,
626 struct obd_device *obd = exp->exp_obd;
627 struct mds_obd *mds = &obd->u.mds;
628 struct mds_export_data *med;
629 struct mds_client_data *mcd;
633 med = &exp->exp_mds_data;
637 /* some one reconnect initially, we have to reset
638 * data existing export can have. bug 6102 */
639 if (mcd->mcd_last_xid != 0)
640 CDEBUG(D_HA, "initial reconnect to existing export\n");
641 mcd->mcd_last_transno = 0;
642 mcd->mcd_last_xid = 0;
643 mcd->mcd_last_close_xid = 0;
644 mcd->mcd_last_result = 0;
645 mcd->mcd_last_data = 0;
648 if (!(flags & OBD_OPT_MDS_CONNECTION)) {
649 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
650 atomic_inc(&mds->mds_real_clients);
651 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
652 obd->obd_name, exp->exp_client_uuid.uuid,
653 atomic_read(&mds->mds_real_clients));
654 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
656 if (mds->mds_md_name)
657 rc = mds_md_connect(obd, mds->mds_md_name);
662 static int mds_init_export(struct obd_export *exp)
664 struct mds_export_data *med = &exp->exp_mds_data;
666 INIT_LIST_HEAD(&med->med_open_head);
667 spin_lock_init(&med->med_open_lock);
671 static int mds_destroy_export(struct obd_export *export)
673 struct obd_device *obd = export->exp_obd;
674 struct mds_export_data *med = &export->exp_mds_data;
675 struct lvfs_run_ctxt saved;
679 mds_free_export_data(med);
680 target_destroy_export(export);
682 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
685 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
687 /* Close any open files (which may also cause orphan unlinking). */
688 spin_lock(&med->med_open_lock);
689 while (!list_empty(&med->med_open_head)) {
690 struct list_head *tmp = med->med_open_head.next;
691 struct mds_file_data *mfd =
692 list_entry(tmp, struct mds_file_data, mfd_list);
693 struct lustre_id sid;
695 BDEVNAME_DECLARE_STORAGE(btmp);
697 /* bug 1579: fix force-closing for 2.5 */
698 struct dentry *dentry = mfd->mfd_dentry;
700 list_del(&mfd->mfd_list);
701 spin_unlock(&med->med_open_lock);
703 down(&dentry->d_inode->i_sem);
704 rc = mds_read_inode_sid(obd, dentry->d_inode, &sid);
705 up(&dentry->d_inode->i_sem);
707 CERROR("Can't read inode self id, inode %lu, "
708 "rc %d\n", dentry->d_inode->i_ino, rc);
709 memset(&sid, 0, sizeof(sid));
712 /* If you change this message, be sure to update
713 * replay_single:test_46 */
714 CERROR("force closing client file handle for %.*s (%s:"
715 DLID4")\n", dentry->d_name.len, dentry->d_name.name,
716 ll_bdevname(dentry->d_inode->i_sb, btmp),
719 /* child inode->i_alloc_sem protects orphan_dec_test and
720 * is_orphan race, mds_mfd_close drops it */
721 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
722 rc = mds_mfd_close(NULL, 0, obd, mfd,
723 !(export->exp_flags & OBD_OPT_FAILOVER));
725 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
726 spin_lock(&med->med_open_lock);
728 spin_unlock(&med->med_open_lock);
729 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
733 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
737 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
739 unsigned long irqflags;
740 struct obd_device *obd;
745 LASSERT(exp != NULL);
746 obd = class_exp2obd(exp);
748 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
749 exp->exp_handle.h_cookie);
755 * suppress any inter-mds requests durring disconnecting lmv if this is
756 * detected --force mode. This is needed to avoid endless recovery.
758 if (atomic_read(&mds->mds_real_clients) > 0 &&
759 !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
760 flags |= OBD_OPT_FORCE;
762 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
763 && !atomic_read(&mds->mds_real_clients)) {
764 /* there was no client at all */
765 mds_md_disconnect(obd, flags);
768 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
769 && atomic_dec_and_test(&mds->mds_real_clients)) {
770 /* time to drop LMV connections */
771 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
772 "Disconnnect from LMV now\n",
773 obd->obd_name, exp->exp_client_uuid.uuid);
774 mds_md_disconnect(obd, flags);
777 spin_lock_irqsave(&exp->exp_lock, irqflags);
778 exp->exp_flags = flags;
779 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
781 /* disconnect early so that clients can't keep using export */
782 rc = class_disconnect(exp, flags);
783 ldlm_cancel_locks_for_export(exp);
785 /* complete all outstanding replies */
786 spin_lock_irqsave(&exp->exp_lock, irqflags);
787 while (!list_empty(&exp->exp_outstanding_replies)) {
788 struct ptlrpc_reply_state *rs =
789 list_entry(exp->exp_outstanding_replies.next,
790 struct ptlrpc_reply_state, rs_exp_list);
791 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
793 spin_lock(&svc->srv_lock);
794 list_del_init(&rs->rs_exp_list);
795 ptlrpc_schedule_difficult_reply(rs);
796 spin_unlock(&svc->srv_lock);
798 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
802 static int mds_getstatus(struct ptlrpc_request *req)
804 struct mds_obd *mds = mds_req2mds(req);
805 struct mds_body *body;
809 size = sizeof(*body);
811 rc = lustre_pack_reply(req, 1, &size, NULL);
812 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
813 CERROR("mds: out of memory for message: size=%d\n", size);
814 req->rq_status = -ENOMEM; /* superfluous? */
818 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
819 body->valid |= OBD_MD_FID;
821 memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
824 * the last_committed and last_xid fields are filled in for all replies
825 * already - no need to do so here also.
830 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
831 void *data, int flag)
836 if (flag == LDLM_CB_CANCELING) {
837 /* Don't need to do anything here. */
841 /* XXX layering violation! -phil */
842 lock_res_and_lock(lock);
845 * get this: if mds_blocking_ast is racing with mds_intent_policy, such
846 * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
847 * then by the time we get the lock, we might not be the correct
848 * blocking function anymore. So check, and return early, if so.
850 if (lock->l_blocking_ast != mds_blocking_ast) {
851 unlock_res_and_lock(lock);
855 lock->l_flags |= LDLM_FL_CBPENDING;
856 do_ast = (!lock->l_readers && !lock->l_writers);
857 unlock_res_and_lock(lock);
860 struct lustre_handle lockh;
863 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
864 ldlm_lock2handle(lock, &lockh);
865 rc = ldlm_cli_cancel(&lockh);
867 CERROR("ldlm_cli_cancel: %d\n", rc);
869 LDLM_DEBUG(lock, "Lock still has references, will be "
875 static int mds_convert_md(struct obd_device *obd, struct inode *inode,
876 void *md, int size, int mea)
880 if (S_ISREG(inode->i_mode)) {
881 rc = mds_convert_lov_ea(obd, inode, md, size);
882 } else if (S_ISDIR(inode->i_mode)) {
884 rc = mds_convert_mea_ea(obd, inode, md, size);
886 rc = mds_convert_lov_ea(obd, inode, md, size);
889 CERROR("Invalid EA format (nor LOV or MEA) "
890 "is detected. Inode %lu/%u\n",
891 inode->i_ino, inode->i_generation);
897 int mds_get_md(struct obd_device *obd, struct inode *inode,
898 void *md, int *size, int lock, int mea)
907 rc = fsfilt_get_md(obd, inode, md, *size,
908 (mea ? EA_MEA : EA_LOV));
910 CERROR("Error %d reading eadata for ino %lu\n",
914 rc = mds_convert_md(obd, inode, md,
929 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
930 * Call with lock=0 if the caller has already taken the i_sem. */
931 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
932 struct mds_body *body, struct inode *inode, int lock, int mea)
934 struct mds_obd *mds = &obd->u.mds;
939 lmm = lustre_msg_buf(msg, offset, 0);
941 /* Some problem with getting eadata when I sized the reply
943 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
947 lmm_size = msg->buflens[offset];
949 /* I don't really like this, but it is a sanity check on the client
950 * MD request. However, if the client doesn't know how much space
951 * to reserve for the MD, it shouldn't be bad to have too much space.
953 if (lmm_size > mds->mds_max_mdsize) {
954 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
955 inode->i_ino, lmm_size, mds->mds_max_mdsize);
959 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, mea);
961 body->valid |= S_ISDIR(inode->i_mode) ?
962 OBD_MD_FLDIREA : OBD_MD_FLEASIZE;
965 body->valid |= OBD_MD_MEA;
967 body->eadatasize = lmm_size;
974 int mds_pack_link(struct dentry *dentry, struct ptlrpc_request *req,
975 struct mds_body *repbody, int reply_off)
977 struct inode *inode = dentry->d_inode;
982 symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
983 LASSERT(symname != NULL);
984 len = req->rq_repmsg->buflens[reply_off + 1];
986 rc = inode->i_op->readlink(dentry, symname, len);
988 CERROR("readlink failed: %d\n", rc);
989 } else if (rc != len - 1) {
990 CERROR ("Unexpected readlink rc %d: expecting %d\n",
994 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
995 repbody->valid |= OBD_MD_LINKNAME;
996 repbody->eadatasize = rc + 1;
997 symname[rc] = 0; /* NULL terminate */
1004 int mds_pack_xattr(struct dentry *dentry, struct ptlrpc_request *req,
1005 struct mds_body *repbody, int req_off, int reply_off)
1007 struct inode *inode = dentry->d_inode;
1013 ea_name = lustre_msg_string(req->rq_reqmsg, req_off + 1, 0);
1014 len = req->rq_repmsg->buflens[reply_off + 1];
1016 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
1020 if (!strcmp(ea_name, XATTR_NAME_LUSTRE_ACL)) {
1021 struct rmtacl_upcall_desc desc;
1023 if (len != LUSTRE_ACL_SIZE_MAX || !value) {
1024 CERROR("no reply buffer prepared\n");
1028 memset(&desc, 0, sizeof(desc));
1030 desc.cmd = lustre_msg_string(req->rq_reqmsg, req_off + 2, 0);
1031 desc.cmdlen = req->rq_reqmsg->buflens[req_off + 2];
1032 desc.res = (char *) value;
1033 desc.reslen = LUSTRE_ACL_SIZE_MAX;
1035 mds_do_remote_acl_upcall(&desc);
1037 if (desc.upcall_status)
1038 RETURN(desc.upcall_status);
1040 if (desc.reslen > LUSTRE_ACL_SIZE_MAX) {
1041 CERROR("downcall claim reslen %u\n", desc.reslen);
1044 /* like remote setfacl, steal "flags" in mds_body as the
1045 * exececution status
1047 repbody->flags = desc.status;
1048 repbody->valid |= OBD_MD_FLXATTR;
1049 repbody->eadatasize = desc.reslen;
1054 if (inode->i_op && inode->i_op->getxattr)
1055 rc = inode->i_op->getxattr(dentry, ea_name, value, len);
1058 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1059 CERROR("getxattr failed: %d", rc);
1061 repbody->valid |= OBD_MD_FLXATTR;
1062 repbody->eadatasize = rc;
1069 int mds_pack_xattr_list(struct dentry *dentry, struct ptlrpc_request *req,
1070 struct mds_body *repbody, int reply_off)
1072 struct inode *inode = dentry->d_inode;
1077 len = req->rq_repmsg->buflens[reply_off + 1];
1079 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
1082 if (inode->i_op && inode->i_op->getxattr)
1083 rc = inode->i_op->listxattr(dentry, value, len);
1086 CERROR("listxattr failed: %d", rc);
1088 repbody->valid |= OBD_MD_FLXATTRLIST;
1089 repbody->eadatasize = rc;
1096 int mds_pack_posix_acl(struct lustre_msg *repmsg, int *offset,
1097 struct mds_body *body, struct inode *inode)
1099 struct dentry de = { .d_inode = inode };
1100 __u32 buflen, *sizep;
1102 int size, pack_off = *offset;
1105 sizep = lustre_msg_buf(repmsg, pack_off++, 4);
1107 CERROR("can't locate returned acl size buf\n");
1111 if (!inode->i_op->getxattr)
1114 buflen = repmsg->buflens[pack_off];
1115 buf = lustre_msg_buf(repmsg, pack_off++, buflen);
1118 size = inode->i_op->getxattr(&de, XATTR_NAME_ACL_ACCESS, buf, buflen);
1119 if (size == -ENODATA || size == -EOPNOTSUPP)
1124 body->valid |= OBD_MD_FLACL;
1126 *sizep = cpu_to_le32(size);
1130 int mds_pack_remote_perm(struct ptlrpc_request *req, int *reply_off,
1131 struct mds_body *body, struct inode *inode)
1133 struct lustre_sec_desc *lsd;
1134 struct mds_remote_perm *perm;
1135 int pack_off = *reply_off;
1138 LASSERT(inode->i_op);
1139 LASSERT(inode->i_op->permission);
1140 LASSERT(req->rq_export->exp_mds_data.med_remote);
1142 perm = (struct mds_remote_perm *)
1143 lustre_msg_buf(req->rq_repmsg, pack_off++, sizeof(perm));
1145 CERROR("no remote perm buf at offset %d\n", pack_off - 1);
1149 memset(perm, 0, sizeof(*perm));
1151 /* obtain authenticated uid/gid and LSD permissions, which
1152 * might be different from current process context, from LSD
1154 lsd = mds_get_lsd(current->uid);
1156 CWARN("can't LSD of uid %u\n", current->uid);
1160 perm->mrp_auth_uid = lsd->lsd_uid;
1161 perm->mrp_auth_gid = lsd->lsd_gid;
1163 lsd_perms = mds_lsd_get_perms(lsd, 1, 0, req->rq_peer.peer_id.nid);
1164 if (lsd_perms & LSD_PERM_SETUID)
1165 perm->mrp_allow_setuid = 1;
1166 if (lsd_perms & LSD_PERM_SETGID)
1167 perm->mrp_allow_setgid = 1;
1171 /* permission bits of current user
1172 * XXX this is low efficient, could we do it in one blow?
1174 if (inode->i_op->permission(inode, MAY_EXEC, NULL) == 0)
1175 perm->mrp_perm |= MAY_EXEC;
1176 if (inode->i_op->permission(inode, MAY_WRITE, NULL) == 0)
1177 perm->mrp_perm |= MAY_WRITE;
1178 if (inode->i_op->permission(inode, MAY_READ, NULL) == 0)
1179 perm->mrp_perm |= MAY_READ;
1181 body->valid |= (OBD_MD_FLACL | OBD_MD_FLRMTACL);
1183 *reply_off = pack_off;
1188 int mds_pack_acl(struct ptlrpc_request *req, int *reply_off,
1189 struct mds_body *body, struct inode *inode)
1193 if (!req->rq_export->exp_mds_data.med_remote)
1194 rc = mds_pack_posix_acl(req->rq_repmsg, reply_off, body, inode);
1196 rc = mds_pack_remote_perm(req, reply_off, body, inode);
1201 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
1202 struct ptlrpc_request *req, int req_off,
1203 struct mds_body *reqbody, int reply_off,
1204 struct mds_req_sec_desc *rsd)
1206 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1207 struct inode *inode = dentry->d_inode;
1208 struct mds_body *body;
1209 int rc = 0, offset = 0;
1212 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
1215 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
1216 LASSERT(body != NULL); /* caller prepped reply */
1218 if (dentry->d_flags & DCACHE_CROSS_REF) {
1219 mds_pack_dentry2body(obd, body, dentry,
1220 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1221 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
1226 mds_pack_inode2body(obd, body, inode,
1227 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1229 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
1230 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
1232 /* guessing what kind og attribute do we need. */
1233 int is_mea = (S_ISDIR(inode->i_mode) &&
1234 (reqbody->valid & OBD_MD_MEA) != 0);
1236 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1,
1237 body, inode, 1, is_mea);
1239 /* if we have LOV EA data, the OST holds size, atime, mtime. */
1240 if (!(body->valid & OBD_MD_FLEASIZE) &&
1241 !(body->valid & OBD_MD_FLDIREA))
1242 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1243 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1244 } else if (S_ISLNK(inode->i_mode) &&
1245 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
1246 rc = mds_pack_link(dentry, req, body, reply_off);
1247 } else if (reqbody->valid & OBD_MD_FLXATTR) {
1248 rc = mds_pack_xattr(dentry, req, body, req_off, reply_off);
1249 } else if (reqbody->valid & OBD_MD_FLXATTRLIST) {
1250 rc = mds_pack_xattr_list(dentry, req, body, reply_off);
1253 offset = reply_off + ((reqbody->valid & OBD_MD_FLEASIZE) ? 2 : 1);
1254 if (reqbody->valid & OBD_MD_FLACL) {
1255 rc = mds_pack_acl(req, &offset, body, inode);
1258 if (reqbody->valid & OBD_MD_FLKEY) {
1259 rc = mds_pack_gskey(obd, req->rq_repmsg, &offset,
1263 mds_pack_audit(obd, inode, body);
1265 if (reqbody->valid & OBD_MD_CAPA) {
1266 struct lustre_capa *req_capa;
1268 LASSERT(!(reqbody->valid & ~OBD_MD_CAPA));
1269 LASSERT(S_ISREG(inode->i_mode));
1271 req_capa = lustre_swab_reqbuf(req, req_off + 1,
1273 lustre_swab_lustre_capa);
1274 if (req_capa == NULL) {
1275 CERROR("Can't unpack capa\n");
1279 offset = reply_off + 1;
1280 rc = mds_pack_capa(obd, reqbody, req_capa, req->rq_repmsg,
1285 mds_body_do_reverse_map(med, body);
1290 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
1291 struct dentry *dentry,
1294 int rc = 0, size[1] = {sizeof(struct mds_body)};
1297 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1298 CERROR("failed MDS_GETATTR_PACK test\n");
1299 req->rq_status = -ENOMEM;
1303 rc = lustre_pack_reply(req, 1, size, NULL);
1305 CERROR("lustre_pack_reply failed: rc %d\n", rc);
1306 GOTO(out, req->rq_status = rc);
1314 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct dentry *de,
1317 struct inode *inode = de->d_inode;
1318 struct mds_obd *mds = mds_req2mds(req);
1319 struct mds_body *body;
1320 int rc = 0, size[4] = {sizeof(*body)}, bufcount = 1;
1323 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1324 LASSERT(body != NULL); /* checked by caller */
1325 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
1327 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
1328 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
1331 down(&inode->i_sem);
1332 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
1333 ((body->valid & OBD_MD_MEA) ? EA_MEA : EA_LOV));
1336 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1337 CERROR("error getting inode %lu MD: rc = %d\n",
1340 } else if (rc > mds->mds_max_mdsize) {
1342 CERROR("MD size %d larger than maximum possible %u\n",
1343 rc, mds->mds_max_mdsize);
1345 size[bufcount] = rc;
1348 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
1349 if (inode->i_size + 1 != body->eadatasize)
1350 CERROR("symlink size: %Lu, reply space: %d\n",
1351 inode->i_size + 1, body->eadatasize);
1352 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
1354 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
1355 inode->i_size + 1, body->eadatasize);
1356 } else if ((body->valid & OBD_MD_FLXATTR)) {
1357 char *ea_name = lustre_msg_string(req->rq_reqmsg,
1361 if (!strcmp(ea_name, XATTR_NAME_LUSTRE_ACL)) {
1362 size[bufcount] = LUSTRE_ACL_SIZE_MAX;
1364 if (inode->i_op && inode->i_op->getxattr)
1365 rc = inode->i_op->getxattr(de, ea_name,
1369 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1370 CERROR("error get inode %lu EA: %d\n",
1374 size[bufcount] = min_t(int,
1375 body->eadatasize, rc);
1379 } else if (body->valid & OBD_MD_FLXATTRLIST) {
1381 if (inode->i_op && inode->i_op->getxattr)
1382 rc = inode->i_op->listxattr(de, NULL, 0);
1385 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1386 CERROR("error getting inode %lu EA: rc = %d\n",
1390 size[bufcount] = min_t(int, body->eadatasize, rc);
1395 /* may co-exist with OBD_MD_FLEASIZE */
1396 if (body->valid & OBD_MD_FLACL) {
1397 if (req->rq_export->exp_mds_data.med_remote) {
1398 size[bufcount++] = sizeof(struct mds_remote_perm);
1400 size[bufcount++] = sizeof(int);
1401 size[bufcount++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
1405 if (body->valid & OBD_MD_FLKEY) {
1406 size[bufcount++] = sizeof(int);
1407 size[bufcount++] = sizeof(struct crypto_key);
1410 if (body->valid & OBD_MD_CAPA)
1411 size[bufcount++] = sizeof(struct lustre_capa);
1413 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1414 CERROR("failed MDS_GETATTR_PACK test\n");
1415 req->rq_status = -ENOMEM;
1416 GOTO(out, rc = -ENOMEM);
1419 rc = lustre_pack_reply(req, bufcount, size, NULL);
1421 CERROR("out of memory\n");
1422 GOTO(out, req->rq_status = rc);
1430 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
1431 char *name, int namelen)
1433 struct mea *mea = NULL;
1434 int mea_size, rc = 0;
1437 rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
1440 if (mea != NULL && mea->mea_count) {
1442 * dir is already splitted, check if requested filename should
1443 * live at this MDS or at another one.
1445 int i = mea_name2idx(mea, name, namelen - 1);
1446 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1448 "inapropriate MDS(%d) for %s. should be "
1449 "%lu(%d)\n", mea->mea_master, name,
1450 (unsigned long)id_group(&mea->mea_ids[i]), i);
1456 OBD_FREE(mea, mea_size);
1460 int mds_getattr_size(struct obd_device *obd, struct dentry *dentry,
1461 struct ptlrpc_request *req, struct mds_body *body)
1463 struct inode *inode = dentry->d_inode;
1466 LASSERT(body != NULL);
1468 if (dentry->d_inode == NULL || !S_ISREG(inode->i_mode))
1471 if (obd->obd_recovering) {
1472 CDEBUG(D_INODE, "size for "DLID4" is unknown yet (recovering)\n",
1477 if (atomic_read(&inode->i_writecount)) {
1478 /* some one has opened the file for write.
1479 * mds doesn't know actual size */
1480 CDEBUG(D_INODE, "MDS doesn't know actual size for "DLID4"\n",
1484 CDEBUG(D_INODE, "MDS returns "LPD64"/"LPD64" for"DLID4"\n",
1485 body->size, body->blocks, OLID4(&body->id1));
1486 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1490 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
1491 struct lustre_handle *child_lockh, int child_part)
1493 struct obd_device *obd = req->rq_export->exp_obd;
1494 struct mds_obd *mds = &obd->u.mds;
1495 struct ldlm_reply *rep = NULL;
1496 struct lvfs_run_ctxt saved;
1497 struct mds_req_sec_desc *rsd;
1498 struct mds_body *body;
1499 struct dentry *dparent = NULL, *dchild = NULL;
1500 struct lvfs_ucred uc = {NULL, NULL,};
1501 struct lustre_handle parent_lockh[2] = {{0}, {0}};
1502 unsigned int namesize;
1503 int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
1507 LASSERT(!strcmp(obd->obd_type->typ_name, OBD_MDS_DEVICENAME));
1508 MD_COUNTER_INCREMENT(obd, getattr_lock);
1510 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1512 CERROR("Can't unpack security desc\n");
1516 /* swab now, before anyone looks inside the request. */
1517 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1518 lustre_swab_mds_body);
1520 CERROR("Can't swab mds_body\n");
1521 GOTO(cleanup, rc = -EFAULT);
1524 LASSERT_REQSWAB(req, offset + 1);
1525 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
1527 CERROR("Can't unpack name\n");
1528 GOTO(cleanup, rc = -EFAULT);
1530 namesize = req->rq_reqmsg->buflens[offset + 1];
1532 /* namesize less than 2 means we have empty name, probably came from
1533 revalidate by cfid, so no point in having name to be set */
1537 LASSERT (offset == 1 || offset == 3);
1539 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1545 rc = mds_init_ucred(&uc, req, rsd);
1547 if (child_lockh->cookie == 0)
1548 mds_audit_auth(req, &uc, AUDIT_STAT, &body->id1,
1553 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1554 cleanup_phase = 1; /* kernel context */
1555 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1557 LASSERT(namesize > 0);
1558 if (child_lockh->cookie != 0) {
1559 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1563 if (body->valid == OBD_MD_FLID) {
1564 struct mds_body *mds_reply;
1565 int size = sizeof(*mds_reply);
1569 dparent = mds_id2dentry(obd, &body->id1, NULL);
1570 if (IS_ERR(dparent)) {
1571 rc = PTR_ERR(dparent);
1575 * the user requested ONLY the inode number, so do a raw lookup.
1577 rc = lustre_pack_reply(req, 1, &size, NULL);
1579 CERROR("out of memory\n");
1583 dir = dparent->d_inode;
1584 LASSERT(dir->i_op->lookup_raw != NULL);
1585 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1587 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1588 sizeof(*mds_reply));
1590 id_ino(&mds_reply->id1) = inum;
1591 mds_reply->valid = OBD_MD_FLID;
1595 if (resent_req == 0) {
1596 LASSERT(id_fid(&body->id1) != 0);
1598 rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1599 parent_lockh, &dparent,
1601 MDS_INODELOCK_UPDATE,
1604 child_lockh, &dchild,
1605 LCK_PR, child_part);
1609 cleanup_phase = 2; /* dchild, dparent, locks */
1612 * let's make sure this name should leave on this mds
1615 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1619 /* we have no dentry here, drop LOOKUP bit */
1620 /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1621 child_part &= ~MDS_INODELOCK_LOOKUP;
1622 CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1623 obd->obd_name, OLID4(&body->id1));
1625 dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1626 LCK_PR, parent_lockh,
1627 &update_mode, NULL, 0,
1628 MDS_INODELOCK_UPDATE);
1629 if (IS_ERR(dchild)) {
1630 CERROR("can't find inode with id "DLID4", err = %d\n",
1631 OLID4(&body->id1), (int)PTR_ERR(dchild));
1632 GOTO(cleanup, rc = PTR_ERR(dchild));
1634 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
1637 struct ldlm_lock *granted_lock;
1639 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1640 granted_lock = ldlm_handle2lock(child_lockh);
1642 LASSERTF(granted_lock != NULL, LPU64"/%lu lockh "LPX64"\n",
1643 id_fid(&body->id1), (unsigned long)id_group(&body->id1),
1644 child_lockh->cookie);
1647 /* usual named request */
1648 dparent = mds_id2dentry(obd, &body->id1, NULL);
1649 LASSERT(!IS_ERR(dparent));
1650 dchild = ll_lookup_one_len(name, dparent, namesize - 1);
1651 LASSERT(!IS_ERR(dchild));
1653 /* client wants to get attr. by id */
1654 dchild = mds_id2dentry(obd, &body->id1, NULL);
1655 LASSERT(!IS_ERR(dchild));
1657 LDLM_LOCK_PUT(granted_lock);
1660 cleanup_phase = 2; /* dchild, dparent, locks */
1662 if (!DENTRY_VALID(dchild)) {
1663 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1665 * in the intent case, the policy clears this error: the
1666 * disposition is enough.
1671 intent_set_disposition(rep, DISP_LOOKUP_POS);
1674 if (req->rq_repmsg == NULL) {
1675 if (dchild->d_flags & DCACHE_CROSS_REF)
1676 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1678 rc = mds_getattr_pack_msg(req, dchild, offset);
1680 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1685 rc = mds_getattr_internal(obd, dchild, req, offset, body,
1688 GOTO(cleanup, rc); /* returns the lock to the client */
1690 /* probably MDS knows actual size? */
1691 body = lustre_msg_buf(req->rq_repmsg, reply_offset, sizeof(*body));
1692 LASSERT(body != NULL);
1693 mds_getattr_size(obd, dchild, req, body);
1698 switch (cleanup_phase) {
1700 if (resent_req == 0) {
1701 if (rc && DENTRY_VALID(dchild))
1702 ldlm_lock_decref(child_lockh, LCK_PR);
1704 ldlm_lock_decref(parent_lockh, LCK_PR);
1706 if (parent_lockh[1].cookie != 0)
1707 ldlm_lock_decref(parent_lockh + 1, update_mode);
1712 /* audit stuff for getattr */
1713 if (dchild->d_inode)
1714 mds_audit_stat(req, &body->id1, dchild, rc);
1718 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1720 mds_exit_ucred(&uc);
1725 static int mds_getattr(struct ptlrpc_request *req, int offset)
1727 struct obd_device *obd = req->rq_export->exp_obd;
1728 struct lvfs_run_ctxt saved;
1730 struct mds_req_sec_desc *rsd;
1731 struct mds_body *body;
1732 struct lvfs_ucred uc = {NULL, NULL,};
1736 MD_COUNTER_INCREMENT(obd, getattr);
1738 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1740 CERROR("Can't unpack security desc\n");
1744 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1745 lustre_swab_mds_body);
1747 CERROR ("Can't unpack body\n");
1751 rc = mds_init_ucred(&uc, req, rsd);
1753 mds_exit_ucred(&uc);
1757 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1758 de = mds_id2dentry(obd, &body->id1, NULL);
1760 rc = req->rq_status = PTR_ERR(de);
1764 rc = mds_getattr_pack_msg(req, de, offset);
1766 CERROR("mds_getattr_pack_msg: %d\n", rc);
1770 req->rq_status = mds_getattr_internal(obd, de, req, offset, body,
1776 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1777 mds_exit_ucred(&uc);
1781 static int mds_access_check(struct ptlrpc_request *req, int offset)
1783 struct obd_device *obd = req->rq_export->exp_obd;
1784 struct lvfs_run_ctxt saved;
1786 struct mds_req_sec_desc *rsd;
1787 struct mds_body *body;
1788 struct lvfs_ucred uc;
1789 int rep_size[2] = {sizeof(*body),
1790 sizeof(struct mds_remote_perm)};
1791 int rc = 0, rep_offset;
1794 if (!req->rq_export->exp_mds_data.med_remote) {
1795 CERROR("from local client "LPU64"\n", req->rq_peer.peer_id.nid);
1799 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1801 CERROR("Can't unpack security desc\n");
1805 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1806 lustre_swab_mds_body);
1808 CERROR ("Can't unpack body\n");
1812 MD_COUNTER_INCREMENT(obd, access_check);
1814 rc = mds_init_ucred(&uc, req, rsd);
1816 CERROR("init ucred error: %d\n", rc);
1819 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1821 de = mds_id2dentry(obd, &body->id1, NULL);
1823 CERROR("grab ino "LPU64": err %ld\n",
1824 body->id1.li_stc.u.e3s.l3s_ino, PTR_ERR(de));
1825 GOTO(out_pop, rc = PTR_ERR(de));
1828 rc = lustre_pack_reply(req, 2, rep_size, NULL);
1830 CERROR("pack reply error: %d\n", rc);
1831 GOTO(out_dput, rc = -EINVAL);
1834 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1838 rc = mds_pack_remote_perm(req, &rep_offset, body, de->d_inode);
1845 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1846 mds_exit_ucred(&uc);
1850 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1851 unsigned long max_age)
1856 spin_lock(&obd->obd_osfs_lock);
1857 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1859 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1860 spin_unlock(&obd->obd_osfs_lock);
1865 static int mds_statfs(struct ptlrpc_request *req)
1867 struct obd_device *obd = req->rq_export->exp_obd;
1868 int rc, size = sizeof(struct obd_statfs);
1871 /* This will trigger a watchdog timeout */
1872 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1873 (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1875 rc = lustre_pack_reply(req, 1, &size, NULL);
1876 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1877 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1881 OBD_COUNTER_INCREMENT(obd, statfs);
1883 /* We call this so that we can cache a bit - 1 jiffie worth */
1884 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1887 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1893 req->rq_status = rc;
1897 static int mds_sync(struct ptlrpc_request *req, int offset)
1899 struct obd_device *obd = req->rq_export->exp_obd;
1900 struct mds_obd *mds = &obd->u.mds;
1901 struct mds_body *body;
1902 int rc, size = sizeof(*body);
1905 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1906 lustre_swab_mds_body);
1908 GOTO(out, rc = -EPROTO);
1910 rc = lustre_pack_reply(req, 1, &size, NULL);
1911 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1912 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1916 if (id_ino(&body->id1) == 0) {
1917 /* an id of zero is taken to mean "sync whole filesystem" */
1918 rc = fsfilt_sync(obd, mds->mds_sb);
1922 /* just any file to grab fsync method - "file" arg unused */
1923 struct file *file = mds->mds_rcvd_filp;
1924 struct mds_body *rep_body;
1927 de = mds_id2dentry(obd, &body->id1, NULL);
1929 GOTO(out, rc = PTR_ERR(de));
1931 rc = file->f_op->fsync(NULL, de, 1);
1935 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1936 mds_pack_inode2body(obd, rep_body, de->d_inode,
1937 (body->valid & OBD_MD_FID) ? 1 : 0);
1943 req->rq_status = rc;
1947 /* mds_readpage does not take a DLM lock on the inode, because the client must
1948 * already have a PR lock.
1950 * If we were to take another one here, a deadlock will result, if another
1951 * thread is already waiting for a PW lock. */
1952 static int mds_readpage(struct ptlrpc_request *req, int offset)
1954 struct obd_device *obd = req->rq_export->exp_obd;
1955 struct vfsmount *mnt;
1958 struct mds_req_sec_desc *rsd;
1959 struct mds_body *body, *repbody;
1960 struct lvfs_run_ctxt saved;
1961 int rc, size = sizeof(*repbody);
1962 struct lvfs_ucred uc = {NULL, NULL,};
1965 rc = lustre_pack_reply(req, 1, &size, NULL);
1966 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1967 CERROR("mds: out of memory\n");
1968 GOTO(out, rc = -ENOMEM);
1971 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1973 CERROR("Can't unpack security desc\n");
1974 GOTO (out, rc = -EFAULT);
1977 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1978 lustre_swab_mds_body);
1980 CERROR("Can't unpack body\n");
1981 GOTO (out, rc = -EFAULT);
1984 rc = mds_init_ucred(&uc, req, rsd);
1986 mds_audit_auth(req, &uc, AUDIT_READDIR, &body->id1,
1991 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1992 de = mds_id2dentry(obd, &body->id1, &mnt);
1994 GOTO(out_pop, rc = PTR_ERR(de));
1996 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1998 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1999 /* note: in case of an error, dentry_open puts dentry */
2001 GOTO(out_pop, rc = PTR_ERR(file));
2003 /* body->size is actually the offset -eeb */
2004 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
2005 CERROR("offset "LPU64" not on a block boundary of %lu\n",
2006 body->size, de->d_inode->i_blksize);
2007 GOTO(out_file, rc = -EFAULT);
2010 /* body->nlink is actually the #bytes to read -eeb */
2011 if (body->nlink & (de->d_inode->i_blksize - 1)) {
2012 CERROR("size %u is not multiple of blocksize %lu\n",
2013 body->nlink, de->d_inode->i_blksize);
2014 GOTO(out_file, rc = -EFAULT);
2017 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
2018 repbody->size = file->f_dentry->d_inode->i_size;
2019 repbody->valid = OBD_MD_FLSIZE;
2021 /* to make this asynchronous make sure that the handling function
2022 doesn't send a reply when this function completes. Instead a
2023 callback function would send the reply */
2024 /* body->size is actually the offset -eeb */
2025 rc = mds_sendpage(req, file, body->size, body->nlink);
2029 filp_close(file, 0);
2031 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2033 mds_exit_ucred(&uc);
2034 req->rq_status = rc;
2038 int mds_read_md(struct obd_device *obd, struct lustre_id *id,
2039 char **data, int *datalen)
2041 struct dentry *dentry;
2042 struct mds_obd *mds = &obd->u.mds;
2043 int rc = 0, mea = 0;
2050 dentry = mds_id2dentry(obd, id, NULL);
2052 GOTO(out, rc = PTR_ERR(dentry));
2054 if (!dentry->d_inode) {
2055 CERROR("Can't find object "DLID4".\n",
2057 GOTO(out_dentry, rc = -EINVAL);
2059 if (S_ISDIR(dentry->d_inode->i_mode)) {
2060 *datalen = obd_packmd(mds->mds_md_exp, NULL, NULL);
2063 *datalen = obd_packmd(mds->mds_dt_exp, NULL, NULL);
2066 OBD_ALLOC(ea, *datalen);
2069 GOTO(out_dentry, rc = PTR_ERR(dentry));
2072 down(&dentry->d_inode->i_sem);
2073 rc = fsfilt_get_md(obd, dentry->d_inode, *data, *datalen,
2074 (mea ? EA_MEA : EA_LOV));
2075 up(&dentry->d_inode->i_sem);
2078 CERROR("Error %d reading eadata for ino %lu\n",
2079 rc, dentry->d_inode->i_ino);
2085 EXPORT_SYMBOL(mds_read_md);
2087 int mds_reint(struct ptlrpc_request *req, int offset,
2088 struct lustre_handle *lockh)
2090 struct mds_update_record *rec;
2091 struct mds_req_sec_desc *rsd;
2095 OBD_ALLOC(rec, sizeof(*rec));
2099 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
2101 CERROR("Can't unpack security desc\n");
2102 GOTO(out, rc = -EFAULT);
2105 rc = mds_update_unpack(req, offset, rec);
2106 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
2107 CERROR("invalid record\n");
2108 GOTO(out, req->rq_status = -EINVAL);
2111 rc = mds_init_ucred(&rec->ur_uc, req, rsd);
2113 audit_op code = AUDIT_NONE;
2114 char * au_name = NULL;
2116 switch (rec->ur_opcode) {
2118 code = AUDIT_SETATTR;
2121 code = AUDIT_CREATE;
2122 au_name = rec->ur_name;
2123 au_len = rec->ur_namelen;
2129 code = AUDIT_UNLINK;
2132 code = AUDIT_RENAME;
2135 au_name = rec->ur_name;
2136 au_len = rec->ur_namelen;
2140 CERROR("Wrong opcode in reint\n");
2144 mds_audit_auth(req, &rec->ur_uc, code, rec->ur_id1,
2149 /* rc will be used to interrupt a for loop over multiple records */
2150 rc = mds_reint_rec(rec, offset, req, lockh);
2152 /* audit stuff for OPEN */
2153 if (offset == 3 && rec->ur_opcode == REINT_OPEN)
2154 mds_audit_open(req, rec, rc);
2157 mds_exit_ucred(&rec->ur_uc);
2158 OBD_FREE(rec, sizeof(*rec));
2162 static int mds_filter_recovery_request(struct ptlrpc_request *req,
2163 struct obd_device *obd, int *process)
2165 switch (req->rq_reqmsg->opc) {
2166 case MDS_CONNECT: /* This will never get here, but for completeness. */
2167 case OST_CONNECT: /* This will never get here, but for completeness. */
2168 case MDS_DISCONNECT:
2169 case OST_DISCONNECT:
2174 case MDS_SYNC: /* used in unmounting */
2179 *process = target_queue_recovery_request(req, obd);
2183 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
2185 /* XXX what should we set rq_status to here? */
2186 req->rq_status = -EAGAIN;
2187 RETURN(ptlrpc_error(req));
2191 static char *reint_names[] = {
2192 [REINT_SETATTR] "setattr",
2193 [REINT_CREATE] "create",
2194 [REINT_LINK] "link",
2195 [REINT_UNLINK] "unlink",
2196 [REINT_RENAME] "rename",
2197 [REINT_OPEN] "open",
2200 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER | \
2201 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
2202 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
2205 static void reconstruct_create(struct ptlrpc_request *req)
2207 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2208 struct mds_client_data *mcd = med->med_mcd;
2209 struct dentry *dentry;
2210 struct ost_body *body;
2211 struct lustre_id id;
2215 /* copy rc, transno and disp; steal locks */
2216 mds_req_from_mcd(req, mcd);
2217 if (req->rq_status) {
2225 id_ino(&id) = mcd->mcd_last_data;
2226 LASSERT(id_ino(&id) != 0);
2228 dentry = mds_id2dentry(req2obd(req), &id, NULL);
2229 if (IS_ERR(dentry)) {
2230 CERROR("can't find inode "LPU64"\n", id_ino(&id));
2231 req->rq_status = PTR_ERR(dentry);
2236 CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
2237 req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
2238 dentry->d_inode->i_generation);
2240 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2241 obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
2242 body->oa.o_id = dentry->d_inode->i_ino;
2243 body->oa.o_generation = dentry->d_inode->i_generation;
2244 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2246 down(&dentry->d_inode->i_sem);
2247 rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
2248 up(&dentry->d_inode->i_sem);
2250 CERROR("Can't read inode self id, inode %lu, "
2251 "rc %d\n", dentry->d_inode->i_ino, rc);
2255 body->oa.o_fid = id_fid(&id);
2256 body->oa.o_mds = id_group(&id);
2262 static int mds_inode_init_acl(struct obd_device *obd, void *handle,
2263 struct dentry *de, void *xattr, int xattr_size)
2265 struct inode *inode = de->d_inode;
2266 struct posix_acl *acl;
2273 LASSERT(xattr_size > 0);
2275 if (!inode->i_op->getxattr || !inode->i_op->setxattr) {
2276 CERROR("backend fs dosen't support xattr\n");
2280 /* set default acl */
2281 if (S_ISDIR(inode->i_mode)) {
2282 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_DEFAULT,
2283 xattr, xattr_size, 0);
2285 CERROR("set default acl err: %d\n", rc);
2290 /* set access acl */
2291 acl = posix_acl_from_xattr(xattr, xattr_size);
2292 if (acl == NULL || IS_ERR(acl)) {
2293 CERROR("insane attr data\n");
2294 return PTR_ERR(acl);
2297 if (posix_acl_valid(acl)) {
2298 CERROR("default acl not valid: %d\n", rc);
2303 mode = inode->i_mode;
2304 rc = posix_acl_create_masq(acl, &mode);
2306 CERROR("create masq err %d\n", rc);
2310 if (inode->i_mode != mode) {
2311 struct iattr iattr = { .ia_valid = ATTR_MODE,
2315 rc2 = fsfilt_setattr(obd, de, handle, &iattr, 0);
2317 CERROR("setattr mode err: %d\n", rc2);
2324 /* we didn't change acl except mode bits of some
2325 * entries, so should be fit into original size.
2327 rc = posix_acl_to_xattr(acl, xattr, xattr_size);
2330 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_ACCESS,
2331 xattr, xattr_size, 0);
2333 CERROR("set access acl err: %d\n", rc);
2336 posix_acl_release(acl);
2340 static int mdt_obj_create(struct ptlrpc_request *req)
2342 struct obd_device *obd = req->rq_export->exp_obd;
2343 struct mds_obd *mds = &obd->u.mds;
2344 struct ost_body *body, *repbody;
2347 char idname[LL_ID_NAMELEN];
2348 int size = sizeof(*repbody);
2349 struct inode *parent_inode;
2350 struct lvfs_run_ctxt saved;
2351 int rc, cleanup_phase = 0;
2352 struct dentry *new = NULL;
2353 struct dentry_params dp;
2354 int mealen, flags = 0;
2355 struct lvfs_ucred uc;
2356 struct lustre_id id;
2358 void *handle = NULL;
2359 unsigned long cr_inum = 0;
2363 DEBUG_REQ(D_HA, req, "create remote object");
2364 parent_inode = mds->mds_unnamed_dir->d_inode;
2366 body = lustre_swab_reqbuf(req, 0, sizeof(*body),
2367 lustre_swab_ost_body);
2371 /* acl data is packed transparently, no swab here */
2372 LASSERT(req->rq_reqmsg->bufcount >= 2);
2373 acl_size = req->rq_reqmsg->buflens[1];
2375 acl = lustre_msg_buf(req->rq_reqmsg, 1, acl_size);
2377 CERROR("No default acl buf?\n");
2382 rc = lustre_pack_reply(req, 1, &size, NULL);
2386 MDS_CHECK_RESENT(req, reconstruct_create(req));
2389 uc.luc_ginfo = NULL;
2390 uc.luc_uid = body->oa.o_uid;
2391 uc.luc_gid = body->oa.o_gid;
2392 uc.luc_fsuid = body->oa.o_uid;
2393 uc.luc_fsgid = body->oa.o_gid;
2395 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2396 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2398 /* in REPLAY case inum should be given (client or other MDS fills it) */
2399 if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2400 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2402 * this is re-create request from MDS holding directory name.
2403 * we have to lookup given ino/gen first. if it exists (good
2404 * case) then there is nothing to do. if it does not then we
2405 * have to recreate it.
2407 id_ino(&id) = body->oa.o_id;
2408 id_gen(&id) = body->oa.o_generation;
2410 new = mds_id2dentry(obd, &id, NULL);
2411 if (!IS_ERR(new) && new->d_inode) {
2412 struct lustre_id sid;
2414 CDEBUG(D_OTHER, "mkdir repairing %lu/%lu\n",
2415 (unsigned long)id_ino(&id),
2416 (unsigned long)id_gen(&id));
2418 obdo_from_inode(&repbody->oa, new->d_inode,
2419 FILTER_VALID_FLAGS);
2421 repbody->oa.o_id = new->d_inode->i_ino;
2422 repbody->oa.o_generation = new->d_inode->i_generation;
2423 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2426 down(&new->d_inode->i_sem);
2427 rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2428 up(&new->d_inode->i_sem);
2430 CERROR("Can't read inode self id "
2431 "inode %lu, rc %d.\n",
2432 new->d_inode->i_ino, rc);
2436 repbody->oa.o_fid = id_fid(&sid);
2437 repbody->oa.o_mds = id_group(&sid);
2438 LASSERT(id_fid(&sid) != 0);
2441 * here we could use fid passed in body->oa.o_fid and
2442 * thus avoid mds_read_inode_sid().
2444 cr_inum = new->d_inode->i_ino;
2445 GOTO(cleanup, rc = 0);
2449 down(&parent_inode->i_sem);
2450 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2451 if (IS_ERR(handle)) {
2452 up(&parent_inode->i_sem);
2453 CERROR("fsfilt_start() failed, rc = %d\n",
2454 (int)PTR_ERR(handle));
2455 GOTO(cleanup, rc = PTR_ERR(handle));
2457 cleanup_phase = 1; /* transaction */
2460 rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2461 new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2463 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2464 obd->obd_name, idname, (int) PTR_ERR(new));
2465 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2466 up(&parent_inode->i_sem);
2467 RETURN(PTR_ERR(new));
2468 } else if (new->d_inode) {
2469 CERROR("%s: name exists. repeat\n", obd->obd_name);
2472 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2473 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2474 fid = body->oa.o_fid;
2476 fid = mds_alloc_fid(obd);
2478 new->d_fsdata = (void *)&dp;
2483 dp.p_group = mds->mds_num;
2485 if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2486 (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2487 LASSERT(body->oa.o_id != 0);
2488 dp.p_inum = body->oa.o_id;
2489 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2490 (unsigned long)body->oa.o_id,
2491 (unsigned long)body->oa.o_generation);
2494 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2497 rc = mds_inode_init_acl(obd, handle, new,
2500 up(&parent_inode->i_sem);
2504 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2505 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2506 new->d_inode->i_generation = body->oa.o_generation;
2507 mark_inode_dirty(new->d_inode);
2510 * avoiding asserts in cache flush case, as
2511 * @body->oa.o_id should be zero.
2513 if (body->oa.o_id) {
2514 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2515 "BUG 3550: failed to recreate obj "
2516 LPU64" -> %lu\n", body->oa.o_id,
2517 new->d_inode->i_ino);
2519 LASSERTF(body->oa.o_generation ==
2520 new->d_inode->i_generation,
2521 "BUG 3550: failed to recreate obj/gen "
2522 LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2523 body->oa.o_generation,
2524 new->d_inode->i_ino,
2525 new->d_inode->i_generation);
2529 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2530 repbody->oa.o_id = new->d_inode->i_ino;
2531 repbody->oa.o_generation = new->d_inode->i_generation;
2532 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2534 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2535 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2536 LASSERT(body->oa.o_id != 0);
2537 LASSERT(body->oa.o_fid != 0);
2540 mds_inode2id(obd, &id, new->d_inode, fid);
2541 mds_update_inode_ids(obd, new->d_inode, handle, &id,
2542 body->oa.o_valid & OBD_MD_FLID ?
2543 NULL : obdo_id(&body->oa));
2545 /* initializing o_fid after it is allocated. */
2546 repbody->oa.o_fid = id_fid(&id);
2547 repbody->oa.o_mds = id_group(&id);
2549 rc = fsfilt_del_dir_entry(obd, new);
2550 up(&parent_inode->i_sem);
2552 CERROR("can't remove name for object: %d\n", rc);
2556 cleanup_phase = 2; /* created directory object */
2558 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2559 (unsigned long)new->d_inode->i_ino,
2560 (unsigned long)new->d_inode->i_generation,
2561 (unsigned)new->d_inode->i_mode);
2562 cr_inum = new->d_inode->i_ino;
2564 up(&parent_inode->i_sem);
2565 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2569 if (body->oa.o_valid & OBD_MD_FLID) {
2570 /* this is new object for splitted dir. We have to prevent
2571 * recursive splitting on it -bzzz */
2572 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2574 OBD_ALLOC(mea, mealen);
2576 GOTO(cleanup, rc = -ENOMEM);
2578 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2579 mea->mea_master = body->oa.o_mds; /* master mds num */
2582 obdo2id(&mea->mea_ids[body->oa.o_mds], &body->oa);
2584 down(&new->d_inode->i_sem);
2585 rc = fsfilt_set_md(obd, new->d_inode, handle,
2586 mea, mealen, EA_MEA);
2587 up(&new->d_inode->i_sem);
2589 CERROR("fsfilt_set_md() failed, "
2592 OBD_FREE(mea, mealen);
2594 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2595 obd->obd_name, new->d_inode->i_ino,
2596 new->d_inode->i_generation, flags);
2597 } else if (body->oa.o_easize) {
2598 /* we pass LCK_EX to split routine to signal that we have
2599 * exclusive access to the directory. simple because nobody
2600 * knows it already exists -bzzz */
2601 rc = mds_try_to_split_dir(obd, new, NULL,
2602 body->oa.o_easize, LCK_EX);
2604 CERROR("Can't split directory %lu, error = %d.\n",
2605 new->d_inode->i_ino, rc);
2613 if (rc == 0 && (body->oa.o_flags & OBD_FL_REINT)) {
2614 rc = mds_fidmap_add(obd, &id);
2616 CERROR("can't create fid->ino mapping, "
2623 switch (cleanup_phase) {
2624 case 2: /* object has been created, but we'll may want to replay it later */
2626 ptlrpc_require_repack(req);
2627 case 1: /* transaction */
2628 rc = mds_finish_transno(mds, parent_inode, handle,
2633 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2637 static int mdt_get_info(struct ptlrpc_request *req)
2639 struct obd_export *exp = req->rq_export;
2644 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2646 DEBUG_REQ(D_HA, req, "no get_info key");
2649 keylen = req->rq_reqmsg->buflens[0];
2651 if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2652 (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2653 (keylen < strlen("lovdesc") || strcmp(key, "lovdesc") != 0) &&
2654 (keylen < strlen("getext") || strcmp(key, "getext") != 0) &&
2655 (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2658 if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2659 struct lustre_id *reply;
2660 int size = sizeof(*reply);
2662 rc = lustre_pack_reply(req, 1, &size, NULL);
2666 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2667 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2668 } else if (keylen >= strlen("lovdesc") && !strcmp(key, "lovdesc")) {
2669 struct lov_desc *reply;
2670 int size = sizeof(*reply);
2672 rc = lustre_pack_reply(req, 1, &size, NULL);
2676 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2677 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2678 } else if (keylen >= strlen("getext") && !strcmp(key, "getext")) {
2679 struct fid_extent *reply;
2680 int size = sizeof(*reply);
2682 rc = lustre_pack_reply(req, 1, &size, NULL);
2686 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2687 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2690 int size = sizeof(*reply);
2692 rc = lustre_pack_reply(req, 1, &size, NULL);
2696 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2697 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2700 req->rq_repmsg->status = 0;
2704 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2705 void *key, __u32 vallen, void *val)
2707 struct obd_device *obd;
2708 struct mds_obd *mds;
2712 obd = class_exp2obd(exp);
2714 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2715 exp->exp_handle.h_cookie);
2720 if (keylen == 5 && memcmp(key, "audit", 5) == 0) {
2721 rc = mds_set_audit(obd, val);
2723 } else if (keylen >= strlen("ids") && memcmp(key, "ids", keylen) == 0) {
2724 struct lustre_id *ids = (struct lustre_id *)val;
2726 struct inode *inode;
2730 de = mds_id2dentry(obd, ids, NULL);
2733 CERROR("lookup by an id error rc=%d\n ", rc);
2736 inode = de->d_inode;
2738 GOTO(out_put, rc = -ENOENT);
2740 down(&inode->i_sem);
2741 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
2742 if (IS_ERR(handle)) {
2744 GOTO(out_put, rc = PTR_ERR(handle));
2747 rc = mds_update_inode_ids(obd, inode, handle, NULL, ids + 1);
2749 err = fsfilt_commit(obd, mds->mds_sb, inode, handle,
2752 CERROR("error committing transaction: %d\n", err);
2761 if (keylen >= strlen("crypto_type") &&
2762 memcmp(key, "crypto_type", keylen) == 0) {
2763 rc = mds_set_crypto_type(obd, val, vallen);
2767 if (keylen >= strlen("setext") && !memcmp(key, "setext", keylen)) {
2768 struct fid_extent *ext = val;
2770 CDEBUG(D_IOCTL, "set last fid to extent ["LPD64"-"LPD64"]\n",
2771 ext->fe_start, ext->fe_width);
2773 /* set lastfid into fid extent start. All next object creates
2774 * will use that fid. */
2775 mds_set_last_fid(obd, ext->fe_start);
2777 /* setting the same extent to OSC to avoid ids intersecting in
2778 * object ids, as all cache MDSs have the same group 0. */
2779 rc = obd_set_info(mds->mds_dt_exp, strlen("setext"),
2780 "setext", sizeof(*ext), ext);
2782 CERROR("can't set extent ["LPD64"-"LPD64"] to %s, "
2783 "err %d\n", ext->fe_start, ext->fe_width,
2784 mds->mds_dt_exp->exp_obd->obd_name, rc);
2788 CDEBUG(D_IOCTL, "invalid key\n");
2792 static int mdt_set_info(struct ptlrpc_request *req)
2795 struct obd_export *exp = req->rq_export;
2796 int keylen, rc = 0, vallen = 0;
2799 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2801 DEBUG_REQ(D_HA, req, "no set_info key");
2804 keylen = req->rq_reqmsg->buflens[0];
2806 if ((keylen == strlen("crypto_type") &&
2807 memcmp(key, "crypto_type", keylen) == 0)) {
2808 rc = lustre_pack_reply(req, 0, NULL, NULL);
2812 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2813 vallen = req->rq_reqmsg->buflens[1];
2815 rc = obd_set_info(exp, keylen, key, vallen, val);
2816 req->rq_repmsg->status = 0;
2818 } else if (keylen == 5 && memcmp(key, "audit", 5) == 0) {
2819 struct audit_attr_msg msg, *p;
2822 rc = lustre_pack_reply(req, 0, NULL, NULL);
2826 p = lustre_swab_reqbuf(req, 1, sizeof(msg),
2827 lustre_swab_audit_attr);
2830 CDEBUG(D_INFO, "Get new audit setting 0x%x\n", (__u32)msg.attr);
2831 rc = obd_set_info(exp, keylen, key, sizeof(msg), &msg);
2833 req->rq_repmsg->status = rc;
2835 } else if (keylen == strlen("ids") &&
2836 memcmp(key, "ids", keylen) == 0) {
2837 struct lustre_id *id, ids[2];
2839 rc = lustre_pack_reply(req, 0, NULL, NULL);
2842 id = lustre_swab_reqbuf(req, 1, sizeof(struct lustre_id),
2843 lustre_swab_lustre_id);
2845 id = lustre_swab_reqbuf(req, 2, sizeof(struct lustre_id),
2846 lustre_swab_lustre_id);
2849 rc = obd_set_info(exp, keylen, key, vallen, ids);
2850 req->rq_repmsg->status = rc;
2852 } else if (keylen == strlen("setext") &&
2853 memcmp(key, "setext", keylen) == 0) {
2854 rc = lustre_pack_reply(req, 0, NULL, NULL);
2858 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2859 vallen = req->rq_reqmsg->buflens[1];
2861 rc = obd_set_info(exp, keylen, key, vallen, val);
2862 req->rq_repmsg->status = 0;
2866 CDEBUG(D_IOCTL, "invalid key\n");
2870 static void mds_revoke_export_locks(struct obd_export *exp)
2872 struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2873 struct list_head rpc_list;
2874 struct ldlm_lock *lock, *next;
2875 struct ldlm_lock_desc desc;
2877 /* don't do this for local client */
2878 if (!exp->u.eu_mds_data.med_remote)
2881 /* don't revoke locks during recovery */
2882 if (exp->exp_obd->obd_recovering)
2886 INIT_LIST_HEAD(&rpc_list);
2888 spin_lock(&exp->exp_ldlm_data.led_lock);
2889 list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2891 lock_res_and_lock(lock);
2892 if (lock->l_req_mode != lock->l_granted_mode) {
2893 unlock_res_and_lock(lock);
2897 LASSERT(lock->l_resource);
2898 if (lock->l_resource->lr_type != LDLM_IBITS &&
2899 lock->l_resource->lr_type != LDLM_PLAIN) {
2900 unlock_res_and_lock(lock);
2904 if (lock->l_flags & LDLM_FL_AST_SENT) {
2905 unlock_res_and_lock(lock);
2909 LASSERT(lock->l_blocking_ast);
2910 LASSERT(!lock->l_blocking_lock);
2912 lock->l_flags |= LDLM_FL_AST_SENT;
2913 unlock_res_and_lock(lock);
2915 list_move(&lock->l_export_chain, &rpc_list);
2917 spin_unlock(&exp->exp_ldlm_data.led_lock);
2919 while (!list_empty(&rpc_list)) {
2920 lock = list_entry(rpc_list.next, struct ldlm_lock,
2922 list_del_init(&lock->l_export_chain);
2924 /* the desc just pretend to exclusive */
2925 ldlm_lock2desc(lock, &desc);
2926 desc.l_req_mode = LCK_EX;
2927 desc.l_granted_mode = 0;
2929 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2934 static int mds_msg_check_version(struct lustre_msg *msg)
2940 case MDS_DISCONNECT:
2942 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2944 CERROR("bad opc %u version %08x, expecting %08x\n",
2945 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2950 case MDS_GETATTR_LOCK:
2951 case MDS_ACCESS_CHECK:
2955 case MDS_DONE_WRITING:
2958 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2960 CERROR("bad opc %u version %08x, expecting %08x\n",
2961 msg->opc, msg->version, LUSTRE_MDS_VERSION);
2965 case LDLM_BL_CALLBACK:
2966 case LDLM_CP_CALLBACK:
2967 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2969 CERROR("bad opc %u version %08x, expecting %08x\n",
2970 msg->opc, msg->version, LUSTRE_DLM_VERSION);
2972 case OBD_LOG_CANCEL:
2973 case LLOG_ORIGIN_HANDLE_OPEN:
2974 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2975 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2976 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2977 case LLOG_ORIGIN_HANDLE_CLOSE:
2979 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2981 CERROR("bad opc %u version %08x, expecting %08x\n",
2982 msg->opc, msg->version, LUSTRE_LOG_VERSION);
2988 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2990 CERROR("bad opc %u version %08x, expecting %08x\n",
2991 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2994 case SEC_INIT_CONTINUE:
2999 CERROR("MDS unknown opcode %d\n", msg->opc);
3007 int mds_handle(struct ptlrpc_request *req)
3009 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
3010 struct obd_device *obd = NULL;
3011 struct mds_obd *mds = NULL; /* quell gcc overwarning */
3015 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
3017 rc = mds_msg_check_version(req->rq_reqmsg);
3019 CERROR("MDS drop mal-formed request\n");
3023 /* Security opc should NOT trigger any recovery events */
3024 if (req->rq_reqmsg->opc == SEC_INIT ||
3025 req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
3026 if (req->rq_export) {
3027 mds_req_add_idmapping(req,
3028 &req->rq_export->exp_mds_data);
3029 mds_revoke_export_locks(req->rq_export);
3032 } else if (req->rq_reqmsg->opc == SEC_FINI) {
3033 if (req->rq_export) {
3034 mds_req_del_idmapping(req,
3035 &req->rq_export->exp_mds_data);
3036 mds_revoke_export_locks(req->rq_export);
3041 LASSERT(current->journal_info == NULL);
3042 /* XXX identical to OST */
3043 if (req->rq_reqmsg->opc != MDS_CONNECT) {
3044 struct mds_export_data *med;
3047 if (req->rq_export == NULL) {
3048 CERROR("operation %d on unconnected MDS from %s\n",
3049 req->rq_reqmsg->opc,
3051 req->rq_status = -ENOTCONN;
3052 GOTO(out, rc = -ENOTCONN);
3055 med = &req->rq_export->exp_mds_data;
3056 obd = req->rq_export->exp_obd;
3059 /* sanity check: if the xid matches, the request must
3060 * be marked as a resent or replayed */
3061 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
3062 req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
3063 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
3064 (MSG_RESENT | MSG_REPLAY),
3065 "rq_xid "LPU64" matches last_xid, "
3066 "expected RESENT flag\n",
3069 /* else: note the opposite is not always true; a
3070 * RESENT req after a failover will usually not match
3071 * the last_xid, since it was likely never
3072 * committed. A REPLAYed request will almost never
3073 * match the last xid, however it could for a
3074 * committed, but still retained, open. */
3076 spin_lock_bh(&obd->obd_processing_task_lock);
3077 recovering = obd->obd_recovering;
3078 spin_unlock_bh(&obd->obd_processing_task_lock);
3080 rc = mds_filter_recovery_request(req, obd,
3082 if (rc || should_process == 0) {
3084 } else if (should_process < 0) {
3085 req->rq_status = should_process;
3086 rc = ptlrpc_error(req);
3092 switch (req->rq_reqmsg->opc) {
3094 DEBUG_REQ(D_INODE, req, "connect");
3095 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
3096 rc = target_handle_connect(req);
3098 struct mds_export_data *med;
3100 LASSERT(req->rq_export);
3101 med = &req->rq_export->u.eu_mds_data;
3102 mds_init_export_data(req, med);
3103 mds_req_add_idmapping(req, med);
3105 /* Now that we have an export, set mds. */
3106 obd = req->rq_export->exp_obd;
3107 mds = mds_req2mds(req);
3111 case MDS_DISCONNECT:
3112 DEBUG_REQ(D_INODE, req, "disconnect");
3113 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
3114 rc = target_handle_disconnect(req);
3115 req->rq_status = rc; /* superfluous? */
3119 DEBUG_REQ(D_INODE, req, "getstatus");
3120 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
3121 rc = mds_getstatus(req);
3125 DEBUG_REQ(D_INODE, req, "getattr");
3126 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
3127 rc = mds_getattr(req, MDS_REQ_REC_OFF);
3130 case MDS_ACCESS_CHECK:
3131 DEBUG_REQ(D_INODE, req, "access_check");
3132 OBD_FAIL_RETURN(OBD_FAIL_MDS_ACCESS_CHECK_NET, 0);
3133 rc = mds_access_check(req, MDS_REQ_REC_OFF);
3136 case MDS_GETATTR_LOCK: {
3137 struct lustre_handle lockh;
3138 DEBUG_REQ(D_INODE, req, "getattr_lock");
3139 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
3141 /* If this request gets a reconstructed reply, we won't be
3142 * acquiring any new locks in mds_getattr_lock, so we don't
3146 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
3147 MDS_INODELOCK_UPDATE);
3148 /* this non-intent call (from an ioctl) is special */
3149 req->rq_status = rc;
3150 if (rc == 0 && lockh.cookie)
3151 ldlm_lock_decref(&lockh, LCK_PR);
3155 DEBUG_REQ(D_INODE, req, "statfs");
3156 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
3157 rc = mds_statfs(req);
3161 DEBUG_REQ(D_INODE, req, "readpage");
3162 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
3163 rc = mds_readpage(req, MDS_REQ_REC_OFF);
3165 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
3166 if (req->rq_reply_state) {
3167 lustre_free_reply_state (req->rq_reply_state);
3168 req->rq_reply_state = NULL;
3175 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
3178 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
3179 mds->mds_max_cookiesize};
3182 /* NB only peek inside req now; mds_reint() will swab it */
3184 CERROR ("Can't inspect opcode\n");
3189 if (lustre_msg_swabbed (req->rq_reqmsg))
3192 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
3193 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
3194 reint_names[opc] == NULL) ? reint_names[opc] :
3197 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
3199 if (opc == REINT_UNLINK || opc == REINT_RENAME)
3201 else if (opc == REINT_OPEN)
3206 /* for SETATTR: I have different reply setting for
3207 * remote setfacl, so delay the reply buffer allocation.
3209 if (opc != REINT_SETATTR) {
3210 rc = lustre_pack_reply(req, bufcount, size, NULL);
3215 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
3216 fail = OBD_FAIL_MDS_REINT_NET_REP;
3221 DEBUG_REQ(D_INODE, req, "close");
3222 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
3223 rc = mds_close(req, MDS_REQ_REC_OFF);
3226 case MDS_DONE_WRITING:
3227 DEBUG_REQ(D_INODE, req, "done_writing");
3228 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
3229 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
3233 DEBUG_REQ(D_INODE, req, "pin");
3234 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
3235 rc = mds_pin(req, MDS_REQ_REC_OFF);
3239 DEBUG_REQ(D_INODE, req, "sync");
3240 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
3241 rc = mds_sync(req, MDS_REQ_REC_OFF);
3244 DEBUG_REQ(D_INODE, req, "parseid");
3245 rc = mds_parse_id(req);
3248 DEBUG_REQ(D_INODE, req, "ping");
3249 rc = target_handle_ping(req);
3252 case OBD_LOG_CANCEL:
3253 CDEBUG(D_INODE, "log cancel\n");
3254 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
3255 rc = -ENOTSUPP; /* la la la */
3259 DEBUG_REQ(D_INODE, req, "enqueue");
3260 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
3261 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
3262 ldlm_server_blocking_ast, NULL);
3263 fail = OBD_FAIL_LDLM_REPLY;
3266 DEBUG_REQ(D_INODE, req, "convert");
3267 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
3268 rc = ldlm_handle_convert(req);
3270 case LDLM_BL_CALLBACK:
3271 case LDLM_CP_CALLBACK:
3272 DEBUG_REQ(D_INODE, req, "callback");
3273 CERROR("callbacks should not happen on MDS\n");
3275 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
3277 case LLOG_ORIGIN_HANDLE_OPEN:
3278 DEBUG_REQ(D_INODE, req, "llog_init");
3279 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3280 rc = llog_origin_handle_open(req);
3282 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
3283 DEBUG_REQ(D_INODE, req, "llog next block");
3284 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3285 rc = llog_origin_handle_next_block(req);
3287 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
3288 DEBUG_REQ(D_INODE, req, "llog prev block");
3289 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3290 rc = llog_origin_handle_prev_block(req);
3292 case LLOG_ORIGIN_HANDLE_READ_HEADER:
3293 DEBUG_REQ(D_INODE, req, "llog read header");
3294 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3295 rc = llog_origin_handle_read_header(req);
3297 case LLOG_ORIGIN_HANDLE_CLOSE:
3298 DEBUG_REQ(D_INODE, req, "llog close");
3299 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3300 rc = llog_origin_handle_close(req);
3303 DEBUG_REQ(D_INODE, req, "ost_create");
3304 rc = mdt_obj_create(req);
3307 DEBUG_REQ(D_INODE, req, "get_info");
3308 rc = mdt_get_info(req);
3311 DEBUG_REQ(D_INODE, req, "set_info");
3312 rc = mdt_set_info(req);
3315 CDEBUG(D_INODE, "write\n");
3316 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
3317 rc = ost_brw_write(req, NULL);
3318 LASSERT(current->journal_info == NULL);
3319 /* mdt_brw sends its own replies */
3323 DEBUG_REQ(D_INODE, req, "llog catinfo");
3324 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3325 rc = llog_catinfo(req);
3328 req->rq_status = -ENOTSUPP;
3329 rc = ptlrpc_error(req);
3333 LASSERT(current->journal_info == NULL);
3337 /* If we're DISCONNECTing, the mds_export_data is already freed */
3338 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
3339 struct mds_export_data *med = &req->rq_export->exp_mds_data;
3340 struct obd_device *obd = list_entry(mds, struct obd_device,
3342 req->rq_repmsg->last_xid =
3343 le64_to_cpu(med->med_mcd->mcd_last_xid);
3345 if (!obd->obd_no_transno) {
3346 req->rq_repmsg->last_committed =
3347 obd->obd_last_committed;
3349 DEBUG_REQ(D_IOCTL, req,
3350 "not sending last_committed update");
3352 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
3354 mds->mds_last_transno, obd->obd_last_committed,
3360 target_send_reply(req, rc, fail);
3364 /* Update the server data on disk. This stores the new mount_count and also the
3365 * last_rcvd value to disk. If we don't have a clean shutdown, then the server
3366 * last_rcvd value may be less than that of the clients. This will alert us
3367 * that we may need to do client recovery.
3369 * Also assumes for mds_last_transno that we are not modifying it (no locking).
3371 int mds_update_server_data(struct obd_device *obd, int force_sync)
3373 struct mds_obd *mds = &obd->u.mds;
3374 struct mds_server_data *msd = mds->mds_server_data;
3375 struct file *filp = mds->mds_rcvd_filp;
3376 struct lvfs_run_ctxt saved;
3381 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3382 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
3384 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
3385 mds->mds_mount_count, mds->mds_last_transno);
3386 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
3388 CERROR("error writing MDS server data: rc = %d\n", rc);
3389 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3394 /* saves last allocated fid counter to file. */
3395 int mds_update_last_fid(struct obd_device *obd, void *handle,
3398 struct mds_obd *mds = &obd->u.mds;
3399 struct file *filp = mds->mds_fid_filp;
3400 struct lvfs_run_ctxt saved;
3406 spin_lock(&mds->mds_last_fid_lock);
3407 last_fid = mds->mds_last_fid;
3408 spin_unlock(&mds->mds_last_fid_lock);
3410 CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
3414 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
3415 handle, mds_commit_last_fid_cb,
3419 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3420 rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
3422 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3425 CERROR("error writing MDS last_fid #"LPU64
3426 ", err = %d\n", last_fid, rc);
3430 CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
3431 "%llu: err = %d\n", last_fid, off, rc);
3436 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
3438 struct mds_obd *mds = &obd->u.mds;
3440 spin_lock(&mds->mds_last_fid_lock);
3441 if (fid > mds->mds_last_fid)
3442 mds->mds_last_fid = fid;
3443 spin_unlock(&mds->mds_last_fid_lock);
3446 void mds_commit_last_transno_cb(struct obd_device *obd,
3447 __u64 transno, void *data,
3450 obd_transno_commit_cb(obd, transno, error);
3453 void mds_commit_last_fid_cb(struct obd_device *obd,
3454 __u64 fid, void *data,
3458 CERROR("%s: fid "LPD64" commit error: %d\n",
3459 obd->obd_name, fid, error);
3463 CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
3464 obd->obd_name, fid);
3467 __u64 mds_alloc_fid(struct obd_device *obd)
3469 struct mds_obd *mds = &obd->u.mds;
3472 spin_lock(&mds->mds_last_fid_lock);
3473 fid = ++mds->mds_last_fid;
3474 spin_unlock(&mds->mds_last_fid_lock);
3480 * reads inode self id from inode EA. Probably later this should be replaced by
3481 * caching inode self id to avoid raeding it every time it is needed.
3483 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
3484 struct lustre_id *id)
3489 LASSERT(id != NULL);
3490 LASSERT(obd != NULL);
3491 LASSERT(inode != NULL);
3493 rc = fsfilt_get_md(obd, inode, &id->li_fid,
3494 sizeof(id->li_fid), EA_SID);
3496 CERROR("fsfilt_get_md() failed, "
3505 id_ino(id) = inode->i_ino;
3506 id_gen(id) = inode->i_generation;
3507 id_type(id) = S_IFMT & inode->i_mode;
3512 int mds_read_inode_pid(struct obd_device *obd, struct inode *inode,
3513 struct lustre_id *id)
3518 LASSERT(inode && id);
3520 rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_PID);
3522 CERROR("get parent id from EA failed, rc=%d\n", rc);
3531 /* updates inode self id in EA. */
3532 int mds_update_inode_ids(struct obd_device *obd, struct inode *inode,
3533 void *handle, struct lustre_id *id,
3534 struct lustre_id *pid)
3540 LASSERT(id == NULL || id_fid(id) != 0);
3541 LASSERT(pid == NULL || id_fid(pid) != 0);
3542 LASSERT(obd != NULL);
3543 LASSERT(inode != NULL);
3546 mds_set_last_fid(obd, id_fid(id));
3547 rc = fsfilt_set_md(obd, inode, handle, &id->li_fid,
3548 sizeof(id->li_fid), EA_SID);
3549 LASSERTF(rc == 0, "failed to update fid: %d\n", rc);
3552 rc = fsfilt_set_md(obd, inode, handle, pid,
3553 sizeof(*pid), EA_PID);
3554 LASSERTF(rc == 0, "failed to update parent fid: %d\n", rc);
3560 /* mount the file system (secretly) */
3561 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3563 struct lustre_cfg* lcfg = buf;
3564 struct mds_obd *mds = &obd->u.mds;
3565 struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3566 char *options = NULL;
3567 struct vfsmount *mnt;
3570 struct crypto_tfm *tfm = NULL;
3574 if (lcfg->lcfg_bufcount < 3)
3575 RETURN(rc = -EINVAL);
3577 if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
3578 RETURN(rc = -EINVAL);
3580 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
3581 if (IS_ERR(obd->obd_fsops))
3582 RETURN(rc = PTR_ERR(obd->obd_fsops));
3584 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3586 page = get_zeroed_page(GFP_KERNEL);
3590 options = (char *)page;
3593 * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3594 * and the rest of options are passed by mount options. Probably this
3595 * should be moved to somewhere else like startup scripts or lconf. */
3596 sprintf(options, "iopen_nopriv");
3598 if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0 && lustre_cfg_buf(lcfg, 4))
3599 sprintf(options + strlen(options), ",%s",
3600 lustre_cfg_string(lcfg, 4));
3602 /* we have to know mdsnum before touching underlying fs -bzzz */
3603 atomic_set(&mds->mds_open_count, 0);
3604 sema_init(&mds->mds_md_sem, 1);
3605 mds->mds_md_connected = 0;
3606 mds->mds_md_name = NULL;
3608 if (LUSTRE_CFG_BUFLEN(lcfg, 5) > 0 && lustre_cfg_buf(lcfg, 5) &&
3609 strncmp(lustre_cfg_string(lcfg, 5), "dumb", LUSTRE_CFG_BUFLEN(lcfg, 5))) {
3612 generate_random_uuid(uuid);
3613 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3615 OBD_ALLOC(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3616 if (mds->mds_md_name == NULL)
3617 RETURN(rc = -ENOMEM);
3619 memcpy(mds->mds_md_name, lustre_cfg_buf(lcfg, 5),
3620 LUSTRE_CFG_BUFLEN(lcfg, 5));
3622 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3623 obd->obd_name, mds->mds_md_name);
3625 rc = mds_md_connect(obd, mds->mds_md_name);
3627 OBD_FREE(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3632 mds->mds_obd_type = MDS_MASTER_OBD;
3634 if (LUSTRE_CFG_BUFLEN(lcfg, 6) > 0 && lustre_cfg_buf(lcfg, 6) &&
3635 strncmp(lustre_cfg_string(lcfg, 6), "dumb",
3636 LUSTRE_CFG_BUFLEN(lcfg, 6))) {
3637 if (!memcmp(lustre_cfg_string(lcfg, 6), "master",
3638 strlen("master"))) {
3639 mds->mds_obd_type = MDS_MASTER_OBD;
3640 } else if (!memcmp(lustre_cfg_string(lcfg, 6), "cache",
3642 mds->mds_obd_type = MDS_CACHE_OBD;
3646 rc = lvfs_mount_fs(lustre_cfg_string(lcfg, 1),
3647 lustre_cfg_string(lcfg, 2),
3648 options, 0, &lvfs_ctxt);
3652 if (rc || !lvfs_ctxt) {
3653 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3657 mnt = lvfs_ctxt->loc_mnt;
3658 mds->mds_lvfs_ctxt = lvfs_ctxt;
3659 ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3661 CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
3663 mds->mds_fidext_thumb = 0;
3664 sema_init(&mds->mds_epoch_sem, 1);
3665 atomic_set(&mds->mds_real_clients, 0);
3666 spin_lock_init(&mds->mds_fidext_lock);
3667 spin_lock_init(&mds->mds_fidmap_lock);
3668 spin_lock_init(&mds->mds_transno_lock);
3669 spin_lock_init(&mds->mds_last_fid_lock);
3670 sema_init(&mds->mds_orphan_recovery_sem, 1);
3671 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3673 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3674 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3676 if (obd->obd_namespace == NULL) {
3677 mds_cleanup(obd, 0);
3678 GOTO(err_put, rc = -ENOMEM);
3680 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3682 tfm = crypto_alloc_tfm(CAPA_HMAC_ALG, 0);
3684 GOTO(err_ns, rc = -ENOSYS);
3686 mds->mds_capa_hmac = tfm;
3687 mds->mds_capa_timeout = CAPA_TIMEOUT;
3688 mds->mds_capa_key_timeout = CAPA_KEY_TIMEOUT;
3690 rc = mds_fs_setup(obd, mnt);
3692 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3697 rc = llog_start_commit_thread();
3702 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 && lustre_cfg_buf(lcfg, 3) &&
3703 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3704 LUSTRE_CFG_BUFLEN(lcfg, 3))) {
3707 generate_random_uuid(uuid);
3708 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3710 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
3711 if (mds->mds_profile == NULL)
3712 GOTO(err_fs, rc = -ENOMEM);
3714 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
3715 LUSTRE_CFG_BUFLEN(lcfg, 3));
3719 * setup root dir and files ID dir if lmv already connected, or there is
3722 if (mds->mds_md_exp || (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 &&
3723 lustre_cfg_buf(lcfg, 3) &&
3724 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3725 LUSTRE_CFG_BUFLEN(lcfg, 3)))) {
3726 rc = mds_fs_setup_rootid(obd);
3730 rc = mds_fs_setup_virtid(obd);
3736 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3737 "mds_ldlm_client", &obd->obd_ldlm_client);
3738 obd->obd_replayable = 1;
3740 mds->mds_crypto_type = NO_CRYPTO;
3742 rc = mds_postsetup(obd);
3749 /* No extra cleanup needed for llog_init_commit_thread() */
3750 mds_fs_cleanup(obd, 0);
3752 crypto_free_tfm(mds->mds_capa_hmac);
3754 ldlm_namespace_free(obd->obd_namespace, 0);
3755 obd->obd_namespace = NULL;
3758 lvfs_umount_fs(mds->mds_lvfs_ctxt);
3762 fsfilt_put_ops(obd->obd_fsops);
3766 static int mds_fs_post_setup(struct obd_device *obd)
3768 struct mds_obd *mds = &obd->u.mds;
3769 struct dentry *dentry;
3773 dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
3774 if (IS_ERR(dentry)) {
3775 CERROR("Can't find ROOT, err = %d\n",
3776 (int)PTR_ERR(dentry));
3777 RETURN(PTR_ERR(dentry));
3779 rc = fsfilt_post_setup(obd, dentry);
3780 //set id2name function handler
3781 fsfilt_set_info(obd, mds->mds_sb, NULL, 7, "id2name",
3782 sizeof(mds_audit_id2name), mds_audit_id2name);
3788 static int mds_postsetup(struct obd_device *obd)
3790 struct mds_obd *mds = &obd->u.mds;
3794 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3795 obd, 0, NULL, &llog_lvfs_ops);
3799 if (mds->mds_profile) {
3800 struct llog_ctxt *lgctxt;
3801 struct lvfs_run_ctxt saved;
3802 struct lustre_profile *lprof;
3803 struct config_llog_instance cfg;
3805 cfg.cfg_instance = NULL;
3806 cfg.cfg_uuid = mds->mds_dt_uuid;
3807 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3809 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3811 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3812 GOTO(err_llog, rc = -EINVAL);
3815 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3816 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3821 lprof = class_get_profile(mds->mds_profile);
3822 if (lprof == NULL) {
3823 CERROR("No profile found: %s\n", mds->mds_profile);
3824 GOTO(err_cleanup, rc = -ENOENT);
3826 rc = mds_dt_connect(obd, lprof->lp_lov);
3828 GOTO(err_cleanup, rc);
3830 rc = mds_md_postsetup(obd);
3832 GOTO(err_cleanup, rc);
3834 rc = mds_fs_post_setup(obd);
3836 CERROR("can not post setup fsfilt\n");
3841 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3842 LLOG_CONFIG_ORIG_CTXT));
3846 int mds_postrecov_common(struct obd_device *obd)
3848 struct mds_obd *mds = &obd->u.mds;
3849 struct llog_ctxt *ctxt;
3850 int rc, item = 0, valsize;
3854 LASSERT(!obd->obd_recovering);
3855 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3856 LASSERT(ctxt != NULL);
3858 /* clean PENDING dir */
3859 rc = mds_cleanup_orphans(obd);
3864 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3865 valsize = sizeof(group);
3866 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
3867 "mds_conn", valsize, &group);
3871 rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3874 CERROR("%s: failed at llog_origin_connect: %d\n",
3879 /* remove the orphaned precreated objects */
3880 rc = mds_dt_clear_orphans(mds, NULL /* all OSTs */);
3885 RETURN(rc < 0 ? rc : item);
3888 /* cleanup all llogging subsystems */
3889 rc = obd_llog_finish(obd, &obd->obd_llogs,
3890 mds->mds_dt_desc.ld_tgt_count);
3892 CERROR("%s: failed to cleanup llogging subsystems\n",
3897 int mds_postrecov(struct obd_device *obd)
3901 rc = mds_postrecov_common(obd);
3903 rc = mds_md_reconnect(obd);
3907 int mds_dt_clean(struct obd_device *obd)
3909 struct mds_obd *mds = &obd->u.mds;
3912 if (mds->mds_profile) {
3914 struct llog_ctxt *llctx;
3915 struct lvfs_run_ctxt saved;
3916 struct config_llog_instance cfg;
3917 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3919 OBD_ALLOC(cln_prof, len);
3920 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3922 cfg.cfg_instance = NULL;
3923 cfg.cfg_uuid = mds->mds_dt_uuid;
3925 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3926 llctx = llog_get_context(&obd->obd_llogs,
3927 LLOG_CONFIG_ORIG_CTXT);
3928 class_config_process_llog(llctx, cln_prof, &cfg);
3929 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3931 OBD_FREE(cln_prof, len);
3932 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3933 mds->mds_profile = NULL;
3938 int mds_md_clean(struct obd_device *obd)
3940 struct mds_obd *mds = &obd->u.mds;
3943 if (mds->mds_md_name) {
3944 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3945 mds->mds_md_name = NULL;
3950 static int mds_precleanup(struct obd_device *obd, int flags)
3956 mds_dt_disconnect(obd, flags);
3958 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3962 extern void lgss_svc_cache_purge_all(void);
3963 static int mds_cleanup(struct obd_device *obd, int flags)
3965 struct mds_obd *mds = &obd->u.mds;
3968 if (mds->mds_sb == NULL)
3971 mds_update_server_data(obd, 1);
3972 mds_update_last_fid(obd, NULL, 1);
3974 if (mds->mds_dt_objids != NULL) {
3975 int size = mds->mds_dt_desc.ld_tgt_count *
3977 OBD_FREE(mds->mds_dt_objids, size);
3979 mds_fs_cleanup(obd, flags);
3983 /* 2 seems normal on mds, (may_umount() also expects 2
3984 fwiw), but we only see 1 at this point in obdfilter. */
3985 lvfs_umount_fs(mds->mds_lvfs_ctxt);
3989 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3991 spin_lock_bh(&obd->obd_processing_task_lock);
3992 if (obd->obd_recovering) {
3993 target_cancel_recovery_timer(obd);
3994 obd->obd_recovering = 0;
3996 spin_unlock_bh(&obd->obd_processing_task_lock);
3999 fsfilt_put_ops(obd->obd_fsops);
4003 lgss_svc_cache_purge_all();
4006 spin_lock(&mds->mds_denylist_lock);
4007 while (!list_empty( &mds->mds_denylist ) ) {
4008 deny_sec_t *p_deny_sec = list_entry(mds->mds_denylist.next,
4010 list_del(&p_deny_sec->list);
4011 OBD_FREE(p_deny_sec, sizeof(*p_deny_sec));
4013 spin_unlock(&mds->mds_denylist_lock);
4015 mds_capa_keys_cleanup(obd);
4017 if (mds->mds_capa_hmac)
4018 crypto_free_tfm(mds->mds_capa_hmac);
4022 static int set_security(const char *value, char **sec)
4024 if (!strcmp(value, "null"))
4026 else if (!strcmp(value, "krb5i"))
4028 else if (!strcmp(value, "krb5p"))
4031 CERROR("Unrecognized security flavor %s\n", value);
4038 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
4040 struct lustre_cfg *lcfg = buf;
4041 struct mds_obd *mds = &obd->u.mds;
4045 switch(lcfg->lcfg_command) {
4046 case LCFG_SET_SECURITY: {
4047 if ((LUSTRE_CFG_BUFLEN(lcfg, 1) == 0) ||
4048 (LUSTRE_CFG_BUFLEN(lcfg, 2) == 0))
4049 GOTO(out, rc = -EINVAL);
4051 if (!strcmp(lustre_cfg_string(lcfg, 1), "mds_sec"))
4052 rc = set_security(lustre_cfg_string(lcfg, 2),
4054 else if (!strcmp(lustre_cfg_string(lcfg, 1), "oss_sec"))
4055 rc = set_security(lustre_cfg_string(lcfg, 2),
4057 else if (!strcmp(lustre_cfg_string(lcfg, 1), "deny_sec")){
4058 spin_lock(&mds->mds_denylist_lock);
4059 rc = add_deny_security(lustre_cfg_string(lcfg, 2),
4060 &mds->mds_denylist);
4061 spin_unlock(&mds->mds_denylist_lock);
4063 CERROR("Unrecognized key\n");
4069 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
4070 GOTO(out, rc = -EINVAL);
4076 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
4078 struct ldlm_lock *new_lock,
4079 struct ldlm_lock **old_lock,
4080 struct lustre_handle *lockh)
4082 struct obd_export *exp = req->rq_export;
4083 struct obd_device *obd = exp->exp_obd;
4084 struct ldlm_request *dlmreq =
4085 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
4086 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
4087 struct list_head *iter;
4089 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
4092 spin_lock(&obd->obd_namespace->ns_hash_lock);
4093 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
4094 struct ldlm_lock *lock;
4095 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
4096 if (lock == new_lock)
4098 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
4099 lockh->cookie = lock->l_handle.h_cookie;
4100 LDLM_DEBUG(lock, "restoring lock cookie");
4101 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
4104 *old_lock = LDLM_LOCK_GET(lock);
4105 spin_unlock(&obd->obd_namespace->ns_hash_lock);
4109 spin_unlock(&obd->obd_namespace->ns_hash_lock);
4111 /* If the xid matches, then we know this is a resent request,
4112 * and allow it. (It's probably an OPEN, for which we don't
4115 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
4119 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
4122 /* This remote handle isn't enqueued, so we never received or
4123 * processed this request. Clear MSG_RESENT, because it can
4124 * be handled like any normal request now. */
4126 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
4128 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
4132 int intent_disposition(struct ldlm_reply *rep, int flag)
4136 return (rep->lock_policy_res1 & flag);
4139 void intent_set_disposition(struct ldlm_reply *rep, int flag)
4143 rep->lock_policy_res1 |= flag;
4146 static int mds_intent_prepare_reply_buffers(struct ptlrpc_request *req,
4147 struct ldlm_intent *it)
4149 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
4150 int rc, reply_buffers;
4151 int repsize[8] = {sizeof(struct ldlm_reply),
4152 sizeof(struct mds_body),
4153 mds->mds_max_mdsize};
4157 if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
4158 /*Ugly here, Actually, we should prepare the reply buffer
4159 *after we know whether these stuff exist or not, which should
4160 * be fixed in future, Now each item is in the fix position,
4161 * the sequence is lsm, acl, crypto ea, capa.*/
4162 repsize[reply_buffers++] = sizeof(int);
4163 /* XXX: mds_remote_perm is stored here too, and for it
4164 * the 'size' is ignored */
4165 repsize[reply_buffers++] =
4166 xattr_acl_size(LL_ACL_MAX_ENTRIES);
4168 repsize[reply_buffers++] = sizeof(int);
4169 repsize[reply_buffers++] = sizeof(struct crypto_key);
4171 /* XXX: if new buffer is to be added, capability reply
4172 * buffer should always been reserved. */
4173 if (it->opc & IT_OPEN)
4174 repsize[reply_buffers++] = sizeof(struct lustre_capa);
4177 rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
4182 static int mds_intent_policy(struct ldlm_namespace *ns,
4183 struct ldlm_lock **lockp, void *req_cookie,
4184 ldlm_mode_t mode, int flags, void *data)
4186 struct ptlrpc_request *req = req_cookie;
4187 struct ldlm_lock *lock = *lockp;
4188 struct ldlm_intent *it;
4189 struct ldlm_reply *rep;
4190 struct lustre_handle lockh[2] = {{0}, {0}};
4191 struct ldlm_lock *new_lock = NULL;
4192 int getattr_part = MDS_INODELOCK_UPDATE;
4195 int offset = MDS_REQ_INTENT_REC_OFF;
4198 LASSERT(req != NULL);
4199 MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
4201 if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
4202 /* No intent was provided */
4203 int size = sizeof(struct ldlm_reply);
4204 rc = lustre_pack_reply(req, 1, &size, NULL);
4209 it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
4210 lustre_swab_ldlm_intent);
4212 CERROR("Intent missing\n");
4213 RETURN(req->rq_status = -EFAULT);
4216 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
4218 rc = mds_intent_prepare_reply_buffers(req, it);
4221 RETURN(req->rq_status = rc);
4223 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
4224 LASSERT(rep != NULL);
4226 intent_set_disposition(rep, DISP_IT_EXECD);
4228 /* execute policy */
4229 switch ((long)it->opc) {
4231 case IT_CREAT|IT_OPEN:
4232 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4234 /* XXX swab here to assert that an mds_open reint
4235 * packet is following */
4236 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4238 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
4240 if (rep->lock_policy_res2) {
4242 * mds_open() returns ENOLCK where it should return
4243 * zero, but it has no lock to return.
4245 if (rep->lock_policy_res2 == ENOLCK)
4246 rep->lock_policy_res2 = 0;
4248 RETURN(ELDLM_LOCK_ABORTED);
4252 * IT_OPEN may return lock on cross-node dentry that we want to
4253 * hold during attr retrival -bzzz
4255 if (lockh[0].cookie == 0)
4256 RETURN(ELDLM_LOCK_ABORTED);
4260 getattr_part = MDS_INODELOCK_LOOKUP;
4263 getattr_part |= MDS_INODELOCK_LOOKUP;
4265 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4266 lock, &new_lock, lockh);
4267 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
4269 /* FIXME: LDLM can set req->rq_status. MDS sets
4270 policy_res{1,2} with disposition and status.
4271 - replay: returns 0 & req->status is old status
4272 - otherwise: returns req->status */
4273 if (intent_disposition(rep, DISP_LOOKUP_NEG))
4274 rep->lock_policy_res2 = 0;
4275 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
4276 rep->lock_policy_res2)
4277 RETURN(ELDLM_LOCK_ABORTED);
4278 if (req->rq_status != 0) {
4280 rep->lock_policy_res2 = req->rq_status;
4281 RETURN(ELDLM_LOCK_ABORTED);
4285 rc = mds_lock_and_check_slave(offset, req, lockh);
4286 if ((rep->lock_policy_res2 = rc)) {
4288 rep->lock_policy_res2 = 0;
4289 RETURN(ELDLM_LOCK_ABORTED);
4293 CERROR("Unhandled intent "LPD64"\n", it->opc);
4297 /* By this point, whatever function we called above must have either
4298 * filled in 'lockh', been an intent replay, or returned an error. We
4299 * want to allow replayed RPCs to not get a lock, since we would just
4300 * drop it below anyways because lock replay is done separately by the
4301 * client afterwards. For regular RPCs we want to give the new lock to
4302 * the client instead of whatever lock it was about to get. */
4303 if (new_lock == NULL)
4304 new_lock = ldlm_handle2lock(&lockh[0]);
4305 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
4308 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
4309 it->opc, lockh[0].cookie);
4311 /* If we've already given this lock to a client once, then we should
4312 * have no readers or writers. Otherwise, we should have one reader
4313 * _or_ writer ref (which will be zeroed below) before returning the
4314 * lock to a client. */
4315 if (new_lock->l_export == req->rq_export) {
4316 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
4318 LASSERT(new_lock->l_export == NULL);
4319 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
4324 if (new_lock->l_export == req->rq_export) {
4325 /* Already gave this to the client, which means that we
4326 * reconstructed a reply. */
4327 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
4329 RETURN(ELDLM_LOCK_REPLACED);
4332 /* Fixup the lock to be given to the client */
4333 lock_res_and_lock(new_lock);
4334 new_lock->l_readers = 0;
4335 new_lock->l_writers = 0;
4337 new_lock->l_export = class_export_get(req->rq_export);
4339 spin_lock(&new_lock->l_export->exp_ldlm_data.led_lock);
4340 list_add(&new_lock->l_export_chain,
4341 &new_lock->l_export->exp_ldlm_data.led_held_locks);
4342 spin_unlock(&new_lock->l_export->exp_ldlm_data.led_lock);
4344 new_lock->l_blocking_ast = lock->l_blocking_ast;
4345 new_lock->l_completion_ast = lock->l_completion_ast;
4347 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
4348 sizeof(lock->l_remote_handle));
4350 new_lock->l_flags &= ~LDLM_FL_LOCAL;
4352 unlock_res_and_lock(new_lock);
4353 LDLM_LOCK_PUT(new_lock);
4355 RETURN(ELDLM_LOCK_REPLACED);
4358 int mds_attach(struct obd_device *dev, obd_count len, void *data)
4360 struct lprocfs_static_vars lvars;
4362 struct mds_obd *mds = &dev->u.mds;
4364 spin_lock_init(&mds->mds_denylist_lock);
4365 INIT_LIST_HEAD(&mds->mds_denylist);
4367 lprocfs_init_multi_vars(0, &lvars);
4369 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
4373 return lprocfs_alloc_md_stats(dev, 0);
4376 int mds_detach(struct obd_device *dev)
4378 lprocfs_free_md_stats(dev);
4379 return lprocfs_obd_detach(dev);
4382 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
4384 struct lprocfs_static_vars lvars;
4386 lprocfs_init_multi_vars(1, &lvars);
4387 return lprocfs_obd_attach(dev, lvars.obd_vars);
4390 int mdt_detach(struct obd_device *dev)
4392 return lprocfs_obd_detach(dev);
4395 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
4397 struct mds_obd *mds = &obd->u.mds;
4402 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4403 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
4404 MDS_SERVICE_WATCHDOG_TIMEOUT,
4405 mds_handle, "mds", obd->obd_proc_entry);
4407 if (!mds->mds_service) {
4408 CERROR("failed to start service\n");
4412 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
4415 GOTO(err_thread, rc);
4417 mds->mds_setattr_service =
4418 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4419 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
4420 MDS_SERVICE_WATCHDOG_TIMEOUT,
4421 mds_handle, "mds_setattr",
4422 obd->obd_proc_entry);
4423 if (!mds->mds_setattr_service) {
4424 CERROR("failed to start getattr service\n");
4425 GOTO(err_thread, rc = -ENOMEM);
4428 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
4429 MDT_NUM_THREADS, "ll_mdt_attr");
4431 GOTO(err_thread2, rc);
4433 mds->mds_readpage_service =
4434 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4435 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
4436 MDS_SERVICE_WATCHDOG_TIMEOUT,
4437 mds_handle, "mds_readpage",
4438 obd->obd_proc_entry);
4439 if (!mds->mds_readpage_service) {
4440 CERROR("failed to start readpage service\n");
4441 GOTO(err_thread2, rc = -ENOMEM);
4444 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
4445 MDT_NUM_THREADS, "ll_mdt_rdpg");
4448 GOTO(err_thread3, rc);
4450 mds->mds_close_service =
4451 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4452 MDS_CLOSE_PORTAL, MDC_REPLY_PORTAL,
4453 MDS_SERVICE_WATCHDOG_TIMEOUT,
4454 mds_handle, "mds_close",
4455 obd->obd_proc_entry);
4456 if (!mds->mds_close_service) {
4457 CERROR("failed to start close service\n");
4458 GOTO(err_thread3, rc = -ENOMEM);
4461 rc = ptlrpc_start_n_threads(obd, mds->mds_close_service,
4462 MDT_NUM_THREADS, "ll_mdt_clos");
4465 GOTO(err_thread4, rc);
4469 ptlrpc_unregister_service(mds->mds_close_service);
4471 ptlrpc_unregister_service(mds->mds_readpage_service);
4473 ptlrpc_unregister_service(mds->mds_setattr_service);
4475 ptlrpc_unregister_service(mds->mds_service);
4479 static int mdt_cleanup(struct obd_device *obd, int flags)
4481 struct mds_obd *mds = &obd->u.mds;
4484 ptlrpc_stop_all_threads(mds->mds_close_service);
4485 ptlrpc_unregister_service(mds->mds_close_service);
4487 ptlrpc_stop_all_threads(mds->mds_readpage_service);
4488 ptlrpc_unregister_service(mds->mds_readpage_service);
4490 ptlrpc_stop_all_threads(mds->mds_setattr_service);
4491 ptlrpc_unregister_service(mds->mds_setattr_service);
4493 ptlrpc_stop_all_threads(mds->mds_service);
4494 ptlrpc_unregister_service(mds->mds_service);
4499 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
4500 __u64 gr, void *data)
4502 struct lustre_id id;
4503 struct obd_device *obd = data;
4507 return mds_id2dentry(obd, &id, NULL);
4510 static int mds_get_info(struct obd_export *exp, __u32 keylen,
4511 void *key, __u32 *valsize, void *val)
4513 struct obd_device *obd;
4514 struct mds_obd *mds;
4518 obd = class_exp2obd(exp);
4522 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
4523 exp->exp_handle.h_cookie);
4527 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
4528 /* get log_context handle. */
4529 struct llog_ctxt *ctxt;
4530 unsigned long *llh_handle = val;
4531 *valsize = sizeof(unsigned long);
4532 ctxt = llog_get_context(&obd->obd_llogs, LLOG_REINT_ORIG_CTXT);
4534 CERROR("Cannot get REINT llog context\n");
4537 *llh_handle = (unsigned long)ctxt;
4540 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
4541 /* get log_context handle. */
4542 unsigned long *sb = val;
4543 *valsize = sizeof(unsigned long);
4544 *sb = (unsigned long)obd->u.mds.mds_sb;
4548 if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
4549 __u32 *mdsize = val;
4550 *valsize = sizeof(*mdsize);
4551 *mdsize = mds->mds_max_mdsize;
4555 if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
4556 __u32 *mdsnum = val;
4557 *valsize = sizeof(*mdsnum);
4558 *mdsnum = mds->mds_num;
4562 if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
4563 struct lustre_id *rootid = val;
4564 *valsize = sizeof(*rootid);
4565 *rootid = mds->mds_rootid;
4569 if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
4570 struct lov_desc *desc = val;
4571 *valsize = sizeof(*desc);
4572 *desc = mds->mds_dt_desc;
4576 if (keylen >= strlen("getext") && strcmp(key, "getext") == 0) {
4577 struct fid_extent *ext = val;
4578 *valsize = sizeof(*ext);
4580 spin_lock(&mds->mds_fidext_lock);
4581 ext->fe_width = MDS_FIDEXT_SIZE;
4582 ext->fe_start = mds->mds_fidext_thumb + 1;
4583 mds->mds_fidext_thumb += MDS_FIDEXT_SIZE;
4584 spin_unlock(&mds->mds_fidext_lock);
4589 rc = fsfilt_get_info(obd, mds->mds_sb, NULL, keylen, key, valsize, val);
4591 CDEBUG(D_IOCTL, "invalid key\n");
4596 struct lvfs_callback_ops mds_lvfs_ops = {
4597 l_id2dentry: mds_lvfs_id2dentry,
4600 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
4601 int objcount, struct obd_ioobj *obj,
4602 int niocount, struct niobuf_remote *nb,
4603 struct niobuf_local *res,
4604 struct obd_trans_info *oti, struct lustre_capa *capa);
4606 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
4607 int objcount, struct obd_ioobj *obj, int niocount,
4608 struct niobuf_local *res, struct obd_trans_info *oti,
4611 /* use obd ops to offer management infrastructure */
4612 static struct obd_ops mds_obd_ops = {
4613 .o_owner = THIS_MODULE,
4614 .o_attach = mds_attach,
4615 .o_detach = mds_detach,
4616 .o_connect = mds_connect,
4617 .o_connect_post = mds_connect_post,
4618 .o_init_export = mds_init_export,
4619 .o_destroy_export = mds_destroy_export,
4620 .o_disconnect = mds_disconnect,
4621 .o_setup = mds_setup,
4622 .o_precleanup = mds_precleanup,
4623 .o_cleanup = mds_cleanup,
4624 .o_process_config = mds_process_config,
4625 .o_postrecov = mds_postrecov,
4626 .o_statfs = mds_obd_statfs,
4627 .o_iocontrol = mds_iocontrol,
4628 .o_create = mds_obd_create,
4629 .o_destroy = mds_obd_destroy,
4630 .o_llog_init = mds_llog_init,
4631 .o_llog_finish = mds_llog_finish,
4632 .o_notify = mds_notify,
4633 .o_get_info = mds_get_info,
4634 .o_set_info = mds_set_info,
4635 .o_preprw = mds_preprw,
4636 .o_commitrw = mds_commitrw,
4639 static struct obd_ops mdt_obd_ops = {
4640 .o_owner = THIS_MODULE,
4641 .o_attach = mdt_attach,
4642 .o_detach = mdt_detach,
4643 .o_setup = mdt_setup,
4644 .o_cleanup = mdt_cleanup,
4647 static int __init mds_init(void)
4649 struct lprocfs_static_vars lvars;
4653 mds_init_lsd_cache();
4654 mds_init_rmtacl_upcall_cache();
4656 lprocfs_init_multi_vars(0, &lvars);
4657 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4658 OBD_MDS_DEVICENAME);
4659 lprocfs_init_multi_vars(1, &lvars);
4660 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4661 OBD_MDT_DEVICENAME);
4663 rc = mds_capa_key_start_thread();
4665 class_unregister_type(OBD_MDT_DEVICENAME);
4666 class_unregister_type(OBD_MDS_DEVICENAME);
4667 mds_cleanup_lsd_cache();
4670 mds_eck_timer.function = mds_capa_key_timer_callback;
4671 mds_eck_timer.data = 0;
4672 init_timer(&mds_eck_timer);
4677 static void /*__exit*/ mds_exit(void)
4679 mds_capa_key_stop_thread();
4680 mds_cleanup_rmtacl_upcall_cache();
4681 mds_cleanup_lsd_cache();
4683 class_unregister_type(OBD_MDS_DEVICENAME);
4684 class_unregister_type(OBD_MDT_DEVICENAME);
4687 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4688 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4689 MODULE_LICENSE("GPL");
4691 module_init(mds_init);
4692 module_exit(mds_exit);