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 "mds_internal.h"
63 static int mds_intent_policy(struct ldlm_namespace *ns,
64 struct ldlm_lock **lockp, void *req_cookie,
65 ldlm_mode_t mode, int flags, void *data);
66 static int mds_postsetup(struct obd_device *obd);
67 static int mds_cleanup(struct obd_device *obd, int flags);
70 /* Assumes caller has already pushed into the kernel filesystem context */
71 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
72 loff_t offset, int count)
74 struct ptlrpc_bulk_desc *desc;
75 struct l_wait_info lwi;
77 int rc = 0, npages, i, tmpcount, tmpsize = 0;
80 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
82 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
83 OBD_ALLOC(pages, sizeof(*pages) * npages);
85 GOTO(out, rc = -ENOMEM);
87 desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
90 GOTO(out_free, rc = -ENOMEM);
92 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
93 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
95 pages[i] = alloc_pages(GFP_KERNEL, 0);
97 GOTO(cleanup_buf, rc = -ENOMEM);
99 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
102 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
103 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
104 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
105 tmpsize, offset, file->f_dentry->d_inode->i_ino,
106 file->f_dentry->d_inode->i_size);
108 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
109 kmap(pages[i]), tmpsize, &offset);
113 GOTO(cleanup_buf, rc = -EIO);
116 LASSERT(desc->bd_nob == count);
118 rc = ptlrpc_start_bulk_transfer(desc);
120 GOTO(cleanup_buf, rc);
122 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
123 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
124 OBD_FAIL_MDS_SENDPAGE, rc = -EIO);
125 GOTO(abort_bulk, rc);
128 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
129 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
130 LASSERT (rc == 0 || rc == -ETIMEDOUT);
133 if (desc->bd_success &&
134 desc->bd_nob_transferred == count)
135 GOTO(cleanup_buf, rc);
137 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
140 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
141 (rc == -ETIMEDOUT) ? "timeout" : "network error",
142 desc->bd_nob_transferred, count,
143 req->rq_export->exp_client_uuid.uuid,
144 req->rq_export->exp_connection->c_remote_uuid.uuid);
146 ptlrpc_fail_export(req->rq_export);
150 ptlrpc_abort_bulk (desc);
152 for (i = 0; i < npages; i++)
154 __free_pages(pages[i], 0);
156 ptlrpc_free_bulk(desc);
158 OBD_FREE(pages, sizeof(*pages) * npages);
163 extern char *ldlm_lockname[];
165 int mds_lock_mode_for_dir(struct obd_device *obd,
166 struct dentry *dentry, int mode)
168 int ret_mode = 0, split;
170 /* any dir access needs couple locks:
171 * 1) on part of dir we gonna lookup/modify in
172 * 2) on a whole dir to protect it from concurrent splitting
173 * and to flush client's cache for readdir()
174 * so, for a given mode and dentry this routine decides what
175 * lock mode to use for lock #2:
176 * 1) if caller's gonna lookup in dir then we need to protect
177 * dir from being splitted only - LCK_CR
178 * 2) if caller's gonna modify dir then we need to protect
179 * dir from being splitted and to flush cache - LCK_CW
180 * 3) if caller's gonna modify dir and that dir seems ready
181 * for splitting then we need to protect it from any
182 * type of access (lookup/modify/split) - LCK_EX -bzzz */
184 split = mds_splitting_expected(obd, dentry);
187 * it is important to check here only for MDS_NO_SPLITTABLE. The reason
188 * is that MDS_NO_SPLITTABLE means dir is not splittable in principle
189 * and another thread will not split it on the quiet. But if we have
190 * MDS_NO_SPLIT_EXPECTED, this means, that dir may be splitted anytime,
191 * but not now (for current thread) and we should consider that it can
192 * happen soon and go that branch which can yield LCK_EX to protect from
193 * possible splitting.
195 if (split == MDS_NO_SPLITTABLE) {
197 * this inode won't be splitted. so we need not to protect from
198 * just flush client's cache on modification.
205 if (mode == LCK_EX) {
207 } else if (mode == LCK_PR) {
209 } else if (mode == LCK_PW) {
211 * caller gonna modify directory. We use concurrent
212 * write lock here to retract client's cache for
215 if (split == MDS_EXPECT_SPLIT) {
217 * splitting possible. serialize any access the
218 * idea is that first one seen dir is splittable
219 * is given exclusive lock and split
220 * directory. caller passes lock mode to
221 * mds_try_to_split_dir() and splitting would be
222 * done with exclusive lock only -bzzz.
224 CDEBUG(D_OTHER, "%s: gonna split %lu/%lu\n",
226 (unsigned long)dentry->d_inode->i_ino,
227 (unsigned long)dentry->d_inode->i_generation);
238 /* only valid locked dentries or errors should be returned */
239 struct dentry *mds_id2locked_dentry(struct obd_device *obd, struct lustre_id *id,
240 struct vfsmount **mnt, int lock_mode,
241 struct lustre_handle *lockh, int *mode,
242 char *name, int namelen, __u64 lockpart)
244 struct dentry *de = mds_id2dentry(obd, id, mnt), *retval = de;
245 ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
246 struct ldlm_res_id res_id = { .name = {0} };
247 int flags = LDLM_FL_ATOMIC_CB, rc;
254 res_id.name[0] = id_fid(id);
255 res_id.name[1] = id_group(id);
258 if (name && IS_PDIROPS(de->d_inode)) {
259 ldlm_policy_data_t cpolicy =
260 { .l_inodebits = { MDS_INODELOCK_UPDATE } };
261 LASSERT(mode != NULL);
262 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
264 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
266 &cpolicy, *mode, &flags,
268 ldlm_completion_ast, NULL, NULL,
269 NULL, 0, NULL, lockh + 1);
270 if (rc != ELDLM_OK) {
272 RETURN(ERR_PTR(-ENOLCK));
275 flags = LDLM_FL_ATOMIC_CB;
277 res_id.name[2] = full_name_hash((unsigned char *)name, namelen);
279 CDEBUG(D_INFO, "take lock on "DLID4":"LPX64"\n",
280 OLID4(id), res_id.name[2]);
283 #warning "No PDIROPS support in the kernel"
285 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
286 LDLM_IBITS, &policy, lock_mode, &flags,
287 mds_blocking_ast, ldlm_completion_ast,
288 NULL, NULL, NULL, 0, NULL, lockh);
289 if (rc != ELDLM_OK) {
291 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
294 ldlm_lock_decref(lockh + 1, *mode);
296 } else if (de->d_inode && de->d_inode->i_nlink == 0) {
297 /* as sometimes we lookup inode by ino/generation through
298 iopen mechanism, it's possible to find already unlinked
299 inode with nlink == 0. let's interpretate the case as
301 CWARN("found already unlinked inode %lu/%u\n",
302 de->d_inode->i_ino, de->d_inode->i_generation);
304 retval = ERR_PTR(-ENOENT);
305 ldlm_lock_decref(lockh, lock_mode);
308 ldlm_lock_decref(lockh + 1, *mode);
315 #ifndef DCACHE_DISCONNECTED
316 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
319 /* Look up an entry by inode number. This function ONLY returns valid dget'd
320 * dentries with an initialized inode or errors */
321 struct dentry *mds_id2dentry(struct obd_device *obd, struct lustre_id *id,
322 struct vfsmount **mnt)
324 struct mds_obd *mds = &obd->u.mds;
325 struct dentry *result;
327 unsigned long ino = 0;
328 __u32 generation = 0;
331 if (!id_ino(id) && id_fid(id)) {
332 struct lustre_id *lid;
334 /* if this is reint case we should use fidmap for resolving
335 * correct local store cookie. */
336 lid = mds_fidmap_lookup(obd, id);
338 RETURN(ERR_PTR(-ENOENT));
340 ino = (unsigned long)id_ino(lid);
341 generation = (__u32)id_gen(lid);
343 CDEBUG(D_DENTRY, "fidmap resolved "DLID4"->"DLID4"\n",
344 OLID4(id), OLID4(lid));
346 } else if (id_ino(id)) {
347 ino = (unsigned long)id_ino(id);
348 generation = (__u32)id_gen(id);
350 CERROR("invalid id for lookup "
351 DLID4"\n", OLID4(id));
356 RETURN(ERR_PTR(-ESTALE));
358 snprintf(idname, sizeof(idname), "0x%lx", ino);
360 CDEBUG(D_DENTRY, "--> mds_id2dentry: ino/gen %lu/%u, sb %p\n",
361 ino, generation, mds->mds_sb);
363 /* under ext3 this is neither supposed to return bad inodes nor NULL
365 result = ll_lookup_one_len(idname, mds->mds_id_de,
370 inode = result->d_inode;
372 RETURN(ERR_PTR(-ENOENT));
374 if (is_bad_inode(inode)) {
375 CERROR("bad inode returned %lu/%u\n",
376 inode->i_ino, inode->i_generation);
378 RETURN(ERR_PTR(-ENOENT));
381 /* here we disabled generation check, as root inode i_generation
382 * of cache mds and real mds are different. */
383 if (id_fid(id) != id_fid(&mds->mds_rootid) && generation != 0 &&
384 inode->i_generation != generation) {
385 /* we didn't find the right inode.. */
386 if (id_group(id) != mds->mds_num) {
387 CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
388 "%u != %u, mds %u != %u, request to wrong MDS?\n",
389 inode->i_ino, (unsigned long)inode->i_nlink,
390 atomic_read(&inode->i_count), inode->i_generation,
391 generation, mds->mds_num, (unsigned)id_group(id));
393 CERROR("bad inode %lu found, link: %lu, ct: %d, generation "
394 "%u != %u, inode is recreated while request handled?\n",
395 inode->i_ino, (unsigned long)inode->i_nlink,
396 atomic_read(&inode->i_count), inode->i_generation,
400 RETURN(ERR_PTR(-ENOENT));
404 *mnt = mds->mds_vfsmnt;
412 int mds_req_add_idmapping(struct ptlrpc_request *req,
413 struct mds_export_data *med)
415 struct mds_req_sec_desc *rsd;
416 struct lustre_sec_desc *lsd;
419 if (!med->med_remote)
422 /* maybe we should do it more completely: invalidate the gss ctxt? */
423 if (req->rq_mapped_uid == MDS_IDMAP_NOTFOUND) {
424 CWARN("didn't find mapped uid\n");
428 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
430 CERROR("Can't unpack security desc\n");
434 lsd = mds_get_lsd(req->rq_mapped_uid);
436 CERROR("can't get LSD(%u), no mapping added\n",
441 rc = mds_idmap_add(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
442 rsd->rsd_gid, lsd->lsd_gid);
448 int mds_req_del_idmapping(struct ptlrpc_request *req,
449 struct mds_export_data *med)
451 struct mds_req_sec_desc *rsd;
452 struct lustre_sec_desc *lsd;
455 if (!med->med_remote)
458 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
460 CERROR("Can't unpack security desc\n");
464 LASSERT(req->rq_mapped_uid != -1);
465 lsd = mds_get_lsd(req->rq_mapped_uid);
467 CERROR("can't get LSD(%u), no idmapping deleted\n",
472 rc = mds_idmap_del(med->med_idmap, rsd->rsd_uid, lsd->lsd_uid,
473 rsd->rsd_gid, lsd->lsd_gid);
478 static int mds_init_export_data(struct ptlrpc_request *req,
479 struct mds_export_data *med)
481 struct obd_connect_data *data, *reply;
482 int ask_remote, ask_local;
485 data = lustre_msg_buf(req->rq_reqmsg, 5, sizeof(*data));
486 reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*data));
487 LASSERT(data && reply);
489 if (med->med_initialized) {
490 CDEBUG(D_SEC, "med already initialized, reconnect?\n");
494 ask_remote = data->ocd_connect_flags & OBD_CONNECT_REMOTE;
495 ask_local = data->ocd_connect_flags & OBD_CONNECT_LOCAL;
497 /* currently the policy is simple: satisfy client as possible
500 if (req->rq_auth_uid == -1) {
502 CWARN("null sec is used, force to be local\n");
506 if (!req->rq_remote_realm)
507 CWARN("local realm asked to be remote\n");
509 } else if (ask_local) {
510 if (req->rq_remote_realm)
511 CWARN("remote realm asked to be local\n");
514 med->med_remote = (req->rq_remote_realm != 0);
517 med->med_nllu = data->ocd_nllu[0];
518 med->med_nllg = data->ocd_nllu[1];
520 med->med_initialized = 1;
522 reply->ocd_connect_flags &= ~(OBD_CONNECT_REMOTE | OBD_CONNECT_LOCAL);
523 if (med->med_remote) {
525 med->med_idmap = mds_idmap_alloc();
528 CERROR("Failed to alloc idmap, following request from "
529 "this client will be refused\n");
531 reply->ocd_connect_flags |= OBD_CONNECT_REMOTE;
532 CDEBUG(D_SEC, "set client as remote\n");
534 reply->ocd_connect_flags |= OBD_CONNECT_LOCAL;
535 CDEBUG(D_SEC, "set client as local\n");
541 static void mds_free_export_data(struct mds_export_data *med)
546 LASSERT(med->med_remote);
547 mds_idmap_free(med->med_idmap);
548 med->med_idmap = NULL;
551 /* Establish a connection to the MDS.
553 * This will set up an export structure for the client to hold state data about
554 * that client, like open files, the last operation number it did on the server,
557 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
558 struct obd_uuid *cluuid, struct obd_connect_data *data,
561 struct mds_export_data *med;
562 struct mds_client_data *mcd;
563 struct obd_export *exp;
567 if (!conn || !obd || !cluuid)
570 /* XXX There is a small race between checking the list and adding a new
571 * connection for the same UUID, but the real threat (list corruption
572 * when multiple different clients connect) is solved.
574 * There is a second race between adding the export to the list, and
575 * filling in the client data below. Hence skipping the case of NULL
576 * mcd above. We should already be controlling multiple connects at the
577 * client, and we can't hold the spinlock over memory allocations
578 * without risk of deadlocking.
580 rc = class_connect(conn, obd, cluuid);
583 exp = class_conn2export(conn);
585 LASSERT(exp != NULL);
586 med = &exp->exp_mds_data;
588 OBD_ALLOC(mcd, sizeof(*mcd));
590 CERROR("%s: out of memory for client data.\n",
592 GOTO(out, rc = -ENOMEM);
595 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
598 rc = mds_client_add(obd, &obd->u.mds, med, -1);
606 OBD_FREE(mcd, sizeof(*mcd));
607 class_disconnect(exp, 0);
609 class_export_put(exp);
614 static int mds_connect_post(struct obd_export *exp, unsigned initial,
617 struct obd_device *obd = exp->exp_obd;
618 struct mds_obd *mds = &obd->u.mds;
619 struct mds_export_data *med;
620 struct mds_client_data *mcd;
624 med = &exp->exp_mds_data;
628 /* some one reconnect initially, we have to reset
629 * data existing export can have. bug 6102 */
630 if (mcd->mcd_last_xid != 0)
631 CDEBUG(D_HA, "initial reconnect to existing export\n");
632 mcd->mcd_last_transno = 0;
633 mcd->mcd_last_xid = 0;
634 mcd->mcd_last_close_xid = 0;
635 mcd->mcd_last_result = 0;
636 mcd->mcd_last_data = 0;
639 if (!(flags & OBD_OPT_MDS_CONNECTION)) {
640 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
641 atomic_inc(&mds->mds_real_clients);
642 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
643 obd->obd_name, exp->exp_client_uuid.uuid,
644 atomic_read(&mds->mds_real_clients));
645 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
647 if (mds->mds_md_name)
648 rc = mds_md_connect(obd, mds->mds_md_name);
653 static int mds_init_export(struct obd_export *exp)
655 struct mds_export_data *med = &exp->exp_mds_data;
657 INIT_LIST_HEAD(&med->med_open_head);
658 spin_lock_init(&med->med_open_lock);
662 static int mds_destroy_export(struct obd_export *export)
664 struct obd_device *obd = export->exp_obd;
665 struct mds_export_data *med = &export->exp_mds_data;
666 struct lvfs_run_ctxt saved;
670 mds_free_export_data(med);
671 target_destroy_export(export);
673 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
676 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
678 /* Close any open files (which may also cause orphan unlinking). */
679 spin_lock(&med->med_open_lock);
680 while (!list_empty(&med->med_open_head)) {
681 struct list_head *tmp = med->med_open_head.next;
682 struct mds_file_data *mfd =
683 list_entry(tmp, struct mds_file_data, mfd_list);
684 struct lustre_id sid;
686 BDEVNAME_DECLARE_STORAGE(btmp);
688 /* bug 1579: fix force-closing for 2.5 */
689 struct dentry *dentry = mfd->mfd_dentry;
691 list_del(&mfd->mfd_list);
692 spin_unlock(&med->med_open_lock);
694 down(&dentry->d_inode->i_sem);
695 rc = mds_read_inode_sid(obd, dentry->d_inode, &sid);
696 up(&dentry->d_inode->i_sem);
698 CERROR("Can't read inode self id, inode %lu, "
699 "rc %d\n", dentry->d_inode->i_ino, rc);
700 memset(&sid, 0, sizeof(sid));
703 /* If you change this message, be sure to update
704 * replay_single:test_46 */
705 CERROR("force closing client file handle for %.*s (%s:"
706 DLID4")\n", dentry->d_name.len, dentry->d_name.name,
707 ll_bdevname(dentry->d_inode->i_sb, btmp),
710 /* child inode->i_alloc_sem protects orphan_dec_test and
711 * is_orphan race, mds_mfd_close drops it */
712 DOWN_WRITE_I_ALLOC_SEM(dentry->d_inode);
713 rc = mds_mfd_close(NULL, 0, obd, mfd,
714 !(export->exp_flags & OBD_OPT_FAILOVER));
716 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
717 spin_lock(&med->med_open_lock);
719 spin_unlock(&med->med_open_lock);
720 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
724 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
728 static int mds_disconnect(struct obd_export *exp, unsigned long flags)
730 unsigned long irqflags;
731 struct obd_device *obd;
736 LASSERT(exp != NULL);
737 obd = class_exp2obd(exp);
739 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
740 exp->exp_handle.h_cookie);
746 * suppress any inter-mds requests durring disconnecting lmv if this is
747 * detected --force mode. This is needed to avoid endless recovery.
749 if (atomic_read(&mds->mds_real_clients) > 0 &&
750 !(exp->exp_flags & OBD_OPT_REAL_CLIENT))
751 flags |= OBD_OPT_FORCE;
753 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
754 && !atomic_read(&mds->mds_real_clients)) {
755 /* there was no client at all */
756 mds_md_disconnect(obd, flags);
759 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
760 && atomic_dec_and_test(&mds->mds_real_clients)) {
761 /* time to drop LMV connections */
762 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
763 "Disconnnect from LMV now\n",
764 obd->obd_name, exp->exp_client_uuid.uuid);
765 mds_md_disconnect(obd, flags);
768 spin_lock_irqsave(&exp->exp_lock, irqflags);
769 exp->exp_flags = flags;
770 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
772 /* disconnect early so that clients can't keep using export */
773 rc = class_disconnect(exp, flags);
774 ldlm_cancel_locks_for_export(exp);
776 /* complete all outstanding replies */
777 spin_lock_irqsave(&exp->exp_lock, irqflags);
778 while (!list_empty(&exp->exp_outstanding_replies)) {
779 struct ptlrpc_reply_state *rs =
780 list_entry(exp->exp_outstanding_replies.next,
781 struct ptlrpc_reply_state, rs_exp_list);
782 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
784 spin_lock(&svc->srv_lock);
785 list_del_init(&rs->rs_exp_list);
786 ptlrpc_schedule_difficult_reply(rs);
787 spin_unlock(&svc->srv_lock);
789 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
793 static int mds_getstatus(struct ptlrpc_request *req)
795 struct mds_obd *mds = mds_req2mds(req);
796 struct mds_body *body;
800 size = sizeof(*body);
802 rc = lustre_pack_reply(req, 1, &size, NULL);
803 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
804 CERROR("mds: out of memory for message: size=%d\n", size);
805 req->rq_status = -ENOMEM; /* superfluous? */
809 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
810 body->valid |= OBD_MD_FID;
812 memcpy(&body->id1, &mds->mds_rootid, sizeof(body->id1));
815 * the last_committed and last_xid fields are filled in for all replies
816 * already - no need to do so here also.
821 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
822 void *data, int flag)
827 if (flag == LDLM_CB_CANCELING) {
828 /* Don't need to do anything here. */
832 /* XXX layering violation! -phil */
833 lock_res_and_lock(lock);
836 * get this: if mds_blocking_ast is racing with mds_intent_policy, such
837 * that mds_blocking_ast is called just before l_i_p takes the ns_lock,
838 * then by the time we get the lock, we might not be the correct
839 * blocking function anymore. So check, and return early, if so.
841 if (lock->l_blocking_ast != mds_blocking_ast) {
842 unlock_res_and_lock(lock);
846 lock->l_flags |= LDLM_FL_CBPENDING;
847 do_ast = (!lock->l_readers && !lock->l_writers);
848 unlock_res_and_lock(lock);
851 struct lustre_handle lockh;
854 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
855 ldlm_lock2handle(lock, &lockh);
856 rc = ldlm_cli_cancel(&lockh);
858 CERROR("ldlm_cli_cancel: %d\n", rc);
860 LDLM_DEBUG(lock, "Lock still has references, will be "
866 static int mds_convert_md(struct obd_device *obd, struct inode *inode,
867 void *md, int size, int mea)
871 if (S_ISREG(inode->i_mode)) {
872 rc = mds_convert_lov_ea(obd, inode, md, size);
873 } else if (S_ISDIR(inode->i_mode)) {
875 rc = mds_convert_mea_ea(obd, inode, md, size);
877 rc = mds_convert_lov_ea(obd, inode, md, size);
880 CERROR("Invalid EA format (nor LOV or MEA) "
881 "is detected. Inode %lu/%u\n",
882 inode->i_ino, inode->i_generation);
888 int mds_get_md(struct obd_device *obd, struct inode *inode,
889 void *md, int *size, int lock, int mea)
898 rc = fsfilt_get_md(obd, inode, md, *size,
899 (mea ? EA_MEA : EA_LOV));
901 CERROR("Error %d reading eadata for ino %lu\n",
905 rc = mds_convert_md(obd, inode, md,
920 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
921 * Call with lock=0 if the caller has already taken the i_sem. */
922 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
923 struct mds_body *body, struct inode *inode, int lock, int mea)
925 struct mds_obd *mds = &obd->u.mds;
930 lmm = lustre_msg_buf(msg, offset, 0);
932 /* Some problem with getting eadata when I sized the reply
934 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
938 lmm_size = msg->buflens[offset];
940 /* I don't really like this, but it is a sanity check on the client
941 * MD request. However, if the client doesn't know how much space
942 * to reserve for the MD, it shouldn't be bad to have too much space.
944 if (lmm_size > mds->mds_max_mdsize) {
945 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
946 inode->i_ino, lmm_size, mds->mds_max_mdsize);
950 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock, mea);
952 body->valid |= S_ISDIR(inode->i_mode) ?
953 OBD_MD_FLDIREA : OBD_MD_FLEASIZE;
956 body->valid |= OBD_MD_MEA;
958 body->eadatasize = lmm_size;
965 int mds_pack_link(struct dentry *dentry, struct ptlrpc_request *req,
966 struct mds_body *repbody, int reply_off)
968 struct inode *inode = dentry->d_inode;
973 symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
974 LASSERT(symname != NULL);
975 len = req->rq_repmsg->buflens[reply_off + 1];
977 rc = inode->i_op->readlink(dentry, symname, len);
979 CERROR("readlink failed: %d\n", rc);
980 } else if (rc != len - 1) {
981 CERROR ("Unexpected readlink rc %d: expecting %d\n",
985 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
986 repbody->valid |= OBD_MD_LINKNAME;
987 repbody->eadatasize = rc + 1;
988 symname[rc] = 0; /* NULL terminate */
995 int mds_pack_xattr(struct dentry *dentry, struct ptlrpc_request *req,
996 struct mds_body *repbody, int req_off, int reply_off)
998 struct inode *inode = dentry->d_inode;
1004 ea_name = lustre_msg_string(req->rq_reqmsg, req_off + 1, 0);
1005 len = req->rq_repmsg->buflens[reply_off + 1];
1007 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
1011 if (!strcmp(ea_name, XATTR_NAME_LUSTRE_ACL)) {
1012 struct rmtacl_upcall_desc desc;
1014 if (len != LUSTRE_ACL_SIZE_MAX || !value) {
1015 CERROR("no reply buffer prepared\n");
1019 memset(&desc, 0, sizeof(desc));
1021 desc.cmd = lustre_msg_string(req->rq_reqmsg, req_off + 2, 0);
1022 desc.cmdlen = req->rq_reqmsg->buflens[req_off + 2];
1023 desc.res = (char *) value;
1024 desc.reslen = LUSTRE_ACL_SIZE_MAX;
1026 mds_do_remote_acl_upcall(&desc);
1028 if (desc.upcall_status)
1029 RETURN(desc.upcall_status);
1031 if (desc.reslen > LUSTRE_ACL_SIZE_MAX) {
1032 CERROR("downcall claim reslen %u\n", desc.reslen);
1035 /* like remote setfacl, steal "flags" in mds_body as the
1036 * exececution status
1038 repbody->flags = desc.status;
1039 repbody->valid |= OBD_MD_FLXATTR;
1040 repbody->eadatasize = desc.reslen;
1045 if (inode->i_op && inode->i_op->getxattr)
1046 rc = inode->i_op->getxattr(dentry, ea_name, value, len);
1049 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1050 CERROR("getxattr failed: %d", rc);
1052 repbody->valid |= OBD_MD_FLXATTR;
1053 repbody->eadatasize = rc;
1060 int mds_pack_xattr_list(struct dentry *dentry, struct ptlrpc_request *req,
1061 struct mds_body *repbody, int reply_off)
1063 struct inode *inode = dentry->d_inode;
1068 len = req->rq_repmsg->buflens[reply_off + 1];
1070 value = lustre_msg_buf(req->rq_repmsg, reply_off + 1, len);
1073 if (inode->i_op && inode->i_op->getxattr)
1074 rc = inode->i_op->listxattr(dentry, value, len);
1077 CERROR("listxattr failed: %d", rc);
1079 repbody->valid |= OBD_MD_FLXATTRLIST;
1080 repbody->eadatasize = rc;
1087 int mds_pack_posix_acl(struct lustre_msg *repmsg, int offset,
1088 struct mds_body *body, struct inode *inode)
1090 struct dentry de = { .d_inode = inode };
1091 __u32 buflen, *sizep;
1096 if (!inode->i_op->getxattr)
1099 sizep = lustre_msg_buf(repmsg, offset, 4);
1101 CERROR("can't locate returned acl size buf\n");
1105 buflen = repmsg->buflens[offset + 1];
1106 buf = lustre_msg_buf(repmsg, offset + 1, buflen);
1108 size = inode->i_op->getxattr(&de, XATTR_NAME_ACL_ACCESS, buf, buflen);
1109 if (size == -ENODATA || size == -EOPNOTSUPP)
1115 *sizep = cpu_to_le32(size);
1116 body->valid |= OBD_MD_FLACL;
1121 int mds_pack_remote_perm(struct ptlrpc_request *req, int reply_off,
1122 struct mds_body *body, struct inode *inode)
1124 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1125 struct lustre_sec_desc *lsd;
1126 struct mds_remote_perm *perm;
1129 LASSERT(inode->i_op);
1130 LASSERT(inode->i_op->permission);
1131 LASSERT(req->rq_export->exp_mds_data.med_remote);
1133 perm = (struct mds_remote_perm *)
1134 lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(perm));
1136 CERROR("no remote perm buf at offset %d\n", reply_off);
1140 memset(perm, 0, sizeof(*perm));
1142 /* obtain authenticated uid/gid and LSD permissions, which
1143 * might be different from current process context, from LSD
1145 lsd = mds_get_lsd(current->uid);
1147 CWARN("can't LSD of uid %u\n", current->uid);
1151 perm->mrp_auth_uid = lsd->lsd_uid;
1152 perm->mrp_auth_gid = lsd->lsd_gid;
1154 lsd_perms = mds_lsd_get_perms(lsd, 1, 0, req->rq_peer.peer_id.nid);
1155 if (lsd_perms & LSD_PERM_SETUID)
1156 perm->mrp_allow_setuid = 1;
1157 if (lsd_perms & LSD_PERM_SETGID)
1158 perm->mrp_allow_setgid = 1;
1162 if (mds_remote_perm_do_reverse_map(med, perm))
1165 /* permission bits of current user
1166 * XXX this is low efficient, could we do it in one blow?
1168 if (inode->i_op->permission(inode, MAY_EXEC, NULL) == 0)
1169 perm->mrp_perm |= MAY_EXEC;
1170 if (inode->i_op->permission(inode, MAY_WRITE, NULL) == 0)
1171 perm->mrp_perm |= MAY_WRITE;
1172 if (inode->i_op->permission(inode, MAY_READ, NULL) == 0)
1173 perm->mrp_perm |= MAY_READ;
1175 body->valid |= (OBD_MD_FLACL | OBD_MD_FLRMTACL);
1180 int mds_pack_acl(struct ptlrpc_request *req, int reply_off,
1181 struct mds_body *body, struct inode *inode)
1185 if (!req->rq_export->exp_mds_data.med_remote)
1186 rc = mds_pack_posix_acl(req->rq_repmsg, reply_off, body, inode);
1188 rc = mds_pack_remote_perm(req, reply_off + 1, body, inode);
1193 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
1194 struct ptlrpc_request *req, int req_off,
1195 struct mds_body *reqbody, int reply_off,
1196 struct mds_req_sec_desc *rsd)
1198 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
1199 struct inode *inode = dentry->d_inode;
1200 struct mds_body *body;
1201 int rc = 0, offset = 0;
1204 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
1207 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
1208 LASSERT(body != NULL); /* caller prepped reply */
1210 if (dentry->d_flags & DCACHE_CROSS_REF) {
1211 mds_pack_dentry2body(obd, body, dentry,
1212 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1213 CDEBUG(D_OTHER, "cross reference: "DLID4"\n",
1218 mds_pack_inode2body(obd, body, inode,
1219 (reqbody->valid & OBD_MD_FID) ? 1 : 0);
1221 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
1222 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
1224 /* guessing what kind og attribute do we need. */
1225 int is_mea = (S_ISDIR(inode->i_mode) &&
1226 (reqbody->valid & OBD_MD_MEA) != 0);
1228 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1,
1229 body, inode, 1, is_mea);
1231 /* if we have LOV EA data, the OST holds size, atime, mtime. */
1232 if (!(body->valid & OBD_MD_FLEASIZE) &&
1233 !(body->valid & OBD_MD_FLDIREA))
1234 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
1235 OBD_MD_FLATIME | OBD_MD_FLMTIME);
1236 } else if (S_ISLNK(inode->i_mode) &&
1237 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
1238 rc = mds_pack_link(dentry, req, body, reply_off);
1239 } else if (reqbody->valid & OBD_MD_FLXATTR) {
1240 rc = mds_pack_xattr(dentry, req, body, req_off, reply_off);
1241 } else if (reqbody->valid & OBD_MD_FLXATTRLIST) {
1242 rc = mds_pack_xattr_list(dentry, req, body, reply_off);
1245 offset = reply_off + ((reqbody->valid & OBD_MD_FLEASIZE) ? 2 : 1);
1246 if (reqbody->valid & OBD_MD_FLACL) {
1247 rc = mds_pack_acl(req, offset, body, inode);
1251 if (reqbody->valid & OBD_MD_FLKEY) {
1252 rc = mds_pack_gskey(obd, req->rq_repmsg, &offset,
1256 mds_pack_audit(obd, inode, body);
1258 if (reqbody->valid & OBD_MD_CAPA) {
1259 struct lustre_capa *req_capa;
1261 LASSERT(!(reqbody->valid & ~OBD_MD_CAPA));
1262 LASSERT(S_ISREG(inode->i_mode));
1264 req_capa = lustre_swab_reqbuf(req, req_off + 1,
1266 lustre_swab_lustre_capa);
1267 if (req_capa == NULL) {
1268 CERROR("Can't unpack capa\n");
1272 offset = reply_off + 1;
1273 rc = mds_pack_capa(obd, med, reqbody, req_capa, req,
1278 mds_body_do_reverse_map(med, body);
1283 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
1284 struct dentry *dentry,
1287 int rc = 0, size[1] = {sizeof(struct mds_body)};
1290 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1291 CERROR("failed MDS_GETATTR_PACK test\n");
1292 req->rq_status = -ENOMEM;
1296 rc = lustre_pack_reply(req, 1, size, NULL);
1298 CERROR("lustre_pack_reply failed: rc %d\n", rc);
1299 GOTO(out, req->rq_status = rc);
1307 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct dentry *de,
1310 struct inode *inode = de->d_inode;
1311 struct mds_obd *mds = mds_req2mds(req);
1312 struct mds_body *body;
1313 int rc = 0, size[4] = {sizeof(*body)}, bufcount = 1;
1316 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1317 LASSERT(body != NULL); /* checked by caller */
1318 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
1320 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
1321 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
1324 down(&inode->i_sem);
1325 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
1326 ((body->valid & OBD_MD_MEA) ? EA_MEA : EA_LOV));
1329 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1330 CERROR("error getting inode %lu MD: rc = %d\n",
1333 } else if (rc > mds->mds_max_mdsize) {
1335 CERROR("MD size %d larger than maximum possible %u\n",
1336 rc, mds->mds_max_mdsize);
1338 size[bufcount] = rc;
1341 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
1342 if (inode->i_size + 1 != body->eadatasize)
1343 CERROR("symlink size: %Lu, reply space: %d\n",
1344 inode->i_size + 1, body->eadatasize);
1345 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
1347 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
1348 inode->i_size + 1, body->eadatasize);
1349 } else if ((body->valid & OBD_MD_FLXATTR)) {
1350 char *ea_name = lustre_msg_string(req->rq_reqmsg,
1354 if (!strcmp(ea_name, XATTR_NAME_LUSTRE_ACL)) {
1355 size[bufcount] = LUSTRE_ACL_SIZE_MAX;
1357 if (inode->i_op && inode->i_op->getxattr)
1358 rc = inode->i_op->getxattr(de, ea_name,
1362 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1363 CERROR("error get inode %lu EA: %d\n",
1367 size[bufcount] = min_t(int,
1368 body->eadatasize, rc);
1372 } else if (body->valid & OBD_MD_FLXATTRLIST) {
1374 if (inode->i_op && inode->i_op->getxattr)
1375 rc = inode->i_op->listxattr(de, NULL, 0);
1378 if (rc != -ENODATA && rc != -EOPNOTSUPP)
1379 CERROR("error getting inode %lu EA: rc = %d\n",
1383 size[bufcount] = min_t(int, body->eadatasize, rc);
1388 /* may co-exist with OBD_MD_FLEASIZE */
1389 if (body->valid & OBD_MD_FLACL) {
1390 if (req->rq_export->exp_mds_data.med_remote) {
1391 size[bufcount++] = sizeof(int);
1392 size[bufcount++] = sizeof(struct mds_remote_perm);
1394 size[bufcount++] = sizeof(int);
1395 size[bufcount++] = xattr_acl_size(LL_ACL_MAX_ENTRIES);
1399 if (body->valid & OBD_MD_FLKEY) {
1400 size[bufcount++] = sizeof(int);
1401 size[bufcount++] = sizeof(struct crypto_key);
1404 if (body->valid & OBD_MD_CAPA)
1405 size[bufcount++] = sizeof(struct lustre_capa);
1407 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
1408 CERROR("failed MDS_GETATTR_PACK test\n");
1409 req->rq_status = -ENOMEM;
1410 GOTO(out, rc = -ENOMEM);
1413 rc = lustre_pack_reply(req, bufcount, size, NULL);
1415 CERROR("out of memory\n");
1416 GOTO(out, req->rq_status = rc);
1424 int mds_check_mds_num(struct obd_device *obd, struct inode *inode,
1425 char *name, int namelen)
1427 struct mea *mea = NULL;
1428 int mea_size, rc = 0;
1431 rc = mds_md_get_attr(obd, inode, &mea, &mea_size);
1434 if (mea != NULL && mea->mea_count) {
1436 * dir is already splitted, check if requested filename should
1437 * live at this MDS or at another one.
1439 int i = mea_name2idx(mea, name, namelen - 1);
1440 if (mea->mea_master != id_group(&mea->mea_ids[i])) {
1442 "inapropriate MDS(%d) for %s. should be "
1443 "%lu(%d)\n", mea->mea_master, name,
1444 (unsigned long)id_group(&mea->mea_ids[i]), i);
1450 OBD_FREE(mea, mea_size);
1454 int mds_getattr_size(struct obd_device *obd, struct dentry *dentry,
1455 struct ptlrpc_request *req, struct mds_body *body)
1457 struct inode *inode = dentry->d_inode;
1460 LASSERT(body != NULL);
1462 if (dentry->d_inode == NULL || !S_ISREG(inode->i_mode))
1465 /* XXX: quite a ugly hack, need to check old code
1466 * drop FLSIZE/FLBLOCKS prior any checking to */
1467 body->valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
1469 if (obd->obd_recovering) {
1470 CDEBUG(D_INODE, "size for "DLID4" is unknown yet (recovering)\n",
1475 if (atomic_read(&inode->i_writecount)) {
1476 /* some one has opened the file for write.
1477 * mds doesn't know actual size */
1478 CDEBUG(D_INODE, "MDS doesn't know actual size for "DLID4"\n",
1482 CDEBUG(D_INODE, "MDS returns "LPD64"/"LPD64" for"DLID4"\n",
1483 body->size, body->blocks, OLID4(&body->id1));
1484 body->valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
1488 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
1489 struct lustre_handle *child_lockh, int child_part)
1491 struct obd_device *obd = req->rq_export->exp_obd;
1492 struct mds_obd *mds = &obd->u.mds;
1493 struct ldlm_reply *rep = NULL;
1494 struct lvfs_run_ctxt saved;
1495 struct mds_req_sec_desc *rsd;
1496 struct mds_body *body;
1497 struct dentry *dparent = NULL, *dchild = NULL;
1498 struct lvfs_ucred uc = {NULL, NULL,};
1499 struct lustre_handle parent_lockh[2] = {{0}, {0}};
1500 unsigned int namesize = 0;
1501 int rc = 0, cleanup_phase = 0, resent_req = 0, update_mode, reply_offset;
1505 LASSERT(!strcmp(obd->obd_type->typ_name, OBD_MDS_DEVICENAME));
1506 MD_COUNTER_INCREMENT(obd, getattr_lock);
1508 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1510 CERROR("Can't unpack security desc\n");
1514 /* swab now, before anyone looks inside the request. */
1515 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1516 lustre_swab_mds_body);
1518 CERROR("Can't swab mds_body\n");
1519 GOTO(cleanup, rc = -EFAULT);
1522 LASSERT_REQSWAB(req, offset + 1);
1523 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
1525 CERROR("Can't unpack name\n");
1526 GOTO(cleanup, rc = -EFAULT);
1528 namesize = req->rq_reqmsg->buflens[offset + 1];
1530 /* namesize less than 2 means we have empty name, probably came from
1531 revalidate by cfid, so no point in having name to be set */
1535 LASSERT (offset == 1 || offset == 3);
1537 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
1543 rc = mds_init_ucred(&uc, req, rsd);
1545 if (child_lockh->cookie == 0)
1546 mds_audit_auth(req, &uc, AUDIT_STAT, &body->id1,
1547 name, namesize - 1);
1551 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1552 cleanup_phase = 1; /* kernel context */
1553 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
1555 LASSERT(namesize > 0);
1556 if (child_lockh->cookie != 0) {
1557 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1561 if (body->valid == OBD_MD_FLID) {
1562 struct mds_body *mds_reply;
1563 int size = sizeof(*mds_reply);
1567 dparent = mds_id2dentry(obd, &body->id1, NULL);
1568 if (IS_ERR(dparent)) {
1569 rc = PTR_ERR(dparent);
1573 * the user requested ONLY the inode number, so do a raw lookup.
1575 rc = lustre_pack_reply(req, 1, &size, NULL);
1577 CERROR("out of memory\n");
1581 dir = dparent->d_inode;
1582 LASSERT(dir->i_op->lookup_raw != NULL);
1583 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
1585 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
1586 sizeof(*mds_reply));
1588 id_ino(&mds_reply->id1) = inum;
1589 mds_reply->valid = OBD_MD_FLID;
1593 if (resent_req == 0) {
1594 LASSERT(id_fid(&body->id1) != 0);
1596 rc = mds_get_parent_child_locked(obd, mds, &body->id1,
1597 parent_lockh, &dparent,
1599 MDS_INODELOCK_UPDATE,
1602 child_lockh, &dchild,
1603 LCK_PR, child_part);
1607 cleanup_phase = 2; /* dchild, dparent, locks */
1610 * let's make sure this name should leave on this mds
1613 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1617 /* we have no dentry here, drop LOOKUP bit */
1618 /* FIXME: we need MDS_INODELOCK_LOOKUP or not. */
1619 child_part &= ~MDS_INODELOCK_LOOKUP;
1620 CDEBUG(D_OTHER, "%s: retrieve attrs for "DLID4"\n",
1621 obd->obd_name, OLID4(&body->id1));
1623 dchild = mds_id2locked_dentry(obd, &body->id1, NULL,
1624 LCK_PR, parent_lockh,
1625 &update_mode, NULL, 0,
1626 MDS_INODELOCK_UPDATE);
1627 if (IS_ERR(dchild)) {
1628 CERROR("can't find inode with id "DLID4", err = %d\n",
1629 OLID4(&body->id1), (int)PTR_ERR(dchild));
1630 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 mds_audit(req, dchild, name, namesize - 1,
1703 if (rc && DENTRY_VALID(dchild))
1704 ldlm_lock_decref(child_lockh, LCK_PR);
1706 ldlm_lock_decref(parent_lockh, LCK_PR);
1708 if (parent_lockh[1].cookie != 0)
1709 ldlm_lock_decref(parent_lockh + 1, update_mode);
1717 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1719 mds_exit_ucred(&uc);
1724 static int mds_getattr(struct ptlrpc_request *req, int offset)
1726 struct obd_device *obd = req->rq_export->exp_obd;
1727 struct lvfs_run_ctxt saved;
1729 struct mds_req_sec_desc *rsd;
1730 struct mds_body *body;
1731 struct lvfs_ucred uc = {NULL, NULL,};
1735 MD_COUNTER_INCREMENT(obd, getattr);
1737 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1739 CERROR("Can't unpack security desc\n");
1743 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1744 lustre_swab_mds_body);
1746 CERROR ("Can't unpack body\n");
1750 rc = mds_init_ucred(&uc, req, rsd);
1752 mds_exit_ucred(&uc);
1756 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1757 de = mds_id2dentry(obd, &body->id1, NULL);
1759 rc = req->rq_status = PTR_ERR(de);
1763 rc = mds_getattr_pack_msg(req, de, offset);
1765 CERROR("mds_getattr_pack_msg: %d\n", rc);
1769 req->rq_status = mds_getattr_internal(obd, de, req, offset, body,
1775 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1776 mds_exit_ucred(&uc);
1780 static int mds_access_check(struct ptlrpc_request *req, int offset)
1782 struct obd_device *obd = req->rq_export->exp_obd;
1783 struct lvfs_run_ctxt saved;
1785 struct mds_req_sec_desc *rsd;
1786 struct mds_body *body;
1787 struct lvfs_ucred uc;
1788 int rep_size[2] = {sizeof(*body),
1789 sizeof(struct mds_remote_perm)};
1793 if (!req->rq_export->exp_mds_data.med_remote) {
1794 CERROR("from local client "LPU64"\n", req->rq_peer.peer_id.nid);
1798 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1800 CERROR("Can't unpack security desc\n");
1804 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1805 lustre_swab_mds_body);
1807 CERROR ("Can't unpack body\n");
1811 MD_COUNTER_INCREMENT(obd, access_check);
1813 rc = mds_init_ucred(&uc, req, rsd);
1815 CERROR("init ucred error: %d\n", rc);
1818 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1820 de = mds_id2dentry(obd, &body->id1, NULL);
1822 CERROR("grab ino "LPU64": err %ld\n",
1823 body->id1.li_stc.u.e3s.l3s_ino, PTR_ERR(de));
1824 GOTO(out_pop, rc = PTR_ERR(de));
1827 rc = lustre_pack_reply(req, 2, rep_size, NULL);
1829 CERROR("pack reply error: %d\n", rc);
1830 GOTO(out_dput, rc = -EINVAL);
1833 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1836 rc = mds_pack_remote_perm(req, 1, body, de->d_inode);
1843 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1844 mds_exit_ucred(&uc);
1848 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1849 unsigned long max_age)
1854 spin_lock(&obd->obd_osfs_lock);
1855 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1857 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1858 spin_unlock(&obd->obd_osfs_lock);
1863 static int mds_statfs(struct ptlrpc_request *req)
1865 struct obd_device *obd = req->rq_export->exp_obd;
1866 int rc, size = sizeof(struct obd_statfs);
1869 /* This will trigger a watchdog timeout */
1870 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1871 (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1873 rc = lustre_pack_reply(req, 1, &size, NULL);
1874 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1875 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1879 OBD_COUNTER_INCREMENT(obd, statfs);
1881 /* We call this so that we can cache a bit - 1 jiffie worth */
1882 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1885 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1891 req->rq_status = rc;
1895 static int mds_sync(struct ptlrpc_request *req, int offset)
1897 struct obd_device *obd = req->rq_export->exp_obd;
1898 struct mds_obd *mds = &obd->u.mds;
1899 struct mds_body *body;
1900 int rc, size = sizeof(*body);
1903 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1904 lustre_swab_mds_body);
1906 GOTO(out, rc = -EPROTO);
1908 rc = lustre_pack_reply(req, 1, &size, NULL);
1909 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1910 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1914 if (id_ino(&body->id1) == 0) {
1915 /* an id of zero is taken to mean "sync whole filesystem" */
1916 rc = fsfilt_sync(obd, mds->mds_sb);
1920 /* just any file to grab fsync method - "file" arg unused */
1921 struct file *file = mds->mds_rcvd_filp;
1922 struct mds_body *rep_body;
1925 de = mds_id2dentry(obd, &body->id1, NULL);
1927 GOTO(out, rc = PTR_ERR(de));
1929 rc = file->f_op->fsync(NULL, de, 1);
1933 rep_body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep_body));
1934 mds_pack_inode2body(obd, rep_body, de->d_inode,
1935 (body->valid & OBD_MD_FID) ? 1 : 0);
1941 req->rq_status = rc;
1945 /* mds_readpage does not take a DLM lock on the inode, because the client must
1946 * already have a PR lock.
1948 * If we were to take another one here, a deadlock will result, if another
1949 * thread is already waiting for a PW lock. */
1950 static int mds_readpage(struct ptlrpc_request *req, int offset)
1952 struct obd_device *obd = req->rq_export->exp_obd;
1953 struct vfsmount *mnt;
1956 struct mds_req_sec_desc *rsd;
1957 struct mds_body *body, *repbody;
1958 struct lvfs_run_ctxt saved;
1959 int rc, size = sizeof(*repbody);
1960 struct lvfs_ucred uc = {NULL, NULL,};
1963 rc = lustre_pack_reply(req, 1, &size, NULL);
1964 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1965 CERROR("mds: out of memory\n");
1966 GOTO(out, rc = -ENOMEM);
1969 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1971 CERROR("Can't unpack security desc\n");
1972 GOTO (out, rc = -EFAULT);
1975 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1976 lustre_swab_mds_body);
1978 CERROR("Can't unpack body\n");
1979 GOTO (out, rc = -EFAULT);
1982 rc = mds_init_ucred(&uc, req, rsd);
1984 mds_audit_auth(req, &uc, AUDIT_READDIR, &body->id1,
1989 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1990 de = mds_id2dentry(obd, &body->id1, &mnt);
1992 GOTO(out_pop, rc = PTR_ERR(de));
1994 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1996 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1997 /* note: in case of an error, dentry_open puts dentry */
1999 GOTO(out_pop, rc = PTR_ERR(file));
2001 /* body->size is actually the offset -eeb */
2002 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
2003 CERROR("offset "LPU64" not on a block boundary of %lu\n",
2004 body->size, de->d_inode->i_blksize);
2005 GOTO(out_file, rc = -EFAULT);
2008 /* body->nlink is actually the #bytes to read -eeb */
2009 if (body->nlink & (de->d_inode->i_blksize - 1)) {
2010 CERROR("size %u is not multiple of blocksize %lu\n",
2011 body->nlink, de->d_inode->i_blksize);
2012 GOTO(out_file, rc = -EFAULT);
2015 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
2016 repbody->size = file->f_dentry->d_inode->i_size;
2017 repbody->valid = OBD_MD_FLSIZE;
2019 /* to make this asynchronous make sure that the handling function
2020 doesn't send a reply when this function completes. Instead a
2021 callback function would send the reply */
2022 /* body->size is actually the offset -eeb */
2023 rc = mds_sendpage(req, file, body->size, body->nlink);
2027 filp_close(file, 0);
2029 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2031 mds_exit_ucred(&uc);
2032 req->rq_status = rc;
2036 int mds_read_md(struct obd_device *obd, struct lustre_id *id,
2037 char **data, int *datalen)
2039 struct dentry *dentry;
2040 struct mds_obd *mds = &obd->u.mds;
2041 int rc = 0, mea = 0;
2048 dentry = mds_id2dentry(obd, id, NULL);
2050 GOTO(out, rc = PTR_ERR(dentry));
2052 if (!dentry->d_inode) {
2053 CERROR("Can't find object "DLID4".\n",
2055 GOTO(out_dentry, rc = -EINVAL);
2057 if (S_ISDIR(dentry->d_inode->i_mode)) {
2058 *datalen = obd_packmd(mds->mds_md_exp, NULL, NULL);
2061 *datalen = obd_packmd(mds->mds_dt_exp, NULL, NULL);
2064 OBD_ALLOC(ea, *datalen);
2067 GOTO(out_dentry, rc = PTR_ERR(dentry));
2070 down(&dentry->d_inode->i_sem);
2071 rc = fsfilt_get_md(obd, dentry->d_inode, *data, *datalen,
2072 (mea ? EA_MEA : EA_LOV));
2073 up(&dentry->d_inode->i_sem);
2076 CERROR("Error %d reading eadata for ino %lu\n",
2077 rc, dentry->d_inode->i_ino);
2083 EXPORT_SYMBOL(mds_read_md);
2085 int mds_reint(struct ptlrpc_request *req, int offset,
2086 struct lustre_handle *lockh)
2088 struct mds_update_record *rec;
2089 struct mds_req_sec_desc *rsd;
2093 OBD_ALLOC(rec, sizeof(*rec));
2097 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
2099 CERROR("Can't unpack security desc\n");
2100 GOTO(out, rc = -EFAULT);
2103 rc = mds_update_unpack(req, offset, rec);
2104 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
2105 CERROR("invalid record\n");
2106 GOTO(out, req->rq_status = -EINVAL);
2109 rc = mds_init_ucred(&rec->ur_uc, req, rsd);
2111 mds_audit_reint(req, rec);
2115 /* rc will be used to interrupt a for loop over multiple records */
2116 rc = mds_reint_rec(rec, offset, req, lockh);
2119 mds_exit_ucred(&rec->ur_uc);
2120 OBD_FREE(rec, sizeof(*rec));
2124 static int mds_filter_recovery_request(struct ptlrpc_request *req,
2125 struct obd_device *obd, int *process)
2127 switch (req->rq_reqmsg->opc) {
2128 case MDS_CONNECT: /* This will never get here, but for completeness. */
2129 case OST_CONNECT: /* This will never get here, but for completeness. */
2130 case MDS_DISCONNECT:
2131 case OST_DISCONNECT:
2136 case MDS_SYNC: /* used in unmounting */
2141 *process = target_queue_recovery_request(req, obd);
2145 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
2147 /* XXX what should we set rq_status to here? */
2148 req->rq_status = -EAGAIN;
2149 RETURN(ptlrpc_error(req));
2153 static char *reint_names[] = {
2154 [REINT_SETATTR] "setattr",
2155 [REINT_CREATE] "create",
2156 [REINT_LINK] "link",
2157 [REINT_UNLINK] "unlink",
2158 [REINT_RENAME] "rename",
2159 [REINT_OPEN] "open",
2162 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER | \
2163 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ| \
2164 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME| \
2167 static void reconstruct_create(struct ptlrpc_request *req)
2169 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2170 struct mds_client_data *mcd = med->med_mcd;
2171 struct dentry *dentry;
2172 struct ost_body *body;
2173 struct lustre_id id;
2177 /* copy rc, transno and disp; steal locks */
2178 mds_req_from_mcd(req, mcd);
2179 if (req->rq_status) {
2187 id_ino(&id) = mcd->mcd_last_data;
2188 LASSERT(id_ino(&id) != 0);
2190 dentry = mds_id2dentry(req2obd(req), &id, NULL);
2191 if (IS_ERR(dentry)) {
2192 CERROR("can't find inode "LPU64"\n", id_ino(&id));
2193 req->rq_status = PTR_ERR(dentry);
2198 CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
2199 req->rq_xid, id_ino(&id), dentry->d_inode->i_ino,
2200 dentry->d_inode->i_generation);
2202 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
2203 obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
2204 body->oa.o_id = dentry->d_inode->i_ino;
2205 body->oa.o_generation = dentry->d_inode->i_generation;
2206 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2208 down(&dentry->d_inode->i_sem);
2209 rc = mds_read_inode_sid(req2obd(req), dentry->d_inode, &id);
2210 up(&dentry->d_inode->i_sem);
2212 CERROR("Can't read inode self id, inode %lu, "
2213 "rc %d\n", dentry->d_inode->i_ino, rc);
2217 body->oa.o_fid = id_fid(&id);
2218 body->oa.o_mds = id_group(&id);
2224 static int mds_inode_init_acl(struct obd_device *obd, void *handle,
2225 struct dentry *de, void *xattr, int xattr_size)
2227 struct inode *inode = de->d_inode;
2228 struct posix_acl *acl;
2235 LASSERT(xattr_size > 0);
2237 if (!inode->i_op->getxattr || !inode->i_op->setxattr) {
2238 CERROR("backend fs dosen't support xattr\n");
2242 /* set default acl */
2243 if (S_ISDIR(inode->i_mode)) {
2244 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_DEFAULT,
2245 xattr, xattr_size, 0);
2247 CERROR("set default acl err: %d\n", rc);
2252 /* set access acl */
2253 acl = posix_acl_from_xattr(xattr, xattr_size);
2254 if (acl == NULL || IS_ERR(acl)) {
2255 CERROR("insane attr data\n");
2256 return PTR_ERR(acl);
2259 if (posix_acl_valid(acl)) {
2260 CERROR("default acl not valid: %d\n", rc);
2265 mode = inode->i_mode;
2266 rc = posix_acl_create_masq(acl, &mode);
2268 CERROR("create masq err %d\n", rc);
2272 if (inode->i_mode != mode) {
2273 struct iattr iattr = { .ia_valid = ATTR_MODE,
2277 rc2 = fsfilt_setattr(obd, de, handle, &iattr, 0);
2279 CERROR("setattr mode err: %d\n", rc2);
2286 /* we didn't change acl except mode bits of some
2287 * entries, so should be fit into original size.
2289 rc = posix_acl_to_xattr(acl, xattr, xattr_size);
2292 rc = inode->i_op->setxattr(de, XATTR_NAME_ACL_ACCESS,
2293 xattr, xattr_size, 0);
2295 CERROR("set access acl err: %d\n", rc);
2298 posix_acl_release(acl);
2302 static int mdt_obj_create(struct ptlrpc_request *req)
2304 struct obd_device *obd = req->rq_export->exp_obd;
2305 struct mds_obd *mds = &obd->u.mds;
2306 struct ost_body *body, *repbody;
2309 char idname[LL_ID_NAMELEN];
2310 int size = sizeof(*repbody);
2311 struct inode *parent_inode;
2312 struct lvfs_run_ctxt saved;
2313 int rc, cleanup_phase = 0;
2314 struct dentry *new = NULL;
2315 struct dentry_params dp;
2316 int mealen, flags = 0;
2317 struct lvfs_ucred uc;
2318 struct lustre_id id;
2320 void *handle = NULL;
2321 unsigned long cr_inum = 0;
2325 DEBUG_REQ(D_HA, req, "create remote object");
2326 parent_inode = mds->mds_unnamed_dir->d_inode;
2328 body = lustre_swab_reqbuf(req, 0, sizeof(*body),
2329 lustre_swab_ost_body);
2333 /* acl data is packed transparently, no swab here */
2334 LASSERT(req->rq_reqmsg->bufcount >= 2);
2335 acl_size = req->rq_reqmsg->buflens[1];
2337 acl = lustre_msg_buf(req->rq_reqmsg, 1, acl_size);
2339 CERROR("No default acl buf?\n");
2344 rc = lustre_pack_reply(req, 1, &size, NULL);
2348 MDS_CHECK_RESENT(req, reconstruct_create(req));
2351 uc.luc_ginfo = NULL;
2352 uc.luc_uid = body->oa.o_uid;
2353 uc.luc_gid = body->oa.o_gid;
2354 uc.luc_fsuid = body->oa.o_uid;
2355 uc.luc_fsgid = body->oa.o_gid;
2357 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2358 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
2360 /* in REPLAY case inum should be given (client or other MDS fills it) */
2361 if (body->oa.o_id && ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2362 (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))) {
2364 * this is re-create request from MDS holding directory name.
2365 * we have to lookup given ino/gen first. if it exists (good
2366 * case) then there is nothing to do. if it does not then we
2367 * have to recreate it.
2369 id_ino(&id) = body->oa.o_id;
2370 id_gen(&id) = body->oa.o_generation;
2372 new = mds_id2dentry(obd, &id, NULL);
2373 if (!IS_ERR(new) && new->d_inode) {
2374 struct lustre_id sid;
2376 CDEBUG(D_OTHER, "mkdir repairing %lu/%lu\n",
2377 (unsigned long)id_ino(&id),
2378 (unsigned long)id_gen(&id));
2380 obdo_from_inode(&repbody->oa, new->d_inode,
2381 FILTER_VALID_FLAGS);
2383 repbody->oa.o_id = new->d_inode->i_ino;
2384 repbody->oa.o_generation = new->d_inode->i_generation;
2385 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
2388 down(&new->d_inode->i_sem);
2389 rc = mds_read_inode_sid(obd, new->d_inode, &sid);
2390 up(&new->d_inode->i_sem);
2392 CERROR("Can't read inode self id "
2393 "inode %lu, rc %d.\n",
2394 new->d_inode->i_ino, rc);
2398 repbody->oa.o_fid = id_fid(&sid);
2399 repbody->oa.o_mds = id_group(&sid);
2400 LASSERT(id_fid(&sid) != 0);
2403 * here we could use fid passed in body->oa.o_fid and
2404 * thus avoid mds_read_inode_sid().
2406 cr_inum = new->d_inode->i_ino;
2407 GOTO(cleanup, rc = 0);
2411 down(&parent_inode->i_sem);
2412 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
2413 if (IS_ERR(handle)) {
2414 up(&parent_inode->i_sem);
2415 CERROR("fsfilt_start() failed, rc = %d\n",
2416 (int)PTR_ERR(handle));
2417 GOTO(cleanup, rc = PTR_ERR(handle));
2419 cleanup_phase = 1; /* transaction */
2422 rc = sprintf(idname, "%u.%u", ll_insecure_random_int(), current->pid);
2423 new = lookup_one_len(idname, mds->mds_unnamed_dir, rc);
2425 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
2426 obd->obd_name, idname, (int) PTR_ERR(new));
2427 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
2428 up(&parent_inode->i_sem);
2429 RETURN(PTR_ERR(new));
2430 } else if (new->d_inode) {
2431 CERROR("%s: name exists. repeat\n", obd->obd_name);
2434 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2435 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2436 fid = body->oa.o_fid;
2438 fid = mds_alloc_fid(obd);
2440 new->d_fsdata = (void *)&dp;
2445 dp.p_group = mds->mds_num;
2447 if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
2448 (body->oa.o_flags & OBD_FL_RECREATE_OBJS)) {
2449 LASSERT(body->oa.o_id != 0);
2450 dp.p_inum = body->oa.o_id;
2451 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
2452 (unsigned long)body->oa.o_id,
2453 (unsigned long)body->oa.o_generation);
2456 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
2459 rc = mds_inode_init_acl(obd, handle, new,
2462 up(&parent_inode->i_sem);
2466 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2467 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2468 new->d_inode->i_generation = body->oa.o_generation;
2469 mark_inode_dirty(new->d_inode);
2472 * avoiding asserts in cache flush case, as
2473 * @body->oa.o_id should be zero.
2475 if (body->oa.o_id) {
2476 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
2477 "BUG 3550: failed to recreate obj "
2478 LPU64" -> %lu\n", body->oa.o_id,
2479 new->d_inode->i_ino);
2481 LASSERTF(body->oa.o_generation ==
2482 new->d_inode->i_generation,
2483 "BUG 3550: failed to recreate obj/gen "
2484 LPU64"/%u -> %lu/%u\n", body->oa.o_id,
2485 body->oa.o_generation,
2486 new->d_inode->i_ino,
2487 new->d_inode->i_generation);
2491 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
2492 repbody->oa.o_id = new->d_inode->i_ino;
2493 repbody->oa.o_generation = new->d_inode->i_generation;
2494 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER | OBD_MD_FID;
2496 if ((body->oa.o_flags & OBD_FL_RECREATE_OBJS) ||
2497 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2498 LASSERT(body->oa.o_id != 0);
2499 LASSERT(body->oa.o_fid != 0);
2502 mds_inode2id(obd, &id, new->d_inode, fid);
2503 mds_update_inode_ids(obd, new->d_inode, handle, &id,
2504 body->oa.o_valid & OBD_MD_FLID ?
2505 NULL : obdo_id(&body->oa));
2507 /* initializing o_fid after it is allocated. */
2508 repbody->oa.o_fid = id_fid(&id);
2509 repbody->oa.o_mds = id_group(&id);
2511 rc = fsfilt_del_dir_entry(obd, new);
2512 up(&parent_inode->i_sem);
2514 CERROR("can't remove name for object: %d\n", rc);
2518 cleanup_phase = 2; /* created directory object */
2520 CDEBUG(D_OTHER, "created dirobj: %lu/%lu mode %o\n",
2521 (unsigned long)new->d_inode->i_ino,
2522 (unsigned long)new->d_inode->i_generation,
2523 (unsigned)new->d_inode->i_mode);
2524 cr_inum = new->d_inode->i_ino;
2526 up(&parent_inode->i_sem);
2527 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
2531 if (body->oa.o_valid & OBD_MD_FLID) {
2532 /* this is new object for splitted dir. We have to prevent
2533 * recursive splitting on it -bzzz */
2534 mealen = obd_size_diskmd(mds->mds_md_exp, NULL);
2536 OBD_ALLOC(mea, mealen);
2538 GOTO(cleanup, rc = -ENOMEM);
2540 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
2541 mea->mea_master = body->oa.o_mds; /* master mds num */
2544 obdo2id(&mea->mea_ids[body->oa.o_mds], &body->oa);
2546 down(&new->d_inode->i_sem);
2547 rc = fsfilt_set_md(obd, new->d_inode, handle,
2548 mea, mealen, EA_MEA);
2549 up(&new->d_inode->i_sem);
2551 CERROR("fsfilt_set_md() failed, "
2554 OBD_FREE(mea, mealen);
2556 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
2557 obd->obd_name, new->d_inode->i_ino,
2558 new->d_inode->i_generation, flags);
2559 } else if (body->oa.o_easize) {
2560 /* we pass LCK_EX to split routine to signal that we have
2561 * exclusive access to the directory. simple because nobody
2562 * knows it already exists -bzzz */
2563 rc = mds_try_to_split_dir(obd, new, NULL,
2564 body->oa.o_easize, LCK_EX);
2566 CERROR("Can't split directory %lu, error = %d.\n",
2567 new->d_inode->i_ino, rc);
2575 if (rc == 0 && (body->oa.o_flags & OBD_FL_REINT)) {
2576 rc = mds_fidmap_add(obd, &id);
2578 CERROR("can't create fid->ino mapping, "
2585 switch (cleanup_phase) {
2586 case 2: /* object has been created, but we'll may want to replay it later */
2588 ptlrpc_require_repack(req);
2589 case 1: /* transaction */
2590 rc = mds_finish_transno(mds, parent_inode, handle,
2595 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
2599 static int mdt_get_info(struct ptlrpc_request *req)
2601 struct obd_export *exp = req->rq_export;
2606 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2608 DEBUG_REQ(D_HA, req, "no get_info key");
2611 keylen = req->rq_reqmsg->buflens[0];
2613 if ((keylen < strlen("mdsize") || strcmp(key, "mdsize") != 0) &&
2614 (keylen < strlen("mdsnum") || strcmp(key, "mdsnum") != 0) &&
2615 (keylen < strlen("lovdesc") || strcmp(key, "lovdesc") != 0) &&
2616 (keylen < strlen("getext") || strcmp(key, "getext") != 0) &&
2617 (keylen < strlen("rootid") || strcmp(key, "rootid") != 0))
2620 if (keylen >= strlen("rootid") && !strcmp(key, "rootid")) {
2621 struct lustre_id *reply;
2622 int size = sizeof(*reply);
2624 rc = lustre_pack_reply(req, 1, &size, NULL);
2628 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2629 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2630 } else if (keylen >= strlen("lovdesc") && !strcmp(key, "lovdesc")) {
2631 struct lov_desc *reply;
2632 int size = sizeof(*reply);
2634 rc = lustre_pack_reply(req, 1, &size, NULL);
2638 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2639 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2640 } else if (keylen >= strlen("getext") && !strcmp(key, "getext")) {
2641 struct fid_extent *reply;
2642 int size = sizeof(*reply);
2644 rc = lustre_pack_reply(req, 1, &size, NULL);
2648 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2649 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2652 int size = sizeof(*reply);
2654 rc = lustre_pack_reply(req, 1, &size, NULL);
2658 reply = lustre_msg_buf(req->rq_repmsg, 0, size);
2659 rc = obd_get_info(exp, keylen, key, (__u32 *)&size, reply);
2662 req->rq_repmsg->status = 0;
2666 static int mds_set_info(struct obd_export *exp, __u32 keylen,
2667 void *key, __u32 vallen, void *val)
2669 struct obd_device *obd;
2670 struct mds_obd *mds;
2674 obd = class_exp2obd(exp);
2676 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2677 exp->exp_handle.h_cookie);
2682 if (keylen == 5 && memcmp(key, "audit", 5) == 0) {
2683 rc = mds_set_audit(obd, val);
2685 } else if (keylen >= strlen("ids") && memcmp(key, "ids", keylen) == 0) {
2686 struct lustre_id *ids = (struct lustre_id *)val;
2688 struct inode *inode;
2692 de = mds_id2dentry(obd, ids, NULL);
2695 CERROR("lookup by an id error rc=%d\n ", rc);
2698 inode = de->d_inode;
2700 GOTO(out_put, rc = -ENOENT);
2702 down(&inode->i_sem);
2703 handle = fsfilt_start(obd, inode, FSFILT_OP_SETATTR, NULL);
2704 if (IS_ERR(handle)) {
2706 GOTO(out_put, rc = PTR_ERR(handle));
2709 rc = mds_update_inode_ids(obd, inode, handle, NULL, ids + 1);
2711 err = fsfilt_commit(obd, mds->mds_sb, inode, handle,
2714 CERROR("error committing transaction: %d\n", err);
2723 if (keylen >= strlen("crypto_type") &&
2724 memcmp(key, "crypto_type", keylen) == 0) {
2725 rc = mds_set_crypto_type(obd, val, vallen);
2729 if (keylen >= strlen("setext") && !memcmp(key, "setext", keylen)) {
2730 struct fid_extent *ext = val;
2732 CDEBUG(D_IOCTL, "set last fid to extent ["LPD64"-"LPD64"]\n",
2733 ext->fe_start, ext->fe_width);
2735 /* set lastfid into fid extent start. All next object creates
2736 * will use that fid. */
2737 mds_set_last_fid(obd, ext->fe_start);
2739 /* setting the same extent to OSC to avoid ids intersecting in
2740 * object ids, as all cache MDSs have the same group 0. */
2741 rc = obd_set_info(mds->mds_dt_exp, strlen("setext"),
2742 "setext", sizeof(*ext), ext);
2744 CERROR("can't set extent ["LPD64"-"LPD64"] to %s, "
2745 "err %d\n", ext->fe_start, ext->fe_width,
2746 mds->mds_dt_exp->exp_obd->obd_name, rc);
2750 CDEBUG(D_IOCTL, "invalid key\n");
2754 static int mdt_set_info(struct ptlrpc_request *req)
2757 struct obd_export *exp = req->rq_export;
2758 int keylen, rc = 0, vallen = 0;
2761 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
2763 DEBUG_REQ(D_HA, req, "no set_info key");
2766 keylen = req->rq_reqmsg->buflens[0];
2768 if ((keylen == strlen("crypto_type") &&
2769 memcmp(key, "crypto_type", keylen) == 0)) {
2770 rc = lustre_pack_reply(req, 0, NULL, NULL);
2774 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2775 vallen = req->rq_reqmsg->buflens[1];
2777 rc = obd_set_info(exp, keylen, key, vallen, val);
2778 req->rq_repmsg->status = 0;
2780 } else if (keylen == 5 && memcmp(key, "audit", 5) == 0) {
2781 struct audit_attr_msg msg, *p;
2784 rc = lustre_pack_reply(req, 0, NULL, NULL);
2788 p = lustre_swab_reqbuf(req, 1, sizeof(msg),
2789 lustre_swab_audit_attr);
2792 CDEBUG(D_INFO, "Get new audit setting 0x%x\n", (__u32)msg.attr);
2793 rc = obd_set_info(exp, keylen, key, sizeof(msg), &msg);
2795 req->rq_repmsg->status = rc;
2797 } else if (keylen == strlen("ids") &&
2798 memcmp(key, "ids", keylen) == 0) {
2799 struct lustre_id *id, ids[2];
2801 rc = lustre_pack_reply(req, 0, NULL, NULL);
2804 id = lustre_swab_reqbuf(req, 1, sizeof(struct lustre_id),
2805 lustre_swab_lustre_id);
2807 id = lustre_swab_reqbuf(req, 2, sizeof(struct lustre_id),
2808 lustre_swab_lustre_id);
2811 rc = obd_set_info(exp, keylen, key, vallen, ids);
2812 req->rq_repmsg->status = rc;
2814 } else if (keylen == strlen("setext") &&
2815 memcmp(key, "setext", keylen) == 0) {
2816 rc = lustre_pack_reply(req, 0, NULL, NULL);
2820 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
2821 vallen = req->rq_reqmsg->buflens[1];
2823 rc = obd_set_info(exp, keylen, key, vallen, val);
2824 req->rq_repmsg->status = 0;
2828 CDEBUG(D_IOCTL, "invalid key\n");
2832 static void mds_revoke_export_locks(struct obd_export *exp)
2834 struct list_head *locklist = &exp->exp_ldlm_data.led_held_locks;
2835 struct list_head rpc_list;
2836 struct ldlm_lock *lock, *next;
2837 struct ldlm_lock_desc desc;
2839 /* don't do this for local client */
2840 if (!exp->u.eu_mds_data.med_remote)
2843 /* don't revoke locks during recovery */
2844 if (exp->exp_obd->obd_recovering)
2848 INIT_LIST_HEAD(&rpc_list);
2850 spin_lock(&exp->exp_ldlm_data.led_lock);
2851 list_for_each_entry_safe(lock, next, locklist, l_export_chain) {
2853 lock_res_and_lock(lock);
2854 if (lock->l_req_mode != lock->l_granted_mode) {
2855 unlock_res_and_lock(lock);
2859 LASSERT(lock->l_resource);
2860 if (lock->l_resource->lr_type != LDLM_IBITS &&
2861 lock->l_resource->lr_type != LDLM_PLAIN) {
2862 unlock_res_and_lock(lock);
2866 if (lock->l_flags & LDLM_FL_AST_SENT) {
2867 unlock_res_and_lock(lock);
2871 LASSERT(lock->l_blocking_ast);
2872 LASSERT(!lock->l_blocking_lock);
2874 lock->l_flags |= LDLM_FL_AST_SENT;
2875 unlock_res_and_lock(lock);
2877 list_move(&lock->l_export_chain, &rpc_list);
2879 spin_unlock(&exp->exp_ldlm_data.led_lock);
2881 while (!list_empty(&rpc_list)) {
2882 lock = list_entry(rpc_list.next, struct ldlm_lock,
2884 list_del_init(&lock->l_export_chain);
2886 /* the desc just pretend to exclusive */
2887 ldlm_lock2desc(lock, &desc);
2888 desc.l_req_mode = LCK_EX;
2889 desc.l_granted_mode = 0;
2891 lock->l_blocking_ast(lock, &desc, NULL, LDLM_CB_BLOCKING);
2896 static int mds_msg_check_version(struct lustre_msg *msg)
2902 case MDS_DISCONNECT:
2904 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2906 CERROR("bad opc %u version %08x, expecting %08x\n",
2907 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2912 case MDS_GETATTR_LOCK:
2913 case MDS_ACCESS_CHECK:
2917 case MDS_DONE_WRITING:
2921 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
2923 CERROR("bad opc %u version %08x, expecting %08x\n",
2924 msg->opc, msg->version, LUSTRE_MDS_VERSION);
2928 case LDLM_BL_CALLBACK:
2929 case LDLM_CP_CALLBACK:
2930 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
2932 CERROR("bad opc %u version %08x, expecting %08x\n",
2933 msg->opc, msg->version, LUSTRE_DLM_VERSION);
2935 case OBD_LOG_CANCEL:
2936 case LLOG_ORIGIN_HANDLE_OPEN:
2937 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2938 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2939 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2940 case LLOG_ORIGIN_HANDLE_CLOSE:
2942 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
2944 CERROR("bad opc %u version %08x, expecting %08x\n",
2945 msg->opc, msg->version, LUSTRE_LOG_VERSION);
2951 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
2953 CERROR("bad opc %u version %08x, expecting %08x\n",
2954 msg->opc, msg->version, LUSTRE_OBD_VERSION);
2957 case SEC_INIT_CONTINUE:
2962 CERROR("MDS unknown opcode %d\n", msg->opc);
2970 int mds_handle(struct ptlrpc_request *req)
2972 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
2973 struct obd_device *obd = NULL;
2974 struct mds_obd *mds = NULL; /* quell gcc overwarning */
2978 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
2980 rc = mds_msg_check_version(req->rq_reqmsg);
2982 CERROR("MDS drop mal-formed request\n");
2986 /* Security opc should NOT trigger any recovery events */
2987 if (req->rq_reqmsg->opc == SEC_INIT ||
2988 req->rq_reqmsg->opc == SEC_INIT_CONTINUE) {
2989 if (req->rq_export) {
2990 mds_req_add_idmapping(req,
2991 &req->rq_export->exp_mds_data);
2992 mds_revoke_export_locks(req->rq_export);
2995 } else if (req->rq_reqmsg->opc == SEC_FINI) {
2996 if (req->rq_export) {
2997 mds_req_del_idmapping(req,
2998 &req->rq_export->exp_mds_data);
2999 mds_revoke_export_locks(req->rq_export);
3004 LASSERT(current->journal_info == NULL);
3005 /* XXX identical to OST */
3006 if (req->rq_reqmsg->opc != MDS_CONNECT) {
3007 struct mds_export_data *med;
3010 if (req->rq_export == NULL) {
3011 CERROR("operation %d on unconnected MDS from %s\n",
3012 req->rq_reqmsg->opc,
3014 req->rq_status = -ENOTCONN;
3015 GOTO(out, rc = -ENOTCONN);
3018 med = &req->rq_export->exp_mds_data;
3019 obd = req->rq_export->exp_obd;
3022 /* sanity check: if the xid matches, the request must
3023 * be marked as a resent or replayed */
3024 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
3025 req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid)) {
3026 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
3027 (MSG_RESENT | MSG_REPLAY),
3028 "rq_xid "LPU64" matches last_xid, "
3029 "expected RESENT flag\n",
3032 /* else: note the opposite is not always true; a
3033 * RESENT req after a failover will usually not match
3034 * the last_xid, since it was likely never
3035 * committed. A REPLAYed request will almost never
3036 * match the last xid, however it could for a
3037 * committed, but still retained, open. */
3039 spin_lock_bh(&obd->obd_processing_task_lock);
3040 recovering = obd->obd_recovering;
3041 spin_unlock_bh(&obd->obd_processing_task_lock);
3043 rc = mds_filter_recovery_request(req, obd,
3045 if (rc || should_process == 0) {
3047 } else if (should_process < 0) {
3048 req->rq_status = should_process;
3049 rc = ptlrpc_error(req);
3055 switch (req->rq_reqmsg->opc) {
3057 DEBUG_REQ(D_INODE, req, "connect");
3058 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
3059 rc = target_handle_connect(req);
3061 struct mds_export_data *med;
3063 LASSERT(req->rq_export);
3064 med = &req->rq_export->u.eu_mds_data;
3065 mds_init_export_data(req, med);
3066 mds_req_add_idmapping(req, med);
3068 /* Now that we have an export, set mds. */
3069 obd = req->rq_export->exp_obd;
3070 mds = mds_req2mds(req);
3074 case MDS_DISCONNECT:
3075 DEBUG_REQ(D_INODE, req, "disconnect");
3076 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
3077 rc = target_handle_disconnect(req);
3078 req->rq_status = rc; /* superfluous? */
3082 DEBUG_REQ(D_INODE, req, "getstatus");
3083 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
3084 rc = mds_getstatus(req);
3088 DEBUG_REQ(D_INODE, req, "getattr");
3089 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
3090 rc = mds_getattr(req, MDS_REQ_REC_OFF);
3093 case MDS_ACCESS_CHECK:
3094 DEBUG_REQ(D_INODE, req, "access_check");
3095 OBD_FAIL_RETURN(OBD_FAIL_MDS_ACCESS_CHECK_NET, 0);
3096 rc = mds_access_check(req, MDS_REQ_REC_OFF);
3099 case MDS_GETATTR_LOCK: {
3100 struct lustre_handle lockh;
3101 DEBUG_REQ(D_INODE, req, "getattr_lock");
3102 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_LOCK_NET, 0);
3104 /* If this request gets a reconstructed reply, we won't be
3105 * acquiring any new locks in mds_getattr_lock, so we don't
3109 rc = mds_getattr_lock(req, MDS_REQ_REC_OFF, &lockh,
3110 MDS_INODELOCK_UPDATE);
3111 /* this non-intent call (from an ioctl) is special */
3112 req->rq_status = rc;
3113 if (rc == 0 && lockh.cookie)
3114 ldlm_lock_decref(&lockh, LCK_PR);
3118 DEBUG_REQ(D_INODE, req, "statfs");
3119 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
3120 rc = mds_statfs(req);
3124 DEBUG_REQ(D_INODE, req, "readpage");
3125 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
3126 rc = mds_readpage(req, MDS_REQ_REC_OFF);
3128 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
3129 if (req->rq_reply_state) {
3130 lustre_free_reply_state (req->rq_reply_state);
3131 req->rq_reply_state = NULL;
3138 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
3141 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
3142 mds->mds_max_cookiesize};
3145 /* NB only peek inside req now; mds_reint() will swab it */
3147 CERROR ("Can't inspect opcode\n");
3152 if (lustre_msg_swabbed (req->rq_reqmsg))
3155 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
3156 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
3157 reint_names[opc] == NULL) ? reint_names[opc] :
3160 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
3162 if (opc == REINT_UNLINK || opc == REINT_RENAME)
3164 else if (opc == REINT_OPEN)
3169 /* for SETATTR: I have different reply setting for
3170 * remote setfacl, so delay the reply buffer allocation.
3172 if (opc != REINT_SETATTR) {
3173 rc = lustre_pack_reply(req, bufcount, size, NULL);
3178 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
3179 fail = OBD_FAIL_MDS_REINT_NET_REP;
3184 DEBUG_REQ(D_INODE, req, "close");
3185 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
3186 rc = mds_close(req, MDS_REQ_REC_OFF);
3189 case MDS_DONE_WRITING:
3190 DEBUG_REQ(D_INODE, req, "done_writing");
3191 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
3192 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
3196 DEBUG_REQ(D_INODE, req, "pin");
3197 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
3198 rc = mds_pin(req, MDS_REQ_REC_OFF);
3202 DEBUG_REQ(D_INODE, req, "sync");
3203 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
3204 rc = mds_sync(req, MDS_REQ_REC_OFF);
3207 DEBUG_REQ(D_INODE, req, "parseid");
3208 rc = mds_parse_id(req);
3211 DEBUG_REQ(D_INODE, req, "ping");
3212 rc = target_handle_ping(req);
3215 case OBD_LOG_CANCEL:
3216 CDEBUG(D_INODE, "log cancel\n");
3217 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
3218 rc = -ENOTSUPP; /* la la la */
3222 DEBUG_REQ(D_INODE, req, "enqueue");
3223 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
3224 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
3225 ldlm_server_blocking_ast, NULL);
3226 fail = OBD_FAIL_LDLM_REPLY;
3229 DEBUG_REQ(D_INODE, req, "convert");
3230 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
3231 rc = ldlm_handle_convert(req);
3233 case LDLM_BL_CALLBACK:
3234 case LDLM_CP_CALLBACK:
3235 DEBUG_REQ(D_INODE, req, "callback");
3236 CERROR("callbacks should not happen on MDS\n");
3238 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
3240 case LLOG_ORIGIN_HANDLE_OPEN:
3241 DEBUG_REQ(D_INODE, req, "llog_init");
3242 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3243 rc = llog_origin_handle_open(req);
3245 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
3246 DEBUG_REQ(D_INODE, req, "llog next block");
3247 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3248 rc = llog_origin_handle_next_block(req);
3250 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
3251 DEBUG_REQ(D_INODE, req, "llog prev block");
3252 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3253 rc = llog_origin_handle_prev_block(req);
3255 case LLOG_ORIGIN_HANDLE_READ_HEADER:
3256 DEBUG_REQ(D_INODE, req, "llog read header");
3257 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3258 rc = llog_origin_handle_read_header(req);
3260 case LLOG_ORIGIN_HANDLE_CLOSE:
3261 DEBUG_REQ(D_INODE, req, "llog close");
3262 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3263 rc = llog_origin_handle_close(req);
3266 DEBUG_REQ(D_INODE, req, "ost_create");
3267 rc = mdt_obj_create(req);
3270 DEBUG_REQ(D_INODE, req, "get_info");
3271 rc = mdt_get_info(req);
3274 DEBUG_REQ(D_INODE, req, "set_info");
3275 rc = mdt_set_info(req);
3278 CDEBUG(D_INODE, "write\n");
3279 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
3280 rc = ost_brw_write(req, NULL);
3281 LASSERT(current->journal_info == NULL);
3282 /* mdt_brw sends its own replies */
3286 DEBUG_REQ(D_INODE, req, "llog catinfo");
3287 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
3288 rc = llog_catinfo(req);
3291 req->rq_status = -ENOTSUPP;
3292 rc = ptlrpc_error(req);
3296 LASSERT(current->journal_info == NULL);
3298 /* If we're DISCONNECTing, the mds_export_data is already freed */
3299 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
3300 struct mds_export_data *med = &req->rq_export->exp_mds_data;
3301 struct obd_device *obd = list_entry(mds, struct obd_device,
3303 req->rq_repmsg->last_xid =
3304 le64_to_cpu(med->med_mcd->mcd_last_xid);
3306 if (!obd->obd_no_transno) {
3307 req->rq_repmsg->last_committed =
3308 obd->obd_last_committed;
3310 DEBUG_REQ(D_IOCTL, req,
3311 "not sending last_committed update");
3313 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
3315 mds->mds_last_transno, obd->obd_last_committed,
3321 target_send_reply(req, rc, fail);
3325 /* Update the server data on disk. This stores the new mount_count and also the
3326 * last_rcvd value to disk. If we don't have a clean shutdown, then the server
3327 * last_rcvd value may be less than that of the clients. This will alert us
3328 * that we may need to do client recovery.
3330 * Also assumes for mds_last_transno that we are not modifying it (no locking).
3332 int mds_update_server_data(struct obd_device *obd, int force_sync)
3334 struct mds_obd *mds = &obd->u.mds;
3335 struct mds_server_data *msd = mds->mds_server_data;
3336 struct file *filp = mds->mds_rcvd_filp;
3337 struct lvfs_run_ctxt saved;
3342 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3343 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
3345 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
3346 mds->mds_mount_count, mds->mds_last_transno);
3347 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off, force_sync);
3349 CERROR("error writing MDS server data: rc = %d\n", rc);
3350 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3355 /* saves last allocated fid counter to file. */
3356 int mds_update_last_fid(struct obd_device *obd, void *handle,
3359 struct mds_obd *mds = &obd->u.mds;
3360 struct file *filp = mds->mds_fid_filp;
3361 struct lvfs_run_ctxt saved;
3367 spin_lock(&mds->mds_last_fid_lock);
3368 last_fid = mds->mds_last_fid;
3369 spin_unlock(&mds->mds_last_fid_lock);
3371 CDEBUG(D_SUPER, "MDS last_fid is #"LPU64"\n",
3375 fsfilt_add_journal_cb(obd, mds->mds_sb, last_fid,
3376 handle, mds_commit_last_fid_cb,
3380 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3381 rc = fsfilt_write_record(obd, filp, &last_fid, sizeof(last_fid),
3383 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3386 CERROR("error writing MDS last_fid #"LPU64
3387 ", err = %d\n", last_fid, rc);
3391 CDEBUG(D_SUPER, "wrote fid #"LPU64" at idx "
3392 "%llu: err = %d\n", last_fid, off, rc);
3397 void mds_set_last_fid(struct obd_device *obd, __u64 fid)
3399 struct mds_obd *mds = &obd->u.mds;
3401 spin_lock(&mds->mds_last_fid_lock);
3402 if (fid > mds->mds_last_fid)
3403 mds->mds_last_fid = fid;
3404 spin_unlock(&mds->mds_last_fid_lock);
3407 void mds_commit_last_transno_cb(struct obd_device *obd,
3408 __u64 transno, void *data,
3411 obd_transno_commit_cb(obd, transno, error);
3414 void mds_commit_last_fid_cb(struct obd_device *obd,
3415 __u64 fid, void *data,
3419 CERROR("%s: fid "LPD64" commit error: %d\n",
3420 obd->obd_name, fid, error);
3424 CDEBUG(D_HA, "%s: fid "LPD64" committed\n",
3425 obd->obd_name, fid);
3428 __u64 mds_alloc_fid(struct obd_device *obd)
3430 struct mds_obd *mds = &obd->u.mds;
3433 spin_lock(&mds->mds_last_fid_lock);
3434 fid = ++mds->mds_last_fid;
3435 spin_unlock(&mds->mds_last_fid_lock);
3441 * reads inode self id from inode EA. Probably later this should be replaced by
3442 * caching inode self id to avoid raeding it every time it is needed.
3444 int mds_read_inode_sid(struct obd_device *obd, struct inode *inode,
3445 struct lustre_id *id)
3450 LASSERT(id != NULL);
3451 LASSERT(obd != NULL);
3452 LASSERT(inode != NULL);
3454 rc = fsfilt_get_md(obd, inode, &id->li_fid,
3455 sizeof(id->li_fid), EA_SID);
3457 CERROR("fsfilt_get_md() failed, "
3466 id_ino(id) = inode->i_ino;
3467 id_gen(id) = inode->i_generation;
3468 id_type(id) = S_IFMT & inode->i_mode;
3473 int mds_read_inode_pid(struct obd_device *obd, struct inode *inode,
3474 struct lustre_id *id)
3479 LASSERT(inode && id);
3481 rc = fsfilt_get_md(obd, inode, id, sizeof(*id), EA_PID);
3483 CERROR("get parent id from EA failed, rc=%d\n", rc);
3492 /* updates inode self id in EA. */
3493 int mds_update_inode_ids(struct obd_device *obd, struct inode *inode,
3494 void *handle, struct lustre_id *id,
3495 struct lustre_id *pid)
3501 LASSERT(id == NULL || id_fid(id) != 0);
3502 LASSERT(pid == NULL || id_fid(pid) != 0);
3503 LASSERT(obd != NULL);
3504 LASSERT(inode != NULL);
3507 mds_set_last_fid(obd, id_fid(id));
3508 rc = fsfilt_set_md(obd, inode, handle, &id->li_fid,
3509 sizeof(id->li_fid), EA_SID);
3510 LASSERTF(rc == 0, "failed to update fid: %d\n", rc);
3513 rc = fsfilt_set_md(obd, inode, handle, pid,
3514 sizeof(*pid), EA_PID);
3515 LASSERTF(rc == 0, "failed to update parent fid: %d\n", rc);
3521 /* mount the file system (secretly) */
3522 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
3524 struct lustre_cfg* lcfg = buf;
3525 struct mds_obd *mds = &obd->u.mds;
3526 struct lvfs_obd_ctxt *lvfs_ctxt = NULL;
3527 char *options = NULL;
3528 struct vfsmount *mnt;
3531 struct crypto_tfm *tfm = NULL;
3535 if (lcfg->lcfg_bufcount < 3)
3536 RETURN(rc = -EINVAL);
3538 if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
3539 RETURN(rc = -EINVAL);
3541 obd->obd_fsops = fsfilt_get_ops(lustre_cfg_string(lcfg, 2));
3542 if (IS_ERR(obd->obd_fsops))
3543 RETURN(rc = PTR_ERR(obd->obd_fsops));
3545 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
3547 page = get_zeroed_page(GFP_KERNEL);
3551 options = (char *)page;
3554 * here we use "iopen_nopriv" hardcoded, because it affects MDS utility
3555 * and the rest of options are passed by mount options. Probably this
3556 * should be moved to somewhere else like startup scripts or lconf. */
3557 sprintf(options, "iopen_nopriv");
3559 if (LUSTRE_CFG_BUFLEN(lcfg, 4) > 0 && lustre_cfg_buf(lcfg, 4))
3560 sprintf(options + strlen(options), ",%s",
3561 lustre_cfg_string(lcfg, 4));
3563 /* we have to know mdsnum before touching underlying fs -bzzz */
3564 atomic_set(&mds->mds_open_count, 0);
3565 sema_init(&mds->mds_md_sem, 1);
3566 mds->mds_md_connected = 0;
3567 mds->mds_md_name = NULL;
3569 if (LUSTRE_CFG_BUFLEN(lcfg, 5) > 0 && lustre_cfg_buf(lcfg, 5) &&
3570 strncmp(lustre_cfg_string(lcfg, 5), "dumb", LUSTRE_CFG_BUFLEN(lcfg, 5))) {
3573 generate_random_uuid(uuid);
3574 class_uuid_unparse(uuid, &mds->mds_md_uuid);
3576 OBD_ALLOC(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3577 if (mds->mds_md_name == NULL)
3578 RETURN(rc = -ENOMEM);
3580 memcpy(mds->mds_md_name, lustre_cfg_buf(lcfg, 5),
3581 LUSTRE_CFG_BUFLEN(lcfg, 5));
3583 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
3584 obd->obd_name, mds->mds_md_name);
3586 rc = mds_md_connect(obd, mds->mds_md_name);
3588 OBD_FREE(mds->mds_md_name, LUSTRE_CFG_BUFLEN(lcfg, 5));
3593 mds->mds_obd_type = MDS_MASTER_OBD;
3595 if (LUSTRE_CFG_BUFLEN(lcfg, 6) > 0 && lustre_cfg_buf(lcfg, 6) &&
3596 strncmp(lustre_cfg_string(lcfg, 6), "dumb",
3597 LUSTRE_CFG_BUFLEN(lcfg, 6))) {
3598 if (!memcmp(lustre_cfg_string(lcfg, 6), "master",
3599 strlen("master"))) {
3600 mds->mds_obd_type = MDS_MASTER_OBD;
3601 } else if (!memcmp(lustre_cfg_string(lcfg, 6), "cache",
3603 mds->mds_obd_type = MDS_CACHE_OBD;
3607 rc = lvfs_mount_fs(lustre_cfg_string(lcfg, 1),
3608 lustre_cfg_string(lcfg, 2),
3609 options, 0, &lvfs_ctxt);
3613 if (rc || !lvfs_ctxt) {
3614 CERROR("lvfs_mount_fs failed: rc = %d\n", rc);
3618 mnt = lvfs_ctxt->loc_mnt;
3619 mds->mds_lvfs_ctxt = lvfs_ctxt;
3620 ll_clear_rdonly(ll_sbdev(mnt->mnt_sb));
3622 CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
3624 mds->mds_fidext_thumb = 0;
3625 sema_init(&mds->mds_epoch_sem, 1);
3626 atomic_set(&mds->mds_real_clients, 0);
3627 spin_lock_init(&mds->mds_fidext_lock);
3628 spin_lock_init(&mds->mds_fidmap_lock);
3629 spin_lock_init(&mds->mds_transno_lock);
3630 spin_lock_init(&mds->mds_last_fid_lock);
3631 sema_init(&mds->mds_orphan_recovery_sem, 1);
3632 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
3634 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
3635 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER);
3637 if (obd->obd_namespace == NULL) {
3638 mds_cleanup(obd, 0);
3639 GOTO(err_put, rc = -ENOMEM);
3641 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
3643 tfm = crypto_alloc_tfm(CAPA_HMAC_ALG, 0);
3645 GOTO(err_ns, rc = -ENOSYS);
3647 mds->mds_capa_hmac = tfm;
3648 mds->mds_capa_timeout = CAPA_TIMEOUT;
3649 mds->mds_capa_key_timeout = CAPA_KEY_TIMEOUT;
3651 rc = mds_fs_setup(obd, mnt);
3653 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
3658 rc = llog_start_commit_thread();
3663 if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 && lustre_cfg_buf(lcfg, 3) &&
3664 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3665 LUSTRE_CFG_BUFLEN(lcfg, 3))) {
3668 generate_random_uuid(uuid);
3669 class_uuid_unparse(uuid, &mds->mds_dt_uuid);
3671 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
3672 if (mds->mds_profile == NULL)
3673 GOTO(err_fs, rc = -ENOMEM);
3675 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
3676 LUSTRE_CFG_BUFLEN(lcfg, 3));
3680 * setup root dir and files ID dir if lmv already connected, or there is
3683 if (mds->mds_md_exp || (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0 &&
3684 lustre_cfg_buf(lcfg, 3) &&
3685 strncmp(lustre_cfg_string(lcfg, 3), "dumb",
3686 LUSTRE_CFG_BUFLEN(lcfg, 3)))) {
3687 rc = mds_fs_setup_rootid(obd);
3691 rc = mds_fs_setup_virtid(obd);
3697 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
3698 "mds_ldlm_client", &obd->obd_ldlm_client);
3699 obd->obd_replayable = 1;
3701 mds->mds_crypto_type = NO_CRYPTO;
3703 rc = mds_postsetup(obd);
3710 /* No extra cleanup needed for llog_init_commit_thread() */
3711 mds_fs_cleanup(obd, 0);
3713 crypto_free_tfm(mds->mds_capa_hmac);
3715 ldlm_namespace_free(obd->obd_namespace, 0);
3716 obd->obd_namespace = NULL;
3719 lvfs_umount_fs(mds->mds_lvfs_ctxt);
3723 fsfilt_put_ops(obd->obd_fsops);
3727 static int mds_fs_post_setup(struct obd_device *obd)
3729 struct mds_obd *mds = &obd->u.mds;
3730 struct dentry *dentry;
3734 dentry = mds_id2dentry(obd, &mds->mds_rootid, NULL);
3735 if (IS_ERR(dentry)) {
3736 CERROR("Can't find ROOT, err = %d\n",
3737 (int)PTR_ERR(dentry));
3738 RETURN(PTR_ERR(dentry));
3740 rc = fsfilt_post_setup(obd, dentry);
3741 //set id2name function handler
3742 fsfilt_set_info(obd, mds->mds_sb, NULL, 7, "id2name",
3743 sizeof(mds_audit_id2name), mds_audit_id2name);
3749 static int mds_postsetup(struct obd_device *obd)
3751 struct mds_obd *mds = &obd->u.mds;
3755 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
3756 obd, 0, NULL, &llog_lvfs_ops);
3760 if (mds->mds_profile) {
3761 struct llog_ctxt *lgctxt;
3762 struct lvfs_run_ctxt saved;
3763 struct lustre_profile *lprof;
3764 struct config_llog_instance cfg;
3766 cfg.cfg_instance = NULL;
3767 cfg.cfg_uuid = mds->mds_dt_uuid;
3768 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3770 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
3772 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3773 GOTO(err_llog, rc = -EINVAL);
3776 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
3777 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3782 lprof = class_get_profile(mds->mds_profile);
3783 if (lprof == NULL) {
3784 CERROR("No profile found: %s\n", mds->mds_profile);
3785 GOTO(err_cleanup, rc = -ENOENT);
3787 rc = mds_dt_connect(obd, lprof->lp_lov);
3789 GOTO(err_cleanup, rc);
3791 rc = mds_md_postsetup(obd);
3793 GOTO(err_cleanup, rc);
3795 rc = mds_fs_post_setup(obd);
3797 CERROR("can not post setup fsfilt\n");
3802 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
3803 LLOG_CONFIG_ORIG_CTXT));
3807 int mds_postrecov_common(struct obd_device *obd)
3809 struct mds_obd *mds = &obd->u.mds;
3810 struct llog_ctxt *ctxt;
3811 int rc, item = 0, valsize;
3812 struct timeval tstart, now;
3816 CDEBUG(D_ERROR, "%s: post-recovery\n", obd->obd_name);
3817 LASSERT(!obd->obd_recovering);
3818 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
3819 LASSERT(ctxt != NULL);
3821 /* clean PENDING dir */
3822 do_gettimeofday(&tstart);
3823 rc = mds_cleanup_orphans(obd);
3824 do_gettimeofday(&now);
3825 CDEBUG(D_ERROR, "%s: cleaned %d MDS orphans in %us\n",
3826 obd->obd_name, rc, (unsigned) (now.tv_sec - tstart.tv_sec));
3831 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
3832 valsize = sizeof(group);
3833 rc = obd_set_info(mds->mds_dt_exp, strlen("mds_conn"),
3834 "mds_conn", valsize, &group);
3836 CERROR("%s: failed to obd_set_info(): %d\n",
3841 rc = llog_connect(ctxt, obd->u.mds.mds_dt_desc.ld_tgt_count,
3844 CERROR("%s: failed at llog_origin_connect: %d\n",
3849 /* remove the orphaned precreated objects */
3850 do_gettimeofday(&tstart);
3851 rc = mds_dt_clear_orphans(mds, NULL /* all OSTs */);
3852 do_gettimeofday(&now);
3853 CDEBUG(D_ERROR, "%s: cleaned OSS orphans in %us: %d\n",
3854 obd->obd_name, (unsigned) (now.tv_sec - tstart.tv_sec), rc);
3856 CERROR("%s: can't clear OSS orphans: %d\n",
3862 RETURN(rc < 0 ? rc : item);
3865 /* cleanup all llogging subsystems */
3866 rc = obd_llog_finish(obd, &obd->obd_llogs,
3867 mds->mds_dt_desc.ld_tgt_count);
3869 CERROR("%s: failed to cleanup llogging subsystems\n",
3874 int mds_postrecov(struct obd_device *obd)
3878 rc = mds_postrecov_common(obd);
3880 rc = mds_md_reconnect(obd);
3884 int mds_dt_clean(struct obd_device *obd)
3886 struct mds_obd *mds = &obd->u.mds;
3889 if (mds->mds_profile) {
3891 struct llog_ctxt *llctx;
3892 struct lvfs_run_ctxt saved;
3893 struct config_llog_instance cfg;
3894 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
3896 OBD_ALLOC(cln_prof, len);
3898 CERROR("can't allocate memory, processing cleanup "
3899 "profile is skipped\n");
3903 sprintf(cln_prof, "%s-clean", mds->mds_profile);
3905 cfg.cfg_instance = NULL;
3906 cfg.cfg_uuid = mds->mds_dt_uuid;
3908 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3909 llctx = llog_get_context(&obd->obd_llogs,
3910 LLOG_CONFIG_ORIG_CTXT);
3911 class_config_process_llog(llctx, cln_prof, &cfg);
3912 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
3914 OBD_FREE(cln_prof, len);
3916 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
3917 mds->mds_profile = NULL;
3922 int mds_md_clean(struct obd_device *obd)
3924 struct mds_obd *mds = &obd->u.mds;
3927 if (mds->mds_md_name) {
3928 OBD_FREE(mds->mds_md_name, strlen(mds->mds_md_name) + 1);
3929 mds->mds_md_name = NULL;
3934 static int mds_precleanup(struct obd_device *obd, int flags)
3940 mds_dt_disconnect(obd, flags);
3942 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
3946 extern void lgss_svc_cache_purge_all(void);
3947 static int mds_cleanup(struct obd_device *obd, int flags)
3949 struct mds_obd *mds = &obd->u.mds;
3952 if (mds->mds_sb == NULL)
3955 mds_update_server_data(obd, 1);
3956 mds_update_last_fid(obd, NULL, 1);
3958 if (mds->mds_dt_objids != NULL) {
3959 int size = mds->mds_dt_desc.ld_tgt_count *
3961 OBD_FREE(mds->mds_dt_objids, size);
3963 mds_fs_cleanup(obd, flags);
3967 /* 2 seems normal on mds, (may_umount() also expects 2
3968 fwiw), but we only see 1 at this point in obdfilter. */
3969 lvfs_umount_fs(mds->mds_lvfs_ctxt);
3973 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
3975 spin_lock_bh(&obd->obd_processing_task_lock);
3976 if (obd->obd_recovering) {
3977 target_cancel_recovery_timer(obd);
3978 obd->obd_recovering = 0;
3980 spin_unlock_bh(&obd->obd_processing_task_lock);
3983 fsfilt_put_ops(obd->obd_fsops);
3987 lgss_svc_cache_purge_all();
3990 spin_lock(&mds->mds_denylist_lock);
3991 while (!list_empty( &mds->mds_denylist ) ) {
3992 deny_sec_t *p_deny_sec = list_entry(mds->mds_denylist.next,
3994 list_del(&p_deny_sec->list);
3995 OBD_FREE(p_deny_sec, sizeof(*p_deny_sec));
3997 spin_unlock(&mds->mds_denylist_lock);
3999 mds_capa_keys_cleanup(obd);
4001 if (mds->mds_capa_hmac)
4002 crypto_free_tfm(mds->mds_capa_hmac);
4006 static int set_security(const char *value, char **sec)
4008 if (!strcmp(value, "null"))
4010 else if (!strcmp(value, "krb5i"))
4012 else if (!strcmp(value, "krb5p"))
4015 CERROR("Unrecognized security flavor %s\n", value);
4022 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
4024 struct lustre_cfg *lcfg = buf;
4025 struct mds_obd *mds = &obd->u.mds;
4029 switch(lcfg->lcfg_command) {
4030 case LCFG_SET_SECURITY: {
4031 if ((LUSTRE_CFG_BUFLEN(lcfg, 1) == 0) ||
4032 (LUSTRE_CFG_BUFLEN(lcfg, 2) == 0))
4033 GOTO(out, rc = -EINVAL);
4035 if (!strcmp(lustre_cfg_string(lcfg, 1), "mds_sec"))
4036 rc = set_security(lustre_cfg_string(lcfg, 2),
4038 else if (!strcmp(lustre_cfg_string(lcfg, 1), "oss_sec"))
4039 rc = set_security(lustre_cfg_string(lcfg, 2),
4041 else if (!strcmp(lustre_cfg_string(lcfg, 1), "deny_sec")){
4042 spin_lock(&mds->mds_denylist_lock);
4043 rc = add_deny_security(lustre_cfg_string(lcfg, 2),
4044 &mds->mds_denylist);
4045 spin_unlock(&mds->mds_denylist_lock);
4047 CERROR("Unrecognized key\n");
4053 CERROR("Unknown command: %d\n", lcfg->lcfg_command);
4054 GOTO(out, rc = -EINVAL);
4060 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
4062 struct ldlm_lock *new_lock,
4063 struct ldlm_lock **old_lock,
4064 struct lustre_handle *lockh)
4066 struct obd_export *exp = req->rq_export;
4067 struct obd_device *obd = exp->exp_obd;
4068 struct ldlm_request *dlmreq =
4069 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
4070 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
4071 struct list_head *iter;
4073 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
4076 spin_lock(&obd->obd_namespace->ns_hash_lock);
4077 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
4078 struct ldlm_lock *lock;
4079 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
4080 if (lock == new_lock)
4082 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
4083 lockh->cookie = lock->l_handle.h_cookie;
4084 LDLM_DEBUG(lock, "restoring lock cookie");
4085 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
4088 *old_lock = LDLM_LOCK_GET(lock);
4089 spin_unlock(&obd->obd_namespace->ns_hash_lock);
4093 spin_unlock(&obd->obd_namespace->ns_hash_lock);
4095 /* If the xid matches, then we know this is a resent request,
4096 * and allow it. (It's probably an OPEN, for which we don't
4099 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
4103 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
4106 /* This remote handle isn't enqueued, so we never received or
4107 * processed this request. Clear MSG_RESENT, because it can
4108 * be handled like any normal request now. */
4110 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
4112 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
4116 int intent_disposition(struct ldlm_reply *rep, int flag)
4120 return (rep->lock_policy_res1 & flag);
4123 void intent_set_disposition(struct ldlm_reply *rep, int flag)
4127 rep->lock_policy_res1 |= flag;
4130 static int mds_intent_prepare_reply_buffers(struct ptlrpc_request *req,
4131 struct ldlm_intent *it)
4133 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
4134 struct mds_export_data *med = &req->rq_export->u.eu_mds_data;
4135 int rc, reply_buffers;
4136 int repsize[8] = {sizeof(struct ldlm_reply),
4137 sizeof(struct mds_body),
4138 mds->mds_max_mdsize};
4142 if (it->opc & ( IT_OPEN | IT_GETATTR | IT_LOOKUP | IT_CHDIR )) {
4143 /*Ugly here, Actually, we should prepare the reply buffer
4144 *after we know whether these stuff exist or not, which should
4145 * be fixed in future, Now each item is in the fix position,
4146 * the sequence is lsm, acl, crypto ea, capa.*/
4147 repsize[reply_buffers++] = sizeof(int);
4148 if (med->med_remote)
4149 repsize[reply_buffers++] =
4150 sizeof(struct mds_remote_perm);
4152 repsize[reply_buffers++] =
4153 xattr_acl_size(LL_ACL_MAX_ENTRIES);
4155 repsize[reply_buffers++] = sizeof(int);
4156 repsize[reply_buffers++] = sizeof(struct crypto_key);
4158 /* XXX: if new buffer is to be added, capability reply
4159 * buffer should always been reserved. */
4160 if (it->opc & IT_OPEN)
4161 repsize[reply_buffers++] = sizeof(struct lustre_capa);
4164 rc = lustre_pack_reply(req, reply_buffers, repsize, NULL);
4169 static int mds_intent_policy(struct ldlm_namespace *ns,
4170 struct ldlm_lock **lockp, void *req_cookie,
4171 ldlm_mode_t mode, int flags, void *data)
4173 struct ptlrpc_request *req = req_cookie;
4174 struct ldlm_lock *lock = *lockp;
4175 struct ldlm_intent *it;
4176 struct ldlm_reply *rep;
4177 struct lustre_handle lockh[2] = {{0}, {0}};
4178 struct ldlm_lock *new_lock = NULL;
4179 int getattr_part = MDS_INODELOCK_UPDATE;
4182 int offset = MDS_REQ_INTENT_REC_OFF;
4185 LASSERT(req != NULL);
4186 MD_COUNTER_INCREMENT(req->rq_export->exp_obd, intent_lock);
4188 if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
4189 /* No intent was provided */
4190 int size = sizeof(struct ldlm_reply);
4191 rc = lustre_pack_reply(req, 1, &size, NULL);
4196 it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
4197 lustre_swab_ldlm_intent);
4199 CERROR("Intent missing\n");
4200 RETURN(req->rq_status = -EFAULT);
4203 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
4205 rc = mds_intent_prepare_reply_buffers(req, it);
4208 RETURN(req->rq_status = rc);
4210 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
4211 LASSERT(rep != NULL);
4213 intent_set_disposition(rep, DISP_IT_EXECD);
4215 /* execute policy */
4216 switch ((long)it->opc) {
4218 case IT_CREAT|IT_OPEN:
4219 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4221 /* XXX swab here to assert that an mds_open reint
4222 * packet is following */
4223 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4225 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
4227 if (rep->lock_policy_res2) {
4229 * mds_open() returns ENOLCK where it should return
4230 * zero, but it has no lock to return.
4232 if (rep->lock_policy_res2 == ENOLCK)
4233 rep->lock_policy_res2 = 0;
4235 RETURN(ELDLM_LOCK_ABORTED);
4239 * IT_OPEN may return lock on cross-node dentry that we want to
4240 * hold during attr retrival -bzzz
4242 if (lockh[0].cookie == 0)
4243 RETURN(ELDLM_LOCK_ABORTED);
4247 getattr_part = MDS_INODELOCK_LOOKUP;
4250 getattr_part |= MDS_INODELOCK_LOOKUP;
4252 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
4253 lock, &new_lock, lockh);
4254 rep->lock_policy_res2 = mds_getattr_lock(req, offset, lockh,
4256 /* FIXME: LDLM can set req->rq_status. MDS sets
4257 policy_res{1,2} with disposition and status.
4258 - replay: returns 0 & req->status is old status
4259 - otherwise: returns req->status */
4260 if (intent_disposition(rep, DISP_LOOKUP_NEG))
4261 rep->lock_policy_res2 = 0;
4262 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
4263 rep->lock_policy_res2)
4264 RETURN(ELDLM_LOCK_ABORTED);
4265 if (req->rq_status != 0) {
4267 rep->lock_policy_res2 = req->rq_status;
4268 RETURN(ELDLM_LOCK_ABORTED);
4272 rc = mds_lock_and_check_slave(offset, req, lockh);
4273 if ((rep->lock_policy_res2 = rc)) {
4275 rep->lock_policy_res2 = 0;
4276 RETURN(ELDLM_LOCK_ABORTED);
4280 CERROR("Unhandled intent "LPD64"\n", it->opc);
4284 /* By this point, whatever function we called above must have either
4285 * filled in 'lockh', been an intent replay, or returned an error. We
4286 * want to allow replayed RPCs to not get a lock, since we would just
4287 * drop it below anyways because lock replay is done separately by the
4288 * client afterwards. For regular RPCs we want to give the new lock to
4289 * the client instead of whatever lock it was about to get. */
4290 if (new_lock == NULL)
4291 new_lock = ldlm_handle2lock(&lockh[0]);
4292 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
4295 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
4296 it->opc, lockh[0].cookie);
4298 /* If we've already given this lock to a client once, then we should
4299 * have no readers or writers. Otherwise, we should have one reader
4300 * _or_ writer ref (which will be zeroed below) before returning the
4301 * lock to a client. */
4302 if (new_lock->l_export == req->rq_export) {
4303 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
4305 LASSERT(new_lock->l_export == NULL);
4306 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
4311 if (new_lock->l_export == req->rq_export) {
4312 /* Already gave this to the client, which means that we
4313 * reconstructed a reply. */
4314 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
4316 RETURN(ELDLM_LOCK_REPLACED);
4319 /* Fixup the lock to be given to the client */
4320 lock_res_and_lock(new_lock);
4321 new_lock->l_readers = 0;
4322 new_lock->l_writers = 0;
4324 new_lock->l_export = class_export_get(req->rq_export);
4326 spin_lock(&new_lock->l_export->exp_ldlm_data.led_lock);
4327 list_add(&new_lock->l_export_chain,
4328 &new_lock->l_export->exp_ldlm_data.led_held_locks);
4329 spin_unlock(&new_lock->l_export->exp_ldlm_data.led_lock);
4331 new_lock->l_blocking_ast = lock->l_blocking_ast;
4332 new_lock->l_completion_ast = lock->l_completion_ast;
4334 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
4335 sizeof(lock->l_remote_handle));
4337 new_lock->l_flags &= ~LDLM_FL_LOCAL;
4339 unlock_res_and_lock(new_lock);
4340 LDLM_LOCK_PUT(new_lock);
4342 RETURN(ELDLM_LOCK_REPLACED);
4345 int mds_attach(struct obd_device *dev, obd_count len, void *data)
4347 struct lprocfs_static_vars lvars;
4349 struct mds_obd *mds = &dev->u.mds;
4351 spin_lock_init(&mds->mds_denylist_lock);
4352 INIT_LIST_HEAD(&mds->mds_denylist);
4354 lprocfs_init_multi_vars(0, &lvars);
4356 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
4360 return lprocfs_alloc_md_stats(dev, 0);
4363 int mds_detach(struct obd_device *dev)
4365 lprocfs_free_md_stats(dev);
4366 return lprocfs_obd_detach(dev);
4369 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
4371 struct lprocfs_static_vars lvars;
4373 lprocfs_init_multi_vars(1, &lvars);
4374 return lprocfs_obd_attach(dev, lvars.obd_vars);
4377 int mdt_detach(struct obd_device *dev)
4379 return lprocfs_obd_detach(dev);
4382 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
4384 struct mds_obd *mds = &obd->u.mds;
4389 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4390 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
4391 MDS_SERVICE_WATCHDOG_TIMEOUT,
4392 mds_handle, "mds", obd->obd_proc_entry);
4394 if (!mds->mds_service) {
4395 CERROR("failed to start service\n");
4399 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
4402 GOTO(err_thread, rc);
4404 mds->mds_setattr_service =
4405 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4406 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
4407 MDS_SERVICE_WATCHDOG_TIMEOUT,
4408 mds_handle, "mds_setattr",
4409 obd->obd_proc_entry);
4410 if (!mds->mds_setattr_service) {
4411 CERROR("failed to start getattr service\n");
4412 GOTO(err_thread, rc = -ENOMEM);
4415 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
4416 MDT_NUM_THREADS, "ll_mdt_attr");
4418 GOTO(err_thread2, rc);
4420 mds->mds_readpage_service =
4421 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4422 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
4423 MDS_SERVICE_WATCHDOG_TIMEOUT,
4424 mds_handle, "mds_readpage",
4425 obd->obd_proc_entry);
4426 if (!mds->mds_readpage_service) {
4427 CERROR("failed to start readpage service\n");
4428 GOTO(err_thread2, rc = -ENOMEM);
4431 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
4432 MDT_NUM_THREADS, "ll_mdt_rdpg");
4435 GOTO(err_thread3, rc);
4437 mds->mds_close_service =
4438 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
4439 MDS_CLOSE_PORTAL, MDC_REPLY_PORTAL,
4440 MDS_SERVICE_WATCHDOG_TIMEOUT,
4441 mds_handle, "mds_close",
4442 obd->obd_proc_entry);
4443 if (!mds->mds_close_service) {
4444 CERROR("failed to start close service\n");
4445 GOTO(err_thread3, rc = -ENOMEM);
4448 rc = ptlrpc_start_n_threads(obd, mds->mds_close_service,
4449 MDT_NUM_THREADS, "ll_mdt_clos");
4452 GOTO(err_thread4, rc);
4456 ptlrpc_unregister_service(mds->mds_close_service);
4458 ptlrpc_unregister_service(mds->mds_readpage_service);
4460 ptlrpc_unregister_service(mds->mds_setattr_service);
4462 ptlrpc_unregister_service(mds->mds_service);
4466 static int mdt_cleanup(struct obd_device *obd, int flags)
4468 struct mds_obd *mds = &obd->u.mds;
4471 ptlrpc_stop_all_threads(mds->mds_close_service);
4472 ptlrpc_unregister_service(mds->mds_close_service);
4474 ptlrpc_stop_all_threads(mds->mds_readpage_service);
4475 ptlrpc_unregister_service(mds->mds_readpage_service);
4477 ptlrpc_stop_all_threads(mds->mds_setattr_service);
4478 ptlrpc_unregister_service(mds->mds_setattr_service);
4480 ptlrpc_stop_all_threads(mds->mds_service);
4481 ptlrpc_unregister_service(mds->mds_service);
4486 static struct dentry *mds_lvfs_id2dentry(__u64 ino, __u32 gen,
4487 __u64 gr, void *data)
4489 struct lustre_id id;
4490 struct obd_device *obd = data;
4494 return mds_id2dentry(obd, &id, NULL);
4497 static int mds_get_info(struct obd_export *exp, __u32 keylen,
4498 void *key, __u32 *valsize, void *val)
4500 struct obd_device *obd;
4501 struct mds_obd *mds;
4505 obd = class_exp2obd(exp);
4509 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
4510 exp->exp_handle.h_cookie);
4514 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
4515 /* get log_context handle. */
4516 struct llog_ctxt *ctxt;
4517 unsigned long *llh_handle = val;
4518 *valsize = sizeof(unsigned long);
4519 ctxt = llog_get_context(&obd->obd_llogs, LLOG_REINT_ORIG_CTXT);
4521 CERROR("Cannot get REINT llog context\n");
4524 *llh_handle = (unsigned long)ctxt;
4527 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
4528 /* get log_context handle. */
4529 unsigned long *sb = val;
4530 *valsize = sizeof(unsigned long);
4531 *sb = (unsigned long)obd->u.mds.mds_sb;
4535 if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
4536 __u32 *mdsize = val;
4537 *valsize = sizeof(*mdsize);
4538 *mdsize = mds->mds_max_mdsize;
4542 if (keylen >= strlen("mdsnum") && strcmp(key, "mdsnum") == 0) {
4543 __u32 *mdsnum = val;
4544 *valsize = sizeof(*mdsnum);
4545 *mdsnum = mds->mds_num;
4549 if (keylen >= strlen("rootid") && strcmp(key, "rootid") == 0) {
4550 struct lustre_id *rootid = val;
4551 *valsize = sizeof(*rootid);
4552 *rootid = mds->mds_rootid;
4556 if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) {
4557 struct lov_desc *desc = val;
4558 *valsize = sizeof(*desc);
4559 *desc = mds->mds_dt_desc;
4563 if (keylen >= strlen("getext") && strcmp(key, "getext") == 0) {
4564 struct fid_extent *ext = val;
4565 *valsize = sizeof(*ext);
4567 spin_lock(&mds->mds_fidext_lock);
4568 ext->fe_width = MDS_FIDEXT_SIZE;
4569 ext->fe_start = mds->mds_fidext_thumb + 1;
4570 mds->mds_fidext_thumb += MDS_FIDEXT_SIZE;
4571 spin_unlock(&mds->mds_fidext_lock);
4576 rc = fsfilt_get_info(obd, mds->mds_sb, NULL, keylen, key, valsize, val);
4578 CDEBUG(D_IOCTL, "invalid key\n");
4583 struct lvfs_callback_ops mds_lvfs_ops = {
4584 l_id2dentry: mds_lvfs_id2dentry,
4587 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
4588 int objcount, struct obd_ioobj *obj,
4589 int niocount, struct niobuf_remote *nb,
4590 struct niobuf_local *res,
4591 struct obd_trans_info *oti, struct lustre_capa *capa);
4593 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
4594 int objcount, struct obd_ioobj *obj, int niocount,
4595 struct niobuf_local *res, struct obd_trans_info *oti,
4598 /* use obd ops to offer management infrastructure */
4599 static struct obd_ops mds_obd_ops = {
4600 .o_owner = THIS_MODULE,
4601 .o_attach = mds_attach,
4602 .o_detach = mds_detach,
4603 .o_connect = mds_connect,
4604 .o_connect_post = mds_connect_post,
4605 .o_init_export = mds_init_export,
4606 .o_destroy_export = mds_destroy_export,
4607 .o_disconnect = mds_disconnect,
4608 .o_setup = mds_setup,
4609 .o_precleanup = mds_precleanup,
4610 .o_cleanup = mds_cleanup,
4611 .o_process_config = mds_process_config,
4612 .o_postrecov = mds_postrecov,
4613 .o_statfs = mds_obd_statfs,
4614 .o_iocontrol = mds_iocontrol,
4615 .o_create = mds_obd_create,
4616 .o_destroy = mds_obd_destroy,
4617 .o_llog_init = mds_llog_init,
4618 .o_llog_finish = mds_llog_finish,
4619 .o_notify = mds_notify,
4620 .o_get_info = mds_get_info,
4621 .o_set_info = mds_set_info,
4622 .o_preprw = mds_preprw,
4623 .o_commitrw = mds_commitrw,
4626 static struct obd_ops mdt_obd_ops = {
4627 .o_owner = THIS_MODULE,
4628 .o_attach = mdt_attach,
4629 .o_detach = mdt_detach,
4630 .o_setup = mdt_setup,
4631 .o_cleanup = mdt_cleanup,
4634 static int __init mds_init(void)
4636 struct lprocfs_static_vars lvars;
4640 mds_init_lsd_cache();
4641 mds_init_rmtacl_upcall_cache();
4643 lprocfs_init_multi_vars(0, &lvars);
4644 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
4645 OBD_MDS_DEVICENAME);
4646 lprocfs_init_multi_vars(1, &lvars);
4647 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
4648 OBD_MDT_DEVICENAME);
4650 rc = mds_capa_key_start_thread();
4652 class_unregister_type(OBD_MDT_DEVICENAME);
4653 class_unregister_type(OBD_MDS_DEVICENAME);
4654 mds_cleanup_lsd_cache();
4657 mds_eck_timer.function = mds_capa_key_timer_callback;
4658 mds_eck_timer.data = 0;
4659 init_timer(&mds_eck_timer);
4664 static void /*__exit*/ mds_exit(void)
4666 mds_capa_key_stop_thread();
4667 mds_cleanup_rmtacl_upcall_cache();
4668 mds_cleanup_lsd_cache();
4670 class_unregister_type(OBD_MDS_DEVICENAME);
4671 class_unregister_type(OBD_MDT_DEVICENAME);
4674 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
4675 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
4676 MODULE_LICENSE("GPL");
4678 module_init(mds_init);
4679 module_exit(mds_exit);