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/ext3_fs.h>
43 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
44 # include <linux/smp_lock.h>
45 # include <linux/buffer_head.h>
46 # include <linux/workqueue.h>
47 # include <linux/mount.h>
49 # include <linux/locks.h>
51 #include <linux/obd_lov.h>
52 #include <linux/obd_ost.h>
53 #include <linux/lustre_mds.h>
54 #include <linux/lustre_fsfilt.h>
55 #include <linux/lprocfs_status.h>
56 #include <linux/lustre_commit_confd.h>
59 #include <linux/lustre_smfs.h>
60 #include <linux/lustre_snap.h>
62 #include "mds_internal.h"
64 static int mds_intent_policy(struct ldlm_namespace *ns,
65 struct ldlm_lock **lockp, void *req_cookie,
66 ldlm_mode_t mode, int flags, void *data);
67 static int mds_postsetup(struct obd_device *obd);
68 static int mds_cleanup(struct obd_device *obd, int flags);
71 /* Assumes caller has already pushed into the kernel filesystem context */
72 static int mds_sendpage(struct ptlrpc_request *req, struct file *file,
73 loff_t offset, int count)
75 struct ptlrpc_bulk_desc *desc;
76 struct l_wait_info lwi;
78 int rc = 0, npages, i, tmpcount, tmpsize = 0;
81 LASSERT((offset & (PAGE_SIZE - 1)) == 0); /* I'm dubious about this */
83 npages = (count + PAGE_SIZE - 1) >> PAGE_SHIFT;
84 OBD_ALLOC(pages, sizeof(*pages) * npages);
86 GOTO(out, rc = -ENOMEM);
88 desc = ptlrpc_prep_bulk_exp(req, npages, BULK_PUT_SOURCE,
91 GOTO(out_free, rc = -ENOMEM);
93 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
94 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
96 pages[i] = alloc_pages(GFP_KERNEL, 0);
98 GOTO(cleanup_buf, rc = -ENOMEM);
100 ptlrpc_prep_bulk_page(desc, pages[i], 0, tmpsize);
103 for (i = 0, tmpcount = count; i < npages; i++, tmpcount -= tmpsize) {
104 tmpsize = tmpcount > PAGE_SIZE ? PAGE_SIZE : tmpcount;
105 CDEBUG(D_EXT2, "reading %u@%llu from dir %lu (size %llu)\n",
106 tmpsize, offset, file->f_dentry->d_inode->i_ino,
107 file->f_dentry->d_inode->i_size);
109 rc = fsfilt_readpage(req->rq_export->exp_obd, file,
110 kmap(pages[i]), tmpsize, &offset);
114 GOTO(cleanup_buf, rc = -EIO);
117 LASSERT(desc->bd_nob == count);
119 rc = ptlrpc_start_bulk_transfer(desc);
121 GOTO(cleanup_buf, rc);
123 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
124 CERROR("obd_fail_loc=%x, fail operation rc=%d\n",
125 OBD_FAIL_MDS_SENDPAGE, rc);
126 GOTO(abort_bulk, rc);
129 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
130 rc = l_wait_event(desc->bd_waitq, !ptlrpc_bulk_active(desc), &lwi);
131 LASSERT (rc == 0 || rc == -ETIMEDOUT);
134 if (desc->bd_success &&
135 desc->bd_nob_transferred == count)
136 GOTO(cleanup_buf, rc);
138 rc = -ETIMEDOUT; /* XXX should this be a different errno? */
141 DEBUG_REQ(D_ERROR, req, "bulk failed: %s %d(%d), evicting %s@%s\n",
142 (rc == -ETIMEDOUT) ? "timeout" : "network error",
143 desc->bd_nob_transferred, count,
144 req->rq_export->exp_client_uuid.uuid,
145 req->rq_export->exp_connection->c_remote_uuid.uuid);
147 ptlrpc_fail_export(req->rq_export);
151 ptlrpc_abort_bulk (desc);
153 for (i = 0; i < npages; i++)
155 __free_pages(pages[i], 0);
157 ptlrpc_free_bulk(desc);
159 OBD_FREE(pages, sizeof(*pages) * npages);
164 int mds_lock_mode_for_dir(struct obd_device *obd,
165 struct dentry *dentry, int mode)
169 /* any dir access needs couple locks:
170 * 1) on part of dir we gonna lookup/modify in
171 * 2) on a whole dir to protect it from concurrent splitting
172 * and to flush client's cache for readdir()
173 * so, for a given mode and dentry this routine decides what
174 * lock mode to use for lock #2:
175 * 1) if caller's gonna lookup in dir then we need to protect
176 * dir from being splitted only - LCK_CR
177 * 2) if caller's gonna modify dir then we need to protect
178 * dir from being splitted and to flush cache - LCK_CW
179 * 3) if caller's gonna modify dir and that dir seems ready
180 * for splitting then we need to protect it from any
181 * type of access (lookup/modify/split) - LCK_EX -bzzz */
183 split = mds_splitting_expected(obd, dentry);
185 /* it is important to check here only for MDS_NO_SPLITTABLE.
186 * The reason is that MDS_NO_SPLITTABLE means dir is not splittable
187 * in principle and another thread will not split it on the quiet.
188 * But if we have MDS_NO_SPLIT_EXPECTED, this means, that dir may be
189 * splitted anytime, but not now (forcurrent thread) and we should
190 * consider that it can happen soon and go that branch which can yield
191 * LCK_EX to protect from possible splitting. */
192 if (split == MDS_NO_SPLITTABLE) {
193 /* this inode won't be splitted. so we need not to protect from
194 * just flush client's cache on modification */
202 if (mode == LCK_PR) {
204 } else if (mode == LCK_PW) {
205 /* caller gonna modify directory.we use concurrent
206 write lock here to retract client's cache for readdir */
208 if (split == MDS_EXPECT_SPLIT) {
209 /* splitting possible. serialize any access
210 * the idea is that first one seen dir is
211 * splittable is given exclusive lock and
212 * split directory. caller passes lock mode
213 * to mds_try_to_split_dir() and splitting
214 * would be done with exclusive lock only -bzzz */
215 CDEBUG(D_OTHER, "%s: gonna split %u/%u\n",
217 (unsigned) dentry->d_inode->i_ino,
218 (unsigned) dentry->d_inode->i_generation);
226 /* only valid locked dentries or errors should be returned */
227 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
228 struct vfsmount **mnt, int lock_mode,
229 struct lustre_handle *lockh, int *mode,
230 char *name, int namelen, __u64 lockpart)
232 struct mds_obd *mds = &obd->u.mds;
233 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
234 struct ldlm_res_id res_id = { .name = {0} };
236 ldlm_policy_data_t policy = { .l_inodebits = { lockpart } };
243 res_id.name[0] = de->d_inode->i_ino;
244 res_id.name[1] = de->d_inode->i_generation;
247 if (name && IS_PDIROPS(de->d_inode)) {
248 ldlm_policy_data_t cpolicy =
249 { .l_inodebits = { MDS_INODELOCK_UPDATE } };
250 LASSERT(mode != NULL);
251 *mode = mds_lock_mode_for_dir(obd, de, lock_mode);
253 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace,
255 &cpolicy, *mode, &flags,
257 ldlm_completion_ast, NULL, NULL,
258 NULL, 0, NULL, lockh + 1);
259 if (rc != ELDLM_OK) {
261 RETURN(ERR_PTR(-ENOLCK));
266 res_id.name[2] = full_name_hash(name, namelen);
268 CDEBUG(D_INFO, "take lock on %lu:%u:"LPX64"\n",
269 de->d_inode->i_ino, de->d_inode->i_generation,
273 #warning "No PDIROPS support in the kernel"
275 rc = ldlm_cli_enqueue(NULL, NULL, obd->obd_namespace, res_id,
276 LDLM_IBITS, &policy, lock_mode, &flags,
277 mds_blocking_ast, ldlm_completion_ast, NULL, NULL,
278 NULL, 0, NULL, lockh);
279 if (rc != ELDLM_OK) {
281 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
284 ldlm_lock_decref(lockh + 1, LCK_CW);
291 #ifndef DCACHE_DISCONNECTED
292 #define DCACHE_DISCONNECTED DCACHE_NFSD_DISCONNECTED
296 /* Look up an entry by inode number. */
297 /* this function ONLY returns valid dget'd dentries with an initialized inode
299 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
300 struct vfsmount **mnt)
303 unsigned long ino = fid->id;
304 __u32 generation = fid->generation;
306 struct dentry *result;
309 RETURN(ERR_PTR(-ESTALE));
311 snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
313 CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
314 ino, generation, mds->mds_sb);
316 /* under ext3 this is neither supposed to return bad inodes
318 result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
322 inode = result->d_inode;
324 RETURN(ERR_PTR(-ENOENT));
326 /* here we disabled generation check, as root inode i_generation
327 * of cache mds and real mds are different. */
328 if (inode->i_ino != mds->mds_rootfid.id && generation &&
329 inode->i_generation != generation) {
330 /* we didn't find the right inode.. */
331 CERROR("bad inode %lu, link: %lu ct: %d or generation %u/%u\n",
332 inode->i_ino, (unsigned long)inode->i_nlink,
333 atomic_read(&inode->i_count), inode->i_generation,
336 RETURN(ERR_PTR(-ENOENT));
340 *mnt = mds->mds_vfsmnt;
348 /* Establish a connection to the MDS.
350 * This will set up an export structure for the client to hold state data
351 * about that client, like open files, the last operation number it did
352 * on the server, etc.
354 static int mds_connect(struct lustre_handle *conn, struct obd_device *obd,
355 struct obd_uuid *cluuid, unsigned long connect_flags)
357 struct obd_export *exp;
358 struct mds_export_data *med; /* */
359 struct mds_client_data *mcd;
363 if (!conn || !obd || !cluuid)
366 /* XXX There is a small race between checking the list and adding a
367 * new connection for the same UUID, but the real threat (list
368 * corruption when multiple different clients connect) is solved.
370 * There is a second race between adding the export to the list,
371 * and filling in the client data below. Hence skipping the case
372 * of NULL mcd above. We should already be controlling multiple
373 * connects at the client, and we can't hold the spinlock over
374 * memory allocations without risk of deadlocking.
376 rc = class_connect(conn, obd, cluuid);
379 exp = class_conn2export(conn);
381 med = &exp->exp_mds_data;
383 OBD_ALLOC(mcd, sizeof(*mcd));
385 CERROR("mds: out of memory for client data\n");
386 GOTO(out, rc = -ENOMEM);
389 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
392 rc = mds_client_add(obd, &obd->u.mds, med, -1);
396 if (!(connect_flags & OBD_OPT_MDS_CONNECTION)) {
397 struct mds_obd *mds = &obd->u.mds;
398 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)) {
399 atomic_inc(&mds->mds_real_clients);
400 CDEBUG(D_OTHER,"%s: peer from %s is real client (%d)\n",
401 obd->obd_name, exp->exp_client_uuid.uuid,
402 atomic_read(&mds->mds_real_clients));
403 exp->exp_flags |= OBD_OPT_REAL_CLIENT;
405 if (mds->mds_lmv_name)
406 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
411 OBD_FREE(mcd, sizeof(*mcd));
412 class_disconnect(exp, 0);
414 class_export_put(exp);
419 static int mds_init_export(struct obd_export *exp)
421 struct mds_export_data *med = &exp->exp_mds_data;
423 INIT_LIST_HEAD(&med->med_open_head);
424 spin_lock_init(&med->med_open_lock);
428 static int mds_destroy_export(struct obd_export *export)
430 struct mds_export_data *med;
431 struct obd_device *obd = export->exp_obd;
432 struct lvfs_run_ctxt saved;
436 med = &export->exp_mds_data;
437 target_destroy_export(export);
439 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
442 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
444 /* Close any open files (which may also cause orphan unlinking). */
445 spin_lock(&med->med_open_lock);
446 while (!list_empty(&med->med_open_head)) {
447 struct list_head *tmp = med->med_open_head.next;
448 struct mds_file_data *mfd =
449 list_entry(tmp, struct mds_file_data, mfd_list);
450 BDEVNAME_DECLARE_STORAGE(btmp);
452 /* bug 1579: fix force-closing for 2.5 */
453 struct dentry *dentry = mfd->mfd_dentry;
455 list_del(&mfd->mfd_list);
456 spin_unlock(&med->med_open_lock);
458 /* If you change this message, be sure to update
459 * replay_single:test_46 */
460 CERROR("force closing client file handle for %*s (%s:%lu)\n",
461 dentry->d_name.len, dentry->d_name.name,
462 ll_bdevname(dentry->d_inode->i_sb, btmp),
463 dentry->d_inode->i_ino);
464 rc = mds_mfd_close(NULL, 0, obd, mfd,
465 !(export->exp_flags & OBD_OPT_FAILOVER));
468 CDEBUG(D_INODE, "Error closing file: %d\n", rc);
469 spin_lock(&med->med_open_lock);
471 spin_unlock(&med->med_open_lock);
472 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
475 mds_client_free(export, !(export->exp_flags & OBD_OPT_FAILOVER));
480 static int mds_disconnect(struct obd_export *exp, int flags)
482 struct obd_device *obd;
484 unsigned long irqflags;
489 class_export_get(exp);
491 obd = class_exp2obd(exp);
493 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
494 exp->exp_handle.h_cookie);
499 if (!(exp->exp_flags & OBD_OPT_REAL_CLIENT)
500 && !atomic_read(&mds->mds_real_clients)) {
501 /* there was no client at all */
502 mds_lmv_disconnect(obd, flags);
505 if ((exp->exp_flags & OBD_OPT_REAL_CLIENT)
506 && atomic_dec_and_test(&mds->mds_real_clients)) {
507 /* time to drop LMV connections */
508 CDEBUG(D_OTHER, "%s: last real client %s disconnected. "
509 "Disconnnect from LMV now\n",
510 obd->obd_name, exp->exp_client_uuid.uuid);
511 mds_lmv_disconnect(obd, flags);
514 spin_lock_irqsave(&exp->exp_lock, irqflags);
515 exp->exp_flags = flags;
516 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
518 /* Disconnect early so that clients can't keep using export */
519 rc = class_disconnect(exp, flags);
520 ldlm_cancel_locks_for_export(exp);
522 /* complete all outstanding replies */
523 spin_lock_irqsave(&exp->exp_lock, irqflags);
524 while (!list_empty(&exp->exp_outstanding_replies)) {
525 struct ptlrpc_reply_state *rs =
526 list_entry(exp->exp_outstanding_replies.next,
527 struct ptlrpc_reply_state, rs_exp_list);
528 struct ptlrpc_service *svc = rs->rs_srv_ni->sni_service;
530 spin_lock(&svc->srv_lock);
531 list_del_init(&rs->rs_exp_list);
532 ptlrpc_schedule_difficult_reply(rs);
533 spin_unlock(&svc->srv_lock);
535 spin_unlock_irqrestore(&exp->exp_lock, irqflags);
537 class_export_put(exp);
541 static int mds_getstatus(struct ptlrpc_request *req)
543 struct mds_obd *mds = mds_req2mds(req);
544 struct mds_body *body;
545 int rc, size = sizeof(*body);
548 rc = lustre_pack_reply(req, 1, &size, NULL);
549 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
550 CERROR("mds: out of memory for message: size=%d\n", size);
551 req->rq_status = -ENOMEM; /* superfluous? */
555 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
556 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
558 /* the last_committed and last_xid fields are filled in for all
559 * replies already - no need to do so here also.
564 int mds_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc,
565 void *data, int flag)
570 if (flag == LDLM_CB_CANCELING) {
571 /* Don't need to do anything here. */
575 /* XXX layering violation! -phil */
576 l_lock(&lock->l_resource->lr_namespace->ns_lock);
577 /* Get this: if mds_blocking_ast is racing with mds_intent_policy,
578 * such that mds_blocking_ast is called just before l_i_p takes the
579 * ns_lock, then by the time we get the lock, we might not be the
580 * correct blocking function anymore. So check, and return early, if
582 if (lock->l_blocking_ast != mds_blocking_ast) {
583 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
587 lock->l_flags |= LDLM_FL_CBPENDING;
588 do_ast = (!lock->l_readers && !lock->l_writers);
589 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
592 struct lustre_handle lockh;
595 LDLM_DEBUG(lock, "already unused, calling ldlm_cli_cancel");
596 ldlm_lock2handle(lock, &lockh);
597 rc = ldlm_cli_cancel(&lockh);
599 CERROR("ldlm_cli_cancel: %d\n", rc);
601 LDLM_DEBUG(lock, "Lock still has references, will be "
607 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
615 rc = fsfilt_get_md(obd, inode, md, *size);
620 CERROR("Error %d reading eadata for ino %lu\n",
625 if (S_ISREG(inode->i_mode))
626 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
627 if (S_ISDIR(inode->i_mode))
628 rc = mds_convert_mea_ea(obd, inode, md, lmm_size);
642 /* Call with lock=1 if you want mds_pack_md to take the i_sem.
643 * Call with lock=0 if the caller has already taken the i_sem. */
644 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
645 struct mds_body *body, struct inode *inode, int lock)
647 struct mds_obd *mds = &obd->u.mds;
653 lmm = lustre_msg_buf(msg, offset, 0);
655 /* Some problem with getting eadata when I sized the reply
657 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
661 lmm_size = msg->buflens[offset];
663 /* I don't really like this, but it is a sanity check on the client
664 * MD request. However, if the client doesn't know how much space
665 * to reserve for the MD, it shouldn't be bad to have too much space.
667 if (lmm_size > mds->mds_max_mdsize) {
668 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
669 inode->i_ino, lmm_size, mds->mds_max_mdsize);
673 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
675 if (S_ISDIR(inode->i_mode))
676 body->valid |= OBD_MD_FLDIREA;
678 body->valid |= OBD_MD_FLEASIZE;
679 body->eadatasize = lmm_size;
686 void mds_squash_root(struct mds_obd *mds, struct mds_req_sec_desc *rsd,
689 if (!mds->mds_squash_uid ||
690 (rsd->rsd_uid && rsd->rsd_fsuid))
693 if (*peernid == mds->mds_nosquash_nid)
696 CDEBUG(D_OTHER, "squash req from 0x%llx, (%d:%d/%x)=>(%d:%d/%x)\n",
698 rsd->rsd_fsuid, rsd->rsd_fsgid, rsd->rsd_cap,
699 mds->mds_squash_uid, mds->mds_squash_gid,
700 (rsd->rsd_cap & ~CAP_FS_MASK));
702 rsd->rsd_uid = mds->mds_squash_uid;
703 rsd->rsd_fsuid = mds->mds_squash_uid;
704 rsd->rsd_fsgid = mds->mds_squash_gid;
705 /* XXX should we remove all capabilities? */
706 rsd->rsd_cap &= ~CAP_FS_MASK;
709 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
710 struct ptlrpc_request *req,
711 struct mds_body *reqbody, int reply_off)
713 struct mds_body *body;
714 struct inode *inode = dentry->d_inode;
718 if (inode == NULL && !(dentry->d_flags & DCACHE_CROSS_REF))
721 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
722 LASSERT(body != NULL); /* caller prepped reply */
724 if (dentry->d_flags & DCACHE_CROSS_REF) {
725 CDEBUG(D_OTHER, "cross reference: %lu/%lu/%lu\n",
726 (unsigned long) dentry->d_mdsnum,
727 (unsigned long) dentry->d_inum,
728 (unsigned long) dentry->d_generation);
729 body->valid |= OBD_MD_FLID | OBD_MD_MDS;
730 body->fid1.id = dentry->d_inum;
731 body->fid1.mds = dentry->d_mdsnum;
732 body->fid1.generation = dentry->d_generation;
735 mds_pack_inode2fid(obd, &body->fid1, inode);
736 mds_pack_inode2body(obd, body, inode);
738 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
739 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
740 rc = mds_pack_md(obd, req->rq_repmsg, reply_off + 1, body,
743 /* If we have LOV EA data, the OST holds size, atime, mtime */
744 if (!(body->valid & OBD_MD_FLEASIZE) &&
745 !(body->valid & OBD_MD_FLDIREA))
746 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
747 OBD_MD_FLATIME | OBD_MD_FLMTIME);
748 } else if (S_ISLNK(inode->i_mode) &&
749 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
750 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off + 1,0);
753 LASSERT (symname != NULL); /* caller prepped reply */
754 len = req->rq_repmsg->buflens[reply_off + 1];
756 rc = inode->i_op->readlink(dentry, symname, len);
758 CERROR("readlink failed: %d\n", rc);
759 } else if (rc != len - 1) {
760 CERROR ("Unexpected readlink rc %d: expecting %d\n",
764 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
765 body->valid |= OBD_MD_LINKNAME;
766 body->eadatasize = rc + 1;
767 symname[rc] = 0; /* NULL terminate */
775 static int mds_getattr_pack_msg_cf(struct ptlrpc_request *req,
776 struct dentry *dentry,
779 int rc = 0, size[1] = {sizeof(struct mds_body)};
782 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
783 CERROR("failed MDS_GETATTR_PACK test\n");
784 req->rq_status = -ENOMEM;
785 GOTO(out, rc = -ENOMEM);
788 rc = lustre_pack_reply(req, 1, size, NULL);
790 CERROR("out of memory\n");
791 GOTO(out, req->rq_status = rc);
799 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
802 struct mds_obd *mds = mds_req2mds(req);
803 struct mds_body *body;
804 int rc = 0, size[2] = {sizeof(*body)}, bufcount = 1;
807 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*body));
808 LASSERT(body != NULL); /* checked by caller */
809 LASSERT_REQSWABBED(req, offset); /* swabbed by caller */
811 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
812 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
815 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0);
817 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
821 CERROR("error getting inode %lu MD: rc = %d\n",
824 } else if (rc > mds->mds_max_mdsize) {
826 CERROR("MD size %d larger than maximum possible %u\n",
827 rc, mds->mds_max_mdsize);
832 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
833 if (inode->i_size + 1 != body->eadatasize)
834 CERROR("symlink size: %Lu, reply space: %d\n",
835 inode->i_size + 1, body->eadatasize);
836 size[bufcount] = min_t(int, inode->i_size+1, body->eadatasize);
838 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
839 inode->i_size + 1, body->eadatasize);
842 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
843 CERROR("failed MDS_GETATTR_PACK test\n");
844 req->rq_status = -ENOMEM;
845 GOTO(out, rc = -ENOMEM);
848 rc = lustre_pack_reply(req, bufcount, size, NULL);
850 CERROR("out of memory\n");
851 GOTO(out, req->rq_status = rc);
859 int mds_check_mds_num(struct obd_device *obd, struct inode* inode,
860 char *name, int namelen)
862 struct mea *mea = NULL;
863 int mea_size, rc = 0;
866 rc = mds_get_lmv_attr(obd, inode, &mea, &mea_size);
870 /* dir is already splitted, check if requested filename
871 * should live at this MDS or at another one */
873 i = mea_name2idx(mea, name, namelen - 1);
874 if (mea->mea_master != mea->mea_fids[i].mds) {
876 "inapropriate MDS(%d) for %s. should be %d(%d)\n",
877 mea->mea_master, name, mea->mea_fids[i].mds, i);
883 OBD_FREE(mea, mea_size);
887 static int mds_getattr_name(struct ptlrpc_request *req, int offset,
888 struct lustre_handle *child_lockh, int child_part)
890 struct obd_device *obd = req->rq_export->exp_obd;
891 struct mds_obd *mds = &obd->u.mds;
892 struct ldlm_reply *rep = NULL;
893 struct lvfs_run_ctxt saved;
894 struct mds_req_sec_desc *rsd;
895 struct mds_body *body;
896 struct dentry *dparent = NULL, *dchild = NULL;
897 struct lvfs_ucred uc;
898 struct lustre_handle parent_lockh[2];
899 int namesize, update_mode;
900 int rc = 0, cleanup_phase = 0, resent_req = 0, reply_offset;
901 struct clonefs_info *clone_info = NULL;
905 LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
907 MDS_UPDATE_COUNTER((&obd->u.mds), MDS_GETATTR_NAME_COUNT);
909 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
911 CERROR("Can't unpack security desc\n");
914 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
916 /* Swab now, before anyone looks inside the request */
917 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
918 lustre_swab_mds_body);
920 CERROR("Can't swab mds_body\n");
921 GOTO(cleanup, rc = -EFAULT);
924 LASSERT_REQSWAB(req, offset + 1);
925 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
927 CERROR("Can't unpack name\n");
928 GOTO(cleanup, rc = -EFAULT);
930 namesize = req->rq_reqmsg->buflens[offset + 1];
932 LASSERT (offset == 1 || offset == 3);
933 /* if requests were at offset 3, the getattr reply goes back at 1 */
935 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
941 rc = mds_init_ucred(&uc, rsd);
943 CERROR("can't init ucred\n");
947 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
948 cleanup_phase = 1; /* kernel context */
949 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
951 LASSERT(namesize > 0);
953 /* we have no dentry here, drop LOOKUP bit */
954 child_part &= ~MDS_INODELOCK_LOOKUP;
955 CDEBUG(D_OTHER, "%s: request to retrieve attrs for %lu/%lu\n",
956 obd->obd_name, (unsigned long) body->fid1.id,
957 (unsigned long) body->fid1.generation);
958 dchild = mds_fid2locked_dentry(obd, &body->fid1, NULL, LCK_PR,
959 parent_lockh, &update_mode,
960 NULL, 0, child_part);
961 if (IS_ERR(dchild)) {
962 CERROR("can't find inode: %d\n", (int) PTR_ERR(dchild));
963 GOTO(cleanup, rc = PTR_ERR(dchild));
965 memcpy(child_lockh, parent_lockh, sizeof(parent_lockh[0]));
967 if (parent_lockh[1].cookie)
968 ldlm_lock_decref(parent_lockh + 1, update_mode);
975 /* FIXME: handle raw lookup */
976 if (body->valid == OBD_MD_FLID) {
977 struct mds_body *mds_reply;
978 int size = sizeof(*mds_reply);
981 dparent = mds_fid2dentry(mds, &body->fid1, NULL);
982 if (IS_ERR(dparent)) {
983 rc = PTR_ERR(dparent);
986 // The user requested ONLY the inode number, so do a raw lookup
987 rc = lustre_pack_reply(req, 1, &size, NULL);
989 CERROR("out of memory\n");
993 dir = dparent->d_inode;
994 LASSERT(dir->i_op->lookup_raw != NULL);
995 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
997 mds_reply = lustre_msg_buf(req->rq_repmsg, 0,
999 mds_reply->fid1.id = inum;
1000 mds_reply->valid = OBD_MD_FLID;
1005 if (child_lockh->cookie != 0) {
1006 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1010 if (resent_req == 0) {
1011 rc = mds_get_parent_child_locked(obd, mds, &body->fid1,
1012 parent_lockh, &dparent,
1013 LCK_PR, MDS_INODELOCK_LOOKUP,
1014 &update_mode, name, namesize,
1015 child_lockh, &dchild, LCK_PR,
1016 child_part, clone_info);
1020 /* let's make sure this name should leave on this mds node */
1021 rc = mds_check_mds_num(obd, dparent->d_inode, name, namesize);
1025 struct ldlm_lock *granted_lock;
1026 struct ll_fid child_fid;
1027 struct ldlm_resource *res;
1028 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
1029 granted_lock = ldlm_handle2lock(child_lockh);
1030 LASSERT(granted_lock);
1032 res = granted_lock->l_resource;
1033 child_fid.id = res->lr_name.name[0];
1034 child_fid.generation = res->lr_name.name[1];
1035 dchild = mds_fid2dentry(mds, &child_fid, NULL);
1037 LDLM_LOCK_PUT(granted_lock);
1040 cleanup_phase = 2; /* dchild, dparent, locks */
1044 if (!DENTRY_VALID(dchild)) {
1045 intent_set_disposition(rep, DISP_LOOKUP_NEG);
1046 /* in the intent case, the policy clears this error:
1047 the disposition is enough */
1051 intent_set_disposition(rep, DISP_LOOKUP_POS);
1054 if (req->rq_repmsg == NULL) {
1055 if (dchild->d_flags & DCACHE_CROSS_REF)
1056 rc = mds_getattr_pack_msg_cf(req, dchild, offset);
1058 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
1060 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1065 rc = mds_getattr_internal(obd, dchild, req, body, reply_offset);
1066 GOTO(cleanup, rc); /* returns the lock to the client */
1069 switch (cleanup_phase) {
1071 if (resent_req == 0) {
1072 if (rc && DENTRY_VALID(dchild))
1073 ldlm_lock_decref(child_lockh, LCK_PR);
1075 ldlm_lock_decref(parent_lockh, LCK_PR);
1077 if (parent_lockh[1].cookie != 0)
1078 ldlm_lock_decref(parent_lockh + 1,
1087 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1088 mds_exit_ucred(&uc);
1094 static int mds_getattr(struct ptlrpc_request *req, int offset)
1096 struct mds_obd *mds = mds_req2mds(req);
1097 struct obd_device *obd = req->rq_export->exp_obd;
1098 struct lvfs_run_ctxt saved;
1100 struct mds_req_sec_desc *rsd;
1101 struct mds_body *body;
1102 struct lvfs_ucred uc;
1106 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1108 CERROR("Can't unpack security desc\n");
1112 body = lustre_swab_reqbuf(req, offset, sizeof (*body),
1113 lustre_swab_mds_body);
1115 CERROR ("Can't unpack body\n");
1119 MDS_UPDATE_COUNTER(mds, MDS_GETATTR_COUNT);
1121 rc = mds_init_ucred(&uc, rsd);
1123 CERROR("can't init ucred\n");
1127 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1128 de = mds_fid2dentry(mds, &body->fid1, NULL);
1130 rc = req->rq_status = -ENOENT;
1131 GOTO(out_pop, PTR_ERR(de));
1134 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1136 CERROR ("mds_getattr_pack_msg: %d\n", rc);
1140 req->rq_status = mds_getattr_internal(obd, de, req, body, 0);
1145 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1146 mds_exit_ucred(&uc);
1150 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1151 unsigned long max_age)
1155 spin_lock(&obd->obd_osfs_lock);
1156 rc = fsfilt_statfs(obd, obd->u.mds.mds_sb, max_age);
1158 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1159 spin_unlock(&obd->obd_osfs_lock);
1164 static int mds_statfs(struct ptlrpc_request *req)
1166 struct obd_device *obd = req->rq_export->exp_obd;
1167 int rc, size = sizeof(struct obd_statfs);
1170 rc = lustre_pack_reply(req, 1, &size, NULL);
1171 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1172 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1176 MDS_UPDATE_COUNTER((&obd->u.mds), MDS_STATFS_COUNT);
1178 /* We call this so that we can cache a bit - 1 jiffie worth */
1179 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, 0, size),
1182 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1188 req->rq_status = rc;
1192 static int mds_sync(struct ptlrpc_request *req, int offset)
1194 struct obd_device *obd = req->rq_export->exp_obd;
1195 struct mds_obd *mds = &obd->u.mds;
1196 struct mds_body *body;
1197 int rc, size = sizeof(*body);
1200 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
1202 GOTO(out, rc = -EPROTO);
1204 rc = lustre_pack_reply(req, 1, &size, NULL);
1205 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1206 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1210 if (body->fid1.id == 0) {
1211 /* a fid of zero is taken to mean "sync whole filesystem" */
1212 rc = fsfilt_sync(obd, mds->mds_sb);
1216 /* just any file to grab fsync method - "file" arg unused */
1217 struct file *file = mds->mds_rcvd_filp;
1220 de = mds_fid2dentry(mds, &body->fid1, NULL);
1222 GOTO(out, rc = PTR_ERR(de));
1224 rc = file->f_op->fsync(NULL, de, 1);
1229 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*body));
1230 mds_pack_inode2fid(obd, &body->fid1, de->d_inode);
1231 mds_pack_inode2body(obd, body, de->d_inode);
1234 req->rq_status = rc;
1238 /* mds_readpage does not take a DLM lock on the inode, because the client must
1239 * already have a PR lock.
1241 * If we were to take another one here, a deadlock will result, if another
1242 * thread is already waiting for a PW lock. */
1243 static int mds_readpage(struct ptlrpc_request *req, int offset)
1245 struct obd_device *obd = req->rq_export->exp_obd;
1246 struct mds_obd *mds = &obd->u.mds;
1247 struct vfsmount *mnt;
1250 struct mds_req_sec_desc *rsd;
1251 struct mds_body *body, *repbody;
1252 struct lvfs_run_ctxt saved;
1253 int rc, size = sizeof(*repbody);
1254 struct lvfs_ucred uc;
1257 rc = lustre_pack_reply(req, 1, &size, NULL);
1258 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK)) {
1259 CERROR("mds: out of memory\n");
1260 GOTO(out, rc = -ENOMEM);
1263 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1265 CERROR("Can't unpack security desc\n");
1266 GOTO (out, rc = -EFAULT);
1268 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
1270 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1271 lustre_swab_mds_body);
1273 CERROR("Can't unpack body\n");
1274 GOTO (out, rc = -EFAULT);
1277 rc = mds_init_ucred(&uc, rsd);
1279 CERROR("can't init ucred\n");
1283 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1284 de = mds_fid2dentry(mds, &body->fid1, &mnt);
1286 GOTO(out_pop, rc = PTR_ERR(de));
1288 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1290 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1291 /* note: in case of an error, dentry_open puts dentry */
1293 GOTO(out_pop, rc = PTR_ERR(file));
1295 /* body->size is actually the offset -eeb */
1296 if ((body->size & (de->d_inode->i_blksize - 1)) != 0) {
1297 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1298 body->size, de->d_inode->i_blksize);
1299 GOTO(out_file, rc = -EFAULT);
1302 /* body->nlink is actually the #bytes to read -eeb */
1303 if (body->nlink & (de->d_inode->i_blksize - 1)) {
1304 CERROR("size %u is not multiple of blocksize %lu\n",
1305 body->nlink, de->d_inode->i_blksize);
1306 GOTO(out_file, rc = -EFAULT);
1309 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*repbody));
1310 repbody->size = file->f_dentry->d_inode->i_size;
1311 repbody->valid = OBD_MD_FLSIZE;
1313 /* to make this asynchronous make sure that the handling function
1314 doesn't send a reply when this function completes. Instead a
1315 callback function would send the reply */
1316 /* body->size is actually the offset -eeb */
1317 rc = mds_sendpage(req, file, body->size, body->nlink);
1320 filp_close(file, 0);
1322 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1323 mds_exit_ucred(&uc);
1325 req->rq_status = rc;
1329 int mds_reint(struct ptlrpc_request *req, int offset,
1330 struct lustre_handle *lockh)
1332 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1333 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
1334 struct mds_req_sec_desc *rsd;
1338 OBD_ALLOC(rec, sizeof(*rec));
1342 rsd = lustre_swab_mds_secdesc(req, MDS_REQ_SECDESC_OFF);
1344 CERROR("Can't unpack security desc\n");
1345 GOTO(out, rc = -EFAULT);
1347 mds_squash_root(mds, rsd, &req->rq_peer.peer_id.nid);
1349 rc = mds_update_unpack(req, offset, rec);
1350 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1351 CERROR("invalid record\n");
1352 GOTO(out, req->rq_status = -EINVAL);
1355 rc = mds_init_ucred(&rec->ur_uc, rsd);
1357 CERROR("can't init ucred\n");
1361 /* rc will be used to interrupt a for loop over multiple records */
1362 rc = mds_reint_rec(rec, offset, req, lockh);
1364 mds_exit_ucred(&rec->ur_uc);
1366 OBD_FREE(rec, sizeof(*rec));
1370 static int mds_filter_recovery_request(struct ptlrpc_request *req,
1371 struct obd_device *obd, int *process)
1373 switch (req->rq_reqmsg->opc) {
1374 case MDS_CONNECT: /* This will never get here, but for completeness. */
1375 case OST_CONNECT: /* This will never get here, but for completeness. */
1376 case MDS_DISCONNECT:
1377 case OST_DISCONNECT:
1382 case MDS_SYNC: /* used in unmounting */
1387 *process = target_queue_recovery_request(req, obd);
1391 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1393 /* XXX what should we set rq_status to here? */
1394 req->rq_status = -EAGAIN;
1395 RETURN(ptlrpc_error(req));
1399 static char *reint_names[] = {
1400 [REINT_SETATTR] "setattr",
1401 [REINT_CREATE] "create",
1402 [REINT_LINK] "link",
1403 [REINT_UNLINK] "unlink",
1404 [REINT_RENAME] "rename",
1405 [REINT_OPEN] "open",
1408 #define FILTER_VALID_FLAGS (OBD_MD_FLTYPE | OBD_MD_FLMODE | OBD_MD_FLGENER |\
1409 OBD_MD_FLSIZE | OBD_MD_FLBLOCKS | OBD_MD_FLBLKSZ|\
1410 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME|\
1413 static void reconstruct_create(struct ptlrpc_request *req)
1415 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1416 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
1417 struct mds_client_data *mcd = med->med_mcd;
1418 struct dentry *dentry;
1419 struct ost_body *body;
1423 /* copy rc, transno and disp; steal locks */
1424 mds_req_from_mcd(req, mcd);
1425 if (req->rq_status) {
1431 fid.id = mcd->mcd_last_data;
1434 LASSERT(fid.id != 0);
1436 dentry = mds_fid2dentry(mds, &fid, NULL);
1437 if (IS_ERR(dentry)) {
1438 CERROR("can't find inode "LPU64"\n", fid.id);
1439 req->rq_status = PTR_ERR(dentry);
1444 CWARN("reconstruct reply for x"LPU64" (remote ino) "LPU64" -> %lu/%u\n",
1445 req->rq_xid, fid.id, dentry->d_inode->i_ino,
1446 dentry->d_inode->i_generation);
1448 body = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*body));
1449 obdo_from_inode(&body->oa, dentry->d_inode, FILTER_VALID_FLAGS);
1450 body->oa.o_id = dentry->d_inode->i_ino;
1451 body->oa.o_generation = dentry->d_inode->i_generation;
1452 body->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1459 static int mdt_obj_create(struct ptlrpc_request *req)
1461 struct obd_device *obd = req->rq_export->exp_obd;
1462 struct mds_obd *mds = &obd->u.mds;
1463 struct ost_body *body, *repbody;
1464 char fidname[LL_FID_NAMELEN];
1465 struct inode *parent_inode;
1466 struct lvfs_run_ctxt saved;
1467 struct dentry *new = NULL;
1468 struct dentry_params dp;
1469 int mealen, flags = 0, rc, size = sizeof(*repbody), cleanup_phase = 0;
1470 struct lvfs_ucred uc;
1472 void *handle = NULL;
1473 unsigned long cr_inum = 0;
1476 DEBUG_REQ(D_HA, req, "create remote object");
1478 parent_inode = mds->mds_unnamed_dir->d_inode;
1480 body = lustre_swab_reqbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
1484 rc = lustre_pack_reply(req, 1, &size, NULL);
1488 MDS_CHECK_RESENT(req, reconstruct_create(req));
1490 /* FIXME don't understand what security context we should
1491 * take here. check this -ericm
1493 uc.luc_fsuid = body->oa.o_uid;
1494 uc.luc_fsgid = body->oa.o_gid;
1495 uc.luc_uid = body->oa.o_uid;
1496 uc.luc_ghash = mds_get_group_entry(mds, body->oa.o_uid);
1497 uc.luc_ginfo = NULL;
1498 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1500 repbody = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*repbody));
1502 if (body->oa.o_flags & OBD_FL_RECREATE_OBJS) {
1503 /* this is re-create request from MDS holding directory name.
1504 * we have to lookup given ino/generation first. if it exists
1505 * (good case) then there is nothing to do. if it does not
1506 * then we have to recreate it */
1508 fid.id = body->oa.o_id;
1509 fid.generation = body->oa.o_generation;
1510 new = mds_fid2dentry(mds, &fid, NULL);
1511 if (!IS_ERR(new) && new->d_inode) {
1512 CDEBUG(D_HA, "mkdir() repairing is on its way: %lu/%lu\n",
1513 (unsigned long) fid.id,
1514 (unsigned long) fid.generation);
1515 obdo_from_inode(&repbody->oa, new->d_inode,
1516 FILTER_VALID_FLAGS);
1517 repbody->oa.o_id = new->d_inode->i_ino;
1518 repbody->oa.o_generation = new->d_inode->i_generation;
1519 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1521 cr_inum = new->d_inode->i_ino;
1522 GOTO(cleanup, rc = 0);
1524 CWARN("hmm. for some reason dir %lu/%lu (or reply) got lost\n",
1525 (unsigned long) fid.id, (unsigned long) fid.generation);
1526 LASSERT(new->d_inode == NULL ||
1527 new->d_inode->i_generation != fid.generation);
1531 down(&parent_inode->i_sem);
1532 handle = fsfilt_start(obd, parent_inode, FSFILT_OP_MKDIR, NULL);
1533 if (IS_ERR(handle)) {
1534 up(&parent_inode->i_sem);
1535 GOTO(cleanup, rc = PTR_ERR(handle));
1537 cleanup_phase = 1; /* transaction */
1540 rc = sprintf(fidname, "%u.%u", ll_insecure_random_int(), current->pid);
1541 new = lookup_one_len(fidname, mds->mds_unnamed_dir, rc);
1543 CERROR("%s: can't lookup new inode (%s) for mkdir: %d\n",
1544 obd->obd_name, fidname, (int) PTR_ERR(new));
1545 fsfilt_commit(obd, mds->mds_sb, new->d_inode, handle, 0);
1546 up(&parent_inode->i_sem);
1547 RETURN(PTR_ERR(new));
1548 } else if (new->d_inode) {
1549 CERROR("%s: name exists. repeat\n", obd->obd_name);
1553 new->d_fsdata = (void *) &dp;
1557 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1558 DEBUG_REQ(D_HA, req, "replay create obj %lu/%lu",
1559 (unsigned long) body->oa.o_id,
1560 (unsigned long) body->oa.o_generation);
1561 dp.p_inum = body->oa.o_id;
1563 rc = vfs_mkdir(parent_inode, new, body->oa.o_mode);
1565 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1566 new->d_inode->i_generation = body->oa.o_generation;
1567 mark_inode_dirty(new->d_inode);
1568 LASSERTF(body->oa.o_id == new->d_inode->i_ino,
1569 "BUG 3550: failed to recreate obj "
1571 body->oa.o_id, new->d_inode->i_ino);
1572 LASSERTF(body->oa.o_generation ==
1573 new->d_inode->i_generation,
1574 "BUG 3550: failed to recreate obj/gen "
1575 LPU64"/%u -> %lu/%u\n",
1576 body->oa.o_id, body->oa.o_generation,
1577 new->d_inode->i_ino,
1578 new->d_inode->i_generation);
1581 obdo_from_inode(&repbody->oa, new->d_inode, FILTER_VALID_FLAGS);
1582 repbody->oa.o_id = new->d_inode->i_ino;
1583 repbody->oa.o_generation = new->d_inode->i_generation;
1584 repbody->oa.o_valid |= OBD_MD_FLID | OBD_MD_FLGENER;
1586 rc = fsfilt_del_dir_entry(obd, new);
1587 up(&parent_inode->i_sem);
1590 CERROR("can't remove name for object: %d\n", rc);
1594 cleanup_phase = 2; /* created directory object */
1596 cr_inum = new->d_inode->i_ino;
1597 CDEBUG(D_OTHER, "created dirobj: %lu/%u mode %o\n",
1598 new->d_inode->i_ino, new->d_inode->i_generation,
1599 new->d_inode->i_mode);
1601 up(&parent_inode->i_sem);
1602 CERROR("%s: can't create dirobj: %d\n", obd->obd_name, rc);
1606 if (body->oa.o_valid & OBD_MD_FLID) {
1607 /* this is new object for splitted dir. we have to
1608 * prevent recursive splitting on it -bzzz */
1609 mealen = obd_size_diskmd(mds->mds_lmv_exp, NULL);
1610 OBD_ALLOC(mea, mealen);
1612 GOTO(cleanup, rc = -ENOMEM);
1613 mea->mea_magic = MEA_MAGIC_ALL_CHARS;
1614 mea->mea_master = 0;
1616 down(&new->d_inode->i_sem);
1617 rc = fsfilt_set_md(obd, new->d_inode, handle, mea, mealen);
1618 up(&new->d_inode->i_sem);
1619 OBD_FREE(mea, mealen);
1620 CDEBUG(D_OTHER, "%s: mark non-splittable %lu/%u - %d\n",
1621 obd->obd_name, new->d_inode->i_ino,
1622 new->d_inode->i_generation, flags);
1623 } else if (body->oa.o_easize) {
1624 /* we pass LCK_EX to split routine to signal that we have
1625 * exclusive access to the directory. simple because nobody
1626 * knows it already exists -bzzz */
1627 mds_try_to_split_dir(obd, new, NULL, body->oa.o_easize, LCK_EX);
1631 switch (cleanup_phase) {
1632 case 2: /* object has been created, but we'll may want to replay it later */
1634 ptlrpc_require_repack(req);
1635 case 1: /* transaction */
1636 rc = mds_finish_transno(mds, parent_inode, handle, req, rc, cr_inum);
1640 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1641 mds_put_group_entry(mds, uc.luc_ghash);
1645 static int mdt_get_info(struct ptlrpc_request *req)
1648 struct obd_export *exp = req->rq_export;
1649 int keylen, rc = 0, size = sizeof(obd_id);
1653 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1655 DEBUG_REQ(D_HA, req, "no get_info key");
1658 keylen = req->rq_reqmsg->buflens[0];
1660 if (keylen < strlen("mdsize") || memcmp(key, "mdsize", 6) != 0)
1663 rc = lustre_pack_reply(req, 1, &size, NULL);
1667 reply = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*reply));
1668 rc = obd_get_info(exp, keylen, key, &size, reply);
1669 req->rq_repmsg->status = 0;
1673 static int mds_set_info(struct obd_export *exp, __u32 keylen,
1674 void *key, __u32 vallen, void *val)
1676 struct obd_device *obd;
1677 struct mds_obd *mds;
1681 obd = class_exp2obd(exp);
1683 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
1684 exp->exp_handle.h_cookie);
1688 #define KEY_IS(str) \
1689 (keylen == strlen(str) && memcmp(key, str, keylen) == 0)
1692 if (KEY_IS("mds_num")) {
1695 CDEBUG(D_IOCTL, "set mds num %d\n", *(int*)val);
1696 mds->mds_num = *(int*)val;
1697 group = FILTER_GROUP_FIRST_MDS + mds->mds_num;
1698 valsize = sizeof(group);
1699 /*mds number has been changed, so the corresponding obdfilter exp
1700 *need to be changed too*/
1701 rc = obd_set_info(mds->mds_osc_exp, strlen("mds_conn"), "mds_conn",
1705 CDEBUG(D_IOCTL, "invalid key\n");
1709 static int mdt_set_info(struct ptlrpc_request *req)
1712 struct obd_export *exp = req->rq_export;
1713 int keylen, rc = 0, vallen;
1716 key = lustre_msg_buf(req->rq_reqmsg, 0, 1);
1718 DEBUG_REQ(D_HA, req, "no set_info key");
1721 keylen = req->rq_reqmsg->buflens[0];
1723 if (keylen == strlen("mds_num") &&
1724 memcmp(key, "mds_num", keylen) == 0) {
1725 rc = lustre_pack_reply(req, 0, NULL, NULL);
1728 val = lustre_msg_buf(req->rq_reqmsg, 1, 0);
1730 vallen = req->rq_reqmsg->buflens[1];
1732 rc = obd_set_info(exp, keylen, key, vallen, val);
1733 req->rq_repmsg->status = 0;
1735 } else if (keylen == strlen("client") &&
1736 memcmp(key, "client", keylen) == 0) {
1737 rc = lustre_pack_reply(req, 0, NULL, NULL);
1740 rc = obd_set_info(exp, keylen, key, sizeof(obd_id), NULL);
1741 req->rq_repmsg->status = 0;
1744 CDEBUG(D_IOCTL, "invalid key\n");
1748 static int mds_msg_check_version(struct lustre_msg *msg)
1754 case MDS_DISCONNECT:
1756 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1758 CERROR("bad opc %u version %08x, expecting %08x\n",
1759 msg->opc, msg->version, LUSTRE_OBD_VERSION);
1764 case MDS_GETATTR_NAME:
1768 case MDS_DONE_WRITING:
1771 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
1773 CERROR("bad opc %u version %08x, expecting %08x\n",
1774 msg->opc, msg->version, LUSTRE_MDS_VERSION);
1778 case LDLM_BL_CALLBACK:
1779 case LDLM_CP_CALLBACK:
1780 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1782 CERROR("bad opc %u version %08x, expecting %08x\n",
1783 msg->opc, msg->version, LUSTRE_DLM_VERSION);
1785 case OBD_LOG_CANCEL:
1786 case LLOG_ORIGIN_HANDLE_OPEN:
1787 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1788 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1789 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1790 case LLOG_ORIGIN_HANDLE_CLOSE:
1792 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1794 CERROR("bad opc %u version %08x, expecting %08x\n",
1795 msg->opc, msg->version, LUSTRE_LOG_VERSION);
1801 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1803 CERROR("bad opc %u version %08x, expecting %08x\n",
1804 msg->opc, msg->version, LUSTRE_OBD_VERSION);
1807 CERROR("MDS unknown opcode %d\n", msg->opc);
1815 int mds_handle(struct ptlrpc_request *req)
1817 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1819 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1820 struct obd_device *obd = NULL;
1823 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1825 rc = mds_msg_check_version(req->rq_reqmsg);
1827 CERROR("MDS drop mal-formed request\n");
1831 LASSERT(current->journal_info == NULL);
1832 /* XXX identical to OST */
1833 if (req->rq_reqmsg->opc != MDS_CONNECT) {
1834 struct mds_export_data *med;
1837 if (req->rq_export == NULL) {
1838 CERROR("lustre_mds: operation %d on unconnected MDS\n",
1839 req->rq_reqmsg->opc);
1840 req->rq_status = -ENOTCONN;
1841 GOTO(out, rc = -ENOTCONN);
1844 med = &req->rq_export->exp_mds_data;
1845 obd = req->rq_export->exp_obd;
1848 /* sanity check: if the xid matches, the request must
1849 * be marked as a resent or replayed */
1850 if (req->rq_xid == med->med_mcd->mcd_last_xid)
1851 LASSERTF(lustre_msg_get_flags(req->rq_reqmsg) &
1852 (MSG_RESENT | MSG_REPLAY),
1853 "rq_xid "LPU64" matches last_xid, "
1854 "expected RESENT flag\n",
1856 /* else: note the opposite is not always true; a
1857 * RESENT req after a failover will usually not match
1858 * the last_xid, since it was likely never
1859 * committed. A REPLAYed request will almost never
1860 * match the last xid, however it could for a
1861 * committed, but still retained, open. */
1863 spin_lock_bh(&obd->obd_processing_task_lock);
1864 recovering = obd->obd_recovering;
1865 spin_unlock_bh(&obd->obd_processing_task_lock);
1867 rc = mds_filter_recovery_request(req, obd,
1869 if (rc || should_process == 0) {
1871 } else if (should_process < 0) {
1872 req->rq_status = should_process;
1873 rc = ptlrpc_error(req);
1879 switch (req->rq_reqmsg->opc) {
1881 DEBUG_REQ(D_INODE, req, "connect");
1882 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1883 rc = target_handle_connect(req);
1885 /* Now that we have an export, set mds. */
1886 mds = mds_req2mds(req);
1889 case MDS_DISCONNECT:
1890 DEBUG_REQ(D_INODE, req, "disconnect");
1891 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1892 rc = target_handle_disconnect(req);
1893 req->rq_status = rc; /* superfluous? */
1897 DEBUG_REQ(D_INODE, req, "getstatus");
1898 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1899 rc = mds_getstatus(req);
1903 DEBUG_REQ(D_INODE, req, "getattr");
1904 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1905 rc = mds_getattr(req, MDS_REQ_REC_OFF);
1908 case MDS_GETATTR_NAME: {
1909 struct lustre_handle lockh;
1910 DEBUG_REQ(D_INODE, req, "getattr_name");
1911 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1913 /* If this request gets a reconstructed reply, we won't be
1914 * acquiring any new locks in mds_getattr_name, so we don't
1918 rc = mds_getattr_name(req, MDS_REQ_REC_OFF, &lockh,
1919 MDS_INODELOCK_UPDATE);
1920 /* this non-intent call (from an ioctl) is special */
1921 req->rq_status = rc;
1922 if (rc == 0 && lockh.cookie)
1923 ldlm_lock_decref(&lockh, LCK_PR);
1927 DEBUG_REQ(D_INODE, req, "statfs");
1928 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1929 rc = mds_statfs(req);
1933 DEBUG_REQ(D_INODE, req, "readpage");
1934 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1935 rc = mds_readpage(req, MDS_REQ_REC_OFF);
1937 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1938 if (req->rq_reply_state) {
1939 lustre_free_reply_state (req->rq_reply_state);
1940 req->rq_reply_state = NULL;
1948 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, MDS_REQ_REC_OFF,
1951 int size[3] = {sizeof(struct mds_body), mds->mds_max_mdsize,
1952 mds->mds_max_cookiesize};
1955 /* NB only peek inside req now; mds_reint() will swab it */
1957 CERROR ("Can't inspect opcode\n");
1962 if (lustre_msg_swabbed (req->rq_reqmsg))
1965 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1966 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1967 reint_names[opc] == NULL) ? reint_names[opc] :
1970 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1972 if (opc == REINT_UNLINK || opc == REINT_RENAME)
1974 else if (opc == REINT_OPEN)
1979 rc = lustre_pack_reply(req, bufcount, size, NULL);
1983 rc = mds_reint(req, MDS_REQ_REC_OFF, NULL);
1984 fail = OBD_FAIL_MDS_REINT_NET_REP;
1989 DEBUG_REQ(D_INODE, req, "close");
1990 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1991 rc = mds_close(req, MDS_REQ_REC_OFF);
1994 case MDS_DONE_WRITING:
1995 DEBUG_REQ(D_INODE, req, "done_writing");
1996 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1997 rc = mds_done_writing(req, MDS_REQ_REC_OFF);
2001 DEBUG_REQ(D_INODE, req, "pin");
2002 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
2003 rc = mds_pin(req, MDS_REQ_REC_OFF);
2007 DEBUG_REQ(D_INODE, req, "sync");
2008 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
2009 rc = mds_sync(req, MDS_REQ_REC_OFF);
2013 DEBUG_REQ(D_INODE, req, "ping");
2014 rc = target_handle_ping(req);
2017 case OBD_LOG_CANCEL:
2018 CDEBUG(D_INODE, "log cancel\n");
2019 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
2020 rc = -ENOTSUPP; /* la la la */
2024 DEBUG_REQ(D_INODE, req, "enqueue");
2025 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
2026 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2027 ldlm_server_blocking_ast, NULL);
2030 DEBUG_REQ(D_INODE, req, "convert");
2031 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
2032 rc = ldlm_handle_convert(req);
2034 case LDLM_BL_CALLBACK:
2035 case LDLM_CP_CALLBACK:
2036 DEBUG_REQ(D_INODE, req, "callback");
2037 CERROR("callbacks should not happen on MDS\n");
2039 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
2041 case LLOG_ORIGIN_HANDLE_OPEN:
2042 DEBUG_REQ(D_INODE, req, "llog_init");
2043 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2044 rc = llog_origin_handle_open(req);
2046 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
2047 DEBUG_REQ(D_INODE, req, "llog next block");
2048 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2049 rc = llog_origin_handle_next_block(req);
2051 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
2052 DEBUG_REQ(D_INODE, req, "llog prev block");
2053 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2054 rc = llog_origin_handle_prev_block(req);
2056 case LLOG_ORIGIN_HANDLE_READ_HEADER:
2057 DEBUG_REQ(D_INODE, req, "llog read header");
2058 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2059 rc = llog_origin_handle_read_header(req);
2061 case LLOG_ORIGIN_HANDLE_CLOSE:
2062 DEBUG_REQ(D_INODE, req, "llog close");
2063 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2064 rc = llog_origin_handle_close(req);
2067 DEBUG_REQ(D_INODE, req, "ost_create");
2068 rc = mdt_obj_create(req);
2071 DEBUG_REQ(D_INODE, req, "get_info");
2072 rc = mdt_get_info(req);
2075 DEBUG_REQ(D_INODE, req, "set_info");
2076 rc = mdt_set_info(req);
2079 CDEBUG(D_INODE, "write\n");
2080 OBD_FAIL_RETURN(OBD_FAIL_OST_BRW_NET, 0);
2081 rc = ost_brw_write(req, NULL);
2082 LASSERT(current->journal_info == NULL);
2083 /* mdt_brw sends its own replies */
2087 DEBUG_REQ(D_INODE, req, "llog catinfo");
2088 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
2089 rc = llog_catinfo(req);
2092 req->rq_status = -ENOTSUPP;
2093 rc = ptlrpc_error(req);
2097 LASSERT(current->journal_info == NULL);
2101 /* If we're DISCONNECTing, the mds_export_data is already freed */
2102 if (!rc && req->rq_reqmsg->opc != MDS_DISCONNECT) {
2103 struct mds_export_data *med = &req->rq_export->exp_mds_data;
2104 struct obd_device *obd = list_entry(mds, struct obd_device,
2106 req->rq_repmsg->last_xid =
2107 le64_to_cpu(med->med_mcd->mcd_last_xid);
2109 if (!obd->obd_no_transno) {
2110 req->rq_repmsg->last_committed =
2111 obd->obd_last_committed;
2113 DEBUG_REQ(D_IOCTL, req,
2114 "not sending last_committed update");
2116 CDEBUG(D_INFO, "last_transno "LPU64", last_committed "LPU64
2118 mds->mds_last_transno, obd->obd_last_committed,
2123 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
2124 if (obd && obd->obd_recovering) {
2125 DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
2126 return target_queue_final_reply(req, rc);
2128 /* Lost a race with recovery; let the error path DTRT. */
2129 rc = req->rq_status = -ENOTCONN;
2132 target_send_reply(req, rc, fail);
2136 /* Update the server data on disk. This stores the new mount_count and
2137 * also the last_rcvd value to disk. If we don't have a clean shutdown,
2138 * then the server last_rcvd value may be less than that of the clients.
2139 * This will alert us that we may need to do client recovery.
2141 * Also assumes for mds_last_transno that we are not modifying it (no locking).
2143 int mds_update_server_data(struct obd_device *obd, int force_sync)
2145 struct mds_obd *mds = &obd->u.mds;
2146 struct mds_server_data *msd = mds->mds_server_data;
2147 struct file *filp = mds->mds_rcvd_filp;
2148 struct lvfs_run_ctxt saved;
2153 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2154 msd->msd_last_transno = cpu_to_le64(mds->mds_last_transno);
2156 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
2157 mds->mds_mount_count, mds->mds_last_transno);
2158 rc = fsfilt_write_record(obd, filp, msd, sizeof(*msd), &off,force_sync);
2160 CERROR("error writing MDS server data: rc = %d\n", rc);
2161 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2166 /* mount the file system (secretly) */
2167 static int mds_setup(struct obd_device *obd, obd_count len, void *buf)
2169 struct lustre_cfg* lcfg = buf;
2170 struct mds_obd *mds = &obd->u.mds;
2171 char *options = NULL;
2172 struct vfsmount *mnt;
2177 dev_clear_rdonly(2);
2179 if (!lcfg->lcfg_inlbuf1 || !lcfg->lcfg_inlbuf2)
2180 RETURN(rc = -EINVAL);
2182 obd->obd_fsops = fsfilt_get_ops(lcfg->lcfg_inlbuf2);
2183 if (IS_ERR(obd->obd_fsops))
2184 RETURN(rc = PTR_ERR(obd->obd_fsops));
2186 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
2188 page = __get_free_page(GFP_KERNEL);
2192 options = (char *)page;
2193 memset(options, 0, PAGE_SIZE);
2195 /* here we use "iopen_nopriv" hardcoded, because it affects MDS utility
2196 * and the rest of options are passed by mount options. Probably this
2197 * should be moved to somewhere else like startup scripts or lconf. */
2198 sprintf(options, "iopen_nopriv");
2200 if (lcfg->lcfg_inllen4 > 0 && lcfg->lcfg_inlbuf4)
2201 sprintf(options + strlen(options), ",%s",
2202 lcfg->lcfg_inlbuf4);
2204 /* we have to know mdsnum before touching underlying fs -bzzz */
2205 sema_init(&mds->mds_lmv_sem, 1);
2206 mds->mds_lmv_connected = 0;
2207 if (lcfg->lcfg_inllen5 > 0 && lcfg->lcfg_inlbuf5 &&
2208 strcmp(lcfg->lcfg_inlbuf5, "dumb")) {
2211 CDEBUG(D_OTHER, "MDS: %s is master for %s\n",
2212 obd->obd_name, lcfg->lcfg_inlbuf5);
2214 generate_random_uuid(uuid);
2215 class_uuid_unparse(uuid, &mds->mds_lmv_uuid);
2217 OBD_ALLOC(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2218 if (mds->mds_lmv_name == NULL)
2219 RETURN(rc = -ENOMEM);
2221 memcpy(mds->mds_lmv_name, lcfg->lcfg_inlbuf5,
2222 lcfg->lcfg_inllen5);
2224 rc = mds_lmv_connect(obd, mds->mds_lmv_name);
2226 OBD_FREE(mds->mds_lmv_name, lcfg->lcfg_inllen5);
2231 /* FIXME-WANGDI: this should be reworked when we will use lmv along
2232 * with cobd, because correct mdsnum is set in mds_lmv_connect(). */
2233 if (lcfg->lcfg_inllen6 > 0 && lcfg->lcfg_inlbuf6 && !mds->mds_lmv_obd &&
2234 strcmp(lcfg->lcfg_inlbuf6, "dumb")) {
2235 if (!memcmp(lcfg->lcfg_inlbuf6, "master", strlen("master")) &&
2236 mds->mds_num == 0) {
2237 mds->mds_num = REAL_MDS_NUMBER;
2238 } else if (!memcmp(lcfg->lcfg_inlbuf6, "cache", strlen("cache")) &&
2239 mds->mds_num == 0) {
2240 mds->mds_num = CACHE_MDS_NUMBER;
2244 mnt = do_kern_mount(lcfg->lcfg_inlbuf2, 0,
2245 lcfg->lcfg_inlbuf1, options);
2251 CERROR("do_kern_mount failed: rc = %d\n", rc);
2255 CDEBUG(D_SUPER, "%s: mnt = %p\n", lcfg->lcfg_inlbuf1, mnt);
2257 sema_init(&mds->mds_orphan_recovery_sem, 1);
2258 sema_init(&mds->mds_epoch_sem, 1);
2259 spin_lock_init(&mds->mds_transno_lock);
2260 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2261 atomic_set(&mds->mds_real_clients, 0);
2263 obd->obd_namespace = ldlm_namespace_new(obd->obd_name,
2264 LDLM_NAMESPACE_SERVER);
2265 if (obd->obd_namespace == NULL) {
2266 mds_cleanup(obd, 0);
2267 GOTO(err_put, rc = -ENOMEM);
2269 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
2271 rc = mds_fs_setup(obd, mnt);
2273 CERROR("MDS filesystem method init failed: rc = %d\n", rc);
2277 rc = llog_start_commit_thread();
2281 /* this check for @dumb string is needed to handle mounting MDS with
2282 * smfs. Read lconf:MDSDEV.write_conf() for more details. */
2283 if (lcfg->lcfg_inllen3 > 0 && lcfg->lcfg_inlbuf3 &&
2284 strcmp(lcfg->lcfg_inlbuf3, "dumb")) {
2287 generate_random_uuid(uuid);
2288 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
2290 OBD_ALLOC(mds->mds_profile, lcfg->lcfg_inllen3);
2291 if (mds->mds_profile == NULL)
2292 GOTO(err_fs, rc = -ENOMEM);
2294 memcpy(mds->mds_profile, lcfg->lcfg_inlbuf3,
2295 lcfg->lcfg_inllen3);
2298 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
2299 "mds_ldlm_client", &obd->obd_ldlm_client);
2300 obd->obd_replayable = 1;
2302 mds->mds_counters = lprocfs_alloc_mds_counters();
2304 rc = mds_postsetup(obd);
2311 /* No extra cleanup needed for llog_init_commit_thread() */
2312 mds_fs_cleanup(obd, 0);
2314 ldlm_namespace_free(obd->obd_namespace, 0);
2315 obd->obd_namespace = NULL;
2318 mntput(mds->mds_vfsmnt);
2322 fsfilt_put_ops(obd->obd_fsops);
2326 static int mds_postsetup(struct obd_device *obd)
2328 struct mds_obd *mds = &obd->u.mds;
2332 rc = obd_llog_setup(obd, &obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT,
2333 obd, 0, NULL, &llog_lvfs_ops);
2337 if (mds->mds_profile) {
2338 struct llog_ctxt *lgctxt;
2339 struct lvfs_run_ctxt saved;
2340 struct lustre_profile *lprof;
2341 struct config_llog_instance cfg;
2343 cfg.cfg_instance = NULL;
2344 cfg.cfg_uuid = mds->mds_lov_uuid;
2345 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2347 lgctxt = llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT);
2349 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2350 GOTO(err_llog, rc = -EINVAL);
2353 rc = class_config_process_llog(lgctxt, mds->mds_profile, &cfg);
2354 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2359 lprof = class_get_profile(mds->mds_profile);
2360 if (lprof == NULL) {
2361 CERROR("No profile found: %s\n", mds->mds_profile);
2362 GOTO(err_cleanup, rc = -ENOENT);
2364 rc = mds_lov_connect(obd, lprof->lp_osc);
2366 GOTO(err_cleanup, rc);
2368 rc = mds_lmv_postsetup(obd);
2370 GOTO(err_cleanup, rc);
2378 obd_llog_cleanup(llog_get_context(&obd->obd_llogs,
2379 LLOG_CONFIG_ORIG_CTXT));
2383 int mds_postrecov(struct obd_device *obd)
2385 struct mds_obd *mds = &obd->u.mds;
2386 struct llog_ctxt *ctxt;
2390 LASSERT(!obd->obd_recovering);
2391 ctxt = llog_get_context(&obd->obd_llogs, LLOG_UNLINK_ORIG_CTXT);
2392 LASSERT(ctxt != NULL);
2394 /* set nextid first, so we are sure it happens */
2395 rc = mds_lov_set_nextid(obd);
2397 CERROR("%s: mds_lov_set_nextid failed\n", obd->obd_name);
2401 /* clean PENDING dir */
2402 rc = mds_cleanup_orphans(obd);
2407 rc = llog_connect(ctxt, obd->u.mds.mds_lov_desc.ld_tgt_count,
2410 CERROR("%s: failed at llog_origin_connect: %d\n",
2415 /* remove the orphaned precreated objects */
2416 rc = mds_lov_clearorphans(mds, NULL /* all OSTs */);
2421 RETURN(rc < 0 ? rc : item);
2424 /* cleanup all llogging subsystems */
2425 rc = obd_llog_finish(obd, &obd->obd_llogs,
2426 mds->mds_lov_desc.ld_tgt_count);
2428 CERROR("%s: failed to cleanup llogging subsystems\n",
2433 int mds_lov_clean(struct obd_device *obd)
2435 struct mds_obd *mds = &obd->u.mds;
2437 if (mds->mds_profile) {
2439 struct config_llog_instance cfg;
2440 struct lvfs_run_ctxt saved;
2441 int len = strlen(mds->mds_profile) + sizeof("-clean") + 1;
2443 OBD_ALLOC(cln_prof, len);
2444 sprintf(cln_prof, "%s-clean", mds->mds_profile);
2446 cfg.cfg_instance = NULL;
2447 cfg.cfg_uuid = mds->mds_lov_uuid;
2449 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2450 class_config_process_llog(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT),
2452 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2454 OBD_FREE(cln_prof, len);
2455 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2456 mds->mds_profile = NULL;
2461 int mds_lmv_clean(struct obd_device *obd)
2463 struct mds_obd *mds = &obd->u.mds;
2465 if (mds->mds_lmv_name) {
2466 OBD_FREE(mds->mds_lmv_name, strlen(mds->mds_lmv_name) + 1);
2467 mds->mds_lmv_name = NULL;
2472 static int mds_precleanup(struct obd_device *obd, int flags)
2478 mds_lov_disconnect(obd, flags);
2480 obd_llog_cleanup(llog_get_context(&obd->obd_llogs, LLOG_CONFIG_ORIG_CTXT));
2484 static int mds_cleanup(struct obd_device *obd, int flags)
2486 struct mds_obd *mds = &obd->u.mds;
2489 if (mds->mds_sb == NULL)
2492 mds_update_server_data(obd, 1);
2493 if (mds->mds_lov_objids != NULL) {
2494 OBD_FREE(mds->mds_lov_objids,
2495 mds->mds_lov_desc.ld_tgt_count * sizeof(obd_id));
2497 mds_fs_cleanup(obd, flags);
2501 /* 2 seems normal on mds, (may_umount() also expects 2
2502 fwiw), but we only see 1 at this point in obdfilter. */
2503 if (atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count) > 2)
2504 CERROR("%s: mount busy, mnt_count %d != 2\n", obd->obd_name,
2505 atomic_read(&obd->u.mds.mds_vfsmnt->mnt_count));
2507 mntput(mds->mds_vfsmnt);
2511 ldlm_namespace_free(obd->obd_namespace, flags & OBD_OPT_FORCE);
2513 if (mds->mds_counters) {
2514 lprocfs_free_mds_counters(mds->mds_counters);
2517 spin_lock_bh(&obd->obd_processing_task_lock);
2518 if (obd->obd_recovering) {
2519 target_cancel_recovery_timer(obd);
2520 obd->obd_recovering = 0;
2522 spin_unlock_bh(&obd->obd_processing_task_lock);
2525 dev_clear_rdonly(2);
2526 fsfilt_put_ops(obd->obd_fsops);
2531 static void fixup_handle_for_resent_req(struct ptlrpc_request *req,
2533 struct ldlm_lock *new_lock,
2534 struct lustre_handle *lockh)
2536 struct obd_export *exp = req->rq_export;
2537 struct obd_device *obd = exp->exp_obd;
2538 struct ldlm_request *dlmreq =
2539 lustre_msg_buf(req->rq_reqmsg, offset, sizeof (*dlmreq));
2540 struct lustre_handle remote_hdl = dlmreq->lock_handle1;
2541 struct list_head *iter;
2543 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2546 l_lock(&obd->obd_namespace->ns_lock);
2547 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2548 struct ldlm_lock *lock;
2549 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2550 if (lock == new_lock)
2552 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2553 lockh->cookie = lock->l_handle.h_cookie;
2554 DEBUG_REQ(D_HA, req, "restoring lock cookie "LPX64,
2556 l_unlock(&obd->obd_namespace->ns_lock);
2560 l_unlock(&obd->obd_namespace->ns_lock);
2562 /* If the xid matches, then we know this is a resent request,
2563 * and allow it. (It's probably an OPEN, for which we don't
2565 if (req->rq_xid == exp->exp_mds_data.med_mcd->mcd_last_xid)
2568 /* This remote handle isn't enqueued, so we never received or
2569 * processed this request. Clear MSG_RESENT, because it can
2570 * be handled like any normal request now. */
2572 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2574 DEBUG_REQ(D_HA, req, "no existing lock with rhandle "LPX64,
2578 int intent_disposition(struct ldlm_reply *rep, int flag)
2582 return (rep->lock_policy_res1 & flag);
2585 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2589 rep->lock_policy_res1 |= flag;
2592 static int mds_intent_policy(struct ldlm_namespace *ns,
2593 struct ldlm_lock **lockp, void *req_cookie,
2594 ldlm_mode_t mode, int flags, void *data)
2596 struct ptlrpc_request *req = req_cookie;
2597 struct ldlm_lock *lock = *lockp;
2598 struct ldlm_intent *it;
2599 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2600 struct ldlm_reply *rep;
2601 struct lustre_handle lockh[2] = {{0}, {0}};
2602 struct ldlm_lock *new_lock;
2603 int getattr_part = MDS_INODELOCK_UPDATE;
2604 int rc, repsize[4] = { sizeof(struct ldlm_reply),
2605 sizeof(struct mds_body),
2606 mds->mds_max_mdsize,
2607 mds->mds_max_cookiesize };
2608 int offset = MDS_REQ_INTENT_REC_OFF;
2611 LASSERT(req != NULL);
2613 if (req->rq_reqmsg->bufcount <= MDS_REQ_INTENT_IT_OFF) {
2614 /* No intent was provided */
2615 int size = sizeof(struct ldlm_reply);
2616 rc = lustre_pack_reply(req, 1, &size, NULL);
2621 it = lustre_swab_reqbuf(req, MDS_REQ_INTENT_IT_OFF, sizeof(*it),
2622 lustre_swab_ldlm_intent);
2624 CERROR("Intent missing\n");
2625 RETURN(req->rq_status = -EFAULT);
2628 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2630 rc = lustre_pack_reply(req, 3, repsize, NULL);
2632 RETURN(req->rq_status = rc);
2634 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof (*rep));
2635 intent_set_disposition(rep, DISP_IT_EXECD);
2637 fixup_handle_for_resent_req(req, MDS_REQ_INTENT_LOCKREQ_OFF,
2640 /* execute policy */
2641 switch ((long)it->opc) {
2643 case IT_CREAT|IT_OPEN:
2644 /* XXX swab here to assert that an mds_open reint
2645 * packet is following */
2646 rep->lock_policy_res2 = mds_reint(req, offset, lockh);
2648 /* We abort the lock if the lookup was negative and
2649 * we did not make it to the OPEN portion */
2650 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2651 RETURN(ELDLM_LOCK_ABORTED);
2652 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2653 !intent_disposition(rep, DISP_OPEN_OPEN))
2655 /* IT_OPEN may return lock on cross-node dentry
2656 * that we want to hold during attr retrival -bzzz */
2657 if (rc != 0 || lockh[0].cookie == 0)
2658 RETURN(ELDLM_LOCK_ABORTED);
2661 getattr_part = MDS_INODELOCK_LOOKUP;
2664 getattr_part |= MDS_INODELOCK_LOOKUP;
2666 rep->lock_policy_res2 = mds_getattr_name(req, offset, lockh,
2668 /* FIXME: LDLM can set req->rq_status. MDS sets
2669 policy_res{1,2} with disposition and status.
2670 - replay: returns 0 & req->status is old status
2671 - otherwise: returns req->status */
2672 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2673 rep->lock_policy_res2 = 0;
2674 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2675 rep->lock_policy_res2)
2676 RETURN(ELDLM_LOCK_ABORTED);
2677 if (req->rq_status != 0) {
2679 rep->lock_policy_res2 = req->rq_status;
2680 RETURN(ELDLM_LOCK_ABORTED);
2684 rc = mds_lock_and_check_slave(offset, req, lockh);
2685 if ((rep->lock_policy_res2 = rc)) {
2687 rep->lock_policy_res2 = 0;
2688 RETURN(ELDLM_LOCK_ABORTED);
2692 CERROR("Unhandled intent "LPD64"\n", it->opc);
2696 /* By this point, whatever function we called above must have either
2697 * filled in 'lockh', been an intent replay, or returned an error. We
2698 * want to allow replayed RPCs to not get a lock, since we would just
2699 * drop it below anyways because lock replay is done separately by the
2700 * client afterwards. For regular RPCs we want to give the new lock to
2701 * the client instead of whatever lock it was about to get. */
2702 new_lock = ldlm_handle2lock(&lockh[0]);
2703 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2706 LASSERT(new_lock != NULL);
2708 /* If we've already given this lock to a client once, then we should
2709 * have no readers or writers. Otherwise, we should have one reader
2710 * _or_ writer ref (which will be zeroed below) before returning the
2711 * lock to a client. */
2712 if (new_lock->l_export == req->rq_export) {
2713 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2715 LASSERT(new_lock->l_export == NULL);
2716 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2721 if (new_lock->l_export == req->rq_export) {
2722 /* Already gave this to the client, which means that we
2723 * reconstructed a reply. */
2724 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2726 RETURN(ELDLM_LOCK_REPLACED);
2729 /* Fixup the lock to be given to the client */
2730 l_lock(&new_lock->l_resource->lr_namespace->ns_lock);
2731 new_lock->l_readers = 0;
2732 new_lock->l_writers = 0;
2734 new_lock->l_export = class_export_get(req->rq_export);
2735 list_add(&new_lock->l_export_chain,
2736 &new_lock->l_export->exp_ldlm_data.led_held_locks);
2738 new_lock->l_blocking_ast = lock->l_blocking_ast;
2739 new_lock->l_completion_ast = lock->l_completion_ast;
2741 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2742 sizeof(lock->l_remote_handle));
2744 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2746 LDLM_LOCK_PUT(new_lock);
2747 l_unlock(&new_lock->l_resource->lr_namespace->ns_lock);
2749 RETURN(ELDLM_LOCK_REPLACED);
2752 int mds_attach(struct obd_device *dev, obd_count len, void *data)
2754 struct lprocfs_static_vars lvars;
2756 lprocfs_init_multi_vars(0, &lvars);
2757 return lprocfs_obd_attach(dev, lvars.obd_vars);
2760 int mds_detach(struct obd_device *dev)
2762 return lprocfs_obd_detach(dev);
2765 int mdt_attach(struct obd_device *dev, obd_count len, void *data)
2767 struct lprocfs_static_vars lvars;
2769 lprocfs_init_multi_vars(1, &lvars);
2770 return lprocfs_obd_attach(dev, lvars.obd_vars);
2773 int mdt_detach(struct obd_device *dev)
2775 return lprocfs_obd_detach(dev);
2778 static int mdt_setup(struct obd_device *obd, obd_count len, void *buf)
2780 struct mds_obd *mds = &obd->u.mds;
2785 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2786 MDS_REQUEST_PORTAL, MDC_REPLY_PORTAL,
2787 mds_handle, "mds", obd->obd_proc_entry);
2789 if (!mds->mds_service) {
2790 CERROR("failed to start service\n");
2794 rc = ptlrpc_start_n_threads(obd, mds->mds_service, MDT_NUM_THREADS,
2797 GOTO(err_thread, rc);
2799 mds->mds_setattr_service =
2800 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2801 MDS_SETATTR_PORTAL, MDC_REPLY_PORTAL,
2802 mds_handle, "mds_setattr",
2803 obd->obd_proc_entry);
2804 if (!mds->mds_setattr_service) {
2805 CERROR("failed to start getattr service\n");
2806 GOTO(err_thread, rc = -ENOMEM);
2809 rc = ptlrpc_start_n_threads(obd, mds->mds_setattr_service,
2810 MDT_NUM_THREADS, "ll_mdt_attr");
2812 GOTO(err_thread2, rc);
2814 mds->mds_readpage_service =
2815 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2816 MDS_READPAGE_PORTAL, MDC_REPLY_PORTAL,
2817 mds_handle, "mds_readpage",
2818 obd->obd_proc_entry);
2819 if (!mds->mds_readpage_service) {
2820 CERROR("failed to start readpage service\n");
2821 GOTO(err_thread2, rc = -ENOMEM);
2824 rc = ptlrpc_start_n_threads(obd, mds->mds_readpage_service,
2825 MDT_NUM_THREADS, "ll_mdt_rdpg");
2828 GOTO(err_thread3, rc);
2833 ptlrpc_unregister_service(mds->mds_readpage_service);
2835 ptlrpc_unregister_service(mds->mds_setattr_service);
2837 ptlrpc_unregister_service(mds->mds_service);
2841 static int mdt_cleanup(struct obd_device *obd, int flags)
2843 struct mds_obd *mds = &obd->u.mds;
2846 ptlrpc_stop_all_threads(mds->mds_readpage_service);
2847 ptlrpc_unregister_service(mds->mds_readpage_service);
2849 ptlrpc_stop_all_threads(mds->mds_setattr_service);
2850 ptlrpc_unregister_service(mds->mds_setattr_service);
2852 ptlrpc_stop_all_threads(mds->mds_service);
2853 ptlrpc_unregister_service(mds->mds_service);
2858 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2861 struct obd_device *obd = data;
2864 fid.generation = gen;
2865 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2868 static int mds_get_info(struct obd_export *exp, __u32 keylen,
2869 void *key, __u32 *vallen, void *val)
2871 struct obd_device *obd;
2872 struct mds_obd *mds;
2875 obd = class_exp2obd(exp);
2877 CDEBUG(D_IOCTL, "invalid client cookie "LPX64"\n",
2878 exp->exp_handle.h_cookie);
2882 if (keylen >= strlen("reint_log") && memcmp(key, "reint_log", 9) == 0) {
2883 /*Get log_context handle*/
2884 unsigned long *llh_handle = val;
2885 *vallen = sizeof(unsigned long);
2886 *llh_handle = (unsigned long)obd->obd_llog_ctxt[LLOG_REINT_ORIG_CTXT];
2889 if (keylen >= strlen("cache_sb") && memcmp(key, "cache_sb", 8) == 0) {
2890 /*Get log_context handle*/
2891 unsigned long *sb = val;
2892 *vallen = sizeof(unsigned long);
2893 *sb = (unsigned long)obd->u.mds.mds_sb;
2898 if (keylen >= strlen("mdsize") && memcmp(key, "mdsize", keylen) == 0) {
2899 __u32 *mdsize = val;
2900 *vallen = sizeof(*mdsize);
2901 *mdsize = mds->mds_max_mdsize;
2905 CDEBUG(D_IOCTL, "invalid key\n");
2909 struct lvfs_callback_ops mds_lvfs_ops = {
2910 l_fid2dentry: mds_lvfs_fid2dentry,
2913 int mds_preprw(int cmd, struct obd_export *exp, struct obdo *oa,
2914 int objcount, struct obd_ioobj *obj,
2915 int niocount, struct niobuf_remote *nb,
2916 struct niobuf_local *res,
2917 struct obd_trans_info *oti);
2918 int mds_commitrw(int cmd, struct obd_export *exp, struct obdo *oa,
2919 int objcount, struct obd_ioobj *obj, int niocount,
2920 struct niobuf_local *res, struct obd_trans_info *oti,
2923 /* use obd ops to offer management infrastructure */
2924 static struct obd_ops mds_obd_ops = {
2925 .o_owner = THIS_MODULE,
2926 .o_attach = mds_attach,
2927 .o_detach = mds_detach,
2928 .o_connect = mds_connect,
2929 .o_init_export = mds_init_export,
2930 .o_destroy_export = mds_destroy_export,
2931 .o_disconnect = mds_disconnect,
2932 .o_setup = mds_setup,
2933 .o_precleanup = mds_precleanup,
2934 .o_cleanup = mds_cleanup,
2935 .o_postrecov = mds_postrecov,
2936 .o_statfs = mds_obd_statfs,
2937 .o_iocontrol = mds_iocontrol,
2938 .o_create = mds_obd_create,
2939 .o_destroy = mds_obd_destroy,
2940 .o_llog_init = mds_llog_init,
2941 .o_llog_finish = mds_llog_finish,
2942 .o_notify = mds_notify,
2943 .o_get_info = mds_get_info,
2944 .o_set_info = mds_set_info,
2945 .o_preprw = mds_preprw,
2946 .o_commitrw = mds_commitrw,
2949 static struct obd_ops mdt_obd_ops = {
2950 .o_owner = THIS_MODULE,
2951 .o_attach = mdt_attach,
2952 .o_detach = mdt_detach,
2953 .o_setup = mdt_setup,
2954 .o_cleanup = mdt_cleanup,
2955 .o_attach = mdt_attach,
2956 .o_detach = mdt_detach,
2959 static int __init mds_init(void)
2961 struct lprocfs_static_vars lvars;
2963 mds_group_hash_init();
2965 lprocfs_init_multi_vars(0, &lvars);
2966 class_register_type(&mds_obd_ops, NULL, lvars.module_vars,
2968 lprocfs_init_multi_vars(1, &lvars);
2969 class_register_type(&mdt_obd_ops, NULL, lvars.module_vars,
2975 static void /*__exit*/ mds_exit(void)
2977 mds_group_hash_cleanup();
2979 class_unregister_type(LUSTRE_MDS_NAME);
2980 class_unregister_type(LUSTRE_MDT_NAME);
2983 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2984 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2985 MODULE_LICENSE("GPL");
2987 module_init(mds_init);
2988 module_exit(mds_exit);