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) 2001-2005 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 the Lustre file system, http://www.lustre.org
14 * Lustre is a trademark of Cluster File Systems, Inc.
16 * You may have signed or agreed to another license before downloading
17 * this software. If so, you are bound by the terms and conditions
18 * of that agreement, and the following does not apply to you. See the
19 * LICENSE file included with this distribution for more information.
21 * If you did not agree to a different license, then this copy of Lustre
22 * is open source software; you can redistribute it and/or modify it
23 * under the terms of version 2 of the GNU General Public License as
24 * published by the Free Software Foundation.
26 * In either case, Lustre is distributed in the hope that it will be
27 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
28 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
29 * license text for more details.
33 # define EXPORT_SYMTAB
35 #define DEBUG_SUBSYSTEM S_MDS
37 #include <lustre_mds.h>
38 #include <linux/module.h>
39 #include <linux/init.h>
40 #include <linux/random.h>
42 #include <linux/jbd.h>
43 #include <linux/smp_lock.h>
44 #include <linux/buffer_head.h>
45 #include <linux/workqueue.h>
46 #include <linux/mount.h>
48 #include <linux/lustre_acl.h>
49 #include <obd_class.h>
50 #include <lustre_dlm.h>
52 #include <lustre_fsfilt.h>
53 #include <lprocfs_status.h>
54 #include <lustre_commit_confd.h>
55 #include <lustre_quota.h>
56 #include <lustre_disk.h>
57 #include <lustre_param.h>
59 #include "mds_internal.h"
62 CFS_MODULE_PARM(mds_num_threads, "i", int, 0444,
63 "number of MDS service threads to start");
65 static int mds_intent_policy(struct ldlm_namespace *ns,
66 struct ldlm_lock **lockp, void *req_cookie,
67 ldlm_mode_t mode, int flags, void *data);
68 static int mds_postsetup(struct obd_device *obd);
69 static int mds_cleanup(struct obd_device *obd);
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 & ~CFS_PAGE_MASK) == 0); /* I'm dubious about this */
83 npages = (count + CFS_PAGE_SIZE - 1) >> CFS_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 > CFS_PAGE_SIZE ? CFS_PAGE_SIZE : tmpcount;
96 OBD_PAGE_ALLOC(pages[i], CFS_ALLOC_STD);
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 > CFS_PAGE_SIZE ? CFS_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 i_size_read(file->f_dentry->d_inode));
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",
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 class_fail_export(req->rq_export);
151 ptlrpc_abort_bulk (desc);
153 for (i = 0; i < npages; i++)
155 OBD_PAGE_FREE(pages[i]);
157 ptlrpc_free_bulk(desc);
159 OBD_FREE(pages, sizeof(*pages) * npages);
164 /* only valid locked dentries or errors should be returned */
165 struct dentry *mds_fid2locked_dentry(struct obd_device *obd, struct ll_fid *fid,
166 struct vfsmount **mnt, int lock_mode,
167 struct lustre_handle *lockh,
170 struct mds_obd *mds = &obd->u.mds;
171 struct dentry *de = mds_fid2dentry(mds, fid, mnt), *retval = de;
172 struct ldlm_res_id res_id = { .name = {0} };
173 int flags = LDLM_FL_ATOMIC_CB, rc;
174 ldlm_policy_data_t policy = { .l_inodebits = { lockpart} };
180 res_id.name[0] = de->d_inode->i_ino;
181 res_id.name[1] = de->d_inode->i_generation;
182 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id,
183 LDLM_IBITS, &policy, lock_mode, &flags,
184 ldlm_blocking_ast, ldlm_completion_ast,
185 NULL, NULL, 0, NULL, lockh);
186 if (rc != ELDLM_OK) {
188 retval = ERR_PTR(-EIO); /* XXX translate ldlm code */
194 /* Look up an entry by inode number. */
195 /* this function ONLY returns valid dget'd dentries with an initialized inode
197 struct dentry *mds_fid2dentry(struct mds_obd *mds, struct ll_fid *fid,
198 struct vfsmount **mnt)
201 unsigned long ino = fid->id;
202 __u32 generation = fid->generation;
204 struct dentry *result;
207 RETURN(ERR_PTR(-ESTALE));
209 snprintf(fid_name, sizeof(fid_name), "0x%lx", ino);
211 CDEBUG(D_DENTRY, "--> mds_fid2dentry: ino/gen %lu/%u, sb %p\n",
212 ino, generation, mds->mds_obt.obt_sb);
214 /* under ext3 this is neither supposed to return bad inodes
216 result = ll_lookup_one_len(fid_name, mds->mds_fid_de, strlen(fid_name));
220 inode = result->d_inode;
222 RETURN(ERR_PTR(-ENOENT));
224 if (inode->i_generation == 0 || inode->i_nlink == 0) {
225 LCONSOLE_WARN("Found inode with zero generation or link -- this"
226 " may indicate disk corruption (inode: %lu/%u, "
227 "link %lu, count %d)\n", inode->i_ino,
228 inode->i_generation,(unsigned long)inode->i_nlink,
229 atomic_read(&inode->i_count));
231 RETURN(ERR_PTR(-ENOENT));
234 if (generation && inode->i_generation != generation) {
235 /* we didn't find the right inode.. */
236 CDEBUG(D_INODE, "found wrong generation: inode %lu, link: %lu, "
237 "count: %d, generation %u/%u\n", inode->i_ino,
238 (unsigned long)inode->i_nlink,
239 atomic_read(&inode->i_count), inode->i_generation,
242 RETURN(ERR_PTR(-ENOENT));
246 *mnt = mds->mds_vfsmnt;
253 static int mds_connect_internal(struct obd_export *exp,
254 struct obd_connect_data *data)
256 struct obd_device *obd = exp->exp_obd;
258 data->ocd_connect_flags &= MDT_CONNECT_SUPPORTED;
259 data->ocd_ibits_known &= MDS_INODELOCK_FULL;
261 /* If no known bits (which should not happen, probably,
262 as everybody should support LOOKUP and UPDATE bits at least)
263 revert to compat mode with plain locks. */
264 if (!data->ocd_ibits_known &&
265 data->ocd_connect_flags & OBD_CONNECT_IBITS)
266 data->ocd_connect_flags &= ~OBD_CONNECT_IBITS;
268 if (!obd->u.mds.mds_fl_acl)
269 data->ocd_connect_flags &= ~OBD_CONNECT_ACL;
271 if (!obd->u.mds.mds_fl_user_xattr)
272 data->ocd_connect_flags &= ~OBD_CONNECT_XATTR;
274 exp->exp_connect_flags = data->ocd_connect_flags;
275 data->ocd_version = LUSTRE_VERSION_CODE;
276 exp->exp_mds_data.med_ibits_known = data->ocd_ibits_known;
279 if (obd->u.mds.mds_fl_acl &&
280 ((exp->exp_connect_flags & OBD_CONNECT_ACL) == 0)) {
281 CWARN("%s: MDS requires ACL support but client does not\n",
288 static int mds_reconnect(struct obd_export *exp, struct obd_device *obd,
289 struct obd_uuid *cluuid,
290 struct obd_connect_data *data)
295 if (exp == NULL || obd == NULL || cluuid == NULL)
298 rc = mds_connect_internal(exp, data);
303 /* Establish a connection to the MDS.
305 * This will set up an export structure for the client to hold state data
306 * about that client, like open files, the last operation number it did
307 * on the server, etc.
309 static int mds_connect(const struct lu_env *env,
310 struct lustre_handle *conn, struct obd_device *obd,
311 struct obd_uuid *cluuid, struct obd_connect_data *data)
313 struct obd_export *exp;
314 struct mds_export_data *med;
315 struct mds_client_data *mcd = NULL;
319 if (!conn || !obd || !cluuid)
322 /* XXX There is a small race between checking the list and adding a
323 * new connection for the same UUID, but the real threat (list
324 * corruption when multiple different clients connect) is solved.
326 * There is a second race between adding the export to the list,
327 * and filling in the client data below. Hence skipping the case
328 * of NULL mcd above. We should already be controlling multiple
329 * connects at the client, and we can't hold the spinlock over
330 * memory allocations without risk of deadlocking.
332 rc = class_connect(conn, obd, cluuid);
335 exp = class_conn2export(conn);
337 med = &exp->exp_mds_data;
339 rc = mds_connect_internal(exp, data);
343 OBD_ALLOC(mcd, sizeof(*mcd));
345 GOTO(out, rc = -ENOMEM);
347 memcpy(mcd->mcd_uuid, cluuid, sizeof(mcd->mcd_uuid));
350 rc = mds_client_add(obd, exp, -1);
356 OBD_FREE(mcd, sizeof(*mcd));
359 class_disconnect(exp);
361 class_export_put(exp);
367 int mds_init_export(struct obd_export *exp)
369 struct mds_export_data *med = &exp->exp_mds_data;
371 INIT_LIST_HEAD(&med->med_open_head);
372 spin_lock_init(&med->med_open_lock);
374 spin_lock(&exp->exp_lock);
375 exp->exp_connecting = 1;
376 spin_unlock(&exp->exp_lock);
381 static int mds_destroy_export(struct obd_export *export)
383 struct mds_export_data *med;
384 struct obd_device *obd = export->exp_obd;
385 struct mds_obd *mds = &obd->u.mds;
386 struct lvfs_run_ctxt saved;
387 struct lov_mds_md *lmm;
388 struct llog_cookie *logcookies;
392 med = &export->exp_mds_data;
393 target_destroy_export(export);
395 if (obd_uuid_equals(&export->exp_client_uuid, &obd->obd_uuid))
398 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
399 /* Close any open files (which may also cause orphan unlinking). */
401 OBD_ALLOC(lmm, mds->mds_max_mdsize);
403 CWARN("%s: allocation failure during cleanup; can not force "
404 "close file handles on this service.\n", obd->obd_name);
405 GOTO(out, rc = -ENOMEM);
408 OBD_ALLOC(logcookies, mds->mds_max_cookiesize);
409 if (logcookies == NULL) {
410 CWARN("%s: allocation failure during cleanup; can not force "
411 "close file handles on this service.\n", obd->obd_name);
412 OBD_FREE(lmm, mds->mds_max_mdsize);
413 GOTO(out_lmm, rc = -ENOMEM);
416 spin_lock(&med->med_open_lock);
417 while (!list_empty(&med->med_open_head)) {
418 struct list_head *tmp = med->med_open_head.next;
419 struct mds_file_data *mfd =
420 list_entry(tmp, struct mds_file_data, mfd_list);
421 int lmm_size = mds->mds_max_mdsize;
422 umode_t mode = mfd->mfd_dentry->d_inode->i_mode;
425 /* Remove mfd handle so it can't be found again.
426 * We are consuming the mfd_list reference here. */
427 mds_mfd_unlink(mfd, 0);
428 spin_unlock(&med->med_open_lock);
430 /* If you change this message, be sure to update
431 * replay_single:test_46 */
432 CDEBUG(D_INODE|D_IOCTL, "%s: force closing file handle for "
433 "%.*s (ino %lu)\n", obd->obd_name,
434 mfd->mfd_dentry->d_name.len,mfd->mfd_dentry->d_name.name,
435 mfd->mfd_dentry->d_inode->i_ino);
437 rc = mds_get_md(obd, mfd->mfd_dentry->d_inode, lmm, &lmm_size, 1);
439 CWARN("mds_get_md failure, rc=%d\n", rc);
441 valid |= OBD_MD_FLEASIZE;
443 /* child orphan sem protects orphan_dec_test and
444 * is_orphan race, mds_mfd_close drops it */
445 MDS_DOWN_WRITE_ORPHAN_SEM(mfd->mfd_dentry->d_inode);
446 rc = mds_mfd_close(NULL, REQ_REC_OFF, obd, mfd,
447 !(export->exp_flags & OBD_OPT_FAILOVER),
448 lmm, lmm_size, logcookies,
449 mds->mds_max_cookiesize,
453 CDEBUG(D_INODE|D_IOCTL, "Error closing file: %d\n", rc);
455 if (valid & OBD_MD_FLCOOKIE) {
456 rc = mds_osc_destroy_orphan(obd, mode, lmm,
457 lmm_size, logcookies, 1);
459 CDEBUG(D_INODE, "%s: destroy of orphan failed,"
460 " rc = %d\n", obd->obd_name, rc);
463 valid &= ~OBD_MD_FLCOOKIE;
466 spin_lock(&med->med_open_lock);
468 spin_unlock(&med->med_open_lock);
470 OBD_FREE(logcookies, mds->mds_max_cookiesize);
472 OBD_FREE(lmm, mds->mds_max_mdsize);
474 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
475 mds_client_free(export);
479 static int mds_disconnect(struct obd_export *exp)
485 class_export_get(exp);
487 /* Disconnect early so that clients can't keep using export */
488 rc = class_disconnect(exp);
489 if (exp->exp_obd->obd_namespace != NULL)
490 ldlm_cancel_locks_for_export(exp);
492 /* complete all outstanding replies */
493 spin_lock(&exp->exp_lock);
494 while (!list_empty(&exp->exp_outstanding_replies)) {
495 struct ptlrpc_reply_state *rs =
496 list_entry(exp->exp_outstanding_replies.next,
497 struct ptlrpc_reply_state, rs_exp_list);
498 struct ptlrpc_service *svc = rs->rs_service;
500 spin_lock(&svc->srv_lock);
501 list_del_init(&rs->rs_exp_list);
502 ptlrpc_schedule_difficult_reply(rs);
503 spin_unlock(&svc->srv_lock);
505 spin_unlock(&exp->exp_lock);
507 class_export_put(exp);
511 static int mds_getstatus(struct ptlrpc_request *req)
513 struct mds_obd *mds = mds_req2mds(req);
514 struct mds_body *body;
515 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
518 rc = lustre_pack_reply(req, 2, size, NULL);
519 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_GETSTATUS_PACK)) {
520 CERROR("mds: out of memory for message\n");
521 req->rq_status = -ENOMEM; /* superfluous? */
525 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*body));
526 memcpy(&body->fid1, &mds->mds_rootfid, sizeof(body->fid1));
528 /* the last_committed and last_xid fields are filled in for all
529 * replies already - no need to do so here also.
534 /* get the LOV EA from @inode and store it into @md. It can be at most
535 * @size bytes, and @size is updated with the actual EA size.
536 * The EA size is also returned on success, and -ve errno on failure.
537 * If there is no EA then 0 is returned. */
538 int mds_get_md(struct obd_device *obd, struct inode *inode, void *md,
545 LOCK_INODE_MUTEX(inode);
546 rc = fsfilt_get_md(obd, inode, md, *size, "lov");
549 CERROR("Error %d reading eadata for ino %lu\n",
553 rc = mds_convert_lov_ea(obd, inode, md, lmm_size);
565 UNLOCK_INODE_MUTEX(inode);
571 /* Call with lock=1 if you want mds_pack_md to take the i_mutex.
572 * Call with lock=0 if the caller has already taken the i_mutex. */
573 int mds_pack_md(struct obd_device *obd, struct lustre_msg *msg, int offset,
574 struct mds_body *body, struct inode *inode, int lock)
576 struct mds_obd *mds = &obd->u.mds;
582 lmm = lustre_msg_buf(msg, offset, 0);
584 /* Some problem with getting eadata when I sized the reply
586 CDEBUG(D_INFO, "no space reserved for inode %lu MD\n",
590 lmm_size = lustre_msg_buflen(msg, offset);
592 /* I don't really like this, but it is a sanity check on the client
593 * MD request. However, if the client doesn't know how much space
594 * to reserve for the MD, it shouldn't be bad to have too much space.
596 if (lmm_size > mds->mds_max_mdsize) {
597 CWARN("Reading MD for inode %lu of %d bytes > max %d\n",
598 inode->i_ino, lmm_size, mds->mds_max_mdsize);
602 rc = mds_get_md(obd, inode, lmm, &lmm_size, lock);
604 if (S_ISDIR(inode->i_mode))
605 body->valid |= OBD_MD_FLDIREA;
607 body->valid |= OBD_MD_FLEASIZE;
608 body->eadatasize = lmm_size;
615 #ifdef CONFIG_FS_POSIX_ACL
617 int mds_pack_posix_acl(struct inode *inode, struct lustre_msg *repmsg,
618 struct mds_body *repbody, int repoff)
620 struct dentry de = { .d_inode = inode };
624 LASSERT(repbody->aclsize == 0);
625 LASSERT(lustre_msg_bufcount(repmsg) > repoff);
627 buflen = lustre_msg_buflen(repmsg, repoff);
631 if (!inode->i_op || !inode->i_op->getxattr)
634 rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
635 lustre_msg_buf(repmsg, repoff, buflen),
639 repbody->aclsize = rc;
640 else if (rc != -ENODATA) {
641 CERROR("buflen %d, get acl: %d\n", buflen, rc);
646 repbody->valid |= OBD_MD_FLACL;
650 #define mds_pack_posix_acl(inode, repmsg, repbody, repoff) 0
653 int mds_pack_acl(struct mds_export_data *med, struct inode *inode,
654 struct lustre_msg *repmsg, struct mds_body *repbody,
657 return mds_pack_posix_acl(inode, repmsg, repbody, repoff);
660 static int mds_getattr_internal(struct obd_device *obd, struct dentry *dentry,
661 struct ptlrpc_request *req,
662 struct mds_body *reqbody, int reply_off)
664 struct mds_body *body;
665 struct inode *inode = dentry->d_inode;
672 body = lustre_msg_buf(req->rq_repmsg, reply_off, sizeof(*body));
673 LASSERT(body != NULL); /* caller prepped reply */
675 mds_pack_inode2fid(&body->fid1, inode);
676 body->flags = reqbody->flags; /* copy MDS_BFLAG_EXT_FLAGS if present */
677 mds_pack_inode2body(body, inode);
680 if ((S_ISREG(inode->i_mode) && (reqbody->valid & OBD_MD_FLEASIZE)) ||
681 (S_ISDIR(inode->i_mode) && (reqbody->valid & OBD_MD_FLDIREA))) {
682 rc = mds_pack_md(obd, req->rq_repmsg, reply_off, body,
685 /* If we have LOV EA data, the OST holds size, atime, mtime */
686 if (!(body->valid & OBD_MD_FLEASIZE) &&
687 !(body->valid & OBD_MD_FLDIREA))
688 body->valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
689 OBD_MD_FLATIME | OBD_MD_FLMTIME);
691 lustre_shrink_reply(req, reply_off, body->eadatasize, 0);
692 if (body->eadatasize)
694 } else if (S_ISLNK(inode->i_mode) &&
695 (reqbody->valid & OBD_MD_LINKNAME) != 0) {
696 char *symname = lustre_msg_buf(req->rq_repmsg, reply_off, 0);
699 LASSERT (symname != NULL); /* caller prepped reply */
700 len = lustre_msg_buflen(req->rq_repmsg, reply_off);
702 rc = inode->i_op->readlink(dentry, symname, len);
704 CERROR("readlink failed: %d\n", rc);
705 } else if (rc != len - 1) {
706 CERROR ("Unexpected readlink rc %d: expecting %d\n",
710 CDEBUG(D_INODE, "read symlink dest %s\n", symname);
711 body->valid |= OBD_MD_LINKNAME;
712 body->eadatasize = rc + 1;
713 symname[rc] = 0; /* NULL terminate */
717 } else if (reqbody->valid == OBD_MD_FLFLAGS &&
718 reqbody->flags & MDS_BFLAG_EXT_FLAGS) {
721 /* We only return the full set of flags on ioctl, otherwise we
722 * get enough flags from the inode in mds_pack_inode2body(). */
723 rc = fsfilt_iocontrol(obd, inode, NULL, EXT3_IOC_GETFLAGS,
726 body->flags = flags | MDS_BFLAG_EXT_FLAGS;
729 if (reqbody->valid & OBD_MD_FLMODEASIZE) {
730 struct mds_obd *mds = mds_req2mds(req);
731 body->max_cookiesize = mds->mds_max_cookiesize;
732 body->max_mdsize = mds->mds_max_mdsize;
733 body->valid |= OBD_MD_FLMODEASIZE;
739 #ifdef CONFIG_FS_POSIX_ACL
740 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
741 (reqbody->valid & OBD_MD_FLACL)) {
742 rc = mds_pack_acl(&req->rq_export->exp_mds_data,
743 inode, req->rq_repmsg,
746 lustre_shrink_reply(req, reply_off, body->aclsize, 0);
755 static int mds_getattr_pack_msg(struct ptlrpc_request *req, struct inode *inode,
758 struct mds_obd *mds = mds_req2mds(req);
759 struct mds_body *body;
760 int rc, bufcount = 2;
761 int size[4] = { sizeof(struct ptlrpc_body), sizeof(*body) };
764 LASSERT(offset == REQ_REC_OFF); /* non-intent */
766 body = lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*body));
767 LASSERT(body != NULL); /* checked by caller */
768 LASSERT(lustre_req_swabbed(req, offset)); /* swabbed by caller */
770 if ((S_ISREG(inode->i_mode) && (body->valid & OBD_MD_FLEASIZE)) ||
771 (S_ISDIR(inode->i_mode) && (body->valid & OBD_MD_FLDIREA))) {
772 LOCK_INODE_MUTEX(inode);
773 rc = fsfilt_get_md(req->rq_export->exp_obd, inode, NULL, 0,
775 UNLOCK_INODE_MUTEX(inode);
776 CDEBUG(D_INODE, "got %d bytes MD data for inode %lu\n",
779 if (rc != -ENODATA) {
780 CERROR("error getting inode %lu MD: rc = %d\n",
785 } else if (rc > mds->mds_max_mdsize) {
787 CERROR("MD size %d larger than maximum possible %u\n",
788 rc, mds->mds_max_mdsize);
793 } else if (S_ISLNK(inode->i_mode) && (body->valid & OBD_MD_LINKNAME)) {
794 if (i_size_read(inode) + 1 != body->eadatasize)
795 CERROR("symlink size: %Lu, reply space: %d\n",
796 i_size_read(inode) + 1, body->eadatasize);
797 size[bufcount] = min_t(int, i_size_read(inode) + 1,
800 CDEBUG(D_INODE, "symlink size: %Lu, reply space: %d\n",
801 i_size_read(inode) + 1, body->eadatasize);
804 #ifdef CONFIG_FS_POSIX_ACL
805 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
806 (body->valid & OBD_MD_FLACL)) {
807 struct dentry de = { .d_inode = inode };
810 if (inode->i_op && inode->i_op->getxattr) {
811 rc = inode->i_op->getxattr(&de, MDS_XATTR_NAME_ACL_ACCESS,
815 if (rc != -ENODATA) {
816 CERROR("got acl size: %d\n", rc);
826 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_GETATTR_PACK)) {
827 CERROR("failed MDS_GETATTR_PACK test\n");
828 req->rq_status = -ENOMEM;
832 rc = lustre_pack_reply(req, bufcount, size, NULL);
834 CERROR("lustre_pack_reply failed: rc %d\n", rc);
842 static int mds_getattr_lock(struct ptlrpc_request *req, int offset,
843 int child_part, struct lustre_handle *child_lockh)
845 struct obd_device *obd = req->rq_export->exp_obd;
846 struct mds_obd *mds = &obd->u.mds;
847 struct ldlm_reply *rep = NULL;
848 struct lvfs_run_ctxt saved;
849 struct mds_body *body;
850 struct dentry *dparent = NULL, *dchild = NULL;
851 struct lvfs_ucred uc = {0,};
852 struct lustre_handle parent_lockh;
854 int rc = 0, cleanup_phase = 0, resent_req = 0;
858 LASSERT(!strcmp(obd->obd_type->typ_name, LUSTRE_MDS_NAME));
860 /* Swab now, before anyone looks inside the request */
861 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
862 lustre_swab_mds_body);
864 CERROR("Can't swab mds_body\n");
868 lustre_set_req_swabbed(req, offset + 1);
869 name = lustre_msg_string(req->rq_reqmsg, offset + 1, 0);
871 CERROR("Can't unpack name\n");
874 namesize = lustre_msg_buflen(req->rq_reqmsg, offset + 1);
875 /* namesize less than 2 means we have empty name, probably came from
876 revalidate by cfid, so no point in having name to be set */
880 rc = mds_init_ucred(&uc, req, offset);
884 LASSERT(offset == REQ_REC_OFF || offset == DLM_INTENT_REC_OFF);
885 /* if requests were at offset 2, the getattr reply goes back at 1 */
886 if (offset == DLM_INTENT_REC_OFF) {
887 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF,
889 offset = DLM_REPLY_REC_OFF;
892 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
893 cleanup_phase = 1; /* kernel context */
894 intent_set_disposition(rep, DISP_LOOKUP_EXECD);
896 /* FIXME: handle raw lookup */
898 if (body->valid == OBD_MD_FLID) {
899 struct mds_body *mds_reply;
900 int size = sizeof(*mds_reply);
902 // The user requested ONLY the inode number, so do a raw lookup
903 rc = lustre_pack_reply(req, 1, &size, NULL);
905 CERROR("out of memory\n");
909 rc = dir->i_op->lookup_raw(dir, name, namesize - 1, &inum);
911 mds_reply = lustre_msg_buf(req->rq_repmsg, offset,
913 mds_reply->fid1.id = inum;
914 mds_reply->valid = OBD_MD_FLID;
919 if (lustre_handle_is_used(child_lockh)) {
920 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
924 if (resent_req == 0) {
926 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RESEND, obd_timeout * 2);
927 rc = mds_get_parent_child_locked(obd, &obd->u.mds,
931 MDS_INODELOCK_UPDATE,
933 child_lockh, &dchild,
936 /* For revalidate by fid we always take UPDATE lock */
937 dchild = mds_fid2locked_dentry(obd, &body->fid2, NULL,
942 rc = PTR_ERR(dchild);
947 struct ldlm_lock *granted_lock;
948 struct ll_fid child_fid;
949 struct ldlm_resource *res;
950 DEBUG_REQ(D_DLMTRACE, req, "resent, not enqueuing new locks");
951 granted_lock = ldlm_handle2lock(child_lockh);
952 LASSERTF(granted_lock != NULL, LPU64"/%u lockh "LPX64"\n",
953 body->fid1.id, body->fid1.generation,
954 child_lockh->cookie);
957 res = granted_lock->l_resource;
958 child_fid.id = res->lr_name.name[0];
959 child_fid.generation = res->lr_name.name[1];
960 dchild = mds_fid2dentry(&obd->u.mds, &child_fid, NULL);
961 LASSERT(!IS_ERR(dchild));
962 LDLM_LOCK_PUT(granted_lock);
965 cleanup_phase = 2; /* dchild, dparent, locks */
967 if (dchild->d_inode == NULL) {
968 intent_set_disposition(rep, DISP_LOOKUP_NEG);
969 /* in the intent case, the policy clears this error:
970 the disposition is enough */
971 GOTO(cleanup, rc = -ENOENT);
973 intent_set_disposition(rep, DISP_LOOKUP_POS);
976 if (req->rq_repmsg == NULL) {
977 rc = mds_getattr_pack_msg(req, dchild->d_inode, offset);
979 CERROR ("mds_getattr_pack_msg: %d\n", rc);
984 rc = mds_getattr_internal(obd, dchild, req, body, offset);
985 GOTO(cleanup, rc); /* returns the lock to the client */
988 switch (cleanup_phase) {
990 if (resent_req == 0) {
991 if (rc && dchild->d_inode)
992 ldlm_lock_decref(child_lockh, LCK_CR);
994 ldlm_lock_decref(&parent_lockh, LCK_CR);
1000 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1002 mds_exit_ucred(&uc, mds);
1003 if (req->rq_reply_state == NULL) {
1004 req->rq_status = rc;
1005 lustre_pack_reply(req, 1, NULL, NULL);
1011 static int mds_getattr(struct ptlrpc_request *req, int offset)
1013 struct mds_obd *mds = mds_req2mds(req);
1014 struct obd_device *obd = req->rq_export->exp_obd;
1015 struct lvfs_run_ctxt saved;
1017 struct mds_body *body;
1018 struct lvfs_ucred uc = {0,};
1022 OBD_COUNTER_INCREMENT(obd, getattr);
1024 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1025 lustre_swab_mds_body);
1029 rc = mds_init_ucred(&uc, req, offset);
1031 GOTO(out_ucred, rc);
1033 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1034 de = mds_fid2dentry(mds, &body->fid1, NULL);
1036 rc = req->rq_status = PTR_ERR(de);
1040 rc = mds_getattr_pack_msg(req, de->d_inode, offset);
1042 CERROR("mds_getattr_pack_msg: %d\n", rc);
1046 req->rq_status = mds_getattr_internal(obd, de, req, body,
1052 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1054 if (req->rq_reply_state == NULL) {
1055 req->rq_status = rc;
1056 lustre_pack_reply(req, 1, NULL, NULL);
1058 mds_exit_ucred(&uc, mds);
1062 static int mds_obd_statfs(struct obd_device *obd, struct obd_statfs *osfs,
1067 spin_lock(&obd->obd_osfs_lock);
1068 rc = fsfilt_statfs(obd, obd->u.obt.obt_sb, max_age);
1070 memcpy(osfs, &obd->obd_osfs, sizeof(*osfs));
1071 spin_unlock(&obd->obd_osfs_lock);
1076 static int mds_statfs(struct ptlrpc_request *req)
1078 struct obd_device *obd = req->rq_export->exp_obd;
1079 int rc, size[2] = { sizeof(struct ptlrpc_body),
1080 sizeof(struct obd_statfs) };
1083 /* This will trigger a watchdog timeout */
1084 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_STATFS_LCW_SLEEP,
1085 (MDS_SERVICE_WATCHDOG_TIMEOUT / 1000) + 1);
1086 OBD_COUNTER_INCREMENT(obd, statfs);
1088 rc = lustre_pack_reply(req, 2, size, NULL);
1089 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_STATFS_PACK)) {
1090 CERROR("mds: statfs lustre_pack_reply failed: rc = %d\n", rc);
1094 /* We call this so that we can cache a bit - 1 jiffie worth */
1095 rc = mds_obd_statfs(obd, lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1096 size[REPLY_REC_OFF]),
1097 cfs_time_current_64() - HZ);
1099 CERROR("mds_obd_statfs failed: rc %d\n", rc);
1105 req->rq_status = rc;
1109 static int mds_sync(struct ptlrpc_request *req, int offset)
1111 struct obd_device *obd = req->rq_export->exp_obd;
1112 struct mds_obd *mds = &obd->u.mds;
1113 struct mds_body *body;
1114 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
1117 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1118 lustre_swab_mds_body);
1120 GOTO(out, rc = -EFAULT);
1122 rc = lustre_pack_reply(req, 2, size, NULL);
1123 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_SYNC_PACK)) {
1124 CERROR("fsync lustre_pack_reply failed: rc = %d\n", rc);
1128 if (body->fid1.id == 0) {
1129 /* a fid of zero is taken to mean "sync whole filesystem" */
1130 rc = fsfilt_sync(obd, obd->u.obt.obt_sb);
1135 de = mds_fid2dentry(mds, &body->fid1, NULL);
1137 GOTO(out, rc = PTR_ERR(de));
1139 /* The file parameter isn't used for anything */
1140 if (de->d_inode->i_fop && de->d_inode->i_fop->fsync)
1141 rc = de->d_inode->i_fop->fsync(NULL, de, 1);
1143 body = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1145 mds_pack_inode2fid(&body->fid1, de->d_inode);
1146 mds_pack_inode2body(body, de->d_inode);
1153 req->rq_status = rc;
1157 /* mds_readpage does not take a DLM lock on the inode, because the client must
1158 * already have a PR lock.
1160 * If we were to take another one here, a deadlock will result, if another
1161 * thread is already waiting for a PW lock. */
1162 static int mds_readpage(struct ptlrpc_request *req, int offset)
1164 struct obd_device *obd = req->rq_export->exp_obd;
1165 struct mds_obd *mds = &obd->u.mds;
1166 struct vfsmount *mnt;
1169 struct mds_body *body, *repbody;
1170 struct lvfs_run_ctxt saved;
1171 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
1172 struct lvfs_ucred uc = {0,};
1175 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_READPAGE_PACK))
1178 rc = lustre_pack_reply(req, 2, size, NULL);
1180 CERROR("error packing readpage reply: rc %d\n", rc);
1184 body = lustre_swab_reqbuf(req, offset, sizeof(*body),
1185 lustre_swab_mds_body);
1187 GOTO (out, rc = -EFAULT);
1189 rc = mds_init_ucred(&uc, req, offset);
1193 push_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1194 de = mds_fid2dentry(&obd->u.mds, &body->fid1, &mnt);
1196 GOTO(out_pop, rc = PTR_ERR(de));
1198 CDEBUG(D_INODE, "ino %lu\n", de->d_inode->i_ino);
1200 file = dentry_open(de, mnt, O_RDONLY | O_LARGEFILE);
1201 /* note: in case of an error, dentry_open puts dentry */
1203 GOTO(out_pop, rc = PTR_ERR(file));
1205 /* body->size is actually the offset -eeb */
1206 if ((body->size & (de->d_inode->i_sb->s_blocksize - 1)) != 0) {
1207 CERROR("offset "LPU64" not on a block boundary of %lu\n",
1208 body->size, de->d_inode->i_sb->s_blocksize);
1209 GOTO(out_file, rc = -EFAULT);
1212 /* body->nlink is actually the #bytes to read -eeb */
1213 if (body->nlink & (de->d_inode->i_sb->s_blocksize - 1)) {
1214 CERROR("size %u is not multiple of blocksize %lu\n",
1215 body->nlink, de->d_inode->i_sb->s_blocksize);
1216 GOTO(out_file, rc = -EFAULT);
1219 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1221 repbody->size = i_size_read(file->f_dentry->d_inode);
1222 repbody->valid = OBD_MD_FLSIZE;
1224 /* to make this asynchronous make sure that the handling function
1225 doesn't send a reply when this function completes. Instead a
1226 callback function would send the reply */
1227 /* body->size is actually the offset -eeb */
1228 rc = mds_sendpage(req, file, body->size, body->nlink);
1231 filp_close(file, 0);
1233 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, &uc);
1235 mds_exit_ucred(&uc, mds);
1236 req->rq_status = rc;
1240 int mds_reint(struct ptlrpc_request *req, int offset,
1241 struct lustre_handle *lockh)
1243 struct mds_update_record *rec; /* 116 bytes on the stack? no sir! */
1246 OBD_ALLOC(rec, sizeof(*rec));
1250 rc = mds_update_unpack(req, offset, rec);
1251 if (rc || OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNPACK)) {
1252 CERROR("invalid record\n");
1253 GOTO(out, req->rq_status = -EINVAL);
1256 /* rc will be used to interrupt a for loop over multiple records */
1257 rc = mds_reint_rec(rec, offset, req, lockh);
1259 OBD_FREE(rec, sizeof(*rec));
1263 int mds_filter_recovery_request(struct ptlrpc_request *req,
1264 struct obd_device *obd, int *process)
1266 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1267 case MDS_CONNECT: /* This will never get here, but for completeness. */
1268 case OST_CONNECT: /* This will never get here, but for completeness. */
1269 case MDS_DISCONNECT:
1270 case OST_DISCONNECT:
1275 case MDS_DONE_WRITING:
1276 case MDS_SYNC: /* used in unmounting */
1282 *process = target_queue_recovery_request(req, obd);
1286 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1291 EXPORT_SYMBOL(mds_filter_recovery_request);
1293 static char *reint_names[] = {
1294 [REINT_SETATTR] "setattr",
1295 [REINT_CREATE] "create",
1296 [REINT_LINK] "link",
1297 [REINT_UNLINK] "unlink",
1298 [REINT_RENAME] "rename",
1299 [REINT_OPEN] "open",
1302 static int mds_set_info_rpc(struct obd_export *exp, struct ptlrpc_request *req)
1305 int keylen, vallen, rc = 0;
1308 key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1);
1310 DEBUG_REQ(D_HA, req, "no set_info key");
1313 keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF);
1315 val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, 0);
1316 vallen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1);
1318 rc = lustre_pack_reply(req, 1, NULL, NULL);
1321 lustre_msg_set_status(req->rq_repmsg, 0);
1323 if (KEY_IS("read-only")) {
1324 if (val == NULL || vallen < sizeof(__u32)) {
1325 DEBUG_REQ(D_HA, req, "no set_info val");
1330 exp->exp_connect_flags |= OBD_CONNECT_RDONLY;
1332 exp->exp_connect_flags &= ~OBD_CONNECT_RDONLY;
1340 static int mds_handle_quotacheck(struct ptlrpc_request *req)
1342 struct obd_quotactl *oqctl;
1346 oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1347 lustre_swab_obd_quotactl);
1351 rc = lustre_pack_reply(req, 1, NULL, NULL);
1353 CERROR("mds: out of memory while packing quotacheck reply\n");
1357 req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1361 static int mds_handle_quotactl(struct ptlrpc_request *req)
1363 struct obd_quotactl *oqctl, *repoqc;
1364 int rc, size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) };
1367 oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1368 lustre_swab_obd_quotactl);
1372 rc = lustre_pack_reply(req, 2, size, NULL);
1376 repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc));
1378 req->rq_status = obd_quotactl(req->rq_export, oqctl);
1383 int mds_msg_check_version(struct lustre_msg *msg)
1387 switch (lustre_msg_get_opc(msg)) {
1389 case MDS_DISCONNECT:
1392 case SEC_CTX_INIT_CONT:
1394 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1396 CERROR("bad opc %u version %08x, expecting %08x\n",
1397 lustre_msg_get_opc(msg),
1398 lustre_msg_get_version(msg),
1399 LUSTRE_OBD_VERSION);
1403 case MDS_GETATTR_NAME:
1410 case MDS_DONE_WRITING:
1416 case MDS_QUOTACHECK:
1422 rc = lustre_msg_check_version(msg, LUSTRE_MDS_VERSION);
1424 CERROR("bad opc %u version %08x, expecting %08x\n",
1425 lustre_msg_get_opc(msg),
1426 lustre_msg_get_version(msg),
1427 LUSTRE_MDS_VERSION);
1431 case LDLM_BL_CALLBACK:
1432 case LDLM_CP_CALLBACK:
1433 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1435 CERROR("bad opc %u version %08x, expecting %08x\n",
1436 lustre_msg_get_opc(msg),
1437 lustre_msg_get_version(msg),
1438 LUSTRE_DLM_VERSION);
1440 case OBD_LOG_CANCEL:
1441 case LLOG_ORIGIN_HANDLE_CREATE:
1442 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1443 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1444 case LLOG_ORIGIN_HANDLE_CLOSE:
1445 case LLOG_ORIGIN_HANDLE_DESTROY:
1446 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1448 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1450 CERROR("bad opc %u version %08x, expecting %08x\n",
1451 lustre_msg_get_opc(msg),
1452 lustre_msg_get_version(msg),
1453 LUSTRE_LOG_VERSION);
1456 CERROR("MDS unknown opcode %d\n", lustre_msg_get_opc(msg));
1461 EXPORT_SYMBOL(mds_msg_check_version);
1463 int mds_handle(struct ptlrpc_request *req)
1465 int should_process, fail = OBD_FAIL_MDS_ALL_REPLY_NET;
1467 struct mds_obd *mds = NULL; /* quell gcc overwarning */
1468 struct obd_device *obd = NULL;
1471 OBD_FAIL_RETURN(OBD_FAIL_MDS_ALL_REQUEST_NET | OBD_FAIL_ONCE, 0);
1473 LASSERT(current->journal_info == NULL);
1475 rc = mds_msg_check_version(req->rq_reqmsg);
1477 CERROR("MDS drop mal-formed request\n");
1481 /* XXX identical to OST */
1482 if (lustre_msg_get_opc(req->rq_reqmsg) != MDS_CONNECT) {
1483 struct mds_export_data *med;
1486 if (req->rq_export == NULL) {
1487 CERROR("operation %d on unconnected MDS from %s\n",
1488 lustre_msg_get_opc(req->rq_reqmsg),
1489 libcfs_id2str(req->rq_peer));
1490 req->rq_status = -ENOTCONN;
1491 GOTO(out, rc = -ENOTCONN);
1494 med = &req->rq_export->exp_mds_data;
1495 obd = req->rq_export->exp_obd;
1496 mds = mds_req2mds(req);
1498 /* sanity check: if the xid matches, the request must
1499 * be marked as a resent or replayed */
1500 if (req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_xid) ||
1501 req->rq_xid == le64_to_cpu(med->med_mcd->mcd_last_close_xid))
1502 if (!(lustre_msg_get_flags(req->rq_reqmsg) &
1503 (MSG_RESENT | MSG_REPLAY))) {
1504 CERROR("rq_xid "LPU64" matches last_xid, "
1505 "expected RESENT flag\n",
1507 req->rq_status = -ENOTCONN;
1508 GOTO(out, rc = -EFAULT);
1510 /* else: note the opposite is not always true; a
1511 * RESENT req after a failover will usually not match
1512 * the last_xid, since it was likely never
1513 * committed. A REPLAYed request will almost never
1514 * match the last xid, however it could for a
1515 * committed, but still retained, open. */
1517 /* Check for aborted recovery. */
1518 spin_lock_bh(&obd->obd_processing_task_lock);
1519 recovering = obd->obd_recovering;
1520 spin_unlock_bh(&obd->obd_processing_task_lock);
1522 rc = mds_filter_recovery_request(req, obd,
1524 if (rc || !should_process)
1526 else if (should_process < 0) {
1527 req->rq_status = should_process;
1528 rc = ptlrpc_error(req);
1534 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1536 DEBUG_REQ(D_INODE, req, "connect");
1537 OBD_FAIL_RETURN(OBD_FAIL_MDS_CONNECT_NET, 0);
1538 rc = target_handle_connect(req);
1540 /* Now that we have an export, set mds. */
1542 * XXX nikita: these assignments are useless: mds is
1543 * never used below, and obd is only used for
1544 * MSG_LAST_REPLAY case, which never happens for
1547 obd = req->rq_export->exp_obd;
1548 mds = mds_req2mds(req);
1552 case MDS_DISCONNECT:
1553 DEBUG_REQ(D_INODE, req, "disconnect");
1554 OBD_FAIL_RETURN(OBD_FAIL_MDS_DISCONNECT_NET, 0);
1555 rc = target_handle_disconnect(req);
1556 req->rq_status = rc; /* superfluous? */
1560 DEBUG_REQ(D_INODE, req, "getstatus");
1561 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETSTATUS_NET, 0);
1562 rc = mds_getstatus(req);
1566 DEBUG_REQ(D_INODE, req, "getattr");
1567 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NET, 0);
1568 rc = mds_getattr(req, REQ_REC_OFF);
1572 DEBUG_REQ(D_INODE, req, "setxattr");
1573 OBD_FAIL_RETURN(OBD_FAIL_MDS_SETXATTR_NET, 0);
1574 rc = mds_setxattr(req);
1578 DEBUG_REQ(D_INODE, req, "getxattr");
1579 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETXATTR_NET, 0);
1580 rc = mds_getxattr(req);
1583 case MDS_GETATTR_NAME: {
1584 struct lustre_handle lockh = { 0 };
1585 DEBUG_REQ(D_INODE, req, "getattr_name");
1586 OBD_FAIL_RETURN(OBD_FAIL_MDS_GETATTR_NAME_NET, 0);
1588 /* If this request gets a reconstructed reply, we won't be
1589 * acquiring any new locks in mds_getattr_lock, so we don't
1592 rc = mds_getattr_lock(req, REQ_REC_OFF, MDS_INODELOCK_UPDATE,
1594 /* this non-intent call (from an ioctl) is special */
1595 req->rq_status = rc;
1596 if (rc == 0 && lustre_handle_is_used(&lockh))
1597 ldlm_lock_decref(&lockh, LCK_CR);
1601 DEBUG_REQ(D_INODE, req, "statfs");
1602 OBD_FAIL_RETURN(OBD_FAIL_MDS_STATFS_NET, 0);
1603 rc = mds_statfs(req);
1607 DEBUG_REQ(D_INODE, req, "readpage");
1608 OBD_FAIL_RETURN(OBD_FAIL_MDS_READPAGE_NET, 0);
1609 rc = mds_readpage(req, REQ_REC_OFF);
1611 if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_SENDPAGE)) {
1618 __u32 *opcp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF,
1621 int size[4] = { sizeof(struct ptlrpc_body),
1622 sizeof(struct mds_body),
1623 mds->mds_max_mdsize,
1624 mds->mds_max_cookiesize };
1627 /* NB only peek inside req now; mds_reint() will swab it */
1629 CERROR ("Can't inspect opcode\n");
1634 if (lustre_msg_swabbed(req->rq_reqmsg))
1637 DEBUG_REQ(D_INODE, req, "reint %d (%s)", opc,
1638 (opc < sizeof(reint_names) / sizeof(reint_names[0]) ||
1639 reint_names[opc] == NULL) ? reint_names[opc] :
1642 OBD_FAIL_RETURN(OBD_FAIL_MDS_REINT_NET, 0);
1644 if (opc == REINT_UNLINK || opc == REINT_RENAME)
1646 else if (opc == REINT_OPEN)
1651 rc = lustre_pack_reply(req, bufcount, size, NULL);
1655 rc = mds_reint(req, REQ_REC_OFF, NULL);
1656 fail = OBD_FAIL_MDS_REINT_NET_REP;
1661 DEBUG_REQ(D_INODE, req, "close");
1662 OBD_FAIL_RETURN(OBD_FAIL_MDS_CLOSE_NET, 0);
1663 rc = mds_close(req, REQ_REC_OFF);
1666 case MDS_DONE_WRITING:
1667 DEBUG_REQ(D_INODE, req, "done_writing");
1668 OBD_FAIL_RETURN(OBD_FAIL_MDS_DONE_WRITING_NET, 0);
1669 rc = mds_done_writing(req, REQ_REC_OFF);
1673 DEBUG_REQ(D_INODE, req, "pin");
1674 OBD_FAIL_RETURN(OBD_FAIL_MDS_PIN_NET, 0);
1675 rc = mds_pin(req, REQ_REC_OFF);
1679 DEBUG_REQ(D_INODE, req, "sync");
1680 OBD_FAIL_RETURN(OBD_FAIL_MDS_SYNC_NET, 0);
1681 rc = mds_sync(req, REQ_REC_OFF);
1685 DEBUG_REQ(D_INODE, req, "set_info");
1686 rc = mds_set_info_rpc(req->rq_export, req);
1689 case MDS_QUOTACHECK:
1690 DEBUG_REQ(D_INODE, req, "quotacheck");
1691 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACHECK_NET, 0);
1692 rc = mds_handle_quotacheck(req);
1696 DEBUG_REQ(D_INODE, req, "quotactl");
1697 OBD_FAIL_RETURN(OBD_FAIL_MDS_QUOTACTL_NET, 0);
1698 rc = mds_handle_quotactl(req);
1702 DEBUG_REQ(D_INODE, req, "ping");
1703 rc = target_handle_ping(req);
1706 case OBD_LOG_CANCEL:
1707 CDEBUG(D_INODE, "log cancel\n");
1708 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOG_CANCEL_NET, 0);
1709 rc = -ENOTSUPP; /* la la la */
1713 DEBUG_REQ(D_INODE, req, "enqueue");
1714 OBD_FAIL_RETURN(OBD_FAIL_LDLM_ENQUEUE, 0);
1715 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1716 ldlm_server_blocking_ast, NULL);
1717 fail = OBD_FAIL_LDLM_REPLY;
1720 DEBUG_REQ(D_INODE, req, "convert");
1721 OBD_FAIL_RETURN(OBD_FAIL_LDLM_CONVERT, 0);
1722 rc = ldlm_handle_convert(req);
1724 case LDLM_BL_CALLBACK:
1725 case LDLM_CP_CALLBACK:
1726 DEBUG_REQ(D_INODE, req, "callback");
1727 CERROR("callbacks should not happen on MDS\n");
1729 OBD_FAIL_RETURN(OBD_FAIL_LDLM_BL_CALLBACK, 0);
1731 case LLOG_ORIGIN_HANDLE_CREATE:
1732 DEBUG_REQ(D_INODE, req, "llog_init");
1733 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1734 rc = llog_origin_handle_create(req);
1736 case LLOG_ORIGIN_HANDLE_DESTROY:
1737 DEBUG_REQ(D_INODE, req, "llog_init");
1738 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1739 rc = llog_origin_handle_destroy(req);
1741 case LLOG_ORIGIN_HANDLE_NEXT_BLOCK:
1742 DEBUG_REQ(D_INODE, req, "llog next block");
1743 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1744 rc = llog_origin_handle_next_block(req);
1746 case LLOG_ORIGIN_HANDLE_PREV_BLOCK:
1747 DEBUG_REQ(D_INODE, req, "llog prev block");
1748 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1749 rc = llog_origin_handle_prev_block(req);
1751 case LLOG_ORIGIN_HANDLE_READ_HEADER:
1752 DEBUG_REQ(D_INODE, req, "llog read header");
1753 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1754 rc = llog_origin_handle_read_header(req);
1756 case LLOG_ORIGIN_HANDLE_CLOSE:
1757 DEBUG_REQ(D_INODE, req, "llog close");
1758 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1759 rc = llog_origin_handle_close(req);
1762 DEBUG_REQ(D_INODE, req, "llog catinfo");
1763 OBD_FAIL_RETURN(OBD_FAIL_OBD_LOGD_NET, 0);
1764 rc = llog_catinfo(req);
1767 req->rq_status = -ENOTSUPP;
1768 rc = ptlrpc_error(req);
1772 LASSERT(current->journal_info == NULL);
1774 /* If we're DISCONNECTing, the mds_export_data is already freed */
1775 if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != MDS_DISCONNECT) {
1776 struct mds_export_data *med = &req->rq_export->exp_mds_data;
1778 /* I don't think last_xid is used for anyway, so I'm not sure
1779 if we need to care about last_close_xid here.*/
1780 lustre_msg_set_last_xid(req->rq_repmsg,
1781 le64_to_cpu(med->med_mcd->mcd_last_xid));
1783 target_committed_to_req(req);
1789 target_send_reply(req, rc, fail);
1793 /* Update the server data on disk. This stores the new mount_count and
1794 * also the last_rcvd value to disk. If we don't have a clean shutdown,
1795 * then the server last_rcvd value may be less than that of the clients.
1796 * This will alert us that we may need to do client recovery.
1798 * Also assumes for mds_last_transno that we are not modifying it (no locking).
1800 int mds_update_server_data(struct obd_device *obd, int force_sync)
1802 struct mds_obd *mds = &obd->u.mds;
1803 struct lr_server_data *lsd = mds->mds_server_data;
1804 struct file *filp = mds->mds_rcvd_filp;
1805 struct lvfs_run_ctxt saved;
1810 CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
1811 mds->mds_mount_count, mds->mds_last_transno);
1813 spin_lock(&mds->mds_transno_lock);
1814 lsd->lsd_last_transno = cpu_to_le64(mds->mds_last_transno);
1815 spin_unlock(&mds->mds_transno_lock);
1817 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1818 rc = fsfilt_write_record(obd, filp, lsd, sizeof(*lsd), &off,force_sync);
1819 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
1821 CERROR("error writing MDS server data: rc = %d\n", rc);
1825 static void fsoptions_to_mds_flags(struct mds_obd *mds, char *options)
1835 while (*p && *p != ',')
1839 if (len == sizeof("user_xattr") - 1 &&
1840 memcmp(options, "user_xattr", len) == 0) {
1841 mds->mds_fl_user_xattr = 1;
1842 LCONSOLE_INFO("Enabling user_xattr\n");
1843 } else if (len == sizeof("nouser_xattr") - 1 &&
1844 memcmp(options, "nouser_xattr", len) == 0) {
1845 mds->mds_fl_user_xattr = 0;
1846 LCONSOLE_INFO("Disabling user_xattr\n");
1847 } else if (len == sizeof("acl") - 1 &&
1848 memcmp(options, "acl", len) == 0) {
1849 #ifdef CONFIG_FS_POSIX_ACL
1850 mds->mds_fl_acl = 1;
1851 LCONSOLE_INFO("Enabling ACL\n");
1853 CWARN("ignoring unsupported acl mount option\n");
1855 } else if (len == sizeof("noacl") - 1 &&
1856 memcmp(options, "noacl", len) == 0) {
1857 #ifdef CONFIG_FS_POSIX_ACL
1858 mds->mds_fl_acl = 0;
1859 LCONSOLE_INFO("Disabling ACL\n");
1866 static int mds_lov_presetup (struct mds_obd *mds, struct lustre_cfg *lcfg)
1871 rc = llog_start_commit_thread();
1875 if (lcfg->lcfg_bufcount >= 4 && LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
1878 ll_generate_random_uuid(uuid);
1879 class_uuid_unparse(uuid, &mds->mds_lov_uuid);
1881 OBD_ALLOC(mds->mds_profile, LUSTRE_CFG_BUFLEN(lcfg, 3));
1882 if (mds->mds_profile == NULL)
1885 strncpy(mds->mds_profile, lustre_cfg_string(lcfg, 3),
1886 LUSTRE_CFG_BUFLEN(lcfg, 3));
1891 /* mount the file system (secretly). lustre_cfg parameters are:
1897 static int mds_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
1899 struct lprocfs_static_vars lvars;
1900 struct mds_obd *mds = &obd->u.mds;
1901 struct lustre_mount_info *lmi;
1902 struct vfsmount *mnt;
1903 struct lustre_sb_info *lsi;
1904 struct obd_uuid uuid;
1911 /* setup 1:/dev/loop/0 2:ext3 3:mdsA 4:errors=remount-ro,iopen_nopriv */
1913 CLASSERT(offsetof(struct obd_device, u.obt) ==
1914 offsetof(struct obd_device, u.mds.mds_obt));
1916 if (lcfg->lcfg_bufcount < 3)
1919 if (LUSTRE_CFG_BUFLEN(lcfg, 1) == 0 || LUSTRE_CFG_BUFLEN(lcfg, 2) == 0)
1922 lmi = server_get_mount(obd->obd_name);
1924 CERROR("Not mounted in lustre_fill_super?\n");
1928 /* We mounted in lustre_fill_super.
1929 lcfg bufs 1, 2, 4 (device, fstype, mount opts) are ignored.*/
1931 lsi = s2lsi(lmi->lmi_sb);
1932 fsoptions_to_mds_flags(mds, lsi->lsi_ldd->ldd_mount_opts);
1933 fsoptions_to_mds_flags(mds, lsi->lsi_lmd->lmd_opts);
1935 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
1936 if (IS_ERR(obd->obd_fsops))
1937 GOTO(err_put, rc = PTR_ERR(obd->obd_fsops));
1939 CDEBUG(D_SUPER, "%s: mnt = %p\n", lustre_cfg_string(lcfg, 1), mnt);
1941 LASSERT(!lvfs_check_rdonly(lvfs_sbdev(mnt->mnt_sb)));
1943 sema_init(&mds->mds_epoch_sem, 1);
1944 spin_lock_init(&mds->mds_transno_lock);
1945 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
1946 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
1947 mds->mds_atime_diff = MAX_ATIME_DIFF;
1948 mds->mds_evict_ost_nids = 1;
1950 sprintf(ns_name, "mds-%s", obd->obd_uuid.uuid);
1951 obd->obd_namespace = ldlm_namespace_new(ns_name, LDLM_NAMESPACE_SERVER,
1952 LDLM_NAMESPACE_GREEDY);
1953 if (obd->obd_namespace == NULL) {
1955 GOTO(err_ops, rc = -ENOMEM);
1957 ldlm_register_intent(obd->obd_namespace, mds_intent_policy);
1959 lprocfs_init_vars(mds, &lvars);
1960 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
1961 lprocfs_alloc_obd_stats(obd, LPROC_MDS_LAST) == 0) {
1962 /* Init private stats here */
1963 mds_stats_counter_init(obd->obd_stats);
1964 obd->obd_proc_exports = proc_mkdir("exports",
1965 obd->obd_proc_entry);
1968 rc = mds_fs_setup(obd, mnt);
1970 CERROR("%s: MDS filesystem method init failed: rc = %d\n",
1975 rc = mds_lov_presetup(mds, lcfg);
1979 ptlrpc_init_client(LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
1980 "mds_ldlm_client", &obd->obd_ldlm_client);
1981 obd->obd_replayable = 1;
1983 rc = lquota_setup(mds_quota_interface_ref, obd);
1988 mds->mds_group_hash = upcall_cache_init(obd->obd_name);
1989 if (IS_ERR(mds->mds_group_hash)) {
1990 rc = PTR_ERR(mds->mds_group_hash);
1991 mds->mds_group_hash = NULL;
1992 GOTO(err_qctxt, rc);
1996 /* Don't wait for mds_postrecov trying to clear orphans */
1997 obd->obd_async_recov = 1;
1998 rc = mds_postsetup(obd);
1999 /* Bug 11557 - allow async abort_recov start
2000 FIXME can remove most of this obd_async_recov plumbing
2001 obd->obd_async_recov = 0;
2004 GOTO(err_qctxt, rc);
2006 uuid_ptr = fsfilt_uuid(obd, obd->u.obt.obt_sb);
2007 if (uuid_ptr != NULL) {
2008 class_uuid_unparse(uuid_ptr, &uuid);
2014 label = fsfilt_get_label(obd, obd->u.obt.obt_sb);
2015 if (obd->obd_recovering) {
2016 LCONSOLE_WARN("MDT %s now serving %s (%s%s%s), but will be in "
2017 "recovery until %d %s reconnect, or if no clients"
2018 " reconnect for %d:%.02d; during that time new "
2019 "clients will not be allowed to connect. "
2020 "Recovery progress can be monitored by watching "
2021 "/proc/fs/lustre/mds/%s/recovery_status.\n",
2022 obd->obd_name, lustre_cfg_string(lcfg, 1),
2023 label ?: "", label ? "/" : "", str,
2024 obd->obd_max_recoverable_clients,
2025 (obd->obd_max_recoverable_clients == 1) ?
2026 "client" : "clients",
2027 (int)(OBD_RECOVERY_TIMEOUT) / 60,
2028 (int)(OBD_RECOVERY_TIMEOUT) % 60,
2031 LCONSOLE_INFO("MDT %s now serving %s (%s%s%s) with recovery "
2032 "%s\n", obd->obd_name, lustre_cfg_string(lcfg, 1),
2033 label ?: "", label ? "/" : "", str,
2034 obd->obd_replayable ? "enabled" : "disabled");
2037 if (ldlm_timeout == LDLM_TIMEOUT_DEFAULT)
2043 lquota_cleanup(mds_quota_interface_ref, obd);
2045 /* No extra cleanup needed for llog_init_commit_thread() */
2046 mds_fs_cleanup(obd);
2048 upcall_cache_cleanup(mds->mds_group_hash);
2049 mds->mds_group_hash = NULL;
2052 lprocfs_obd_cleanup(obd);
2053 lprocfs_free_obd_stats(obd);
2054 ldlm_namespace_free(obd->obd_namespace, 0);
2055 obd->obd_namespace = NULL;
2057 fsfilt_put_ops(obd->obd_fsops);
2059 server_put_mount(obd->obd_name, mnt);
2060 obd->u.obt.obt_sb = NULL;
2064 static int mds_lov_clean(struct obd_device *obd)
2066 struct mds_obd *mds = &obd->u.mds;
2067 struct obd_device *osc = mds->mds_osc_obd;
2070 if (mds->mds_profile) {
2071 class_del_profile(mds->mds_profile);
2072 OBD_FREE(mds->mds_profile, strlen(mds->mds_profile) + 1);
2073 mds->mds_profile = NULL;
2076 /* There better be a lov */
2080 RETURN(PTR_ERR(osc));
2082 obd_register_observer(osc, NULL);
2084 /* Give lov our same shutdown flags */
2085 osc->obd_force = obd->obd_force;
2086 osc->obd_fail = obd->obd_fail;
2088 /* Cleanup the lov */
2089 obd_disconnect(mds->mds_osc_exp);
2090 class_manual_cleanup(osc);
2091 mds->mds_osc_exp = NULL;
2096 static int mds_postsetup(struct obd_device *obd)
2098 struct mds_obd *mds = &obd->u.mds;
2102 rc = llog_setup(obd, NULL, LLOG_CONFIG_ORIG_CTXT, obd, 0, NULL,
2107 rc = llog_setup(obd, NULL, LLOG_LOVEA_ORIG_CTXT, obd, 0, NULL,
2112 if (mds->mds_profile) {
2113 struct lustre_profile *lprof;
2114 /* The profile defines which osc and mdc to connect to, for a
2115 client. We reuse that here to figure out the name of the
2116 lov to use (and ignore lprof->lp_md).
2117 The profile was set in the config log with
2118 LCFG_MOUNTOPT profilenm oscnm mdcnm */
2119 lprof = class_get_profile(mds->mds_profile);
2120 if (lprof == NULL) {
2121 CERROR("No profile found: %s\n", mds->mds_profile);
2122 GOTO(err_cleanup, rc = -ENOENT);
2124 rc = mds_lov_connect(obd, lprof->lp_dt);
2126 GOTO(err_cleanup, rc);
2133 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2134 llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2138 int mds_postrecov(struct obd_device *obd)
2146 LASSERT(!obd->obd_recovering);
2147 LASSERT(llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT) != NULL);
2149 /* set nextid first, so we are sure it happens */
2150 mutex_down(&obd->obd_dev_sem);
2151 rc = mds_lov_set_nextid(obd);
2152 mutex_up(&obd->obd_dev_sem);
2154 CERROR("%s: mds_lov_set_nextid failed %d\n",
2159 /* clean PENDING dir */
2160 if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2161 rc = mds_cleanup_pending(obd);
2165 /* FIXME Does target_finish_recovery really need this to block? */
2166 /* Notify the LOV, which will in turn call mds_notify for each tgt */
2167 /* This means that we have to hack obd_notify to think we're obd_set_up
2168 during mds_lov_connect. */
2169 obd_notify(obd->u.mds.mds_osc_obd, NULL,
2170 obd->obd_async_recov ? OBD_NOTIFY_SYNC_NONBLOCK :
2171 OBD_NOTIFY_SYNC, NULL);
2173 /* quota recovery */
2174 lquota_recovery(mds_quota_interface_ref, obd);
2180 /* We need to be able to stop an mds_lov_synchronize */
2181 static int mds_lov_early_clean(struct obd_device *obd)
2183 struct mds_obd *mds = &obd->u.mds;
2184 struct obd_device *osc = mds->mds_osc_obd;
2186 if (!osc || (!obd->obd_force && !obd->obd_fail))
2189 CDEBUG(D_HA, "abort inflight\n");
2190 return (obd_precleanup(osc, OBD_CLEANUP_EARLY));
2193 static int mds_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage)
2199 case OBD_CLEANUP_EARLY:
2201 case OBD_CLEANUP_EXPORTS:
2202 /*XXX Use this for mdd mds cleanup, so comment out
2203 *this target_cleanup_recovery for this tmp MDD MDS
2205 if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2206 target_cleanup_recovery(obd);
2207 mds_lov_early_clean(obd);
2209 case OBD_CLEANUP_SELF_EXP:
2210 mds_lov_disconnect(obd);
2212 llog_cleanup(llog_get_context(obd, LLOG_CONFIG_ORIG_CTXT));
2213 llog_cleanup(llog_get_context(obd, LLOG_LOVEA_ORIG_CTXT));
2214 rc = obd_llog_finish(obd, 0);
2216 case OBD_CLEANUP_OBD:
2222 static int mds_cleanup(struct obd_device *obd)
2224 struct mds_obd *mds = &obd->u.mds;
2225 lvfs_sbdev_type save_dev;
2228 if (obd->u.obt.obt_sb == NULL)
2230 save_dev = lvfs_sbdev(obd->u.obt.obt_sb);
2232 if (mds->mds_osc_exp)
2233 /* lov export was disconnected by mds_lov_clean;
2234 we just need to drop our ref */
2235 class_export_put(mds->mds_osc_exp);
2237 lprocfs_obd_cleanup(obd);
2238 lprocfs_free_obd_stats(obd);
2240 lquota_cleanup(mds_quota_interface_ref, obd);
2242 mds_update_server_data(obd, 1);
2243 if (mds->mds_lov_objids != NULL)
2244 OBD_FREE(mds->mds_lov_objids, mds->mds_lov_objids_size);
2245 mds_fs_cleanup(obd);
2248 upcall_cache_cleanup(mds->mds_group_hash);
2249 mds->mds_group_hash = NULL;
2252 server_put_mount(obd->obd_name, mds->mds_vfsmnt);
2253 obd->u.obt.obt_sb = NULL;
2255 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
2257 spin_lock_bh(&obd->obd_processing_task_lock);
2258 if (obd->obd_recovering) {
2259 target_cancel_recovery_timer(obd);
2260 obd->obd_recovering = 0;
2262 spin_unlock_bh(&obd->obd_processing_task_lock);
2264 fsfilt_put_ops(obd->obd_fsops);
2266 LCONSOLE_INFO("MDT %s has stopped.\n", obd->obd_name);
2271 static void fixup_handle_for_resent_req(struct ptlrpc_request *req, int offset,
2272 struct ldlm_lock *new_lock,
2273 struct ldlm_lock **old_lock,
2274 struct lustre_handle *lockh)
2276 struct obd_export *exp = req->rq_export;
2277 struct ldlm_request *dlmreq =
2278 lustre_msg_buf(req->rq_reqmsg, offset, sizeof(*dlmreq));
2279 struct lustre_handle remote_hdl = dlmreq->lock_handle[0];
2280 struct list_head *iter;
2282 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT))
2285 spin_lock(&exp->exp_ldlm_data.led_lock);
2286 list_for_each(iter, &exp->exp_ldlm_data.led_held_locks) {
2287 struct ldlm_lock *lock;
2288 lock = list_entry(iter, struct ldlm_lock, l_export_chain);
2289 if (lock == new_lock)
2291 if (lock->l_remote_handle.cookie == remote_hdl.cookie) {
2292 lockh->cookie = lock->l_handle.h_cookie;
2293 LDLM_DEBUG(lock, "restoring lock cookie");
2294 DEBUG_REQ(D_DLMTRACE, req,"restoring lock cookie "LPX64,
2297 *old_lock = LDLM_LOCK_GET(lock);
2298 spin_unlock(&exp->exp_ldlm_data.led_lock);
2302 spin_unlock(&exp->exp_ldlm_data.led_lock);
2304 /* If the xid matches, then we know this is a resent request,
2305 * and allow it. (It's probably an OPEN, for which we don't
2308 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_xid))
2312 le64_to_cpu(exp->exp_mds_data.med_mcd->mcd_last_close_xid))
2315 /* This remote handle isn't enqueued, so we never received or
2316 * processed this request. Clear MSG_RESENT, because it can
2317 * be handled like any normal request now. */
2319 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2321 DEBUG_REQ(D_DLMTRACE, req, "no existing lock with rhandle "LPX64,
2325 int intent_disposition(struct ldlm_reply *rep, int flag)
2329 return (rep->lock_policy_res1 & flag);
2332 void intent_set_disposition(struct ldlm_reply *rep, int flag)
2336 rep->lock_policy_res1 |= flag;
2339 static int mds_intent_policy(struct ldlm_namespace *ns,
2340 struct ldlm_lock **lockp, void *req_cookie,
2341 ldlm_mode_t mode, int flags, void *data)
2343 struct ptlrpc_request *req = req_cookie;
2344 struct ldlm_lock *lock = *lockp;
2345 struct ldlm_intent *it;
2346 struct mds_obd *mds = &req->rq_export->exp_obd->u.mds;
2347 struct ldlm_reply *rep;
2348 struct lustre_handle lockh = { 0 };
2349 struct ldlm_lock *new_lock = NULL;
2350 int getattr_part = MDS_INODELOCK_UPDATE;
2351 int repsize[5] = { [MSG_PTLRPC_BODY_OFF] = sizeof(struct ptlrpc_body),
2352 [DLM_LOCKREPLY_OFF] = sizeof(struct ldlm_reply),
2353 [DLM_REPLY_REC_OFF] = sizeof(struct mds_body),
2354 [DLM_REPLY_REC_OFF+1] = mds->mds_max_mdsize };
2355 int repbufcnt = 4, rc;
2358 LASSERT(req != NULL);
2360 if (lustre_msg_bufcount(req->rq_reqmsg) <= DLM_INTENT_IT_OFF) {
2361 /* No intent was provided */
2362 rc = lustre_pack_reply(req, 2, repsize, NULL);
2367 it = lustre_swab_reqbuf(req, DLM_INTENT_IT_OFF, sizeof(*it),
2368 lustre_swab_ldlm_intent);
2370 CERROR("Intent missing\n");
2371 RETURN(req->rq_status = -EFAULT);
2374 LDLM_DEBUG(lock, "intent policy, opc: %s", ldlm_it2str(it->opc));
2376 if ((req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) &&
2377 (it->opc & (IT_OPEN | IT_GETATTR | IT_LOOKUP)))
2378 /* we should never allow OBD_CONNECT_ACL if not configured */
2379 repsize[repbufcnt++] = LUSTRE_POSIX_ACL_MAX_SIZE;
2380 else if (it->opc & IT_UNLINK)
2381 repsize[repbufcnt++] = mds->mds_max_cookiesize;
2383 rc = lustre_pack_reply(req, repbufcnt, repsize, NULL);
2385 RETURN(req->rq_status = rc);
2387 rep = lustre_msg_buf(req->rq_repmsg, DLM_LOCKREPLY_OFF, sizeof(*rep));
2388 intent_set_disposition(rep, DISP_IT_EXECD);
2391 /* execute policy */
2392 switch ((long)it->opc) {
2394 case IT_CREAT|IT_OPEN:
2395 mds_counter_incr(req->rq_export, LPROC_MDS_OPEN);
2396 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock, NULL,
2398 /* XXX swab here to assert that an mds_open reint
2399 * packet is following */
2400 rep->lock_policy_res2 = mds_reint(req, DLM_INTENT_REC_OFF,
2403 /* We abort the lock if the lookup was negative and
2404 * we did not make it to the OPEN portion */
2405 if (!intent_disposition(rep, DISP_LOOKUP_EXECD))
2406 RETURN(ELDLM_LOCK_ABORTED);
2407 if (intent_disposition(rep, DISP_LOOKUP_NEG) &&
2408 !intent_disposition(rep, DISP_OPEN_OPEN))
2411 /* If there was an error of some sort or if we are not
2412 * returning any locks */
2413 if (rep->lock_policy_res2 ||
2414 !intent_disposition(rep, DISP_OPEN_LOCK))
2415 RETURN(ELDLM_LOCK_ABORTED);
2418 getattr_part = MDS_INODELOCK_LOOKUP;
2420 getattr_part |= MDS_INODELOCK_LOOKUP;
2421 OBD_COUNTER_INCREMENT(req->rq_export->exp_obd, getattr);
2423 fixup_handle_for_resent_req(req, DLM_LOCKREQ_OFF, lock,
2426 /* INODEBITS_INTEROP: if this lock was converted from a
2427 * plain lock (client does not support inodebits), then
2428 * child lock must be taken with both lookup and update
2429 * bits set for all operations.
2431 if (!(req->rq_export->exp_connect_flags & OBD_CONNECT_IBITS))
2432 getattr_part = MDS_INODELOCK_LOOKUP |
2433 MDS_INODELOCK_UPDATE;
2435 rep->lock_policy_res2 = mds_getattr_lock(req,DLM_INTENT_REC_OFF,
2436 getattr_part, &lockh);
2437 /* FIXME: LDLM can set req->rq_status. MDS sets
2438 policy_res{1,2} with disposition and status.
2439 - replay: returns 0 & req->status is old status
2440 - otherwise: returns req->status */
2441 if (intent_disposition(rep, DISP_LOOKUP_NEG))
2442 rep->lock_policy_res2 = 0;
2443 if (!intent_disposition(rep, DISP_LOOKUP_POS) ||
2444 rep->lock_policy_res2)
2445 RETURN(ELDLM_LOCK_ABORTED);
2446 if (req->rq_status != 0) {
2448 rep->lock_policy_res2 = req->rq_status;
2449 RETURN(ELDLM_LOCK_ABORTED);
2453 CERROR("Unhandled intent "LPD64"\n", it->opc);
2457 /* By this point, whatever function we called above must have either
2458 * filled in 'lockh', been an intent replay, or returned an error. We
2459 * want to allow replayed RPCs to not get a lock, since we would just
2460 * drop it below anyways because lock replay is done separately by the
2461 * client afterwards. For regular RPCs we want to give the new lock to
2462 * the client instead of whatever lock it was about to get. */
2463 if (new_lock == NULL)
2464 new_lock = ldlm_handle2lock(&lockh);
2465 if (new_lock == NULL && (flags & LDLM_FL_INTENT_ONLY))
2468 LASSERTF(new_lock != NULL, "op "LPX64" lockh "LPX64"\n",
2469 it->opc, lockh.cookie);
2471 /* If we've already given this lock to a client once, then we should
2472 * have no readers or writers. Otherwise, we should have one reader
2473 * _or_ writer ref (which will be zeroed below) before returning the
2474 * lock to a client. */
2475 if (new_lock->l_export == req->rq_export) {
2476 LASSERT(new_lock->l_readers + new_lock->l_writers == 0);
2478 LASSERT(new_lock->l_export == NULL);
2479 LASSERT(new_lock->l_readers + new_lock->l_writers == 1);
2484 if (new_lock->l_export == req->rq_export) {
2485 /* Already gave this to the client, which means that we
2486 * reconstructed a reply. */
2487 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
2489 RETURN(ELDLM_LOCK_REPLACED);
2492 /* Fixup the lock to be given to the client */
2493 lock_res_and_lock(new_lock);
2494 new_lock->l_readers = 0;
2495 new_lock->l_writers = 0;
2497 new_lock->l_export = class_export_get(req->rq_export);
2498 spin_lock(&req->rq_export->exp_ldlm_data.led_lock);
2499 list_add(&new_lock->l_export_chain,
2500 &new_lock->l_export->exp_ldlm_data.led_held_locks);
2501 spin_unlock(&req->rq_export->exp_ldlm_data.led_lock);
2503 new_lock->l_blocking_ast = lock->l_blocking_ast;
2504 new_lock->l_completion_ast = lock->l_completion_ast;
2506 memcpy(&new_lock->l_remote_handle, &lock->l_remote_handle,
2507 sizeof(lock->l_remote_handle));
2509 new_lock->l_flags &= ~LDLM_FL_LOCAL;
2511 unlock_res_and_lock(new_lock);
2512 LDLM_LOCK_PUT(new_lock);
2514 RETURN(ELDLM_LOCK_REPLACED);
2517 static int mdt_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2519 struct mds_obd *mds = &obd->u.mds;
2520 struct lprocfs_static_vars lvars;
2521 int mds_min_threads;
2522 int mds_max_threads;
2526 lprocfs_init_vars(mdt, &lvars);
2527 lprocfs_obd_setup(obd, lvars.obd_vars);
2529 sema_init(&mds->mds_health_sem, 1);
2531 if (mds_num_threads) {
2532 /* If mds_num_threads is set, it is the min and the max. */
2533 if (mds_num_threads > MDS_THREADS_MAX)
2534 mds_num_threads = MDS_THREADS_MAX;
2535 if (mds_num_threads < MDS_THREADS_MIN)
2536 mds_num_threads = MDS_THREADS_MIN;
2537 mds_max_threads = mds_min_threads = mds_num_threads;
2539 /* Base min threads on memory and cpus */
2540 mds_min_threads = num_possible_cpus() * num_physpages >>
2541 (27 - CFS_PAGE_SHIFT);
2542 if (mds_min_threads < MDS_THREADS_MIN)
2543 mds_min_threads = MDS_THREADS_MIN;
2544 /* Largest auto threads start value */
2545 if (mds_min_threads > 32)
2546 mds_min_threads = 32;
2547 mds_max_threads = min(MDS_THREADS_MAX, mds_min_threads * 4);
2551 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2552 MDS_MAXREPSIZE, MDS_REQUEST_PORTAL,
2553 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2554 mds_handle, LUSTRE_MDS_NAME,
2555 obd->obd_proc_entry, NULL,
2556 mds_min_threads, mds_max_threads, "ll_mdt", 0);
2558 if (!mds->mds_service) {
2559 CERROR("failed to start service\n");
2560 GOTO(err_lprocfs, rc = -ENOMEM);
2563 rc = ptlrpc_start_threads(obd, mds->mds_service);
2565 GOTO(err_thread, rc);
2567 mds->mds_setattr_service =
2568 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2569 MDS_MAXREPSIZE, MDS_SETATTR_PORTAL,
2570 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2571 mds_handle, "mds_setattr",
2572 obd->obd_proc_entry, NULL,
2573 mds_min_threads, mds_max_threads,
2575 if (!mds->mds_setattr_service) {
2576 CERROR("failed to start getattr service\n");
2577 GOTO(err_thread, rc = -ENOMEM);
2580 rc = ptlrpc_start_threads(obd, mds->mds_setattr_service);
2582 GOTO(err_thread2, rc);
2584 mds->mds_readpage_service =
2585 ptlrpc_init_svc(MDS_NBUFS, MDS_BUFSIZE, MDS_MAXREQSIZE,
2586 MDS_MAXREPSIZE, MDS_READPAGE_PORTAL,
2587 MDC_REPLY_PORTAL, MDS_SERVICE_WATCHDOG_TIMEOUT,
2588 mds_handle, "mds_readpage",
2589 obd->obd_proc_entry, NULL,
2590 MDS_THREADS_MIN_READPAGE, mds_max_threads,
2592 if (!mds->mds_readpage_service) {
2593 CERROR("failed to start readpage service\n");
2594 GOTO(err_thread2, rc = -ENOMEM);
2597 rc = ptlrpc_start_threads(obd, mds->mds_readpage_service);
2600 GOTO(err_thread3, rc);
2602 ping_evictor_start();
2607 ptlrpc_unregister_service(mds->mds_readpage_service);
2608 mds->mds_readpage_service = NULL;
2610 ptlrpc_unregister_service(mds->mds_setattr_service);
2611 mds->mds_setattr_service = NULL;
2613 ptlrpc_unregister_service(mds->mds_service);
2614 mds->mds_service = NULL;
2616 lprocfs_obd_cleanup(obd);
2620 static int mdt_cleanup(struct obd_device *obd)
2622 struct mds_obd *mds = &obd->u.mds;
2625 ping_evictor_stop();
2627 down(&mds->mds_health_sem);
2628 ptlrpc_unregister_service(mds->mds_readpage_service);
2629 ptlrpc_unregister_service(mds->mds_setattr_service);
2630 ptlrpc_unregister_service(mds->mds_service);
2631 mds->mds_readpage_service = NULL;
2632 mds->mds_setattr_service = NULL;
2633 mds->mds_service = NULL;
2634 up(&mds->mds_health_sem);
2636 lprocfs_obd_cleanup(obd);
2641 static int mdt_health_check(struct obd_device *obd)
2643 struct mds_obd *mds = &obd->u.mds;
2646 down(&mds->mds_health_sem);
2647 rc |= ptlrpc_service_health_check(mds->mds_readpage_service);
2648 rc |= ptlrpc_service_health_check(mds->mds_setattr_service);
2649 rc |= ptlrpc_service_health_check(mds->mds_service);
2650 up(&mds->mds_health_sem);
2653 * health_check to return 0 on healthy
2654 * and 1 on unhealthy.
2662 static struct dentry *mds_lvfs_fid2dentry(__u64 id, __u32 gen, __u64 gr,
2665 struct obd_device *obd = data;
2668 fid.generation = gen;
2669 return mds_fid2dentry(&obd->u.mds, &fid, NULL);
2672 static int mds_health_check(struct obd_device *obd)
2674 struct obd_device_target *odt = &obd->u.obt;
2675 #ifdef USE_HEALTH_CHECK_WRITE
2676 struct mds_obd *mds = &obd->u.mds;
2680 if (odt->obt_sb->s_flags & MS_RDONLY)
2683 #ifdef USE_HEALTH_CHECK_WRITE
2684 LASSERT(mds->mds_health_check_filp != NULL);
2685 rc |= !!lvfs_check_io_health(obd, mds->mds_health_check_filp);
2690 static int mds_process_config(struct obd_device *obd, obd_count len, void *buf)
2692 struct lustre_cfg *lcfg = buf;
2693 struct lprocfs_static_vars lvars;
2696 lprocfs_init_vars(mds, &lvars);
2698 rc = class_process_proc_param(PARAM_MDT, lvars.obd_vars, lcfg, obd);
2702 struct lvfs_callback_ops mds_lvfs_ops = {
2703 l_fid2dentry: mds_lvfs_fid2dentry,
2706 /* use obd ops to offer management infrastructure */
2707 static struct obd_ops mds_obd_ops = {
2708 .o_owner = THIS_MODULE,
2709 .o_connect = mds_connect,
2710 .o_reconnect = mds_reconnect,
2711 .o_init_export = mds_init_export,
2712 .o_destroy_export = mds_destroy_export,
2713 .o_disconnect = mds_disconnect,
2714 .o_setup = mds_setup,
2715 .o_precleanup = mds_precleanup,
2716 .o_cleanup = mds_cleanup,
2717 .o_postrecov = mds_postrecov,
2718 .o_statfs = mds_obd_statfs,
2719 .o_iocontrol = mds_iocontrol,
2720 .o_create = mds_obd_create,
2721 .o_destroy = mds_obd_destroy,
2722 .o_llog_init = mds_llog_init,
2723 .o_llog_finish = mds_llog_finish,
2724 .o_notify = mds_notify,
2725 .o_health_check = mds_health_check,
2726 .o_process_config = mds_process_config,
2729 static struct obd_ops mdt_obd_ops = {
2730 .o_owner = THIS_MODULE,
2731 .o_setup = mdt_setup,
2732 .o_cleanup = mdt_cleanup,
2733 .o_health_check = mdt_health_check,
2736 quota_interface_t *mds_quota_interface_ref;
2737 extern quota_interface_t mds_quota_interface;
2739 static __attribute__((unused)) int __init mds_init(void)
2742 struct lprocfs_static_vars lvars;
2744 request_module("lquota");
2745 mds_quota_interface_ref = PORTAL_SYMBOL_GET(mds_quota_interface);
2746 rc = lquota_init(mds_quota_interface_ref);
2748 if (mds_quota_interface_ref)
2749 PORTAL_SYMBOL_PUT(mds_quota_interface);
2752 init_obd_quota_ops(mds_quota_interface_ref, &mds_obd_ops);
2754 lprocfs_init_vars(mds, &lvars);
2755 class_register_type(&mds_obd_ops, NULL,
2756 lvars.module_vars, LUSTRE_MDS_NAME, NULL);
2757 lprocfs_init_vars(mdt, &lvars);
2758 mdt_obd_ops = mdt_obd_ops; //make compiler happy
2759 // class_register_type(&mdt_obd_ops, NULL,
2760 // lvars.module_vars, LUSTRE_MDT_NAME, NULL);
2765 static __attribute__((unused)) void /*__exit*/ mds_exit(void)
2767 lquota_exit(mds_quota_interface_ref);
2768 if (mds_quota_interface_ref)
2769 PORTAL_SYMBOL_PUT(mds_quota_interface);
2771 class_unregister_type(LUSTRE_MDS_NAME);
2772 // class_unregister_type(LUSTRE_MDT_NAME);
2774 /*mds still need lov setup here*/
2775 static int mds_cmd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
2777 struct mds_obd *mds = &obd->u.mds;
2778 struct lvfs_run_ctxt saved;
2780 struct vfsmount *mnt;
2781 struct lustre_sb_info *lsi;
2782 struct lustre_mount_info *lmi;
2783 struct dentry *dentry;
2788 CDEBUG(D_INFO, "obd %s setup \n", obd->obd_name);
2789 if (strncmp(obd->obd_name, MDD_OBD_NAME, strlen(MDD_OBD_NAME)))
2792 if (lcfg->lcfg_bufcount < 5) {
2793 CERROR("invalid arg for setup %s\n", MDD_OBD_NAME);
2796 dev = lustre_cfg_string(lcfg, 4);
2797 lmi = server_get_mount(dev);
2798 LASSERT(lmi != NULL);
2800 lsi = s2lsi(lmi->lmi_sb);
2802 /* FIXME: MDD LOV initialize objects.
2803 * we need only lmi here but not get mount
2804 * OSD did mount already, so put mount back
2806 atomic_dec(&lsi->lsi_mounts);
2809 obd->obd_fsops = fsfilt_get_ops(MT_STR(lsi->lsi_ldd));
2810 mds_init_ctxt(obd, mnt);
2812 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2813 dentry = simple_mkdir(current->fs->pwd, "OBJECTS", 0777, 1);
2814 if (IS_ERR(dentry)) {
2815 rc = PTR_ERR(dentry);
2816 CERROR("cannot create OBJECTS directory: rc = %d\n", rc);
2817 GOTO(err_putfs, rc);
2819 mds->mds_objects_dir = dentry;
2821 dentry = lookup_one_len("__iopen__", current->fs->pwd,
2822 strlen("__iopen__"));
2823 if (IS_ERR(dentry)) {
2824 rc = PTR_ERR(dentry);
2825 CERROR("cannot lookup __iopen__ directory: rc = %d\n", rc);
2826 GOTO(err_objects, rc);
2829 mds->mds_fid_de = dentry;
2830 if (!dentry->d_inode || is_bad_inode(dentry->d_inode)) {
2832 CERROR("__iopen__ directory has no inode? rc = %d\n", rc);
2836 /* open and test the lov objd file */
2837 file = filp_open(LOV_OBJID, O_RDWR | O_CREAT, 0644);
2840 CERROR("cannot open/create %s file: rc = %d\n", LOV_OBJID, rc);
2841 GOTO(err_fid, rc = PTR_ERR(file));
2843 mds->mds_lov_objid_filp = file;
2844 if (!S_ISREG(file->f_dentry->d_inode->i_mode)) {
2845 CERROR("%s is not a regular file!: mode = %o\n", LOV_OBJID,
2846 file->f_dentry->d_inode->i_mode);
2847 GOTO(err_lov_objid, rc = -ENOENT);
2850 rc = mds_lov_presetup(mds, lcfg);
2852 GOTO(err_objects, rc);
2854 /* Don't wait for mds_postrecov trying to clear orphans */
2855 obd->obd_async_recov = 1;
2856 rc = mds_postsetup(obd);
2857 /* Bug 11557 - allow async abort_recov start
2858 FIXME can remove most of this obd_async_recov plumbing
2859 obd->obd_async_recov = 0;
2863 GOTO(err_objects, rc);
2865 mds->mds_max_mdsize = sizeof(struct lov_mds_md);
2866 mds->mds_max_cookiesize = sizeof(struct llog_cookie);
2869 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2872 if (mds->mds_lov_objid_filp &&
2873 filp_close((struct file *)mds->mds_lov_objid_filp, 0))
2874 CERROR("can't close %s after error\n", LOV_OBJID);
2876 dput(mds->mds_fid_de);
2878 dput(mds->mds_objects_dir);
2880 fsfilt_put_ops(obd->obd_fsops);
2884 static int mds_cmd_cleanup(struct obd_device *obd)
2886 struct mds_obd *mds = &obd->u.mds;
2887 struct lvfs_run_ctxt saved;
2892 LCONSOLE_WARN("%s: shutting down for failover; client state "
2893 "will be preserved.\n", obd->obd_name);
2895 push_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2896 if (mds->mds_lov_objid_filp) {
2897 rc = filp_close((struct file *)mds->mds_lov_objid_filp, 0);
2898 mds->mds_lov_objid_filp = NULL;
2900 CERROR("%s file won't close, rc=%d\n", LOV_OBJID, rc);
2902 if (mds->mds_objects_dir != NULL) {
2903 l_dput(mds->mds_objects_dir);
2904 mds->mds_objects_dir = NULL;
2907 if (mds->mds_lov_objids != NULL)
2908 OBD_FREE(mds->mds_lov_objids, mds->mds_lov_objids_size);
2910 shrink_dcache_parent(mds->mds_fid_de);
2911 dput(mds->mds_fid_de);
2912 LL_DQUOT_OFF(obd->u.obt.obt_sb);
2913 fsfilt_put_ops(obd->obd_fsops);
2915 pop_ctxt(&saved, &obd->obd_lvfs_ctxt, NULL);
2920 static int mds_cmd_health_check(struct obd_device *obd)
2925 static struct obd_ops mds_cmd_obd_ops = {
2926 .o_owner = THIS_MODULE,
2927 .o_setup = mds_cmd_setup,
2928 .o_cleanup = mds_cmd_cleanup,
2929 .o_precleanup = mds_precleanup,
2930 .o_create = mds_obd_create,
2931 .o_destroy = mds_obd_destroy,
2932 .o_llog_init = mds_llog_init,
2933 .o_llog_finish = mds_llog_finish,
2934 .o_notify = mds_notify,
2935 .o_postrecov = mds_postrecov,
2936 // .o_health_check = mds_cmd_health_check,
2939 static int __init mds_cmd_init(void)
2941 struct lprocfs_static_vars lvars;
2943 lprocfs_init_vars(mds, &lvars);
2944 class_register_type(&mds_cmd_obd_ops, NULL, lvars.module_vars,
2945 LUSTRE_MDS_NAME, NULL);
2950 static void /*__exit*/ mds_cmd_exit(void)
2952 class_unregister_type(LUSTRE_MDS_NAME);
2955 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
2956 MODULE_DESCRIPTION("Lustre Metadata Server (MDS)");
2957 MODULE_LICENSE("GPL");
2959 module_init(mds_cmd_init);
2960 module_exit(mds_cmd_exit);